Flotation process working condition identification method and system based on dynamic time series graph neural network

Abstract

The present invention discloses a flotation process working condition identification method and system based on a dynamic time-series graph neural network, comprising the following steps: selecting important surface features reflecting the foam image based on empirical knowledge of the actual froth flotation production process; constructing a dynamic graph structure for the froth flotation cell using the surface features as graph nodes, and using an adjacency matrix to represent the interaction relationship between feature variables; mining the changes in foam image feature variables over time from the perspectives of long-term and short-term relationships; and training the network by updating the graph network content based on an established message-passing graph neural network model to obtain working condition classification results. The present invention utilizes surface feature variables to establish a graph network to overcome the differences and complexity between flotation cells, highlighting the impact of the relationship between variables on flotation working conditions, helping workers to identify the working conditions of the flotation cells and promoting efficient and stable operation of the flotation process.

Description

Technical Field

[0001] The present invention mainly relates to the technical field of froth flotation working condition monitoring, and in particular to a lead-zinc froth flotation video acquisition and working condition intelligent monitoring based on graph neural network. Background Art

[0002] Mineral resources, as a key material foundation for social development and the national economy, play an indispensable role in modern industry. Froth flotation, a widely used beneficiation method in lead and zinc ore processing, has a lengthy process flow and complex mechanisms across multiple flotation cells. Therefore, operating condition monitoring and control technologies during flotation are crucial, directly impacting the efficient output of lead and zinc ore. Currently, due to the lack of accurate and effective operating condition monitoring methods and equipment, flotation plants generally rely on manual observation of froth patterns to identify the current operating status of flotation cells. However, manual judgment of operating conditions has numerous drawbacks. Differences in worker experience lead to subjective and inconsistent results, providing invalid or even erroneous information for the production process, often resulting in large fluctuations in concentrate grade and low recoveries of valuable minerals. Experienced workers are in short supply, and accumulating operational experience requires time. Manual judgment increases workload and production costs. To address this issue and move beyond the current reliance on experienced technicians, an intelligent, automated approach is urgently needed to replace traditional manual monitoring methods and promote the sustainable development of the froth flotation industry.

[0003] Foam images are an important source of data in the froth flotation process, containing a wealth of information. Previous studies have shown that the visual characteristics of the foam surface are closely correlated with flotation production indicators and provide a direct reflection of operating conditions. By analyzing and processing foam images, automatic recognition, segmentation, and feature extraction can be achieved, enabling real-time monitoring and control of the flotation process. With the development of deep learning in the field of froth flotation, the use of convolutional networks to process foam images has become an effective method. However, these methods rely on large amounts of training data to improve model performance, resulting in high investment costs. Furthermore, they cannot incorporate human experience and tend to overlook correlations between foam features.

[0004] Therefore, an efficient, low-cost and accurate method for automatic identification of flotation conditions is needed.

[0005] Glossary:

[0006] Adaptive Dynamic Graph Structure Learning: Adaptive dynamic graph structure learning is a dynamic graph construction method that automatically adjusts the graph structure (such as node and edge connectivity) based on the characteristics of the input data and task requirements to capture dynamic changes and complex relationships in the data. The graph structure is not fixed but dynamically adjusted based on data characteristics and task requirements. Node embeddings and adjacency matrices are generated through learnable parameters and nonlinear transformations.

[0007] Dynamic graph: A type of graph structure in which the connections between nodes and edges change over time and can reflect the temporal dynamics in the data.

[0008] Temporal convolution: Apply convolution operations on the time dimension to process time series data and capture time series dependencies. Summary of the Invention

[0009] The technical problem to be solved by the present invention is to propose a flotation process working condition identification method and system based on a dynamic time sequence graph neural network in response to the shortcomings of the existing technology.

