A method for predicting the grade of a zinc flotation concentrate

Abstract

The application discloses a zinc flotation concentrate grade prediction method, and belongs to the technical field of froth flotation. The method obtains a froth video of a zinc flotation process and samples the froth video according to a preset time interval to obtain a froth image sequence. Froth size, froth color and texture features are extracted from the froth image to form multi-dimensional semantic feature time series data. Distribution feature analysis is performed on the time series data to obtain actual distribution features. The run-of-mine grade and the flotation reagent addition amount under the current working condition are input into a pre-trained expert knowledge reasoning model to output target distribution constraint parameters. Distribution consistency weighted evaluation is performed based on the deviation between the actual distribution features and the target distribution constraint parameters, and the time series data is reconstructed based on the matching weight to obtain non-equal time interval semantic feature time series data. The zinc flotation concentrate grade prediction value is output through a prediction model. The application can improve the accuracy and robustness of concentrate grade prediction under complex flotation conditions.

Description

Technical Field

[0001] This invention belongs to the field of froth flotation technology, specifically relating to a method for predicting the grade of zinc flotation concentrate. Background Technology

[0002] Foam flotation is a mineral processing method that separates valuable minerals from gangue based on the differences in hydrophilicity and hydrophobicity of mineral surfaces, with the ultimate goal of obtaining high-grade concentrate products. In zinc flotation, to consistently and stably obtain high-grade concentrate products, on-site operators typically observe the color, texture, and other characteristics of the flotation cell foam surface with the naked eye, judge the current flotation status based on expert experience, and adjust reagents according to the concentrate grade.

[0003] Currently, flotation sites primarily rely on fluorescence analyzers to detect concentrate grade. However, this equipment is expensive and difficult to maintain. Typically, only one analyzer is used across the entire flotation site, yet it needs to cover grade detection at multiple key points. This results in long detection cycles (approximately 20 minutes per test) at each point, poor timeliness of grade detection, and difficulty in meeting the needs of real-time reagent adjustments. Therefore, achieving real-time prediction of concentrate grade is crucial for optimizing flotation production indicators, reducing detection costs, and ensuring the economic benefits of the flotation plant.

[0004] Existing methods for predicting concentrate grade in flotation processes mainly rely on establishing a correlation model between froth image features and concentrate grade. A typical process involves: acquiring real-time froth video using an industrial camera, sampling a series of froth images from the video, extracting the froth image feature sequence, and then using a data-driven model to establish a mapping relationship between features and grade, thereby achieving grade prediction.

[0005] However, the existing methods described above have the following shortcomings: First, the flotation process is complex and the on-site environment is noisy. The froth images sampled from the video are easily affected by environmental interference factors such as changes in lighting, slurry splashing, and bubble merging. Some sampled froth images may not reflect the actual flotation conditions, resulting in noisy data mixed in with the extracted semantic feature sequences of the froth images. This leads to insufficient accuracy and reliability of the feature sequences, ultimately limiting the accuracy of concentrate grade prediction.

[0006] Second, existing methods typically input all sampled frames into the prediction model with equal weight, lacking a mechanism for identifying and constraining data quality. When there are anomalous data points in the sampled sequence that do not conform to the current operating conditions, the model cannot automatically identify and mitigate their impact, resulting in poor robustness of the prediction results.

[0007] Third, the rich expert experience and knowledge accumulated by on-site operators over a long period of time, such as the expected changing trends and distribution range of foam characteristics under different ore grades and reagent conditions, have not yet been effectively utilized.

[0008] Therefore, how to integrate these expert experiences into the prediction process, constrain the semantic feature sequence of foam images, and improve the quality and physical consistency of the feature sequence is a challenge currently facing the field of intelligent monitoring of the flotation process. Summary of the Invention

[0009] This invention provides a method for predicting the grade of zinc flotation concentrate, which overcomes the above-mentioned defects in the prior art.

[0010] This invention provides a method for predicting the grade of zinc flotation concentrate, achieved through the following specific technical means, including the following steps: Acquire foam videos of the zinc flotation process and sample the foam videos at preset time intervals to obtain a continuous foam image sequence; Semantic features are extracted from each frame of the foam image sequence to obtain multidimensional semantic feature time-series data including foam size features, foam color features, and texture features. The distribution characteristics of the multidimensional semantic feature time series data are analyzed to obtain the actual distribution characteristics used to characterize the changing trends and fluctuation patterns of each semantic feature; Obtain the feed grade and flotation reagent addition amount under the current flotation conditions, and input the feed grade and flotation reagent addition amount into a pre-trained expert knowledge reasoning model to output the target distribution constraint parameters corresponding to each semantic feature under the current conditions. Based on the deviation between the actual distribution characteristics and the target distribution constraint parameters, the distribution consistency weighted evaluation is performed on the multidimensional semantic feature time series data to generate matching weights for each data point, and the time series data is reconstructed based on the matching weights to obtain non-equal time interval semantic feature time series data. The non-equal time interval semantic feature time series data and their corresponding time intervals are input into the time series prediction model, and the predicted value of zinc flotation concentrate grade at the current moment is output.

