CN-121563982-B - Visual analysis system for detecting grade of phosphorite flotation foam layer

CN121563982BCN 121563982 BCN121563982 BCN 121563982BCN-121563982-B

Abstract

The invention relates to the technical field of mineral processing visual detection, and discloses a visual analysis system for detecting the grade of a phosphorite flotation froth layer. The system comprises an image acquisition and decomposition module, a parallel feature extraction module, a dynamic feature fusion module, a foam evolution analysis module and a grade decision output module. The system performs multi-scale decomposition on the foam image, extracts physical and semantic features in parallel, and builds a dynamic fusion network based on bidirectional mapping for iterative interaction to generate a multi-mode feature descriptor. The self-organization growth of the foam evolution map is driven, the evolution track of the key foam graphic element is positioned and tracked, a foam grade state vector is formed, and finally, the regulation and control decision is output by combining external control parameters. The system realizes the deep fusion of physical and semantic features and the deep analysis of the dynamic evolution process of the foam, improves the accuracy and predictability of foam grade state sensing, and provides an effective means for the precise control of the flotation process.

Inventors

ZHANG YUNLONG
WANG LI
ZHU HUALEI
LIN JIBO

Assignees

山东鑫海矿业技术装备股份有限公司
烟台鑫海矿业研究设计有限公司

Dates

Publication Date: 20260512
Application Date: 20260122

Claims (8)

