A mineral flotation condition identification and prediction device and method

Abstract

The present invention provides a device for identifying and predicting mineral flotation conditions, comprising an image acquisition module, an ultrasonic ranging sensor, a main control module, a 5G communication module, a data transmission and positioning module, a display module, and a power supply module. The image acquisition module is responsible for acquiring image information of the current flotation conditions, while the ultrasonic ranging sensor above the flotation cell is used to measure the liquid level. By deploying the recognition algorithm on a worker's helmet, the present invention enables real-time identification of the current operating conditions and prediction of mineral flotation trends during operation, with the results displayed on a wearable recognition device. By wearing the device, workers can foresee fluctuations in the flotation process and understand the operating status of the flotation cell, enabling them to take appropriate measures in advance, further improving the stability and production efficiency of the mineral flotation process, reducing safety risks, and minimizing economic losses.

Description

Technical Field

[0001] The present invention relates to the technical field of mineral flotation, and in particular to a device and method for identifying and predicting mineral flotation working conditions. Background Art

[0002] The mineral flotation process adds organic solvents to separate concentrate particles from the slurry by adhering to the surface of the froth. However, factors such as low flotation tank agitator speed leading to a large number of large particles in the slurry, motor failure leading to obstructed slurry flow and accumulation, and improper use of frothers often lead to abnormal operating conditions such as settling and bubbling. These can lead to safety and economic problems such as overloaded flotation equipment causing fires, chemical leaks endangering worker health, and significant reductions in production efficiency. Therefore, timely identification of mineral flotation operating conditions is necessary to reduce economic losses and improve production efficiency. Existing operating condition identification methods typically rely on the experience of professional technicians or measure physical indicators such as the flotation tank's liquid level and pH value. Only when the liquid level or pH value in the flotation tank exceeds a set threshold can an abnormal operating condition be identified and warned.

[0003] Traditional methods for determining flotation conditions based on empirical knowledge have significant limitations and come with high labor costs. Individual differences in inspectors' experience in determining flotation conditions often lead to inconsistencies and subjective biases in the results, posing a significant risk of misjudgments or omissions. Furthermore, this reliance on manual monitoring is inefficient and unreliable, making it difficult to meet the precise detection requirements of modern industry.

[0004] However, methods that identify working conditions by setting thresholds for physical indicators often face challenges when faced with highly time-varying and dynamically complex flotation processes. Due to the volatility of flotation conditions, thresholds are difficult to set accurately. Setting the threshold too low can lead to frequent false alarms, wasting human resources; setting it too high can result in warnings being issued too late, resulting in significant economic losses. Furthermore, while emerging working condition identification methods based on image data can improve the accuracy of detecting current flotation conditions, they still struggle to predict trends in mineral flotation conditions.

[0005] Therefore, the working condition recognition algorithm based on multimodal data fusion of image data and liquid level information data is becoming a very promising direction in the field of mineral flotation working condition recognition and prediction. Summary of the Invention

[0006] In order to make up for the deficiencies of the prior art, the embodiments of the present application propose a method for identifying and predicting mineral flotation conditions to solve the problems existing in the prior art.

[0007] In order to solve the above technical problems, the present invention provides the following technical solutions:

[0008] A mineral flotation condition identification and prediction device includes an image acquisition module, an ultrasonic ranging sensor, a main control module, a 5G communication module, a data transmission and positioning module, a display module, and a power module. The image acquisition module is responsible for acquiring image information of the current flotation condition. The ultrasonic ranging sensor above the flotation tank is used to measure the height of the liquid level. The image and height data are transmitted to the main control module via the MIPI CSI interface and the built-in Bluetooth module in the data transmission and positioning module, respectively. The 5G communication module is used to realize real-time data upload and remote monitoring between the flotation equipment and the remote server. The main control module applies the working condition identification and prediction algorithm to identify the current flotation condition and predict the trend. The identification and prediction results are presented through the VR glasses in the display module.

