A method for detecting pure natural royal jelly freeze-dried powder

Abstract

The application discloses a kind of pure natural royal jelly freeze-dried powder detection methods, it is related to quality detection technical field, the method includes the following steps: collecting the multi-dimensional original data of royal jelly freeze-dried powder, data is handled to generate data matrix, from data matrix extraction key feature index set, to index set is analyzed and evaluated to generate authenticity score, generate digital detection report and quality spectrum, based on detection result is continuously optimized.The application, by fusing near-infrared spectrum, microscopic image, physicochemical activity and block chain traceability and other multi-source heterogeneous data, intelligent evaluation is carried out using knowledge graph and lightweight multi-modal network, realizes the multi-dimensional, high-precision analysis of the pure nature, naturalness and adulteration of royal jelly freeze-dried powder.

Description

Technical Field

[0001] This invention relates to the field of quality testing technology, and in particular to a method for testing pure natural royal jelly freeze-dried powder. Background Technology

[0002] Pure natural royal jelly freeze-dried powder is made from fresh royal jelly using modern freeze-drying technology. It can be stably stored at room temperature for a long time, preserving the active nutrients of royal jelly to the maximum extent. Its core functions are to enhance immunity, combat fatigue, provide antioxidants, and help regulate the nervous system and bodily functions, making it a highly effective and convenient natural tonic.

[0003] Currently, the quality testing of pure natural royal jelly freeze-dried powder mainly relies on experience-based judgment, single instrument analysis, or simple comparison of physicochemical indicators. It is difficult to deeply integrate and synergistically analyze the multi-source heterogeneous characteristics of the sample (such as chemical composition, microstructure, and bioactivity), resulting in a single detection dimension and one-sided identification characteristics, making it impossible to comprehensively and objectively assess its purity and naturalness.

[0004] Therefore, a method for detecting pure natural royal jelly freeze-dried powder is proposed to solve the above problems. Summary of the Invention

[0005] The main objective of this invention is to provide a method for detecting pure natural royal jelly freeze-dried powder, so as to solve the problems mentioned in the background above.

[0006] To achieve the above objectives, the technical solution adopted by the present invention is as follows: a method for detecting pure natural royal jelly freeze-dried powder, the method comprising the following steps:

[0007] S1. Collect multi-dimensional raw data of the freeze-dried royal jelly powder to be tested, including near-infrared spectral data, visible light microscopic image data, and physicochemical activity index data;

[0008] S2. Preprocess and extract features from the multi-dimensional raw data to generate a standardized fused data matrix;

[0009] S3. Based on the royal jelly quality knowledge graph, extract a set of key feature indicators related to purity, freshness and adulteration from the standardized fusion data matrix;

[0010] S4. Construct and utilize a lightweight multimodal fusion network to perform parallel analysis and authenticity evaluation of the key feature index set, and output preliminary scores for each dimension.

[0011] S5. Based on the credible traceability database and the historical best product database, the initial score is dynamically weighted and calibrated to generate a comprehensive authenticity score.

[0012] S6. Based on the analysis results of the comprehensive authenticity score and key feature index set, generate a digital inspection report and quality map with traceability chain;

[0013] S7. Based on batch detection results, continuously optimize the parameters and dynamic weighting strategy of the multimodal fusion network, and update the adulteration feature library and threshold.

[0014] Preferably, the collection of multi-dimensional raw data in step S1 includes the following steps:

[0015] S11. Use a near-infrared spectrometer to collect the chemical composition index data of the freeze-dried powder and record the characteristic absorption peaks;

[0016] S12. Take microscopic images of the freeze-dried powder using a digital microscope, including crystal morphology and color distribution;

[0017] S13. Use a rapid test kit to obtain basic physicochemical data of the reconstituted lyophilized powder.

[0018] Preferably, the preprocessing and feature extraction in S2 includes the following steps:

[0019] S21. Smooth, baseline correct and normalize the near-infrared spectral data, perform background segmentation, contrast enhancement and size normalization on the microscopic image, and extract local binary pattern texture features.

[0020] S22. Compare the physicochemical activity index data with the standard reference values ​​and calculate the relative activity deviation.

[0021] S23. The processed spectral feature vector, image morphology feature vector, and activity feature vector are concatenated according to the sample ID to construct a standardized fusion data matrix.

[0022] Preferably, the extraction of the key feature index set in S3 includes the following steps:

[0023] S31. Construct a knowledge graph of royal jelly quality. The nodes of the knowledge graph of royal jelly quality include the standard component range, typical morphological characteristics, activity index thresholds and known adulteration patterns.

[0024] S32. Using the Node2Vec graph embedding method, the standardized fusion data matrix is ​​mapped to the knowledge graph, and core features and abnormal features are identified through similarity matching.

[0025] S33. Extract the feature peak area ratio, texture consistency parameters calculated based on the local binary mode variance of the image, and the average relative deviation of the physicochemical indicators relative to the standard values ​​as the activity deviation degree, and encode them into a structured feature vector.

[0026] Preferably, the evaluation using a lightweight multimodal fusion network in step S4 includes the following steps:

[0027] S41. Construct a dual-branch lightweight network, in which a one-dimensional convolutional layer processes spectral features, a lightweight CNN processes image features, and active features are processed through a fully connected layer.

[0028] S42. Input the structured feature vector into the corresponding branch and calculate the initial scores for component authenticity based on spectral and activity features and structural regularity based on image morphological features, respectively.

[0029] S43. The primary scores output from each branch are concatenated, fused through a fully connected layer, and then normalized using the Sigmoid function to output the normalized primary scores for each dimension.

[0030] Preferably, the construction of the lightweight multimodal fusion network in step S4 further includes the following steps:

[0031] S411. Initialize the image branch weights using pre-trained MobileNetV2 and fine-tune them with a small number of samples;

[0032] S412. Introduce data augmentation to improve network robustness and reduce reliance on synthetic data;

[0033] S413. Use Grad-CAM class activation mapping to generate visual heatmaps of key feature regions to improve interpretability.

[0034] Preferably, the dynamic weighting and confidence calibration in step S5 includes the following steps:

[0035] S51. Retrieve basic traceability information of raw materials from the trusted traceability database and calculate the simple trust score;

[0036] S52. Query the historical best product database to obtain benchmark data of similar honey sources and seasons as a reference.

[0037] S53. Using the primary score, credibility score, and the Euclidean distance between the current sample's key features and historical benchmark features as the benchmark deviation, a weighted average fusion method is used to calculate the comprehensive authenticity score.