1. A visual analysis system for detecting grade of a phosphorite flotation froth layer, comprising the following modules: The image acquisition and decomposition module is used for continuously acquiring an original foam image stream from an imaging device on the surface of the flotation cell, performing multi-scale image decomposition on the original foam image stream and generating an image pyramid level set containing different resolutions; the parallel feature extraction module is used for parallelly executing foam physical feature extraction and foam semantic feature extraction in the image pyramid level set to generate a physical feature set and a semantic feature set; The dynamic feature fusion module establishes a bidirectional mapping relation between a foam physical feature set and a foam semantic feature set, and constructs a dynamic feature fusion network according to the bidirectional mapping relation, and specifically comprises the following steps: constructing a bidirectional mapping module formed by multi-head crossed attention layers; Inputting the physical feature set to a query vector generator of the bidirectional mapping module, and inputting the semantic feature set to a key vector generator and a value vector generator of the bidirectional mapping module; In the bidirectional mapping module, the attention score distribution of each feature vector in the physical feature set relative to all feature vectors in the semantic feature set is calculated, and the attention mapping from physical to semantic is generated; meanwhile, the attention score distribution of each feature vector in the semantic feature set relative to all feature vectors in the physical feature set is calculated, and semantic-to-physical attention mapping is generated; Splicing the physical-to-semantic attention map and the semantic-to-physical attention map, and transforming through a learnable projection matrix to generate a bidirectional attention fusion matrix; Taking the bidirectional attention fusion matrix as an initial connection weight of a dynamic feature fusion network, and initializing a feature interaction layer in the network; In the feature interaction layer, the physical features and the semantic features are subjected to information exchange and complementation of multiple iterations based on the association strength defined by the bidirectional attention fusion matrix; the method comprises the steps of carrying out iterative interaction and fusion on physical characteristics and semantic characteristics through a dynamic characteristic fusion network to generate a fused multi-mode foam characteristic descriptor, and specifically comprises the following steps: in each iteration of the dynamic feature fusion network, a feature interaction operation is performed, which includes two parallel sub-processes: the first sub-process is to project the physical feature representation of the current round to a semantic feature space through the corresponding weight slice in the bidirectional attention fusion matrix, and generate semantic space projection of the physical feature; The second sub-process projects the semantic feature representation of the current round to a physical feature space through another group of corresponding weight slices in the bidirectional attention fusion matrix to generate a physical space projection of the semantic feature; performing element-by-element gating fusion on semantic space projection of physical features and original current round semantic features, and updating semantic feature representation; Performing element-by-element gating fusion on the physical space projection of the semantic features and the original current round physical features, and updating the physical feature representation; after the preset iteration turns are completed, the finally updated physical characteristic representation and the finally updated semantic characteristic representation are connected in channel dimension; applying global average pooling and full-connection layer transformation to the connected feature representation, and compressing the feature representation into a feature vector with a fixed length, wherein the feature vector is a multi-mode foam feature descriptor; the foam evolution analysis module is used for driving the self-organization growth and dynamic updating of a foam evolution map based on the multi-mode foam feature descriptors, positioning foam primitives with grade indicators in the foam evolution map, marking the evolution tracks of the foam primitives, integrating the marked foam primitives and the evolution tracks of the foam primitives, and generating a foam grade state vector; and the grade decision output module is used for receiving external process control parameters and outputting a grade regulation decision sequence according to the real-time matching degree of the foam grade state vector and the process control parameters.
2. A visual analysis system for detecting grade of a froth layer in a phosphorite flotation as claimed in claim 1, wherein the performing of multi-scale image decomposition on the raw froth image stream generates a set of image pyramid levels comprising different resolutions, in particular comprising: Carrying out convolution processing on each frame of image in the original foam image stream by using a Gaussian differential filter bank to obtain a basic response atlas; performing iterative downsampling operation on the basic response atlas according to a preset downsampling rate sequence, and reserving a response image of a current scale after downsampling each time; The response images reserved in each iteration are sequenced and stored according to the corresponding scale level to form an image pyramid level set, and each level in the image pyramid level set corresponds to a specific spatial resolution; each level in the image pyramid level set is distributed with an independent feature extraction pipeline, and the feature extraction pipelines of different levels work in parallel; Within each feature extraction pipeline, parameters of feature extraction operators are adaptively configured according to spatial resolution characteristics of the assigned hierarchy.
3. The visual analysis system for detecting the grade of a foam layer in a phosphorite flotation process according to claim 2, wherein the steps of performing foam physical feature extraction and foam semantic feature extraction in parallel in the image pyramid level set to generate a physical feature set and a semantic feature set specifically comprise: In the physical feature extraction flow, for each level of the image pyramid level set, the following operations are performed in synchronization: Calculating the geometric outline of each foam region in the hierarchical image, and fitting out the minimum circumscribed polygon; Measuring the area, perimeter and compactness of the minimum circumscribing polygon as the geometric feature of each foam region; analyzing the pixel intensity distribution of each foam region in the hierarchical image, and calculating the color moment and texture co-occurrence matrix statistic of the pixel intensity distribution as the apparent characteristic of each foam region; performing cross-scale alignment and merging on geometric features and apparent features of foam areas from all layers to form a physical feature set; In the semantic feature extraction process, for each level of the image pyramid level set, the following operations are performed synchronously: Forward propagating the hierarchical image by using a pre-trained convolutional neural network encoder, and extracting deep feature mapping; Carrying out space pyramid pooling on the deep feature mapping to generate semantic coding vectors with fixed dimensions; inputting the semantic coding vector to an attention mechanism module, and calculating semantic significance weights of different foam areas in the hierarchical image; according to the semantic significance weight, carrying out weighted aggregation on semantic codes of the foam region to generate aggregated semantic features of the current processing level; And splicing and dimension-reducing the aggregated semantic features from all the layers to form a semantic feature set.
