Flotation froth state identification method

Abstract

The invention provides a flotation froth state recognition method based on bimodal fusion. A dual-mode parallel detection framework of a shadow area analysis method and a ResNet deep learning model is innovatively constructed by collecting flotation cell video stream data in real time and adopting an image preprocessing module to eliminate environmental interference and extract a liquid level target area; the abnormal quantitative scoring of the running working condition is realized; and the second detection module extracts deep visual features based on a pre-trained ResNet model, and outputs a sinking groove abnormal probability score. After bimodal detection results are fused through a dynamic weight distribution strategy, a fuzzy logic algorithm is adopted to construct a third-order membership function and an inference rule to carry out score smooth optimization, and finally an output anomaly type is judged in combination with a threshold value. The method breaks through the sensitivity defect of a single detection threshold, and solves the problems of low efficiency of traditional manual observation and large error of single-mode detection.

Description

Technical Field

[0001] The present invention relates to a method for identifying the state of flotation foam, belonging to the technical field of flotation. Background Art

[0002] In the flotation beneficiation process, the foam state plays a key role in the separation efficiency of useful minerals and waste rocks in the pulp, directly affecting the beneficiation efficiency, production cost control, and environmental protection. The traditional flotation process realizes the refined control of the beneficiation process by monitoring the foam state in real time and dynamically adjusting the liquid level, air intake, and reagent addition of the flotation cell. However, due to the non-rigid, viscous, and variable dynamic characteristics of the liquid surface and foam in the flotation cell, accurately determining abnormal working conditions has always been a difficult point restricting the improvement of the automation and intelligent level of the flotation process.

[0003] At present, there are mainly two methods for detecting abnormal working conditions: one is the shadow area detection method, which monitors the change of the liquid surface shadow through image segmentation and area calculation, mainly used to judge the "overflow abnormal working condition"; the other is to use the ResNet deep learning algorithm to extract and classify the features of the liquid surface image, and then identify the "subsidence abnormal working condition". These two methods can achieve high detection accuracy in their respective specific scenarios. For example, in the overflow working condition, there is more pulp in the flotation cell, and the shadow area of the liquid surface decreases significantly. The shadow area detection method can quickly capture the abnormal signal; while in the subsidence working condition, due to the decrease in foam density and the increase in liquid surface brightness, the detection method based on ResNet can effectively distinguish subtle feature differences and complete the abnormal determination.

[0004] However, the actual working conditions often have complex and variable transitional or boundary states. At this time, the two detection methods may output inconsistent detection results due to different focuses on processing grayscale, texture, and edge information in the on-site image, resulting in fuzzy judgment of the abnormal state by the system. For example, during the transition of the flotation cell liquid surface from the overflow state to the normal state, the shadow area detection may not reach the abnormal threshold, while the ResNet detection may output an abnormal signal in advance due to local feature mutations; conversely, in the subsidence boundary state, there may also be similar judgment differences between the two. This problem of inconsistent output of multiple algorithms not only affects the accurate identification of abnormal working conditions but also may mislead the automatic adjustment of subsequent process parameters, thereby reducing the overall flotation efficiency and having an adverse impact on production costs and environmental protection.

[0005] Therefore, how to achieve consistent output between the shadow area detection and the ResNet deep learning detection on the premise of ensuring the high accuracy of individual detection methods has become a technical contradiction urgently to be solved in the prior art. Summary of the Invention

[0006] The technical problem to be solved by the present invention is to design a method for identifying the flotation foam state to overcome the deficiencies of the prior art.

[0007] The technical solution of the present invention is to provide a method for identifying the flotation foam state, and the method includes the following steps:

[0008] S1. Obtain the real-time video image of the foam in the flotation cell;

[0009] S2. Preprocess the video image, and the preprocessing includes noise removal, image enhancement, and extraction of the region of interest;

[0010] S3. Synchronously use the first anomaly detection module and the second anomaly detection module to perform preliminary anomaly condition detection on the preprocessed image, where:

[0011] The first anomaly detection module extracts the liquid surface shadow by using the shadow area detection method, and outputs the running groove anomaly condition detection score Score_shadow based on the comparison between the liquid surface shadow area and the preset threshold; the second anomaly detection module uses the ResNet deep learning algorithm to extract features from the liquid surface image and outputs the sinking groove anomaly condition detection score Score_ResNet;

[0012] S4. According to the preset weight factors W1 and W2, perform weighted fusion on the detection scores of the first anomaly detection module and the second anomaly detection module, and calculate the preliminary comprehensive anomaly score S;

[0013] S5. Perform smoothing processing on the preliminary comprehensive anomaly score S by using the fuzzy logic algorithm to obtain the smoothed comprehensive score S';

[0014] S6. According to the smoothed comprehensive score S' and the preset threshold, determine whether there is an abnormal condition in the flotation cell, and determine the type of anomaly based on the relative detection scores of the first and second anomaly detection modules.

