Intelligent acceptance and sorting method for sea-weed raw material

Abstract

The application discloses a seaweed raw material intelligent acceptance and sorting method, and belongs to the image recognition technical field based on machine learning. The method comprises the following steps: collecting multi-modal data, hyperspectral texture maps and three-dimensional point cloud data of raw materials through a deployment device; constructing three data processing mechanisms; respectively extracting pulse timing features, generating steady-state pattern features, and extracting geometric features, and then fusing the three features through a parameterized delayed coupled oscillator; performing factor spin mapping and non-equilibrium free energy minimum solving on the space-time fused feature vector; dynamically exploring and performing robust optimization on the comprehensive score of seaweed quality grades to obtain a real-time sorting threshold adjustment amount; performing air nozzle mapping blowing and data aggregation processing on the real-time sorting threshold adjustment amount to obtain a sorting bin material execution instruction and execute the same, and finally generating an acceptance report for uploading. The application aims to realize precise, robust and automated intelligent sorting of seaweed raw materials.

Description

Technical Field

[0001] This invention relates to the field of image recognition technology based on machine learning, and in particular to a method for intelligent acceptance and sorting of seaweed raw materials. Background Technology

[0002] Nori is a processed seafood product made from seaweed such as Porphyra yezoensis and Porphyra tenera through washing, spreading, baking, and seasoning. In recent years, with the expansion of diverse consumption scenarios such as ready-made dishes, children's complementary foods, and rice ingredients, the scale of China's seaweed farming and deep processing industry has continued to climb. Coastal production areas have formed a complete upstream and downstream industrial chain, with an annual processing volume of over one million tons of dried seaweed raw materials. The quality of seaweed raw materials directly determines the color, taste, shelf life, and food safety level of the finished nori. Affected by seawater temperature, red tides, and farming cycles, the raw materials entering the factory are mixed with moldy algae, debris, farming nets, shells, hair, plastic filaments, and other foreign objects. The thickness, moisture content, and protein content of seaweed vary significantly at different harvesting periods. If the raw materials are not accurately graded and mixed for processing, batch quality defects such as uneven baking, color difference in finished products, and excessive peroxide value will occur. Therefore, the industry's demand for standardization and automation of raw material quality grading, foreign object removal, and safety acceptance continues to surge.

[0003] Current sorting technologies still have several problems in intelligent sorting: 1. The lack of modeling ability for the dynamic optical properties and micro-texture physical evolution process of seaweed leads to insufficient physical perception; 2. The inability to model high-order statistical dependencies between factors and batch distribution drift leads to low sorting accuracy when multiple defects coexist; 3. The inability to perform threshold adaptive optimization based on incoming material quality trend prediction easily leads to missed detections. Summary of the Invention

[0004] This invention discloses an intelligent acceptance and sorting method for seaweed raw materials. By constructing a hybrid intelligent architecture that integrates pulse-coupled neural fields, differentiable reaction diffusion equations, and non-equilibrium spin glass complex free energy decision-making, it solves the technical problems existing in the prior art. Therefore, the present invention provides the following technical solution: The present invention provides an intelligent acceptance and sorting method for seaweed raw materials, the method comprising: Collect multimodal data, hyperspectral texture maps, and 3D point cloud data of the raw materials; Construct a two-layer pulse-coupled neural field, a differentiable reaction diffusion equation, and a dynamic graph convolutional network; Pulse-coupled neural fields are used to process multimodal data and extract pulse temporal features. Differentiable reaction diffusion equations are used to process hyperspectral texture maps and generate steady-state pattern features. Dynamic graph convolutional networks are used to process 3D point cloud data and extract geometric features. The three features are fused through a parameterized delay-coupled oscillator to generate a spatiotemporal fusion feature vector. Using a spin glass free energy minimization method based on learnable duplicate breaking and quantum perturbation attention, factor spin mapping and non-equilibrium free energy minimization are performed on the spatiotemporal fusion feature vector to obtain a comprehensive score for seaweed quality grade. Dynamic exploration and robust optimization of the comprehensive score of seaweed quality grade were carried out to obtain the real-time sorting threshold adjustment amount; The real-time sorting threshold adjustment is processed by air nozzle mapping and data aggregation to obtain and execute the material execution instructions for the sorting bin, and finally, an acceptance report is generated and uploaded.

[0005] Furthermore, the multimodal data includes a hyperspectral data cube, an RGB image, and batch attribute information; the hyperspectral texture map is derived from the hyperspectral data cube; and the three-dimensional point cloud data includes geometric point clouds.

[0006] Furthermore, the process of processing multimodal data using pulse-coupled neural fields includes: By differentiable spectral unmixing, the hyperspectral cube is transformed into an abundance map and stitched with the RGB image to obtain the coding layer pulse sequence, which is used as the input of the double-layer pulse-coupled neural field coding layer. By adjusting the synaptic weights of converging layer neurons on the pulse sequence of the coding layer, pulse firing rate, firing delay, and burst interval are extracted to form pulse timing features.

[0007] Furthermore, the process of processing hyperspectral texture maps using differentiable reaction diffusion equations includes: Using multi-scale Gabor filtering and convolution dimensionality reduction methods, the local texture of the hyperspectral cube is calculated to generate a single-channel hyperspectral texture map, and the state field of the reaction-diffusion system is initialized. The coefficients of the equations for the hyperspectral texture map and the thickness gradient field are dynamically generated and spatiotemporally evolved to extract steady-state pattern features.

[0008] Furthermore, the process of processing 3D point cloud data using a dynamic graph convolutional network includes: A geodesic dynamic graph convolutional network with learnable kernel width is used to perform shape topological encoding on the contour map generated from 3D point cloud data, and extract 3D morphological feature vectors.

[0009] Furthermore, the factor spin mapping and the solution for minimizing nonequilibrium free energy include: Using a parallel lightweight decoding head and a GBDT regressor improved by feature reproduction adaptive regularization, the spatiotemporal fusion feature vector is decoded to predict five continuous quality factors, which are then converted into spin vectors by parameterized tanh mapping. Quantum noise-enhanced inter-node attention calculation and symmetry are performed on the spin vector to obtain the factor-symmetric interaction matrix and external field vector; By substituting the symmetric interaction matrix and external field vector into the spin glass Hamiltonian, and using a meta-network to adaptively predict the temperature and Parisi fracture parameters based on factor skewness and kurtosis, a free energy variational functional is constructed.

