Intelligent dosing method and system for lead-zinc ore flotation

Abstract

The present invention relates to the field of artificial intelligence technology and discloses an intelligent dosing method and system for lead-zinc ore flotation to improve the efficiency, stability, and economic benefits of the flotation process. The method comprises: calculating first and second cross-modal features, fusing the first and second cross-modal features with bubble image features and element grade features through residual connections, and then processing them to obtain a multimodal feature sequence; inputting the multimodal feature sequence for each time step into an LSTM network to obtain a corresponding LSTM hidden state sequence; then calculating the temporal attention weight of the hidden state sequence for each time step within a time window, and performing weighted summation on the corresponding LSTM hidden state sequence based on the temporal attention weight to obtain a context vector within the time window; finally, sharing the weighted summed context vector as input to the fully connected layers corresponding to the two detection heads to respectively obtain the probability distribution of the predicted reagent type and the dosage value of each reagent.

Description

Technical Field

[0001] The present invention relates to the field of artificial intelligence and automatic control technology, and in particular to an intelligent dosing method and system for lead-zinc ore flotation. Background Art

[0002] In the mining industry, lead and zinc ore flotation is a core process designed to efficiently separate valuable metals from waste rock within the ore. Optimizing the dosage of reagents during flotation is crucial for improving target mineral recovery and process efficiency. However, existing flotation dosing methods suffer from significant deficiencies in accuracy, real-time performance, and intelligent operation, severely limiting the stability and economic benefits of the flotation process.

[0003] First, traditional dosing relies heavily on the operator's experience and subjective judgment. This manual approach is susceptible to variations in operator skill, work status, and experience, leading to unstable and inaccurate dosing, which in turn reduces flotation performance. Insufficient dosing can result in reduced target mineral recovery, while excessive dosing can lead to resource waste, increased production costs, and environmental pollution.

[0004] Secondly, although some semi-automated dosing equipment (such as self-service dosing machines) has emerged in the industry, it still relies heavily on manual experience. Operators must observe bubble images and elemental grade data to preset dosing intervals and dosages. This manual setup lacks real-time awareness of dynamic changes in the flotation process, making it impossible to make timely adjustments based on changes in slurry properties during the flotation process, making it difficult to achieve true closed-loop control and efficient operation.

[0005] Furthermore, the existing dosing process lacks intelligence and cannot meet the dynamic control requirements of flotation processes. During flotation, pulp properties and flotation conditions can fluctuate continuously with changes in raw material quality or environmental factors. Traditional dosing methods, due to their high hysteresis, struggle to intelligently adjust the type and dosage of reagents, leading to increased volatility in flotation results and particularly low efficiency in large-scale continuous production.

[0006] In summary, existing flotation dosing methods have significant room for improvement in terms of accuracy, real-time performance, and intelligence. Therefore, a new dosing method and device that combines advanced technology with intelligent control algorithms is urgently needed to achieve real-time monitoring and dynamic regulation of the flotation process, thereby improving the efficiency, stability, and economic benefits of the flotation process. Summary of the Invention

[0007] The present invention aims to disclose an intelligent dosing method and system for lead-zinc ore flotation, so as to improve the efficiency, stability and economic benefits of the flotation process.

[0008] To achieve the above-mentioned purpose, the intelligent dosing method for lead-zinc ore flotation disclosed in the present invention includes:

[0009] Step S1: During the flotation process of lead-zinc ore, bubble image features are extracted based on the bubble image, and at the same time, grade data of at least two metal elements of the slurry sample in the process are collected in real time through an online analyzer to construct element grade features;

[0010] Step S2: Calculate the query vector, key vector, and value vector of the bubble image feature and the element grade feature based on the attention mechanism respectively;

[0011] Step S3: A first cross-modal feature is calculated using a cross-attention mechanism based on a key vector and a value vector based on the element quality feature and a query vector based on the bubble image feature. Simultaneously, a second cross-modal feature is calculated using a cross-attention mechanism based on a key vector and a value vector based on the bubble image feature and a query vector based on the element quality feature. The first cross-modal feature, the second cross-modal feature, the bubble image feature, and the element quality feature are fused using a residual connection, and then layer-normalized to obtain a final multimodal feature sequence.