[0011] A further technical solution is that the semantic features include: foam size features, foam color features, and texture features extracted based on the gray-level co-occurrence matrix.

[0012] A further technical solution is that the semantic features are extracted by: obtaining the foam size feature through an image segmentation method based on a marker watershed; obtaining the foam color feature through color space transformation; and calculating the texture feature through a gray-level co-occurrence matrix, including inverse moment, correlation, contrast, and energy.

[0013] A further technical solution is that the actual distribution characteristics include the first-order difference mean and the second-order difference mean, wherein: the first-order difference mean is used to characterize the average rate of change of semantic features; and the second-order difference mean is used to characterize the degree of fluctuation of the trend of semantic feature change.

[0014] A further technical solution is that the expert knowledge reasoning model is a fuzzy logic interval reasoning model, whose output includes the first-order difference mean interval and the second-order difference mean interval of each semantic feature, which are used to characterize the dynamic distribution range constraint of the semantic features under the current working condition.

[0015] A further technical solution is that the weighted evaluation of distribution consistency includes: for each semantic feature data point, calculating the deviation value between its actual distribution characteristics and the target distribution constraint parameters; A matching weight function is constructed based on the deviation value, and the matching weight function satisfies the following: the smaller the deviation value, the greater the matching weight; the larger the deviation value, the smaller the matching weight. The semantic feature time series data is weighted and reconstructed based on the matching weights, so that high-weight data points are retained or enhanced, and low-weight data points are weakened or downsampled, thereby generating non-equal time interval semantic feature time series data.

[0016] A further technical solution is that the time series prediction model is a GRU-D model based on a combination of gated cyclic units and a time decay mechanism, which is used to process time series data with semantic features of non-equal time intervals and output concentrate grade prediction results.

[0017] A further technical solution is that the training process of the expert knowledge reasoning model includes: acquiring historical foam video data and extracting corresponding semantic feature time-series data and distribution features; acquiring historical ore grade and historical flotation reagent addition amount; using the ore grade and reagent addition amount as input and the semantic feature distribution features as output, training the fuzzy logic interval reasoning model; The training process of the GRU-D model includes: generating semantic feature time series data based on historical bubble videos; using the expert knowledge reasoning model to perform distribution consistency weighted reconstruction on the semantic feature time series data to obtain non-uniform time interval semantic feature time series data; and using the non-uniform time interval semantic feature sequence and time interval as input and the corresponding concentrate grade as output to train the GRU-D model.

[0018] Compared with the prior art, the present invention has the following beneficial effects: This invention introduces an expert knowledge constraint mechanism. By constructing a fuzzy logic interval reasoning model, the expert experience accumulated by on-site operators over a long period of time is encoded into a computable knowledge system. Using the ore grade and the amount of flotation reagent added as the working condition input, the target distribution constraint parameters (i.e., the first-order difference mean interval and the second-order difference mean interval) that each semantic feature should satisfy under the current working condition are inferred. From the perspective of expert knowledge, the dynamic change range of the foam characteristics under the current flotation working condition is quantified. This mechanism makes the prediction process no longer completely dependent on data-driven, but organically integrates domain knowledge with the data model.

[0019] This invention proposes a data reconstruction strategy based on distribution consistency weighted evaluation, replacing the traditional methods of simple equal-weighted input or crude data point deletion. It calculates the deviation between the actual distribution characteristics of each data point and the target distribution constraint parameters, and then performs weighted reconstruction sampling based on matching weights. This approach effectively weakens the impact of outlier data points on prediction results while retaining useful information to a certain extent, exhibiting better information utilization and robustness compared to directly deleting data points.

[0020] Compared to traditional methods for predicting concentrate grade, this invention establishes a correlation between foam semantic features and concentrate grade, and builds a single data-driven model for prediction. This effectively introduces knowledge constraints, acquires expert experience knowledge from historical data, and applies knowledge constraints to the foam image sequences sampled from foam videos. Considering the numerous interference factors in flotation plants and the complex flotation environment, foam surface features may be affected by the environment and subject to interference. Through knowledge constraints, some foam image interference that does not belong to the current operating conditions is effectively removed, improving the accuracy of the data model input. Furthermore, to effectively process non-equal interval semantic feature sequences, a GRU-D model is built to achieve real-time prediction of concentrate grade, effectively maintaining the stability of on-site operation and economic benefits of the flotation plant. Attached Figure Description

[0021] Figure 1 This is the data flow diagram of the present invention; Figure 2 This is a system module structure diagram of the present invention. Detailed Implementation

[0022] The specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings. It should be understood that the specific embodiments of the present invention are merely exemplary and should not be construed as limiting the scope of protection of the present invention. Those skilled in the art, based on their understanding of the technical solutions of the present invention, can make various modifications and variations to the specific embodiments, but all such modifications and variations should fall within the scope of protection of the present invention.

