A classification method and system for mineral flotation foam images based on deep learning

Abstract

The present invention belongs to the field of mineral flotation technology and discloses a method and system for classifying mineral flotation foam images based on deep learning. The method includes: constructing a mineral flotation foam preliminary classification model for classifying mineral flotation foam images based on a deep learning network; using a laser camera to acquire mineral flotation foam images at different time periods during production, and using the extracted parameters of denoised standard texture features, color features, and underlying characteristics of foam size distribution as an input set to the mineral flotation foam preliminary classification model; training the mineral flotation foam preliminary classification model to obtain clear features of mineral flotation foam images at different time periods, and then using an improved image classification method to classify them, and obtaining the mineral flotation foam operating status at different time periods during production based on the classification results. The present invention improves concentrate grade and mineral recovery rate by approximately 20%.

Description

Technical Field

[0001] The present invention belongs to the technical field of mineral flotation, and in particular relates to a method and system for classifying mineral flotation foam images based on deep learning. Background Art

[0002] The mineral flotation production process is typically controlled by experienced workers who observe the state of the flotation foam. This operational uncertainty makes it difficult to maintain optimal flotation operation. Using digital image processing technology to classify and interpret flotation foam images, obtain operating information for mineral flotation, and then optimize control, is an effective method for improving economic efficiency. Current research on flotation foam image classification and recognition primarily describes the flotation image by extracting underlying characteristic parameters such as texture, color, and foam size distribution. Existing technologies then use methods such as neural networks or support vector machines for classification and recognition, thereby deriving operating conditions to guide production.

[0003] However, processing the image as a whole and describing the foam image solely using its underlying features makes it difficult for humans to understand. This means that classification results based on underlying image features suffer from a semantic gap. Furthermore, because local information in foam images is not utilized, and industrial sites are subject to noise interference from lighting, dust, and other factors, two completely different images may have similar global underlying feature descriptions. This results in low classification accuracy using neural networks or support vector machines based on the underlying features of foam images, leading to frequent misoperation of production operations based on working conditions and difficulties in maintaining stable and optimal production operations.

[0004] To address the above issues, a prior art invention patent (CN102855492A) discloses a classification method for mineral flotation foam images, which includes three main stages: (1) generating a foam state vocabulary based on texture and color features; (2) bag-of-words description of the foam image; and (3) classifying the foam image using a vector space model. This application abstracts the foam image into text, which can then be classified and identified at the semantic level, just like text processing. By fully utilizing the local information of the image, the foam image is divided into blocks, and the abstraction process is performed twice to obtain a semantic description of the foam image, thus solving the semantic gap problem.

[0005] However, existing technologies are only theoretically feasible. In practice, the vector inner product method prefers a linear solution, but in actual production, foam images may be random. The K-nearest neighbor method also has the disadvantages of high computational and spatial complexity, making the foam classification process computationally intensive.

[0006] Depth search provides a wider application space for the field of image processing. It can accurately classify objects with unclear image source information and interfered images to obtain accurate information. Summary of the Invention

[0007] To overcome the problems existing in the related art, the disclosed embodiments of the present invention provide a method and system for classifying mineral flotation foam images based on deep learning.

[0008] The technical solution is as follows: A method for classifying mineral flotation foam images based on deep learning, comprising the following steps:

[0009] S1, based on the deep learning network, builds a mineral flotation foam preliminary classification model for classifying mineral flotation foam images;

[0010] S2, using a laser camera to obtain mineral flotation foam images at different times during production, and using the extracted denoised standard texture features, color features, and parameters of the underlying characteristics of foam size distribution as input sets to the mineral flotation foam preliminary selection model;

[0011] S3, trains the mineral flotation foam preliminary model containing the input set to obtain clear mineral flotation foam images at different time periods, uses the improved image classification method to classify, and obtains the mineral flotation foam working status at different time periods in the production situation according to the classification results.

[0012] In step S1, a mineral flotation foam preliminary classification model for classifying mineral flotation foam images is constructed based on a deep learning network, including:

[0013] S101, selecting a mineral flotation foam image with abnormal foam production, setting a mineral flotation foam image encoder and performing recognition, and reporting to a mineral flotation foam image decoder, so that the mineral flotation foam image decoder cyclically detects the status of each mineral flotation foam image encoder;

[0014] S102 , by adding a constraint on the total amount of mineral flotation foam generated, the amount of mineral flotation foam generated under different production process conditions is distributed, the local data of the mineral flotation foam production section is updated, and the data is transmitted to the mineral flotation foam image decoder.