[0009] A method for identifying and predicting mineral flotation conditions, using the above-mentioned device, comprises the following steps:

[0010] Step 1: Construct a mineral flotation condition dataset. Various working condition images are collected through the actual flotation process at the industrial site and a self-made mineral flotation condition recognition platform. The labels are reclassified into sink, normal, and bubbling flume. The self-made working condition dataset is then divided into a validation set and a training set.

[0011] Step 2: Build a model. Based on the improved YOLOV8, build the mineral flotation condition recognition model SD-LSTM-Model. First, use SD-Darknet as the backbone network to extract features, including: strided convolution layer, SD-C2f cross-stage feature fusion layer, SD-SPPF pooling layer, fully connected layer and activation function. Then, the features are fused and enhanced through the neck network. Finally, three decoupled head detectors and prediction modules are called for detection and prediction.

[0012] Step 3: Feature extraction: The SD-C2f module is used to connect feature maps at different levels to extract dense features at different scales, focusing on more important channel features. The SD-SPPF fast spatial pyramid pooling layer is then used to extract deep information in the spatial and channel dimensions, improving the network's ability to extract detailed information from mineral flotation photos.

[0013] Step 4: Feature fusion: Extract the spatial features of the image data in the YOLO detection module, fuse the spatial features with the time series measured by the ultrasonic distance sensor through the self-attention mechanism, and feed them into the LSTM network to predict the mineral flotation conditions at the next moment.

[0014] Step 5: Detection and prediction: The working condition photos taken by the equipment camera and the data obtained by the ultrasonic rangefinder are input into the detector to identify the working condition and predict the working trend of mineral flotation so that the staff can take countermeasures in advance.

[0015] As a further technical solution of the present invention, the dataset is constructed by collecting samples required for model training based on a self-made mineral flotation condition identification platform. Label Img is used to redefine labels into settling tank, normal tank, and bubbling tank according to different working conditions, totaling 50,000 labels. The redefined labels and samples constitute a new mineral flotation condition identification dataset.

[0016] As a further technical solution of the present invention: the model construction is specifically: building a deep convolutional neural network model SD-LSTM-Model improved based on Yolov8, including: a deep convolutional network SD-Darknet, an SD-C2f module, an SD-SPPF module, a feature fusion module and a decoupling head detector, and an LSTM-based prediction module. When a picture is input into the model, the backbone network extracts the feature information of the picture through strided convolution. The SD-C2f module integrates the channel attention mechanism of SE, allowing the neural network to pay more attention to important channel features, and then extracts the depth detail information through the SD-SPPF module. Through feature fusion and enhancement, the decoupling detection head is used to detect the video, output the recognition result, extract the spatial image features in the detection result, and fuse the liquid level height data collected by the ultrasonic sensor. The data is input into the LSTM network for prediction.

[0017] As a further technical solution of the present invention: the feature extraction: first, through a series of standard convolutions, the basic features of the image are extracted, and based on the improvement of the C2f module of Darknet-53, the SD-C2f module is designed to divide the input feature map into channels, so as to process each channel separately to improve the discrimination between features. Then, N Bottleneck2 modules are used to process the feature map after channel separation, and the processed feature map is spliced ​​with the feature map after separation of other channels to obtain the connection between the features of different channels, and then processed by the Bottlencek2 module in sequence to reduce the number of parameters and speed up the calculation. The above feature extraction process is repeated N times, and the N processed feature maps are spliced. The SE attention mechanism module is introduced to adaptively adjust and proofread the feature weights, so that the network can dynamically adjust the attention of the channel and enhance the utilization of effective information. Finally, the feature map is processed using the residual network to obtain feature information rich in the gradient flow and avoid the problems of gradient explosion and network performance degradation.

