Intelligent quality detection method and system for micro-mineral bio-organic fertilizer

Abstract

The invention relates to the technical field of material testing, and particularly discloses an intelligent quality detection method and system for a micro-mineral bio-organic fertilizer, local thermal excitation is applied to a fertilizer sample through a micro-area thermal pulse excitation device, a gas release kinetic curve and a spectrum change track are synchronously collected, and a gas spectrum coupling data cube is constructed; performing differential transformation and modal decomposition on the time sequence feature set, extracting transient response feature vectors and generating an activity response distribution diagram; establishing component-activity correlation analysis, and decoupling biological metabolic activity and matrix background interference through feature separation and a comparative learning strategy to obtain a dynamic metabolic fingerprint; constructing a multi-dimensional feature space by using the dynamic metabolic fingerprints and the apparent characteristic parameters, and performing information aggregation and feature reconstruction by using a graph convolutional network to generate a comprehensive quality index; and finally, realizing quality grade judgment based on a quality characteristic pyramid structure, and establishing a dynamic early warning mechanism by analyzing time sequence evolution characteristics of the activity response distribution diagram.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of material testing, in particular to a quality intelligent detection method and system for micro-mineral biological organic fertilizer. BACKGROUND

[0002] As a new type of fertilizer integrating organic matter, microbial agent and trace elements, the quality of micro-mineral biological organic fertilizer is highly dependent on the number and metabolic capacity of active microorganisms. However, due to the complex composition of the fertilizer, including grass carbon, humic acid, attapulgite and other auxiliary materials, traditional detection methods relying on chemical analysis or single sensors (such as near-infrared spectroscopy or gas sensors) cannot accurately and non-destructively evaluate the microbial activity. The existing technology usually judges the quality through offline sampling, laboratory culture or static spectral analysis, which has the problems of complicated process, poor timeliness, easy to be affected by subjective experience, etc., and cannot meet the needs of real-time monitoring and intelligent management of fertilizer quality in modern agricultural production.

[0003] The existing technology has the following disadvantages:

[0004] The porous adsorbent auxiliary materials (such as zeolite and attapulgite) in the fertilizer can strongly adsorb the trace characteristic gases (such as volatile organic acids) produced by microbial metabolism, resulting in serious attenuation of the gas sensor signal and causing "false negative" misjudgment. At the same time, the organic matter such as grass carbon and humic acid forms a strong background spectral signal in the near-infrared band, completely burying the weak spectral changes caused by microbial metabolism. This matrix interference makes the traditional data analysis methods (such as static threshold judgment or simple regression model) unable to effectively distinguish the real metabolic signal of microorganisms from the matrix background noise, resulting in the detection system misjudging high-activity fertilizer as "unqualified" or misjudging poor-quality fertilizer contaminated by bacteria as "high-activity", which seriously restricts the reliability and applicability of quality detection. SUMMARY

[0005] The purpose of the present application is to provide a quality intelligent detection method and system for micro-mineral biological organic fertilizer to solve the problems in the background.

[0006] The purpose of the present application can be achieved by the following technical solutions:

[0007] A quality intelligent detection method for micro-mineral biological organic fertilizer, comprising the following steps:

[0008] S1: Construct a time sequence feature set containing gas release kinetics curve and spectral change trajectory by applying a preset mode of energy excitation and synchronously collecting the multi-source dynamic response signal of the fertilizer sample within the excitation period;

[0009] S2: Perform differential transformation and modal decomposition on the time sequence feature set, extract transient response feature vectors representing biological metabolic intensity, and generate corresponding activity response distribution maps;

[0010] S3: Establish a component-activity association model based on the activity response distribution map, decouple biological metabolic activity from matrix background interference, and obtain a dynamic metabolic fingerprint reflecting the real activity of microorganisms;

[0011] S4: Construct a multi-dimensional feature space by combining the dynamic metabolic fingerprint with the apparent characteristic parameters of the sample, and use a graph convolution network to aggregate information and reconstruct features of the characteristic nodes to generate a comprehensive quality index;

[0012] S5: Generate quality grade determination results based on the comprehensive quality index, and analyze the time sequence evolution characteristics of the activity response distribution map to establish a dynamic quality warning threshold.

[0013] As a further scheme of the present application: the time sequence feature set containing gas release kinetics curves and spectral change trajectories is constructed, specifically including:

[0014] A micro-thermal pulse excitation device is used to apply local thermal excitation to the surface of the sample, with a duration of milliseconds and a temperature change amplitude of 3-5 degrees Celsius. High-spectral image sequences of specific areas on the surface of the sample are collected before and after excitation;

[0015] A gas sensor array arranged around the sample space is used to synchronously monitor the instantaneous changes in the concentrations of multiple characteristic gases at a millisecond sampling frequency, and record the gas release rate curves and their spectral characteristics;

[0016] The high-spectral image sequences and the gas release rate curves are time-stamped and aligned, the spectral reflectance change gradient at each sampling time is extracted, and a gas-spectrum coupling data cube containing spatiotemporal features is constructed;

[0017] The gas-spectrum coupling data cube is subjected to multi-scale decomposition, and fast response components and slow background components are extracted, and the fast response components are reconstructed into a time sequence feature set reflecting the dynamics of microbial metabolism.

[0018] As a further scheme of the present application: the time sequence feature set is subjected to differential transformation and modal decomposition, and transient response feature vectors representing biological metabolic intensity are extracted, specifically including:

[0019] The gas release kinetics curves in the time sequence feature set are subjected to high-order differential processing, the first-order differential sequence and the second-order differential sequence are calculated, and the mutation points and change trends of the gas release rate are captured;

[0020] Adaptive intrinsic mode decomposition is performed on the spectrum change trajectory, and the non-stationary signal is decomposed into multiple intrinsic mode components with different time scales, and a characteristic mode matching the microbial metabolic cycle is screened;

[0021] The amplitude feature of the mutation point and the energy distribution parameter of the characteristic mode are extracted, and a multi-dimensional transient response feature vector containing time domain mutation features and frequency domain energy features is constructed;

[0022] The multi-dimensional transient response feature vector is processed by dimension reduction to obtain a core feature subset representing the biological metabolic intensity.

