Multi-source information driven froth flotation control parameter dynamic optimization method and system

Abstract

The application provides a multi-source information driven froth flotation control parameter dynamic optimization method and system, which comprises the following steps: inputting an RGB image sequence and a 3D point cloud sequence into a multi-modal data fusion and digital representation module to generate a multivariate digital representation sequence; inputting the multivariate digital representation sequence and a multi-source sensor time sequence into a multi-step time sequence prediction module to generate a multivariate digital representation prediction sequence and a multi-source sensor prediction sequence; inputting the multivariate digital representation prediction sequence and the multi-source sensor prediction sequence into a set value real-time prediction module to predict and output a sensor control unit that needs to be modified at the current time and a specific modification value; and inputting the multi-source sensor prediction sequence and the modification value of the set value of the sensor control unit at the current time into a PID parameter real-time prediction module to predict and output a PID parameter that needs to be adjusted by the sensor control unit that needs to be modified at the current time. The application can dynamically optimize the froth flotation control parameter.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of froth flotation, in particular to a multi-source information driven froth flotation control parameter dynamic optimization method and system. BACKGROUND

[0002] As a common mineral separation method, froth flotation effectively realizes the separation and extraction of minerals by utilizing the difference in hydrophobicity of mineral surfaces, so that the target minerals adhere to the bubbles and float. As a key technology for recovering titanium ore resources, the fine and accurate process regulation of froth flotation is particularly important. By reasonably setting the control parameters of the flotation tank, such as appropriate sensor control unit set values and PID parameters, the recovery efficiency of titanium ore resources can be effectively improved. Although many methods for adjusting the control parameters of froth flotation have emerged, these methods have some problems. On the one hand, the existing adjustment of froth flotation control parameters mainly relies on the regular inspection of on-site operators, and the related parameters of each control unit in the flotation process are manually adjusted according to the on-site production conditions. However, this traditional adjustment method makes it difficult for on-site operators to respond to changes in the production process in a timely and accurate manner, lacks real-time feedback mechanism, and cannot identify and respond to abnormal fluctuations in the flotation process in the first time. On the other hand, the existing control parameter adjustment method usually relies on the manual experience of operators, and adjusts the control parameters through repeated experiments and experience accumulation. However, this method shows weak adaptability and robustness when dealing with complex, dynamic and nonlinear changes in the flotation process. Moreover, manual experience often ignores the coupling relationship between multiple variables in the flotation process, making it difficult to achieve accurate adjustment of control parameters, thereby limiting the improvement of flotation efficiency and the stability of product quality. SUMMARY

[0003] In order to solve the technical problems existing in the prior art, the present application provides a multi-source information driven froth flotation control parameter dynamic optimization method and system, and the technical solution is as follows:

[0004] On the one hand, a multi-source information driven froth flotation control parameter dynamic optimization method is provided, which comprises:

[0005] S1, collecting the RGB image and the corresponding 3D point cloud of the froth in the flotation working area at a time before the current time, and constructing the RGB image sequence and the 3D point cloud sequence in time sequence;

[0006] S2, inputting the RGB image sequence and the 3D point cloud sequence into the multi-modal data fusion and digital representation module to generate a multi-variable digital representation sequence;

[0007] S3, input the multivariate digital characterization sequence and the multi-source sensor time sequence involved in the flotation process into a multi-step time sequence prediction module to generate a multivariate digital characterization prediction sequence and a multi-source sensor prediction sequence after a period of time from the current time;

[0008] S4, input the multivariate digital characterization prediction sequence and the multi-source sensor prediction sequence into a set value real-time prediction module to predict and output the sensor control unit that needs to be modified in the current time and the specific modification value;

[0009] S5, input the multi-source sensor prediction sequence and the modification value of the sensor control unit set value at the current time into a PID parameter real-time prediction module to predict and output the PID parameter that needs to be adjusted by the sensor control unit that needs to be modified in the current time, and realize dynamic optimization of the flotation control parameter driven by multi-source information.

[0010] Optionally, the multi-modal data fusion and digital characterization module is composed of a module encoder and a module decoder, the encoder and the decoder are each composed of four serial processing units, which are used to extract and fuse deep semantic features of RGB images and 3D point clouds layer by layer, each processing unit includes a residual network block and a cross attention module, the residual network block stably extracts deep high-dimensional features of the foam in the RGB image through multi-layer convolution, the cross attention module performs interactive fusion between multi-modal data by taking the RGB image feature as Query and the 3D point cloud feature extracted by three-dimensional convolution at the same time as Key and Value, in the first processing unit and the fourth processing unit of the module encoder, respectively introduce a skip connection to integrate the fused high-dimensional features into the corresponding decoder, realize the reservation of semantic information, and promote the fusion of multi-scale information, finally generate the multivariate digital characterization sequence through the module decoder.

[0011] Optionally, the multi-step time sequence prediction module extracts and captures time-dependent features of the multivariate digital characterization sequence and the multi-source sensor time sequence respectively, and the processing process is as follows:

[0012] First, the input sequence is subjected to word embedding to realize high-dimensional mapping of sequence features and obtain a high-dimensional feature vector;

[0013] After that, the high-dimensional feature vector is sent into the Seq2Seq encoder, which is stacked by multiple LSTM units, each of which is responsible for receiving the high-dimensional feature vector output by the word embedding and combining the hidden state of the previous time step to generate the hidden state of the current time step and the sequence representation feature, the hidden state is used to pass feature information between LSTM units, and the sequence representation feature is input into the self-attention layer with shared weights after normalization processing, which is used to further integrate the high-dimensional features within the sequence, the shared weights mean that all sequences share the same set of parameters, and the global association between the digital representation sequence and the sensor sequence is established to enhance the information interaction ability between sequences and obtain high-level semantic features.

