Foam flotation working condition identification method based on multi-branch reinforcement learning feature selection

Abstract

The invention is suitable for the technical field of data identification, and provides a froth flotation working condition identification method based on multi-branch reinforcement learning feature selection, and the method comprises the steps: extracting a plurality of froth image features from froth images of a plurality of froth flotation processes under different working conditions; clustering the plurality of foam image features based on a maximum information coefficient index to obtain a plurality of feature clusters; constructing a multi-branch neural network, and performing reinforcement learning on the multi-branch neural network based on the multiple foam image features to obtain an optimal feature subset; the multi-branch neural network comprises a sharing module and a plurality of branch modules; and identifying the working condition of the target froth flotation process needing working condition identification based on the optimal feature subset. The method can improve the recognition effect of the froth flotation working condition.

Description

Technical Field

[0001] The present application belongs to the field of data recognition technology, and in particular, relates to a method for identifying froth flotation conditions based on multi-branch reinforcement learning feature selection. Background Art

[0002] The froth flotation process is an important process in the nonferrous metallurgical industry. Its purpose is to separate the ideal minerals from the ore raw materials according to the different hydrophobicity of the ore particles. In the froth flotation process, the visual characteristics of the froth surface are closely related to the flotation production indicators, working conditions, and operating variables. However, since the froth flotation process includes multiple cascaded flotation cells, it has the characteristics of long process flow, complex reaction process, and difficult online monitoring. At present, production operations mainly rely on manual observation of the foam surface state. However, due to the strong subjectivity and arbitrariness of the manual method, it is difficult to maintain stability in the flotation production process, serious waste of mineral raw materials, and excessive consumption of flotation reagents.

[0003] In recent years, with the continuous development of industrial visual inspection technology and industrial big data technology, it has become possible to monitor the surface state of the foam in the flotation process online, and then realize data-driven working condition identification and comprehensive optimization control of multiple production operation parameters. The accurate description of the foam surface characteristics is the premise of flotation production process monitoring based on machine vision. At present, through the existing feature extraction technology, a series of characteristics such as the color, size, texture, and speed of the flotation process can be obtained, realizing the quantitative description of the visual characteristics of the foam surface and the objective evaluation of the flotation production state. However, due to the lack of knowledge of the corresponding mechanism, these manually extracted visual features are likely to contain some redundant information. These redundant, irrelevant, and even noisy features will not only increase the complexity of the working condition identification model and the difficulty of training, but may also reduce the final working condition identification effect. Summary of the invention

[0004] The embodiment of the present application provides a method for identifying froth flotation conditions based on multi-branch reinforcement learning feature selection, which can solve the problem of poor identification effect of froth flotation conditions.

[0005] The embodiment of the present application provides a method for identifying froth flotation conditions based on multi-branch reinforcement learning feature selection, including:

[0006] Step 1, for each froth flotation process in a plurality of froth flotation processes under different working conditions, extracting a plurality of froth image features from a froth image of the froth flotation process;

[0007] Step 2, clustering multiple foam image features based on the maximum information coefficient index to obtain multiple feature clusters;

[0008] Step 3, construct a multi-branch neural network, and perform reinforcement learning on the multi-branch neural network based on multiple foam image features to obtain an optimal feature subset; the optimal feature subset includes at least part of the multiple foam image features; the multi-branch neural network includes a shared module and multiple branch modules, the multiple branch modules correspond to multiple feature clusters one by one, the shared module is used to calculate the environmental state input into the multi-branch neural network and output an implicit representation, the environmental state is a binary code of a feature subset, the branch module is used to process the implicit representation output by the shared module and output the value of each action in the action space, each foam image feature in the feature cluster corresponding to the branch module and an empty action are all an action in the action space of the branch module;

[0009] Step 4: Identify the operating condition of the target froth flotation process that needs to be identified based on the optimal feature subset.

[0010] Optionally, step 2 includes:

[0011] Step 2.1, for each of the multiple foam image features, respectively, using the maximum information coefficient index to calculate the correlation between the foam image feature and the working condition label;

[0012] Step 2.2, taking the foam image features whose correlation with the working condition label is greater than a preset threshold as the cluster center of a feature cluster;

[0013] Step 2.3, for each other foam image feature among the multiple foam image features except the cluster center, calculate the maximum information coefficient between other foam image features and each cluster center, and assign other foam image features to the feature cluster corresponding to the maximum information coefficient with the largest value.

[0014] Optionally, the correlation between the foam image features and the working condition labels is calculated using the maximum information coefficient index, including:

[0015] The correlation MIC (f i ,y):

[0016]

[0017] Where n is the number of froth flotation processes, B(n) is the upper limit of the grid division a*b, MI(f i ,y) is the i-th foam image feature f i The maximum information coefficient between the working condition label y.