[0010] Furthermore, the factor spin mapping and the solution for minimizing nonequilibrium free energy also include: Using a free energy variational functional, gradient descent minimization is performed on magnetization and overlap to obtain equilibrium order parameters. The optimal spin configuration and corresponding order reference configuration of the sample are then reconstructed based on the equilibrium order parameters. The optimal spin configuration of the sample and the corresponding grade reference configuration are used to calculate the similarity and adaptive penalty correction to obtain the comprehensive score of seaweed quality grade.

[0011] Furthermore, the dynamic exploration and adversarial robustness optimization includes: The comprehensive score of each piece of seaweed quality grade is concatenated with the five-dimensional quality factors to form a six-dimensional state vector, and a fixed-length sliding window is constructed along the processing sequence to form a quality state window matrix; The LSTM time-series evolution of the quality state window matrix with chaotic gated bias modulation outputs the mean, variance, and volatility index of the future quality score. A heuristic information matrix is ​​constructed based on the mean and variance of future quality scores. The probability of action selection is calculated by combining the pheromone concentration, and the sorting threshold adjustment action is output.

[0012] Furthermore, the dynamic exploration and adversarial robustness optimization also includes: Construct a Critic network to evaluate the sorting threshold adjustment action by combining asymmetric penalties for false rejections and missed detections, grade ratio deviation and prediction consistency rewards, and calculate the robust TD error by minimizing the Q value against perturbations, and output the action value. The loss function for the Critic network is defined as follows: In the formula L Q Let be the loss function of the Critic network; s, a, and r are the reinforcement learning state vector, sorting threshold adjustment action, and immediate reward at the current time step, respectively; s 1 The state to which the action is performed is the next state; γ is the discount factor. To calculate the expectation of the mini-batch transfer tuples sampled from the experience replay pool; Q(s,a;ψ) is the state-action value estimate of the current Critic network output; ψ - The parameters of the target Critic network; Q(s) 1 ,a 1 ;ψ - ) is to counteract the robustness term; λ 正则For learnable robust regularization coefficients; η is the infinite norm of the Jacobian matrix of Q-value with respect to state input s; η is the adversarial perturbation vector; min is the minimum function; agrmax is the function of the maximum value of the independent variable. This is the optimal action for the next state as perceived by the current network. The minimum value of Q after applying the worst-case perturbation to the next state; The cumulative reward maximization optimization of the action value evaluated by Critic is performed under the constraint of predictive consistency. The ant colony chaos parameters are dynamically updated and the real-time sorting threshold adjustment is output.

[0013] Furthermore, the nozzle mapping and data aggregation processing includes: The real-time sorting threshold adjustment is converted into material execution instructions for the sorting bins of the air nozzle array, completing the four-level pneumatic sorting and warehousing of seaweed raw materials. Compile statistical analysis of execution results, generate an acceptance report, and upload it.

[0014] Compared with the prior art, the present invention achieves at least one of the following beneficial effects: 1. A two-layer pulse-coupled neural field is constructed to encode the hyperspectral abundance map and RGB image with pulse sequences. At the same time, a parameterized differentiable reaction diffusion equation is constructed to dynamically generate equations based on the hyperspectral texture map and thickness gradient field and perform spatiotemporal evolution. The steady-state pattern of morphogenesis is extracted. The pulse sequence and steady-state pattern are fused through a delayed coupled oscillator to generate a spatiotemporal fusion feature vector that simultaneously contains chromatographic time-series dynamics, texture morphogenesis mode and geometric topology information. This solves the problem of insufficient physical perception caused by the lack of modeling ability of existing technology for the dynamic optical properties and physical evolution process of seaweed dynamics and microtexture.

[0015] 2. By mapping multidimensional quality factors to spin variables, a symmetric interaction matrix and external field vector between factors are dynamically generated. Temperature and Parisi fracture parameters are adaptively predicted based on batch factor skewness and kurtosis through a meta-network. A differentiable free energy variational functional is constructed and the optimal spin configuration is solved by minimizing it. Non-equilibrium entropy is introduced to generate a penalty term to correct the grade decision. This achieves the solution of minimizing the free energy of complex nonlinear coupling between factors and batch adaptive non-equilibrium state, and outputs accurate quality grade labels and continuous comprehensive scores. This solves the problem of low sorting accuracy when multiple defects coexist, which is caused by the inability of existing technologies to model high-order statistical dependencies between factors and batch distribution drift.

[0016] 3. The quality scoring sequence of continuous sheet materials is constructed as a one-dimensional differentiable cellular automaton. LSTM with a forget gate and input gate modulated by a hyperbolic chaotic mapping is used to predict future quality states. In the action selection for threshold adjustment, a pheromone ant colony reinforcement learning framework is introduced to adaptively adjust the selection temperature using Q-value gradients. A robust reward function is constructed using an asymmetric learnable penalty factor. Chaotic parameters, pheromone evaporation coefficients, and reinforcement learning network parameters are updated collaboratively through meta-gradients. The output is a sorting threshold adjustment amount that is optimized in real time according to batch quality fluctuations, ensuring dynamic balance in sorting. This solves the problem that existing technologies cannot adaptively optimize thresholds based on incoming material quality trends, which easily leads to missed detections.

[0017] 4. Learnable kurtosis and offset parameters are introduced into the factor spin mapping. The asymmetric coupling strength between factors is automatically learned through the quantum perturbation graph attention mechanism. The meta-network adjusts the replica fracture parameter and system temperature online according to batch skewness and kurtosis, realizing unsupervised dynamic allocation of factor importance. This eliminates the subjective bias of human experience weighting and enables the automatic adjustment of the decision contribution of each factor for different seaweed varieties and quality fluctuations. This solves the limitations of fixed weights and linear separability in the existing multi-index evaluation technology. Attached Figure Description

[0018] To more clearly illustrate the technical solutions in the embodiments of this application, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0019] Figure 1 This is a flowchart illustrating the intelligent acceptance and sorting method for seaweed raw materials according to the present invention. Figure 2 This is a pie chart showing the grade distribution of the present invention; Figure 3 This is a Pareto chart showing the defects of the present invention. Detailed Implementation

[0020] In the following description, specific details such as particular system architectures and techniques are set forth for illustrative purposes and not for limitation, in order to provide a thorough understanding of the embodiments of this application. However, those skilled in the art will understand that this application may also be implemented in other embodiments without these specific details. In other instances, detailed descriptions of well-known systems, apparatuses, circuits, and methods have been omitted so as not to obscure the description of this application with unnecessary detail.

