A method for constructing a flotation key indicator monitoring model based on data fusion

Abstract

A kind of flotation key index monitoring model construction method based on data fusion, the present application proposes a kind of flotation key performance index monitoring model construction method based on foam video data and strong semantic data fusion, foam video is collected based on the digital image acquisition system set in flotation field, first, extract deep feature information and strong semantic data information from foam video and industrial field, then the information of two is effectively fused, the coupling relationship between the multi-source heterogeneous data with different data representation is analyzed, and the flotation key performance index monitoring model is constructed.The present application solves the problems such as weak semantic information of deep feature vector of foam video obtained by traditional convolutional neural network method, weak generalization performance, model training is prone to overfitting and other problems, can more accurately judge working condition and effectively guide field operation.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of froth flotation, and particularly relates to a method for constructing a froth flotation key performance index monitoring model in a zinc flotation process. BACKGROUND

[0002] Froth flotation is one of the most commonly used separation methods in the ore dressing process. Froth flotation realizes the separation of valuable mineral particles and waste rock or other non-valuable minerals by utilizing the differences in the physical and chemical properties (mainly hydrophobicity) of the mineral surface, has the advantages of strong flexibility, wide application range, good ore dressing effect, etc., and can help improve the recovery rate and metal grade of the ore. However, due to the extremely complex physical and chemical processes involved in the flotation process, the flotation process has the characteristics of long process flow, unclear internal mechanism, many external interference factors, and serious nonlinearity. Although a large number of studies have been conducted by domestic and foreign scholars in recent years, so far there is no effective and applicable mathematical model that can describe the flotation process, and the analysis and modeling of the flotation process are still a difficult problem to be solved. For a long time, the flotation process mainly relies on manual visual observation of the froth state on the surface of the flotation tank to complete the on-site operation. This production method is heavily dependent on the experience and knowledge of workers, is highly subjective, and has low accuracy, resulting in frequent fluctuations in flotation production indicators, serious loss of mineral raw materials, large consumption of reagents, low resource recovery rate, and the like. Froth video is an indicator of the state and changes of the froth on the surface of the flotation tank, and is the main basis for monitoring the key performance indicators of flotation. However, the froth video has a long duration, contains rich information, and has large data redundancy. The semantic information of the froth video deep feature vector obtained by using the traditional convolutional neural network method is weak, the generalization performance is not strong, and the model is prone to overfitting during training. Strong semantic data (such as temperature, pressure, flow, etc. in the field) is a data set with clear physical and knowledge meaning in the industrial process, which is complementary to weak semantic data such as froth video, and can be obtained by industrial field sensors or flotation field professional knowledge. In view of the fact that the traditional deep learning does not consider the information supplementing effect of the strong semantic data of the industrial process on the monitoring of the key performance indicators of flotation, a new method for constructing a flotation key performance index monitoring model based on the fusion of froth video and strong semantic data is proposed. SUMMARY

[0003] The purpose of the present application is to provide a flotation key performance indicator monitoring model construction method based on foam video and strong semantic data fusion. In the flotation industry process, foam video is an important data reflecting the current flotation working condition and is the basis for flotation key performance indicator monitoring. In view of the long duration of the foam video, the rich information contained, the large data redundancy, the weak semantic information of the foam video deep feature vector obtained by using the traditional convolutional neural network method, the weak generalization performance, and the easy overfitting of the model during training, a new flotation key performance indicator monitoring model construction method based on foam video data and strong semantic data fusion is proposed. The method first extracts deep feature information and strong semantic data information from foam video and industrial field, then effectively fuses the two information, analyzes the coupling relationship between multi-source heterogeneous data with different data representation, constructs the flotation key performance indicator monitoring model, and improves the representation ability and generalization ability of the model. The technical scheme steps are as follows:

[0004] Step one: collect zinc flotation foam video data by using the image acquisition system of the flotation field, pretreat the collected data, and constitute a zinc flotation data set;