[0015] In step S101, a mineral flotation froth image encoder is set and recognized, and reported to a mineral flotation froth image decoder, so that the mineral flotation froth image decoder cyclically detects the status of each mineral flotation froth image encoder, including:

[0016] (1) The mineral flotation foam image encoder is connected to the mineral flotation foam image decoder, and the image classification offset function is minimized by finding the parameters of the mineral flotation foam preliminary model. The first Y mineral flotation foam image encoders with the largest offset are selected to perform local update; wherein, the image classification offset function expression is:

[0017]

[0018] Where A(d0) is the image classification offset function, E is the number of mineral flotation foam image encoders, ξ is the sample batch, and B e is the local dataset, h(d,ξ) is the e and the offset function of the randomly selected sample batch ξ in the model parameters d0, O e is the data ratio of the e-th mineral flotation foam image encoder, A e (d) is the local offset function of the mineral flotation foam image e, d * are model parameters;

[0019] (2) Select the current global mineral flotation foam primary selection model for global update and define the mineral flotation foam image selection strategy function λ gy , mapping the global mineral flotation foam preliminary model d(g0) to the selected mineral flotation foam image set C(λ gy ,d0), the expression is:

[0020]

[0021] In the formula, Top is selected, For the updated active mineral flotation froth image encoder set, A y is the offset function of the active mineral flotation foam image encoder, d(g0) is the global mineral flotation foam primary selection model, C(λ gy ,d0) is the selected mineral flotation foam image set, λ gy Selecting strategy functions for mineral flotation froth images;

[0022] (3) The update of the cyclic detection g of the mineral flotation foam image encoder selection cluster is expressed as:

[0023]

[0024] Where, d e (g+1, G) is the local mineral flotation foam primary selection model parameter of the mineral flotation foam image e at time g+1, Y is the size of the active mineral flotation foam image encoder set, v is the batch size, is the stochastic gradient on the batch ξ(g,G), is the global mineral flotation foam primary selection model parameter, dy (g, G) is the local mineral flotation foam preliminary selection model parameter of the mineral flotation foam image encoder y at time g, η(g, G) is the current local mineral flotation foam preliminary selection model parameter of the mineral flotation foam image encoder at time g, is the downward gradient, ξ(g,G) is the batch of mineral flotation foam image e at time g, are the parameters of the global mineral flotation foam primary selection model when g+1.

[0025] In step (1), the first Y mineral flotation froth image encoders with the largest offset are selected to perform local updates, including:

[0026] The mineral flotation foam production section sends the local lost data of the production section to the mineral flotation foam image decoder, which tracks each response; the mineral flotation foam image encoder selects the mineral flotation foam image decoder, processes it according to the latest status of the mineral flotation foam production section and the channel status information, and broadcasts it to the mineral flotation foam production section before aggregation; the latest offset value record of the unselected mineral flotation foam image encoder is set to 1, and the selected mineral flotation foam production section uploads the corresponding mineral flotation foam preliminary model parameters, which are expressed as follows:

[0027]

[0028] Where, are the parameters of the global mineral flotation foam primary selection model at g, η g is the parameter of the local mineral flotation foam primary selection model at g, d y (g) is the local mineral flotation foam primary selection model parameter at time g, and ξ(g) is the mineral flotation batch at time g.

[0029] In step S102, the local data of the mineral flotation foam production section is updated and transmitted to the mineral flotation foam image decoder, including:

[0030] (1) Modeling a system connecting the mineral flotation foam production section and the parameter mineral flotation foam image decoder through error control, and using orthogonal frequency division multiplexing technology for transmission;

[0031] (2) In the distributed classifier, the mineral flotation foam image decoder receives the superimposed signals from its neighbors: in the b-th round of communication of the classification, the local update is sent to the mineral flotation foam image decoder through error control, with S sub-channels with N different production process conditions; the sub-channels are arrive Assigned to the mineral flotation foam production section y∈[Y], y is the sub-mineral flotation foam image encoder of a production section in the mineral flotation foam image encoder set in the mineral flotation foam production, and Y is the mineral flotation foam image encoder set in the mineral flotation foam production; for the amount of mineral flotation foam generated Mineral flotation foam production section The parallel wireless Gaussian channel capacity between the CMOS and the mineral flotation froth image decoder is determined using the following formula:

[0032]

[0033] Where U(g) is the amount of mineral flotation foam generated in communication round g allocated to the mineral flotation foam production section y;

[0034] (3) Distribution of mineral flotation foam generation on sub-channels y,i [s] Adapt to the corresponding channel coefficient T y [s] perform gradient aggregation through spatial compression, and the mineral flotation foam image decoder aggregates messages according to the aggregation rule to obtain a new global model update.