[0023] As a further scheme of the present application, the corresponding activity response distribution map is generated, specifically including:

[0024] The transient response feature vector is reorganized according to the sampling space position to construct a feature intensity matrix based on a spatial grid;

[0025] The radial basis function interpolation method is used to perform spatial interpolation on the feature intensity matrix to generate a continuously distributed activity intensity contour map;

[0026] According to the gradient change characteristics of each region in the activity intensity contour map, the response regions of different activity levels are divided, and the metabolic hot spot regions are labeled;

[0027] The spatial distribution of the gas release feature and the spectrum change feature is superimposed and fused to generate an activity response distribution map that comprehensively reflects the spatial distribution and metabolic intensity of microorganisms.

[0028] As a further scheme of the present application, the component-activity correlation model is established based on the activity response distribution map, specifically including:

[0029] The spatial gradient feature and the intensity distribution feature in the activity response distribution map are taken as input to construct a double-channel feature matrix containing substrate background features and metabolic activity features;

[0030] An adversarial twin network model is established, and the generator network is used to learn to separate the substrate background features from the double-channel feature matrix, and the discriminator network is used to identify the integrity of the metabolic activity features;

[0031] The separated metabolic activity features are input into a residual mapping network to establish a nonlinear mapping relationship between metabolic activity intensity and microbial quantity, and output a pure activity feature vector free from substrate interference;

[0032] The residual mapping network is optimized by a contrastive learning strategy, so that the output features of the same activity level remain consistent under different substrate backgrounds.

[0033] As a further scheme of the present application, the dynamic metabolic fingerprint reflecting the real activity of microorganisms is obtained, specifically including:

[0034] Reorganize the pure active feature vector in time series to construct a dynamic metabolic trajectory matrix;

[0035] Perform singular value decomposition on the dynamic metabolic trajectory matrix to extract intrinsic metabolic patterns representing the main characteristics of the metabolic process;

[0036] Perform dynamic time warping matching between intrinsic metabolic patterns and standard metabolic templates to calculate metabolic pathway similarity indexes;

[0037] Based on the energy distribution of the metabolic pathway similarity index and the intrinsic metabolic pattern, generate a dynamic metabolic fingerprint with unique identification.

[0038] As a further scheme of the present application: the generated comprehensive quality indicator specifically includes:

[0039] Each feature dimension of the dynamic metabolic fingerprint and the sample apparent characteristic parameter are jointly constructed as a feature node, and a node connection edge is established according to the mutual information between the features, forming a dynamic feature topology graph;

[0040] Through the graph convolution layer, the neighborhood information of each feature node is aggregated layer by layer, while retaining the local features and fusing the global correlation features;

[0041] According to the contribution of the feature node to the quality evaluation, the weight in the information transmission process is dynamically adjusted;

[0042] Global pooling is performed on the feature nodes processed by the multi-layer graph convolution, and a comprehensive quality indicator containing local features and global correlations is output.

[0043] As a further scheme of the present application: the quality grade determination result is generated based on the comprehensive quality indicator, specifically including:

[0044] Construct a quality feature pyramid-based grade decision tree, and input the comprehensive quality indicator to the corresponding decision node according to the hierarchical structure of the quality feature pyramid;

[0045] At each decision node, an adaptive membership function is used to fuzzy quantify the input features, and the probability distribution of each quality grade is calculated;

[0046] Fuse and analyze the high-frequency fluctuations of the bottom micro-features and the stable trends of the top macro-features;

[0047] According to the comparison result of the final confidence score and the preset grade threshold, output the quality grade determination result with confidence evaluation.

[0048] As a further scheme of the present application: the time sequence evolution characteristics of the active response distribution graph are analyzed, and a dynamic quality early warning threshold is established, specifically including:

[0049] extracting spatial distribution characteristic parameters of the active response distribution map in the continuous detection process, including active intensity gradient, metabolic hotspot area, and spatial heterogeneity index;

[0050] calculating differential characteristics of the spatial distribution characteristic parameters over time, identifying trend changes and abnormal fluctuation patterns in the active evolution process;

[0051] based on the acceleration characteristics of the trend changes and the statistical characteristics of the abnormal fluctuations, a dynamic early warning threshold curve is constructed;

[0052] by comparing and analyzing the historical early warning data and the current monitoring data, the morphological parameters of the early warning threshold curve are optimized, when the active evolution characteristics exceed the early warning threshold, a graded early warning signal is triggered and the abnormal characteristic map is recorded.

[0053] A quality intelligent detection system for micro-mineral organic fertilizer, comprising:

[0054] a dynamic excitation and multi-source signal acquisition module, which acquires multi-source dynamic response signals of the fertilizer sample in an excitation cycle by applying energy excitation of a preset mode and synchronously, constructs a time sequence feature set containing a gas release kinetics curve and a spectral change trajectory;

[0055] a transient feature extraction and visualization module, which is used for differential transformation and modal decomposition of the time sequence feature set, extracts a transient response feature vector representing biological metabolic intensity, and generates a corresponding active response distribution map;

[0056] a metabolic fingerprint decoupling and analysis module, which establishes a component-activity correlation model based on the active response distribution map, decouples biological metabolic activity from matrix background interference, and obtains a dynamic metabolic fingerprint reflecting the real activity of microorganisms;

[0057] a multi-modal feature fusion and quality index generation module, which constructs a multi-dimensional feature space of the dynamic metabolic fingerprint and sample apparent characteristic parameters, aggregates and reconstructs feature nodes using a graph convolution network, and generates a comprehensive quality index;

[0058] an intelligent decision and dynamic early warning module, which generates a quality grade determination result based on the comprehensive quality index, analyzes the time sequence evolution characteristics of the active response distribution map, and establishes a dynamic quality early warning threshold.