[0038] Preferably, the weighted average fusion method in step S53 further includes the following steps:

[0039] S531. Assign time decay weights to recent batch data;

[0040] S532. For missing or contradictory source data, use the mean or mode to fill in the missing data;

[0041] S533. When the score is seriously inconsistent with the traceability, the manual review process is triggered and the abnormal case is recorded.

[0042] Preferably, the generation of the digital inspection report and quality map in step S6 includes the following steps:

[0043] S61. Based on the score and feature set, a rule matching method is used to identify potential quality defects;

[0044] S62. Integrate detection results, feature data and source tracing information to generate a static visualization map;

[0045] S63. Convert the results into a structured report template and generate a report file containing a unique hash value for verification.

[0046] Preferably, the optimization model and update strategy in S7 includes the following steps:

[0047] S71. Based on the expert review results, the network parameters are updated using the elastic weight consolidation incremental learning algorithm.

[0048] S72. When abnormal patterns occur continuously, the adulteration feature library and thresholds are updated after manual review.

[0049] S73. The updated model and feature library are periodically pushed to the detection terminal to form a semi-automatic iterative system.

[0050] The present invention has the following beneficial effects:

[0051] 1. In this invention, by systematically fusing multi-source heterogeneous data such as near-infrared spectroscopy, microscopic images, and physicochemical activity indicators, and extracting key identification features based on the royal jelly quality knowledge graph, it is possible to achieve synergistic analysis and comprehensive quantification of the chemical composition, physical morphology, and biological activity of samples. This breaks through the limitations of traditional methods that rely on a single detection dimension, and solves the problem of one-sided identification dimensions and difficulty in comprehensively assessing the purity and naturalness of products caused by relying on a single detection method, thereby improving the objectivity and comprehensiveness of the detection.

[0052] 2. In this invention, by introducing a dynamic weighted calibration mechanism that combines blockchain traceability information with historical benchmark data, the initial score output by the multimodal network is reliably corrected and personalized. This allows the final comprehensive authenticity score to dynamically adapt to the differences in the characteristics of raw materials from different origins and seasons. This overcomes the shortcomings of existing technologies that use fixed thresholds or standards, resulting in rigid evaluation and an inability to integrate reliable external information. It also solves the problems of rigidity in traditional fixed threshold evaluation methods and the inability to effectively utilize traceability information for accurate judgment, thereby enhancing the accuracy and reliability of the test results.

[0053] 3. In this invention, by generating digital inspection reports and visualized quality maps with complete traceability chains, and continuously adaptively optimizing the inspection model and feature library based on batch feedback results, it can provide intuitive and auditable decision-making basis for quality control, and ensure that the system's inspection capabilities continue to evolve with new adulteration patterns. This changes the passive situation in traditional inspection processes where report formats are singular, system updates rely on manual experience, and it is difficult to cope with new adulteration methods. It solves the problems of delayed feedback and rigid system that cannot improve itself in traditional methods, and improves the practicality and intelligence level of the inspection system. Attached Figure Description

[0054] Figure 1 This is a flowchart of a method for detecting pure natural royal jelly freeze-dried powder according to the present invention;

[0055] Figure 2 This is a flowchart illustrating the process of collecting multi-dimensional raw data in a method for detecting pure natural royal jelly freeze-dried powder according to the present invention.

[0056] Figure 3 This is a flowchart of the preprocessing and feature extraction process in a method for detecting pure natural royal jelly freeze-dried powder according to the present invention.

[0057] Figure 4 This is a flowchart illustrating the extraction of key feature index set in a method for detecting pure natural royal jelly freeze-dried powder according to the present invention.

[0058] Figure 5 This is a flowchart illustrating the evaluation using a lightweight multimodal fusion network in a method for detecting pure natural royal jelly freeze-dried powder according to the present invention.

[0059] Figure 6 This is a flowchart of dynamic weighting and reliability calibration in a method for detecting pure natural royal jelly freeze-dried powder according to the present invention.

[0060] Figure 7 This is a flowchart illustrating the generation of a digital test report and quality spectrum in a method for detecting pure natural royal jelly freeze-dried powder according to the present invention.

[0061] Figure 8 This is a flowchart of the optimization model and update strategy in the detection method of pure natural royal jelly freeze-dried powder of the present invention. Detailed Implementation

[0062] 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.

[0063] Please see Figures 1-8 This invention provides a technical solution: a method for detecting pure natural royal jelly freeze-dried powder, the method comprising the following steps:

[0064] S1. Collect multi-dimensional raw data of the freeze-dried royal jelly powder to be tested, including near-infrared spectral data, visible light microscopic image data, and physicochemical activity index data;

[0065] S2. Preprocess and extract features from the multi-dimensional raw data to generate a standardized fused data matrix;

[0066] S3. Based on the royal jelly quality knowledge graph, extract a set of key feature indicators related to purity, freshness and adulteration from the standardized fusion data matrix;

[0067] S4. Construct and utilize a lightweight multimodal fusion network to perform parallel analysis and authenticity evaluation of the key feature index set, and output preliminary scores for each dimension.

[0068] S5. Based on the credible traceability database and the historical best product database, the initial score is dynamically weighted and calibrated to generate a comprehensive authenticity score.

[0069] S6. Based on the analysis results of the comprehensive authenticity score and key feature index set, generate a digital inspection report and quality map with traceability chain;

[0070] S7. Based on batch detection results, continuously optimize the parameters and dynamic weighting strategy of the multimodal fusion network, and update the adulteration feature library and threshold.

[0071] The process of collecting multi-dimensional raw data in S1 includes the following steps:

[0072] S11. Use a near-infrared spectrometer to collect the chemical composition index data of the freeze-dried powder and record the characteristic absorption peaks;

[0073] S111. Set the scanning range of the near-infrared spectrometer to 10000 cm⁻¹. -1 Up to 4000cm -1 The resolution is 8cm. -1 Each sample was scanned 32 times to improve the signal-to-noise ratio;

[0074] S112. Collect diffuse reflectance spectral data of the sample and use the built-in background subtraction algorithm to eliminate ambient light interference to obtain the original absorbance spectral curve. ,in Wave number;

[0075] S113. Automatically identify and record the location at 5165cm from the spectral curve. -1 (Moisture-OH first-order overtone), 6920cm-1 (Protein / Amino Acid NH Combination Frequency) and 8250cm -1 The core characteristic absorption peak near the (second harmonic of lipid CH) is calculated, and its peak height is calculated. With half-peak width The calculation formula is as follows:

[0076] Peak height: ,in The absorbance at the peak point. The baseline absorbance is determined by interpolation.