[0018] Optionally, reinforcement learning is performed on a multi-branch neural network based on multiple foam image features to obtain an optimal feature subset, including:

[0019] Step 3.1, using the constructed multi-branch neural network as an intelligent agent;

[0020] Step 3.2, controlling the intelligent agent to interact with the environment based on the multiple foam image features, and outputting a feature subset;

[0021] Step 3.3, based on the data used to train the multi-branch neural network during the interaction process, update the parameters of the multi-branch neural network, and use the multi-branch neural network after the parameter update as the intelligent agent in step 3.2, and return to execute step 3.2 until the number of interactions between the intelligent agent and the environment reaches the preset maximum number of interactions;

[0022] Step 3.4, the feature subset with the highest recognition accuracy among all feature subsets output by the interaction between the agent and the environment is taken as the optimal feature subset.

[0023] Optionally, step 3.2 includes:

[0024] Step 3.21, set the feature subset to an empty set, generate a binary code of the feature subset, and use the binary code as the environment state s t ;

[0025] Step 3.22, the environment state s t Input the agent for processing and obtain the value Q output by all branch modules j (s t ,a t j ), j = 1, 2, ..., M, M is the number of branch modules;

[0026] Step 3.23, for each branch module, take the action corresponding to the value with the largest value among the values ​​output by the branch module as the execution action

[0027] Step 3.24, add the foam image features corresponding to all the executed actions to the feature subset to obtain the updated feature subset, and use the binary code of the updated feature subset as the updated environment state s t+1 ;

[0028] Step 3.25, based on multiple foam image features, calculate the difference in recognition accuracy between the updated feature subset and the feature subset before the update, and use the difference as the reward signal r t+1 ;

[0029] Step 3.26, update the environment state s before t , all execution actions Reward signal r t+1 and the updated environment state s t+1Stored in memory as a piece of data for training a multi-branch neural network;

[0030] Step 3.27, update the environment state s t+1 As the environmental state s in step 3.22 t , and return to execute step 3.22 until the actions corresponding to the values ​​with the largest values ​​among the values ​​output by all branch modules are all empty actions, and output the feature subset when the actions corresponding to the values ​​with the largest values ​​among the values ​​output by all branch modules are all empty actions.

[0031] Optionally, updating parameters of the multi-branch neural network based on the data used to train the multi-branch neural network during the interaction process includes:

[0032] Step 3.31, extract b pieces of data from the data stored in the memory, denoted as B = {s t ,a t ,s t+1 ,r t+1} b ;

[0033] Step 3.32, by formula Q j eval =Q j (s t ,a t j ) Calculate the estimated Q value Q of the j-th branch module j eval , and using the formula Q j target =r t+1 +γ*max a Q j (s t+1 ,a) Calculate the target Q value Q of the j-th branch module j target , calculate the loss function of this branch Among them, γ is the attenuation factor, a is the variable of the max function, (s t ,a t ,s t+1 ,r t+1 )~B;

[0034] Step 3.33, using the formula Calculate the loss value loss of the multi-branch neural network, and update the parameters of the multi-branch neural network based on the loss value loss.

[0035] Optionally, both the shared module and the branch module are double-layer fully connected layers.

[0036] The above solution of the present application has the following beneficial effects:

[0037] In an embodiment of the present application, a multi-branch neural network with a shared module is established as a decision network for feature selection, wherein different branch modules are responsible for selecting representative features from their corresponding feature clusters. At the same time, the present application constructs a reinforcement learning environment based on a foam flotation process of known working conditions, and performs reinforcement learning training on the constructed multi-branch neural network to obtain the most important features for foam flotation condition identification and prediction, thereby improving the foam flotation condition identification effect when the foam flotation condition is identified based on these important features.

[0038] Other beneficial effects of the present application will be described in detail in the subsequent specific implementation section. BRIEF DESCRIPTION OF THE DRAWINGS

[0039] In order to more clearly illustrate the technical solutions in the embodiments of the present application, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present application. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative work.

[0040] Figure 1 A flowchart of a method for identifying froth flotation conditions based on multi-branch reinforcement learning feature selection provided in one embodiment of the present application;

[0041] Figure 2 A schematic diagram of the structure of a multi-branch neural network provided in one embodiment of the present application;

[0042] Figure 3 This is a performance evaluation result diagram of different feature selection methods in an example of this application. DETAILED DESCRIPTION

[0043] In the following description, specific details such as specific system structures, technologies, etc. are provided for the purpose of illustration rather than limitation, so as to provide a thorough understanding of the embodiments of the present application. However, it should be clear to those skilled in the art that the present application may also be implemented in other embodiments without these specific details. In other cases, detailed descriptions of well-known systems, devices, circuits, and methods are omitted to prevent unnecessary details from obstructing the description of the present application.

[0044] It should be understood that when used in the present specification and the appended claims, the term "comprising" indicates the presence of described features, integers, steps, operations, elements and / or components, but does not exclude the presence or addition of one or more other features, integers, steps, operations, elements, components and / or combinations thereof.