[0021] It should be understood that, when used in this application specification and the appended claims, the term "comprising" indicates the presence of the described features, integrals, steps, operations, elements and / or components, but does not exclude the presence or addition of one or more other features, integrals, steps, operations, elements, components and / or a collection thereof.

[0022] It should also be understood that the term “and / or” as used in this application specification and the appended claims means any combination of one or more of the associated listed items and all possible combinations, and includes such combinations.

[0023] As used in this application specification and the appended claims, the term "if" may be interpreted, depending on the context, as "when," "once," "in response to determination," or "in response to detection." Similarly, the phrase "if determined" or "if detected [the described condition or event]" may be interpreted, depending on the context, as meaning "once determined," "in response to determination," "once detected [the described condition or event]," or "in response to detection [the described condition or event]."

[0024] Furthermore, in the description of this application and the appended claims, the terms "first," "second," "third," etc., are used only to distinguish descriptions and should not be construed as indicating or implying relative importance.

[0025] References to "one embodiment" or "some embodiments" as described in this specification mean that one or more embodiments of this application include a specific feature, structure, or characteristic described in connection with that embodiment. Therefore, the phrases "in one embodiment," "in some embodiments," "in other embodiments," "in still other embodiments," etc., appearing in different parts of this specification do not necessarily refer to the same embodiment, but rather mean "one or more, but not all, embodiments," unless otherwise specifically emphasized. The terms "comprising," "including," "having," and variations thereof mean "including but not limited to," unless otherwise specifically emphasized.

[0026] Example 1: Please see the appendix Figure 1 The present invention relates to an intelligent acceptance and sorting method for seaweed raw materials, the method comprising: Collect multimodal data, hyperspectral texture maps, and 3D point cloud data of the raw materials; Construct a two-layer pulse-coupled neural field, a differentiable reaction diffusion equation, and a dynamic graph convolutional network; Pulse-coupled neural fields are used to process multimodal data and extract pulse temporal features. Differentiable reaction diffusion equations are used to process hyperspectral texture maps and generate steady-state pattern features. Dynamic graph convolutional networks are used to process 3D point cloud data and extract geometric features. The three features are fused through a parameterized delay-coupled oscillator to generate a spatiotemporal fusion feature vector. Using a spin glass free energy minimization method based on learnable duplicate breaking and quantum perturbation attention, factor spin mapping and non-equilibrium free energy minimization are performed on the spatiotemporal fusion feature vector to obtain a comprehensive score for seaweed quality grade. Dynamic exploration and robust optimization of the comprehensive score of seaweed quality grade were carried out to obtain the real-time sorting threshold adjustment amount; The real-time sorting threshold adjustment is processed by air nozzle mapping and data aggregation to obtain and execute the material execution instructions for the sorting bin, and finally, an acceptance report is generated and uploaded.

[0027] Its core technologies include: 1- Neuromorphic and physical-inspired multimodal dynamic feature encoding; 2- Non-equilibrium hybrid factor decision-making based on spin glass and duplicate theory; 3- Dynamic threshold optimization of chaotic cell prediction and pheromone enhancement learning.

[0028] The core concept of this invention is as follows: It abandons traditional static convolution feature extraction and uses pulse-coupled neural fields and differentiable reaction diffusion equations, combined with dynamic graph convolution, to map hyperspectral, RGB, and point cloud data into unified features containing temporal dynamics and morphogenesis. It models seaweed quality evaluation as a physical spin system with complex interactions, using statistical mechanics replication methods to solve for optimal decisions, fundamentally breaking the factor independence assumption. It predicts quality trends through differentiable cellular automata and combines chaotic ant colony and reinforcement learning to dynamically adjust sorting thresholds online, achieving automated optimal control. The overall architecture is a closed-loop link, a hybrid architecture, exhibiting dynamic adaptability to the detection rate and sorting of complex defects.

[0029] In specific implementation, the multimodal data includes a hyperspectral data cube, an RGB image, and batch attribute information; the hyperspectral texture map is derived from the hyperspectral data cube; and the three-dimensional point cloud data includes geometric point clouds.

[0030] Its data acquisition hardware includes: 1-line array hyperspectral camera; 2-color line scan camera; 3-laser triangulation rangefinder; 4-conveyor belt encoder; 5-PLC and batch information module; installed above the seaweed sorting conveyor belt, the following sensors are installed in sequence along the material's forward direction, and the fields of view are ensured to overlap on the same transverse section.

[0031] In specific implementation, the process of using pulse-coupled neural fields to process multimodal data includes: By differentiable spectral unmixing, the hyperspectral cube is transformed into an abundance map and stitched with the RGB image to obtain the coding layer pulse sequence, which is used as the input of the double-layer pulse-coupled neural field coding layer. By adjusting the synaptic weights of converging layer neurons on the pulse sequence of the coding layer, pulse firing rate, firing delay, and burst interval are extracted to form pulse timing features.

[0032] It involves converting a hyperspectral cube into an abundance map and stitching it with an RGB image through differentiable spectral unmixing, including: Step 1) Perform differentiable spectral unmixing on the hyperspectral cube to obtain the abundance map (the unmixing uses a linear model). Step 2) Based on Step 1), the abundance map and the RGB image are stitched together by channel to form a multispectral-visual joint input; Step 3) Based on Step 2), construct a two-layer two-dimensional pulse-coupled neural network. The first layer is the encoding layer, consisting of W×H spiking neurons, each corresponding to a spatial location of the joint input. The second layer is the convergence layer, with neurons arranged spatially in the same way as the encoding layer. It receives the pulse output from the encoding layer and generates higher-order temporal features. Both layers of neurons use the parametric leaky integral firing (LIF) model. The membrane potential of the encoding layer neurons satisfies the following equation: In the formula τ is the time derivative of the membrane potential (used to describe how the membrane potential continuously evolves under the combined influence of resting leakage, synaptic input, and noise); m V is the membrane time constant of the neuron; ij (t) represents the membrane potential of the coding layer neuron; V 静息 This is the resting potential; For the synaptic input current; η ij (t) represents the independent Gaussian white noise term; where τ m The local humidity sensitivity coefficient corresponding to the spatial location is dynamically generated (derived from the unmixed moisture endmember abundance through linear mapping); the synaptic input current consists of a feedforward input and a neighborhood recursive connection, and its calculation formula is as follows: In the formula W 前馈 X is the feedforward input weight matrix; ij is the multimodal joint input vector at spatial location (i,j) (obtained by concatenating the hyperspectral abundance vector with the RGB three-channel values); N(i,j) is the local neighborhood set at spatial location (i,j); (k,l) is the spatial coordinates of a neighboring neuron within the neighborhood set; w ij,kl d represents the recursive synaptic weights from neurons (k,l) to (i,j); ij,kl S represents the conduction delay from neuron (k,l) to (i,j); kl (tdij,kl ) represents the delayed neighborhood pulse.