[0012] Step S4: In any time window, the multimodal feature sequence of each time step is input into the LSTM network to obtain the LSTM hidden state sequence corresponding to each time step;

[0013] Step S5: Calculate the temporal attention weight of the hidden state sequence corresponding to each time step in the time window, and perform weighted summation on the corresponding LSTM hidden state sequence according to the temporal attention weight to obtain the context vector in the time window;

[0014] Step S6: Share the weighted summed context vector as input to the fully connected layers corresponding to the two detection heads to respectively obtain the probability distribution of the predicted drug type and the dosage value of each drug.

[0015] Preferably, in step S1, based on the input 224×224×3 original image, 64-dimensional image features are extracted through a pre-trained ResNet-18 feature extraction network.

[0016] Preferably, in step S2, the query vector, key vector and value vector of the bubble image feature and the element grade feature based on the attention mechanism are calculated by independent linear transformation. The specific calculation process is:

[0017]

[0018] in, It is a 64-dimensional image feature; are different weight matrices that can be trained, and ; are respectively the query vector, key vector and value vector of the bubble image feature based on the attention mechanism;

[0019]

[0020] in, It is the 22-dimensional element grade characteristic, are different weight matrices that can be trained, and ; They are respectively the query vector, key vector and value vector of the element grade feature based on the attention mechanism.

[0021] Preferably, the calculation formula of step S3 specifically includes:

[0022] ; ;

[0023] in, is the first cross-modal feature, is the second cross-modal feature, and is the temperature coefficient of the hyperparameter, is the matrix transpose symbol.

[0024] Preferably, the calculation formula of step S5 specifically includes:

[0025] ; ; ;

[0026] in, is the time window length, The corresponding time step of the LSTM network The output hidden state sequence; is the trainable weight matrix; is a trainable bias term; is the time step Weights in time windows; for Normalized weights after conversion based on the Softmax function; The context vector within the time window is obtained by weighted summation.

[0027] Preferably, the specific calculation formula of step S6 includes:

[0028] ;

[0029] in, is the classification task weight matrix; is the bias term for the classification task; Linear output for classification tasks; is the probability of drug type; is a natural constant, Represents standard deviation.

[0030] ;

[0031] in, is the regression task weight matrix; is the regression task bias; is the predicted value of drug dosage.

[0032] Preferably, in the process of training the model for implementing steps S2 to S6, the loss function is composed of the weighted sum of the multi-label binary cross entropy loss of the drug type information and the drug dosage mean square loss, and the loss gradient of the model back propagation is obtained by deriving the loss.

[0033] To achieve the above-mentioned purpose, the present invention also discloses an intelligent dosing system for lead-zinc ore flotation, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor implements the above-mentioned method when executing the computer program.

[0034] The present invention can improve the efficiency, stability and economic benefits of the flotation process and has the following beneficial effects:

[0035] 1. The multi-dimensional bubble image features and elemental grade features are first calculated using an attention mechanism to calculate query vectors, key vectors, and value vectors. A unique cross-modal attention calculation is then performed on the two types of features. Specifically, a first cross-modal feature is calculated using the key vector and value vector based on the elemental grade features and the query vector based on the bubble image features through a cross-attention mechanism. Simultaneously, a second cross-modal feature is calculated using the key vector and value vector based on the bubble image features and the query vector based on the elemental grade features through a cross-attention mechanism. These first and second cross-modal features are fused with the bubble image features and the elemental grade features through a residual connection, and then layer-normalized to obtain the final multimodal feature sequence. Compared to omitting the unique cross-modal attention calculation for the two types of features during the fusion process of bubble image features and elemental grade features, this multimodal feature sequence optimizes the interaction between features and better captures more key features based on overall synergy at each time step, tailored to the real-time state of the flotation process.

