Tailings ash content value prediction system and method based on bionic neural network

Abstract

The application discloses the technical field of coal mine ash value prediction, and discloses a tailing ash value prediction system and method based on a bionic neural network, which comprises the following steps: acquiring a tailing coal slurry water collection image; performing multi-dimensional feature extraction on the tailing coal slurry water collection image; performing dynamic weighting on the extracted multi-dimensional features through a multi-layer relationship attention mechanism; inputting the dynamically weighted features into a pre-trained neural network to predict the ash value; and outputting an ash value prediction result. The application uses the meta-learning theory to construct the multi-layer relationship attention mechanism in view of the shortcomings of the traditional training method, breaks through the traditional single and fixed neural network training mode, and realizes accurate prediction of the ash value of the tailing coal slurry water.

Description

Technical Field

[0001] This invention relates to a tailings ash content prediction system and method based on biomimetic neural networks, belonging to the field of coal mine ash content prediction technology. Background Technology

[0002] After the coal slurry entering the flotation machine is cleaned and recovered, the remaining portion is tailings. Ash is a residue composed of non-combustible products of coal slurry, gangue (useless minerals in the ore), mineral fragments, etc. It usually contains a small amount of valuable components that have not been fully recovered, such as fine-grained coal and coal-containing components mixed with gangue. These valuable components can be separated and recovered using flotation reagents. If the raw materials for flotation reagents can be precisely proportioned with the ash content known in advance, the recovery rate will be higher. Therefore, it is particularly necessary to know the ash content of the tailings in advance.

[0003] In recent years, with in-depth research, methods such as camera photography have been used to obtain features of coal slurry water regarding light reflection and sampled images, such as energy values ​​and grayscale values. By analyzing the relationship between these features, conditions, and ash content, and based on a large amount of data, machine learning linear regression methods are used, with features as conditions and the actually measured ash content as the target value, to determine the target ash content value through feature conditions. However, machine learning linear regression methods can only capture the linear relationship between conditions and the target, and the neural network training method is singular and fixed, making it unable to accurately predict the ash content value of tailings coal slurry water. Summary of the Invention

[0004] The purpose of this invention is to overcome the shortcomings of the prior art and provide a tailings ash content prediction system and method based on biomimetic neural networks. In view of the shortcomings of traditional training methods, this invention uses meta-learning theory to construct a multi-layer relational attention mechanism, which breaks through the traditional single and fixed neural network training method and realizes accurate prediction of tailings coal slurry water ash content.

[0005] To achieve the above objectives, the present invention is implemented using the following technical solution:

[0006] In a first aspect, the present invention provides a method for predicting the ash content of tailings based on a biomimetic neural network, comprising:

[0007] Acquire images of tailings slurry water;

[0008] Multidimensional feature extraction was performed on images of tailings coal slurry water.

[0009] Dynamic weighting is performed based on the extracted multidimensional features through a multi-layer relational attention mechanism;

[0010] The dynamically weighted features are input into a pre-trained neural network to predict gray values.

[0011] Output the gray value prediction results.

[0012] Furthermore, the multidimensional features include grayscale features, color features, texture features, and edge features, wherein:

[0013] The grayscale feature includes the grayscale mean;

[0014] The color features include the mean of the R / G / B channels, the standard deviation of the R / G / B channels, and the variance of the R / G / B channels;

[0015] The edge features are Sobel edge features, including edge mean, variance, and energy;

[0016] The texture features are GLCM texture features, including contrast, homogeneity, correlation, energy, and entropy.

[0017] Furthermore, the multi-layered relational attention mechanism includes: a feature-target relational attention module, a feature-feature relational attention module, a feature-feature relational attention module, and a cross-modal multi-interaction relational attention module, wherein:

[0018] The feature-target relationship attention module is:

[0019] In the formula: For feature-target attention weights, For the Sigmoid function, For the first linear layer, For the second linear layer, For activation function, This is the input feature vector.

[0020] Furthermore, the feature-feature relationship attention module is as follows:

[0021] In the formula: For relation weights, For the Sigmoid function, For the input feature vector, Let be the projection matrix. It is a relation matrix. The number of heads receiving multi-head attention.

[0022] Furthermore, the mathematical model for the feature-feature relation attention module is as follows:

[0023] In the formula: This is the new relationship matrix after the interaction between the heads. For normalized exponential functions, This is a multi-head relationship weight matrix. For element-wise multiplication, It is a learnable bilinear matrix. For the Sigmoid function, For gating weights, The gated weight matrix is ​​used to... Mapping to gate weights Dimensions It is a function of average value. This is the final result of the feature-feature relationship attention module.

[0024] Furthermore, the mathematical model of the cross-modal multi-interaction relationship attention module is as follows:

[0025] In the formula: The projection vector of the feature-target relationship. The mapping function that projects feature-target attention weights onto the shared interaction space. For feature-target attention weights, Let be the matrix used for the linear transformation of the feature-target attention weights. This is the projection vector of the relation-relation weights. This is a multi-head relationship weight matrix. The mapping function that projects relation-relation weights onto the shared interaction space. This is the matrix used to perform a linear transformation on the multi-head relationship weight matrix. The adjustment factor is obtained through cross-modal interactive calculation. The final output is the weight of the cross-modal interaction relationship. It is a linear rectified function. It is an interactive core.