[0023] The present invention provides a method for predicting the grade of zinc flotation concentrate, comprising the following steps, with reference to... Figure 1 : S1: Foam Video Acquisition and Semantic Feature Extraction This invention first uses an industrial camera to capture real-time video of the foam surface during the zinc flotation process, obtaining foam video including the foam's movement. In a preferred embodiment, the industrial camera is fixedly mounted above the flotation cell to ensure clear capture of the overall state of the foam surface. The acquired foam video is sampled at preset time intervals to obtain a continuous sequence of foam images.

[0024] During the sampling process, a fixed time interval t is used to sample frames of the foam video, and the time interval between two adjacent frames remains consistent. In a preferred embodiment of the present invention, the time interval t is set to 1 second, that is, one frame is sampled per second. For a sampling sequence containing n consecutive foam images, the corresponding semantic feature data needs to be extracted from each frame. When the sampling time interval is 1 second, the time span covered by the n frames is (n-1) seconds.

[0025] Semantic features are extracted from each frame of the foam image. The semantic features extracted in this invention include three categories: foam size features, foam color features, and texture features, totaling six specific feature parameters, which together constitute multidimensional semantic feature time-series data. These semantic features can effectively reflect the physical state of the foam surface, and the physical state of the foam surface is closely related to flotation conditions (such as feed grade, reagent addition amount, etc.). Therefore, these features can serve as important input parameters for predicting concentrate grade.

[0026] Foam size features are used to characterize the size distribution of foam. This invention employs a marked watershed-based image segmentation method to extract foam size features. The method first preprocesses the foam image, including grayscale conversion and morphological operations; then, it determines the seed points of the foam using a marking technique; next, it uses a watershed algorithm to segment the foam regions, separating interconnected foam into independent regions; finally, it calculates the area of ​​each independent foam region and uses it as the foam size feature data C1. Larger foam sizes typically indicate a higher slurry level or more pronounced bubble merging, possibly related to lower reagent concentrations or specific ore properties; smaller foam sizes may indicate stronger reagent effects or more vigorous slurry agitation.

[0027] Foam color features are used to characterize the color distribution on the foam surface. This invention employs a color space transformation method to extract foam color features. Specifically, the original foam image is converted from the RGB color space to the HSV or Lab color space, and the hue (H component or a component and b component) is extracted as the foam color feature data C2. The HSV color space separates color information into three components: hue, saturation, and lightness. Among these, the hue component effectively reflects the mineral coverage on the foam surface. Darker foam colors generally indicate a higher foam load, which may correspond to a higher concentrate grade; lighter foam colors may indicate a lower foam load.

[0028] Texture features are used to characterize the roughness and uniformity of the foam surface. This invention uses the Gray Level Co-occurrence Matrix (GLCM) method to extract four texture feature parameters. The GLCM describes the texture characteristics of an image by statistically analyzing the frequency of gray-level value pairs in a specific direction.

[0029] The first texture feature is the inverse moment, which characterizes the uniformity of local texture changes in an image. The larger the inverse moment value, the more uniform the texture.

[0030] The second texture feature is correlation, which characterizes the similarity of image textures along the row or column direction. The magnitude of the correlation value reflects the regularity of the foam surface texture.

[0031] The third texture feature is contrast, which characterizes the clarity or local variation of image texture. The higher the contrast value, the coarser the texture.

[0032] The fourth texture feature is energy. Energy represents the uniformity of the image texture and the regularity of the gray-level distribution. The larger the energy value, the more regular the texture.

[0033] The above four texture features were extracted using the gray-level co-occurrence matrix, resulting in texture feature data C3, C4, C5, and C6.

[0034] The six semantic feature data are combined sequentially to form the multidimensional semantic feature time-series data C of the current bubble video, denoted as: C = {C1, C2, C3, C4, C5, C6} Where C1 represents foam size characteristics, C2 represents foam color characteristics, C3 represents inverse difference moment, C4 represents correlation, C5 represents contrast, and C6 represents energy.

[0035] In actual flotation processes, the real-time changes in the surface state of the froth can reflect the dynamic characteristics of the flotation conditions. By continuously extracting semantic features from multiple frames of froth images to form semantic feature time-series data, the temporal evolution pattern of the froth state can be captured, providing a rich data foundation for subsequent concentrate grade prediction.

[0036] S2: Distribution Characteristics Analysis After acquiring the multidimensional semantic feature time-series data C, this invention performs distribution feature analysis on the time-series data of each semantic feature, calculating the actual distribution features used to characterize the changing trends and fluctuation patterns of each semantic feature. This invention uses the first-order difference mean and the second-order difference mean as distribution feature descriptors, which can characterize the dynamic characteristics of the semantic feature time-series data from different perspectives.