[0036] 2. By calculating the temporal attention weight of the hidden state sequence corresponding to each time step within the time window and performing weighted summation on the corresponding LSTM hidden state sequence based on the temporal attention weight to obtain the context vector within the time window, the key operation time point (i.e., the time step with the largest temporal attention weight in the hidden state sequence obtained by LSTM) can be accurately identified, thereby improving the subsequent prediction accuracy of the probability distribution of drug types and the dosage value of each drug.

[0037] 3. Based on the same context vector within the time window, the probability distribution of drug types and the dosage value of each drug are predicted, which solves the problem of insufficient synergy caused by the separation of the two types of tasks in traditional methods.

[0038] The present invention will be described in further detail below with reference to the accompanying drawings. BRIEF DESCRIPTION OF THE DRAWINGS

[0039] The accompanying drawings, which constitute part of this application, are intended to provide a further understanding of the present invention. The exemplary embodiments of the present invention and their descriptions are intended to explain the present invention and do not constitute an undue limitation of the present invention. In the accompanying drawings:

[0040] Figure 1 The present invention is a flow chart of an intelligent dosing method for lead-zinc ore flotation disclosed in an embodiment of the present invention. DETAILED DESCRIPTION

[0041] The embodiments of the present invention are described in detail below with reference to the accompanying drawings. However, the present invention can be implemented in many different ways as defined and covered by the claims.

[0042] Example 1

[0043] This embodiment discloses an intelligent dosing method for lead-zinc ore flotation, such as Figure 1 Shown, including:

[0044] Step S1: During the flotation process of lead-zinc ore, bubble image features are extracted based on the bubble image, and at the same time, grade data of at least two metal elements of the slurry sample in the process are collected in real time through an online analyzer to construct element grade features.

[0045] Preferably, in this step, based on the input 224×224×3 original image, 64-dimensional image features are extracted through a pre-trained ResNet-18 feature extraction network.

[0046] For the original training dataset of bubble images, several data augmentation methods can be randomly selected to expand the training set. These augmentation methods include, but are not limited to, rotation, flipping, and cropping. By implementing data augmentation on bubble images, we can not only introduce diverse transformations and noise, thereby improving the model's generalization capabilities and enabling it to better adapt to bubble images in different scenarios and conditions; we can also generate richer training samples, effectively expanding the training dataset and alleviating the problem of data scarcity.

[0047] In specific deployment, on the flotation process production line, an online grade analyzer can be used to collect grade data of multiple metal elements (refer to the subsequent specific examples) in real time from slurry samples in key processes such as roughing, cleaning, and scavenging to construct a multi-dimensional elemental grade characteristic.

[0048] Step S2: Calculate the query vector, key vector, and value vector of the bubble image features and element quality features based on the attention mechanism respectively.

[0049] In this step, the query vector, key vector, and value vector of the bubble image features and element grade features based on the attention mechanism can be calculated through independent linear transformations. The specific calculation process is as follows:

[0050]

[0051] in, It is a 64-dimensional image feature; are different weight matrices that can be trained, and ; They are respectively the query vector, key vector and value vector of the bubble image features based on the attention mechanism.

[0052]

[0053] in, It is the element grade characteristic of 22 dimensions, are different weight matrices that can be trained, and ; They are respectively the query vector, key vector and value vector of the element grade feature based on the attention mechanism.

[0054] Step S3: A first cross-modal feature is calculated based on a key vector and a value vector based on the element quality feature and a query vector based on the bubble image feature through a cross-attention mechanism. At the same time, a second cross-modal feature is calculated based on a key vector and a value vector based on the bubble image feature and a query vector based on the element quality feature through a cross-attention mechanism. The first cross-modal feature, the second cross-modal feature, the bubble image feature and the element quality feature are fused through a residual connection, and the final multimodal feature sequence is obtained by layer normalization.

[0055] Referring to the subsequent steps, each feature in the multimodal feature sequence fused in this step is aligned based on the time step.

[0056] The specific calculation formula for this step includes:

[0057] ; ;

[0058] in, is the first cross-modal feature, is the second cross-modal feature, and is the hyperparameter temperature coefficient. Optionally, , , superscript is the matrix transpose symbol.