[0026] Furthermore, the neural network includes a neural network model, a similarity calculation model, a weight adjustment model, and a learning rate adjustment model, wherein:

[0027] The neural network model is as follows:

[0028] In the formula: This is the predicted output value of the neural network for the target gray value. Activation function For the input feature vector, For the final layer;

[0029] The similarity calculation model is as follows:

[0030] In the formula: For the current sample The overall similarity score with historical samples, For the features of the current sample vector, For the vector features of historical samples, This is the feature-target relation weight vector for the current sample. Historical Sample Feature-target relation weight vector, This is the feature-feature relationship weight matrix for the current sample. Historical Sample Feature-feature relation weight matrix The cosine similarity function is used. For historical memory The weighting coefficients, For current samples and historical memory The overall similarity For historical memory The prediction confidence level; Let m be the total number of samples stored in the historical memory bank, and m be the summation index of historical samples. This represents the combined similarity between the current sample and historical sample m. The prediction confidence level for historical sample m;

[0031] The weights of the sampled image features are adjusted by judging the difference between the predicted value and the verification target value. The weight adjustment model is as follows: In the formula: The set of all learnable parameters of a neural network. To represent parameters To optimize the objective by minimizing the value of subsequent expressions, For loss function, For function mapping, For the first The feature vectors of each training sample The prediction function of a neural network model. For the first The true gray value of each training sample The regularization coefficient is used. For the computation function of the attention mechanism, Here is the weight matrix for the attention mechanism. For query vector, For key vectors, For value vectors, For normalized exponential functions, This is the transpose of the key vector matrix. Let be the dimension of the key vector. For weighted elements;

[0032] The learning rate adjustment model is as follows:

[0033] In the formula: The step size is The learning rate The initial learning rate, The target correlation influence coefficient, The average attention weights are the feature importance. The coefficient representing the influence of the relationship between conditions. Let V be the variance of the characteristic relationship.

[0034] Secondly, the present invention provides a tailings ash content prediction system based on a biomimetic neural network, comprising:

[0035] Image acquisition module: used to acquire images of tailings slurry water;

[0036] Feature extraction module: used to extract multidimensional features from images of tailings coal slurry water.

[0037] Dynamic weighting module: used to dynamically weight based on extracted multidimensional features through a multi-layer relational attention mechanism;

[0038] Gray value prediction module: used to input dynamically weighted features into a pre-trained neural network to predict gray values;

[0039] Ash value output module: Used to output ash value prediction results.

[0040] Thirdly, the present invention provides a tailings ash content prediction device based on a biomimetic neural network, including a processor and a storage medium.

[0041] The storage medium is used to store instructions;

[0042] The processor is configured to operate according to the instructions to perform the steps of the method according to any of the foregoing.

[0043] Fourthly, the present invention provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps of any of the methods described above.

[0044] Compared with the prior art, the beneficial effects achieved by the present invention are as follows:

[0045] I. This invention addresses the shortcomings of traditional training methods by constructing a multi-layer relational attention mechanism using meta-learning theory. First, multi-dimensional features are extracted from the collected images of tailings coal slurry water. Then, the extracted multi-dimensional features are dynamically weighted through the multi-layer relational attention mechanism. Finally, the dynamically weighted features are input into a pre-trained neural network for gray value prediction. This breaks through the traditional single and fixed neural network training method and achieves accurate prediction of the gray value of tailings coal slurry water.

[0046] Second, this invention constructs a multi-connected attention mechanism based on the basic neuronal connectivity. It utilizes error feedback between neurons and synaptic plasticity to adjust the strength of connections through electrical signal feedback, and establishes a backpropagation mechanism and a dynamic adjustment mechanism for neuronal weights based on the neural network loss function, utilizing the biological principles of connection methods. Furthermore, it establishes a feature-target and feature-feature multi-layered attention relationship mechanism by strengthening information transmission paths and speed through multi-layered connections between neurons, groups, and domains, guided by attention. Finally, it establishes a dynamic learning rate regulation mechanism guided by the brain's adjustment of neurotransmitter release and neuronal switching based on task responsibility, thereby improving the learning efficiency of the neural network. Attached Figure Description

[0047] The accompanying drawings, which form part of this specification, are used to provide a further understanding of the invention. The illustrative embodiments of the invention and their descriptions are used to explain the invention and do not constitute an undue limitation of the invention. In the drawings:

[0048] Figure 1 This is a flowchart illustrating the tailings ash content prediction method based on a biomimetic neural network provided in Embodiment 1 of the present invention. Detailed Implementation

[0049] The present invention will now be described in detail with reference to the accompanying drawings and embodiments. It should be noted that, unless otherwise specified, the embodiments and features described herein can be combined with each other.

[0050] The following detailed description is exemplary and intended to provide further detailed explanation of the invention. Unless otherwise specified, all technical terms used in this invention have the same meaning as commonly understood by one of ordinary skill in the art. The terminology used in this invention is for describing particular embodiments only and is not intended to limit the scope of exemplary embodiments according to the invention.