[0014] After that, the high-level semantic features are output to the multi-layer perceptron, and the processing results are sent to the Seq2Seq decoder, which has the same unit structure as the Seq2Seq encoder but runs in the opposite direction, and is used to step-by-step reconstruct or predict the time series, the Seq2Seq decoder not only integrates the context information and global dependency between sequences extracted by the encoder, but also combines the output state of the previous time step to achieve effective modeling and generation of target sequences.

[0015] Finally, after normalization, the prediction output corresponding to the sequence is generated: the multivariate digital representation sequence corresponds to the multivariate digital representation prediction sequence, and the multi-source sensor time series sequence corresponds to the multi-source sensor prediction sequence.

[0016] Optionally, the processing process of the set value real-time prediction module is as follows:

[0017] The multivariate digital representation prediction sequence is input into the self-attention layer respectively to extract the key feature representation corresponding to each sequence, further enhance the expression ability and discriminability of the feature, and then perform channel splicing to form a high-dimensional feature used to guide the sensor control unit, the high-dimensional feature is output through the gating network to obtain the weight parameter W i for each sensor control unit, the weight parameter W i simulates the visual perception of the on-site operator to guide the decision-making of the sensor control unit that needs to modify the set value, W i takes the value of 0 or 1, representing the on or off of the corresponding sensor control unit, wherein on means that the set value of the corresponding sensor control unit needs to be modified, and off means that the set value of the corresponding sensor control unit does not need to be modified, for the sensor control unit that needs to be modified, the specific modification of the set value is performed through the gating decision unit.

[0018] Optionally, the processing process of the gating network is as follows:

[0019] The high-dimensional features are sequentially passed through a fully connected layer, an ELU activation function and a fully connected layer again to extract more discriminative feature expressions, and are stacked and retained through a skip connection to preserve original feature information.

[0020] The feature distribution is then stabilized through a normalization operation, and a 1x1 convolution is used to compress the channel dimension to output the weight parameters W corresponding to each sensor control unit i ;

[0021] The gating decision unit is composed of 4 stacked structure units, which realize mapping modeling from sensor prediction sequences to modified set values, and the processing process of each structure unit is as follows:

[0022] Taking each sensor prediction sequence as input, the original data is projected to a unified semantic space through a word embedding layer, and after stabilizing the feature distribution through layer normalization, the self-attention layer is input to capture the sequence's own dependency relationship, and the original feature information is retained through the introduction of a skip connection in this process.

[0023] Then, after layer normalization, multilayer perception and skip connection, further feature compression and reconstruction are realized.

[0024] Optionally, the PID parameter real-time prediction module predicts the PID parameters that need to be adjusted by the sensor control unit that needs to modify the set value at the current time by fusing the deviation sequence between the sensor prediction sequence and the modified value of the sensor control unit set value at the current time, the difference between the previous and subsequent times, and the cumulative error information. Through deep learning modeling of these information, adaptive optimization and dynamic control of the control strategy are realized, and the stability and response ability of the system under complex working conditions are improved, wherein the PID parameters K P , K I , K D correspond to the current error, the historical cumulative error and the future predicted error respectively, and the specific processing process is as follows:

[0025] The difference between each prediction value in the sensor prediction sequence and the modified value of the sensor control unit set value at the current time is calculated to generate a sensor prediction deviation sequence e(t), and Δe(t) and ∫e(t)dt are further calculated based on this.

[0026] In addition, a sensor parameter knowledge base is introduced to obtain additional knowledge covariates through word embedding, and the knowledge covariates accurately guide the generation of PID parameters according to the characteristics of each sensor control unit.

[0027] The multi-source information composed of the modification value of the set value of the sensing control unit at the current time, e(t), Delta e(t), integral e(t)dt, knowledge covariant is uniformly input to the PID parameter prediction module, the effective fusion between the characteristics is realized by introducing the cross attention mechanism, the perception ability of the model to the key information is enhanced, and the fused characteristics are further subjected to nonlinear mapping and parameter extraction through the full connection layer to predict the PID parameter to be adjusted by the sensing control unit at the current time

[0028] Optionally, the loss function of the prediction of the PID parameter to be adjusted by the sensing control unit at the current time which needs to be set value modified is calculated in a display modeling manner, specifically:

[0029] By taking the current value of the sensor sequence as the starting value, combining the modification value of the set value of the sensing control unit at the current time predicted by the model as the target endpoint, the PID parameter predicted is used to generate a system response curve, then the system response curve is sampled to obtain an output change sequence of the system, and the output change sequence is compared with a preset self-built sequence label, and a loss function L is calculated through the error between the two sequences total , realizing more accurate prediction of the PID parameter with physical meaning and dynamic characteristics;

[0030] The calculation of the loss function L total significantly expands the calculation dimension of the loss function by introducing the system response curve, from the original K P , K I , K D 3 parameter values to multiple response curve sampling points, not only increases the calculation data, but also makes the loss calculation and subsequent back propagation more stable and accurate; at the same time, the process information of the system dynamic response is also integrated, realizing higher accuracy in model feature description, and the loss function L total fully integrates the control-related index parameters, and the formula is as follows:

[0031] L total =α1L tracking +α2L overshoot +α3L response

[0032] Wherein, alpha1, alpha2, alpha3 are weight coefficients corresponding to the sub-loss function, and the specific value is determined through adaptive learning in the model training process;