[0045] It should also be understood that the term “and / or” used in the specification and appended claims refers to any and all possible combinations of one or more of the associated listed items, and includes these combinations.

[0046] As used in the specification and appended claims of this application, the term "if" can be interpreted as "when" or "uponce" or "in response to determining" or "in response to detecting", depending on the context. Similarly, the phrase "if it is determined" or "if [described condition or event] is detected" can be interpreted as meaning "uponce it is determined" or "in response to determining" or "uponce [described condition or event] is detected" or "in response to detecting [described condition or event]", depending on the context.

[0047] In addition, in the description of the present application specification and the appended claims, the terms "first", "second", "third", etc. are only used to distinguish the descriptions and cannot be understood as indicating or implying relative importance.

[0048] References to "one embodiment" or "some embodiments" etc. described in the specification of this application mean that one or more embodiments of the present application include specific features, structures or characteristics described in conjunction with the embodiment. Therefore, the statements "in one embodiment", "in some embodiments", "in some other embodiments", "in some other embodiments", etc. that appear in different places in this specification do not necessarily refer to the same embodiment, but mean "one or more but not all embodiments", unless otherwise specifically emphasized in other ways. The terms "including", "comprising", "having" and their variations all mean "including but not limited to", unless otherwise specifically emphasized in other ways.

[0049] In response to the problem of poor recognition effect of foam flotation conditions at present, an embodiment of the present application provides a foam flotation condition recognition method based on multi-branch reinforcement learning feature selection. The method establishes a multi-branch neural network with a shared module as a decision network for feature selection, wherein different branch modules are responsible for selecting representative features from their corresponding feature clusters. At the same time, the present application constructs a reinforcement learning environment based on the foam flotation process of known conditions, and performs reinforcement learning training on the constructed multi-branch neural network to obtain the most important features for foam flotation condition recognition and prediction, thereby improving the foam flotation condition recognition effect when foam flotation condition recognition is performed based on these important features.

[0050] To facilitate understanding of the froth flotation condition identification method provided in the present application, the relevant content of feature selection is first introduced here.

[0051] Feature selection is an important data analysis technique and a necessary data preprocessing step in machine learning. By performing feature selection on the original data feature set, a large number of interfering features that are irrelevant or redundant to the target variable can be removed, thereby effectively reducing the data dimension, improving the accuracy of the target variable prediction, and reducing the subsequent processing cost of the original data set. At the same time, it is easier for researchers to understand and analyze, thereby better revealing the useful information hidden in the original data.

[0052] The essence of feature selection is to use a certain search scheme to find the best feature subset from the original feature space. This process is mainly affected by the search strategy. Common search strategies can be divided into three categories: complete search, random search, and sequential search. Complete search searches all possible feature subsets in the original feature space. In theory, the complete search strategy can ensure that the feature subset with the strongest discrimination ability is found. However, when the number of features is too large, the complete search will incur a huge computational cost. Random search starts from a randomly generated candidate feature subset and gradually approaches the global optimal feature subset using a certain intelligent optimization algorithm. Commonly used search methods include genetic algorithms and particle swarm optimization algorithms. Although intelligent optimization algorithms show more outstanding performance than other search strategies in feature selection problems, they often face the challenges of local optimal stagnation and high computational cost when dealing with large-scale feature selection problems. Sequential search adds or removes the number of candidate features each time based on well-founded heuristic rules. Compared with complete search and random search, sequential search can effectively reduce the computational complexity of the feature selection process, but there is a risk of falling into local optimality. At the same time, this method requires a fitness function with monotonic properties and requires strong professional background knowledge.

[0053] Feature selection is the optimal selection of decision variables in a discrete feature decision space. It has natural similarities with the "action selection" of reinforcement learning. The characteristics of deep reinforcement learning "offline training, online decision-making" make it possible to solve the feature selection problem in real time online. Therefore, using reinforcement learning to solve the traditional feature selection problem is a good choice. In reinforcement learning, there are two objects that can interact: the agent and the environment. The agent selects the appropriate action based on the state of environmental feedback. The environment will execute the action selected by the agent and generate a reward signal based on the result of the action. The agent optimizes its own strategy to maximize the total reward. Compared with traditional feature selection algorithms, feature selection algorithms based on reinforcement learning have a series of advantages such as fast solution speed and strong generalization ability. However, the feature selection algorithm based on reinforcement learning currently ignores the implicit prior information of the data set when constructing the decision network, so there are problems such as complex decision process, long decision process, and difficulty in training the decision network.

[0054] The following is an illustrative description of the froth flotation condition identification method based on multi-branch reinforcement learning feature selection provided in the present application in conjunction with specific embodiments.