[0051] Example 1:

[0052] This embodiment proposes a tailings ash content prediction method based on biomimetic neural networks. Addressing the shortcomings of traditional training methods, it utilizes meta-learning theory, mimicking the learning process of primate brains in learning instructions and observing things. Starting from the "minimal consciousness system," it simulates the hierarchical perception mechanism of the visual system and dynamically adjusts the connection strength, mode, and relationship of brain neurons. That is, no pre-programming is required during the training process; goal-oriented behavior can be achieved simply by adjusting the connection strength, connection mode, and selection of neurons, i.e., neural network training nodes, through feedback. This breaks through the traditional single and fixed training method. This scheme establishes predicted ash content based on MAE difference feedback, multi-layer relationships and attention mechanisms, and heuristic reasoning. It adds intra-feature-feature relationships, relationship-relationships (i.e., relationships between features), and cross-modal multi-interaction relationships between relationship-relationships and feature-target relationships to the feature-target relationship, and incorporates an attention mechanism into these relationships. Utilizing the difference between the predicted result and the actual ash content, feedback is fed back to adjust the neural network and the parameters of the aforementioned relationships and the neural network itself. Multiple iterations are performed, and the training results and parameters of each iteration are stored in the neural network. In the next prediction, the corresponding empirical relationships are retrieved based on the characteristics of the prediction body to participate in the prediction, thereby increasing prediction accuracy. This invention, based on establishing multiple neural network parameter relationships and combining a multi-layer relationship attention mechanism, adjusts various parameters of the neural network and can remember each training result and its characteristics, achieving accurate prediction of the ash content of tailings coal slurry. Specifically, it includes the following:

[0053] I. A Bionic Neural Network-Based Gray Value Prediction Method Based on MAE Feedback, Multi-Layer Relational Attention Mechanism, and Heuristic Reasoning

[0054] The algorithm establishes relationships between multiple parameters of a neural network and utilizes a multi-head attention mechanism. After each training iteration, the prediction result is evaluated using the MAE prediction method, and the evaluation is fed back to the neural network to adjust the relationships and related parameters. The training results are stored for quick retrieval in the next prediction. After multiple iterations, the optimal result is saved. The specific computational scheme includes the following:

[0055] 1. Preprocessing and parameter extraction of images sampled from the camera: Use OpenCV to read the images and convert them into a processable format. The texture feature (GLCM) and edge feature (Sobel) operators are both image features. Tensors are used to convert them into tensors usable by the PyTorch architecture, so that the GPU can participate in the calculation and reduce CPU overhead.

[0056] 2. The model is used to predict the gray value of the sampled image and the extracted parameters. Then, the neural network is used for training. After obtaining the training feature results, they are loaded into the prediction terminal to predict the gray value.

[0057] II. Implementation of the Connection between Bionic Neural Network Based on 18-Dimensional Features and Minimal Consciousness System

[0058] 1. Biosensory mapping of 18-dimensional feature processing

[0059] Based on the relationship between gray value, gray mean, and other lighting characteristics, 18-dimensional features were extracted. This physically covers all visual dimensions related to gray value prediction.

[0060] 1) Basic Visual Layer (Grayscale / Color Features) – 10 dimensions in total: grayscale mean, R, G, and B channel mean, R, G, and B channel standard deviation, and R, G, and B channel variance. Grayscale mean: Directly reflects the overall grayscale level; grayscale value is directly proportional to brightness. RGB channel statistics: Identifies mineral type tendencies and reflective interference.

[0061] 2) Intermediate feature layer (texture / edge features):

[0062] Sobel edge features (3D, edge mean, variance, energy): captures edge information by calculating image gradients. Its statistical features directly reflect the physical properties of particles in coal slurry water, quantify particle sharpness and contour clarity. The ash content is proportional to the edge strength and variance, and the extreme ash content is positively correlated with the edge energy.

[0063] GLCM (Gray Co-occurrence Matrix) texture features (5-dimensional, contrast, homogeneity, correlation, energy, entropy): Quantizes texture by statistically analyzing the spatial relationship of pixel pairs. Its core parameters directly reflect the physical state of coal slurry and analyze the particle distribution. Contrast and correlation are directly proportional to gray content, homogeneity is inversely proportional to gray content, and extreme gray content values ​​are directly proportional to energy.

[0064] III. Four types of enhanced connection computation modules for neural nodes

[0065] 1. Feature-Target Relationship Attention Module k-target (k-target refers to the correlation between the feature (key) and the target (target). "k" is an abbreviation for "key". It is directly derived from the core Query-Key-Value (QKV) model of the attention mechanism. k refers to the key (feature) and target refers to the target) neurons selectively focus on key stimuli.

[0066] Mathematical model: (Synaptic weight generation)

[0067] In the formula, the input feature vector First, through the first linear layer Compressing dimensions, and then... The activation function introduces non-linearity, which is then passed through a second linear layer. Restore the original dimensions, then use the Sigmoid function. Generate attention weights And apply the attention features to the original features, and use the original features With attention weight Successive multiplication Finally, attention weights are applied to the original features.

[0068] The FeatureAttention module uses linear transformations and activation functions to generate attention weights along the feature dimension through two layers of linear transformations. This calculates the relationship between 18 features (such as the mean and standard deviation of RGB channels) and the target gray value, dynamically adjusting the weight of each feature. For example:

[0069] 1) High mean in the red channel: This may correspond to a specific region in the image (such as impurities in coal). The model will enhance the influence of this feature on gray value prediction through attention weights.