[0059] Fused multimodal feature sequence It can be expressed as:

[0060] ;

[0061] in, is the weight coefficient, and LayerNorm is the normalization function. Optionally, , which means that we give greater weight to the element grade feature, making it dominant in the feature calculation.

[0062] Step S4: In any time window, the multimodal feature sequence of each time step is input into the LSTM network to obtain the LSTM hidden state sequence corresponding to each time step.

[0063] For example, input multimodal feature sequence , then after LSTM processing, the output ,in, . Optionally, the time window T=5.

[0064] In this step, because the hidden state dimension of the LSTM is independent of the input dimension, that is, the hidden state dimension can be customized, this embodiment maps the 86-dimensional feature (i.e., the sum of 64 + 22) to 128 dimensions, which can enhance the expressiveness of the feature and facilitate the subsequent processing of the temporal attention mechanism and task head (classification and regression).

[0065] Step S5: Calculate the temporal attention weight of the hidden state sequence corresponding to each time step in the time window, and perform weighted summation on the corresponding LSTM hidden state sequence according to the temporal attention weight to obtain the context vector in the time window.

[0066] The calculation formula for this step specifically includes:

[0067] ; ; ;

[0068] in, is the time window length, The corresponding time step of the LSTM network The output hidden state sequence; is the trainable weight matrix; is a trainable bias term; is the time step Weights in time windows; for Normalized weights after conversion based on the Softmax function; The context vector within the time window is obtained by weighted summation.

[0069] During the dimensionality processing of this embodiment, although both bubble images and grade data serve as the basis for dosing, the grade data carries a greater weight. Therefore, the image feature dimensionality is intentionally reduced to prevent the elemental grade features from being overwhelmed during the cross-attention calculation between the bubble image features and the elemental grade features. Furthermore, during the cross-modal feature attention calculation, the elemental grade feature key vector is divided by 0.5 to strengthen its attention weight, allowing the elemental grade features to play a leading role. Conversely, dividing the bubble image feature key vector by 2 weakens its attention weight, allowing the bubble image features to play a supporting role.

[0070] Step S6: Share the weighted summed context vector as input to the fully connected layers corresponding to the two detection heads to respectively obtain the probability distribution of the predicted drug type and the dosage value of each drug.

[0071] Preferably, the specific calculation formula of this step includes:

[0072] ;

[0073] in, is the classification task weight matrix; is the bias term for the classification task; Linear output for classification tasks; is the probability of drug type; is a natural constant, Represents standard deviation.

[0074] ;

[0075] in, is the regression task weight matrix; is the regression task bias; is the predicted value of drug dosage.

[0076] In this embodiment, in the process of training the model for implementing steps S2 to S6, the loss function can be composed of the weighted sum of the multi-label binary classification cross entropy loss of the drug type information and the drug dosage mean square loss, and the loss gradient of the model back propagation is obtained by deriving the loss.

[0077] Optionally, multi-label binary classification cross entropy loss of drug type information , its formal expression can be:

[0078] ;

[0079] in, represents the number of batch samples, Index of drug types (taking 10 drugs in total as an example), For the In the sample The true label of the drug, the value is 0 or 1; For the In the sample The predicted probability of the drug is [0,1].

[0080] Optionally, calculate the mean square loss of drug dose The formal expression can be:

[0081] ;

[0082] in, Indicates the In the sample The actual dosage value of the drug, No. In the sample The predicted dose value of the drug.

[0083] The total loss function can be: ; where λ is the loss weight coefficient, which can be set to 0.5. Furthermore, the final loss is derived to obtain the loss gradient, which is then backpropagated through the network model to optimize the model parameters. The calculation formula can be: ;in, For the The loss gradient parameters of the network model to be learned during the iteration; Represents the learning rate.