[0037] The first-order difference mean is used to characterize the average rate of change of semantic features. The formula for calculating the first-order difference mean of the j-th feature is: ; The first-order difference represents the change in semantic feature values ​​between two adjacent sampling points. A positive value indicates an upward trend in the feature value, while a negative value indicates a downward trend. The larger the absolute value, the more drastic the change. By averaging all first-order differences, we obtain the first-order difference mean, which characterizes the average rate of change of the semantic feature within the sampling period. When the first-order difference mean is close to zero, it indicates that the semantic feature remains relatively stable overall; when the first-order difference mean is significantly positive, it indicates that the semantic feature exhibits an upward trend; and when the first-order difference mean is significantly negative, it indicates that the semantic feature exhibits a downward trend.

[0038] The second-order difference mean is used to characterize the degree of fluctuation in the trend of semantic feature changes. For the j-th feature, the formula for calculating the second-order difference mean is: ; The second-order difference represents the change in the first-order difference, i.e., the change in the rate of change of semantic features. When the second-order difference is positive, it indicates that the first-order difference is increasing, meaning the rate of change is accelerating; when the second-order difference is negative, it indicates that the first-order difference is decreasing, meaning the rate of change is slowing down. The mean of the second-order difference characterizes the overall fluctuation of the trend of semantic feature change; the larger its absolute value, the more unstable and volatile the change of semantic features.

[0039] The first-order difference mean and the second-order difference mean describe the distribution characteristics of semantic feature time-series data from two different perspectives. The first-order difference mean reflects the linear trend of semantic features, while the second-order difference mean reflects the stability of this trend. This distribution feature representation method can effectively capture the dynamic changes in the flotation state, providing physically meaningful feature representations for subsequent knowledge constraints and concentrate grade prediction.

[0040] S3: Expert Knowledge Reasoning This invention constructs an expert knowledge reasoning model to infer the target distribution constraint parameters that each semantic feature should satisfy based on the current flotation operating parameters. The input of the expert knowledge reasoning model is the feed grade and the amount of flotation reagents added under the current flotation operating conditions, and the output is the target distribution constraint parameters of each semantic feature, that is, the interval range of the first-order difference mean and the interval range of the second-order difference mean of each semantic feature.

[0041] The expert knowledge reasoning model is implemented using a fuzzy logic interval reasoning model. The fuzzy logic system can effectively handle the fuzziness and uncertainty in the flotation process, encoding the operator's expert experience into rules. In a preferred embodiment of the invention, the number of rules e in the fuzzy logic system is set to 3, that is, it contains 3 fuzzy reasoning rules.

[0042] The fuzzy logic interval inference model has four input variables: ore grade, copper sulfate addition, No. 2 oil addition, and butyl xanthate addition. Ore grade reflects the properties of the incoming ore and is a key factor affecting the froth state and concentrate grade. Copper sulfate, as an activator, enhances the hydrophobicity of the mineral surface, promoting the adhesion of useful minerals to bubbles. No. 2 oil, as a foaming agent, generates stable bubbles and maintains the foam layer structure. Butyl xanthate, as a collector, selectively adheres to the surface of useful minerals, making them hydrophobic.

[0043] The output of the fuzzy logic interval inference model is the target distribution constraint parameters for each semantic feature, specifically including: the first-order difference mean interval and the second-order difference mean interval for foam size feature, the first-order difference mean interval and the second-order difference mean interval for foam color feature, the first-order difference mean interval and the second-order difference mean interval for inverse moment, the first-order difference mean interval and the second-order difference mean interval for correlation, the first-order difference mean interval and the second-order difference mean interval for contrast, and the first-order difference mean interval and the second-order difference mean interval for energy.

[0044] Compared to single-value outputs, interval outputs better reflect the fuzzy nature of flotation conditions. Flotation is a complex physicochemical process influenced by multiple factors; even with identical operating parameters, the distribution of foam states can vary within a certain range. Using interval constraints instead of point estimations provides greater tolerance for subsequent knowledge constraints.

[0045] The training process for the expert knowledge reasoning model is as follows: First, historical foam video data is collected. k historical foam video samples are obtained, each containing a complete sequence of foam images and corresponding flotation parameters.

[0046] Then, steps S1 and S2 are performed on each historical foam video sample to extract semantic feature time-series data and calculate distribution features, obtaining the mean of 6 first-order differences and the mean of 6 second-order differences for each sample. At the same time, the ore grade value and the dosage of three reagents (copper sulfate, No. 2 oil, and butyl xanthate) corresponding to each sample are obtained.

[0047] Next, the fuzzy logic interval reasoning model was trained using the ore grade and the dosage of three reagents (a total of four input variables) as model inputs, and six first-order difference means and six second-order difference means (a total of twelve output variables) as model outputs. During training, the model learned the mapping relationship between ore grade, reagent dosage, and semantic feature distribution, and encoded and stored expert experience knowledge in the form of fuzzy rules.