[0035] In step (3), the mineral flotation foam generation capacity on the sub-channel is constrained as follows:

[0036]

[0037] Where, is the amount of mineral flotation foam generated on the sub-channel, V0 is the total amount of initial mineral flotation foam generated, Constraint quantity for mineral flotation froth image encoder;

[0038] The channel coefficients are distributed equally on the sub-channels, and the constraint on the amount of froth generated by mineral flotation in each channel is reduced to:

[0039]

[0040] Where S is the number of sub-channels;

[0041] Mineral flotation foam production section The amount of mineral flotation foam generated on the i-th sub-channel is as follows:

[0042]

[0043] Where O0 is the proportional factor for controlling the average amount of released mineral flotation foam, is the value constrained by the proportional factor of the change in the amount of froth produced by mineral flotation, is the change value of the amount of mineral flotation foam generated on the i-th sub-channel at time g and in the mineral flotation foam production section in a mineral flotation foam image decoder;

[0044]

[0045] Each mineral flotation foam production section is based on the mineral flotation foam production amount V in the χ iteration. g Send the processed signal, V g satisfy:

[0046]

[0047] Where, is a linear decrease function.

[0048] Furthermore, the linear descent function The expression is:

[0049]

[0050] Where min_iter is the predefined minimum number of communication rounds, and mum_iter is the current number of communication rounds;

[0051] The exact value of the proportional factor o0 for controlling the amount of froth generated by the average released mineral flotation is defined by the following formula:

[0052]

[0053] Where s is the number of sub-channels with a linear trend, and Ei(θ) is the energy of mineral flotation foam generated at an angle of θ on the i-th sub-channel.

[0054] Furthermore, the mineral flotation froth image decoder aggregates messages according to aggregation rules to obtain a new global model update, including:

[0055] Find the geometric median of a set of points, given point D (g) ∈{d y (g),y∈[Y]}, the fixed point l mean is minimized Point D * , where D (g) is the node in the global model at time g, d y (g) is the parameter of the local mineral flotation foam primary selection model at time g, β y is the aggregation coefficient of mineral flotation foam image decoder, ‖‖ is the Euclidean norm;

[0056]

[0057] Where, is the aggregate distance weight of the mineral flotation foam image decoder at time g, ω y (g) is the aggregate distance value of the mineral flotation froth image decoder at time g, Yes y (g) weight;

[0058]

[0059] Where min{} is the minimization set, and I is the smoothing factor of the distance;

[0060] The objective function is used to find the vector with the smallest distance to all update messages and is defined as:

[0061]

[0062] Where, is the fixed point distance vector in the global model, and D is the fixed point in the global model;

[0063] In order to meet the constraint of average mineral flotation foam production, the channel inversion application at the mineral flotation foam production section is:

[0064]

[0065] Where x y (g) is the inversion parameter of the local mineral flotation foam primary selection model at time g, is the amount of local mineral flotation foam generated at g, k y (g) is the updated information, x′ y (g) is x y (g) Channel inversion, Scaling factor for mineral flotation foam generation.

[0066] In step S3, the improved image classification method is used for classification, including: calculating the correlation coefficient H between each standard texture feature, color feature, bottom feature node of foam size distribution and the mineral flotation foam working state f , the calculation formula is as follows:

[0067]

[0068] Where m is the overall state of mineral flotation foam working conditions, n is the total category of standard texture features, color features, and bottom feature nodes of foam size distribution, S ab is the corresponding classification count, Q abIt is the classification count of the expected standard texture features, color features, and foam size distribution underlying features obtained based on the original mineral flotation foam preliminary selection model containing the input set. a is the current status of a mineral flotation foam working condition, and b is the category of the standard texture features, color features, and foam size distribution underlying feature nodes.

[0069] Another object of the present invention is to provide a mineral flotation foam image classification system based on deep learning, which implements the mineral flotation foam image classification method based on deep learning. The system includes:

[0070] A mineral flotation foam preliminary selection model construction module is used to construct a mineral flotation foam preliminary selection model for classifying mineral flotation foam images based on a deep learning network;

[0071] The image data input module is used to use a laser camera to obtain mineral flotation foam images at different times during production, and to use the extracted denoised standard texture features, color features, and foam size distribution underlying feature parameters as input sets to input into the mineral flotation foam primary selection model;

[0072] The mineral flotation foam classification module is used to train the mineral flotation foam preliminary model containing the input set, obtain clear mineral flotation foam images at different time periods, use the improved image classification method to perform classification, and obtain the mineral flotation foam working status at different time periods in the production situation based on the classification results.