[0018] As a further technical solution of the present invention: the SE channel attention mechanism module includes Squeeze and Excitation operations, which are designed to make the neural network pay more attention to the important features of the current task. First, the global spatial information is compressed into a channel descriptor, that is, a global flat pooling operation is performed on each channel, and the features containing the global information with dimensions H*W*C are compressed into a feature vector Z of 1*1*C. The channel features of C feature maps are compressed into a numerical value, and channel-level statistics Z with contextual information are generated to solve the problem of channel dependency. C is defined as follows:

[0019]

[0020] Among them, Z C is the Cth element of Z, F sq Refers to the global average pooling of channels, U C are the C output feature maps obtained;

[0021] In order to utilize the information gathered in the compression operation, the Excitation excitation operation is used to fully capture the dependencies between channels. A two-layer fully connected gate mechanism is adopted. The first fully connected layer reduces the C channels of the feature map to C / m channels to reduce the amount of computation. After processing through a ReLU nonlinear activation layer, it is input into the second fully connected layer to restore the number of channels to C channels and obtain the weight P through Sigmoid activation. Finally, the 1*1*C dimension P is obtained, which is used to represent the weights of the C feature maps in the feature map U, where m refers to the compression ratio. P is expressed as follows:

[0022] P=F ex (z, W)=σ(g(z, W))=σ(W2δ(W1z)) (2)

[0023] Finally, in order to weight the obtained attention weight P to the features of each channel, it is necessary to multiply each feature map in the feature map U by the corresponding weight to obtain the final output T of the SE attention mechanism module, which is expressed as follows:

[0024] T=F scale (U c , P)=U c P (3)

[0025] The feature map extracted by the SD-Darknet backbone network is input into the SD-SPPF spatial feature pyramid pooling layer. The role of SPPF spatial pyramid pooling is to apply the maximum pooling operation of different scales to the entire feature map. First, the input feature map is divided into different grid areas, and then the maximum pooling operation is applied to each area to generate a fixed-length feature vector. Finally, the feature vectors generated at these different scales are spliced ​​together to form a fixed-length feature vector to meet the fixed-dimensional input requirements when the fully connected layer implements classification.

[0026] The SD-SPPF module first uses DW depth-wise separable convolution to extract features, and then forms a fixed-length feature vector through maximum pooling and splicing. DW convolution consists of depth-wise convolution and point-wise convolution. In depth-wise convolution, the input channel and a separate filter are first convolved to obtain a feature map consistent with the number of channels, thereby realizing the extraction of spatial features.

[0027] In point-by-point convolution, the 1x1 convolution kernel is used to increase the dimension of the output of the depth-wise convolution and compress it to the required number of channels to extract channel features. Assuming that the convolution kernel size is D K *D K , the input channel is M, the output channel is N, and the output feature map size is D F *D F ,The comparison value of the depth-separable convolution and the standard convolution parameter λ:

[0028]

[0029] Generally, N is large enough to be ignored. When we use the most common 3*3 convolution kernel, the number of parameters and computational complexity of depthwise separable convolution can be reduced to one-ninth of the original, making it easier to deploy on helmet devices for working condition identification.

[0030] As a further technical solution of the present invention: the feature fusion specifically comprises: using the SD-LSTM-Model model to extract the spatial features of the current liquid level, i.e., the liquid level height, from the image, capturing the relationship between the liquid level height and the spatial features, and obtaining the fused feature representation.

[0031] As a further technical solution of the present invention: the detection and prediction specifically involves: using the working condition photos taken by the recognition device camera and the liquid level height data obtained by the ultrasonic rangefinder as input, the designed SD-LSTM-Model network is used to identify and predict the current mineral flotation working conditions, and the results are displayed on VR glasses of a self-made flotation working condition identification and trend prediction device.

[0032] One or more technical solutions provided in the embodiments of this application have at least the following technical effects or advantages:

[0033] By deploying this recognition algorithm on workers' helmets, they can identify current working conditions and predict mineral flotation trends in real time during operation, with the results displayed on a wearable recognition device. By wearing the device, workers can foresee fluctuations in the flotation process and understand the operating status of the flotation cells, allowing them to take appropriate measures in advance, further improving the stability and production efficiency of the mineral flotation process, reducing safety risks, and minimizing economic losses. BRIEF DESCRIPTION OF THE DRAWINGS

[0034] Figure 1 It is the structural diagram of the flotation condition identification and trend prediction equipment.

[0035] Figure 2 It is a flow chart for flotation condition identification and prediction.

[0036] Figure 3 This is the SD-LSTM-Model structure diagram.