4. A visual analysis system for detecting grade of a foam layer of phosphorite flotation according to claim 3, wherein the self-organizing growth and dynamic updating of the foam evolution spectrum are driven based on the multi-mode foam characteristic descriptor, and the visual analysis system specifically comprises: taking a time-continuous multi-mode foam characteristic descriptor sequence as a potential node sequence of a foam evolution map; initializing an empty graph structure, and inserting a multi-mode foam feature descriptor at a first moment as a starting graph node; For the multi-mode foam feature descriptors at each subsequent moment, calculating cosine similarity of the multi-mode foam feature descriptors in a multi-mode feature space with all existing graph nodes in the graph; If the maximum value of the cosine similarity exceeds a preset growth threshold, connecting the descriptor at the current moment with the existing graph node with the highest similarity to form an evolution edge, wherein the weight of the evolution edge is the cosine similarity; If the maximum value of the cosine similarity does not exceed the growth threshold, using the descriptor at the current moment as a seed to create a new graph node in the graph; periodically scanning weights of all evolution edges in the graph, and deleting the evolution edge if the weight of one edge is lower than a preset maintenance threshold value in a plurality of continuous time windows; if a graph node is no longer connected to any other graph node after the deletion operation, the graph node is removed from the foam evolution graph.
5. A visual analysis system for detecting grade of a foam layer of phosphorite flotation according to claim 4, wherein the positioning of foam primitives with grade indication in a foam evolution map and marking of the evolution track thereof specifically comprises: in the foam evolution map, calculating the centrality measurement of each map node, wherein the centrality measurement comprehensively considers the centrality of each node, the centrality of the feature vector and the stability score of the corresponding multi-mode foam feature descriptor; screening out graph nodes with the centrality measurement higher than a preset importance threshold value, and marking the graph nodes as candidate key nodes; backtracking a historical connection edge of each candidate key node in the foam evolution map, and tracking a precursor node sequence from which the candidate key node is derived until the candidate key node is backtracked to a starting map node or another candidate key node, wherein the traced node and edge form a candidate evolution track; Extracting multi-mode foam feature descriptors of all graph nodes contained in each candidate evolution track, and calculating covariance matrixes of descriptor changes on the candidate evolution tracks; analyzing the principal components of the covariance matrix, projecting the track to a principal component space, and calculating the smoothness and direction consistency of the projected track; And formally marking the candidate evolution tracks with smoothness and direction consistency higher than the preset track quality threshold as the evolution tracks of the foam primitives with grade indication, and formally marking the associated candidate key nodes as the grade indication primitives.
6. The visual analysis system for detecting the grade of a foam layer of phosphorite flotation according to claim 5, wherein the integrating of marked foam primitives and evolution tracks thereof generates a foam grade state vector, and specifically comprises: For each marked grade indicating graphic element, reading a corresponding multi-mode foam characteristic descriptor; extracting track characteristics including track length, average weight of all evolution edges on the track, smoothness score of the track and direction consistency score of the track for each marked evolution track; Creating an initial empty foam grade state vector template, wherein the dimension of the foam grade state vector template is determined by the number of grade indicative graphic elements and the number of marked evolution tracks; The multi-mode foam feature descriptors of each grade indicating graphic element are subjected to dimension reduction and standardization through a special embedded layer, and are converted into graphic element state subvectors with fixed dimensions; processing the track characteristics of each marked evolution track through a track coding network, and converting the track characteristics into track state subvectors with fixed dimensions; Sequentially connecting all the primitive state subvectors with all the track state subvectors according to a predefined splicing sequence to form a high-dimensional composite vector; and inputting the high-dimensional composite vector into a self-encoder with a bottleneck structure for compression and denoising, and outputting the intermediate bottleneck layer of the self-encoder into a finally generated low-dimensional dense foam grade state vector.
7. The visual analysis system for detecting the grade of a foam layer in a phosphorite flotation as claimed in claim 6, wherein the receiving the external process control parameters and outputting a grade control decision sequence according to the real-time matching degree of the foam grade state vector and the process control parameters specifically comprises: receiving external process control parameters from a flotation process control system in real time through an industrial data interface, the parameters including, but not limited to, medicament flow, aeration, pulp level; encoding external process control parameters into a process parameter vector, the dimensions of which are matched with the dimensions of the foam grade state vector; calculating the real-time matching degree between the foam grade state vector and the process parameter vector, wherein the matching degree is obtained by calculating the dot product of the two vectors and then normalizing the dot product by a Sigmoid function; Establishing a preset decision rule base, wherein the decision rule base comprises a plurality of matching degree intervals and candidate regulation and control actions associated with each interval; Searching a matching degree interval to which the real-time matching degree belongs according to the calculated real-time matching degree, and activating a candidate regulation and control action set corresponding to the matching degree interval; selecting one regulation action with the highest success rate from the activated candidate regulation action sets according to the sequence of the historical execution success rate as a preferred decision at the current moment; continuity checking is carried out on the optimal decision at the current moment and decisions at a plurality of last moments in a short-term decision cache, and if conflict exists, sub-optimal decisions which are consistent with historical decisions are selected from a candidate regulation action set to be replaced; and adding the final decision subjected to the continuity check to the tail end of the grade regulation decision sequence, and sending the grade regulation decision sequence to a flotation process control system in real time for execution.
8. The visual analysis system for detecting grade of a phosphate rock flotation froth layer according to claim 7, wherein the analyzing the principal components of the covariance matrix, projecting trajectories into a principal component space, and calculating smoothness and directional consistency of the projected trajectories comprises: Performing feature decomposition on the covariance matrix, and extracting feature values and corresponding feature vectors; arranging the characteristic values in a descending order, and selecting characteristic vectors corresponding to a plurality of the maximum characteristic values as main component directions; projecting the multi-mode foam feature descriptors of all the graph nodes on the candidate evolution track to the main component direction respectively to obtain the coordinates of each descriptor in a main component space; connecting the coordinate points according to time sequence to form a projected track; Calculating direction vectors between adjacent coordinate points on the projected track, and calculating cosine values of included angles between all adjacent direction vectors; Calculating an average value of cosine values of all included angles, and taking the average value as a measure of direction consistency of the track after projection; And calculating the distance deviation of a line segment formed by each coordinate point on the projected track and the adjacent points before and after the coordinate point, and solving a root mean square value for all the distance deviations to be used as the smoothness score of the projected track.