[0055] like Figure 1 As shown, the froth flotation condition identification method based on multi-branch reinforcement learning feature selection provided in the embodiment of the present application includes the following steps:

[0056] Step 1: for each froth flotation process in a plurality of froth flotation processes under different working conditions, extract a plurality of froth image features from a froth image of the froth flotation process.

[0057] The above foam image can be acquired by a visual acquisition device. After the foam image is acquired, the foam surface visual features (i.e., foam image features) can be extracted from the foam image by feature extraction technology to form raw data. It should be noted that, in order to facilitate subsequent model training, the raw data also records the working condition labels of each foam flotation process, such as good, poor, bad, etc.

[0058] Specifically, multiple foam image features may include flotation foam color features (such as grayscale mean, red mean, etc.), flotation foam dynamic features (such as foam flow rate, etc.), flotation foam size features (such as foam size mean, aspect ratio, etc.), flotation foam statistical related features (such as foam size mean, variance, skewness, etc.), and flotation foam texture features (such as foam coarseness, etc.).

[0059] Because the industrial field environment of data collection is complex, such data inevitably contains some missing values ​​and outliers. Therefore, the collected data is preprocessed for the missing values ​​and outliers in the data. First, the invalid samples and feature variables with duplicate information and too many missing values ​​or outliers are deleted. Select appropriate standards to fill in a small number of missing values ​​in the data. Generally, the mean or median is filled for continuous feature variables, and the mode is filled for categorical feature variables. In addition, in order to facilitate the training of subsequent models, the continuous numerical features are normalized and the categorical features are encoded.

[0060] Step 2: Cluster multiple foam image features based on the maximum information coefficient index to obtain multiple feature clusters.

[0061] In some embodiments of the present application, the specific implementation of step 2 above includes the following steps:

[0062] Step 2.1, for each of the multiple foam image features, respectively, calculate the correlation between the foam image feature and the working condition label using the maximum information coefficient index.

[0063] Specifically, the correlation MIC (f) between the i-th foam image feature and the working condition label can be calculated by the following formula: i ,y):

[0064]

[0065] Among them, the number of foam image features is 38, n is the number of foam flotation processes, B(n) is the upper limit of the grid division a*b, MI(f i ,y) is the i-th foam image feature f i The maximum information coefficient between the foam image feature f and the working condition label y. It should be noted that when calculating the maximum information coefficient, the foam image feature f i and the operating condition label y are considered as a random variable.

[0066] Step 2.2, taking the foam image features whose correlation with the working condition label is greater than a preset threshold as the clustering center of a feature cluster.

[0067] In some embodiments of the present application, the above preset threshold can be 0.1*MIC max , where MIC max It represents the maximum value of the maximum information coefficient between all features and working condition labels.

[0068] Step 2.3, for each other foam image feature among the multiple foam image features except the cluster center, calculate the maximum information coefficient between other foam image features and each cluster center, and assign other foam image features to the feature cluster corresponding to the maximum information coefficient with the largest value.

[0069] That is, for the remaining foam image features in the data set, the maximum information coefficient between them and each cluster center feature is calculated, and they are assigned to the feature cluster with the largest maximum information coefficient value.

[0070] Step 3: construct a multi-branch neural network, and perform reinforcement learning on the multi-branch neural network based on multiple foam image features to obtain the optimal feature subset.

[0071] The above reinforcement learning is to select the most important features for the recognition and prediction of froth flotation conditions. Specifically, the optimal feature subset includes at least part of the multiple froth image features (i.e., the most important features), such as flotation froth color features, flotation froth dynamic features, flotation froth texture features, etc.

[0072] like Figure 2As shown, the multi-branch neural network includes a shared module (i.e., a shared layer in the figure) and multiple branch modules (i.e., branches 1, 2 to M in the figure, where M is the number of branch modules). The number of branch modules is consistent with the number of feature clusters, and multiple branch modules correspond to multiple feature clusters one by one. The input layer in the figure is the environment state input to the multi-branch neural network, and the output layer is the value of the action output by the branch module.

[0073] The shared module consists of a double-layer fully connected layer. The shared module is used to calculate the environment state input into the multi-branch neural network and output an implicit representation. The environment state is a binary code of a feature subset (initial empty set). Specifically, an n-bit binary code is used to represent the environment state, where n is equal to the number of foam image features in the original data. A bit coded as 1 indicates that the corresponding foam image feature is included in the feature subset, and a bit coded as 0 indicates that it is not included in the feature subset. When an environment state is provided as input, the shared module calculates an implicit representation.

[0074] Each branch module is also composed of a double-layer fully connected layer. For each branch module: the branch module is used to process the implicit representation output by the shared module and output the value of each action in the action space (i.e., the action space of the molecular module). Each foam image feature and an empty action in the feature cluster corresponding to the branch module are an action in the action space of the branch module.

[0075] That is, each branch module takes the implicit representation output by the shared module as input and outputs the value of each action dimension. For each branch module, its action space includes all foam image features in its corresponding feature cluster and an empty action, where the empty action means not selecting any foam image features in the feature cluster. Then, these decomposed action values ​​are used to generate joint action tuples.