[0070] 2) Large color variance: This may indicate complex image texture. The model will adjust the importance of this feature based on historical feedback.

[0071] 2. Feature-feature relationship attention (kk, the internal correlation between a feature key and other feature keys)

[0072] Neuronal cluster interactions simulate the collaborative processing of neuronal clusters in the cerebral cortex, such as the association weights between visual processing neurons and motor neurons.

[0073] Mathematical model: (Multi-cluster parallel processing)

[0074] In the formula, the input By different projection matrices Mapping to multiple subspaces, learning different relation representations, each projected feature and learnable relation matrix Multiply, then pass through the Sigmoid function That is, generating relation weights The final weighted features of each head By averaging and merging, the model can capture feature relationships from different subspaces. This indicates the number of heads in the multi-head attention mechanism. Multi-head attention is the core mechanism of Transformer. This attention calculation involves feeding the features of the input heads (i.e., the features of the prediction volume) into multiple subspaces. Each input feature, such as RGB mean, standard deviation, and contrast, learns different attention patterns in the mapping network and calculates attention in its own independent subspace. Finally, the multiple results are concatenated, achieving the effect of capturing complex feature combinations. The ConditionalRelationAttention module models complex relationships between features using a multi-head mechanism, for example:

[0075] Dependencies between texture features: Regions with high contrast may also have low homogeneity; these two features together reflect the particle distribution characteristics of coal. Multi-head attention mechanisms can learn this multi-dimensional dependency in parallel.

[0076] 3. Relationship-relationship (i.e., feature-feature relationship): Attention kk—kk, where kk represents the internal correlation between a feature (key) and other features (keys), and kk—kk represents the relationship between these correlations. This involves the interactive coordination of high-level neural circuits.

[0077] Mathematical model:

[0078] In the formula: This is the new relationship matrix (intermediate variable) after the interaction between the heads. For normalized exponential functions, This is a multi-head relationship weight matrix. This is used for element-wise multiplication, where weights are multiplied by the matrix positionally. It is a learnable bilinear matrix (used to model the interactions between different attention heads). For the Sigmoid function, For gating weights, The gated weight matrix is ​​used to... Mapping to gate weights Dimensions It is a function of average value. It is an enhanced relation matrix that integrates multi-level feature relation information, and is the final result of the feature-feature relation attention module.

[0079] Input multi-head relation weight matrix and bilinear matrix Multiply by the attention heads to calculate the interactions between them, resulting in a new relation matrix. and through Normalization. This is achieved through gating weights. Weighted fusion of multi-head relationship weight matrix and inter-head interaction relationship This enables dynamic adjustments. The RelationAttention module handles the relationships between different attention heads (representing different feature clusters). For example:

[0080] The interaction between the color feature cluster and the texture feature cluster: the first attention head may focus on color features (RGB mean, variance), and the second head may focus on texture features (contrast, correlation). This module learns the interaction weights between these two feature clusters through a bilinear matrix.

[0081] 4. Cross-modal interaction (k-target—kk—kk):

[0082] (k-target—kk—kk, representing the relationship between feature-target relationship attention k-target and relationship-relationship kk—kk) Integration and decision-making of multi-source information, simulating the integration of visual, tactile and other multimodal information by the prefrontal cortex.

[0083] Mathematical model:

[0084] In the formula: The projection vector of the feature-target relationship. The mapping function that projects feature-target attention weights onto the shared interaction space. For feature-target attention weights, Let be the matrix used for the linear transformation of the feature-target attention weights. This is the projection vector of the relation-relation weights. This is a multi-head relationship weight matrix. The mapping function that projects relation-relation weights onto the shared interaction space. This is the matrix used to perform a linear transformation on the multi-head relationship weight matrix. The adjustment factor is obtained through cross-modal interactive calculation. The final output cross-modal interaction relationship weights are composed of a multi-head relationship weight matrix and an adjustment factor. Element-by-element multiplication ( )get, It is a linear rectified function. It is an interactive core.

[0085] First, the feature-target attention weights are... The relation-relation weights are projected onto the shared interaction space and then processed through tensor product. That is, calculating the interaction between two types of attention—feature-target relationship and relationship-relationship—then adjusting it using an interaction kernel, and finally... Activate the generation of adjustment factors And multiply them step by step using the multi-head relationship weight matrix. Module modeling features - target weights and relationships - relationship interactions.

[0086] The synergistic effect of color and texture features on grayscale values: When the red channel mean is high, the model may enhance the influence of texture features (such as contrast) on grayscale values. This interaction is calculated through tensor product and interaction kernel matrix.

[0087] IV. Learning Behavior and Neural Networks

[0088] The above conditional features, enhanced by the attention mechanism, are then input into the DynamicWeightNet neural network and mapped to multiple nodes and hidden layers. Neural network model: In the formula: This is the predicted output value of the neural network for the target gray value. For activation function, For the input feature vector, This is the final layer. Neural networks can learn complex feature relationships, given multiple input features. conduct The weight parameters of the final layer are summed using the activation function. The nonlinear characteristics Matrix multiplication or vector dot product operation.