[0033] L tracking is a tracking error sub-loss, used to reduce the deviation between the output change sequence and the self-built sequence label y ref (t), and improve the accuracy of the system, and T is the time series length, and the formula is:

[0034] L overshoot is a overshoot loss, a part of the modified value for suppressing the output change sequence from exceeding the set value, thereby improving the stability of the system, y set is a modified value of the set value, and the formula is:

[0035] L response is a response speed loss, by encouraging the system to quickly reach the target value, to improve dynamic performance, t reach is the modified value of the time when the set value is first reached, and the formula is:

[0036] In another aspect, a multi-source information driven froth flotation control parameter dynamic optimization system is provided, the system comprising:

[0037] A collection module is configured to collect RGB images and corresponding 3D point clouds of froth in a flotation work area at a time period before the current time, and construct an RGB image sequence and a 3D point cloud sequence in chronological order;

[0038] A first generation module is configured to input the RGB image sequence and the 3D point cloud sequence into a multi-modal data fusion and digital representation module to generate a multi-variable digital representation sequence;

[0039] A second generation module is configured to input the multi-variable digital representation sequence and a multi-source sensor time sequence involved in the flotation process into a multi-step time sequence prediction module to generate a multi-variable digital representation prediction sequence and a multi-source sensor prediction sequence at a time period after the current time;

[0040] A first prediction module is configured to input the multi-variable digital representation prediction sequence and the multi-source sensor prediction sequence into a set value real-time prediction module to predict an output of a sensor control unit that needs to be modified at the current time, and a specific modified value;

[0041] A second prediction module is configured to input the multi-source sensor prediction sequence and a modified value of a set value of a sensor control unit at the current time into a PID parameter real-time prediction module to predict a PID parameter that needs to be adjusted by the sensor control unit that needs to be modified at the current time, thereby realizing multi-source information driven froth flotation control parameter dynamic optimization.

[0042] In another aspect, an electronic device is provided, the electronic device comprising a processor and a memory, the memory having stored therein at least one instruction, the at least one instruction being loaded and executed by the processor to implement the above-mentioned multi-source information driven froth flotation control parameter dynamic optimization method.

[0043] In another aspect, a computer-readable storage medium is provided, the storage medium storing at least one instruction, the at least one instruction being loaded and executed by a processor to implement the above-mentioned multi-source information driven dynamic optimization method of froth flotation control parameters.

[0044] The technical solutions provided by the present application have at least the following beneficial effects:

[0045] 1) The present application breaks through the limitations of traditional single sensor or RGB image analysis, and realizes multi-dimensional perception of froth flotation by integrating RGB, 3D point cloud data and other multi-source information, providing more comprehensive and detailed control parameter adjustment.

[0046] 2) The present application simulates the visual perception of on-site operators with the predicted sequence of digital representation, and uses it as a guide to combine the predicted sequence trend of multi-source sensors to intelligently decide the change range and amplitude of the set value of the sensing control unit, so as to dynamically select the sensing control unit and its corresponding modification value of the set value to be adjusted, and realize real-time optimization and regulation based on working condition characterization information.

[0047] 3) The present application simulates the input mechanism of the classic PID controller by fusing the deviation sequence between the predicted sequence and the modification value of the sensing control unit set value, the difference between the previous and subsequent time points, and the cumulative error and other characteristic information, and realizes adaptive optimization and dynamic regulation of the control strategy through deep learning modeling of these characteristics, thereby improving the stability and response ability of the system under complex working conditions.

[0048] 4) The present application proposes a control system training optimization scheme for display modeling and its corresponding loss function design, which maps the predicted PID parameters to the corresponding response curve to expand the calculation dimension of the loss function and improve the robustness of the model feature description, thereby realizing more accurate parameter prediction, and at the same time, combining the PID control law and dynamic response characteristics to construct a customized loss function, so that the model training process is more in line with the performance index requirements of the control system. BRIEF DESCRIPTION OF DRAWINGS

[0049] In order to more clearly illustrate the technical solutions in the embodiments of the present application, the drawings needed in the embodiment description will be briefly introduced. Obviously, the drawings in the following description are only some embodiments of the present application, and other drawings can be obtained by those skilled in the art without creative labor.

[0050] Figure 1 is a multi-source information driven dynamic optimization method flowchart of froth flotation control parameters provided by the present application;

[0051] Figure 2 is a multi-modal data fusion and digital representation module structural block diagram provided by the embodiment of the application;

[0052] Figure 3 is a multi-step time series prediction module structural block diagram provided by the embodiment of the application;

[0053] Figure 4 is a Seq2Seq encoder-decoder structural block diagram provided by the embodiment of the application;

[0054] Figure 5 is a set value real-time prediction module structural block diagram provided by the embodiment of the application;

[0055] Figure 6 is a gating network structural block diagram provided by the embodiment of the application;

[0056] Figure 7 is a gating decision unit structural block diagram provided by the embodiment of the application;

[0057] Figure 8 is a PID parameter real-time prediction module structural block diagram provided by the embodiment of the application;

[0058] Figure 9 is a display modeling flowchart provided by the embodiment of the application;

[0059] Figure 10 is a multi-source information driven foam flotation control parameter dynamic optimization system block diagram provided by the embodiment of the application;

[0060] Figure 11 is a structural schematic diagram of an electronic device provided by the embodiment of the application. DETAILED DESCRIPTION

[0061] In order to make the technical problems, technical solutions and advantages of the application clearer, the following will be described in detail with reference to the drawings and specific embodiments.