[0048] After the model training is completed, the current ore grade and the amount of the three reagents added are input into the trained fuzzy logic interval reasoning model, and the target distribution constraint parameters of each semantic feature under the current working condition can be inferred, that is, the interval range of the first-order difference mean and the interval range of the second-order difference mean of each semantic feature.

[0049] S4: Weighted Evaluation and Reconstruction of Distribution Consistency This invention performs a weighted evaluation of the distribution consistency of multidimensional semantic feature time series data based on the deviation between the actual distribution characteristics and the target distribution constraint parameters, generates matching weights for each data point, and reconstructs the time series data based on the matching weights to obtain non-equal time interval semantic feature time series data.

[0050] The core idea of ​​distribution consistency weighted evaluation is that data points in time series data whose distribution characteristics match the target distribution constraint parameters are more likely to belong to the current operating conditions and should be retained or sampled more thoroughly; while data points with a lower matching degree may be outliers caused by noise interference or do not belong to the current operating conditions and should be weakened or downsampled.

[0051] The specific process for weighted evaluation of distribution consistency is as follows: First, for each semantic feature data point, the deviation between its actual distribution feature and the target distribution constraint parameters is calculated. The deviation is calculated as follows: for each data point, the first-order difference mean and the second-order difference mean are calculated based on its neighboring data points, and then compared with the interval boundaries in the target distribution constraint parameters to calculate the degree of deviation between the actual distribution feature and the target distribution interval.

[0052] Then, a matching weight function is constructed based on the deviation value. The matching weight function satisfies the following properties: the smaller the deviation value, the larger the corresponding matching weight; the larger the deviation value, the smaller the corresponding matching weight. The matching weight function can take various forms, such as linear functions, exponential functions, or Gaussian functions. The output range of the matching weight function is usually normalized to the interval [0,1], with a weight value of 1 indicating a perfect match and a weight value of 0 indicating a complete mismatch.

[0053] Next, the semantic feature time-series data is weighted and reconstructed based on the matching weights. Specifically, for high-weight data points, the probability of them being selected into the reconstruction sequence or their contribution weight in the model input is increased; for low-weight data points, the probability of them being selected into the reconstruction sequence is decreased or they are downsampled. In the final reconstruction sequence, the time intervals between different data points are no longer equal, forming non-equal time interval semantic feature time-series data.

[0054] In a preferred embodiment of the present invention, when the matching weights are in binary form (i.e., the weight values ​​are only 0 or 1), the weighted evaluation of distribution consistency is equivalent to proportional reduction according to a preset reduction ratio. Preferably, the reduction ratio is set to 0.4, that is, 40% of the data points with the lowest matching degree are deleted, and 60% of the data points with the highest matching degree are retained. Because some data points in the middle positions are deleted, the reconstructed semantic feature sequence is no longer at equal time intervals, naturally forming a non-equal interval sequence. .

[0055] In this embodiment, the knowledge constraint mechanism effectively removes interfering data points that do not belong to the current operating conditions, improving the quality of the input data. Specifically, the flotation site environment is complex and may have interfering factors such as bubble bursting, reagent spraying, and equipment vibration. These interferences may cause the features of the sampled flotation image to deviate from the actual operating conditions. Through knowledge constraints, these abnormal data points can be identified and eliminated, making the final data sequence used for prediction more accurately reflect the current flotation operating conditions.

[0056] S5: Time Series Forecasting This invention employs a GRU-D model based on a combination of gated cyclic units and a time decay mechanism as a time series prediction model to process time series data with semantic features of non-equal time intervals and output concentrate grade prediction results.

[0057] The GRU-D model, short for Gated Recurrent Unit with Decay, is an extended version of the standard GRU model. Its core improvement lies in the introduction of a time decay mechanism, enabling it to effectively handle non-uniformly spaced time series data. In the standard GRU model, the hidden state is updated at every time step. However, for non-uniformly spaced data, the time intervals between adjacent time steps may differ; the longer the time interval, the more historical information decays. The GRU-D model explicitly models the impact of time intervals on the hidden state through a decay factor, allowing the model to better handle time series data with uneven time intervals.

[0058] The decay mechanism of the GRU-D model is proportional to the time interval; that is, the longer the time interval between adjacent data points, the greater the decay of the hidden state; the shorter the time interval, the higher the retention of the hidden state. This mechanism conforms to the general rule in time series data that "recent information is more important," and it can also accurately reflect the characteristics of non-equidistant sampling.

[0059] In a preferred embodiment of the present invention, the number of network layers in the GRU-D model is set to 3. Considering the balance between the number of model parameters and computational complexity, the 3-layer network structure can ensure the expressive power of the model while avoiding the problems of overfitting and excessive computational overhead.

[0060] The input to the GRU-D model consists of two parts: non-uniformly spaced semantic feature time-series data and corresponding time interval data. The semantic feature time-series data contains six feature dimensions, and the time interval data records the time interval length between adjacent data points. The model output is the predicted zinc flotation concentrate grade p at the current moment.

[0061] The training process of the GRU-D model is as follows: First, collect k historical bubble video samples.