[0084] Among them, based on this embodiment, when the bubble image (including the enhanced image) and the corresponding dosing information are marked, the bubble image (including the original data and the enhanced data), the original ore Zn grade, the original ore Pb grade, the total lead concentrate Zn grade, the total lead concentrate Pb grade, the zinc concentrate Zn grade, the zinc concentrate Pb grade, the zinc sweep three-tail Zn grade, the zinc sweep three-tail Pb grade, the high lead concentrate Zn grade, the high lead concentrate Pb grade, the low lead concentrate Zn grade, the low lead concentrate Pb grade, the lead fast rough tail Zn grade, the lead fast rough tail Pb grade, the zinc sweep four-bubble Zn grade, the zinc sweep four-bubble Pb grade, the zinc sweep four-tail Zn grade, the zinc sweep four-tail Pb grade, the lead tail Zn grade, the lead tail Pb grade, the zinc fast rough tail Zn grade, the zinc fast rough tail Pb grade and the reagent type sequence and the reagent dosage sequence are established. In the annotation process of a real reagent and dosage sequence, the first 10 elements represent the reagent type sequence, and elements 11 through 20 represent the reagent dosage sequence. For example, the first element in the reagent sequence represents copper sulfate, the second element represents SD-11, the third element represents ethylthiocyanate, the fourth element represents cassiterite activator, the fifth element represents lead nitrate, the sixth element represents 2# oil, the seventh element represents GZT, the eighth element represents xanthate, the ninth element represents SND-21, and the tenth element represents cassiterite collector. Furthermore, these are all binary data, where 0 represents absence and 1 represents presence. Correspondingly, the first element in the dosage sequence represents the weight of copper sulfate, the second element represents the weight of SD-11, the third element represents the weight of ethyl thiocyanate, the fourth element represents the weight of cassiterite activator, the fifth element represents the weight of lead nitrate, the sixth element represents the weight of 2# oil, the seventh element represents the weight of GZT, the eighth element represents the weight of xanthate, the ninth element represents the weight of SND-21, and the tenth element represents the weight of cassiterite collector; among them, the dosage data is usually a floating point number, and the weight unit is kg.

[0085] Example 2

[0086] This embodiment discloses an intelligent dosing system for lead-zinc ore flotation, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, the method disclosed in the above embodiment is implemented.

[0087] In summary, the methods and systems disclosed in the above two embodiments of the present invention can improve the efficiency, stability and economic benefits of the flotation process and have the following beneficial effects:

[0088] 1. The multi-dimensional bubble image features and elemental grade features are first calculated using an attention mechanism to calculate query vectors, key vectors, and value vectors. A unique cross-modal attention calculation is then performed on the two types of features. Specifically, a first cross-modal feature is calculated using the key vector and value vector based on the elemental grade features and the query vector based on the bubble image features through a cross-attention mechanism. Simultaneously, a second cross-modal feature is calculated using the key vector and value vector based on the bubble image features and the query vector based on the elemental grade features through a cross-attention mechanism. These first and second cross-modal features are fused with the bubble image features and the elemental grade features through a residual connection, and then layer-normalized to obtain the final multimodal feature sequence. Compared to omitting the unique cross-modal attention calculation for the two types of features during the fusion process of bubble image features and elemental grade features, this multimodal feature sequence optimizes the interaction between features and better captures more key features based on overall synergy at each time step, tailored to the real-time state of the flotation process.

[0089] 2. By calculating the temporal attention weight of the hidden state sequence corresponding to each time step within the time window and performing weighted summation on the corresponding LSTM hidden state sequence based on the temporal attention weight to obtain the context vector within the time window, the key operation time point (i.e., the time step with the largest temporal attention weight in the hidden state sequence obtained by LSTM) can be accurately identified, thereby improving the subsequent prediction accuracy of the probability distribution of drug types and the dosage value of each drug.

[0090] 3. Based on the same context vector within the time window, the probability distribution of drug types and the dosage value of each drug are predicted, which solves the problem of insufficient synergy caused by the separation of the two types of tasks in traditional methods.

[0091] The foregoing description is merely a preferred embodiment of the present invention and is not intended to limit the present invention. Those skilled in the art will readily appreciate that various modifications and variations of the present invention are possible. Any modifications, equivalent substitutions, or improvements made within the spirit and principles of the present invention are intended to be within the scope of protection of the present invention.