[0059] The beneficial effects of the present application are:

[0060] (1) The present application adopts micro-zone thermal pulse excitation to stimulate specific response of microorganisms, combines gas-spectrum coupling data cube construction and multi-scale signal decomposition, effectively separates microbial metabolic signals and matrix background noise. Through differential transformation and modal decomposition of the time sequence feature set, the transient response characteristics representing the intensity of biological metabolism are further extracted, and through the antagonistic learning and contrast learning strategy in component-activity correlation analysis, the accurate evaluation of the real activity of microorganisms under different formula conditions is realized, solving the detection distortion problem caused by the matrix shielding effect of the traditional method.

[0061] (2) The present application deeply fuses dynamic metabolic fingerprints and apparent characteristic parameters through graph convolution network, constructs a comprehensive quality index, and overcomes the limitations of single index evaluation. Based on the hierarchical decision structure of the quality feature pyramid, the microscopic metabolic characteristics and macroscopic physical properties are comprehensively considered, and through multi-level information fusion and confidence evaluation, the scientificity of quality grade determination is ensured. At the same time, through the time sequence evolution analysis of the activity response distribution diagram, a dynamic early warning mechanism is established, which can early detect abnormal quality trends, realize the transformation from passive detection to active early warning, and provide effective technical support for the quality control of the production process. BRIEF DESCRIPTION OF DRAWINGS

[0062] The present application will be further described below in conjunction with the accompanying drawings.

[0063] Figure 1 is a flow chart of the method of the present application;

[0064] Figure 2 is a flow chart of the system of the present application. DETAILED DESCRIPTION

[0065] The technical solutions in the embodiments of the present application will be described clearly and completely below in conjunction with the accompanying drawings of the embodiments of the present application. Obviously, the described embodiments are only part of the embodiments of the present application, not all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative labor fall within the scope of protection of the present application.

[0066] Please refer to Figure 1 The present application is a quality intelligent detection method for micro-mineral organic fertilizer, which comprises the following steps:

[0067] S1: Through the application of a preset mode of energy excitation and the synchronous acquisition of the multi-source dynamic response signals of the fertilizer sample within the excitation period, a time sequence feature set containing gas release kinetics curve and spectrum change trajectory is constructed;

[0068] S2: Differential transformation and modal decomposition are performed on the time series feature set to extract the transient response feature vector representing the biological metabolic intensity, and the corresponding activity response distribution map is generated;

[0069] S3: A component-activity association model is established based on the activity response distribution map, and the dynamic metabolic fingerprint reflecting the real activity of microorganisms is obtained by decoupling the biological metabolic activity and the matrix background interference;

[0070] S4: The dynamic metabolic fingerprint and the apparent characteristic parameters of the sample are constructed into a multi-dimensional feature space, and the graph convolution network is used to aggregate information and reconstruct features of the characteristic nodes to generate a comprehensive quality index;

[0071] S5: The quality grade determination result is generated based on the comprehensive quality index, and the time series evolution characteristics of the activity response distribution map are analyzed to establish a dynamic quality warning threshold.

[0072] In S1, by applying a preset mode of energy excitation and synchronously collecting the multi-source dynamic response signals of the fertilizer sample within the excitation period, a time series feature set containing gas release kinetic curves and spectral change trajectories is constructed, which specifically includes:

[0073] A micro-thermal pulse excitation device is used to apply local thermal excitation to the sample surface. The device is composed of a precision temperature controller and a micro-heating probe, which can generate a pulse-type thermal excitation with a duration of 50-100 milliseconds, causing the temperature of a specific area on the sample surface to rise by 3-5 degrees Celsius in a short time. The diameter of the thermal excitation area is controlled within 5-8 millimeters, ensuring that the excitation energy is concentrated and will not cause damage to the sample. Before the thermal excitation starts, a reference spectral image of the sample surface is collected by the hyperspectral imaging system; during and after the application of thermal excitation, the hyperspectral image sequence of the sample surface is continuously collected at a rate of 20 frames per second, recording the spectral reflectance changes in the wavelength range of 400-1000 nanometers.

[0074] Secondly, the release dynamics of characteristic gases are synchronously monitored by a gas sensor array arranged in the sample detection cabin. The array contains carbon dioxide sensors, ammonia sensors, and volatile organic compound sensors, each of which has millisecond-level response capability. At the same time the thermal excitation starts, all sensors synchronously record the gas concentration data at a sampling frequency of 100 hertz. By analyzing the rate of change of gas concentration over time, the gas release rate curve is calculated; at the same time, the release rate data is subjected to fast Fourier transform to extract its frequency spectrum features, including the main frequency component and the frequency band energy distribution.

[0075] Then, the collected hyperspectral image sequence and the gas release rate curve are time-synchronized. A unified time axis coordinate system is established, and the gas concentration data at each sampling time and the spectral image at the corresponding time are registered. The spectral reflectance change gradient of each spatial position at each sampling time is extracted from the registered data. The calculation method is to compare the difference between the current time reflectance and the reference reflectance, and combine the change trend of the adjacent sampling points. These spatio-temporal correlated data are organized into a three-dimensional structure, in which two dimensions represent the spatial coordinates of the sample surface, and the third dimension represents time, forming a gas-spectrum coupling data cube.