[0073] Combining all of the above technical solutions, the present invention achieves the following beneficial effects: The foam image classification method can accurately classify and identify foam images, increasing the recognition rate from 88% in the prior art to approximately 93%. This method can then determine whether the current production condition is excellent, good, fair, or poor, thereby guiding and optimizing production. Furthermore, it provides in-depth analysis of objects with unclear image source information, obtaining accurate information. After foam image classification, the current operating condition is identified, which can be used to guide and optimize production, reducing reagent usage by 30-40% and increasing concentrate grade and mineral recovery by approximately 20%. BRIEF DESCRIPTION OF THE DRAWINGS

[0074] The accompanying drawings, which are incorporated in and constitute a part of this specification, illustrate embodiments consistent with the present disclosure and, together with the description, serve to explain the principles of the present disclosure;

[0075] Figure 1 This is a flow chart of a method for classifying mineral flotation foam images based on deep learning provided by an embodiment of the present invention;

[0076] Figure 2This is a flow chart of a mineral flotation foam preliminary selection model for classifying mineral flotation foam images based on a deep learning network provided by an embodiment of the present invention;

[0077] Figure 3 2 is a schematic diagram of a mineral flotation foam image classification system based on deep learning provided by an embodiment of the present invention;

[0078] In the figure: 1. Mineral flotation foam preliminary selection model construction module; 2. Image data input module; 3. Mineral flotation foam classification module. DETAILED DESCRIPTION

[0079] To make the above-mentioned objects, features, and advantages of the present invention more readily apparent, specific embodiments of the present invention are described in detail below with reference to the accompanying drawings. The following description sets forth numerous specific details to facilitate a full understanding of the present invention. However, the present invention can be implemented in many other ways than those described herein, and those skilled in the art may make similar modifications without departing from the scope of the present invention. Therefore, the present invention is not limited to the specific embodiments disclosed below.

[0080] Example 1, as Figure 1 As shown, the classification method of mineral flotation foam images based on deep learning provided by the embodiment of the present invention includes:

[0081] S1, based on the deep learning network, builds a mineral flotation foam preliminary classification model for classifying mineral flotation foam images;

[0082] S2, using a laser camera to obtain mineral flotation foam images at different times during production, and using the extracted denoised standard texture features, color features, and parameters of the underlying characteristics of foam size distribution as input sets to the mineral flotation foam preliminary selection model;

[0083] S3, trains the mineral flotation foam preliminary model containing the input set to obtain clear mineral flotation foam images at different time periods, uses the improved image classification method to classify, and obtains the mineral flotation foam working status at different time periods in the production situation according to the classification results.

[0084] For example, Figure 2 As shown, in step S1, a mineral flotation foam preliminary classification model for classifying mineral flotation foam images is constructed based on a deep learning network, including:

[0085] S101, selecting a mineral flotation foam image with abnormal foam production, setting a mineral flotation foam image encoder and performing recognition, and reporting to a mineral flotation foam image decoder, so that the mineral flotation foam image decoder cyclically detects the status of each mineral flotation foam image encoder;

[0086] S102 , by adding a constraint on the total amount of mineral flotation foam generated, the amount of mineral flotation foam generated under different production process conditions is distributed, the local data of the mineral flotation foam production section is updated, and the data is transmitted to the mineral flotation foam image decoder.

[0087] Exemplarily, step S101 includes:

[0088] (1) The mineral flotation foam image encoder is connected to the mineral flotation foam image decoder, and the image classification offset function is minimized by finding the parameters of the mineral flotation foam preliminary model. The first Y mineral flotation foam image encoders with the largest offset are selected to perform local update; wherein, the image classification offset function expression is:

[0089]

[0090] Where A(d0) is the image classification offset function, E is the number of mineral flotation foam image encoders, ξ is the sample batch, and B e is the local dataset, h(d,ξ) is the e and the offset function of the randomly selected sample batch ξ in the model parameters d0, O e is the data ratio of the e-th mineral flotation foam image encoder, A e (d) is the local offset function of the mineral flotation foam image e, d * are model parameters;

[0091] (2) Select the current global mineral flotation foam primary selection model for global update and define the mineral flotation foam image selection strategy function λ gy , mapping the global mineral flotation foam preliminary model d(g0) to the selected mineral flotation foam image set C(λ gy ,d0), the expression is:

[0092]

[0093] In the formula, Top is selected, For the updated active mineral flotation froth image encoder set, A y is the offset function of the active mineral flotation foam image encoder, d(g0) is the global mineral flotation foam primary selection model, C(λ gy ,d0) is the selected mineral flotation foam image set, λ gy Selecting strategy functions for mineral flotation froth images;