[0015] Further, the shadow area detection method adopted by the first anomaly detection module includes segmenting the liquid surface shadow by using image binarization, morphological erosion and dilation processing, calculating the shadow area based on the segmentation result, and outputting the detection score Score_shadow after comparing with the preset threshold.

[0016] Further, the specific method of using the ResNet deep learning algorithm to extract features from the liquid surface image and output the sinking groove anomaly condition detection score Score_ResNet is as follows:

[0017] 1) Extract the liquid surface area or the target area from the preprocessed liquid surface image;

[0018] 2) Uniformly scale the extracted ROI images to a fixed size;

[0019] 3) Normalize the pixel values of the scaled images, and subtract the mean and divide by the standard deviation according to the requirements of the pre-trained model, so that the image data distribution is consistent with that during training;

[0020] 4) Input the normalized images into the pre-trained ResNet model, and extract deep features through multiple convolutional layers, residual blocks, and global average pooling layers in the model;

[0021] 5) At the last layer of the ResNet model, the fully connected classifier maps the extracted feature vectors to the probability distributions of each category;

[0022] 6) Through the Softmax activation function, output the prediction probabilities of each category, where the probability value corresponding to the abnormal condition of the settling tank is the preliminary score Score_ResNet.

[0023] Furthermore, the calculation method of the preliminary comprehensive anomaly score S is as follows:

[0024] S = W1 * Score_shadow + W2 * Score_ResNet

[0025] where W1 is the weight factor corresponding to Score_shadow, and W2 is the weight factor corresponding to Score_ResNet

[0026] corresponding weight factor.

[0027] Furthermore, the specific method for smoothing the preliminary comprehensive anomaly score using the fuzzy logic algorithm to obtain the smoothed comprehensive score S' is as follows:

[0028] a. Construct an input membership function to map the input S to three fuzzy sets: low anomaly, medium anomaly, and high anomaly;

[0029] b. Construct an output membership function to map the output ΔS to three fuzzy sets: negative adjustment, zero adjustment, and positive adjustment;

[0030] c. Formulate fuzzy inference rules,

[0031] Rule 1: If S is low anomaly, then ΔS = negative adjustment,

[0032] Rule 2: If S is medium anomaly, then ΔS = zero adjustment,

[0033] Rule 3: If S is high anomaly, then ΔS = positive adjustment;

[0034] d. Perform fuzzy inference,

[0035] For a given input S, calculate its membership degrees in three fuzzy sets of low anomaly, medium anomaly, and high anomaly: μ low (S), μ medium (S), and μ high (S);

[0036] For each rule, according to the input membership degrees, clip the corresponding output fuzzy set using the minimum T-norm to obtain the output fuzzy subset of each rule;

[0037] e. Aggregate the output fuzzy subsets of all rules using the maximum operation to form the overall output fuzzy set, that is, aggregate the membership functions of all ΔS;

[0038] f. Use the centroid method to calculate the central value of the aggregated output fuzzy set,

[0039] ΔS crisp = ∫ a b x·μ aggregated (x)dx / ∫ a b μ aggregated (x)dx

[0040] where ΔS crisp represents the central value of the fuzzy set, μ aggregated (x) represents the membership degree of the aggregated output fuzzy set at x, a is the minimum value of the output ΔS, and b is the maximum value of the output ΔS;

[0041] g. Calculate the smoothed comprehensive score S′, S′ = S + ΔS crisp , if S′ < 0, then set S′ = 0; if S′ > 1, then set S′ = 1.