[0076] Finally, the gas-spectrum coupling data cube is processed by multi-scale signal decomposition. A signal change rate-based separation method is used to divide the signal components in the data cube into two parts: fast response component and slow-changing background component. The fast response component mainly corresponds to the transient change signal generated by microbial metabolic activity, with a time constant below seconds; the slow-changing background component is mainly derived from the physical and chemical properties of the substrate material, with a slow change rate. By setting appropriate time windows and change rate thresholds, the fast response component is extracted from the original data and reorganized into a time series feature set reflecting the metabolic dynamics of microorganisms. This time series feature set contains the transient metabolic response characteristics of microorganisms under the action of thermal stimulation, providing a reliable data basis for subsequent activity evaluation.

[0077] In S2, the time series feature set is subjected to differential transformation and modal decomposition to extract the transient response feature vector representing the metabolic intensity of the organism and generate the corresponding activity response distribution map, which includes:

[0078] The gas release kinetics curve in the time series feature set is subjected to differential processing. The central difference method is used to calculate the first-order differential sequence of the gas concentration data, which reflects the change of the gas release rate. On this basis, the second-order differential sequence is calculated, which reflects the acceleration change of the release rate. By setting appropriate thresholds, the extreme points in the first-order differential sequence and the zero-crossing points in the second-order differential sequence are identified, which correspond to the mutation positions in the gas release process. The time coordinates, amplitude sizes and change trends of each mutation point are recorded to form a mutation feature set of gas release.

[0079] The signal decomposition process is performed on the spectral variation trajectory. An adaptive decomposition method based on local extreme points is used to decompose the non-stationary spectral signal into multiple intrinsic mode components. The specific process includes: identifying the local maximum and minimum points of the spectral signal, constructing the upper envelope line and the lower envelope line through cubic spline interpolation respectively, calculating the mean value of the upper and lower envelope lines and subtracting the mean value from the original signal, repeating this process until the intrinsic mode component conditions are met. All the components obtained by decomposition are sorted according to their average period length, and the components with a period in the range of 0.5 seconds to 5 seconds are selected. These components match the time scale of microbial metabolic activity.

[0080] Feature parameters are extracted from the processing results. For the gas release mutation point, its relative amplitude, duration, and pre-post change slope parameters are extracted. For the selected spectral feature modes, the energy value, center frequency, and bandwidth parameters of each mode are calculated. These time-domain and frequency-domain features are combined in a fixed order to form a multi-dimensional transient response feature vector. The dimension of this feature vector may be high, so principal component analysis is used for dimension reduction. The cumulative contribution rate of the selected principal components exceeds 85%, and a core feature subset is formed.

[0081] In generating the activity response distribution map, the transient response feature vector is first reorganized according to the spatial coordinate position at the time of sampling. According to the distribution of sampling points on the sample surface, a two-dimensional grid coordinate system is established, and the feature intensity value of each sampling point is filled into the corresponding grid position to form a feature intensity matrix. For positions with missing data in the grid, a radial basis function interpolation method based on distance weighting is used to fill in the missing data. Specifically, a Gaussian function is used as the basis function, and the shape parameter is adaptively determined according to the spatial distribution density of the data points.

[0082] The activity intensity contour map is generated according to the interpolated feature intensity matrix. In the two-dimensional plane, points with the same feature intensity are connected into smooth curves, and the contour interval is evenly divided according to the global distribution range of the feature intensity. By analyzing the gradient distribution characteristics of the contour map, i.e., the density of the contour lines, the sample area is divided into high activity area, medium activity area, and low activity area. The area with dense contour lines and high intensity value is marked as the metabolic hotspot area.

[0083] The spatial distribution map of gas release features and the spatial distribution map of spectral variation features are superimposed. After coordinate registration of the two feature maps, a weighted average method is used for fusion, and the weight is determined according to the signal-to-noise ratio of each feature. The feature intensity after fusion is visualized through color mapping, with red representing high activity area and blue representing low activity area, generating an activity response distribution map that comprehensively reflects the spatial distribution and metabolic intensity of microorganisms. This distribution map can intuitively display the spatial heterogeneity of microbial activity in the sample, providing a visual basis for quality evaluation.

[0084] In S3, a component-activity association model is established based on the activity response distribution map, and by decoupling biological metabolic activity and matrix background interference, a dynamic metabolic fingerprint reflecting the real activity of microorganisms is obtained.

[0085] The spatial gradient feature and intensity distribution feature in the activity response distribution map are taken as input data to construct a dual-channel feature matrix containing matrix background features and metabolic activity features. The spatial gradient feature is obtained by calculating the feature intensity difference between adjacent sampling points, and the intensity distribution feature is obtained by counting the feature intensity distribution histogram of each region. These two types of features are organized into two independent data channels, where the first channel mainly contains slowly varying background features related to the matrix material, and the second channel mainly contains rapidly changing features related to microbial metabolism.

[0086] An analysis framework containing a feature separation network and a feature discrimination network is constructed. The feature separation network is responsible for learning to distinguish between matrix background features and metabolic activity features from the dual-channel feature matrix, and achieves layer-by-layer separation of features through a multi-layer neural network. The feature discrimination network is responsible for evaluating the integrity of the separated metabolic activity features, ensuring that the complete metabolic activity information is retained while removing the matrix interference. The two networks work cooperatively through alternating training. During the training process, the feature separation network continuously optimizes its separation ability, while the feature discrimination network continuously improves the accuracy of judging the integrity of the features.

[0087] The separated metabolic activity features are input into a deep network with a skip connection structure for further processing. The network contains multiple processing layers, each of which performs nonlinear transformation on the input features while preserving the original feature information through skip connections. In this way, a complex nonlinear mapping relationship between metabolic activity intensity and microbial quantity is established. The output of this network is a pure activity feature vector that has removed the matrix interference, which can more accurately reflect the real metabolic state of microorganisms.