[0062] The embodiment of the application provides a multi-source information driven foam flotation control parameter dynamic optimization method, which can be realized by an electronic device, which can be a terminal or a server. Figure 1 The method flowchart is shown, and the processing flow can include the following steps:

[0063] S1, collect the RGB image and the corresponding 3D point cloud of the foam in the flotation working area before a period of time at the current moment, and construct the RGB image sequence and the 3D point cloud sequence in time sequence;

[0064] The embodiment of the present application collects the RGB image and the corresponding 3D point cloud of the froth in the flotation work area at a time point before the current time tn from different camera devices at the same time point, and constructs the corresponding modal sequence in time sequence. Since the froth has obvious three-dimensional features in the flotation process, it is difficult to obtain sufficient depth information by relying solely on the RGB image, therefore, the embodiment of the present application interacts and fuses the RGB image sequence and the three-dimensional point cloud sequence.

[0065] S2, input the RGB image sequence and the 3D point cloud sequence into the multi-modal data fusion and digital representation module to generate a multi-variable digital representation sequence;

[0066] Optionally, as shown in the figure, Figure 2 The multi-modal data fusion and digital representation module is composed of a module encoder and a module decoder, the encoder and the decoder are both composed of four serial processing units, which are used to extract and fuse the deep semantic features of the RGB image and the 3D point cloud layer by layer, each processing unit includes a residual network block and a cross attention module, the residual network block stably extracts the deep high-dimensional features of the froth in the RGB image through multi-layer convolution, the cross attention module interacts and fuses the multi-modal data by taking the RGB image features as Query and the 3D point cloud features extracted by three-dimensional convolution at the same time as Key and Value, in the first processing unit and the fourth processing unit of the module encoder, the high-dimensional features after fusion are integrated into the corresponding decoder in the subsequent process through the introduction of the skip connection, which realizes the reservation of semantic information and promotes the fusion of multi-scale information, finally, the module decoder generates the multi-variable digital representation sequence (the multi-variable digital representation sequence includes a bubble number representation sequence, a maximum bubble area sequence and other multi-dimensional sequences).

[0067] S3, input the multi-variable digital representation sequence and the time sequence sequence of the multi-source sensor involved in the flotation process into the multi-step time sequence prediction module to generate a multi-variable digital representation prediction sequence and a multi-source sensor prediction sequence after a period of time from the current time;

[0068] Optionally, as shown in the figure, Figure 3 The multi-step time sequence prediction module extracts and captures the time-dependent features of the multi-variable digital representation sequence and the multi-source sensor time sequence (such as the time sequence data collected by the temperature sensor) respectively, and the processing process is as follows:

[0069] First, the input sequence is subjected to word embedding to realize high-dimensional mapping of sequence features and obtain a high-dimensional feature vector;

[0070] Then, the high-dimensional feature vector is sent to the Seq2Seq encoder, as shown in the figure, Figure 4As shown, the Seq2Seq encoder is composed of multiple stacked LSTM units. Each LSTM unit is responsible for receiving the high-dimensional feature vector of the word embedding output and combining it with the hidden state of the previous time step to generate the hidden state and sequence representation features of the current time step. The hidden state is used to pass feature information between LSTM units. The sequence representation features are normalized and then input into a self-attention layer with shared weights to further integrate the high-dimensional features within the sequence. Shared weights mean that all sequences share the same set of parameters, establishing a global association between the digitally represented sequence and the sensor sequence, enhancing the information interaction capability between sequences, and obtaining high-level semantic features.

[0071] Subsequently, the high-level semantic features are output to the multilayer perceptron, and the processing result is sent to the Seq2Seq decoder. The Seq2Seq decoder adopts the same unit structure as the Seq2Seq encoder, but runs in the opposite direction. It is used to progressively reconstruct or predict the time series. The Seq2Seq decoder not only integrates the context information extracted by the encoder and the global dependencies between sequences, but also combines its own output state from the previous time step to achieve effective modeling and generation of the target sequence.

[0072] Finally, after normalization, the corresponding prediction outputs are generated: the multivariate digital representation sequence corresponds to the multivariate digital representation prediction sequence, and the multi-source sensor time series sequence corresponds to the multi-source sensor prediction sequence.

[0073] S4. Input the multivariate digital representation prediction sequence and the multi-source sensor prediction sequence into the set value real-time prediction module, and predict and output the sensor control unit that needs to be modified at the current moment, as well as the specific modification value.

[0074] Optionally, such as Figure 5 As shown, the processing procedure of the real-time prediction module for the setpoint is as follows:

[0075] The multivariate digital representation prediction sequences are input into the attention layer to extract key feature representations corresponding to each sequence, further enhancing the expressive power and discriminative power of the features. Then, channels are concatenated to form high-dimensional features for guiding the sensing control unit (the sensing control unit consists of sensors and corresponding actuators (such as motors, valves, etc.). The sensors are responsible for real-time perception of the process state, and the actuators implement corresponding dynamic adjustments according to the control commands output by the algorithm). The high-dimensional features are passed through a gating network to output weight parameters W for regulating the corresponding weight parameters W of each sensing control unit. i The weight parameter W i The sensor control unit, W, will simulate the visual perception of on-site operators to guide decision-making and require adjustments to setpoints. iValue 0 or 1, respectively, represents the opening or closing of the corresponding sensor control unit, wherein opening means that the corresponding sensor control unit needs to modify the set value, and closing means that the corresponding sensor control unit does not need to modify the set value. For the sensor control unit that needs to be modified, the specific set value modification is performed through the gating decision unit.