[0077] Half-peak width: ,in and These are the values ​​of the two waves corresponding to half the peak height;

[0078] S12. Take microscopic images of the freeze-dried powder using a digital microscope, including crystal morphology and color distribution;

[0079] S13. Use a rapid test kit to obtain basic physicochemical data of the reconstituted lyophilized powder.

[0080] Preprocessing and feature extraction in S2 include the following steps:

[0081] S21. Smooth, baseline correct, and normalize the near-infrared spectral data; perform background segmentation, contrast enhancement, and size normalization on the microscopic images; and extract local binary pattern texture features, including the following steps:

[0082] S211. The collected raw near-infrared spectral data The Savitzky-Golay convolutional smoothing algorithm was used for noise reduction. A window width of 5 points and a second-order polynomial were used for fitting. The smoothing formula is as follows:

[0083] ;

[0084] in, For the smoothed first The absorbance of each data point For the pre-calculated convolution coefficients, Half the width of the window (here) );

[0085] Adaptive baseline correction: Fitting the baseline using an asymmetric least squares algorithm Minimize the weighted error term iteratively Obtain the baseline, where the weights When the data point is higher than the previous fitted baseline, a smaller value p is used (e.g., 0.001); otherwise, 1 is used. The corrected spectrum is... ;

[0086] S213. Normalize the baseline-corrected spectrum using standard normal variables to eliminate the effects of optical path scattering:

[0087] ;

[0088] in, and These represent the mean and standard deviation of absorbance of a single sample spectrum at all wavenumber points, respectively.

[0089] S214. For digital micrographs, an adaptive thresholding algorithm based on the Otsu method is used for background segmentation to extract the crystal foreground region; a contrast-limited adaptive histogram equalization algorithm is used to enhance the contrast within the foreground region; finally, all images are uniformly scaled to 512×512 pixels, and bicubic interpolation is used to preserve geometric features.

[0090] S22. Compare the physicochemical activity index data with the standard reference values ​​and calculate the relative activity deviation, including the following steps:

[0091] S221. Extract measured values ​​of key activity indicators from rapid detection results. (such as 10-HDA content, protein content, and total sugar content), among which The total number of indicators;

[0092] S222. Consult national or industry standards to obtain the standard reference values ​​for the corresponding activity indicators. and its allowed floating range ;

[0093] S223. Calculate the relative activity deviation for each indicator. This value is a normalized deviation measure, calculated using the following formula:

[0094] ;

[0095] S224, relative activity deviation of all indicators Combining to form an active feature vector ;

[0096] S23. The processed spectral feature vector, image morphological feature vector, and activity feature vector are concatenated according to sample ID to construct a standardized fusion data matrix, including the following steps:

[0097] S231, Spectrum normalized from SNV In the preset The average absorbance is extracted from each characteristic wavenumber band (such as the absorption band corresponding to proteins, carbohydrates, and water) to form a spectral feature vector. ;

[0098] S232. From the normalized microscopic image, extract the morphological features (such as equivalent circle diameter, roundness, and aspect ratio) and color features (mean and standard deviation of each channel in L*a*b* space) of each crystal region. Calculate the statistics (mean and standard deviation) for all crystals to construct the image morphological feature vector. ;

[0099] S233, Regarding spectral feature vectors Image morphological feature vector and active feature vector Perform Min-Max normalization on each, mapping them to the interval [0, 1]:

[0100] ;

[0101] in Represents any original feature vector. This is the normalized vector used to eliminate differences in numerical scales among different features, prevent certain features with large numerical ranges from dominating in subsequent model training, and ensure that multimodal features can be fairly and effectively spliced ​​and fused.

[0102] S234. Concatenate the three normalized feature vectors corresponding to the same sample ID sequentially to obtain the fused feature vector of that sample. ;

[0103] S235. Arrange the fused feature vectors of all N samples by row, and finally construct a vector with dimension [missing information]. Standardized fusion data matrix .

[0104] Extracting the key feature index set from S3 includes the following steps:

[0105] S31. Construct a knowledge graph of royal jelly quality. The nodes of the knowledge graph include standard component ranges, typical morphological characteristics, activity index thresholds, and known adulteration patterns. This includes the following steps:

[0106] S311. Define the node entities of the knowledge graph, which mainly include standard components (such as 10-HDA content standard range of 1.4%-2.0%), typical forms (such as pure freeze-dried powder crystal roundness >0.85), activity indicators (such as protein content threshold ≥15%) and known adulteration patterns (such as starch adulteration, syrup adulteration).

[0107] S312. Define semantic relationship edges between nodes, including composition-morphology association (e.g., high 10-HDA usually corresponds to regular crystal morphology), composition-activity association, and morphology-adulteration indicator (e.g., irregular large particles may indicate starch adulteration).

[0108] S313. After structuring authoritative standards, historical testing big data, and expert experience knowledge, store them in a graph database to form a computable quality knowledge graph containing attributes, relationships, and weights. ,in For a set of nodes, Let be the set of edges. This is a matrix of node attributes.

[0109] S32. Using the Node2Vec graph embedding method, the standardized fusion data matrix is ​​mapped to a knowledge graph, and core features and abnormal features are identified through similarity matching, including the following steps:

[0110] S321. Employ the Node2Vec graph embedding algorithm to embed the knowledge graph. Each node in Mapped to a low-dimensional dense vector ( For embedding dimensions (e.g., 128-dimensional), the algorithm generates a sequence of nodes through a biased random walk and optimizes the following objective function using a Skip-gram model:

[0111] ;

[0112] in For mapping functions, For nodes Through strategy The sampled set of network neighbor nodes;

[0113] S322, Regarding standardized fusion data matrices Each sample feature vector The sample embedding vector is obtained by mapping it to the same embedding space as the graph nodes through a fully connected projection network. ,in and These are learnable parameters;

[0114] S323, Calculate the sample embedding vector Embedded vectors of various standard nodes (such as pure standard and starch adulteration patterns) in the knowledge graph Cosine similarity between them:

[0115] ;

[0116] Calculate the similarity between sample features and various standards or patterns in the graph in the direction of vector space. The closer the value is to 1, the more consistent the direction is, indicating that the sample features are more consistent with the quality attributes or defect patterns represented by the node.