[0089] In addition to mapping the four types of enhanced connections mentioned above, and taking into account the characteristics of neural networks, the following three core algorithmic mechanisms are incorporated:

[0090] 1. Experiential memory and heuristic reasoning

[0091] By calculating the similarity between the 18-dimensional feature vectors (basic statistics, color, texture, and edge) of sampled image features and historical memory, a neural network supplemented by an experience base function is used for prediction and reasoning.

[0092] Mathematical model: In the formula, For the current sample The overall similarity score with historical samples, For the features of the current sample vector, For the vector features of historical samples, This is the feature-target relation weight vector for the current sample. Historical Sample Feature-target relation weight vector, This is the feature-feature relationship weight matrix for the current sample. Historical Sample Feature-feature relation weight matrix This is the cosine similarity function.

[0093] During the prediction phase, the weighted average fusion formula for the memory bank is:

[0094] In the formula, For historical memory The weighting coefficients, For current samples and historical memory The overall similarity For historical memory The prediction confidence level; Let m be the total number of samples stored in the historical memory bank, and m be the summation index of historical samples. This represents the combined similarity between the current sample and historical sample m. is the prediction confidence level for historical sample m.

[0095] The 18-dimensional features of historical samples, target-feature attention weights (k_target_weights), feature-feature relation weights (k_k_weights), and their corresponding gray values ​​are stored in a MemoryBank class.

[0096] When making the next prediction, the statistical features (mean, variance) correspond to the initial encoding of light intensity by the bionic retina, and the GLCM texture features correspond to the selectivity of the direction by the bionic V1 area neurons. The two attention weights are used as features to measure the similarity between the prediction target and the two feature values. Together, they constitute multi-dimensional cues for memory retrieval to retrieve experience features. The most similar historical training data is added to the prediction process for prediction.

[0097] 2. Error Feedback and Synaptic Plasticity

[0098] By judging the difference between the predicted value and the target value, the weights of the sampled image features are adjusted, mimicking the principle of adjusting synaptic strength through electrical signal feedback: after... Regularization constrains the attention weights of 18-dimensional features, simulating the "pruning" process of neuronal synapses to preserve key features (such as edges and textures). The model uses attention-based regularization. The loss function is adjusted using feedback.

[0099] Mathematical model: ;

[0100] The core of the attention mechanism is to calculate the importance weights of the input features:

[0101] in, The set of all learnable parameters of a neural network. To represent parameters To optimize the objective by minimizing the value of subsequent expressions, For loss function, For function mapping, For the first The feature vectors of each training sample The prediction function of a neural network model. For the first The true gray value of each training sample The regularization coefficient controls the strength of the regularization term. The larger the value, the stronger the constraint on the weight. This is the weight matrix for the attention mechanism, used to measure the importance of input features (such as the attention score in Transformer, the attention mask in CNN, etc.). For the computation function of the attention mechanism, For query vector, For key vectors, For value vectors, For normalized exponential functions, This is the transpose of the key vector matrix. is the dimension of the key vector.

[0102] norm ( Regularization): That is, the sum of the absolute values ​​of the weighted elements. This is the weight element.

[0103] Regarding the quality of the adjusted predicted values, the reward and punishment measures in biomimetic training utilize the MAE method for judgment. MAE, or Mean Absolute Error, is a regression task evaluation metric that calculates the average of the absolute deviations between the predicted and true values. It is a basic robust loss function primarily used to intuitively measure the performance of the predicted values. The optimization results show that MAE does not directly participate in training or optimizing the model, but The loss mechanism is consistent with the core logic of MEA (both are based on absolute error). When the model minimizes through backpropagation... When there is a loss, the neural network model is made to approach the model range evaluated by MAE, and then the various parameters of the neural network are adjusted so that the predicted value of the trained model gradually approaches the actual value.

[0104] MAE is the average difference between the N predicted values ​​and the N actual target gray values. If the deviation is too large, it needs to be adjusted using methods such as RMSELoss; if the deviation is small, it is adjusted using RegularizedSmoothL1Loss (SmoothL1 loss function with attention regularization). Finally, the state with the minimum MAE value is used to save the neural network and its feature values.

[0105] In the DynamicWeightNet network, Dropout is added to the input layer to randomly drop some neurons, improving generalization ability. BatchNorm is used to normalize the input of each layer, and ReLU is used to regularize the input of each layer, which helps to alleviate the vanishing / exploding gradient problem.

[0106] For example, the model weakens the connections of redundant features (such as background color) and strengthens the weights of key features (such as object edges) through error feedback. Or it weakens reflective noise features (such as RGB fluctuations in water surface highlights) and strengthens the texture features of sediment particles (such as GLCM contrast) through error feedback.

[0107] 3. Dynamic learning rate and meta-learning

[0108] By combining feature importance and feature stability to dynamically adjust the learning rate, it simulates the mechanism by which the biological brain adjusts neural plasticity according to task difficulty. The learning rate is dynamically adjusted according to the stability of sampled image features (such as texture variance) to adapt to samples of different concentrations.

[0109] Mathematical model: ;

[0110] Bionic brain learns logic, The step size is learning rate, The initial learning rate, The target correlation influence coefficient, The average attention weights are the feature importance. The coefficient representing the influence of the relationship between conditions. Let V be the variance of the characteristic relationship.

[0111] 1) Similar to the brain's reinforcement of key neuronal connections, for example, in coal slurry water prediction, the "GLCM contrast" feature has a high attentional weight. (Increase the learning rate to accelerate the learning of the feature).