[0076] Optionally, as shown in Figure 6 , the processing process of the gating network is:

[0077] The high-dimensional features pass through the fully connected layer, the ELU activation function and the fully connected layer again in turn, extract more discriminative feature expressions, and are stacked and reserved through the jump connection to retain the original feature information;

[0078] Then, the feature distribution is stabilized through the normalization operation, and the channel dimension is compressed through the 1x1 convolution, and the weight parameter W i corresponding to each sensor control unit is output.

[0079] As shown in Figure 7 , the gating decision unit is composed of 4 stacked structure units, which realizes the mapping modeling from the sensor prediction sequence to the modified set value, and the processing process of each structure unit is:

[0080] With each sensor prediction sequence as input, the original data is projected to a unified semantic space through the word embedding layer, and after the feature distribution is stabilized through layer normalization, the self-attention layer is input to capture the sequence itself dependency relationship, and the jump connection is introduced in this process to retain the original feature information;

[0081] Then, after layer normalization, multilayer perception and jump connection, further feature compression and reconstruction are realized.

[0082] S5, input the multi-source sensor prediction sequence and the modification value of the current time sensor control unit set value into the PID parameter real-time prediction module, predict the PID parameter that the sensor control unit needs to adjust at the current time, realize the dynamic optimization of the foam flotation control parameter driven by multi-source information.

[0083] 1、Optionally, as shown in Figure 8 , the PID parameter real-time prediction module predicts the PID parameter that the sensor control unit needs to adjust at the current time by fusing the deviation sequence between the sensor prediction sequence and the modification value of the current time sensor control unit set value, the difference between the previous and subsequent time and the cumulative error information. Through deep learning modeling of these information, adaptive optimization and dynamic control of the control strategy are realized, and the stability and response ability of the system under complex working conditions are improved, wherein the PID parameters K P , K I , K DCorresponding to the current error, the historical cumulative error and the future prediction error respectively (the embodiment of the application predicts the PID parameters commonly used in industrial control, which not only can be compatible with the existing control framework, but also can realize higher precision control effect relying on algorithm driving), the specific processing process is as follows:

[0084] The difference between each prediction value in the sensor prediction sequence and the modification value of the current time sensor control unit set value is calculated to generate a sensor prediction deviation sequence e(t), and on this basis, Δe(t) and ∫e(t)dt are further calculated (where Δe(t) can be quantized as e(t)-e(t-1), and ∫e(t)dt can be quantized as the superposition of e(t));

[0085] In addition, a sensor parameter knowledge base (such as the inherent attributes of the corresponding sensor device model) is introduced, and additional knowledge covariates are obtained through word embedding, which accurately guide the generation of PID parameters according to the characteristics of each sensor control unit;

[0086] The multi-source information composed of the modification value of the current time sensor control unit set value, e(t), Δe(t), ∫e(t)dt, and knowledge covariates is uniformly input into the PID parameter prediction module, the cross-attention mechanism is introduced to realize effective fusion between features, and the perception ability of the model to key information is enhanced, and the fused features are further mapped and parameter extracted through the full connection layer, and the PID parameters adjusted by the sensor control unit at the current time are predicted

[0087] Since the prediction target of the overall scheme of the embodiment of the application has two, one is for the prediction of the sensor control unit set value, and the other is for the prediction of the sensor PID parameter. For the prediction of the sensor control unit set value, it has a clear numerical target and a relatively stable mapping relationship, so the mean square error (MSE) loss function can be directly used for training, and the loss function form is as follows:

[0088]

[0089] Where y i and f(x i ) represent the label and prediction value of the set value respectively, and m is the sample number.

[0090] In contrast, the prediction of the sensor PID parameter is more complex, these parameters have a significant impact on the stability and response speed of the system, and the control range of each parameter is wide and the degree of mutual coupling is high, directly using the PID parameter as the label for mean square error calculation, in terms of model convergence speed and prediction accuracy, it faces great challenges, therefore, the embodiment of the application designs a display modeling method to calculate the loss function of the prediction of the PID parameter.

[0091] Optionally, as shown in Figure 9 The loss function of the prediction of the PID parameters that the sensor control unit needs to adjust at the current time point with the set value modification is calculated by a display modeling method, specifically:

[0092] By taking the current value of the sensor sequence as the starting value, combining the modification value of the set value of the sensor control unit at the current time point predicted by the model as the target endpoint, using the predicted PID parameters to generate a system response curve, then sampling the system response curve to obtain the output change sequence of the system, and then comparing the output change sequence with the preset self-built sequence label, the loss function L total is calculated through the error between the two sequences, which realizes more accurate prediction of the PID parameters with physical meaning and dynamic characteristics;

[0093] The calculation of the loss function L total significantly expands the calculation dimension of the loss function (compared to directly calculating the loss function between the predicted PID parameters and the real PID parameters) by introducing the system response curve, from the original K P , K I , K D 3 parameter values to multiple response curve sampling points, not only increasing the calculation data, making the loss calculation and subsequent back propagation more stable and accurate, but also incorporating the process information of the system dynamic response, achieving higher accuracy in model feature description. The loss function L total fully incorporates control-related index parameters, and the formula is as follows:

[0094] L total = α1L tracking + α2L overshoot + α3L response

[0095] Wherein α1, α2, α3 are weight coefficients corresponding to the sub-loss function, and the specific values are determined through adaptive learning in the model training process;