[0010] A method for identifying working conditions of a flotation process based on a dynamic time sequence graph neural network comprises the following steps:

[0011] S1: Obtain the main characteristic variables describing the foam image as foam flotation features; classify the flotation working conditions to obtain keywords describing the foam state of different working conditions;

[0012] S2: Obtain a historical foam flotation video sample set, extract foam flotation features of the video sample set using a machine vision method, and then preprocess to obtain a standardized time series vector Z of the foam flotation features;

[0013] S3: Based on the theory of dynamic graphs, the normalized time series vector Z is used to construct a dynamic graph of froth flotation characteristics;

[0014] S4: The dynamic graph of foam flotation features is used as the input of the message passing graph neural network MPNN that introduces the cumulative fusion of cross-time features. The state update of the nodes in the message passing graph neural network MPNN is realized through message passing and aggregation function. After the c-layer MPNN network is nested, a new dynamic graph G is obtained. Dynamic '; S5: Use pooling operation to compress the new dynamic graph G Dynamic The feature information of nodes and edges is sparsely represented in a graph. The embedded output after pooling is mapped into the probability distribution of different working condition categories through a fully connected layer, thus realizing the working condition classification of froth flotation.

[0015] S6: Construct a cross-entropy loss function as supervision to measure the difference between the model results and the actual working condition labels. Use the loss function to train the message passing graph neural network (MPNN) that integrates the cumulative features across time periods. Update the parameters through backpropagation and repeat steps S4-S5 until the cross-entropy loss function is minimized to obtain the final message passing graph neural network.

[0016] S7: When actually identifying the froth flotation working condition, the apparent feature time series in the froth video is extracted and input into the final message passing graph neural network. After message passing and time series feature extraction, the model outputs the classification result of the current working condition.

[0017] Further improvement, in step S1, the main characteristic variables include relative red component, homogeneity, contrast, energy, entropy, size mean, size variance, velocity mean and velocity variance;

[0018] The keywords describing the foam state in different working conditions include: moderate bubble size and flow rate, stable foam without bursting; low mineral attachment rate, large bubble size and directional foam flow; low mineral attachment rate, large bubble size and fast foam flow rate, poor stability; relatively small bubble size and low foam flow rate; and excessive particles carried in the foam layer, resulting in poor foam stability and severe deformation.

[0019] For further improvement, the specific steps of step S2 are as follows:

[0020] S21: Obtain a video sample set of historical foam floating;

[0021] S22: Slice the video and use machine vision methods to collect the main feature variables of the image frame every a seconds;

[0022] S23: A set of foam features of a video constitutes a set of vectors. After preprocessing the data, the apparent feature time series vector X = {x1, x2, ..., x N}, Where N represents the number of foam features, i represents the type of foam features, and m represents the time length. represents the value of the i-th foam characteristic at time point m, x i The time series vector representing the characteristics of the i-th foam; each set of apparent feature vectors corresponds to an actual working condition category, and these vector sets are used as input to construct the dynamic graph to capture the working condition information.

[0023] Further improvement, in step S22, a=0.6;

[0024] The method for extracting the main characteristic variables is as follows:

[0025] Relative redness:

[0026] Homogeneity:

[0027] Contrast:

[0028] Energy:

[0029] Entropy:

[0030] Size mean m :

[0031] Size variance s v :

[0032] Mean speed v m : Speed ​​variance v v :

[0033] Among them, Red mean Represents the pixel mean of the red channel in a color image; Grey mean represents the mean value of all pixels after converting the color image into a grayscale image; P represents the grayscale co-occurrence matrix, i and j represent different grayscale values ​​in the image, P(i, j) represents the probability of occurrence of a pixel pair consisting of grayscale value i and grayscale value j in the image; n represents the total number of bubbles in the image; p b represents the pixel value of the bth bubble; v b Represents the horizontal velocity of the bth bubble.

[0034] As a further improvement, in step S23, the pre-processing method is as follows:

[0035]

[0036]

[0037] Where, represents the value of the i-th foam characteristic at time point k, represents the mean of the time series vector of the i-th foam characteristic, δ(x i ) represents the standard deviation of the time series vector of the i-th foam characteristic, z i It represents the result of normalization of the time series vector of the i-th foam characteristic. After the above processing, the normalized time series vector Z = {z1, z2, ..., z N}.

[0038] As a further improvement, the steps of step S3 are as follows:

[0039] The foam flotation features are used as the initial attributes of the graph nodes of the dynamic graph. Based on the short-term relationship extraction, an adaptive dynamic graph structure learning method is used to obtain a set of adjacency matrices of the dynamic graph. Temporal convolution is used to process the long-term relationship to obtain the embedded representation of the graph nodes, and a dynamic graph of the foam flotation features is obtained.