[0062] Then, for each historical bubble video sample, steps S1 to S4 are performed: extract semantic feature time series data, calculate distribution features, obtain target distribution constraint parameters through fuzzy logic interval inference model, perform distribution consistency weighted evaluation, and obtain non-equal interval semantic feature time series data.

[0063] Next, the GRU-D model is trained using non-equidistant semantic feature time-series data and their corresponding time intervals as model input, and the corresponding actual measured values ​​of concentrate grade as model output. During training, the model learns the mapping relationship between non-equidistant semantic feature sequences and concentrate grade.

[0064] After the model training is completed, the non-equal interval semantic feature sequence obtained by processing the foam video collected at the current time according to steps S1 to S4 is input into the trained GRU-D model, and the model can output the predicted value of zinc flotation concentrate grade at the current time in real time.

[0065] S6: Real-time Prediction Process The zinc flotation concentrate grade prediction method of the present invention can predict the concentrate grade in real time during the current flotation process after model training is completed. The real-time prediction process includes the following steps: First, acquire the foam video at the current moment. An industrial camera captures video data of the foam surface in the flotation cell in real time and transmits the video stream to the processing terminal.

[0066] Then, the acquired foam video is sampled according to step S1. The foam video is sampled frame by frame at a preset time interval t (preferably 1 second) to obtain a continuous foam image sequence.

[0067] Next, semantic feature extraction and distribution feature analysis are performed on the sampled foam image sequence according to step S2. The foam size feature, foam color feature, and four texture features of each frame image are extracted to form multidimensional semantic feature time series data; the first-order difference mean and second-order difference mean of each semantic feature are calculated to obtain the actual distribution features.

[0068] Then, obtain the current flotation operating parameters, including the feed grade and the amount of three reagents added (copper sulfate, No. 2 oil, and butyl xanthate), and obtain the target distribution constraint parameters through the fuzzy logic interval reasoning model according to step S3.

[0069] Next, a distribution consistency weighted evaluation is performed according to step S4 to reconstruct the semantic feature time series data, resulting in non-equal interval semantic feature time series data.

[0070] Finally, the non-equal interval semantic feature sequence and its time interval are input into the trained GRU-D model, and the predicted value of zinc flotation concentrate grade at the current time is output according to step S5.

[0071] The real-time prediction process can be executed cyclically, updating the prediction results every preset time interval, thereby achieving continuous real-time monitoring of concentrate grade. Compared with traditional fluorescence analyzer detection, the method of this invention has advantages such as short detection cycle (down to the second level), low cost, and no need for frequent maintenance, and can provide timely grade reference information for flotation operations.

[0072] Corresponding to the above-mentioned method for predicting the grade of zinc flotation concentrate, this invention also provides a system for predicting the grade of zinc flotation concentrate. For example... Figure 2 As shown, the system includes the following six functional modules: The foam acquisition module is used to acquire foam video during the zinc flotation process and samples it at preset time intervals to obtain a foam image sequence. The core hardware of this module is an industrial camera, mounted above the flotation cell, which acquires video of the foam surface in real time. The module also includes a video sampling unit that extracts keyframe images from the video stream at set time intervals.

[0073] The semantic feature extraction module is used to extract semantic features from the foam image sequence, obtaining multi-dimensional semantic feature time-series data. This module uses image processing algorithms to analyze each frame of the foam image, extracting foam size features, foam color features, and four texture features. The foam size feature is extracted based on the marker watershed segmentation algorithm, the foam color feature is extracted through color space transformation, and the texture features are calculated using the gray-level co-occurrence matrix. The output of the semantic feature extraction module is 6-dimensional semantic feature time-series data.

[0074] The distribution feature analysis module is used to perform distribution feature analysis on multidimensional semantic feature time series data to obtain the actual distribution features. This module calculates the first-order difference mean and the second-order difference mean for each semantic feature time series data. The first-order difference mean represents the average rate of change of the semantic feature, and the second-order difference mean represents the degree of fluctuation in the trend of semantic feature change. The output of the distribution feature analysis module is 12-dimensional distribution feature data.

[0075] The knowledge reasoning module is used to obtain the feed grade and flotation reagent dosage under the current flotation conditions, and outputs the target distribution constraint parameters of semantic features based on an expert knowledge reasoning model. This module stores a trained fuzzy logic interval reasoning model, takes the feed grade and the dosage of the three reagents as input, and outputs the target distribution constraint parameters (interval range) for each semantic feature. The output of the knowledge reasoning module is a 12-dimensional target distribution constraint parameter.

[0076] The weighted reconstruction module performs a weighted evaluation of the distribution consistency of semantic feature time-series data based on the deviation between the actual distribution characteristics and the target distribution constraint parameters. It then reconstructs the data based on the matching weights to obtain a non-uniformly spaced semantic feature sequence. This module calculates the matching weight for each data point, performs weighted reconstruction sampling on the time-series data, deletes or downsamples data points with low matching degrees, and retains or strengthens the sampling of data points with high matching degrees. The output of the weighted reconstruction module is the non-uniformly spaced semantic feature sequence and its time interval data.