[0042] The beneficial effects of the present invention are: Compared with the prior art,

[0043] 1) The present invention respectively determines the "running groove" and "sinking groove" anomalies by combining the shadow area detection and the ResNet deep learning algorithm, realizing the complementary advantages of the two detection methods, and significantly improving the detection accuracy and reliability of abnormal working conditions;

[0044] 2) The present invention eliminates the problem of inconsistent output results of each detection module under the working condition transition or boundary state by introducing the comprehensive determination based on weight fusion and the fuzzy logic smoothing process, making the system's judgment of abnormal states continuous and stable, thereby providing an accurate and unified basis for subsequent process control. BRIEF DESCRIPTION OF THE DRAWINGS

[0045] Figure 1 is the flow chart of the present invention. Detailed implementation mode

[0046] To make the objectives, technical solutions and advantages of the present invention clearer, the present invention will be further described in detail below with reference to the drawings of the present specification.

[0047] Reference Figure 1 , this embodiment provides a method for identifying the state of flotation foam, including the following steps: S1. Acquisition of real-time video images

[0048] An industrial high-definition camera is installed above the flotation cell to collect real-time video images of the foam surface in the flotation cell. The resolution of the camera is not less than 1920×1080, and dust-proof and vibration-proof designs are combined to adapt to environmental factors such as dust and vibration that may exist in the industrial production site; at the same time, an adjustable lens focal length and aperture are adopted, and clear images of the liquid level of the flotation cell can be collected under different lighting conditions, laying a good foundation for subsequent image analysis.

[0049] S2. Image preprocessing

[0050] The acquired real-time video images usually have problems such as noise interference and uneven illumination. To improve the recognition accuracy of the subsequent detection module for abnormal working conditions, it is necessary to preprocess the images:

[0051] · Denoising processing: Gaussian filtering or median filtering is used to reduce random noise in the image while trying to retain

[0052] key detail information such as the foam edge;

[0053] · Image enhancement: Methods such as histogram equalization and contrast stretching are used to correct the brightness and contrast of the preprocessed image to highlight the foam edge or feature area;

[0054] · Edge detection: The Canny edge detection algorithm is used on the enhanced image to extract the main contour information

[0055] to provide accurate boundary information for subsequent target segmentation;

[0056] · ROI extraction: According to the installation position of the on-site camera and the geometric layout of the flotation cell, a region of interest (ROI) is preset to eliminate irrelevant background or mechanical structure parts, so as to reduce the calculation amount and improve the recognition efficiency.

[0057] S3. Preliminary detection of abnormal working conditions

[0058] In this embodiment, the first abnormal detection module and the second abnormal detection module are synchronously used to judge the abnormality of the preprocessed image, respectively aiming at two main types of abnormalities: "overflow" and "subsidence":

[0059] · The first abnormal detection module (shadow area detection)

[0060] By performing binarization on the image, the shadow area of the flotation foam liquid surface is extracted. Then, morphological erosion and dilation operations are used to remove discrete noise points and small artifacts, and the shadow area of the liquid surface is calculated based on the area of the shadow connected domain. The shadow area is compared with a preset threshold, and the detection score Score_shadow for the abnormal trough-running condition is output. If the shadow area is significantly smaller than the set threshold, it means that the pulp level in the trough is too high and there may be a risk of trough running, and the Score_shadow value will increase accordingly.

[0061] · Second anomaly detection module (ResNet deep learning algorithm)

[0062] After uniformly scaling and normalizing the pixel values of the preprocessed liquid surface image, it is input into a pre-trained ResNet network. This network consists of multiple convolutional layers, residual blocks, and a global average pooling layer, which can extract the deep features of the foam liquid surface and obtain the prediction probabilities of each abnormal category at the output layer. The probability value of the "abnormal sinking trough condition" is Score_ResNet. When the foam is sparse, the brightness increases, and obvious "sinking trough" features appear, the corresponding output probability of Score_ResNet will increase significantly.

[0063] S4. Weighted fusion of detection scores

[0064] Aiming at the differences in sensitivity between the "trough running" and "sinking trough" detection methods under different working conditions, in this embodiment, two weight factors W1 and W2 are set to perform weighted fusion on the detection scores of the first anomaly detection module and the second anomaly detection module respectively, and a preliminary comprehensive anomaly score SSS is obtained. Its calculation formula is:

[0065] S = W1 * Score_shadow + W2 * Score_ResNet

[0066] Among them, W1 is the weight factor for the abnormal trough-running condition, and W2 is the weight factor for the abnormal sinking trough condition. By determining appropriate weights in combination with on-site experience or referring to historical statistical data, the results of the two detection methods can be more reasonably fused in the comprehensive dimension.