[0037] Figure 4 This is the SD-C2f structure diagram.

[0038] Figure 5 This is the SD-SPPF structure diagram.

[0039] Figure 6 This is the result chart of the homemade detection system. DETAILED DESCRIPTION

[0040] The following is a clear and complete description of the technical solutions in the embodiments of the present invention. Obviously, the embodiments described are only some embodiments of the present invention, not all embodiments. Based on the embodiments 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.

[0041] Reference Figure 1-6 , a mineral flotation condition identification and prediction device, consists of a series of modules: image acquisition module, ultrasonic ranging sensor, main control module, 5G communication module, data transmission and positioning module, display module (VR glasses), power module;

[0042] Among them, the image acquisition module is responsible for obtaining image information of the current flotation working condition, while the ultrasonic ranging sensor above the flotation tank is used to measure the height of the liquid level. The image and height data are transmitted to the main control module through the MIPI CSI interface and the built-in Bluetooth module in the data transmission and positioning module respectively to ensure efficient data transmission and precise positioning of the equipment. At the same time, the 5G communication module realizes real-time data upload and remote monitoring between the flotation equipment and the remote server. The main control module then applies the working condition identification and prediction algorithm to identify and predict the trend of the current flotation working condition. Finally, the identification and prediction results are presented through the VR glasses in the display module to help operators grasp the working conditions and their development trends in real time. The structure of the flotation working condition identification and trend prediction equipment is as follows: Figure 1 shown.

[0043] At the same time, the present invention also provides a method for identifying and predicting mineral flotation conditions, comprising the following steps:

[0044] Step 1: Construct a mineral flotation condition dataset. Collect various working condition images through the actual flotation process on industrial site and a self-made mineral flotation condition identification platform. Reclassify the labels into sink, normal, and bubbling flume. Then, divide the self-made working condition dataset into a validation set and a training set.

[0045] Step 2: Model Building: Build the SD-LSTM-Model, a mineral flotation condition recognition model, based on the improved YOLO V8. First, use SD-Darknet as the backbone network to extract features, including strided convolutional layers, SD-C2f cross-stage feature fusion layers, SD-SPPF pooling layers, fully connected layers, and activation functions. Features are then fused and enhanced using the neck network. Finally, three decoupled head detectors and a prediction module are used for detection and prediction.

[0046] Step 3: Feature Extraction: The SD-C2f module connects feature maps at different levels to extract dense features at different scales, focusing on more important channel features. The SD-SPPF fast spatial pyramid pooling layer then extracts deep information in both the spatial and channel dimensions, improving the network's ability to extract detailed information from photos of mineral flotation conditions.

[0047] Step 4, feature fusion stage: Extract the spatial features of the image data in the YOLO detection module, fuse the spatial features with the time series measured by the ultrasonic distance sensor through methods such as the self-attention mechanism, and send them to the LSTM network to predict the mineral flotation conditions at the next moment.

[0048] Step 5, Detection and Prediction Phase: The working condition photos taken by the equipment camera and the data obtained by the ultrasonic rangefinder are input into the detector to identify the working condition and predict the working trend of the mineral flotation so that the staff can take countermeasures in advance.

[0049] (1) Dataset Construction: Using a self-developed mineral flotation condition identification platform, we collected samples required for model training. Using Label Img, we redefined labels into sink, normal, and bubbling flumes according to the operating conditions, totaling 50,000 labels. The redefined labels and samples constituted a new mineral flotation condition identification dataset.

[0050] (2) Model building: Building a deep convolutional neural network model SD-LSTM-Model based on Yolov8 ( Figure 3 ), including: deep convolutional network SD-Darknet, SD-C2f module ( Figure 4 ), SD-SPPF module ( Figure 5 ), a feature fusion module, a decoupled head detector, and an LSTM-based prediction module. When an image is input to the model, the backbone network extracts image features through strided convolution. The SD-C2f module incorporates the SE channel attention mechanism, allowing the neural network to focus on important channel features. The SD-SPPF module then extracts depth details. Through feature fusion and enhancement, the decoupled detection head detects the video and outputs the recognition results. Spatial image features are extracted from the detection results, integrated with liquid level data collected by the ultrasonic sensor, and then input into the LSTM network for prediction.