[0076] As a preferred example, the input dimension of the shared module is 38, which is consistent with the number of foam image features; the number of neurons in the first fully connected layer is 128, and the number of neurons in the second fully connected layer is 64. Each branch module takes the implicit representation output by the shared module as input to output the value of each action dimension. In this embodiment, the input dimension of each branch module is 64; the number of neurons in the first fully connected layer is 32, and the neurons in the second fully connected layer are consistent with the action dimension. For each branch module, its action space includes all foam image features in its corresponding feature cluster and an empty action, where an empty action means that no foam image feature in the feature cluster is selected.

[0077] Step 4: Identify the operating condition of the target froth flotation process that needs to be identified based on the optimal feature subset.

[0078] During actual identification, the foam image of the target foam flotation process can be collected first, and then the values ​​of each foam image feature contained in the optimal feature subset can be extracted from the foam image. The values ​​of these foam image features are then input into a working condition identification model (such as a support vector machine or a neural network) for identification processing to obtain the working condition identification result of the target foam flotation process. The working condition identification result can be: the working condition of the target foam flotation process is good / poor / bad.

[0079] It is worth mentioning that the present application establishes a multi-branch neural network with a shared module as a decision network for feature selection, in which different branch modules are responsible for selecting representative features from their corresponding feature clusters. At the same time, the present application constructs a reinforcement learning environment based on the foam flotation process of known working conditions, and performs reinforcement learning training on the constructed multi-branch neural network to obtain the most important features for the identification and prediction of foam flotation conditions, so that when the foam flotation conditions are identified based on these important features, the foam flotation condition identification effect can be improved.

[0080] The following is an exemplary description of the reinforcement learning process in step 3 above.

[0081] In some embodiments of the present application, the specific implementation method of performing reinforcement learning on a multi-branch neural network based on multiple foam image features to obtain an optimal feature subset in the above step 3 includes the following steps:

[0082] Step 3.1, use the constructed multi-branch neural network as an intelligent agent.

[0083] Step 3.2, based on multiple foam image features, the intelligent agent is controlled to interact with the environment and a feature subset is output.

[0084] In some embodiments of the present application, the specific implementation of step 3.2 above is as follows:

[0085] Step 3.21, set the feature subset to an empty set, generate a binary code of the feature subset, and use the binary code as the environment state s t .

[0086] Step 3.22, the environment state s t Input the agent for processing and obtain the value Q output by all branch modules j (s t ,a t j ), j = 1, 2, ..., M, M is the number of branch modules; Q j (s t ,a t j ) is the value output by the j-th branch module.

[0087] Step 3.23, for each branch module, take the action corresponding to the value with the largest value among the values ​​output by the branch module as the execution action is the execution action of the j-th branch module.

[0088] Step 3.24, add the foam image features corresponding to all the executed actions to the feature subset to obtain the updated feature subset, and use the binary code of the updated feature subset as the updated environment state s t+1 .

[0089] Step 3.25, based on multiple foam image features, calculate the difference in recognition accuracy between the updated feature subset and the feature subset before the update, and use the difference as the reward signal r t+1 . Specifically, for each foam flotation process of the original data, the values ​​of each foam image feature in the updated feature subset can be retrieved from the original data (i.e., the multiple foam image features obtained in the previous article), and then the values ​​of these foam image features are input into the working condition recognition model for recognition processing to obtain the working condition recognition result of the foam flotation process, and the recognition accuracy is obtained based on the degree of deviation between the working condition recognition result and the working condition label, and finally the average recognition accuracy of each foam flotation process is used as the recognition accuracy of the updated feature subset. Similarly, the recognition accuracy of the feature subset before the update can also be obtained in the same way. It should be noted that if the feature subset before the update is an empty set, the recognition accuracy of the feature subset before the update is 0.

[0090] Step 3.26, update the environment state s before t , all execution actions Reward signal r t+1 and the updated environment state s t+1 As a piece of data for training a multi-branch neural network (s t , r t+1 ,s t+1 ) is stored in memory.

[0091] Step 3.27, update the environment state s t+1 As the environmental state s in step 3.22 t , and return to execute step 3.22 until the actions corresponding to the values ​​with the largest values ​​among the values ​​output by all branch modules are all empty actions, and output the feature subset when the actions corresponding to the values ​​with the largest values ​​among the values ​​output by all branch modules are all empty actions.

[0092] Step 3.3, update the parameters of the multi-branch neural network based on the data used to train the multi-branch neural network during the interaction process, and use the multi-branch neural network after parameter update as the intelligent agent in step 3.2, and return to execute step 3.2 until the number of interactions between the intelligent agent and the environment reaches the preset maximum number of interactions.