[0040] For further improvement, the specific steps of step S3 are as follows:

[0041] S31: Normalized apparent feature time series vector Z = {z1,z2,…,z N Abstraction into a dynamic graph: The foam flotation features are regarded as nodes of the dynamic graph, and the mutual influence relationship of the foam flotation features in the feature space is abstracted into an adjacency matrix;

[0042] S32: Discretize the dynamic graph: Divide a set of feature time series vectors into multiple feature time series sub-vectors according to equidistant time slices, and construct a set of multi-time slice sub-graphs based on these feature time series sub-vectors to capture the dynamic relationship of the main feature variables over time and form a discretized dynamic graph; the discretized dynamic graph is composed of a four-tuple: G Dynamic ={V,Z,E,T}, where

[0043] T={t1,t2,…,t s}, t S >0 is the time slice divided equally along the time axis, which is the minimum time unit for exploring the relationship between the main characteristic variables and time; t is the set of equally spaced time slices, s represents the total number of time slices; V = {v1, v2, ..., v N} is the set of graph nodes that remains unchanged across all time slices, N represents the number of foam features, and v N It is the Nth graph node consisting of the Nth type of foam features;

[0044] Z={z1,z2,…,z s}, It is a set of feature vectors based on equidistant time slices, representing the initial embedding of the node; represents the normalized eigenvalue of the Nth foam feature at the sth time slice; is an edge set based on equidistant time slices, represents the time slice t i The edge feature set of the upper node reflects the dynamic changes of the node relationship; s It is a time slice divided equally along the time axis;

[0045] S33: Determine the number of time slices s, and obtain the adjacency matrix representation of the multi-time slice subgraph through an adaptive dynamic graph structure learning method;

[0046] S34: For the normalized apparent feature time series vector Z={z1,z2,…,z N}, and use the time convolution method to process the apparent feature time series vector Z={z1,z2,…,z N}, each node contains the long-term time-varying embedding vector Z′;

[0047] S35: Match the embedded vector Z′ with the multi-time slice subgraph according to the time slice to obtain the foam flotation feature dynamic graph G Dynamic ={V,Z′,E,T}.

[0048] As a further improvement, the message passing graph neural network MPNN that introduces cross-period feature accumulation fusion is obtained by introducing cross-period feature accumulation fusion into the message passing graph neural network. The specific definition formula is as follows:

[0049] (1) Message passing function:

[0050] (2) Aggregate functions:

[0051] Among them, l represents the current network layer number, v is the current node, γ represents the neighbor nodes of node v, and N(v) is the set of neighbor nodes of the current node in the feature space connection relationship. is the feature representation of node v at the t-1th time slice of the lth layer, and are the feature representations at the time slice of the lth layer of node v and γ, respectively. is the feature representation of node v at the t-th time slice of the l+1 layer after the update, is the edge feature of the t-th time slice, is the message vector of node v at the t-th time slice in the l+1 layer; the M(·) operation propagates all node information to their respective neighbor nodes to obtain the message vector of a new layer; the U(·) operation is an aggregation function, which is used to aggregate the neighbor node vector of the current layer, the neighbor node vector of the previous time step and the message vector to obtain the updated node feature representation.

[0052] For further improvement, the specific steps of step S5 are as follows:

[0053] S51: Using graph nodes as channels, we use a two-dimensional convolutional network to compress the dynamic graph G Dynamic The number of nodes of ′ is determined by N l Pooling is N l+1 , reducing the embedding dimension of graph representation;

[0054] Among them, the pooling method is the Diffpool pooling method, which assigns nodes to different clusters through learning, so that the dynamic graph G Dynamic The ′ structure is compressed to generate a sparse graph representation, retaining important graph structure information:

[0055] S l =softmax(GNN pool (A l ,xl ))

[0056] S l is the allocation matrix, which represents the mapping from nodes to clusters. The pooled adjacency matrix and node feature matrix are:

[0057] A l =S lT A l S l ;

[0058] x l+1 =S lΤ x l ;

[0059] Among them, A l is the adjacency matrix of the lth layer, S lT is the allocation matrix S l The transpose of x l and x l+1 Node feature information, GNN pool It is a function composed of the above message passing function and aggregation function;

[0060] Each layer is composed of a message passing module of the message passing graph neural network MPNN that introduces cross-time feature accumulation and fusion and a DiffPool pooling module connected in sequence. After k layers of nesting, the final graph representation x is output. k and the adjacency matrix A k , as a sparse representation of the graph;