[0094] (3) The update of the cyclic detection g of the mineral flotation foam image encoder selection cluster is expressed as:

[0095]

[0096] Where, d e (g+1, G) is the local mineral flotation foam primary selection model parameter of the mineral flotation foam image e at time g+1, Y is the size of the active mineral flotation foam image encoder set, v is the batch size, is the stochastic gradient on the batch ξ(g,G), is the global mineral flotation foam primary selection model parameter, d y (g, G) is the local mineral flotation foam preliminary selection model parameter of the mineral flotation foam image encoder y at time g, η(g, G) is the current local mineral flotation foam preliminary selection model parameter of the mineral flotation foam image encoder at time g, is the downward gradient, ξ(g,G) is the batch of mineral flotation foam image e at time g, are the parameters of the global mineral flotation foam primary selection model when g+1.

[0097] Exemplarily, in step (1), the first Y mineral flotation froth image encoders with the largest offset are selected to perform local update, including:

[0098] The mineral flotation foam production section sends the local lost data of the production section to the mineral flotation foam image decoder, which tracks each response; the mineral flotation foam image encoder selects the mineral flotation foam image decoder, processes it according to the latest status of the mineral flotation foam production section and the channel status information, and broadcasts it to the mineral flotation foam production section before aggregation; the latest offset value record of the unselected mineral flotation foam image encoder is set to 1, and the selected mineral flotation foam production section uploads the corresponding mineral flotation foam preliminary model parameters, which are expressed as follows:

[0099]

[0100] Where, are the parameters of the global mineral flotation foam primary selection model at g, η g is the parameter of the local mineral flotation foam primary selection model at g, d y (g) is the local mineral flotation foam primary selection model parameter at time g, and ξ(g) is the mineral flotation batch at time g.

[0101] Exemplarily, in step S102, updating the local data of the mineral flotation foam production section and transmitting it to the mineral flotation foam image decoder includes:

[0102] (1) Modeling a system connecting the mineral flotation foam production section and the parameter mineral flotation foam image decoder through error control, and using orthogonal frequency division multiplexing technology for transmission;

[0103] (2) In the distributed classifier, the mineral flotation foam image decoder receives the superimposed signals from its neighbors: in the b-th round of communication of the classification, the local update is sent to the mineral flotation foam image decoder through error control, with S sub-channels with N different production process conditions; the sub-channels are arrive Assigned to the mineral flotation foam production section y∈[Y], y is the sub-mineral flotation foam image encoder of a production section in the mineral flotation foam image encoder set in the mineral flotation foam production, and Y is the mineral flotation foam image encoder set in the mineral flotation foam production; for the amount of mineral flotation foam generated Mineral flotation foam production section The parallel wireless Gaussian channel capacity between the CMOS and the mineral flotation froth image decoder is determined using the following formula:

[0104]

[0105] Where U(g) is the amount of mineral flotation foam generated in communication round g allocated to the mineral flotation foam production section y;

[0106] (3) Distribution of mineral flotation foam generation on sub-channels y,i [s] Adapt to the corresponding channel coefficient T y [s] perform gradient aggregation through spatial compression, and the mineral flotation foam image decoder aggregates messages according to the aggregation rule to obtain a new global model update.

[0107] In step (3), the mineral flotation foam generation capacity on the sub-channel is constrained as follows:

[0108]

[0109] Where, is the amount of mineral flotation foam generated on the sub-channel, V0 is the total amount of initial mineral flotation foam generated, Constraint quantity for mineral flotation froth image encoder;

[0110] The channel coefficients are distributed equally on the sub-channels, and the constraint on the amount of froth generated by mineral flotation in each channel is reduced to:

[0111]

[0112] Where S is the number of sub-channels;

[0113] Mineral flotation foam production section The amount of mineral flotation foam generated on the i-th sub-channel is as follows:

[0114]

[0115] Where O0 is the proportional factor for controlling the average amount of released mineral flotation foam, is the value constrained by the proportional factor of the change in the amount of froth produced by mineral flotation, is the change value of the amount of mineral flotation foam generated on the i-th sub-channel at time g and in the mineral flotation foam production section in a mineral flotation foam image decoder;

[0116]

[0117] Each mineral flotation foam production section is based on the mineral flotation foam production amount V in the χ iteration. g Send the processed signal, V g satisfy:

[0118]

[0119] Where, is a linear decrease function.