[0117] S324. Mark the top-K feature dimensions with the highest similarity to pure standard class nodes as core features; mark feature dimensions with similarity to various adulterated pattern nodes exceeding a preset threshold (e.g., 0.7) as anomalous features, and generate a core feature index set. With anomaly feature index set ;

[0118] S33. Extract the feature peak area ratio, texture consistency parameters calculated based on the variance of the local binary mode of the image, and the average relative deviation of the physicochemical indicators relative to the standard values ​​as the activity deviation, and encode them into a structured feature vector, including the following steps:

[0119] S331. From the near-infrared spectral data, calculate the characteristic peak area ratio for the bands related to the identified core and abnormal features. For example, calculate the protein characteristic peak (approximately 6920 cm⁻¹). -1 (Nearby) area With characteristic peaks of carbohydrates (approximately 5800 cm⁻¹) -1 (Nearby) area ratio :

[0120] ;

[0121] The upper and lower limits of integration The peak area is the area under the characteristic absorption peak of a specific chemical component (such as protein or carbohydrate) calculated by integration. Peak area is a more stable indicator of the total amount of the component than peak height because the integral is not sensitive to small shifts in the spectrum and noise. The ratio of characteristic peak area (e.g., protein peak area / carbohydrate peak area) is an important indicator for identifying whether the component ratio is abnormal.

[0122] S332. From the microscopic image, on the segmented crystal foreground region, calculate the texture consistency parameter based on the local binary pattern variance. First, calculate the LBP value of each pixel, and then calculate the variance of the LBP values ​​of the entire foreground region. As a measure of texture consistency, a smaller variance indicates a more uniform and consistent texture.

[0123] ;

[0124] in This represents the total number of foreground pixels. The mean of LBP values;

[0125] S333. Calculate the average relative deviation of physicochemical indicators, and the relative deviation of each activity indicator based on the calculation. Calculate the average absolute value of the deviation of those indicators identified as core activity characteristics (such as 10-HDA, protein). :

[0126] ;

[0127] in The set of activity indexes belonging to the core characteristics is used to calculate the absolute average of their deviations, resulting in a comprehensive activity quality index. A low value indicates that the core active ingredients are close to the standard as a whole, while a high value indicates that there are common problems with key nutrients.

[0128] S334. The extracted quantitative indicators, including the characteristic peak area ratio, are... Texture consistency parameters Average relative deviation and from the core feature set and abnormal feature set Other key statistics selected (such as mean and quantiles) are collectively encoded into a fixed-dimensional structured feature vector. ( (where is the total number of features), which serves as the input for subsequent network evaluation.

[0129] Evaluation using a lightweight multimodal fusion network in S4 includes the following steps:

[0130] S41. Construct a lightweight dual-branch network, where a one-dimensional convolutional layer processes spectral features, a lightweight CNN processes image features, and active features are processed through a fully connected layer, including the following steps:

[0131] S411. Construct the spectral feature processing branch: This branch receives data from the structured feature vector. Spectral correlation section Design a lightweight one-dimensional convolution module with the following structure: a one-dimensional convolutional layer (kernel size = 3, stride = 1, number of output channels = 16, using ReLU activation function), a one-dimensional max pooling layer (pooling size = 2, stride = 2), and a Dropout layer (dropout rate = 0.2). The convolution operation formula is as follows:

[0132] ;

[0133] in This represents a one-dimensional convolution operation. and The weights and biases of the convolutional layers are used to capture the local correlations between spectral bands. ReLU introduces nonlinearity, enabling the network to learn complex mapping relationships and provide a high-level spectral feature representation for subsequent scoring.

[0134] S412. Constructing the image feature processing branch: This branch receives data from the structured feature vector. Image morphology related parts A lightweight fully connected network module is used, with the following structure: a fully connected layer (output dimension = 64, ReLU activation), a batch normalization layer, and a Dropout layer (dropout rate = 0.2). Since the high-dimensional features of the image have already been extracted and compressed in the previous steps... Here, a fully connected layer is used for efficient mapping;

[0135] S413. Construct an activity feature processing submodule: This module receives activity-related feature data. The initial transformation is performed through a fully connected layer (output dimension = 32, ReLU activation);

[0136] S42. Input the structured feature vector into the corresponding branch, and calculate the initial scores for component authenticity based on spectral and activity features and structural regularity based on image morphological features, including the following steps:

[0137] S421, The feature vector Enter the corresponding processing module for each;

[0138] S422. After processing the spectral branch and activity module, the obtained features are concatenated and input into a component authenticity assessment subnetwork. This subnetwork consists of a fully connected layer (output dimension = 1) and is used to calculate a primary score reflecting whether the chemical components and active substances meet the standards of natural purity. The calculation process can be expressed as follows:

[0139] ;

[0140] in, Flatten the convolution output This represents the output of the active feature processing submodule. and As weights and biases, the extracted high-level spectral features and activity features are fused and comprehensively judged through a fully connected layer to output an original score, which reflects the network's preliminary judgment on the authenticity of sample components based on chemical composition and active substance information.

[0141] S423, Features of image branch output The input is fed into a structural regularity evaluation subnetwork, which also consists of a fully connected layer (output dimension = 1) to calculate a primary score reflecting the regularity and uniformity of the microcrystalline morphology and color distribution. :

[0142] ;

[0143] S43. Concatenate the primary scores output from each branch, fuse them through a fully connected layer, and then normalize them using the Sigmoid function to output the normalized primary scores for each dimension. This includes the following steps:

[0144] S431, Transfer the two initial scores and Concatenate into a two-dimensional vector ;

[0145] S432, will Nonlinear fusion and information interaction are performed through a shared feature fusion fully connected layer (output dimension = 8, ReLU activation) to obtain fused features. ;

[0146] S433, will Two independent scoring output heads are input, each consisting of a fully connected layer with a single neuron, which respectively generate the final scalar outputs for the component authenticity dimension and the structural regularity dimension. and ;

[0147] S434. Apply the Sigmoid function to normalize each scalar output, mapping it to the interval [0, 1], to obtain the final normalized primary score. and :

[0148] ;

[0149] S435. Output the primary score vector composed of these two normalized scores. This serves as the input for subsequent dynamic weighted calibration.