[0096] L tracking is the tracking error sub-loss, which is used to reduce the deviation between the output change sequence and the self-built sequence label y ref (t), and improve the accuracy of the system, T is the time series length, and the formula is:

[0097] L overshoot is the overshoot sub-loss, which is used to suppress the part of the output change sequence that exceeds the modification value of the set value, thereby improving the stability of the system, y set is the modification value of the set value, and the formula is:

[0098] L response To respond to the speed sub-loss, by encouraging the system to quickly reach the target value, to improve the dynamic performance, t reach The time when the modified value reaches the set value for the first time is t, and the formula is:

[0099] The overall training process of the overall model of the embodiment of the application is composed of a multi-modal data fusion and digital representation module, a multi-step time series prediction module, a set value real-time prediction module, and a PID parameter real-time prediction module:

[0100] At any current time, the model first models and analyzes the multi-variable digital representation sequence and the multi-source sensor time sequence before the current time, generates a multi-variable digital representation prediction sequence and a multi-source sensor prediction sequence after the current time, and on this basis, the model predicts the sensor control unit that needs to be modified based on the prediction result. The set value at the current time, and the specific modification value;

[0101] Then, the model further predicts the PID parameter that needs to be adjusted by the sensor control unit that needs to be modified at the current time by using the predicted sensor sequence.

[0102] Therefore, when the model generates the set value and the PID parameter at the current time according to the future prediction sequence, the loss function design adopts a differentiated strategy: the loss of the sensor set value adopts the mean square error (MSE) function, and directly returns the error with its label; and the loss of the PID parameter is realized by an explicit modeling method, taking the current value of the sensor sequence as the starting point, combining the modification value of the set value, simulating the system response curve and sampling, comparing the obtained output change sequence with the defined sequence label to calculate the loss L total , to realize effective prediction of the PID parameter with higher accuracy.

[0103] As Figure 10 shown, the embodiment of the application also provides a multi-source information driven foam flotation control parameter dynamic optimization system, which comprises:

[0104] The acquisition and construction module 1010 is used to acquire the RGB image and the corresponding 3D point cloud of the foam in the flotation working area before a period of time at the current time, and construct the RGB image sequence and the 3D point cloud sequence in time sequence;

[0105] The first generation module 1020 is used to input the RGB image sequence and the 3D point cloud sequence into the multi-modal data fusion and digital representation module to generate a multi-variable digital representation sequence;

[0106] The second generation module 1030 is configured to input the multi-variable digitalized characterization sequence and the multi-source sensor time sequence involved in the flotation process into a multi-step time sequence prediction module to generate a multi-variable digitalized characterization prediction sequence and a multi-source sensor prediction sequence after a period of time from the current time.

[0107] The first prediction module 1040 is configured to input the multi-variable digitalized characterization prediction sequence and the multi-source sensor prediction sequence into a set value real-time prediction module to predict an output of a sensing control unit that needs to be modified in the set value at the current time and a specific modification value.

[0108] The second prediction module 1050 is configured to input the multi-source sensor prediction sequence and the modification value of the set value of the sensing control unit at the current time into a PID parameter real-time prediction module to predict an output of a PID parameter that needs to be adjusted for the sensing control unit that needs to be modified in the set value at the current time, so as to realize dynamic optimization of the control parameter of the froth flotation driven by the multi-source information.

[0109] The multi-source information driven froth flotation control parameter dynamic optimization system provided by the embodiment of the present application corresponds to the multi-source information driven froth flotation control parameter dynamic optimization method provided by the embodiment of the present application in terms of functional structure, and will not be described here.

[0110] Figure 11 FIG. 11 is a structural schematic diagram of an electronic device 1100 provided by the embodiment of the present application. The electronic device 1100 can have relatively large differences due to different configurations or performances, and can include one or more processors (central processing units, CPUs) 1101 and one or more memories 1102. The memory 1102 stores at least one instruction, which is loaded and executed by the processor 1101 to realize the steps of the multi-source information driven froth flotation control parameter dynamic optimization method.

[0111] In the exemplary embodiment, a computer readable storage medium, such as a memory including instructions, is also provided. The instructions can be executed by a processor in a terminal to complete the multi-source information driven froth flotation control parameter dynamic optimization method. For example, the computer readable storage medium can be a ROM, a random access memory (RAM), a CD-ROM, a magnetic tape, a floppy disk, an optical data storage device, and the like.

[0112] Those of ordinary skill in the art can understand that all or part of the steps of the above-mentioned embodiments can be completed by hardware, or by a program instructing related hardware, and the program can be stored in a computer readable storage medium, such as a read-only memory, a magnetic disk or an optical disk.

[0113] The above description is only the preferred embodiment of the present application, and is not intended to limit the present application. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present application shall be included in the protection scope of the present application.