[0120] Linear descent function The expression is:

[0121]

[0122] Where min_iter is the predefined minimum number of communication rounds, and mum_iter is the current number of communication rounds;

[0123] The exact value of the proportional factor o0 for controlling the amount of froth generated by the average released mineral flotation is defined by the following formula:

[0124]

[0125] Where s is the number of sub-channels with a linear trend, and Ei(θ) is the energy of mineral flotation foam generated at an angle of θ on the i-th sub-channel.

[0126] For example, for the model aggregation step, the mineral flotation foam image decoder aggregates messages according to certain aggregation rules to obtain a new global model update. Unlike the standard method with one-step average aggregation, the standard system of average aggregation requires a more powerful aggregation rule because the naive average aggregation rule is vulnerable to model or data attacks. The geometric median aggregation rule has a good convergence guarantee for the mineral flotation foam production line, especially when part of the mineral flotation foam production section is attacked, which makes it an ideal tool for the standard system of average aggregation. Find the geometric median of a set of points, given point D (g) ∈{d y (g),y∈[Y]}, the fixed point l mean is minimized Point D, where D (g) is the node in the global model at time g, dy (g) is the parameter of the local mineral flotation foam primary selection model at time g, β y is the aggregation coefficient of mineral flotation foam image decoder, |||| is the Euclidean norm;

[0127]

[0128] Where, is the aggregate distance weight of the mineral flotation foam image decoder at time g, ω y (g) is the aggregate distance value of the mineral flotation froth image decoder at time g, Yes y (g) weight;

[0129]

[0130] Where min{} is the minimization set, and I is the smoothing factor of the distance;

[0131] The objective function is used to find the vector with the smallest distance to all update messages and is defined as:

[0132]

[0133] Where, is the fixed point distance vector in the global model, and D is the fixed point in the global model;

[0134] In order to meet the constraint of average mineral flotation foam production, the channel inversion application at the mineral flotation foam production section is:

[0135]

[0136] Where x y (g) is the inversion parameter of the local mineral flotation foam primary selection model at time g, is the amount of local mineral flotation foam generated at g, k y (g) is the updated information, x′ y (g) is x y (g) Channel inversion, Scaling factor for mineral flotation foam generation.

[0137] For example, in step S3, the improved image classification method is used for classification, including: calculating the correlation coefficient H between each standard texture feature, color feature, foam size distribution bottom feature node and mineral flotation foam working state f , the calculation formula is as follows:

[0138]

[0139] Where m is the overall state of mineral flotation foam working conditions, n is the total category of standard texture features, color features, and bottom feature nodes of foam size distribution, S ab is the corresponding classification count, Q ab It is the classification count of the expected standard texture features, color features, and foam size distribution underlying features obtained based on the original mineral flotation foam preliminary selection model containing the input set. a is the current status of a mineral flotation foam working condition, and b is the category of the standard texture features, color features, and foam size distribution underlying feature nodes.

[0140] Example 2, as Figure 3 As shown, the mineral flotation foam image classification system based on deep learning provided by the embodiment of the present invention includes:

[0141] A mineral flotation foam preliminary selection model construction module 1 is used to construct a mineral flotation foam preliminary selection model for classifying mineral flotation foam images based on a deep learning network;

[0142] Image data input module 2 is used to use a laser camera to obtain mineral flotation foam images at different times during production, and extract the de-noised standard texture features, color features, and foam size distribution underlying feature parameters as input sets to the mineral flotation foam primary selection model;

[0143] The mineral flotation foam classification module 3 is used to train the mineral flotation foam preliminary selection model containing the input set, obtain clear mineral flotation foam images at different time periods, use the improved image classification method to perform classification, and obtain the mineral flotation foam working state at different time periods in the production situation according to the classification results.

[0144] By applying the method proposed in the present invention to actual scenarios, it is shown that the present invention avoids the problem of excessive congestion in classification communication, can quickly classify images, and obtain the mineral flotation foam working conditions at different time periods in the production situation based on the classification results, providing a theoretical basis for actual production.

[0145] The above description is only a preferred specific implementation method of the present invention, but the scope of protection of the present invention is not limited thereto. Any modifications, equivalent substitutions and improvements made by any technician familiar with this technical field within the technical scope disclosed by the present invention and within the spirit and principles of the present invention should be covered by the scope of protection of the present invention.