Description

Visual analysis system for detecting grade of phosphorite flotation foam layer Technical Field The invention relates to the technical field of mineral processing visual detection, in particular to a visual analysis system for detecting the grade of a phosphorite flotation froth layer. Background In the phosphorite flotation process, the apparent state of foam is a key reflection of mineral grade. In the prior art, foam images are captured mainly through an industrial camera, and static physical characteristics such as foam size, shape, texture and the like are extracted by adopting an image processing algorithm, or deep semantic characteristics of the images are extracted by utilizing a trained convolutional neural network model. The technical schemes convert foam visual information into a series of characteristic parameters for establishing a statistical correlation model with grade. Part of the existing schemes adopt traditional machine learning models such as a support vector machine and a random forest, and establish a grade association model based on single-scale foam physical characteristics, but the model has weak generalization capability, and is difficult to adapt to working condition fluctuation such as pulp property, medicament dosage and the like in the flotation process. In addition, the technology focuses on the shallow application of semantic features only, and foam multi-scale structural information is not combined, so that the feature characterization lacks layering sense and cannot cover the visual signals related to grade on the whole. The existing scheme has defects. The physical characteristics are easy to obtain but the characterization dimension is shallow, and the hidden information of complex textures, gloss and the like related to grade is difficult to capture, while the semantic characteristics can express deep patterns, but the physical meaning is fuzzy and is disjointed with the specific physical state of the foam. The two are usually independently analyzed or simply spliced in later period, and the deep fusion and mutual verification of the information cannot be realized, so that the characterization accuracy of the feature descriptors on complex and dynamic foam states is limited. In addition, the traditional method regards continuous image frames as independent samples or only carries out simple time sequence difference, and lacks structural description and tracking on the complete life cycle from generation, evolution to extinction of foam individuals, and cannot explore a grade indication signal with more robustness from dynamic evolution rules. Disclosure of Invention The invention aims to provide a visual analysis system for detecting the grade of a phosphorite flotation froth layer, so as to solve the problems in the background technology. To achieve the above object, the present invention provides a visual analysis system for detecting the grade of a phosphate rock flotation froth layer, the system comprising: The image acquisition and decomposition module is used for continuously acquiring an original foam image stream from an imaging device on the surface of the flotation cell, performing multi-scale image decomposition on the original foam image stream and generating an image pyramid level set containing different resolutions; the parallel feature extraction module is used for parallelly executing foam physical feature extraction and foam semantic feature extraction in the image pyramid level set to generate a physical feature set and a semantic feature set; The dynamic feature fusion module is used for establishing a bidirectional mapping relation between a foam physical feature set and a foam semantic feature set, establishing a dynamic feature fusion network according to the bidirectional mapping relation, and carrying out iterative interaction and fusion on physical features and semantic features through the dynamic feature fusion network to generate a fused multi-mode foam feature descriptor; the foam evolution analysis module is used for driving the self-organization growth and dynamic updating of a foam evolution map based on the multi-mode foam feature descriptors, positioning foam primitives with grade indicators in the foam evolution map, marking the evolution tracks of the foam primitives, integrating the marked foam primitives and the evolution tracks of the foam primitives, and generating a foam grade state vector; and the grade decision output module is used for receiving external process control parameters and outputting a grade regulation decision sequence according to the real-time matching degree of the foam grade state vector and the process control parameters. Preferably, the performing multi-scale image decomposition on the original foam image stream to generate an image pyramid hierarchy set containing different resolutions specifically includes: Carrying out convolution processing on each frame of image in the original foam image stream by usin