[0150] The construction of the lightweight multimodal fusion network in S4 also includes the following steps:

[0151] S411. Initialize the image branch weights using pre-trained MobileNetV2 and fine-tune them with a small number of samples;

[0152] S4111. Obtain a MobileNetV2 model pre-trained on a large natural image dataset (such as ImageNet), remove its top classifier (global average pooling layer and subsequent fully connected layers), and retain its convolutional basis as a feature extractor. The weight matrix of this convolutional basis is denoted as... ;

[0153] S4112. For the microscopic image feature processing task, after the MobileNetV2 convolutional base, a custom adaptation layer sequence is constructed, including a global average pooling layer and a fully connected layer (output dimension = 64, ReLU activation, corresponding to the image branch fully connected layer in S41). The weights of this fully connected layer are denoted as... Random initialization;

[0154] S4113. In the fine-tuning phase, a two-stage training strategy is adopted: first, the weights of the MobileNetV2 convolutional bases are frozen. Only train the weights of the custom adaptation layer And the subsequent part of the network (the component and structure evaluation sub-network), with a learning rate set to (e.g., 0.01); then unfreeze the weights of the last few layers of the convolutional base (e.g., the last 3 inverse residual modules) with a lower learning rate. (like This layer, along with the adaptation layer and subsequent networks, undergoes end-to-end fine-tuning to adapt to the microscopic morphological characteristics of freeze-dried royal jelly powder. The loss function used is mean squared error loss.

[0155] ;

[0156] in and For expert-annotated samples The true component and structure score is calculated by the formula to determine the difference between the model's predicted score and the expert's true score. During model training / fine-tuning, this loss value is minimized by optimization algorithms (such as gradient descent), thereby backpropagating the error and adjusting the network weight parameters so that the model's prediction results continuously approach the expert's judgment criteria, thus improving the model's evaluation accuracy.

[0157] S412. Introduce data augmentation to improve network robustness and reduce reliance on synthetic data;

[0158] S4121. During the training phase, the input image morphological feature vector... The corresponding original image data before preprocessing is subjected to online data augmentation. The main augmentation operations include:

[0159] Random rotation: angle range ;

[0160] Random horizontal and vertical translation: Translation range within the image size Inside;

[0161] Random brightness and contrast adjustment: Brightness adjustment factor in Uniform sampling within the area, contrast adjustment factor at Uniform sampling within the interior;

[0162] S4122. For the enhanced image, re-execute the background segmentation and feature extraction process to generate the enhanced image morphological feature vector. Used to replace the original The input network is trained dynamically during each training iteration, thereby significantly expanding the effective training samples and improving the model's robustness to minor differences in image acquisition (such as focal length and illumination).

[0163] S4123. For spectral and activity data, introduce slight Gaussian noise as an enhancement method: ,in , Let the standard deviation of the corresponding eigenvector be denoted as . 0.01 times;

[0164] S413. Use Grad-CAM class activation mapping to generate visual heatmaps of key feature regions to improve interpretability, including the following steps:

[0165] S4131. After the network has completed training and performed forward propagation on a test sample, select the evaluation dimension that needs to be interpreted (such as structural regularity score). The feature map output by the corresponding last convolutional layer (for the image branch, i.e., the last layer of the MobileNetV2 convolutional base) and the corresponding last convolutional layer. ,in For the number of channels, and The feature map space size;

[0166] S4132. Calculate the target score. For feature maps gradient The gradient is obtained through backpropagation and is denoted as . ;

[0167] S4133, For each channel Calculate gradient weights This weight represents the first The importance of each feature channel to the target score: ;

[0168] in It is the gradient In the passage Spatial location The value at a given location is calculated by averaging the gradients at all locations in each channel c on feature map A to measure the overall contribution of that channel. The larger the gradient, the more sensitive the feature activation of that channel is to changes in the final score, and the higher its importance.

[0169] S4134, Calculation-based activation heatmap By weighted summation of the feature maps of all channels and applying the ReLU function, regions that positively contribute to the score are highlighted:

[0170] ;

[0171] in It is a feature map In the passage ,Location The activation values ​​are used to filter out regions that negatively contribute to the target score using the ReLU function. The resulting heatmap L_Grad-CAM visually highlights which regions in the input image play a key role in the model's judgments on structural regularity, greatly improving the interpretability and transparency of the deep learning model's decision-making process and enhancing the credibility of the results.

[0172] S4135, Generate the heatmap By upsampling to the size of the original input image using bilinear interpolation and superimposing it onto the original microscopic image, a visualization result is generated. This result can intuitively show which regions of the image (such as specific crystal shapes or texture regions) the network focuses on when evaluating structural regularity, thereby improving the transparency and credibility of the model's decision-making process.

[0173] The dynamic weighting and credibility calibration in S5 includes the following steps:

[0174] S51. Retrieve basic traceability information of raw materials from the trusted traceability database and calculate a simple trust score, including the following steps:

[0175] S511. Query the trusted traceability database built on blockchain to retrieve the traceability information tuple of the raw royal jelly corresponding to the batch of the sample to be tested. Including honey source areas Harvesting season Bee species Processing plant Temperature and humidity records for transportation and storage environment And the verification status of digital signatures at each stage. ;

[0176] S512, Design a set of credibility assessment rules and assign a basic weight to each rule. and state factors (0 or 1), calculate the initial credibility score :

[0177] ;

[0178] The rules are as follows:

[0179] Is the honey source area on the certified list? Yes, 0 (No);

[0180] Does the harvesting season coincide with the high-yield and high-quality period of this honey-producing area? );

[0181] : Have all digital signatures at each stage been verified? );

[0182] Is the average temperature during transportation and storage... ( );

[0183] Is the information chain complete and without any missing parts? );

[0184] S513, Introduce a confidence factor based on information integrity. (Value range 0.8-1.0), the initial score is corrected to obtain the simple confidence score. :

[0185] ;

[0186] The actual number of traceability information items provided. To find the total number of items, Factors are assigned to the initial score based on information completeness. The more incomplete the information, the more it is reduced. The lower the value (minimum 0.8), the better the final result. The lower the score, the more it reflects the logic that incomplete information reduces overall credibility;

[0187] S52. Query the historical best product database to obtain benchmark data of similar honey sources and seasons as a reference.

[0188] S521. In the historical best product database, using the current sample's honey source location... and harvest season Historical data was retrieved as the primary filtering criterion.

[0189] S522. Calculate the comprehensive similarity between the current sample and each candidate benchmark sample in the database in terms of key environmental and time dimensions. The reciprocal of the weighted Euclidean distance is used as the metric; the higher the similarity, the smaller the distance.