[0061] S52: Use a fully connected network to map the sparse representation of the graph into the probability distribution of the working condition category;

[0062] The sparse representation x k Input the fully connected layer to obtain the working condition classification results. The specific steps are as follows:

[0063] For the sparse representation x k Perform global pooling to generate the global feature vector x of the graph G :

[0064] x G =pool(x k )

[0065] pool(·) indicates global pooling;

[0066] The global eigenvector x G Input the fully connected layer and obtain the probability distribution y of the working condition classification through the nonlinear transformation function:

[0067] y=softmax(w·x G +b)

[0068] in, is the weight matrix of the fully connected layer; is the bias vector; c is the number of working condition categories, is the probability distribution of the classification results; represents a real number, N k Indicates the number of graph nodes after k-layer graph pooling;

[0069] According to the probability distribution y, the class with the highest probability is selected as the final working condition classification result class: class = argmax(y)

[0070] argmax represents the maximum value function.

[0071] For further improvement, the cross entropy loss function is constructed as follows:

[0072]

[0073] Among them, y is the probability distribution of working condition classification, class true is the one-hot encoding of the true label, and N is the number of samples

[0074] Compared with the prior art, the present invention has the following beneficial effects:

[0075] (1) The dynamic graph model of foam surface characteristics constructed by the flotation process working condition identification method based on dynamic time series graph neural network proposed in the present invention can make the interaction relationship between features explicit, accurately capture the long-term and short-term fluctuations of foam characteristics, learn the high-order embedding of nodes in time series, and enhance the interpretability of graph neural network classification.

[0076] (2) This invention proposes a method for classifying working conditions based on machine vision and utilizing the surface characteristics of foam, which can solve the problem of heavy reliance on experienced operators in actual production processes. To address the problems of strong subjectivity in manual judgment, the subjectivity and imprecision of empirical descriptions, and the difficulty in accumulating experience, a flotation process working condition identification method based on a dynamic time-series graph neural network is adopted to improve the intelligence and objective stability of working condition judgment. BRIEF DESCRIPTION OF THE DRAWINGS

[0077] Figure 1 This is a flow chart of a method for identifying operating conditions of a flotation process based on a dynamic timing graph neural network provided by an embodiment of the present invention.

[0078] Figure 2 It is a graph neural network working condition recognition model based on message passing provided by an embodiment of the present invention. DETAILED DESCRIPTION

[0079] The following will fully and clearly describe the technical solutions of the specific embodiments of the present invention in conjunction with the accompanying drawings of the embodiments of the present invention. However, the specific embodiments described are only part of the implementation cases of the present invention, not all of the implementation cases. Based on the implementation cases of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of the present invention.

[0080] Figure 1 The overall logical framework of the present invention is based on the analysis of the froth flotation process in industry, selecting important characteristic variables to mainly describe the changes in working conditions and classifying the working conditions, which mainly includes the following steps:

[0081] S1: In the actual production process, workers give some simple definitions to the working conditions by observing the color, size, stability and breakage of the foam. Table 1 shows in detail the main characteristic variables that are screened to describe the foam image.

[0082] Table 1

[0083]

[0084] Table 2 shows in detail the characteristic description of the foam state under five different flotation conditions.

[0085] Table 2

[0086]

[0087]

[0088] S2: Obtain historical froth flotation cell video data, use machine vision methods to extract the main feature variables, and obtain the time series vector representation of the feature variables through structured processing.

[0089] (1) Sample a 1-minute video every 6 minutes for 5 consecutive days, collect the features of the image frame every 0.6 seconds, and obtain the foam video collection.

[0090] The feature extraction formula is as follows:

[0091] Relative redness:

[0092] Homogeneity:

[0093] Contrast:

[0094] Energy:

[0095] Entropy:

[0096] Size mean m :

[0097] Size variance s v :

[0098] Mean speed v m :

[0099] Speed ​​variance v v :

[0100] Among them, Red mean Represents the pixel mean of the red channel in a color image; Grey mean represents the mean value of all pixels after converting the color image into a grayscale image; P represents the grayscale co-occurrence matrix, i and j represent different grayscale values ​​in the image, P(i, j) represents the probability of occurrence of a pixel pair consisting of grayscale value i and grayscale value j in the image; n represents the total number of bubbles in the image; p b represents the pixel value of the bth bubble; v b Represents the horizontal velocity of the bth bubble.