Claims

1. A classification method for mineral flotation foam images based on deep learning, characterized in that: The method comprises the following steps: S1, based on the deep learning network, builds a mineral flotation foam preliminary classification model for classifying mineral flotation foam images; S2, using a laser camera to obtain mineral flotation foam images at different times during production, and using the extracted denoised standard texture features, color features, and parameters of the underlying characteristics of foam size distribution as input sets to the mineral flotation foam preliminary selection model; S3, training the mineral flotation foam preliminary model containing the input set to obtain clear mineral flotation foam images at different time periods, classifying them using an improved image classification method, and obtaining the mineral flotation foam working conditions at different time periods in the production situation based on the classification results; In step S1, a mineral flotation foam preliminary classification model for classifying mineral flotation foam images is constructed based on a deep learning network, including: S101, selecting a mineral flotation foam image with abnormal foam production, setting a mineral flotation foam image encoder and performing recognition, and reporting to a mineral flotation foam image decoder, so that the mineral flotation foam image decoder cyclically detects the status of each mineral flotation foam image encoder; S102 , by adding a constraint on the total amount of mineral flotation foam generated, the amount of mineral flotation foam generated under different production process conditions is distributed, the local data of the mineral flotation foam production section is updated, and the data is transmitted to the mineral flotation foam image decoder.

2. The method for classifying mineral flotation foam images based on deep learning according to claim 1, characterized in that: In step S101, a mineral flotation froth image encoder is set and recognized, and reported to a mineral flotation froth image decoder, so that the mineral flotation froth image decoder cyclically detects the status of each mineral flotation froth image encoder, including: (1) The mineral flotation foam image encoder is connected to the mineral flotation foam image decoder, and the parameters of the mineral flotation foam primary selection model are found to minimize the image classification offset function, and the front image with the largest offset is selected. The mineral flotation foam image encoder performs local updates; the image classification offset function expression is: ； Where, is the image classification offset function, froth image encoder number for mineral flotation, is the sample batch, For local datasets, From local dataset and model parameters Randomly selected sample batches The offset function, For the The data ratio of the mineral flotation foam image encoder, Foam images for mineral flotation The local offset function, are model parameters; (2) Select the current global mineral flotation foam preliminary selection model for global update and define the mineral flotation foam image selection strategy function , the global mineral flotation foam primary selection model Map to selected mineral flotation froth image sets , the expression is: ； Where, For selection, For updated active mineral flotation froth image encoder set, is the offset function of the active mineral flotation froth image encoder, It is the global mineral flotation foam preliminary selection model, For a selected set of mineral flotation froth images, Selecting strategy functions for mineral flotation froth images; (3) Cycle detection of mineral flotation foam image encoder selection clustering The update expression is: ； Where, Foam images for mineral flotation exist The parameters of the local mineral flotation foam primary selection model at The size of the active mineral flotation froth image encoder set, is the batch size, For batch The stochastic gradient on Froth image encoder for mineral flotation exist The parameters of the local mineral flotation foam primary selection model at Froth image encoder for mineral flotation Current local mineral flotation foam primary selection model parameters at time, is the downward gradient, For Mineral flotation foam image batches, for Parameters of the global mineral flotation foam primary selection model.

3. The method for classifying mineral flotation foam images based on deep learning according to claim 2, characterized in that: In step (1), select the front with the largest offset The mineral flotation froth image encoder performs local updates, including: The mineral flotation foam production section sends the local lost data of the production section to the mineral flotation foam image decoder, which tracks each response; the mineral flotation foam image encoder selects the mineral flotation foam image decoder, processes it according to the latest status of the mineral flotation foam production section and the channel status information, and broadcasts it to the mineral flotation foam production section before aggregation; the latest offset value record of the unselected mineral flotation foam image encoder is set to 1, and the selected mineral flotation foam production section uploads the corresponding mineral flotation foam preliminary model parameters, which are expressed as follows: ； Where, They are The global mineral flotation foam primary selection model parameters are: For The parameters of the local mineral flotation foam primary selection model at For The parameters of the local mineral flotation foam primary selection model at For Mineral flotation batches at the time.

4. The method for classifying mineral flotation foam images based on deep learning according to claim 1, characterized in that: In step S102, the local data of the mineral flotation foam production section is updated and transmitted to the mineral flotation foam image decoder, including: (1) Modeling a system that connects the mineral flotation foam production section and the parameter mineral flotation foam image decoder through error control, and adopting orthogonal frequency division multiplexing technology for transmission; (2) In the distributed classifier, the mineral flotation foam image decoder receives superimposed signals from its neighbors: During round-trip communication, local updates are sent to the mineral flotation froth image decoder with error control, which has Different production process conditions sub-channels; arrive Assigned to the mineral flotation foam production section , It is a sub-mineral flotation foam image encoder for a certain production section of the mineral flotation foam production. A mineral flotation foam image encoder set for the production of mineral flotation foam; a mineral flotation foam generation amount , mineral flotation foam production section The parallel wireless Gaussian channel capacity between the CMOS and the mineral flotation froth image decoder is determined using the following formula: ； Where, For communication rounds Assigned to the mineral flotation foam production section The amount of mineral flotation foam produced; (3) Distribution of mineral flotation foam generation on sub-channels Adapt to the corresponding channel coefficients By performing gradient aggregation via spatial compression, the mineral flotation froth image decoder aggregates messages according to aggregation rules to obtain a new global model update.