[0190] ; ;

[0191] in, The latitude and longitude of the honey source area Encode the season (e.g., spring = 1). Weighting coefficients (e.g.) This transforms the concepts of similar honey sources and seasons into concrete numerical measures by calculating a weighted distance based on geographical location and seasonal coding. To measure the similarity of growth environments between samples;

[0192] S523, Select The highest front Five (e.g., N=5) historical samples are used as a similar benchmark set. The normalized primary score vectors obtained from these benchmark samples in step S4 are extracted. and the core feature part of the structured feature vector obtained from S33. Calculate the average benchmark score separately. and average benchmark core feature vector As a reference benchmark;

[0193] S53. Using the initial score, credibility score, and the Euclidean distance between the current sample's key features and historical benchmark features as the baseline deviation, a weighted average fusion method is employed to calculate the comprehensive authenticity score, including the following steps:

[0194] S531. Calculate the benchmark deviation between the current sample and the benchmark. Using the core feature vector of the current sample With the average benchmark core feature vector The Euclidean distance is calculated and normalized.

[0195] ; ;

[0196] in This is the set of Euclidean distances between all samples within the historical benchmark set. This normalization ensures that the deviation is in the interval [0,1].

[0197] S532. Setting Dynamic Weights: Weight Vector The testing can be dynamically adjusted based on the testing objectives. For example, when strictly controlling raw materials, the weight of reliability can be increased (e.g., ...). ); Increase the weighting of scores and deviations when focusing on product consistency (e.g. ), and satisfy ;

[0198] S533. Perform weighted average fusion. First, the initial score vector of the current sample... Combined into a single comprehensive primary scoring scalar For example, taking the weighted average or the minimum value (taking the minimum value is more stringent):

[0199] ;

[0200] Then, combine the credibility score (normalized to [0, 1]) and baseline deviation Calculate the overall authenticity score :

[0201] ;

[0202] in This indicates the degree of similarity to the benchmark; the greater the deviation, the smaller the contribution.

[0203] S534. Output the final overall authenticity score. (Value range [0, 1]), the higher the score, the higher the credibility of the sample as pure natural freeze-dried royal jelly powder. The formula linearly integrates evidence from different dimensions—instrument test results, raw material traceability information, and similarity with historical high-quality products—according to configurable weights.

[0204] The weighted average fusion method used in S53 also includes the following steps:

[0205] S531. Assign time decay weights to recent batches of data, including the following steps.

[0206] S5311. When querying the historical best product database in S52, append the collection timestamp to each candidate benchmark data record. ;

[0207] S5312, Let the time of the current testing batch be... Calculate the time decay factor for each candidate benchmark data point. An exponential decay model is used, which assigns higher weights to newer data, resulting in a greater impact on the benchmark.

[0208] ;

[0209] in, The attenuation rate coefficient, The time difference is measured in days;

[0210] S5313, Calculate the average benchmark score in S523. and average benchmark core feature vector In this case, a weighted average based on a time decay factor is used instead of a simple arithmetic average:

[0211] ;

[0212] This step ensures that the reference benchmark better reflects recent product quality characteristics and scoring trends, ultimately yielding... and It is not a simple historical average, but a dynamic benchmark that is closer to the recent quality level. This allows subsequent calibration (calculation of deviation) to be based on the latest high-quality products, making the evaluation more scientific.

[0213] S532. For missing or contradictory source data, imputation using the mean or mode is performed, including the following steps:

[0214] S5321, Retrieve the source information tuple in S511 Then, check each field for missing values ​​(NULL) or logical contradictions (e.g., the harvest season is "summer", but there is no flowering season in the honey source area during that season).

[0215] S5322. For missing values, the following strategy is used to fill them:

[0216] For continuous or numerical data (such as average transport temperature), the average of historical data from the same honey-producing region and season is used. Fill in the gaps: ;

[0217] For categorized data (such as bee species), the mode of historical data from the same nectar source area is used. Fill in the blanks with the categories that appear most frequently.

[0218] S5323. For logically contradictory values, activate the contradiction resolution rule engine. For example, if the harvest season does not match the flowering period of the nectar source, the system automatically queries the standard flowering period table for that nectar source and uses the season closest to the harvest time of the record and in the flowering period as the correction value. Replace;

[0219] S5324. Record all filling and correction operations, generate a data quality report, and compile the complete and consistent traceability information tuples after filling. Used for subsequent credibility scoring Calculation;

[0220] S533. When the score is significantly inconsistent with the traceability, a manual review process is triggered and the abnormal case is recorded, including the following steps:

[0221] S5331, Define the criteria for serious non-compliance, and calculate the overall authenticity score. With source tracing credibility score absolute difference :

[0222] ;

[0223] Set threshold (like ),like If the score is found to be seriously inconsistent with the source tracing, for example, Very high (raw materials are reliable) but Very low (poor test results), or vice versa;

[0224] S5332. When a critical non-compliance condition is triggered, the system automatically performs the following operations:

[0225] Suspend the automatic report generation process;

[0226] Create a manual review work order, package all original data of the sample, preprocessed features, intermediate scores, traceability information and discrepancy details, and push it to the expert review platform;

[0227] Notify the relevant quality inspection personnel;

[0228] S5333. Experts review data on the review platform, conduct manual analysis and judgment, and enter the final ruling (confirming system results, correcting scores, updating traceability information, etc.) into the system.

[0229] S5334. The system stores all data of the sample, the initial system results, the expert ruling results, and the analysis annotations as an anomaly case in a dedicated case library.

[0230] The process of generating a digital inspection report and quality map in S6 includes the following steps:

[0231] S61. Based on the score and feature set, identify potential quality defects using a rule matching method, including the following steps:

[0232] S611, Define the quality defect rule base. Each rule... It is a logical expression composed of atomic conditions (based on scores or feature values) combined using logical operators (AND, OR, NOT) and associated with one or more defect types. and confidence level Atomic conditions can take the form of: (score < threshold) or (feature value ϵ abnormal interval);

[0233] Example rules:

[0234] : →Defect: Suspected contamination or severe degradation during the production process, confidence level: medium;

[0235] : →Defect: Abnormal physical form, possibly due to foreign matter or improper processing; Confidence level: High.

[0236] : →Defect: Insufficient or degraded core active ingredient (10-HDA), confidence level: high;

[0237] S612. Assess the overall authenticity of the current sample. Primary scoring vector Credibility score and from structured feature vectors Key indicators extracted (such as) Using this as input, iterate through the rule base;

[0238] S613. Perform rule matching, for each rule Calculate the truth value (True / False) of its logical expression. For rules that result in True, add their associated defect type and confidence level to the list of potential defects. ;

[0239] S614, to The defects in the data are sorted from highest to lowest confidence level, and this is used as the final result of the potential quality defect analysis.