[0033] The regulation of synaptic weights of convergent layer neurons to coding layer pulse sequences includes: constructing convergent layer LIF neurons to receive coding layer pulse-weighted inputs; their synaptic weights are adjusted online through triple time-dependent plasticity (T-STDP); the T-STDP rule not only considers the time difference between pre- and post-synaptic pulses but also introduces a transiently variable calcium ion concentration proxy variable; the dynamic equation for post-synaptic calcium ion concentration is: In the formula c ij τ is a proxy variable for the postsynaptic calcium ion concentration located at neuron (i,j); c t is the calcium decay time constant; t is a continuous time variable; Let f be the firing time of the f-th impulse of neuron (i,j); δ is the Dirac function; d is the differential operator; dt is a very short time interval; is the first derivative of the postsynaptic calcium ion concentration of neuron (i,j) with respect to time t (used to characterize the net rate of change of calcium concentration at each instant). The linear decay rate is caused by calcium ion efflux or intracellular buffering mechanisms; its synaptic weight update formula set is as follows: , where Δw ij,kl A represents the change in recurrent synaptic weights from neuron (k,l) to (i,j); + and A - These are the enhancement and suppression amplitude factors (controlling long-term enhancement and long-term suppression, respectively); Δt is the postsynaptic pulse duration; τ + and τ - These represent the widths of the enhancement and suppression time windows, respectively. and These are exponentially decaying kernels for the cases Δt>0 and Δt<0, respectively; c kl The calcium ion concentration of the postsynaptic neuron is used as a proxy variable; the converging layer neurons also follow LIF dynamics, and their thresholds are adjusted by a learnable interlayer scaling factor. The pulse firing rate, firing delay and burst interval of the converging layer in time T are recorded to form the pulse timing characteristics.

[0034] In specific implementation, the process of processing hyperspectral texture maps using differentiable reaction diffusion equations includes: Using multi-scale Gabor filtering and convolution dimensionality reduction methods, the local texture of the hyperspectral cube is calculated to generate a single-channel hyperspectral texture map, and the state field of the reaction-diffusion system is initialized. The coefficients of the equations for the hyperspectral texture map and the thickness gradient field are dynamically generated and spatiotemporally evolved to extract steady-state pattern features.

[0035] It calculates the local texture of the hyperspectral cube, using the multi-scale Gabor texture features within a local window of r×r centered at each spatial location (i,j) as the size, to obtain a texture descriptor map. The texture descriptor map is then reduced to a single-channel hyperspectral texture map through a 1×1 convolutional layer. Two state variables, U(x,y,t) and V(x,y,t), are defined to constitute the state field of the reaction-diffusion system. The initial conditions are set as follows: activator concentration field U(x,y,t)=1 and inhibitor concentration field V(x,y,t)=0. The spatiotemporal evolution domain is the two-dimensional grid corresponding to the texture map, thus completing the initialization of the state field of the reaction-diffusion system.

[0036] Its coefficients for the dynamic generation equation of hyperspectral texture map and thickness gradient field are given by the reaction-diffusion partial differential equation, which is defined as: In the formula, U and V are the activator concentration field and the suppressor concentration field, respectively; D u and D v These represent the activator diffusion coefficient and the repressor diffusion coefficient, respectively; F is the feed rate; k is the extinction rate; UV 2 For autocatalytic reaction terms; F(1-U) and (F+k)V are the source and sink terms, respectively; the key parameter modification of this invention is to parameterize the feed rate F and the extinction rate k as functions of the hyperspectral texture map, and to change the activator diffusion coefficient and the suppressor diffusion coefficient to anisotropic tensors, which are dynamically modulated by the local thickness gradient and curvature; using a differentiable numerical solver (such as a semi-implicit spectral method or convolutional recursive network simulation) to iteratively solve within a fixed time range to obtain the final activator concentration field pattern, and to perform global average pooling and multi-scale statistical moment extraction on the steady-state activator concentration field to obtain the morphogenetic texture feature vector as the steady-state pattern feature.

[0037] In specific implementation, the process of using a dynamic graph convolutional network to process 3D point cloud data includes: A geodesic dynamic graph convolutional network with learnable kernel width is used to perform shape topological encoding on the contour map generated from 3D point cloud data, and extract 3D morphological feature vectors.

[0038] Its shape topology encoding of the contour map generated from 3D point cloud data includes: Step 1) Project the 3D point cloud data onto the conveyor belt plane and interpolate to generate a thickness map aligned with the hyperspectral space; Step 2) Based on Step 1), extract the seaweed outline from the thickness map, and sample N marker points on the outline at equal angles or equal arcs to construct a K-nearest neighbor graph. The features of the G nodes are initialized with the coordinates and thickness values ​​of the corresponding points. Step 3) Based on Step 2), shape encoding is performed using parametric geodesic kernel dynamic graph convolution. The feature update formula for the shape encoding nodes is: In the formula Let i be the updated feature vector of node i in the l-th layer; Let i be the original feature vector of node i in the l-th layer; Let i be the feature vector of the neighboring node i; Let be the feature vector of neighboring node j; Let N(i) be the feature difference vector of node i; N(i) be the set of K nearest neighbors of node i; GeoDist(i,j) be the approximate geodesic distance between nodes (i,j) along the seaweed surface; MLP is a multilayer perceptron; the weights for neighborhood aggregation are dynamically adjusted using a Gaussian kernel with a learnable kernel width, and the weight formula is as follows: In the formula w ij p represents the neighborhood aggregation weight. i and p j , i and j are the two-dimensional and three-dimensional spatial coordinate vectors of node (i,j) respectively (used to locate the physical position of the contour marker point in the plane or space of the seaweed raw material conveyor belt); δ is the Gaussian kernel parameter (used to control the rate of weight decay with distance); 2δ 2 The scaling factor is used to smooth the decay curve. After multi-layer graph convolution, global max pooling is performed on the node features to obtain the shape topology feature vector (i.e., the three-dimensional morphological feature vector).