[0088] To improve the stability of the analysis results under different matrix conditions, a contrast learning strategy is used to optimize the above processing process. Specifically, during the training process, samples with the same activity level but different matrix backgrounds are taken as positive sample pairs, and samples with different activity levels are taken as negative sample pairs. By optimizing the network parameters, the feature vectors output by the network after processing the positive sample pairs have high similarity, while the feature vectors of the negative sample pairs have large differences. This processing method can effectively improve the adaptability of the analysis system to different formula fertilizers.

[0089] On the basis of the pure active feature vectors, further construct the dynamic metabolic fingerprint. First, organize the pure active feature vectors of multiple continuous sampling time points in time sequence to construct a dynamic metabolic trajectory matrix. The rows of the matrix correspond to different feature dimensions, and the columns correspond to different sampling time points. Through time series analysis of the matrix, the dynamic change rule of microbial metabolic activity can be obtained.

[0090] Perform matrix decomposition processing on the dynamic metabolic trajectory matrix to extract the intrinsic metabolic mode representing the main characteristics of the metabolic process. The specific process is to calculate the covariance matrix of the matrix, and then solve the eigenvectors and eigenvalues of the covariance matrix. Select the first few eigenvectors with larger eigenvalues as the intrinsic metabolic mode, which represents the most important change rule in the microbial metabolic process.

[0091] Match and compare the extracted intrinsic metabolic mode with the pre-established standard metabolic template. The standard metabolic template is a reference mode established by a large number of experiments on standard samples with known activity levels. The matching process uses the dynamic time warping algorithm, which can handle pattern matching problems of different time scales. By calculating the minimum alignment distance between the two mode sequences, the metabolic pathway similarity index is obtained, which reflects the similarity between the test sample and the standard template.

[0092] Based on the metabolic pathway similarity index and the energy distribution of the intrinsic metabolic mode, generate a dynamic metabolic fingerprint with unique identification. The energy distribution is obtained by calculating the proportion of the eigenvalue corresponding to each intrinsic metabolic mode in the total eigenvalue. Combine the similarity index and the energy distribution parameters according to the fixed coding rule to form a feature vector containing multiple dimensions, which is the dynamic metabolic fingerprint. This fingerprint can comprehensively and accurately reflect the characteristic state of microbial metabolic activity, providing a reliable basis for subsequent quality evaluation.

[0093] In S4, the dynamic metabolic fingerprint and the sample apparent characteristic parameters are constructed into a multi-dimensional feature space, and the graph convolution network is used to aggregate information and reconstruct features of the feature nodes to generate a comprehensive quality index, which includes:

[0094] The feature dimensions of the dynamic metabolic fingerprint and the sample apparent property parameters are jointly constructed as feature nodes. The dynamic metabolic fingerprint contains 12 feature dimensions, respectively from the metabolic pathway similarity index and the energy distribution parameter of the intrinsic metabolic mode. The sample apparent property parameters include 6 parameters such as temperature, humidity, color value and particle size distribution of the sample. These features are uniformly organized into 18 feature nodes, each node containing three attributes of feature name, feature value and feature type. The node connection edges are established according to the mutual information between the features, and the mutual information is obtained by calculating the statistical correlation between each two features. Specifically, the ratio of joint probability and marginal probability of two feature distributions is calculated based on histogram statistics. When the mutual information of two feature nodes exceeds the preset threshold of 0.15, a connection edge is established between the corresponding nodes, and finally a dynamic feature topology graph with 18 nodes and several connection edges is formed.

[0095] The neighborhood information of each feature node is aggregated layer by layer through the graph convolution processing layer. In the first layer processing process, each feature node collects the feature information of its directly adjacent nodes, and fuses the neighborhood features with its own features through weighted summation, and the weight is determined according to the mutual information between nodes. In the second layer processing process, each feature node can collect more distant neighborhood information, and the information transmission range is expanded to two-hop neighborhood. Through this multi-layer processing mechanism, each feature node not only retains its own local features, but also fuses global correlation features. In each layer processing process, linear transformation and nonlinear activation function are combined to update node features, ensuring that the complexity of feature transformation can fully explore the deep relationship between features.

[0096] The weight in the information transmission process is dynamically adjusted according to the contribution of the feature node to the quality evaluation. In the feature information transmission process, an attention allocation mechanism is introduced, which determines the weight by calculating the correlation between the feature node and the quality evaluation target. The specific calculation process includes: first, calculate the cross-correlation coefficient between each feature node and the sample features of known quality levels, and then dynamically adjust the influence of this node in the information aggregation process according to the size of the correlation coefficient. For feature nodes with high correlation with quality evaluation, give them larger weights, so that they play a more important role in feature fusion; for feature nodes with low correlation, appropriately reduce their weights. This dynamic weight adjustment mechanism can adaptively highlight the role of key features and improve the relevance of feature fusion.

[0097] The feature nodes processed by the multi-layer graph convolution are globally pooled to output the comprehensive quality indicator. Global pooling includes two steps of maximum pooling and average pooling: maximum pooling selects the maximum value of each feature dimension from all feature nodes, and average pooling calculates the average value of each feature dimension of all feature nodes. The results of maximum pooling and average pooling are spliced to form a feature vector containing 36 dimensions. The feature vector is standardized to unify the numerical range to 0 to 1. Then the 36-dimensional feature vector is mapped to 3 quality indicator dimensions through a fully connected layer, corresponding to the microbial activity indicator, the nutrient component indicator and the physical property indicator. The three indicators together constitute the comprehensive quality indicator, in which the microbial activity indicator weight is 0.5, the nutrient component indicator weight is 0.3, and the physical property indicator weight is 0.2. The final comprehensive quality score is obtained by weighted summation, and the score range is 0-100 points.