[0240] S62. Integrate detection results, feature data, and source tracing information to generate a static visualization map, including the following steps:

[0241] S621. Design a visualization template, which includes the following core panels:

[0242] Overall rating panel: Displayed in dashboard format The scores are divided into equal parts, and the levels are indicated by color intervals (red / yellow / green).

[0243] Feature Comparison Radar Chart: Select 5-8 core feature indicators (such as 10-HDA content, protein content, texture consistency, crystal roundness, etc.), and compare the current sample's value with the historical benchmark average. The upper and lower limits of the standard range are plotted on the same radar chart to visually demonstrate the deviation.

[0244] Defect indication map: Based on the results of S61, the identified potential defects and their confidence levels are highlighted in the form of a tag cloud or list;

[0245] Source tracking information panel: Displays honey source areas in the form of text and icons. Harvesting season Key traceability information, and can be embedded with simplified geographic location diagrams;

[0246] S622, Data Binding and Rendering: Fill the corresponding data slots in the template with the data required in S621 (scores, feature values, benchmark values, defect text, and source text). Use a visualization library (such as Matplotlib or Plotly's static export function) to generate images for each panel based on the data.

[0247] S623. Panel Composition and Annotation. The generated panel images, sample IDs, detection timestamps, batch numbers, and other metadata are combined according to the template layout to generate a complete, high-resolution static visualization quality map (such as PNG or PDF format).

[0248] S63. Convert the results into a structured report template and generate a report file containing a unique hash value for verification, including the following steps:

[0249] S631. Prepare structured report data. Organize all the information to be output into structured JSON or XML data objects. Its top-level structure includes:

[0250] Meta: {Sample ID, Detection Time, Batch Number, Detection Device ID};

[0251] Scores: {Overall authenticity score, component authenticity score, structural regularity score, credibility score};

[0252] Defects: L_{defects} (list format, containing defect type and confidence level);

[0253] KeyFeatures: Values ​​of key feature metrics;

[0254] Traceability: Traceability information tuple The main fields;

[0255] Visualization: The path to the quality map file generated by S62 or the Base64 encoded image data;

[0256] S632. Report template population: Use predefined report templates (such as HTML or Markdown format) and the template engine to... The data in the template is filled into the corresponding positions to generate a complete and properly formatted structured text report;

[0257] S633. Generate report hash value and calculate the entire... The SHA-256 hash value of the data object is used to ensure the integrity and immutability of the report content. The calculation formula is:

[0258] ;

[0259] in After converting the data object into a standard JSON string, upon receiving the report, the same algorithm can be used to recalculate the hash value of the report data and compare it with the accompanying... If the comparison is consistent, it proves that the report was complete and tamper-proof during transmission and storage; if inconsistent, it proves that the content has been damaged or forged. This provides verifiable integrity and authenticity assurance for digital test reports.

[0260] S634, Final Report Generation and Storage, As a digital fingerprint, it is inserted into the structured report (usually at the end of the report or in the header and footer). Finally, the structured text report containing the hash value is packaged with the static visual quality map to generate a deliverable digital inspection report file package (such as ZIP or PDF format) and stored in the database.

[0261] The optimization model and update strategy in S7 include the following steps:

[0262] S71. Based on the expert review results, the network parameters are updated using the elastic weight consolidation incremental learning algorithm, including the following steps:

[0263] S711. Collect and prepare the incremental learning dataset. This involves collecting new samples (including samples that were incorrectly judged by the system but corrected by experts) and their corresponding expert rating labels from the S533 manual review process. Create a new task dataset The original training set is denoted as the old task dataset. ;

[0264] S712. Calculate the importance weight matrix of network parameters to the old task. In the original model parameters Based on this, calculate the loss function. (e.g., mean squared error loss) for each parameter The diagonal elements of the Fisher information matrix are used as a measure of the importance of this parameter:

[0265] ;

[0266] in For sample features, For labels, importance matrix Each parameter was quantified. The importance of old tasks;

[0267] S713. Construct an elastic weighted consolidation loss function based on the new dataset. When updating network parameters, the objective function not only includes the loss for the new task, but also adds a regularization term to penalize significant modifications to important parameters of the old task:

[0268] ;

[0269] in, It is the loss function for the new task. It is a hyperparameter that weighs the importance of new and old tasks (such as...) ), These are the optimal parameters for the old task;

[0270] S714. Perform incremental learning training, using an optimizer (such as Adam) to minimize... Update network parameters to This process enables the model to learn new knowledge (such as the characteristics of new adulteration patterns) while retaining as much memory as possible of the learned knowledge (known pure and adulterated patterns), effectively mitigating catastrophic forgetting.

[0271] S72. When abnormal patterns occur consecutively, the adulteration feature database and thresholds are updated after manual review, including the following steps:

[0272] S721. Define and detect continuous abnormal patterns; the system monitors the most recent... One (e.g.) If the number of abnormal cases that trigger manual review under S533 exceeds a preset threshold, it indicates that the case belongs to the same or similar potential defect type (classified according to S61). (like If the average distance of its features in the vector space is less than a set value, it is determined that a continuous abnormal pattern has occurred, which may represent a new adulteration method that the system has not fully learned.

[0273] S722. Initiate feature pattern analysis. When a continuous abnormal pattern is detected, the system automatically aggregates the original feature vectors, structured feature vectors, and expert review conclusions of these cases. Using unsupervised clustering algorithms (such as DBSCAN) or principal component analysis, a subset of common abnormal features within the pattern is extracted. and its statistical distribution;

[0274] S723. Expert Review and Confirmation: Submit the pattern analysis report for expert review. Experts will assess whether it is confirmed as a new adulteration pattern, a variant of a known pattern, or a normal fluctuation. If it is confirmed as a new adulteration pattern, it will be labeled, and an initial recognition threshold vector will be set. ;

[0275] S724. Update the knowledge graph and rule base, including new adulteration patterns confirmed by experts and their feature subsets. and threshold As new nodes and relationships, they are added to the royal jelly quality knowledge graph constructed by S31. At the same time, based on the new features and thresholds, new or revised quality defect identification rules are generated, and the rule base of S611 is updated.