[0093] In some embodiments of the present application, the process of updating the parameters of the multi-branch neural network based on the data used to train the multi-branch neural network during the interaction process is as follows:

[0094] Step 3.31, extract b pieces of data from the data stored in the memory, denoted as B = {s t ,a t ,s t+1 ,r t+1} b ; The value of b can be set according to actual conditions.

[0095] Step 3.32, by formula Q j eval =Q j (s t ,a t j ) Calculate the estimated Q value Q of the j-th branch module j eval , and using the formula Q j target =r t+1 +γ*max a Q j (s t+1 ,a) Calculate the target Q value Q of the jth branch module j target , calculate the loss function of this branch Among them, γ is the attenuation factor, a is the variable of the max function, (s t ,a t ,s t+1 ,r t+1 )~B.

[0096] Step 3.33, using the formula Calculate the loss value loss of the multi-branch neural network, and update the parameters of the multi-branch neural network based on the loss value loss. That is, take b pieces of data, calculate a loss value loss for each piece of data, and the final loss function is the sum of all losses. Specifically, the network parameters can be updated through back propagation and using the gradient descent algorithm.

[0097] Step 3.4, the feature subset with the highest recognition accuracy among all feature subsets output by the interaction between the agent and the environment is taken as the optimal feature subset. That is, the feature subset with the highest recognition accuracy among all feature subsets output by the interaction process is taken as the optimal feature subset.

[0098] The following is an illustrative example of the selection process of the optimal feature subset.

[0099] In this example, the froth flotation process is a gold-antimony flotation process. Accordingly, the feature selection process is as follows:

[0100] S1. Froth flotation data collection:

[0101] By collecting foam image data collected from visual acquisition equipment during a gold-antimony flotation process, feature extraction technology is used to extract the visual features of the foam surface and collect the original data. The data includes flotation foam color features, such as grayscale mean, red mean, etc.; flotation foam dynamic features, such as foam flow rate, etc.; flotation foam size features, such as foam size mean, aspect ratio, etc.; flotation foam statistical related features, such as foam size mean, variance, skewness, etc.; flotation foam texture features, such as foam coarseness, etc. This example finally collected a total of 6,000 data samples under 8 different working conditions, including a total of 38 features.

[0102] S2. Data preprocessing:

[0103] In the data collection process of S1, according to the needs of subsequent model training, the continuous numerical features are normalized and the categorical features are encoded. In this example, MinMaxScalar is used to normalize the continuous features, and the categorical features with size relationships are encoded with natural numbers, and the categorical features without size relationships are encoded with One-Hot. The purpose of this step is to construct data that is convenient for model training. Under the premise of achieving the same purpose, other reasonable data preprocessing methods not specifically mentioned in this step of the present application can also be adopted.

[0104] S3. Feature clustering:

[0105] This application first uses the maximum information coefficient (MIC) indicator based on information entropy theory to calculate the correlation between each feature and the label (i.e., the working condition label) in the data set, and sorts them from large to small according to the correlation. The calculation formula of MIC is as follows:

[0106]

[0107] In the formula, n is the data sample size, B(n) is the upper limit of the grid division a*b, and MI(f,y) represents the maximum information coefficient between feature f and working condition label y. At the same time, a threshold is set, and the features with correlation greater than the threshold are set as cluster centers. In this example, the threshold is 0.1*MIC max , where MIC max Indicates the maximum value of the maximum information coefficient between all features and working condition labels. For the remaining features in the data set, calculate the maximum information coefficient between them and each cluster center feature, and assign them to the cluster center with the largest maximum information coefficient value.

[0108] S4. Build a multi-branch neural network architecture:

[0109] In this embodiment, the multi-branch neural network architecture is as follows Figure 2 As shown. The multi-branch decision neural network architecture includes a shared module and several branch modules, where the number of branch modules is consistent with the number of feature clusters. The shared decision module is composed of a neural network consisting of a double-layer fully connected layer. In this embodiment, the input dimension of the shared decision layer is 38, which is consistent with the number of features; the number of neurons in the first fully connected layer is 128, and the number of neurons in the second fully connected layer is 64. Each branch module is also composed of a neural network consisting of a double-layer fully connected layer. Each branch module takes the implicit representation output by the shared decision module as input to output the value of each action dimension. In this embodiment, the input dimension of each branch module is 64; the number of neurons in the first fully connected layer is 32, and the neurons in the output layer are consistent with the action dimension. For each branch module, its action space includes all features in its corresponding feature cluster and an empty action, where an empty action means that no feature in the feature cluster is selected.

[0110] S5. Reinforcement learning feature selection method:

[0111] In this embodiment, a 38-bit binary code is used as the environment state of reinforcement learning, which means the features included in the currently selected feature subset, which is an empty set initially. A bit coded as 1 represents that the corresponding feature is included in the feature subset, and 0 represents that it is not included in the feature subset. The reinforcement learning feature selection process is as follows: 1) Initialize the neural network parameters and set the maximum number of interactions. 2) The agent interacts with the environment. 3) Neural network parameter training. 4) Repeat 2) to 4) until the maximum number of interactions is reached. 5) The feature subset with the highest classification accuracy during the interaction process is used as the final output result.