[0039] The above-mentioned three features, fused through a parameterized delay-coupled oscillator, include: Step 1) Project the pulse timing features, steady-state pattern features, and three-dimensional morphological feature vectors through the fully connected layer into vectors of the same dimension (considered as the initial state of the three coupled oscillators). Step 2) Based on Step 1), construct a parameterized delayed coupled oscillator network. Its coupled oscillator dynamics equations are as follows: In the formula W is the time derivative of the state (the direction and rate of state change are jointly determined by self-evolution, coupled input, and external driving forces); i K is the self-evolution matrix of oscillator i; ij Let τ be the scalar of the coupling strength from oscillator j to i; ij o is the coupling delay from j to i; i (t) is the state vector of the i-th oscillator at time t (i.e., the initial state of the three coupled oscillators); o i (t-τ ij ) represents the state of oscillator j at the delay time; I i (t) represents the external driving input; the core modification of this invention is that the coupling strength is dynamically adjusted by the mutual information between the input features. The mutual information is calculated using differentiable kernel density estimation. After a short time evolution, the synchronization degree of the oscillator state, the phase difference mode, and the final steady-state amplitude are taken as the spatiotemporal fusion feature vector.

[0040] In specific implementation, the factor spin mapping and the solution for minimizing non-equilibrium free energy include: Using a parallel lightweight decoding head and a GBDT regressor improved by feature reproduction adaptive regularization, the spatiotemporal fusion feature vector is decoded to predict five continuous quality factors, which are then converted into spin vectors by parameterized tanh mapping. Quantum noise-enhanced inter-node attention calculation and symmetry are performed on the spin vector to obtain the factor-symmetric interaction matrix and external field vector; By substituting the symmetric interaction matrix and external field vector into the spin glass Hamiltonian, and using a meta-network to adaptively predict the temperature and Parisi fracture parameters based on factor skewness and kurtosis, a free energy variational functional is constructed.

[0041] Its spatiotemporal fusion feature vector decoding predicts five consecutive quality factors, and the tanh mapping transformation includes: Step 1) Construct 5 parallel lightweight decoding heads. Each decoding head consists of two fully connected network layers and an improved GBDT regressor (the improvement lies in adding an adaptive regularization term for feature reproduction to the GBDT splitting criterion). Its five continuous quality factors include: 1-color uniformity factor; 2-shape integrity factor; 3-foreign object defect factor; 4-thickness consistency factor; 5-texture fineness factor. Step 2) Based on Step 1), the five continuous factors are converted into spin variables using a parameterized spin mapping function. The mapping formula is: s i =tanh(γ i (F i -b i )), where s i γ is the spin vector; i F is a learnable kurtosis parameter (used to control the slope of the mapping curve). i b is a continuous value of the i-th factor (i.e., one of the five continuous quality factors); i This is a learnable offset parameter.

[0042] Its quantum noise-enhanced inter-node attention computation and symmetry include: Step 1) Treat the five factors as nodes in a fully connected graph, and use the multi-head graph attention mechanism (GAT) to calculate the interaction strength between nodes. For attention head k, the original attention coefficient between nodes i and j is calculated using the following formula: In the formula The original attention coefficient; z i and z j W represents the embedding vectors for quality factors i and j, respectively; (k) Let a be the linear transformation matrix of the k-th head; (k)Let a be the attention weight vector for the k-th head; (k)T is the transpose of the attention weight vector of the k-th head; LeakyReLU is a modified linear unit with a negative slope; Step 2) Based on Step 1), a learnable transverse field perturbation and Gaussian noise are introduced into the original attention coefficients. The formula for the quantum perturbation term is as follows: In the formula The attention coefficient after perturbation; Γ is the original attention coefficient of the k-th head; Γ is the learnable transverse field strength scalar; ε ij The noise is independent and identically distributed standard Gaussian noise (with random perturbation); then, the normalized attention weights are obtained by applying softmax to the neighboring factors; Step 3) Based on Step 2), after symmetrization using the average multi-head output, the interaction matrix is ​​obtained. The joint calculation formula set for the interaction matrix is ​​as follows: In the formula α ij The attention weights are calculated after averaging across multiple heads; K represents the total number of attention heads. J represents the normalized attention weight of the k-th head quality factor j relative to quality factor i; ij J0 represents the symmetric interaction between mass factors i and j; J0 is a learnable global coupling strength scalar (used to control the strength of the overall coupling of the spin system). This is a symmetry operation (i.e., forcing the interaction matrix to be symmetric).

[0043] The definition formula for its spin glass Hamiltonian is: In the formula, H(s) is the total Hamiltonian of the system; h i h is the external field acting on the mass factor i; i s i This is the coupling term between the external field and the spin; For external field energy; This refers to the pairwise interaction energy (the contribution of the total energy from the coupling between factors, used to drive the coordinated change of the quality factors).

[0044] It uses a meta-network to adaptively predict temperature and Parisi rupture parameters based on factor skewness and kurtosis, including: Step 1) Introducing the replica method: Assume the system is at a finite temperature T, and use the single-step replica symmetric breaking (1RSB) framework. Introduce the Parisi breaking parameter pi to construct a meta-network, where temperature T and breaking parameter m are adaptively predicted through the non-Gaussian properties of batch mass. The calculation formulas are as follows: In the formula, T and P are the finite temperature and Parisi fracture parameters, respectively. A learnable weight vector for predicting rupture parameters; γ 偏κ is the skewness vector of the distribution of each factor within the batch; b is the kurtosis vector of the distribution of each factor within the batch. P The learnable bias is used to predict the fracture parameters; σ is the sigmoid function; softplus is the soft positive activation function; b is a learnable weight vector for predicting temperature. T A learnable bias for predicting temperature; Step 2), based on Step 1), construct the free energy variational functional formula as follows: In the formula F 自由 (q,m) is the unit spin free energy functional; β is the inverse temperature (1 / T); D z and D u These are the integrals over the random coupled field between duplicates and the random field within a single duplicate, respectively; q ij The overlap between replicas of quality factors (i,j) is given by P; P is the Parisi fracture parameter; H 有效 For an effective single-spin Hamiltonian; m i and m j These are the magnetization components for mass factors i and j, respectively; For external energy; This is the interaction energy correction term; is the M-th power of the Boltzmann weight within the copy block; m is the magnetization vector; q is the copy overlap; z is the external Gaussian random field; u is the internal thermal fluctuation Gaussian random field.