[0276] S73. Periodically push the updated model and feature library to the detection terminal to form a semi-automatic iterative system, including the following steps:

[0277] S731, Version Management and Packaging: The central server maintains version numbers for the model, knowledge graph, rule base, and thresholds. Once S71 or S72 completes a valid update, a new version is generated, and the updated lightweight multimodal fusion network parameters are applied. The updated knowledge graph file, the updated rule base file, and the global configuration (including the latest thresholds) are packaged into an update package;

[0278] S732, Differential Push and Secure Transmission: Employing an incremental update strategy, it calculates the difference files between the new version and the current version on each detection terminal, rather than the full data, to reduce transmission load. Using encrypted channels (such as TLS) and digital signature technology, it securely pushes the differential update package to all online detection terminals. The update package includes version information and the publisher's digital signature. ;

[0279] S733, Terminal Authentication and Silent Update: After the terminal receives the update packet, it first verifies the digital signature. To ensure the legitimacy of the update source and verify its authenticity, the update script is automatically executed during system idle periods (such as when there are no detection tasks) to replace the old version with the new model, knowledge base, and rule base, and update the local configuration.

[0280] S734, Feedback and Confirmation Loop: After the terminal update is completed, it sends an update success confirmation signal and the new version number to the central server. The central server records the update status of each terminal.

[0281] It should be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus.

[0282] 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 variations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A method for detecting pure natural royal jelly freeze-dried powder, characterized in that, The method includes the following steps: S1. Collect multi-dimensional raw data of the freeze-dried royal jelly powder to be tested, including near-infrared spectral data, visible light microscopic image data, and physicochemical activity index data; S2. Preprocess and extract features from the multi-dimensional raw data to generate a standardized fused data matrix; S3. Based on the royal jelly quality knowledge graph, extract a set of key feature indicators related to purity, freshness and adulteration from the standardized fusion data matrix; S4. Construct and utilize a lightweight multimodal fusion network to perform parallel analysis and authenticity evaluation of the key feature index set, and output preliminary scores for each dimension. The evaluation using a lightweight multimodal fusion network in S4 includes the following steps: S41. Construct a dual-branch lightweight network, in which a one-dimensional convolutional layer processes spectral features, a lightweight CNN processes image features, and active features are processed through a fully connected layer. S42. Input the structured feature vector into the corresponding branch and calculate the initial scores for component authenticity based on spectral and activity features and structural regularity based on image morphological features, respectively. S43. The initial scores output from each branch are concatenated, fused through a fully connected layer, and then normalized by the Sigmoid function to output the normalized initial scores for each dimension. S5. Based on the credible traceability database and the historical best product database, the initial score is dynamically weighted and calibrated to generate a comprehensive authenticity score. S6. Based on the analysis results of the comprehensive authenticity score and key feature index set, generate a digital inspection report and quality map with traceability chain; S7. Based on batch detection results, continuously optimize the parameters and dynamic weighting strategy of the multimodal fusion network, and update the adulteration feature library and threshold.

2. The method for detecting pure natural royal jelly freeze-dried powder according to claim 1, characterized in that: The collection of multi-dimensional raw data in S1 includes the following steps: S11. Use a near-infrared spectrometer to collect the chemical composition index data of the freeze-dried powder and record the characteristic absorption peaks; S12. Take microscopic images of the freeze-dried powder using a digital microscope, including crystal morphology and color distribution; S13. Use a rapid test kit to obtain basic physicochemical data of the reconstituted lyophilized powder.

3. The method for detecting pure natural royal jelly freeze-dried powder according to claim 1, characterized in that: The preprocessing and feature extraction in S2 includes the following steps: S21. Smooth, baseline correct and normalize the near-infrared spectral data, perform background segmentation, contrast enhancement and size normalization on the microscopic image, and extract local binary pattern texture features. S22. Compare the physicochemical activity index data with the standard reference values ​​and calculate the relative activity deviation. S23. The processed spectral feature vector, image morphology feature vector, and activity feature vector are concatenated according to the sample ID to construct a standardized fusion data matrix.

4. The method for detecting pure natural royal jelly freeze-dried powder according to claim 1, characterized in that: The extraction of key feature indicator set in S3 includes the following steps: S31. Construct a knowledge graph of royal jelly quality. The nodes of the knowledge graph of royal jelly quality include the standard component range, typical morphological characteristics, activity index thresholds and known adulteration patterns. S32. Using the Node2Vec graph embedding method, the standardized fusion data matrix is ​​mapped to the knowledge graph, and core features and abnormal features are identified through similarity matching. S33. Extract the feature peak area ratio, texture consistency parameters calculated based on the local binary mode variance of the image, and the average relative deviation of the physicochemical indicators relative to the standard values ​​as the activity deviation degree, and encode them into a structured feature vector.

5. The method for detecting pure natural royal jelly freeze-dried powder according to claim 1, characterized in that: The construction of the lightweight multimodal fusion network in S4 also includes the following steps: S411. Initialize the image branch weights using pre-trained MobileNetV2 and fine-tune them with a small number of samples; S412. Introduce data augmentation to improve network robustness and reduce reliance on synthetic data; S413. Use Grad-CAM class activation mapping to generate visual heatmaps of key feature regions to improve interpretability.

6. The method for detecting pure natural royal jelly freeze-dried powder according to claim 1, characterized in that: The dynamic weighting and credibility calibration in S5 includes the following steps: S51. Retrieve basic traceability information of raw materials from the trusted traceability database and calculate the simple trust score; S52. Query the historical best product database to obtain benchmark data of similar honey sources and seasons as a reference. S53. Using the primary score, credibility score, and the Euclidean distance between the current sample's key features and historical benchmark features as the benchmark deviation, a weighted average fusion method is used to calculate the comprehensive authenticity score.

7. The method for detecting pure natural royal jelly freeze-dried powder according to claim 6, characterized in that: The weighted average fusion method used in S53 also includes the following steps: S531. Assign time decay weights to recent batch data; S532. For missing or contradictory source data, use the mean or mode to fill in the missing data; S533. When the score is seriously inconsistent with the traceability, the manual review process is triggered and the abnormal case is recorded.

8. The method for detecting pure natural royal jelly freeze-dried powder according to claim 1, characterized in that: The process of generating a digital inspection report and quality map in S6 includes the following steps: S61. Based on the score and feature set, a rule matching method is used to identify potential quality defects; S62. Integrate detection results, feature data and source tracing information to generate a static visualization map; S63. Convert the results into a structured report template and generate a report file containing a unique hash value for verification.

9. The method for detecting pure natural royal jelly freeze-dried powder according to claim 1, characterized in that: The optimization model and update strategy in S7 include the following steps: S71. Based on the expert review results, the network parameters are updated using the elastic weight consolidation incremental learning algorithm. S72. When abnormal patterns occur continuously, the adulteration feature library and thresholds are updated after manual review. S73. The updated model and feature library are periodically pushed to the detection terminal to form a semi-automatic iterative system.

Images