[0098] In S5, the quality grade determination result is generated based on the comprehensive quality indicator, and the time evolution characteristics of the activity response distribution map are analyzed to establish a dynamic quality warning threshold, which specifically includes:

[0099] In terms of quality grade determination, first, a grade decision structure based on the quality feature pyramid is constructed. The structure is divided into three levels: the bottom layer is the microbial activity feature, including 5 indicators such as metabolic rate and activity intensity; the middle layer is the nutrient component feature, including 4 indicators such as organic matter content and trace elements; and the top layer is the physical property feature, including 3 indicators such as particle size and color. The comprehensive quality indicator is decomposed according to this hierarchical structure and input into the corresponding decision node. Each decision node uses a piecewise linear function to preliminarily classify according to the value range of the input feature, and outputs the quality grade of this level.

[0100] Secondly, an adaptive membership function is used to quantitatively process the input features at each decision node. The parameters of the membership function are automatically adjusted according to the feature distribution law of different quality grades in historical data. For the microbial activity feature, a Gaussian membership function is used, and its mean and variance are obtained by statistical learning of the feature values of qualified samples. For the nutrient component feature, a trapezoidal membership function is used, and its boundary points are set according to the content range specified in the national standard. Through membership function calculation, the probability distribution of each feature corresponding to the superior product, qualified product and unqualified product is obtained, and the sum of the probability values is 1.

[0101] Then, the high-frequency fluctuation information of the bottom micro-features is fused with the stable trend information of the top macro-features. The bottom micro-features include the instantaneous change rate of activity intensity and the moving speed of metabolic hotspots, which have a high frequency of 3-5 fluctuations per second. The top macro-features include the overall distribution trend of active regions and the stability of metabolic patterns, which change slowly about 1-2 times per minute. Through time scale transformation, the high-frequency features are processed by moving average, and the low-frequency features are processed by difference refinement, so that they are fused at the same time scale. The fusion weight is dynamically adjusted according to the signal-to-noise ratio of the features, which is determined by calculating the ratio of the variance to the mean of the feature value.

[0102] Finally, the final confidence score is calculated according to the fused features. The confidence score consists of three parts: the matching degree of each level of features and the corresponding quality grade, the consistency degree of the determination results between different levels, and the continuity of the current determination result and the historical determination result. The matching degree is obtained by calculating the Euclidean distance between the feature value and the grade standard value, the consistency is obtained by calculating the correlation coefficient of the determination results of different levels, and the continuity is obtained by calculating the change rate of the determination results at adjacent time points. The three parts are weighted and summed with weights of 0.5, 0.3, and 0.2 to obtain a confidence score of 0-100. The score is compared with the preset grade threshold, and products with a score of 85 or above are determined as excellent, products with a score of 60-85 are determined as qualified, and products with a score of 60 or below are determined as unqualified, while the confidence evaluation report is output.

[0103] In the aspect of dynamic quality early warning, first, the spatial distribution feature parameters of the activity response distribution map in the continuous detection process are extracted. The activity intensity gradient is obtained by calculating the change rate of activity intensity per unit distance, with a sampling point interval of 2 mm. The metabolic hotspot area is obtained by calculating the proportion of the area where the activity intensity is 1.5 times higher than the overall average. The spatial heterogeneity index is obtained by calculating the coefficient of variation of the activity intensity of each sampling point, i.e. the ratio of the standard deviation to the mean. These parameters are collected every 30 seconds to form time series data.

[0104] Next, the differential features of the spatial distribution feature parameters over time are calculated. First-order differential calculation is performed on each feature parameter sequence to obtain the change rate sequence, and second-order differential calculation is performed to obtain the change acceleration sequence. By setting a threshold, abnormal rising or falling trends in the change rate sequence and sudden changes in the change acceleration sequence are identified. Abnormal fluctuation patterns are identified by calculating the consistency of the change direction of 5 consecutive sampling points. When 5 consecutive points show an upward or downward trend, it is determined that there is an abnormal fluctuation.

[0105] A dynamic early warning threshold curve is constructed based on the acceleration characteristics of the trend changes and the statistical characteristics of the abnormal fluctuations. The trend acceleration characteristics are obtained by calculating the change amount of the change rate per unit time. When the acceleration value exceeds twice the standard deviation of the historical average value for 3 consecutive sampling points, it is determined that there is a trend change. The statistical characteristics of abnormal fluctuations include fluctuation amplitude, duration and frequency, which are obtained by statistically analyzing the occurrence of abnormal fluctuations in the past 1 hour. The early warning threshold curve is composed of a basic threshold and an adjustment coefficient. The basic threshold is determined according to the historical normal data statistics, and the adjustment coefficient is dynamically corrected according to the real-time monitored trend acceleration and abnormal fluctuation characteristics.

[0106] Finally, the morphological parameters of the early warning threshold curve are optimized by comparing and analyzing the historical early warning data and the current monitoring data. An early warning case library is established to record the characteristic parameter combination and subsequent quality change when each early warning is triggered. When new monitoring data arrives, the similarity between the new data and the historical data in the case library is matched to find the most similar 10 cases. The sensitivity of the current early warning threshold is adjusted according to the subsequent development of these cases. When the activity evolution characteristics exceed the optimized early warning threshold, different levels of early warning signals are triggered according to the exceeding degree. The first level of early warning is a mild abnormality, the second level of early warning is a moderate abnormality, and the third level of early warning is a serious abnormality. At the same time, the abnormal characteristic spectrum is recorded, including the time, position, duration and evolution trend of the abnormality, which provides a basis for quality traceability.