[0101] (2) Data preprocessing: perform standard normalization on the collected characteristic variables.

[0102]

[0103]

[0104] Where, represents the value of the i-th foam characteristic at time point k, represents the mean of the time series vector of the i-th foam characteristic, δ(x i ) represents the standard deviation of the time series vector of the i-th foam characteristic, z i It represents the result of normalization of the time series vector of the i-th foam characteristic. After the above processing, the normalized time series vector Z = {z1, z2, ..., z N}.

[0105] S3: In this embodiment, based on the concept of the dynamic graph, a set of multiple time slices is established, and a graph structure dynamic learning method is used to obtain the adjacency matrix representation of the subgraph. Based on the temporal subgraph, the long-term and short-term temporal relationships are processed to obtain the embedded representation of the node.

[0106] (1) In this embodiment, the number of time slices s is set to 100, and the time axis is divided into time slices at equal intervals. A set of feature vectors based on the equal-interval time slice division is obtained as the initial embedding content of the graph node.

[0107] Normalized time series vector slice representation Z={z1,z2,…,z 100 Contains all collected feature data, where a single slice contains the values ​​of 9 surface features in a certain time period Time slice set: T = {t1, t2, ..., t 100}

[0108] (2) Based on the slicing of a set of feature vectors and the method of dynamic graph structure learning, the node embedding is converted into a scalar that is easy to divide by utilizing the feature transformation of the learnable matrix parameters and the multiplication operation of the scaling factor. The adaptive selection mechanism is supported to select the first k nodes for graph structure representation, and the graph structure and short-term relationship representation of the multi-time slice subgraph are obtained.

[0109] In this embodiment, the learnable parameters θ1 and θ2 and nonlinear transformation are used to perform feature mapping on the feature variables of each subgraph to obtain node embeddings C1 and C2. The transposed product difference of the node embeddings C1 and C2 is calculated, and a rectified linear unit is introduced to obtain scalarized interaction information. The calculation formula is as follows:

[0110] A=ReLU(σ(C1C2 T -C2C1 T ))

[0111] On the basis of the relationship matrix, an adaptive selection mechanism is used to screen the first k important nodes to form the adjacency matrix A. adapt , the calculation formula is as follows:

[0112] A adapt =argtopk(A)

[0113] Among them, the hyperparameter k is selected as 6. At the beginning of training, the learning parameters are randomly initialized and optimized and adjusted as the model iterative training is carried out to achieve dynamic learning. adapt At the same time, we also learn each time slice χ i The short-term relationship of the apparent characteristic sequence.

[0114] (3) Treat a set of feature vectors X as a continuous time series, process the complete time series using the time convolution method, and obtain an embedding X′ that contains long-term time changes.

[0115] In this embodiment, the feature vector X={χ1,χ2,…,χ S}S=100 is used as the input variable, and the one-dimensional time convolution kernel K={k1,k2,…,k L}L=32 to extract features. That is,

[0116]

[0117] Among them, t is the time step index of the output sequence, the convolution kernel size kernel_size = 9, and the padding sequence padding = 4.

[0118] S4: After the above steps, a time sequence graph with clear structure and node representation is obtained, and the input features are input into the MPNN message passing graph neural network to obtain a graph representation containing neighborhood information and topological structure, which is used for the subsequent analysis of similar patterns of foam image features.

[0119] (1) Specifically, the message function M(·) is formulated as follows:

[0120]

[0121] in, Based on the edge relationship To simplify the calculation, the determined learnable matrix is ​​represented by the connection relationship 0 and 1 of the adjacency matrix.

[0122] (2) During the aggregation process, mean aggregation is selected, and the neighbor node vector of the current layer, the neighbor node vector of the previous time step, and the message vector are calculated through the MLP function. The formula is as follows:

[0123]

[0124] Where |N(v)| is the number of neighbors of node v, x γ is the eigenvector of the neighbor node γ; σ represents the Sigmoid function, w l+1 and b l+1 are the weights and biases of layer l.

[0125] S5: Compress the graph structure and node features to obtain the embedded output, map the output vector into the probability distribution of different working condition categories, and obtain the label of the current working condition.