[0112] 2) When the brain recognizes familiar patterns (such as stable coal slurry texture features), the release of neurotransmitters (such as dopamine) increases, accelerating synaptic plasticity, corresponding to the formula in... Increased activity leads to a higher learning rate; however, when faced with complex tasks (such as the fluctuating characteristics of turbid coal slurry water), the release of neuromodulators decreases, resulting in a lower learning rate to avoid overfitting. Increasing the learning rate leads to a decrease in the learning rate.

[0113] V. Class Consciousness Calculation and Sample Prediction

[0114] 1. Feature weighting mechanism

[0115] The biological intelligent mapping and consciousness-focusing mechanism in gray-level prediction, the selective enhancement of 18-dimensional features (such as gray-level mean, GLCM contrast, edge energy, etc.) through fused_weights, and the biological method of biomimetic prefrontal cortex attention regulation of visual information. Here, fused_weights are equivalent to the plasticity adjustment of neural synapses, enabling the model to automatically focus on key features.

[0116] For example, in turbid coal slurry water, the weights of "texture contrast" and "edge sharpness" are significantly increased, while the weights of interfering features such as background reflection are suppressed.

[0117] 2. Realization of the memory-decision closed loop

[0118] The storage and retrieval of sample experiences. The memory bank mechanism forms a complete cycle of "perception-attention-memory-decision-correction" in gray prediction, realizing class consciousness.

[0119] For example, when encountering a new sample, the system will retrieve the historical sample with the most similar features in the memory (such as texture feature similarity > 85%), and use its gray value as prior knowledge to fuse with the current model prediction.

[0120] The calculation steps of this scheme are as follows: Input data → Feature extraction layer (convolution / recurrent, etc.) → Multi-layer relational attention weighting → Fully connected layer → Memory fusion (prediction stage) → Output predicted value → Loss function error calculation → Backpropagation → Network and parameter adjustment → Output features → Load features + sample to be predicted → Predict gray value. The MAE error feedback process is as follows: Prediction error (MAE) → Regression loss (SmoothL1) → Neural network fc6 → fc5……fc1 → Fusion features; then the fused features split into two branches, one is the backpropagation calculation of fused_weights; the other is the backpropagation calculation of kk-final_weights.

[0121] The generation of the original sampled grayscale image is a physical conversion process from light signal to electrical signal to digital signal. It is the direct carrier of underlying visual information and provides the original data foundation for subsequent feature extraction and model analysis. Color channel comparison quantifies the "information richness" of channels from a statistical perspective. In scenarios such as image optimization, quality assessment, and feature extraction, it provides a quantitative basis for data-driven decisions based on sampled image data. Color channel mean comparison transforms the complex visual information of the sampled image into comparable values, thereby enabling automated feature extraction and decision-making. Color channel distribution is the "statistical fingerprint" of pixel values ​​in the sampled image. The shape of the histogram can intuitively determine the brightness characteristics, detail richness, and potential problems of the channels. Edge image simplifies the visual information of the sampled image into structured contour features, providing support for high-level target recognition and scene understanding. Texture feature comparison is used to quantify the "regularity of surface patterns" in the sampled image to determine the material, state, and category of objects. When color and edge information are insufficient to distinguish targets, texture features often provide more detailed differentiation criteria. Edge intensity distribution represents the distribution of the intensity changes of pixels in the sampled image, used to describe the "strength" of edges at different locations in the image.

[0122] The training feature weight distribution is the result of the feature weighting mechanism, used by the neural network to dynamically weight features. Driven by an attention mechanism, it strengthens the weights of features relevant to the training objective and reduces the weights of irrelevant or weakly related features. This dynamic strengthening mechanism makes the neural network more intelligent, mimicking the thinking process of a biological brain. This section describes the changes in the weights of each feature during each echo, and the MAE (Maximum Averaging Value) used to determine the estimated value. A large MAE indicates a significant difference between the predicted and target values, leading to a decrease in confidence or discarding of the data. Conversely, a small or near-zero MAE increases the confidence in the data. The training feature with the smallest MAE is saved after training. The training-side human-computer interaction interface includes a training storage module, a save module, and a camera display module. The prediction-side display interface includes a predicted grayscale curve display window and a camera display module.

[0123] The training process is as follows:

[0124] (1) First, set the shooting time interval of the sampling camera and the storage location of the sampling image, and observe the condition of the sampling tailings coal slurry platform, whether the image is overexposed, and whether there are stains on the platform.

[0125] (2) Then, in the entry, enter the actual ash content value, and create a sample name under this coal preparation plant using the name of the coal preparation plant. Enter the complete entry into the sample name directory.

[0126] (3) Enter the directory where you want to store the training results under “Output Training Directory” and name the training result .pth file.

[0127] (4) Then the neural network training begins, and the program automatically saves the result of the best MAE value.

[0128] The input terminal operation is as follows:

[0129] (1) Open the software and observe the sample captured by the camera.

[0130] (2) Soft trigger the feed valve, discharge valve, and submersible pump switch as needed.

[0131] (3) Set the sampling interval. And load the neural network that was just trained.

[0132] (4) Click Start Prediction. The predicted ash value will be displayed in the upper right corner of the software interface and in the real-time ash chart in the form of a line.