[0045] In specific implementation, the factor spin mapping and the solution for minimizing non-equilibrium free energy also include: Using a free energy variational functional, gradient descent minimization is performed on magnetization and overlap to obtain equilibrium order parameters. The optimal spin configuration and corresponding order reference configuration of the sample are then reconstructed based on the equilibrium order parameters. The optimal spin configuration of the sample and the corresponding grade reference configuration are used to calculate the similarity and adaptive penalty correction to obtain the comprehensive score of seaweed quality grade.

[0046] It performs gradient descent minimization on magnetization and overlap, and its iterative update rule formula set is as follows: In the formula m (g+1) and q (g+1) These represent the magnetization vector updated in step g+1 and the overlap of the duplicates, respectively; m (g) Let η be the magnetization vector at step g; m The learning rate is the magnetization intensity; ▽ m F 自由 (q,m) is the gradient vector of the free energy functional with respect to the magnetization m; η qThe learning rate is the amount of overlap. The analytical or automatic partial derivative of the free energy with respect to each magnetization component; the optimal effective spin configuration of the sample is reconstructed from the magnetization m.

[0047] Its similarity calculation and adaptive penalty correction include: Step 1), the formula for calculating the similarity (cosine similarity) of effective spin configurations is: , where sim g The cosine similarity to the rank g; s 优 s represents the optimal spin configuration vector for the current sample; (g) Let g be the standard reference spin configuration of the g-th level; where g ∈ [Special, First, Second, Unqualified]; Step 2), based on Step 1), introduces the following formula for the non-equilibrium free energy correction term: In the formula, Pen(s) 优 ) represents the non-equilibrium penalty; λ 平衡 The learnable unbalanced penalty coefficient; The square of the volatility index; Step 3), based on Step 2), calculate the final score using the formula: Score = sim g -Pen(s 优 The score is used as the overall rating for the quality of seaweed.

[0048] In specific implementation, the dynamic exploration and adversarial robustness optimization includes: The comprehensive score of each piece of seaweed quality grade is concatenated with the five-dimensional quality factors to form a six-dimensional state vector, and a fixed-length sliding window is constructed along the processing sequence to form a quality state window matrix; The LSTM time-series evolution of the quality state window matrix with chaotic gated bias modulation outputs the mean, variance, and volatility index of the future quality score. A heuristic information matrix is ​​constructed based on the mean and variance of future quality scores. The probability of action selection is calculated by combining the pheromone concentration, and the sorting threshold adjustment action is output.

[0049] Its LSTM timing evolution with chaotic gated bias modulation of the quality state window matrix includes: Step 1) Define the formula for generating modulation sequences using the Chebyshev-logic cascaded chaotic map as follows: In the formula c k+1 c is the chaotic sequence value at step k+1; k c1 is the chaotic sequence value at step k; c1 is the initial value of the chaotic sequence; μ 映射 λ is the control parameter for the Chebyshev mapping (greater than 1 indicates entry into the chaotic region, and the larger the value, the stronger the chaos); arccos is the inverse cosine function; λ映射 These are the control parameters for logical mapping; is the scaling factor for the logical term; mod is the function for finding the remainder of two numbers; Step 2) Construct a chaotic weighted LSTM for cell evolution (using standard LSTM gated updates). The improvement of this invention lies in the fact that in the calculation of the forget gate and input gate in the standard LSTM gated updates, the bias term no longer uses a fixed formula, but is modulated by the chaotic sequence. The calculation formula set is as follows: In the formula and b is the effective bias of the forget gate and input gate outputs modulated by the chaotic sequence at time t; f and b i These are the learnable forget gate fundamental bias vector and the input gate fundamental bias vector, respectively; a f and a i These are the learnable chaotic modulation amplitude vector and the input gate chaotic modulation amplitude vector, respectively; c k+1 The chaotic sequence value at step k+1 (obtained through step 1); Step 3) Use the historical window as the input sequence of the chaotic weighted LSTM to predict the quality status of the next K sheets in an autoregressive manner. Take the hidden state of the last chaotic weighted LSTM, output the average quality score of the future window through the linear prediction head, and calculate the fluctuation index.

[0050] It combines pheromone concentration to calculate the action selection probability, and uses a chaotic ant colony selection strategy to calculate the selection probability using the following formula: In the formula, p(a) is the probability of choosing action a; τ 信息素 (a) represents the pheromone concentration for action a; H a β is the heuristic information value for action a. p α is the heuristic information weighting index. p The pheromone weight index is used; the improvement of this invention lies in selecting the temperature parameter of the probability, which is adaptively adjusted by the Q-value gradient of the Critic network, and its formula set is as follows: In the formula, T 决策 The decision temperature for the ant colony; T0 is the base temperature; η T This is the temperature attenuation coefficient (used to control the scaling sensitivity of the Critic gradient with respect to temperature); ▽ a Q is the gradient of the action; the threshold of the current step is obtained based on probability sampling, the action is adjusted and executed to obtain a new threshold, which is used as the sorting threshold.

[0051] In practice, the dynamic exploration and adversarial robustness optimization also includes: Construct a Critic network to evaluate the sorting threshold adjustment action by combining asymmetric penalties for false rejections and missed detections, grade ratio deviation and prediction consistency rewards, and calculate the robust TD error by minimizing the Q value against perturbations, and output the action value. The loss function for the Critic network is defined as follows: In the formula L Q Let be the loss function of the Critic network; s, a, and r are the reinforcement learning state vector, sorting threshold adjustment action, and immediate reward at the current time step, respectively; s 1 The state to which the action is performed is the next state; γ is the discount factor. To calculate the expectation of the mini-batch transfer tuples sampled from the experience replay pool; Q(s,a;ψ) is the state-action value estimate of the current Critic network output; ψ - The parameters of the target Critic network; Q(s) 1 ,a 1 ;ψ - ) is to counteract the robustness term; λ 正则 For learnable robust regularization coefficients; η is the infinite norm of the Jacobian matrix of Q-value with respect to state input s; η is the adversarial perturbation vector; min is the minimum function; agrmax is the function of the maximum value of the independent variable. This is the optimal action for the next state as perceived by the current network. The minimum value of Q after applying the worst-case perturbation to the next state; The cumulative reward maximization optimization of the action value evaluated by Critic is performed under the constraint of predictive consistency. The ant colony chaos parameters are dynamically updated and the real-time sorting threshold adjustment is output.