Claims

1. A method for dynamic optimization of control parameters in foam flotation driven by multi-source information, characterized in that, The method includes: S1. Collect RGB images and corresponding 3D point clouds of the flotation working area foams from a period of time before the current moment, and construct RGB image sequence and 3D point cloud sequence in chronological order. S2. Input the RGB image sequence and 3D point cloud sequence into the multimodal data fusion and digital representation module to generate a multivariate digital representation sequence; S3. Input the multivariate digital characterization sequence and the multi-source sensor time series sequence involved in the flotation process into the multi-step time series prediction module to generate a multivariate digital characterization prediction sequence and a multi-source sensor prediction sequence for a period of time after the current moment. S4. Input the multivariate digital representation prediction sequence and the multi-source sensor prediction sequence into the set value real-time prediction module, and predict and output the sensor control unit that needs to be modified at the current moment, as well as the specific modification value. S5. Input the multi-source sensor prediction sequence and the modified value of the current sensor control unit set value into the PID parameter real-time prediction module, predict and output the PID parameter that the sensor control unit that needs to be modified at the current time should adjust, so as to realize the dynamic optimization of the foam flotation control parameters driven by multi-source information. The processing procedure of the real-time prediction module for the set value is as follows: The multivariate digital representation prediction sequences are input into the attention layer to extract key feature representations corresponding to each sequence, further enhancing the expressive power and discriminative power of the features. Then, channel concatenation is performed to form high-dimensional features for guiding the sensor control unit. The high-dimensional features are passed through a gating network to output weight parameters W for regulating the corresponding weights of each sensor control unit. i The weight parameter W i The sensor control unit, W, will simulate the visual perception of on-site operators to guide decision-making and require adjustments to setpoints. i The value is 0 or 1, which represents the on or off state of the corresponding sensor control unit. On means that the set value of the corresponding sensor control unit needs to be modified, and off means that the set value of the corresponding sensor control unit does not need to be modified. For the sensor control unit that needs to be modified, the specific set value is modified through the gating decision unit. The real-time PID parameter prediction module, by fusing the deviation sequence between the sensor prediction sequence and the modified setpoint of the sensor control unit at the current moment, the time difference between previous and subsequent moments, and the cumulative error information, predicts the PID parameters that the sensor control unit should adjust at the current moment when the setpoint needs to be modified. Through deep learning modeling of this information, it achieves adaptive optimization and dynamic control of the control strategy, improving the system's stability and responsiveness under complex operating conditions. The PID parameter K... P K I K D These correspond to the current error, historical cumulative error, and future prediction error, respectively.

2. The method according to claim 1, characterized in that, The multimodal data fusion and digital representation module consists of a module encoder and a module decoder. Each encoder and decoder comprises four cascaded processing units used to extract and fuse deep semantic features from RGB images and 3D point clouds layer by layer. Each processing unit includes a residual network block and a cross-attention module. The residual network block stably extracts deep, high-dimensional features of bubbles in the RGB image through multiple convolutions. The cross-attention module performs interactive fusion between multimodal data by using RGB image features as the query and 3D point cloud features extracted simultaneously through three-dimensional convolution as the key and value. Skip connections are introduced in the first and fourth processing units of the module encoder to integrate the fused high-dimensional features into the corresponding subsequent decoders, thereby preserving semantic information and promoting the fusion of multi-scale information. Finally, the module decoder generates the multivariate digital representation sequence.

3. The method according to claim 1, characterized in that, The multi-step time series prediction module extracts and captures time-dependent features from the multivariate digital representation sequence and the multi-source sensor time series, respectively. The processing procedure is as follows: First, word embedding is performed on the input sequence to achieve high-dimensional mapping of sequence features, resulting in a high-dimensional feature vector. The high-dimensional feature vector is then fed into a Seq2Seq encoder, which consists of multiple stacked LSTM units. Each LSTM unit is responsible for receiving the high-dimensional feature vector output by word embedding and combining it with the hidden state of the previous time step to generate the hidden state and sequence representation features of the current time step. The hidden state is used to pass feature information between LSTM units. The sequence representation features are normalized and then input into a self-attention layer with shared weights to further integrate the high-dimensional features within the sequence. Shared weights mean that all sequences share the same set of parameters, establishing a global association between the digitally represented sequence and the sensor sequence, enhancing the information interaction capability between sequences, and obtaining high-level semantic features. Subsequently, the high-level semantic features are output to the multilayer perceptron, and the processing result is sent to the Seq2Seq decoder. The Seq2Seq decoder adopts the same unit structure as the Seq2Seq encoder, but runs in the opposite direction. It is used to progressively reconstruct or predict the time series. The Seq2Seq decoder not only integrates the context information extracted by the encoder and the global dependencies between sequences, but also combines its own output state from the previous time step to achieve effective modeling and generation of the target sequence. Finally, after normalization, the corresponding prediction outputs are generated: the multivariate digital representation sequence corresponds to the multivariate digital representation prediction sequence, and the multi-source sensor time series sequence corresponds to the multi-source sensor prediction sequence.

4. The method according to claim 1, characterized in that, The processing procedure of the gating network is as follows: The high-dimensional features are sequentially passed through a fully connected layer, an ELU activation function, and a fully connected layer to extract more discriminative feature representations, and the original feature information is preserved by skip connections. Subsequently, the feature distribution is stabilized through normalization, and the channel dimension is compressed using 1×1 convolution to output the weight parameters W used to regulate the corresponding sensor control units. i ; The gated decision unit consists of four stacked structural units, enabling mapping modeling from sensor-predicted sequences to modified setpoints. The processing procedure for each structural unit is as follows: Using the predicted sequences from each sensor as input, the original data is projected onto a unified semantic space through a word embedding layer. After the feature distribution is stabilized by layer normalization, the input is fed into the attention layer to capture the dependencies of the sequence itself. In this process, skip connections are introduced to preserve the original feature information. Then, through layer normalization, multilayer perceptron and skip connections, further feature compression and reconstruction are achieved.