5. The method for classifying mineral flotation foam images based on deep learning according to claim 4, characterized in that: In step (3), the mineral flotation foam generation constraint on the sub-channel is as follows: ； Where, is the amount of mineral flotation foam generated on the sub-channel, is the total amount of initial mineral flotation foam generated, Constraint quantity for mineral flotation froth image encoder; The channel coefficients are distributed equally on the sub-channels, and the constraint on the amount of froth generated by mineral flotation in each channel is reduced to: ； Where, is the number of sub-channels; Mineral flotation foam production section In the The amount of mineral flotation foam produced on each sub-channel is as follows: ； Where, The proportional factor for controlling the amount of froth generated by average release of mineral flotation, is the value constrained by the proportional factor of the change in the amount of froth produced by mineral flotation, for Moment The change value of the amount of mineral flotation foam produced on each sub-channel and in the mineral flotation foam production section in a mineral flotation foam image decoder; ； Each mineral flotation foam production section The amount of froth generated by mineral flotation during iteration Send the processed signal, satisfy: ； Where, is a linear decrease function.

6. The method for classifying mineral flotation foam images based on deep learning according to claim 5, characterized in that: Linear descent function The expression is: ； Where, is the predefined minimum number of communication rounds, is the current communication round number; Proportional factor for controlling the amount of froth generated in average released mineral flotation The exact value of is defined by the following formula: ； Where, is the number of subchannels with a linear trend, For the On a subchannel The energy generated by mineral flotation foam under different angles.

7. The method for classifying mineral flotation foam images based on deep learning according to claim 4, characterized in that: The mineral flotation froth image decoder aggregates messages according to aggregation rules to obtain new global model updates, including: Find the geometric median of a set of points, given a point , the fixed-point l-mean is minimized point ,in, for Nodes in the global model at time t, For The parameters of the local mineral flotation foam primary selection model at is the aggregation coefficient of mineral flotation froth image decoder, is the Euclidean norm; ； Where, for Mineral flotation foam image decoder aggregation distance weight at time, for The aggregated distance value of the mineral flotation foam image decoder at the moment, yes The weight of ； Where, To minimize the set, is the smoothing factor of the distance; The objective function is used to find the vector with the smallest distance to all update messages and is defined as: ； Where, is the fixed-point distance vector in the global model, is a fixed point in the global model; In order to meet the constraint of average mineral flotation foam production, the channel inversion application at the mineral flotation foam production section is: ； Where, For The inversion parameters of the local mineral flotation foam primary selection model at For The amount of local mineral flotation foam generated at To update information, for Channel inversion, Scaling factor for mineral flotation foam generation.

8. The method for classifying mineral flotation foam images based on deep learning according to claim 1, characterized in that: In step S3, the improved image classification method is used for classification, including: calculating the correlation coefficient between each standard texture feature, color feature, bottom feature node of foam size distribution and the mineral flotation foam working state , the calculation formula is as follows: ； Where, It is the overall state of mineral flotation foam working condition, is the total category of the underlying feature nodes of standard texture features, color features, and foam size distribution, is the corresponding classification count, The expected standard texture features, color features, and foam size distribution bottom feature classification counts are obtained based on the original mineral flotation foam primary selection model containing the input set. The current status of the flotation foam working state of a certain mineral, It is the category of the underlying feature node of standard texture feature, color feature, and foam size distribution.

9. A mineral flotation foam image classification system based on deep learning, characterized in that: The system implements the deep learning-based mineral flotation froth image classification method according to any one of claims 1 to 8, and the system comprises: A mineral flotation foam preliminary selection model building module (1) is used to build a mineral flotation foam preliminary selection model for classifying mineral flotation foam images based on a deep learning network; An image data input module (2) is used to use a laser camera to obtain mineral flotation foam images at different time periods during production, and to use the extracted denoised standard texture features, color features, and parameters of the underlying features of the foam size distribution as an input set to input into a mineral flotation foam preliminary selection model; The mineral flotation foam classification module (3) is used to train the mineral flotation foam preliminary selection model containing the input set, obtain clear mineral flotation foam images at different time periods, classify them using the improved image classification method, and obtain the mineral flotation foam working state at different time periods in the production situation according to the classification results.

Images