[0005] Step two: construct a foam video deep feature extraction subnetwork and a foam video strong semantic feature extraction subnetwork; S1: construct a foam video deep feature extraction subnetwork based on a 3D convolutional network and vector compression;

[0006] Extract foam video image frames from the foam video and perform normalization processing to obtain a three-dimensional feature input graph V0 composed of original video image frame sequences; adopt a 3D convolution module to obtain a deep feature representation vector of the foam video; denote F k ,k=1…M as the feature map processed by the 3D convolution module 3DConv_k,k=1...M, and each layer feature map is defined as follows:

[0007]

[0008] Wherein, M is the total number of 3D convolution modules;

[0009] Denote M1 as the strong semantic feature map layer threshold, and F k ,M1≤k≤M is processed by tensor flattening and vector compression to obtain a foam video high-level deep feature representation vector f k , which is defined as follows:

[0010]

[0011] Wherein, VC(·) represents the tensor flattening and vector compression operation;

[0012] S2: construct a foam video strong semantic feature extraction subnetwork based on LSTM;

[0013] In the froth flotation, the industrial process data is collected by sensors in the industrial field; the froth appearance features are obtained by the long-term accumulated experience knowledge of the flotation operators and the professional knowledge in the flotation field; P is the total number of the industrial process data elements, S is the number of the froth appearance features, and the industrial process data vector matched with the froth video image frame is obtained froth appearance feature vector Then the strong semantic feature vector X of the froth video image frame is (t) As shown in the following formula:

[0014]

[0015] Wherein, concat(·) represents the vector splicing operation;

[0016] The strong semantic feature vector is input into the LSTM network in the front and back order of the video frame, and the strong semantic feature vector l of the froth video is obtained q As shown in the following formula:

[0017] l q = LSTM(X (1) ,X (2) ,...,X (Q) )

[0018] Wherein, LSTM(·) represents the long short-term memory network, and Q represents the total number of frames of the current froth video;

[0019] Step three: fusion of video deep features and strong semantic features to solve the flotation key performance indicators;

[0020] S1: network training is performed on the froth video deep feature extraction network based on the 3D convolution network and vector compression, and the pre-training weight of the network is obtained;

[0021] S2: network training is performed on the froth video strong semantic feature extraction network based on the LSTM, and the pre-training weight of the network is obtained;

[0022] S3: the pre-training weights of the above two networks are taken as the initial values, and the network model is trained; S4: the high-level deep feature vector f k and the strong semantic feature vector l q of the froth video are spliced and input into the multilayer perceptron, and the key performance indicators can be solved As shown in the following formula:

[0023]

[0024] Wherein, MLP(·) represents the multilayer perceptron.

[0025] The step two S1 comprises: setting the time interval of foam video image frames extracted from the foam video to 0.2 seconds.

[0026] The step two S1 comprises: VC(·) represents tensor flattening and vector compression operation, wherein the vector compression is performed by a vector compression layer, and the compression of high-dimensional deep vectors is realized by adopting a full connection layer or an autoencoder.

[0027] The step two S1 comprises: setting the strong semantic feature map layer threshold M1 to 4.

[0028] The industrial process data are temperature, pressure and flow collected by sensors in the industrial field; and the foam apparent features are bubble size, foam flow rate and texture.