[0051] (3) Feature extraction: First, a series of standard convolutions are used to extract the basic features of the image. Based on the improvement of the C2f module of Darknet-53, the SD-C2f module is designed. The input feature map is divided into channels to process each channel separately to improve the discrimination between features. Then, N Bottleneck2 modules are used to process the feature map after channel separation. The processed feature map is spliced ​​with the feature map after other channel separation to obtain the connection between different channel features. Then, it is processed by Bottlencek2 modules in sequence to reduce the number of parameters and speed up the calculation. The above feature extraction process is repeated N times, and the N processed feature maps are spliced. The SE attention mechanism module is introduced to adaptively adjust and proofread the feature weights, so that the network can dynamically adjust the attention of the channel and enhance the utilization of effective information. Finally, the residual network is used to process the feature map to obtain feature information rich in gradient flow and avoid the problems of gradient explosion and network performance degradation.

[0052] The SE channel attention mechanism module mainly includes Squeeze (compression) and Excitation (excitation) operations, which aim to make the neural network pay more attention to the important features of the current task. First, the global spatial information is compressed into a channel descriptor, that is, a global flat pooling operation is performed on each channel, and the features containing the global information with dimensions H*W*C are compressed into a feature vector Z of 1*1*C. The channel features of C feature maps are compressed into a single value, generating channel-level statistics Z with contextual information to solve the problem of channel dependency. C is defined as follows:

[0053]

[0054] Among them, Z C is the Cth element of Z, F sq Refers to the global average pooling of channels, U C are the C output feature maps obtained.

[0055] In order to utilize the information gathered in the compression operation, the Excitation excitation operation is used to fully capture the dependencies between channels. A two-layer fully connected gate mechanism is used. The first fully connected layer reduces the C channels of the feature map to C / m channels to reduce the amount of computation. After processing through a ReLU nonlinear activation layer, it is input into the second fully connected layer to restore the number of channels to C channels. The weight P is obtained through Sigmoid activation, and the final 1*1*C dimension P is used to represent the weights of the C feature maps in the feature map U, where m refers to the compression ratio. P is expressed as follows:

[0056] P=F ex (z, W)=σ(g(z, W))=σ(W2δ(W1z)) (2)

[0057] Finally, in order to weight the obtained attention weight P to the features of each channel, it is necessary to multiply each feature map in the feature map U by the corresponding weight to obtain the final output T of the SE attention mechanism module. T is expressed as follows:

[0058] T=F scale (U c , P)=U c P (3)

[0059] The feature maps extracted by the SD-Darknet backbone network are input into the SD-SPPF spatial feature pyramid pooling layer. SPPF spatial pyramid pooling applies maximum pooling operations of different scales to the entire feature map. It first divides the input feature map into different grid regions, then applies the maximum pooling operation to each region to generate a fixed-length feature vector. Finally, these feature vectors generated at different scales are concatenated to form a fixed-length feature vector, which meets the fixed-dimensional input requirement of the fully connected layer for classification.

[0060] The SD-SPPF module first uses DW depthwise separable convolution to extract features, then performs max pooling and concatenation to form a fixed-length feature vector. DW convolution consists of depthwise convolution and pointwise convolution. In depthwise convolution, the input channels are first convolved with individual filters to obtain a feature map with the same number of channels, thus extracting spatial features.

[0061] In point-by-point convolution, a 1x1 convolution kernel is used to increase the dimension of the output of depth-wise convolution and compress it to the required number of channels to extract channel features. Assume that the convolution kernel size is D K *D K , the input channel is M, the output channel is N, and the output feature map size is D F *D F ,The comparison value of the depth-separable convolution and the standard convolution parameter λ:

[0062]

[0063] Generally, N is large enough to be ignored. When we use the most common 3x3 convolution kernel, the number of parameters and computational complexity of depthwise separable convolution can be reduced to one-ninth of the original, making it easier to deploy on helmet-mounted devices for work-condition recognition.