[0067] S5. Fuzzy logic smoothing processing

[0068] Aiming at the characteristics of the variable working conditions of the flotation cell and the easy occurrence of transition states in actual production, in order to avoid drastic fluctuations or inconsistencies in the two detection results in the junction or transition area, in this embodiment, a fuzzy logic algorithm is used to smooth the preliminary comprehensive anomaly score S to obtain the smoothed comprehensive score S'. The specific steps are as follows:

[0069] a. Construct the input membership functions to map the input S to three fuzzy sets: low anomaly, medium anomaly, and high anomaly respectively;

[0070] b. Construct the output membership functions to map the output ΔS to three fuzzy sets: negative adjustment, zero adjustment, and positive adjustment respectively;

[0071] c. Formulate the fuzzy inference rules,

[0072] Rule 1: If S is low anomaly, then ΔS = negative adjustment,

[0073] Rule 2: If S is medium anomaly, then ΔS = zero adjustment,

[0074] Rule 3: If S is high anomaly, then ΔS = positive adjustment;

[0075] d. Conduct fuzzy inference,

[0076] For the given input S, calculate its membership degrees in the three fuzzy sets of low anomaly, medium anomaly, and high anomaly: μ low (S), μ medium (S) and μ high (S);

[0077] For each rule, according to the input membership degree, trim the corresponding output fuzzy set using the minimum T-norm to obtain the output fuzzy subset of each rule;

[0078] e. Aggregate the output fuzzy subsets of all rules using the maximum operation to form the overall output

[0079] fuzzy set, that is, aggregate the membership functions of all ΔS;

[0080] f. Use the centroid method to calculate the central value of the aggregated output fuzzy set,

[0081] ΔS crisp =∫ a b x·μ aggregated (x)dx / ∫ a b μ aggregated (x)dx

[0082] where, ΔS crisp represents the central value of the fuzzy set, μ aggregated (x) represents the membership degree of the aggregated output fuzzy set at x, a is the minimum value of the output ΔS, and b is the maximum value of the output ΔS;

[0083] g. Calculate the smoothed comprehensive score S′, S′ = S + ΔS crisp, if S′ < 0, then set S′ = 0; if S′ > 1, then set S′ = 1.

[0084] S6. Abnormality determination and type identification

[0085] When the smoothed comprehensive score S′ exceeds the set threshold, it is determined that there is an abnormal working condition; if S′ is lower than the threshold, it is considered that the system is in a normal state. If it is determined to be an abnormal working condition, then according to the relative strength of the first anomaly detection module (Score_shadow) and the second anomaly detection module (Score_ResNet), it is identified whether it is "groove running" or "groove sinking". For example, when Score_shadow is significantly higher than Score_ResNet, it is usually determined as a groove running anomaly; on the contrary, it is more inclined to a groove sinking anomaly.

[0086] Furthermore, the specific method of using the ResNet deep learning algorithm to extract features from the liquid surface image and output the detection score Score_ResNet for the groove sinking abnormal working condition is as follows:

[0087] 1) Extract the liquid surface area or target area from the preprocessed liquid surface image;

[0088] 2) Uniformly scale the extracted ROI image to a fixed size;

[0089] 3) Normalize the pixel values of the scaled image, subtract the mean value and divide by the standard deviation according to the requirements of the pre-trained model, so that the image data distribution is consistent with that during training;

[0090] 4) Input the normalized image into the pre-trained ResNet model, and extract deep features through multiple convolutional layers, residual blocks and global average pooling layers in the model;

[0091] 5) At the last layer of the ResNet model, the fully connected classifier maps the extracted feature vectors to the probability distributions of each category;

[0092] 6) Through the Softmax activation function, output the prediction probabilities of each category, where the probability value corresponding to the groove sinking abnormal working condition is the preliminary score Score_ResNet.

[0093] Furthermore, the calculation method of the preliminary comprehensive anomaly score S is as follows:

[0094] S = W1 * Score_shadow + W2 * Score_ResNet

[0095] Among them, W1 is the weight factor corresponding to Score_shadow, and W2 is the weight factor corresponding to Score_ResNet.

[0096] Where the present invention is not elaborated, it is the well-known technology of those skilled in the art. Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention rather than to limit them. Although the present invention has been described in detail with reference to the preferred embodiments, those of ordinary skill in the art should understand that the technical solutions of the present invention can be modified or equivalently replaced without departing from the purpose and scope of the technical solutions of the present invention, and they should all be covered by the scope of the claims of the present invention.