[0112] The process of interaction between the agent and the environment is as follows: 1) Set the current feature subset to an empty set and generate the corresponding binary code as the initial state of the environment s t ; 2) The environmental state s tInput into the multi-branch neural network to obtain the action value output Q of all branch networks j (s t ,a t j ), j = 1, 2, ..., M, where M represents the number of branches; 3) Each branch module takes the action with the greatest value as the execution action Add the corresponding features to the feature subset and update the environment state s t+1 ; 4) Use the difference in classification accuracy between the feature subsets before and after the update as the reward signal r t+1 In this example, support vector machine (SVM) is used as the machine learning machine for classification; 5) s t , r t+1 and t+1 The data is stored in the memory and used to update the neural network parameters. 6) Repeat the process from 2) to 5) until the termination condition is met and the feature subset is output, where the termination condition is that all branches select empty actions.

[0113] The neural network parameter training process is as follows: 1) Randomly extract b pieces of data from the memory, denoted as B = {s t ,a t ,s t+1 ,r t+1} b . 2) Use formula (1) to calculate the estimated Q value of each branch module, and use formula (2) to calculate the target Q value.

[0114] Q j eval =Q j (s t ,a t j ) (1)

[0115] Q j target =r t+1 +γ*max a Q j (s t+1 ,a) (2)

[0116] Where γ in the formula is the attenuation factor. 3) Use formulas (3) and (4) to construct the loss function. 4) Perform back propagation on the reinforcement learning network and use the gradient descent algorithm to update the network parameters. The parameters of reinforcement learning are shown in Table 1.

[0117]

[0118] Table 1 Parameters of reinforcement learning

[0119] Hyperparameters meaning Numeric Lr Gradient descent learning rate 0.001 Epoch Number of interaction rounds 1000 γ Attenuation Factor 0.95 pool_size Amount of data in memory 1024 b Update sampling data volume at one time 32

[0120] S6. Feature selection result evaluation:

[0121] This example is based on the feature selection results obtained after step S5, and uses common feature selection methods including ReliefF, Fisher Score, mRMR, LASSO and Recursive Feature Elimination (RFE) for comparison. In view of the problem background of froth flotation condition identification, the performance of the above methods is evaluated by using 5-fold cross validation as an evaluation strategy based on AUC and Accuracy as evaluation indicators. The performance evaluation results are shown in Figure 2. Figure 3 As shown, Figure 3 The algorithm mentioned in is the feature selection method of this application. It can be seen that the prediction result of the algorithm proposed in this application is better than that of the comparison method. Users can select different evaluation indicators and evaluation strategies to evaluate the performance of the model according to the characteristics of the problem.

[0122] In summary, the feature selection process of this application has the following advantages:

[0123] First, by making full use of the visual features and prediction labels (i.e., working condition labels) in the froth flotation data set, the correlation between features, and clustering the features through the maximum mutual information coefficient, similar features are clustered into a feature cluster, and representative features are selected from each feature cluster to form a feature subset. Compared with the existing froth flotation feature selection algorithm that makes decisions from the overall feature space, the feature selection of this application is more efficient.

[0124] Second, use deep reinforcement learning algorithms for foam flotation feature selection: reinforcement learning does not rely entirely on explicit supervision information. By using appropriately defined reward signals, an agent can learn from its own experience how to find feature subsets related to foam flotation through trial and error, and can dynamically interact with the environment to automatically improve and optimize feature selection strategies. Compared with existing foam flotation feature selection algorithms, it has stronger global search capabilities and can generate feature subsets with better performance.

[0125] Third, a multi-branch neural network is constructed as a decision network for deep reinforcement learning: each branch module shares decision information through a shared module. This multi-branch neural network can select multiple features in one decision. Compared with the existing feature selection algorithm based on reinforcement learning, it has high decision efficiency and fast decision speed, so the computational complexity of neural network training is low.

[0126] The above is a preferred embodiment of the present application. It should be pointed out that for ordinary technicians in this technical field, several improvements and modifications can be made without departing from the principles described in the present application. These improvements and modifications should also be regarded as the scope of protection of the present application.