[0064] (4) Feature fusion stage: Use the SD-LSTM-Model model to extract the spatial features of the current liquid level (including liquid level height, area, etc.) from the image, effectively capture the relationship between the liquid level height and spatial features, and obtain the fused feature representation.

[0065] First, the sensor's liquid level data is mapped to the dimensional space of the query vector through a fully connected layer using a self-attention mechanism, generating a query vector Q (Query). An SD-LSTM model is used to extract the spatial features of the current liquid level (including liquid level and area) from the image. This generates a key vector K (Key) representing the importance of these image features and a value vector V (Value) representing the content of these spatial features. The dot product is used to calculate the similarity between the query and key vectors to generate an attention score matrix. The score matrix is ​​then normalized and converted into a weight matrix representing the importance of each spatial feature to the sensor data. The value vectors are weighted and summed using the attention weights to generate a fused vector 'attention_features' that includes both image and sensor features, weighted by relevance.

[0066] At the same time, the sensor's liquid level features are extracted through a sliding window and horizontally connected to the liquid surface's spatial features to form a concatenated fused feature vector, 'sum_features'. The two fused feature vectors are then processed through a transform-concatenate-transform process. First, the 'attention_features' and 'sum_features' are linearly transformed through a fully connected layer. These transformed feature vectors are then concatenated to form a new fused feature vector, 'combined_features'. This fused feature vector is then linearly transformed through a fully connected layer to capture the relationship between the liquid level and spatial features, ultimately resulting in a fused feature representation.

[0067] The fused features are preprocessed and converted into a three-dimensional tensor suitable for input into the LSTM network to train the LSTM model and predict the change in the liquid level at the next moment. t Positive (liquid level rises), negative (liquid level drops), and the results are displayed on the glasses of the identification helmet.

[0068] (5) Detection and prediction stage: Using the working condition photos taken by the recognition device camera and the liquid level height data obtained by the ultrasonic rangefinder as input, the designed SD-LSTM-Model network is used to identify and predict the current working condition of the mineral flotation, and the results are displayed on the VR glasses of the self-made flotation working condition identification and trend prediction equipment.

[0069] This invention addresses the difficulty of accurately identifying and predicting flotation conditions based on single factors, such as physical indicators or image information. By integrating liquid level and flotation condition image information, this invention constructs a mineral flotation condition identification and prediction model, the SD-LSTM-Model, and deploys the model on a wearable device. By performing multimodal data fusion enhancement on image data and liquid level information, an LSTM network is used to learn about changes in the liquid level within the flotation cell. Based on these changes, the mineral flotation process is predicted. The results and trends are then displayed on VR glasses on the wearable device, allowing operators to observe them in real time and take appropriate preventive measures and adjustments in advance.

[0070] It will be apparent to those skilled in the art that the present invention is not limited to the details of the exemplary embodiments described above and that the invention can be embodied in other specific forms without departing from the spirit or essential characteristics of the invention. Therefore, the embodiments should be considered in all respects as illustrative and non-restrictive, and the scope of the invention is defined by the appended claims rather than the foregoing description, and it is intended that all variations that come within the meaning and range of equivalents of the claims be embraced therein.

[0071] In addition, it should be understood that although this specification is described in terms of implementation methods, not every implementation method contains only one independent technical solution. This narrative method of the specification is only for the sake of clarity. Those skilled in the art should regard the specification as a whole. The technical solutions in each embodiment have also been appropriately combined to form other implementation methods that are easy for those skilled in the art to understand.