[0052] Its asymmetric penalty for missed detections, the combined reward assessment of grade ratio bias and prediction consistency includes: Step 1) Define the reward function as having asymmetric robustness. Statistically analyze the actual sorting results within a subsequent fixed window (taking the next 100 pieces), including: the number of incorrectly rejected high-grade items, the number of missed defective items, and the deviation of each grade's proportion from the standard proportion. The formula for calculating the asymmetric robustness reward is: r t =-λ 误剔 ·N 误剔 -λ 漏检 ·N 漏检 -λ 偏差 ·N 偏差 In the formula r t λ is the immediate reward at step t (the target signal for reinforcement learning); 误剔 , λ 漏检 and λ 偏差These are the learnable error rejection penalty coefficient, the learnable missed detection penalty coefficient, and the fixed percentage deviation penalty coefficient, respectively; N 误剔 N 漏检 and N 偏差 These are the number of incorrectly rejected high-grade products, the number of missed defects, and the grade ratio deviation. Step 2) Simultaneously calculate the prediction consistency regularization term, the formula for which is: L 预测 =||E t+1 -E t || 2 ·||δ 预测 -δ 实际 || 2 In the formula L 预测 For prediction consistency regularization (used to add to the meta-objective function to coordinate the threshold adjustment strategy with prediction confidence); E t+1 The new threshold vector obtained after performing the action; E t This is the threshold vector from the previous step; δ 预测 To predict the quality fluctuation index (obtained by the chaotic weighted LSTM above); δ 实际 This refers to the quality fluctuation index of the actual observed batches.

[0053] Its cumulative reward maximization optimization under the predictive consistency constraint includes: Step 1) Sample mini-batch from the experience replay pool. The priority is composed of two parts, and the formula is as follows: P 优先 =|TD|+κ 方差 Var η [Q(s+η)], where P 优先 η is the priority weight of the empirical samples (determining the probability that the current sample will be sampled from the replay pool); η is the adversarial perturbation vector; TD is the temporal difference error (measuring the unexpectedness of the sample to the current Critic); Q(s+η) is the Q value under the perturbation state; Var η κ represents the variance of the Q-value under the perturbation distribution (reflecting the stability of the decision under that state); 方差 These are the learnable variance weighting coefficients; Step 2) Update the Critic parameter to minimize L Q ; Step 3) Update the ant colony and chaos-related parameters using the meta-gradient method, aiming to maximize the expected value of future cumulative rewards while maintaining prediction consistency. The meta-objective function is: In the formula, J(Θ) is the meta-objective function (the overall objective of optimization); Let π be the expectation along the policy trajectory. Θ The current strategy is implicitly determined by the meta-parameters; Accumulated rewards for discounts; η正则 To predict the coefficient of the consistency regularization term; Step 4) The global pheromone update formula is: P 全局 ←(1-ρ)P 全局 +∑ 精英蚂蚁 ΔP 沉积 In the formula P 全局 ρ is the global pheromone concentration vector (actions with higher pheromone concentrations have a higher probability of being selected); ρ is the pheromone evaporation coefficient (used to control the decay rate of old pheromones after each update); ∑ 精英蚂蚁 To sum over the set of elite ants; ΔP 沉积 The pheromone increment vector deposited by elite ants; where ρ = σ(w p ·δ 预测 +b p ), where w p A learnable weighted scalar (controlling the sensitivity of the volatility exponent to the volatility coefficient); δ 预测 To predict the quality fluctuation index (obtained by the aforementioned chaotic weighted LSTM); w p and b p It can learn to make pheromone evaporate faster when there are large quality fluctuations, accelerate the forgetting of old strategies, and finally output the current optimized new threshold as the sorting threshold.

[0054] In specific implementation, the nozzle mapping and data aggregation processing includes: The real-time sorting threshold adjustment is converted into material execution instructions for the sorting bins of the air nozzle array, completing the four-level pneumatic sorting and warehousing of seaweed raw materials. Compile statistical analysis of execution results, generate an acceptance report, and upload it.

[0055] The real-time sorting threshold adjustment is converted into the sorting bin material execution instruction rule of the air nozzle array as follows: the received sorting threshold is written into the register. When the single piece Score (i.e., the comprehensive score of seaweed quality grade) is greater than or equal to the sorting threshold 1 and g = special grade, it is judged as special grade; when the sorting threshold 2 ≤ Score < sorting threshold 1 or meets the first-level condition, it is judged as first-level; when the Score < sorting threshold 2 and there is no fatal defect, it is judged as second-level; when there are foreign objects / serious defects, it is directly judged as unqualified.

[0056] It generates and uploads an acceptance report, automatically generating a batch acceptance report according to a preset template, including: basic batch information, a pie chart of grade distribution (see attached document). Figure 2 ), Defect Pareto chart (see appendix) Figure 3 The following data are uploaded to the factory's MES system via the OPC-UA protocol: a statistical summary of quality factors, representative hyperspectral pseudo-color images and defect sample image evidence, and a declaration of conformity between the threshold dynamic adjustment process and the standard.

[0057] Those skilled in the art will recognize that the units and algorithm steps of the various examples described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, computer software, or a combination of electronic hardware and software. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement each specific application, but such implementation should not be considered beyond the scope of this application.