[0133] This proposal suggests a feedback-modulated bionic neural network based on a multi-layered connectivity attention mechanism. It mimics the principles of "meta-learning" and "minimal consciousness system," utilizing a four-layer attention mechanism and feedback modulation method to mimic the way brain neurons collaboratively process information.

[0134] Traditional neural networks are trained on a single relationship, relying solely on a fixed learning rate, resulting in inaccurate training results. This patent, based on a four-layer attention mechanism (feature-target, feature-feature, relationship-relationship, and cross-modal), introduces a dynamic learning rate adjustment mechanism. This simulates the brain's process of adjusting neural plasticity according to task difficulty, automatically adjusting the learning rate based on task load. Combined with Dropout, ReLU, and SmoothL1 loss functions, it effectively prevents "dead" states during training, significantly improves training accuracy, avoids repetitive and ineffective training, and enhances the accuracy and stability of tailings ash value prediction.

[0135] Traditional neural networks use fixed weights and lack a memory mechanism. Training relies solely on a program, with no empirical values ​​to reference each training session, resulting in slow and unsustainable training. This patent addresses this by incorporating a biomimetic brain neuron synapse strength / pruning mechanism. It uses the gradient of the loss function to update weights, driving real-time adjustment and optimization of all weights (including attention weights). It mimics the brain's "perception-memory-action" cycle and its ability to quickly generalize from limited experience, remembering each training experience value to achieve heuristic reasoning, improve model generalization, and allow the model to adapt to training tasks in more scenarios. Furthermore, it introduces a MAE (Model-Based Evaluation) mechanism to save the best training features. This endows the neural network with intelligent functionality, making the training process more flexible.

[0136] Compared to complex and potentially health-hazardous detection methods such as X-ray fluorescence spectroscopy and radioisotope methods, as well as the limitations of traditional image analysis methods, this invention is entirely based on camera images and computational models. It is safe, simple to operate, and easy to deploy on-site in coal preparation plants. Through the combined effect of the aforementioned biomimetic neural network structure and algorithm, it ultimately achieves high-precision, online, and real-time prediction of the ash content of tailings coal slurry. This provides a crucial prerequisite for the precise formulation of flotation reagents, potentially significantly improving clean coal recovery rates, reducing resource waste, and guiding further resource utilization of tailings, thus possessing significant industrial application value and economic benefits.

[0137] Example 2:

[0138] The tailings ash content prediction system based on biomimetic neural networks can implement the tailings ash content prediction method based on biomimetic neural networks described in Example 1, including:

[0139] Image acquisition module: used to acquire images of tailings slurry water;

[0140] Feature extraction module: used to extract multidimensional features from images of tailings coal slurry water.

[0141] Dynamic weighting module: used to dynamically weight based on extracted multidimensional features through a multi-layer relational attention mechanism;

[0142] Gray value prediction module: used to input dynamically weighted features into a pre-trained neural network to predict gray values;

[0143] Ash value output module: Used to output ash value prediction results.

[0144] Example 3:

[0145] This invention also provides a tailings ash content prediction device based on a biomimetic neural network, which can realize the tailings ash content prediction method based on a biomimetic neural network described in Embodiment 1, including a processor and a storage medium.

[0146] The storage medium is used to store instructions;

[0147] The processor is configured to operate according to the instructions to perform the steps of the following method:

[0148] Acquire images of tailings slurry water;

[0149] Multidimensional feature extraction was performed on images of tailings coal slurry water.

[0150] Dynamic weighting is performed based on the extracted multidimensional features through a multi-layer relational attention mechanism;

[0151] The dynamically weighted features are input into a pre-trained neural network to predict gray values.

[0152] Output the gray value prediction results.

[0153] Example 4:

[0154] This invention also provides a computer-readable storage medium that implements the tailings ash content prediction method based on a biomimetic neural network as described in Embodiment 1. The medium stores a computer program that, when executed by a processor, performs the steps of the following method:

[0155] Acquire images of tailings slurry water;

[0156] Multidimensional feature extraction was performed on images of tailings coal slurry water.

[0157] Dynamic weighting is performed based on the extracted multidimensional features through a multi-layer relational attention mechanism;

[0158] The dynamically weighted features are input into a pre-trained neural network to predict gray values.

[0159] Output the gray value prediction results.

[0160] As is known from common technical knowledge, this invention can be implemented through other embodiments that do not depart from its spirit or essential characteristics. Therefore, the disclosed embodiments described above are merely illustrative and not exhaustive. All modifications within the scope of this invention or its equivalents are included in this invention.

[0161] Those skilled in the art will understand that embodiments of the present invention can be provided as methods, systems, or computer program products. Therefore, the present invention can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present invention can take the form of a computer program product embodied on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0162] This invention is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of the invention. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart illustrations and / or block diagrams. Figure 1One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0163] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0164] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0165] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and not to limit it. Although the present invention has been described in detail with reference to the above embodiments, those skilled in the art should understand that modifications or equivalent substitutions can still be made to the specific implementation of the present invention. Any modifications or equivalent substitutions that do not depart from the spirit and scope of the present invention should be covered within the scope of protection of the claims of the present invention.