Claims

1. A mineral flotation condition identification and prediction device, comprising an image acquisition module, an ultrasonic ranging sensor, a main control module, a 5G communication module, a data transmission and positioning module, a display module, and a power module, characterized in that: The image acquisition module is responsible for acquiring image information of the current flotation conditions. The ultrasonic ranging sensor above the flotation cell is used to measure the height of the liquid level. The image and height data are transmitted to the main control module via the MIPI CSI interface and the built-in Bluetooth module in the data transmission and positioning module, respectively. The 5G communication module is used to achieve real-time data upload and remote monitoring between the flotation equipment and the remote server. The main control module applies the working condition recognition and prediction algorithm to identify the current flotation conditions and predict trends. The recognition and prediction results are displayed through the VR glasses in the display module. The feature extraction of image information is specifically as follows: first, through a series of standard convolutions, the basic features of the image are extracted. Based on the improvement of the C2f module of Darknet-53, the SD-C2f module is designed to split the input feature map into channels to process each channel separately to improve the discrimination between features. Then, N Bottleneck2 modules are used to process the feature maps after channel separation, and the processed feature maps are spliced ​​with the feature maps after other channel separation to obtain the connection between the features of different channels. Then, they are processed in sequence through the Bottlencek2 module to reduce the number of parameters and speed up the calculation. The above feature extraction process is repeated N times, and the N processed feature maps are spliced. The SE attention mechanism module is introduced to adaptively adjust and proofread the feature weights.

2. A method for identifying and predicting mineral flotation conditions, characterized in that: The device according to claim 1 comprises the following steps: Step 1: Construct a mineral flotation condition dataset. Various working condition images are collected through the actual flotation process at the industrial site and a self-made mineral flotation condition recognition platform. The labels are reclassified into sink, normal, and bubbling flume. The self-made working condition dataset is then divided into a validation set and a training set. Step 2: Build a model. Based on the improved YOLOV8, build the mineral flotation condition recognition model SD-LSTM-Model. First, use SD-Darknet as the backbone network to extract features, including: strided convolution layer, SD-C2f cross-stage feature fusion layer, SD-SPPF pooling layer, fully connected layer and activation function. Then, the features are fused and enhanced through the neck network. Finally, three decoupled head detectors and prediction modules are called for detection and prediction. Step 3: Feature extraction: The SD-C2f module is used to connect feature maps at different levels to extract dense features at different scales, focusing on more important channel features. The SD-SPPF fast spatial pyramid pooling layer is then used to extract deep information in the spatial and channel dimensions, improving the network's ability to extract detailed information from mineral flotation photos. Step 4: Feature fusion: Extract the spatial features of the image data in the YOLO detection module, fuse the spatial features with the time series measured by the ultrasonic distance sensor through the self-attention mechanism, and feed them into the LSTM network to predict the mineral flotation conditions at the next moment. Step 5: Detection and prediction: The working condition photos taken by the equipment camera and the data obtained by the ultrasonic rangefinder are input into the detector to identify the working condition and predict the working trend of the mineral flotation so that the staff can take countermeasures in advance; The feature extraction is specifically as follows: first, the basic features of the image are extracted through a series of standard convolutions, and the SD-C2f module is designed based on the improvement of the C2f module of Darknet-53. The input feature map is segmented into channels to process each channel separately to improve the discrimination between features. Then, N Bottleneck2 modules are used to process the feature maps after channel separation, and the processed feature maps are spliced ​​with feature maps after other channel separation to obtain the connection between features of different channels. Then, the processed feature maps are processed in sequence through the Bottlencek2 module to reduce the number of parameters and speed up the calculation. The above feature extraction process is repeated N times, and the N processed feature maps are spliced. The SE attention mechanism module is introduced to adaptively adjust and proofread the feature weights.

3. A method for identifying and predicting mineral flotation conditions according to claim 2, characterized in that: The specific steps of constructing the dataset are as follows: based on the self-made mineral flotation working condition identification platform, samples required for model training are collected. According to different working conditions, Label Img is used to redefine labels as settling tank, normal tank and bubbling tank, totaling 50,000 labels. The redefined labels and samples constitute a new mineral flotation working condition identification dataset.

4. A method for identifying and predicting mineral flotation conditions according to claim 2, characterized in that: The model is specifically constructed as follows: an improved deep convolutional neural network model SD-LSTM-Model based on Yolov8 is constructed, including: a deep convolutional network SD-Darknet, an SD-C2f module, an SD-SPPF module, a feature fusion module, a decoupling head detector, and an LSTM-based prediction module. When a picture is input into the model, the backbone network extracts the feature information of the picture through strided convolution. The SD-C2f module integrates the channel attention mechanism of SE, allowing the neural network to pay more attention to important channel features. The SD-SPPF module is then used to extract the depth detail information. Through feature fusion and enhancement, the decoupling detection head is used to detect the video, output the recognition result, extract the spatial image features in the detection result, fuse the liquid level height data collected by the ultrasonic sensor, and input the data into the LSTM network for prediction.