5. The method according to claim 1, characterized in that, The specific processing procedure of the real-time prediction module for PID parameters is as follows: Calculate the difference between each predicted value in the sensor prediction sequence and the modified value of the sensor control unit setpoint at the current time, generate the sensor prediction deviation sequence e(t), and further calculate Δe(t) and ∫e(t)dt based on this; In addition, a sensor parameter knowledge base is introduced, and additional knowledge covariates are obtained through word embedding. These knowledge covariates accurately guide the generation of PID parameters based on the characteristics of each sensor control unit. The multi-source information, consisting of the modified setpoint of the current sensor control unit, e(t), Δe(t), ∫e(t)dt, and knowledge covariates, is uniformly input into the PID parameter prediction module. A cross-attention mechanism is introduced to achieve effective feature fusion, enhancing the model's ability to perceive key information. The fused features are further processed through a fully connected layer for nonlinear mapping and parameter extraction, predicting and outputting the PID parameters that the sensor control unit should adjust at the current moment.

6. The method according to claim 1, characterized in that, The loss function for predicting the PID parameters that the sensor control unit, which requires setpoint modification at the current moment, should adjust, is calculated through an explicit modeling method, specifically as follows: Starting with the current value of the sensor sequence and using the predicted modified value of the sensor control unit setpoint at the current moment as the target endpoint, a system response curve is generated using the predicted PID parameters. The system response curve is then sampled to obtain the system's output change sequence. This output change sequence is then compared with a preset self-built sequence label, and the loss function L is calculated based on the error between the two sequences. total This enables more accurate prediction of PID parameters with greater physical meaning and dynamic characteristics; The loss function L total The calculation significantly expands the computational dimension of the loss function by introducing the system response curve, from the original K... P K I K D Extending the three parameter values ​​to multiple response curve sampling points not only increases the amount of computational data, making loss calculation and subsequent backpropagation more robust and accurate, but also incorporates process information of the system's dynamic response, achieving higher accuracy in model feature characterization. The loss function L... total The relevant control parameters are fully incorporated, and the formula is as follows: L total =α1L tracking +α2L overshoot +α3L response Where α1, α2, and α3 are the weight coefficients of the corresponding sub-loss functions, and their specific values ​​are determined through adaptive learning during the model training process; L tracking To track the error sub-loss, used to reduce the output variation sequence With self-built sequence labels y ref The deviation between (t) improves the accuracy of the system, where T is the length of the time series, and the formula is: L overshoot Overtuning loss is used to suppress the portion of the output variation sequence that exceeds a set value, thereby improving system stability. set The formula for modifying the set value is: L response To address the speed sub-loss, dynamic performance is improved by encouraging the system to quickly reach the target value. reach The time when the modified value first reaches the set value is given by the following formula:

7. A multi-source information-driven dynamic optimization system for foam flotation control parameters, characterized in that, The system includes: The acquisition and construction module is used to acquire RGB images and corresponding 3D point clouds of the flotation working area foams before the current moment, and construct them into RGB image sequences and 3D point cloud sequences in chronological order. The first generation module is used to input the RGB image sequence and the 3D point cloud sequence into the multimodal data fusion and digital representation module to generate a multivariate digital representation sequence. The second generation module is used to input the multivariate digital representation sequence and the multi-source sensor time series sequence involved in the flotation process into the multi-step time series prediction module to generate a multivariate digital representation prediction sequence and a multi-source sensor prediction sequence for a period of time after the current moment. The first prediction module is used to input the multivariate digital representation prediction sequence and the multi-source sensor prediction sequence into the set value real-time prediction module, and predict and output the sensor control unit that needs to be modified at the current moment, as well as the specific modification value. The second prediction module is used to input the prediction sequence of the multi-source sensor and the modified value of the set value of the sensor control unit at the current moment into the real-time prediction module of the PID parameter, and predict and output the PID parameter that the sensor control unit that needs to be modified at the current moment should adjust, so as to realize the dynamic optimization of the control parameters of the foam flotation driven by multi-source information. The processing procedure of the real-time prediction module for the set value is as follows: The multivariate digital representation prediction sequences are input into the attention layer to extract key feature representations corresponding to each sequence, further enhancing the expressive power and discriminative power of the features. Then, channel concatenation is performed to form high-dimensional features for guiding the sensor control unit. The high-dimensional features are passed through a gating network to output weight parameters W for regulating the corresponding weights of each sensor control unit. i The weight parameter W i The sensor control unit, W, will simulate the visual perception of on-site operators to guide decision-making and require adjustments to setpoints. i The value is 0 or 1, which represents the on or off state of the corresponding sensor control unit. On means that the set value of the corresponding sensor control unit needs to be modified, and off means that the set value of the corresponding sensor control unit does not need to be modified. For the sensor control unit that needs to be modified, the specific set value is modified through the gating decision unit. The real-time PID parameter prediction module, by fusing the deviation sequence between the sensor prediction sequence and the modified setpoint of the sensor control unit at the current moment, the time difference between previous and subsequent moments, and the cumulative error information, predicts the PID parameters that the sensor control unit should adjust at the current moment when the setpoint needs to be modified. Through deep learning modeling of this information, it achieves adaptive optimization and dynamic control of the control strategy, improving the system's stability and responsiveness under complex operating conditions. The PID parameter K... P K I K D These correspond to the current error, historical cumulative error, and future prediction error, respectively.

8. An electronic device comprising a processor and a memory, wherein the memory stores at least one instruction, characterized in that, The processor loads and executes at least one instruction to implement the multi-source information-driven dynamic optimization method for foam flotation control parameters as described in any one of claims 1-6.

9. A computer-readable storage medium storing at least one instruction, characterized in that, The at least one instruction is loaded and executed by the processor to implement the multi-source information-driven dynamic optimization method for foam flotation control parameters as described in any one of claims 1-6.

Images