[0029] The method is based on the foam video image collection system arranged on site to collect foam video data, extracts deep feature expression of the foam video, adds strong semantic data feature extraction of the industrial process, adopts the mid-term fusion method to fuse the two types of features extracted to obtain more powerful feature expression, enhances the understanding and generalization ability of the model, and is used to more accurately predict the key performance indicators of the flotation process, so as to more accurately judge the working condition and effectively guide the on-site operation. BRIEF DESCRIPTION OF DRAWINGS

[0030] Figure 1 The flowchart of the flotation key performance indicator monitoring model construction method based on fusion of foam video and strong semantic data. DETAILED DESCRIPTION

[0031] Figure 1 The flowchart is the process of the application, and comprises the following processes:

[0032] Step one: collect the zinc flotation foam video data by using the image collection system in the flotation field, and pre-process the collected data, for example:

[0033] 1) remove the error data exceeding the normal change threshold;

[0034] 2) remove the incomplete data;

[0035] Step two: respectively construct a foam video deep feature extraction subnetwork and a foam video strong semantic feature extraction subnetwork;

[0036] S1: construct a foam video deep feature extraction subnetwork based on a 3D convolution network and vector compression;

[0037] For the collected foam video data, 3D convolution is adopted to process the video data. In the 3D convolution, the 3D filter moves in the height, width and channel directions of the data.

[0038] The foam video image frames are extracted from the foam video and normalized to obtain a three-dimensional feature input graph V0 composed of original video image frame sequences.

[0039]

[0040] wherein e ab represents a numerical value corresponding to each pixel point in the video image.

[0041] A 3D convolution module is used to obtain a deep feature representation vector of the foam video. Denote F k , k = 1…M as a feature map processed by the 3D convolution module 3DConv_k, k = 1…M, and each layer of feature map is defined as follows:

[0042]

[0043] wherein M represents the total number of 3D convolution modules.

[0044] With the increase of the number of layers of the 3D convolution module, the semantic information contained in each layer of feature map is stronger. Denote M1 as a threshold of the number of layers of strong semantic feature map, and the high layer feature map F k , M1≤k≤M has higher semantic information, and in this example, the threshold M1 of the number of layers of strong semantic feature map is set to 4. Each video high layer feature map F k , M1≤k≤M is processed by tensor flattening and vector compression. This part of work is performed by the vector compression layer, and the compression of high-dimensional deep vector is realized by using a fully connected layer or an autoencoder.

[0045] After the flattening and compression processing, the foam video high layer deep feature representation vector f k is obtained as shown in the following formula:

[0046]

[0047] wherein VC(·) represents the tensor flattening and vector compression operation.

[0048] S2: constructing a foam video strong semantic feature extraction subnetwork based on LSTM;

[0049] In the froth flotation, some industrial process data (such as temperature, pressure, flow, etc.) can be collected by sensors in the industrial field; and the apparent features of the foam such as bubble size and foam flow rate are obtained by the long-term accumulated experience knowledge of the flotation operator and the professional knowledge in the flotation field.

[0050] Denote P as the total number of industrial process data elements, and S as the number of foam representation features, to obtain the industrial process data vector foam apparent feature vector The strong semantic feature vector X of the image frame (t) As shown in the following formula:

[0051]

[0052] Wherein, concat(·) represents a vector concatenation operation.

[0053] The strong semantic feature vector of each image frame is sequentially sent into the LSTM network to obtain the strong semantic feature vector l of the foam video q As shown in the following formula:

[0054] l q = LSTM(X (1) ,X (2) ,...,X (Q) )

[0055] Wherein, LSTM(·) represents a long short-term memory network, and Q represents the total number of frames of the current foam video.

[0056] Step three: fusion of video deep features and strong semantic features to solve the flotation key performance indicators;

[0057] S1: training the foam video deep feature extraction network based on the 3D convolution network and the vector compression network to obtain the pre-training weight of the network, so that the 3D convolution network has good weight;

[0058] S2: training the foam video strong semantic feature extraction network based on LSTM to obtain the pre-training weight of the network, so that the LSTM network has good weight;

[0059] S3: using the pre-training weights of the above two networks as initial values, training the entire network to obtain the accurate weight parameters of the network, so that the 3D convolution network and the LSTM network maintain the optimal coordination state.

[0060] S4: concatenating the high-level deep feature vector f k and the strong semantic feature vector l q of the foam video, and inputting them into the multi-layer perception machine to solve the key performance indicators As shown in the following formula:

[0061]

[0062] Wherein, MLP(·) represents a multi-layer perception machine.