5. A method for identifying and predicting mineral flotation conditions according to claim 4, characterized in that: The SE attention mechanism module includes Squeeze and Excitation operations, which are designed to make the neural network pay more attention to the important features of the current task. First, the global spatial information is compressed into a channel descriptor, that is, a global flat pooling operation is performed on each channel, and the features with dimensions H*W*C containing global information are compressed into a feature vector Z of 1*1*C. The channel features of C feature maps are compressed into a numerical value, and channel-level statistics Z with contextual information are generated to solve the problem of channel dependency. C is defined as follows: Among them, Z C is the Cth element of Z, F sq Refers to the global average pooling of channels, U C are the C output feature maps obtained; In order to utilize the information gathered in the compression operation, the Excitation excitation operation is used to fully capture the dependencies between channels. A two-layer fully connected gate mechanism is adopted. The first fully connected layer reduces the C channels of the feature map to C / m channels to reduce the amount of computation. After processing through a ReLU nonlinear activation layer, it is input into the second fully connected layer to restore the number of channels to C channels and obtain the weight P through Sigmoid activation. Finally, the 1*1*C dimension P is obtained, which is used to represent the weights of the C feature maps in the feature map U, where m refers to the compression ratio. P is expressed as follows: P=F ex (z,W)=σ(g(z,W))=σ(W2δ(W1z)) (2) Finally, in order to weight the obtained attention weight P to the features of each channel, it is necessary to multiply each feature map in the feature map U by the corresponding weight to obtain the final output T of the SE attention mechanism module, which is expressed as follows: T=F scale (U c ,P)=U c P (3) The feature map extracted by the SD-Darknet backbone network is input into the SD-SPPF spatial feature pyramid pooling layer. The role of SPPF spatial pyramid pooling is to apply the maximum pooling operation of different scales to the entire feature map. First, the input feature map is divided into different grid areas, and then the maximum pooling operation is applied to each area to generate a fixed-length feature vector. Finally, the feature vectors generated at these different scales are spliced ​​together to form a fixed-length feature vector to meet the fixed-dimensional input requirements when the fully connected layer implements classification. The SD-SPPF module first uses DW depth-wise separable convolution to extract features, and then forms a fixed-length feature vector through maximum pooling and splicing. DW convolution consists of depth-wise convolution and point-wise convolution. In depth-wise convolution, the input channel and a separate filter are first convolved to obtain a feature map consistent with the number of channels, thereby realizing the extraction of spatial features. In point-by-point convolution, the 1x1 convolution kernel is used to increase the dimension of the output of the depth-wise convolution and compress it to the required number of channels to extract channel features. Assuming that the convolution kernel size is D K *D K , the input channel is M, the output channel is N, and the output feature map size is D F *D F ,The comparison value of the depth-separable convolution and the standard convolution parameter λ: Generally, N is large enough to be ignored. When we use the most common 3*3 convolution kernel, the number of parameters and computational complexity of depthwise separable convolution can be reduced to one-ninth of the original, making it easier to deploy on helmet devices for working condition identification.

6. A method for identifying and predicting mineral flotation conditions according to claim 2, characterized in that: The feature fusion specifically involves using the SD-LSTM-Model model to extract the spatial features of the current liquid level from the image, capturing the relationship between the liquid level height and the spatial features, and obtaining a fused feature representation.

7. A method for identifying and predicting mineral flotation conditions according to claim 2, characterized in that: The detection and prediction specifically involves: using the working condition photos taken by the recognition device camera and the liquid level height data obtained by the ultrasonic rangefinder as input, the designed SD-LSTM-Model network is used to identify and predict the current mineral flotation working conditions, and the results are displayed on VR glasses, a self-made flotation working condition identification and trend prediction device.

Images