[0077] The prediction module is used to input non-equidistant semantic feature sequences and their corresponding time intervals into the time series prediction model, and outputs the predicted value of the zinc flotation concentrate grade at the current time. This module stores the trained GRU-D model, inputs the reconstructed non-equidistant sequence into the model, and outputs the concentrate grade prediction result in real time. The output of the prediction module is the predicted value of the zinc flotation concentrate grade.

[0078] The modules work collaboratively to achieve a complete process from foam video acquisition to concentrate grade prediction. The foam acquisition module transmits the acquired video data to the semantic feature extraction module; the feature data extracted by the semantic feature extraction module is simultaneously transmitted to the distribution feature analysis module and the weighted reconstruction module; the calculation results of the distribution feature analysis module and the output of the knowledge reasoning module are jointly input into the weighted reconstruction module; the output of the weighted reconstruction module is finally transmitted to the prediction module to generate the concentrate grade prediction result.

[0079] The modules described above can be software modules, hardware modules, or a combination of both. Software modules can be deployed and executed on a processor, while hardware modules may include dedicated image processing chips, FPGAs, etc. The specific implementation of each module can be flexibly configured according to the actual application scenario and system resources.

[0080] The present invention also provides a zinc flotation concentrate grade prediction device, which includes a processor and a memory. The memory stores a computer program, and the processor executes the computer program to implement the steps of the above-described zinc flotation concentrate grade prediction method.

[0081] This electronic device can be used as a standalone prediction terminal or as part of an automated flotation control system, working in conjunction with other equipment. The hardware architecture of the electronic device includes a processor, memory, bus, and external device interfaces.

[0082] The processor is the core computing unit of the electronic device, responsible for executing computer programs stored in memory to implement algorithmic functions such as foam video acquisition, semantic feature extraction, distribution feature analysis, knowledge reasoning, weighted reconstruction, and time series prediction. The processor selection should meet the real-time requirements of the flotation site and possess sufficient computing performance.

[0083] The memory is used to store computer programs and various types of data. The computer programs include the operating system, application programs, and the core algorithm code that implements the method of this invention. The stored data includes historical bubble video data, model parameters, configuration information, etc. The memory can include volatile memory (such as RAM) and non-volatile memory (such as ROM, hard disk), etc.

[0084] The bus connects the processing unit and the storage unit, enabling data transfer between the components. The bus includes a data bus, an address bus, and a control bus, supporting the processor's read and write access to memory and data exchange with external devices.

[0085] The external device interface is used to connect external devices such as industrial cameras, displays, and communication modules. The industrial camera transmits the acquired foam video data to the electronic device through the interface; the display screen is used to show real-time prediction results and system status information; and the communication module is used to interact with other control systems or host computers.

[0086] The electronic equipment of this invention can be applied to various flotation site environments, and the selection and configuration of the equipment can be adapted to the site conditions. The operating environment of the equipment should meet the industrial site requirements in terms of temperature, humidity, electromagnetic compatibility, etc.

[0087] The present invention also provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps of the above-described zinc flotation concentrate grade prediction method.

[0088] Computer-readable storage media may include, but are not limited to: hard disk, solid-state drive, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM), electrically erasable programmable read-only memory (EEPROM), flash memory, magnetic storage device, optical storage device, or any combination thereof.

[0089] The computer program of this invention can exist and be distributed in the form of a program product. The program product can be stored on a computer-readable storage medium or distributed via network download. The program product can be a standalone installation package or a component module embedded in other software products.

[0090] When the computer program is executed by the processor, it implements all the steps of the method of the present invention, including: foam video acquisition and sampling, extraction of semantic features of foam images, distribution feature analysis of semantic feature time series data, expert knowledge reasoning based on ore grade and reagent addition amount, weighted evaluation of distribution consistency and reconstruction of time series data, and concentrate grade prediction based on GRU-D model.

[0091] The storage medium implementation of the present invention enables the zinc flotation concentrate grade prediction method to be deployed and run in software form, reducing the system's hardware dependence and facilitating the promotion and application of the method.

[0092] Those skilled in the art will understand that various modifications and combinations can be made to the above-described specific embodiments without departing from the principles of the present invention. For example: Regarding the sampling interval parameter, the sampling time interval of the foam video can be adjusted according to the actual application requirements. A shorter sampling interval can capture more subtle changes in the foam state, but it will increase the amount of data processing and computational burden; a longer sampling interval can reduce the system load, but may lose some timeliness information.

[0093] The number of rules in a fuzzy logic model can be adjusted based on the scale and distribution characteristics of the training data. Too few rules may result in insufficient model expressive power, while too many rules may lead to overfitting.

[0094] Regarding the form of the matching weight function, linear, exponential, Gaussian, or other forms can be chosen according to actual needs. Different weight functions will affect the data reconstruction effect.