[0063] The application provides a flotation key index monitoring model construction method based on data fusion, which is used for monitoring key indexes of a flotation process, effectively solves the shortcomings of poor generalization ability and weak representation ability of a traditional deep learning model, fully fuses foam video data and industrial process data, and thus can more accurately judge a working condition and effectively guide on-site operation.

Claims

1. A method for constructing a data fusion-based monitoring model of a key flotation index, characterized in that, The method comprises the following steps: Step one: collecting zinc flotation froth video data by using a flotation site image acquisition system, pre-processing the collected data, and constituting a zinc flotation data set; Step two: constructing a froth video deep feature extraction subnetwork and a froth video strong semantic feature extraction subnetwork; S1: constructing a froth video deep feature extraction subnetwork based on a 3D convolution network and vector compression; The foam video is extracted to obtain foam video image frames, and normalization is performed on the foam video image frames to obtain a three-dimensional feature input graph V0 composed of an original video image frame sequence; a 3D convolution module is used to obtain a deep feature representation vector of the foam video; denoted as F k , k = 1…M, the feature map processed by the 3DConv_k, k = 1…M, 3D convolution module, and each layer feature map is defined as follows: Wherein, M is the total number of 3D convolution modules; Let M1 be a strong semantic feature map layer threshold value, F k , M1≤k≤M is processed by tensor flattening and vector compression to obtain a foam video high-level deep feature representation vector f k , defined as follows: Wherein, VC(·) represents a tensor flattening and vector compression operation; S2: constructing a froth video strong semantic feature extraction subnetwork based on LSTM; In froth flotation, industrial process data is collected by sensors in the industrial field; froth appearance features are obtained by long-term accumulated experience knowledge of flotation operators and professional knowledge in flotation field; P is the total number of industrial process data elements, S is the number of froth appearance features, and the industrial process data vector matched with the froth video image frame is obtained Froth appearance feature vector Then the strong semantic feature vector X of the froth video image frame is (t) As shown in the following formula: Wherein, concat(·) represents a vector concatenation operation; The strong semantic feature vector is input to the LSTM network in the front and back order of the video frames to obtain a strong semantic feature vector l of the bubble video q As shown in the following formula: l q = LSTM(X (1) ,X (2) ,…,X (Q) ) Wherein, LSTM(·) represents a long short-term memory network, and Q represents the total number of frames of the current froth video; Step three: fusing video deep features and strong semantic features to solve the flotation key performance indicators; S1: training the froth video deep feature extraction network based on the 3D convolution network and the vector compression to obtain the pre-training weight of the network; S2: training the froth video strong semantic feature extraction network based on LSTM to obtain the pre-training weight of the network; S3: taking the pre-training weights of the above two networks as initial values to train the entire network model; S4: high-level deep feature vector f of the foaming video k and strong semantic feature vector l q concatenate and input into a multi-layer perception machine to solve key performance indicators as shown in the following formula: Wherein, MLP(·) represents a multi-layer perceptron.

2. The method of claim 1, wherein the method is characterized by: The step two S1 comprises: setting the time interval of the froth video image frame extraction froth video to 0.2 seconds.

3. The method of claim 1, wherein the method is characterized by: The step two S1 comprises: VC(·) represents a tensor flattening and vector compression operation, wherein the vector compression is performed by a vector compression layer, and a full connection layer or an autoencoder is used to realize compression of a high-dimensional deep vector.

4. The method of claim 1, wherein the method is characterized by: The step two S1 comprises: setting the strong semantic feature map layer threshold M1 to 4.

5. The method of claim 1, wherein the method is characterized by: The industrial process data is: temperature, pressure and flow rate, collected by sensors in the industrial site; and the froth apparent feature is: bubble size, froth flow rate and texture.

Images