Claims

1. A method for intelligent acceptance and sorting of seaweed raw materials, characterized in that, The method includes: Collect multimodal data, hyperspectral texture maps, and 3D point cloud data of the raw materials; Construct a two-layer pulse-coupled neural field, a differentiable reaction diffusion equation, and a dynamic graph convolutional network; Pulse-coupled neural fields are used to process multimodal data and extract pulse temporal features. Differentiable reaction diffusion equations are used to process hyperspectral texture maps and generate steady-state pattern features. Dynamic graph convolutional networks are used to process 3D point cloud data and extract geometric features. The three features are fused through a parameterized delay-coupled oscillator to generate a spatiotemporal fusion feature vector. Using a spin glass free energy minimization method based on learnable duplicate breaking and quantum perturbation attention, factor spin mapping and non-equilibrium free energy minimization are performed on the spatiotemporal fusion feature vector to obtain a comprehensive score for seaweed quality grade. Dynamic exploration and robust optimization of the comprehensive score of seaweed quality grade were carried out to obtain the real-time sorting threshold adjustment amount; The real-time sorting threshold adjustment is processed by air nozzle mapping and data aggregation to obtain and execute the material execution instructions for the sorting bin, and finally, an acceptance report is generated and uploaded.

2. The intelligent acceptance and sorting method for seaweed raw materials according to claim 1, characterized in that, The multimodal data includes a hyperspectral data cube, an RGB image, and batch attribute information; the hyperspectral texture map is derived from the hyperspectral data cube; and the three-dimensional point cloud data includes geometric point clouds.

3. The intelligent acceptance and sorting method for seaweed raw materials according to claim 1, characterized in that, The process of using pulse-coupled neural fields to process multimodal data includes: By differentiable spectral unmixing, the hyperspectral cube is transformed into an abundance map and stitched with the RGB image to obtain the coding layer pulse sequence, which is used as the input of the double-layer pulse-coupled neural field coding layer. By adjusting the synaptic weights of converging layer neurons on the pulse sequence of the coding layer, pulse firing rate, firing delay, and burst interval are extracted to form pulse timing features.

4. The intelligent acceptance and sorting method for seaweed raw materials according to claim 1, characterized in that, The process of processing hyperspectral texture maps using differentiable reaction diffusion equations includes: Using multi-scale Gabor filtering and convolution dimensionality reduction methods, the local texture of the hyperspectral cube is calculated to generate a single-channel hyperspectral texture map, and the state field of the reaction-diffusion system is initialized. The coefficients of the equations for the hyperspectral texture map and the thickness gradient field are dynamically generated and spatiotemporally evolved to extract steady-state pattern features.

5. The intelligent acceptance and sorting method for seaweed raw materials according to claim 1, characterized in that, The process of using dynamic graph convolutional networks to process 3D point cloud data includes: A geodesic dynamic graph convolutional network with learnable kernel width is used to perform shape topological encoding on the contour map generated from 3D point cloud data, and extract 3D morphological feature vectors.

6. The intelligent acceptance and sorting method for seaweed raw materials according to claim 1, characterized in that, The factor spin mapping and the solution for minimizing nonequilibrium free energy include: Using a parallel lightweight decoding head and a GBDT regressor improved by feature reproduction adaptive regularization, the spatiotemporal fusion feature vector is decoded to predict five continuous quality factors, which are then converted into spin vectors by parameterized tanh mapping. Quantum noise-enhanced inter-node attention calculation and symmetry are performed on the spin vector to obtain the factor-symmetric interaction matrix and external field vector; By substituting the symmetric interaction matrix and external field vector into the spin glass Hamiltonian, and using a meta-network to adaptively predict the temperature and Parisi fracture parameters based on factor skewness and kurtosis, a free energy variational functional is constructed.

7. The intelligent acceptance and sorting method for seaweed raw materials according to claim 6, characterized in that, The factor spin mapping and the solution for minimizing nonequilibrium free energy also include: Using a free energy variational functional, gradient descent minimization is performed on magnetization and overlap to obtain equilibrium order parameters. The optimal spin configuration and corresponding order reference configuration of the sample are then reconstructed based on the equilibrium order parameters. The optimal spin configuration of the sample and the corresponding grade reference configuration are used to calculate the similarity and adaptive penalty correction to obtain the comprehensive score of seaweed quality grade.

8. The intelligent acceptance and sorting method for seaweed raw materials according to claim 1, characterized in that, The dynamic exploration and adversarial robustness optimization includes: The comprehensive score of each piece of seaweed quality grade is concatenated with the five-dimensional quality factors to form a six-dimensional state vector, and a fixed-length sliding window is constructed along the processing sequence to form a quality state window matrix; The LSTM time-series evolution of the quality state window matrix with chaotic gated bias modulation outputs the mean, variance, and volatility index of the future quality score. A heuristic information matrix is ​​constructed based on the mean and variance of future quality scores. The probability of action selection is calculated by combining the pheromone concentration, and the sorting threshold adjustment action is output.

9. The intelligent acceptance and sorting method for seaweed raw materials according to claim 8, characterized in that, The dynamic exploration and adversarial robustness optimization also includes: Construct a Critic network to evaluate the sorting threshold adjustment action by combining asymmetric penalties for false rejections and missed detections, grade ratio deviation and prediction consistency rewards, and calculate the robust TD error by minimizing the Q value against perturbations, and output the action value. The loss function for the Critic network is defined as follows: In the formula L Q Let be the loss function of the Critic network; s, a, and r are the reinforcement learning state vector, sorting threshold adjustment action, and immediate reward at the current time step, respectively; s 1 The state to which the action is performed is the next state; γ is the discount factor. To calculate the expectation of the mini-batch transfer tuples sampled from the experience replay pool; Q(s,a;ψ) is the state-action value estimate of the current Critic network output; ψ - The parameters of the target Critic network; Q(s) 1 ,a 1 ;ψ - ) is to counteract the robustness term; λ 正则 For learnable robust regularization coefficients; η is the infinite norm of the Jacobian matrix of Q-value with respect to state input s; η is the adversarial perturbation vector; min is the minimum function; agrmax is the function of the maximum value of the independent variable. This is the optimal action for the next state as perceived by the current network. The minimum value of Q after applying the worst-case perturbation to the next state; The cumulative reward maximization optimization of the action value evaluated by Critic is performed under the constraint of predictive consistency. The ant colony chaos parameters are dynamically updated and the real-time sorting threshold adjustment is output.

10. The intelligent acceptance and sorting method for seaweed raw materials according to claim 1, characterized in that, The nozzle mapping and data aggregation processing includes: The real-time sorting threshold adjustment is converted into material execution instructions for the sorting bins of the air nozzle array, completing the four-stage pneumatic sorting and warehousing of seaweed raw materials. Compile statistical analysis of execution results, generate an acceptance report, and upload it.

Images