Claims

1. A method for identifying froth flotation conditions based on multi-branch reinforcement learning feature selection, characterized in that: include: Step 1, for each froth flotation process in a plurality of froth flotation processes under different working conditions, extracting a plurality of froth image features from a froth image of the froth flotation process; Step 2, clustering the multiple foam image features based on the maximum information coefficient index to obtain multiple feature clusters; Step 3, constructing a multi-branch neural network, and performing reinforcement learning on the multi-branch neural network based on the multiple foam image features to obtain an optimal feature subset; the optimal feature subset includes at least part of the multiple foam image features; the multi-branch neural network includes a shared module and multiple branch modules, the multiple branch modules correspond to the multiple feature clusters one by one, the shared module is used to calculate the environmental state input to the multi-branch neural network and output an implicit representation, the environmental state is a binary code of a feature subset, the branch module is used to process the implicit representation output by the shared module and output the value of each action in the action space, each foam image feature in the feature cluster corresponding to the branch module and an empty action are all an action in the action space of the branch module; Step 4: identifying the operating condition of the target froth flotation process requiring operating condition identification based on the optimal feature subset.

2. The method for identifying froth flotation conditions according to claim 1, characterized in that: The step 2 comprises: Step 2.1, for each of the plurality of foam image features, respectively, calculating the correlation between the foam image feature and the working condition label using a maximum information coefficient index; Step 2.2, taking the foam image features whose correlation with the working condition label is greater than a preset threshold as the cluster center of a feature cluster; Step 2.3, for each other foam image feature among the multiple foam image features except the cluster center, calculate the maximum information coefficient between the other foam image features and each cluster center, and assign the other foam image features to the feature cluster corresponding to the maximum information coefficient with the largest value.

3. The method for identifying froth flotation conditions according to claim 2, characterized in that: The method of calculating the correlation between the foam image features and the working condition labels respectively by using the maximum information coefficient index includes: The formula is as follows to calculate the i The correlation MIC (f i , y): Where n is the number of froth flotation processes, B(n) is the upper limit of the grid division a*b, MI(f i , y) is the i-th foam image feature f i The maximum information coefficient between the working condition label y.

4. The method for identifying froth flotation conditions according to claim 1, characterized in that: The step of performing reinforcement learning on the multi-branch neural network based on the multiple foam image features to obtain an optimal feature subset includes: Step 3.1, using the constructed multi-branch neural network as an intelligent agent; Step 3.2, controlling the intelligent agent to interact with the environment based on the multiple foam image features, and outputting a feature subset; Step 3.3, updating the parameters of the multi-branch neural network based on the data used to train the multi-branch neural network during the interaction process, and using the multi-branch neural network after the parameter update as the agent in step 3.2, and returning to execute step 3.2 until the number of interactions between the agent and the environment reaches a preset maximum number of interactions; Step 3.4, the feature subset with the highest recognition accuracy among all feature subsets output by the interaction between the agent and the environment is taken as the optimal feature subset.

5. The method for identifying froth flotation conditions according to claim 4, characterized in that: The step 3.2 comprises: Step 3.21, set the feature subset to an empty set, generate a binary code of the feature subset, and use the binary code as the environment state s t ; Step 3.22, the environment state s t Input the agent for processing and obtain the value Q output by all branch modules j (s t , a t j ), j = 1, 2, ..., M, M is the number of branch modules; Step 3.23, for each branch module, the action corresponding to the value with the largest value among the values ​​output by the branch module is taken as the execution action Step 3.24, add the foam image features corresponding to all the executed actions to the feature subset to obtain the updated feature subset, and use the binary code of the updated feature subset as the updated environment state s t+1 ; Step 3.25, based on the multiple foam image features, calculate the difference in recognition accuracy between the updated feature subset and the feature subset before the update, and use the difference as the reward signal r t+1 ; Step 3.26, update the environment state s before t , all execution actions Reward signal r t+1 and the updated environment state s t+1 storing in the memory as a piece of data for training the multi-branch neural network; Step 3.27, update the environment state s t+1 As the environmental state s in step 3.22 t , and return to execute step 3.22 until the actions corresponding to the values ​​with the largest values ​​among the values ​​output by all branch modules are all empty actions, and output the feature subset when the actions corresponding to the values ​​with the largest values ​​among the values ​​output by all branch modules are all empty actions.

6. The method for identifying froth flotation conditions according to claim 5, characterized in that: The updating of the parameters of the multi-branch neural network based on the data used to train the multi-branch neural network during the interaction process includes: Step 3.31, extract b pieces of data from the data stored in the memory, denoted as B = {s t , a t ,s t+1 , r t+1 } b ; Step 3.32, by formula Q j eval =Q j (s t , a t j ) Calculate the estimated Q value Q of the j-th branch module j eval , and using the formula Q j target =r t+1 +γ*max a Q j (s t+1 , a) Calculate the target Q value Q of the j-th branch module j target , calculate the loss function of this branch Among them, γ is the attenuation factor, a is the variable of the max function, (s t , a t ,s t+1 , r t+1 )~B; Step 3.33, using the formula Calculate the loss value loss of the multi-branch neural network, and update the parameters of the multi-branch neural network based on the loss value loss.

7. The method for identifying froth flotation conditions according to claim 1, characterized in that: The shared module and the branch module are both double-layer fully connected layers.

Images