[0126] (1) Using the Diffpool pooling method, nodes are assigned to different clusters through learning, so that the graph structure is compressed, a sparse graph representation is generated, and important graph structure information is retained:

[0127] S l =softmax(GNN pool (A l ,x l ))

[0128] S l is the allocation matrix, which represents the mapping from nodes to clusters. The pooled adjacency matrix and node feature matrix are:

[0129] A l =S lT A l S l ;

[0130] x l+1 =S lΤ x l ;

[0131] This embodiment adopts a k=3-layer nested structure, where each layer is formed by sequentially connecting the MPNN message passing module and the DiffPool pooling module. After the k=3 nested operation, the final graph representation x is output. 3 and the adjacency matrix A 3 , as a sparse representation of the graph.

[0132] (2) Sparse representation x 3 Input the fully connected layer to obtain the working condition classification results. The specific steps are as follows:

[0133] For the sparse representation x 3 Perform global pooling to generate the global feature vector x of the graph G :

[0134] x G =pool(x 3 )

[0135] The global eigenvector h G Input the fully connected layer and obtain the probability distribution of working condition classification through nonlinear transformation and function:

[0136] y=softmax(w·x G +b)

[0137] in, is the weight matrix of the fully connected layer; is the bias vector; c is the number of working condition categories, is the probability distribution of the classification results; represents a real number, N k Indicates the number of graph nodes after k-layer graph pooling.

[0138] According to the probability distribution y, the category with the highest probability is selected as the final working condition classification result:

[0139] class=argmax(y)

[0140] S6: Construct the cross entropy function to calculate the classification results of the model class and classtrue The loss between them is calculated, and then the gradient of the loss function L with respect to the network parameters is calculated through the back propagation algorithm, and the gradient descent (Adam) is used to update the network parameters.

[0141] Among them, the total loss value is:

[0142] Through the above process of building a network model and training, the trained model is used to process the apparent feature data of the flotation tank foam video to give the current working condition classification result.

[0143] This paper proposes a flotation process condition identification method based on a dynamic time-series graph neural network. This method describes the apparent characteristics of foam as a node network, forming an interrelated graph. Using a message-passing graph neural network and Difpool pooling to analyze the interactions between nodes, the flotation process's operating status is analyzed. The influence of the apparent characteristics of foam is revealed in a graph structure. This network model effectively determines the current operating status, intelligently compensating for the subjectivity of manual judgment. Actual performance in the flotation industry demonstrates the model's reliability.

[0144] Obviously, those skilled in the art may make various changes and modifications to this application without departing from the spirit and scope of this application. Thus, if these modifications and variations of this application fall within the scope of the claims of this application and their equivalents, this application is intended to include these modifications and variations.

Claims

1. A flotation process operating condition identification method based on dynamic time series graph neural network, characterized in that: The steps include: S1: Obtain the main characteristic variables describing the foam image as foam flotation features; classify the flotation working conditions to obtain keywords describing the foam state of different working conditions; S2: Obtain a historical foam flotation video sample set, extract foam flotation features of the video sample set using a machine vision method, and then preprocess to obtain a standardized time series vector Z of the foam flotation features; S3: Based on the theory of dynamic graphs, the normalized time series vector Z is used to construct a dynamic graph of froth flotation characteristics; S4: The dynamic graph of foam flotation features is used as the input of the message passing graph neural network MPNN that introduces the cumulative fusion of cross-time features. The state update of the nodes in the message passing graph neural network MPNN is realized through message passing and aggregation function. After the c-layer MPNN network is nested, a new dynamic graph G is obtained. Dynamic '; The message passing graph neural network MPNN that introduces cross-period feature accumulation fusion is obtained by introducing cross-period feature accumulation fusion into the message passing graph neural network. The specific definition formula is as follows: (1) Message passing function: (2) Aggregate functions: Among them, l represents the current network layer number, v is the current node, γ represents the neighbor nodes of node v, and N(v) is the set of neighbor nodes of the current node in the feature space connection relationship. is the feature representation of node v at the t-1th time slice of the lth layer, and are the feature representations at the time slice of the lth layer of node v and γ, respectively. is the feature representation of node v at the t-th time slice of the l+1 layer after the update, is the edge feature of the t-th time slice, is the message vector of node v at the t-th time slice of the l+1 layer; the M(·) operation propagates all node information to their respective neighbor nodes to obtain the message vector of the new layer; the U(·) operation is an aggregation function used to aggregate the neighbor node vector of the current layer, the neighbor node vector of the previous time step, and the message vector to obtain the updated node feature representation; S5: Use pooling operation to compress the new dynamic graph G Dynamic The feature information of nodes and edges is sparsely represented in a graph. The embedded output after pooling is mapped into the probability distribution of different working condition categories through a fully connected layer, thus realizing the working condition classification of froth flotation. S6: Construct a cross-entropy loss function as supervision to measure the difference between the model results and the actual working condition labels. Use the loss function to train the message passing graph neural network (MPNN) that integrates the cumulative features across time periods. Update the parameters through backpropagation and repeat steps S4-S5 until the cross-entropy loss function is minimized to obtain the final message passing graph neural network. S7: When actually identifying the froth flotation working condition, the apparent feature time series in the froth video is extracted and input into the final message passing graph neural network. After message passing and time series feature extraction, the model outputs the classification result of the current working condition.