Claims

1. A tailings ash content prediction method based on biomimetic neural networks, characterized by: include: Acquire images of tailings coal slurry water; Multidimensional feature extraction was performed on images of tailings coal slurry water. Dynamic weighting is performed based on the extracted multidimensional features through a multi-layer relational attention mechanism; The dynamically weighted features are input into a pre-trained neural network to predict gray values. Output the ash content prediction results; The multi-layered relational attention mechanism includes: a feature-target relational attention module, a feature-feature relational attention module, a feature-feature relational attention module, and a cross-modal multi-interaction relational attention module, wherein: The feature-target relationship attention module is: In the formula: For feature-target attention weights, For the Sigmoid function, For the first linear layer, For the second linear layer, For activation function, The input feature vector; The feature-feature relationship attention module is: In the formula: For relation weights, For the Sigmoid function, For the input feature vector, For the projection matrix, It is a relation matrix. The number of heads receiving multi-head attention; The mathematical model for the feature-feature relation attention module is as follows: In the formula: This is the new relationship matrix after the interaction between the heads. For normalized exponential functions, This is a multi-head relationship weight matrix. For element-wise multiplication, It is a learnable bilinear matrix. For the Sigmoid function, For gating weights, The gated weight matrix is ​​used to... Mapping to gate weights Dimensions It is a function of average value. This is the final result of the feature-feature relationship attention module.

2. The tailings ash content prediction method based on biomimetic neural networks according to claim 1, characterized in that, The multidimensional features include grayscale features, color features, texture features, and edge features, wherein: The grayscale feature includes the grayscale mean; The color features include the mean of the R / G / B channels, the standard deviation of the R / G / B channels, and the variance of the R / G / B channels; The edge features are Sobel edge features, including edge mean, variance, and energy; The texture features are GLCM texture features, including contrast, homogeneity, correlation, energy, and entropy.

3. The tailings ash content prediction method based on biomimetic neural networks according to claim 1, characterized in that, The mathematical model for the cross-modal multi-interaction attention module is as follows: In the formula: The projection vector of the feature-target relationship. The mapping function that projects feature-target attention weights onto the shared interaction space. For feature-target attention weights, Let be the matrix used for the linear transformation of the feature-target attention weights. This is the projection vector of the relation-relation weights. This is a multi-head relationship weight matrix. The mapping function that projects relation-relation weights onto the shared interaction space. This is the matrix used to perform a linear transformation on the multi-head relationship weight matrix. The adjustment factor is obtained through cross-modal interactive calculation. The final output is the weight of the cross-modal interaction relationship. It is a linear rectified function. It is an interactive core.

4. The tailings ash content prediction method based on biomimetic neural networks according to claim 1, characterized in that, The neural network includes a neural network model, a similarity calculation model, a weight adjustment model, and a learning rate adjustment model, wherein: The neural network model is as follows: In the formula: This is the predicted output value of the neural network for the target gray value. For activation function, For the input feature vector, For the final layer; The similarity calculation model is as follows: In the formula: For the current sample The overall similarity score with historical samples, For the features of the current sample vector, For the vector features of historical samples, This is the feature-target relation weight vector for the current sample. Historical Sample Feature-target relation weight vector, This is the feature-feature relationship weight matrix for the current sample. Historical Sample Feature-feature relation weight matrix The cosine similarity function is used. For historical memory The weighting coefficients, For current samples and historical memory The overall similarity For historical memory The prediction confidence level; The total number of samples stored in the historical memory bank. The index is the summation of historical samples. For current samples and historical samples The overall similarity Historical Sample The prediction confidence level; The weights of the sampled image features are adjusted by judging the difference between the predicted value and the verification target value. The weight adjustment model is as follows: In the formula: The set of all learnable parameters of a neural network. To represent parameters To optimize the objective by minimizing the value of subsequent expressions, For loss function, For function mapping, For the first The feature vectors of each training sample The prediction function of a neural network model. For the first The true gray value of each training sample The regularization coefficient is used. For the computation function of the attention mechanism, Here is the weight matrix for the attention mechanism. For query vector, For key vectors, For value vectors, For normalized exponential functions, This is the transpose of the key vector matrix. Let be the dimension of the key vector. For weighted elements; The learning rate adjustment model is as follows: In the formula: The step size is The learning rate The initial learning rate, The target correlation influence coefficient, The average attention weights are the feature importance. The coefficient representing the influence of the relationship between conditions. Let V be the variance of the characteristic relationship.

5. A tailings ash content prediction system based on biomimetic neural networks, characterized in that, The tailings ash content prediction method based on biomimetic neural networks as described in claim 1 includes: Image acquisition module: used to acquire images of tailings slurry water; Feature extraction module: used to extract multidimensional features from images of tailings coal slurry water. Dynamic weighting module: used to dynamically weight based on extracted multidimensional features through a multi-layer relational attention mechanism; Gray value prediction module: used to input dynamically weighted features into a pre-trained neural network to predict gray values; Ash value output module: Used to output ash value prediction results.

6. A tailings ash content prediction device based on a biomimetic neural network, characterized in that, Including processor and storage media; The storage medium is used to store instructions; The processor is configured to operate according to the instructions to perform the steps of the method according to any one of claims 1 to 4.

7. A computer-readable storage medium having a computer program stored thereon, characterized in that, When executed by a processor, the program implements the steps of the method according to any one of claims 1 to 4.

Images