[0095] Regarding the network structure of the GRU-D model, parameters such as the number of network layers and the number of hidden units can be adjusted to balance model performance and computational efficiency. All such modifications and extensions should fall within the scope of protection of this invention.

[0096] The zinc flotation concentrate grade prediction method provided by this invention acquires real-time foam video using an industrial camera, extracts semantic features from the foam images, determines target distribution constraints by combining expert knowledge reasoning, performs knowledge-constrained denoising on the semantic feature time-series data, and finally uses a GRU-D model for time-series prediction. Compared with traditional concentrate grade prediction methods, this invention effectively introduces knowledge constraints, improves the quality of input data, and enables real-time and accurate prediction of concentrate grade. This provides timely and accurate reference information for flotation operations, helps maintain the stability of the flotation process, and improves the economic efficiency of flotation plants.

[0097] The specific embodiments of the present invention have been described in detail with reference to the accompanying drawings. However, it should be understood that the scope of protection of the present invention is not limited to these specific embodiments. Any modifications, substitutions, or alterations made by those skilled in the art within the spirit and principles of the present invention should fall within the scope of protection of the present invention.

Claims

1. A method for predicting the grade of zinc flotation concentrate, characterized in that, Includes the following steps: Acquire foam videos of the zinc flotation process and sample the foam videos at preset time intervals to obtain a continuous foam image sequence; Semantic features are extracted from each frame of the foam image sequence to obtain multidimensional semantic feature time-series data including foam size features, foam color features, and texture features. The distribution characteristics of the multidimensional semantic feature time series data are analyzed to obtain the actual distribution characteristics used to characterize the changing trends and fluctuation patterns of each semantic feature; Obtain the feed grade and flotation reagent addition amount under the current flotation conditions, and input the feed grade and flotation reagent addition amount into a pre-trained expert knowledge reasoning model to output the target distribution constraint parameters corresponding to each semantic feature under the current conditions. Based on the deviation between the actual distribution characteristics and the target distribution constraint parameters, the distribution consistency weighted evaluation is performed on the multidimensional semantic feature time series data to generate matching weights for each data point, and the time series data is reconstructed based on the matching weights to obtain non-equal time interval semantic feature time series data. The non-equal time interval semantic feature time series data and their corresponding time intervals are input into the time series prediction model, and the predicted value of zinc flotation concentrate grade at the current moment is output.

2. The method according to claim 1, characterized in that, The semantic features include: Foam size features, foam color features, and texture features extracted based on the gray-level co-occurrence matrix.

3. The method according to claim 2, characterized in that, The methods for extracting the semantic features include: Foam size features were obtained using an image segmentation method based on labeled watersheds; The color characteristics of the foam are obtained through color space transformation; Texture features are calculated using the gray-level co-occurrence matrix, including inverse difference moment, correlation, contrast, and energy.

4. The method according to claim 1, characterized in that, The actual distribution characteristics include the first-order difference mean and the second-order difference mean, wherein: The first-order difference mean is used to characterize the average rate of change of semantic features; The second-order difference mean is used to characterize the degree of fluctuation in the trend of semantic feature changes.

5. The method according to claim 1, characterized in that: The expert knowledge reasoning model is a fuzzy logic interval reasoning model. Its output includes the first-order difference mean interval and the second-order difference mean interval of each semantic feature, which are used to characterize the dynamic distribution range constraint of the semantic features under the current working condition.

6. The method according to claim 1, characterized in that, The weighted evaluation of distribution consistency includes: For each semantic feature data point, calculate the deviation between its actual distribution characteristics and the target distribution constraint parameters; A matching weight function is constructed based on the deviation value, and the matching weight function satisfies the following: the smaller the deviation value, the greater the matching weight; the larger the deviation value, the smaller the matching weight. The semantic feature time series data is weighted and reconstructed based on the matching weights, so that high-weight data points are retained or enhanced, and low-weight data points are weakened or downsampled, thereby generating non-equal time interval semantic feature time series data.

7. The method according to claim 1, characterized in that, The time-series prediction model is a GRU-D model based on a combination of gated cyclic units and a time decay mechanism. It is used to process time-series data with semantic features of non-equal time intervals and output concentrate grade prediction results.

8. The method according to claim 7, characterized in that: The training process of the expert knowledge reasoning model includes: acquiring historical foam video data and extracting corresponding semantic feature time series data and distribution features; acquiring historical ore grade and historical flotation reagent addition amount; using the ore grade and reagent addition amount as input and the semantic feature distribution features as output, training the fuzzy logic interval reasoning model. The training process of the GRU-D model includes: generating semantic feature time series data based on historical bubble videos; using the expert knowledge reasoning model to perform distribution consistency weighted reconstruction on the semantic feature time series data to obtain non-uniform time interval semantic feature time series data; and using the non-uniform time interval semantic feature sequence and time interval as input and the corresponding concentrate grade as output to train the GRU-D model.

Images