2. The flotation process operating condition identification method based on dynamic time sequence graph neural network according to claim 1, characterized in that: In step S1, the main characteristic variables include relative red component, homogeneity, contrast, energy, entropy, size mean, size variance, speed mean and speed variance; The keywords describing the foam state in different working conditions include: moderate bubble size and flow rate, stable foam without bursting; low mineral attachment rate, large bubble size and directional foam flow; low mineral attachment rate, large bubble size and fast foam flow rate, poor stability; relatively small bubble size and low foam flow rate; and excessive particles carried in the foam layer, resulting in poor foam stability and severe deformation.

3. The flotation process operating condition identification method based on dynamic time sequence graph neural network according to claim 1, characterized in that: The specific steps of step S2 are as follows: S21: Obtain a video sample set of historical foam floating; S22: Slice the video and use machine vision methods to collect the main feature variables of the image frame every a seconds; S23: A set of foam features of a video constitutes a set of vectors. After preprocessing the data, the apparent feature time series vector X = {x1, x2, ..., x N }, Where N represents the number of foam features, i represents the type of foam features, and m represents the time length. represents the value of the i-th foam characteristic at time point m, x i The time series vector representing the characteristics of the i-th foam; each set of apparent feature vectors corresponds to an actual working condition category, and these vector sets are used as input to construct the dynamic graph to capture the working condition information.

4. The flotation process operating condition identification method based on dynamic time sequence graph neural network according to claim 3 is characterized in that: In step S22, a=0.6; The method for extracting the main characteristic variables is as follows: Relative redness: Homogeneity: Contrast: Energy: Entropy: Size mean m : Size variance s v : Mean speed v m : Speed ​​variance v v : Among them, Red mean Represents the pixel mean of the red channel in a color image; Grey mean represents the mean value of all pixels after converting the color image into a grayscale image; P represents the grayscale co-occurrence matrix, i and j represent different grayscale values ​​in the image, P(i, j) represents the probability of occurrence of a pixel pair consisting of grayscale value i and grayscale value j in the image; n represents the total number of bubbles in the image; p b represents the pixel value of the bth bubble; v b Represents the horizontal velocity of the bth bubble.

5. The flotation process operating condition identification method based on dynamic time sequence graph neural network according to claim 3, characterized in that: In step S23, the pre-processing method is as follows: Where, represents the value of the i-th foam characteristic at time point k, represents the mean of the time series vector of the i-th foam characteristic, δ(x i ) represents the standard deviation of the time series vector of the i-th foam characteristic, z i It represents the result of normalization of the time series vector of the i-th foam characteristic. After the above processing, the normalized time series vector Z = {z1, z2, ..., z N }.

6. The method for identifying flotation process conditions based on a dynamic time series graph neural network according to claim 1, wherein: The steps of step S3 are as follows: The foam flotation features are used as the initial attributes of the graph nodes of the dynamic graph. Based on the short-term relationship extraction, an adaptive dynamic graph structure learning method is used to obtain a set of adjacency matrices of the dynamic graph. Temporal convolution is used to process the long-term relationship to obtain the embedded representation of the graph nodes, and a dynamic graph of the foam flotation features is obtained.