[0107] Referring to Figure 2 As shown in FIG. 1, a quality intelligent detection system for micro-mineral organic fertilizer includes:

[0108] A dynamic excitation and multi-source signal acquisition module applies a preset mode of energy excitation and synchronously acquires the multi-source dynamic response signals of the fertilizer sample within the excitation period, to construct a time sequence feature set containing a gas release kinetics curve and a spectral change trajectory;

[0109] A transient feature extraction and visualization module is used for differential transformation and modal decomposition of the time sequence feature set, to extract a transient response feature vector representing the biological metabolic intensity and generate a corresponding activity response distribution map;

[0110] A metabolic fingerprint decoupling and analysis module establishes a component-activity correlation model based on the activity response distribution map, to obtain a dynamic metabolic fingerprint reflecting the real activity of microorganisms by decoupling the biological metabolic activity and the matrix background interference;

[0111] A multi-modal feature fusion and quality index generation module, which constructs dynamic metabolic fingerprints and sample apparent characteristic parameters into a multi-dimensional feature space, adopts a graph convolution network to aggregate information and reconstruct features of feature nodes, and generates a comprehensive quality index;

[0112] An intelligent decision-making and dynamic early warning module, which generates a quality grade determination result based on the comprehensive quality index, simultaneously analyzes time sequence evolution characteristics of the activity response distribution graph, and establishes a dynamic quality early warning threshold.

[0113] The working principle of the present application is as follows: a gas-spectrum coupled data cube is constructed by applying a preset mode of micro-zone heat pulse excitation and synchronously collecting gas release kinetics curves and spectral change trajectories of the fertilizer sample; then, a time sequence feature set is subjected to differential transformation and modal decomposition, a transient response feature vector is extracted, and an activity response distribution graph is generated; then, component-activity correlation analysis is established, biological metabolic activity is decoupled from matrix background interference through feature separation and contrast learning strategies, and a dynamic metabolic fingerprint is obtained; further, the dynamic metabolic fingerprint and the apparent characteristic parameters are constructed into a multi-dimensional feature space, a graph convolution network is adopted for feature fusion to generate a comprehensive quality index; finally, quality grade determination is performed based on a multi-level feature pyramid structure, a dynamic early warning mechanism is established by analyzing time sequence evolution characteristics of the activity response distribution graph, early discovery and early warning of quality abnormalities are achieved.

[0114] The above describes one embodiment of the present application in detail, but the content described is only a preferred embodiment of the present application and cannot be considered as limiting the implementation scope of the present application. Any equivalent changes and improvements made within the scope of the present application should still belong to the patent coverage range of the present application.

Claims

1. A method for intelligent quality detection of micro-mineral bio-organic fertilizer, characterized in that, Includes the following steps: S1: By applying energy excitation in a preset mode and simultaneously acquiring multi-source dynamic response signals of fertilizer samples during the excitation period, a time-series feature set containing gas release kinetic curves and spectral change trajectories is constructed. S2: Perform differential transformation and mode decomposition on the time-series feature set to extract transient response feature vectors characterizing the intensity of biological metabolism and generate the corresponding activity response distribution map; S3: Based on the activity response distribution map, a component-activity correlation model is established. By decoupling biological metabolic activity from matrix background interference, a dynamic metabolic fingerprint reflecting the true activity of microorganisms is obtained. S4: Construct a multi-dimensional feature space by combining dynamic metabolic fingerprints and sample apparent characteristic parameters, and use graph convolutional networks to aggregate information and reconstruct features of feature nodes to generate comprehensive quality indicators. S5: Generate quality level judgment results based on comprehensive quality indicators, and analyze the temporal evolution characteristics of the active response distribution map to establish a dynamic quality early warning threshold.

2. The intelligent quality detection method for micro-mineral bio-organic fertilizer according to claim 1, characterized in that, The construction of the time-series feature set, which includes gas release kinetic curves and spectral change trajectories, specifically includes: A micro-area thermal pulse excitation device was used to apply local thermal excitation to the sample surface with a duration of milliseconds and a temperature change range of three to five degrees Celsius. Hyperspectral image sequences of specific areas of the sample surface were acquired before and after the excitation. By arranging an array of gas sensors around the sample space, the instantaneous changes in the concentration of various characteristic gases are monitored synchronously at a millisecond sampling frequency, and the gas release rate curves and their spectral characteristics are recorded. The hyperspectral image sequence was time-stamped with the gas release rate curve, and the gradient of spectral reflectance change at each sampling time was extracted to construct a gas-spectral coupled data cube containing spatiotemporal features. The gas-spectral coupled data cube is decomposed into multi-scale components to extract fast response components and slow-varying background components, and the fast response components are reconstructed into a time-series feature set reflecting the dynamics of microbial metabolism.

3. The intelligent quality detection method for micro-mineral bio-organic fertilizer according to claim 1, characterized in that, The step of performing differential transformation and mode decomposition on the time-series feature set to extract transient response feature vectors characterizing the intensity of biological metabolism specifically includes: Higher-order differential processing is performed on gas release kinetic curves with time-series feature sets to calculate their first-order and second-order differential sequences, thereby capturing abrupt changes and trends in gas release rates. Adaptive intrinsic mode decomposition is performed on the spectral change trajectory to decompose the non-stationary signal into multiple intrinsic mode components at different time scales, and feature modes that match the microbial metabolic cycle are screened. Extract the amplitude features of abrupt change points and the energy distribution parameters of characteristic modes, and construct a multi-dimensional transient response feature vector containing time-domain abrupt change features and frequency-domain energy features; Dimensionality reduction is performed on the multidimensional transient response feature vector to obtain the core feature subset characterizing the intensity of biological metabolism.