Claims

1. An intelligent dosing method for lead-zinc ore flotation, characterized in that: include: Step S1: During the flotation process of lead-zinc ore, bubble image features are extracted based on the bubble image, and at the same time, grade data of at least two metal elements of the slurry sample in the process are collected in real time through an online analyzer to construct element grade features; Step S2: Calculate the query vector, key vector, and value vector of the bubble image feature and the element grade feature based on the attention mechanism respectively; Step S3: A first cross-modal feature is calculated using a cross-attention mechanism based on a key vector and a value vector based on the element quality feature and a query vector based on the bubble image feature. Simultaneously, a second cross-modal feature is calculated using a cross-attention mechanism based on a key vector and a value vector based on the bubble image feature and a query vector based on the element quality feature. The first cross-modal feature, the second cross-modal feature, the bubble image feature, and the element quality feature are fused using a residual connection, and then layer-normalized to obtain a final multimodal feature sequence. Step S4: In any time window, the multimodal feature sequence of each time step is input into the LSTM network to obtain the LSTM hidden state sequence corresponding to each time step; Step S5: Calculate the temporal attention weight of the hidden state sequence corresponding to each time step in the time window, and perform weighted summation on the corresponding LSTM hidden state sequence according to the temporal attention weight to obtain the context vector in the time window; Step S6: Share the weighted summed context vector as input to the fully connected layers corresponding to the two detection heads to respectively obtain the probability distribution of the predicted drug type and the dosage value of each drug.

2. The intelligent dosing method for lead-zinc ore flotation according to claim 1, characterized in that: In step S1, based on the input 224×224×3 original image, 64-dimensional image features are extracted through the pre-trained ResNet-18 feature extraction network.

3. The intelligent dosing method for lead-zinc ore flotation according to claim 2, characterized in that: In step S2, the query vector, key vector, and value vector of the bubble image features and element grade features based on the attention mechanism are calculated through independent linear transformations. The specific calculation process is as follows: in, It is a 64-dimensional image feature; are different weight matrices that can be trained, and ; are respectively the query vector, key vector and value vector of the bubble image feature based on the attention mechanism; The number of dimensions of the corresponding parameter is the corresponding superscript value; in, It is the element grade characteristic of 22 dimensions, are different weight matrices that can be trained, and ; They are respectively the query vector, key vector and value vector of the element grade feature based on the attention mechanism.

4. The intelligent dosing method for lead-zinc ore flotation according to claim 3, characterized in that: The calculation formula of step S3 specifically includes: ； ； in, is the first cross-modal feature, is the second cross-modal feature, and is the temperature coefficient of the hyperparameter, the superscript is the matrix transpose symbol.

5. The intelligent dosing method for lead-zinc ore flotation according to claim 4, characterized in that: The calculation formula of step S5 specifically includes: ； ； ； in, is the time window length, The corresponding time step of the LSTM network The hidden state sequence of the output; is the trainable weight matrix; is a trainable bias term; is the time step Weights in time windows; for Normalized weights after conversion based on the Softmax function; The context vector within the time window is obtained by weighted summation.

6. The intelligent dosing method for lead-zinc ore flotation according to claim 5, characterized in that: The specific calculation formula of step S6 includes: ； in, is the classification task weight matrix; is the bias term for the classification task; Linear output for classification tasks; is the probability of drug type; is a natural constant, represents the standard deviation; ； in, is the regression task weight matrix; is the regression task bias; is the predicted value of drug dosage.

7. The intelligent dosing method for lead-zinc ore flotation according to any one of claims 1 to 6, characterized in that: In the process of training the model for implementing steps S2 to S6, the loss function is composed of the weighted sum of the multi-label binary classification cross entropy loss of the drug type information and the drug dosage mean square loss, and the loss gradient of the model back propagation is obtained by derivatizing the loss.

8. An intelligent dosing system for lead-zinc ore flotation, comprising a memory, a processor, and a computer program stored in the memory and capable of running on the processor, characterized in that: When the processor executes the computer program, the method according to any one of claims 1 to 7 is implemented.

Images