7. The flotation process operating condition identification method based on dynamic time sequence graph neural network according to claim 6, characterized in that: The specific steps of step S3 are as follows: S31: Normalized apparent feature time series vector Z = {z1,z2,…,z N Abstraction into a dynamic graph: The foam flotation features are regarded as nodes of the dynamic graph, and the mutual influence relationship of the foam flotation features in the feature space is abstracted into an adjacency matrix; S32: Discretize the dynamic graph: Divide a set of feature time series vectors into multiple feature time series sub-vectors according to equidistant time slices, and construct a set of multi-time slice sub-graphs based on these feature time series sub-vectors to capture the dynamic relationship of the main feature variables over time and form a discretized dynamic graph; the discretized dynamic graph is composed of a four-tuple: G Dynamic ={V,Z,E,T}, where T={t1,t2,...,t s }, t S >0 is the time slice divided equally along the time axis, which is the minimum time unit for exploring the relationship between the main characteristic variables and time; t is the set of equally spaced time slices, s represents the total number of time slices; V = {v1, v2, ..., v N } is the set of graph nodes that remains unchanged across all time slices, N represents the number of foam features, and v N It is the Nth graph node consisting of the Nth type of foam features; Z={z1,z2,...,z s }, It is a set of feature vectors based on equidistant time slices, representing the initial embedding of the node; represents the normalized eigenvalue of the Nth foam feature at the sth time slice; is an edge set based on equidistant time slices, represents the time slice t i The edge feature set of the upper node reflects the dynamic changes of the node relationship; s It is a time slice divided equally along the time axis; S33: Determine the number of time slices s, and obtain the adjacency matrix representation of the multi-time slice subgraph through an adaptive dynamic graph structure learning method; S34: For the normalized apparent feature time series vector Z={z1,z2,…,z N }, and use the time convolution method to process the apparent feature time series vector Z={z1,z2,…,z N }, each node contains the long-term time-varying embedding vector Z′; S35: Match the embedded vector Z′ with the multi-time slice subgraph according to the time slice to obtain the foam flotation feature dynamic graph G Dynamic ={V,Z′,E,T}.

8. The flotation process operating condition identification method based on dynamic time sequence graph neural network according to claim 1, characterized in that: The specific steps of step S5 are as follows: S51: Using graph nodes as channels, we use a two-dimensional convolutional network to compress the dynamic graph G Dynamic The number of nodes of ′ is determined by N l Pooling is N l+1 , reducing the embedding dimension of graph representation; Among them, the pooling method is the Diffpool pooling method, which assigns nodes to different clusters through learning, so that the dynamic graph G Dynamic The ′ structure is compressed to generate a sparse graph representation, retaining important graph structure information: S l =softmax(GNN pool (A l ,x l )) S l is the allocation matrix, which represents the mapping from nodes to clusters. The pooled adjacency matrix and node feature matrix are: A l =S lT A l S l ; x l+1 =S lΤ x l ; Among them, A l is the adjacency matrix of the lth layer, S lT is the allocation matrix S l The transpose of x l and x l+1 Node feature information, GNN pool It is a function composed of the above message passing function and aggregation function; Each layer is composed of a message passing module of the message passing graph neural network MPNN that introduces cross-time feature accumulation and fusion and a DiffPool pooling module connected in sequence. After k layers of nesting, the final graph representation x is output. k and the adjacency matrix A k , as a sparse representation of the graph; S52: Use a fully connected network to map the sparse representation of the graph into the probability distribution of the working condition category; The sparse representation x k Input the fully connected layer to obtain the working condition classification results. The specific steps are as follows: For the sparse representation x k Perform global pooling to generate the global feature vector x of the graph G : x G =pool(x k ) pool(·) indicates global pooling; The global eigenvector x G Input the fully connected layer and obtain the probability distribution y of the working condition classification through the nonlinear transformation function: y=softmax(w·x G +b) in, is the weight matrix of the fully connected layer; is the bias vector; c is the number of working condition categories, is the probability distribution of the classification results; represents a real number, N k Indicates the number of graph nodes after k-layer graph pooling; According to the probability distribution y, the category with the highest probability is selected as the final working condition classification result class: class=argmax(y) argmax represents the maximum value function.

9. The method for identifying flotation process conditions based on a dynamic time series graph neural network according to claim 1, wherein: The cross entropy loss function is as follows: Among them, y is the probability distribution of working condition classification, class true is the one-hot encoding of the true label, and N is the number of samples.

Images