Claims

1. A flotation foam state identification method, characterized in that: The method comprises the following steps: S1, obtaining a real-time video image of the foam in the flotation tank; S2, preprocessing the video image, wherein the preprocessing includes noise removal, image enhancement, and extraction of a region of interest; S3, synchronously using the first abnormality detection module and the second abnormality detection module to perform preliminary abnormal condition detection on the preprocessed image, wherein: The first abnormality detection module uses a shadow area detection method to extract the shadow of the liquid surface, and outputs the abnormal condition detection score Score_shadow based on the comparison between the shadow area of ​​the liquid surface and the preset threshold; the second abnormality detection module uses the ResNet deep learning algorithm to extract features from the liquid surface image and outputs the abnormal condition detection score Score_ResNet of the sink; S4, performing weighted fusion on the detection scores of the first anomaly detection module and the second anomaly detection module according to preset weight factors W1 and W2, and calculating a preliminary comprehensive anomaly score S; S5, smoothing the preliminary comprehensive abnormality score S using a fuzzy logic algorithm to obtain a smoothed comprehensive score S'; S6. Determine whether the flotation cell has an abnormal operating condition based on the smoothed comprehensive score S' and a preset threshold, and determine the abnormality type based on the relative detection scores of the first and second abnormality detection modules.

2. The flotation foam state identification method according to claim 1, characterized in that: The shadow area detection method adopted by the first anomaly detection module includes segmenting the liquid surface shadow by image binarization, morphological corrosion and dilation processing, calculating the shadow area based on the segmentation result, and outputting the detection score Score_shadow after comparing it with the preset threshold.

3. The flotation foam state identification method according to claim 1, characterized in that: The specific method of using the ResNet deep learning algorithm to extract features from the liquid surface image and output the abnormal condition detection score Score_ResNet of the sink is: 1) extracting the liquid surface area or the target area from the preprocessed liquid surface image; 2) uniformly scaling the extracted ROI images to a fixed size; 3) Normalize the pixel values ​​of the scaled image, subtract the mean and divide by the standard deviation as required by the pre-trained model, so that the image data distribution is consistent with that during training; 4) Input the normalized image into the pre-trained ResNet model, and extract deep features through multiple convolutional layers, residual blocks, and global average pooling layers in the model; 5) In the last layer of the ResNet model, the fully connected classifier maps the extracted feature vectors to the probability distribution of each category; 6) Through the Softmax activation function, the predicted probability of each category is output, where the probability value corresponding to the abnormal working condition of the sink is the preliminary score Score_ResNet.

4. The flotation foam state identification method according to claim 1, characterized in that: The calculation method of the preliminary comprehensive abnormality score S is: S=W1*Score_shadow+W2*Score_ResNet Among them, W1 is the weight factor corresponding to Score_shadow, and W2 is the weight factor corresponding to Score_ResNet The corresponding weight factor.

5. The flotation foam state identification method according to claim 1, characterized in that: The specific method of using fuzzy logic algorithm to smooth the preliminary comprehensive abnormality score to obtain the smoothed comprehensive score S' is: a. Construct input membership functions and map the input S to three fuzzy sets: low anomaly, medium anomaly, and high anomaly; b. Construct the output membership function and map the output ΔS to three fuzzy sets: negative adjustment, zero adjustment and positive adjustment; c. Formulate fuzzy reasoning rules, Rule 1: If S is low abnormality then ΔS = negative adjustment, Rule 2: If S is medium abnormal, ΔS = zero adjustment, Rule 3: If S is abnormally high, ΔS = positive adjustment; d. Perform fuzzy reasoning, For a given input S, calculate its membership in the three fuzzy sets of low anomaly, medium anomaly and high anomaly: μ low (S), μ medium (S) and μ high (S); For each rule, the corresponding output fuzzy set is pruned using the minimum T-norm according to the input membership, and the output fuzzy subset of each rule is obtained; e. Aggregate the output fuzzy subsets of all rules using the maximum operation to form the overall output fuzzy set, that is, aggregate the membership functions of all ΔS; f. Use the centroid method to calculate the center value of the aggregate output fuzzy set. Among them, ΔS crisp Represents the center value of the fuzzy set, μ aggregated (x) represents the membership of the aggregated output fuzzy set at x, a is the minimum value of the output ΔS, and b is the maximum value of the output ΔS; g. Calculate the smoothed comprehensive score S', S' = S + ΔS crisp , if S′<0, set S′=0; if S′>1, set S′=1.

Images