4. The intelligent quality detection method for micro-mineral bio-organic fertilizer according to claim 1, characterized in that, The generation of the corresponding activity response distribution map specifically includes: The transient response feature vectors are reorganized according to the sampling spatial location to construct a feature intensity matrix based on a spatial grid; The radial basis function interpolation method is used to spatially interpolate the characteristic intensity matrix to generate a continuously distributed contour map of activity intensity. Based on the gradient change characteristics of each region in the activity intensity contour map, response regions of different activity levels are divided, and metabolic hotspot regions are marked. By superimposing and fusing the spatial distribution of gas release characteristics and spectral change characteristics, an activity response distribution map that comprehensively reflects the spatial distribution and metabolic intensity of microorganisms is generated.

5. The intelligent quality detection method for micro-mineral bio-organic fertilizer according to claim 1, characterized in that, The component-activity correlation model established based on the activity response distribution map specifically includes: Using the spatial gradient features and intensity distribution features in the activity response distribution map as input, a dual-channel feature matrix containing matrix background features and metabolic activity features is constructed. An adversarial Siamese network model is established. The generator network learns to separate matrix background features from the dual-channel feature matrix, while the discriminator network is used to identify the integrity of metabolic activity features. The separated metabolic activity features are input into the residual mapping network to establish a nonlinear mapping relationship between metabolic activity intensity and microbial quantity, and output a pure activity feature vector that removes matrix interference. The residual mapping network is optimized by a contrastive learning strategy to ensure that the output features of the same activity level remain consistent under different matrix backgrounds.

6. The intelligent quality detection method for micro-mineral bio-organic fertilizer according to claim 1, characterized in that, The obtained dynamic metabolic fingerprint, which reflects the true activity of microorganisms, specifically includes: The pure activity feature vectors are recombined according to the time series to construct a dynamic metabolic trajectory matrix; Singular value decomposition is performed on the dynamic metabolic trajectory matrix to extract intrinsic metabolic patterns that characterize the main features of the metabolic process. The intrinsic metabolic pattern is dynamically time-warped and matched with the standard metabolic template to calculate the metabolic pathway similarity index. Based on the metabolic pathway similarity index and the energy distribution of intrinsic metabolic patterns, a dynamic metabolic fingerprint with unique identifiers is generated.

7. The intelligent quality detection method for micro-mineral bio-organic fertilizer according to claim 1, characterized in that, The generation of comprehensive quality indicators specifically includes: The various feature dimensions of the dynamic metabolic fingerprint and the apparent characteristic parameters of the sample are used to construct feature nodes, and node connection edges are established based on the mutual information between features to form a dynamic feature topology graph. By aggregating the neighborhood information of each feature node layer by layer through graph convolutional layers, global related features are fused while local features are preserved. The weights in the information transmission process are dynamically adjusted based on the contribution of feature nodes to the quality assessment. Global pooling is performed on feature nodes that have undergone multi-layer graph convolution processing, and the output is a comprehensive quality index that includes local features and global correlations.

8. The intelligent quality detection method for micro-mineral bio-organic fertilizer according to claim 1, characterized in that, The generation of quality grade determination results based on comprehensive quality indicators specifically includes: Construct a hierarchical decision tree based on the quality feature pyramid, and input the comprehensive quality indicators into the corresponding decision nodes according to the hierarchical structure of the quality feature pyramid; At each decision node, an adaptive membership function is used to perform fuzzy quantization on the input features, and the probability distribution of each quality level is calculated. The high-frequency fluctuations of the underlying micro-characteristics are integrated with the stable trends of the top-level macro-characteristics for analysis. Based on the comparison between the final confidence score and the preset level threshold, the quality level determination result with confidence assessment is output.

9. The intelligent quality detection method for micro-mineral bio-organic fertilizer according to claim 1, characterized in that, The analysis of the temporal evolution characteristics of the active response distribution map and the establishment of a dynamic quality early warning threshold specifically include: Spatial distribution characteristic parameters of the activity response distribution map during continuous detection were extracted, including activity intensity gradient, area of ​​metabolic hotspot region and spatial heterogeneity index; Calculate the differential characteristics of spatial distribution feature parameters over time to identify trend changes and abnormal fluctuation patterns in the activity evolution process; Based on the acceleration characteristics of trend changes and the statistical characteristics of abnormal fluctuations, a dynamic early warning threshold curve is constructed. By comparing and analyzing historical early warning data with current monitoring data, the morphological parameters of the early warning threshold curve are optimized. When the activity evolution characteristics exceed the early warning threshold, a graded early warning signal is triggered and the abnormal feature map is recorded.

10. A quality intelligent detection system for micro-mineral bio-organic fertilizer, characterized in that, A method for intelligent quality detection of a micro-mineral bio-organic fertilizer according to any one of claims 1-9 includes: The dynamic excitation and multi-source signal acquisition module applies energy excitation in a preset mode and simultaneously acquires multi-source dynamic response signals of fertilizer samples during the excitation period, constructing a time-series feature set including gas release kinetic curves and spectral change trajectories. The transient feature extraction and visualization module is used to perform differential transformation and mode decomposition on the time-series feature set, extract transient response feature vectors that characterize the intensity of biological metabolism, and generate corresponding activity response distribution maps. The metabolic fingerprint decoupling and analysis module establishes a component-activity correlation model based on the activity response distribution map. By decoupling biological metabolic activity from matrix background interference, a dynamic metabolic fingerprint reflecting the true activity of microorganisms is obtained. The multimodal feature fusion and quality index generation module constructs a multidimensional feature space from dynamic metabolic fingerprints and sample phenotypic parameters, and uses a graph convolutional network to aggregate information and reconstruct features from feature nodes to generate a comprehensive quality index. The intelligent decision-making and dynamic early warning module generates quality level judgment results based on comprehensive quality indicators, and analyzes the temporal evolution characteristics of the active response distribution map to establish a dynamic quality early warning threshold.

Images