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

Abstract

The invention provides a multi-source information driven foam flotation control parameter dynamic optimization method and system. The method comprises the steps that an RGB image sequence and a 3D point cloud sequence are input into a multi-modal data fusion and digital representation module to generate a multivariable digital representation sequence; inputting the multivariable digital representation sequence and the multi-source sensor sequential sequence into a multi-step time sequence prediction module to generate a multivariable digital representation prediction sequence and a multi-source sensor prediction sequence; inputting the multivariable digital representation prediction sequence and the multi-source sensor prediction sequence into a set value real-time prediction module, and predicting and outputting a sensing control unit needing to modify the set value at the current moment and a specific modified value; and inputting the prediction sequence of the multi-source sensor and the modified value of the set value of the sensing control unit at the current moment into a PID parameter real-time prediction module, and predicting and outputting a PID parameter which needs to be adjusted by the sensing control unit which needs to be subjected to set value modification at the current moment. According to the method, the foam flotation control parameters can be dynamically optimized.

Description

Technical Field

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

[0002] Froth flotation, a commonly used mineral separation method, exploits differences in surface hydrophobicity to cause target minerals to attach to bubbles and float upward, effectively achieving separation and extraction. As a key technology for recovering titanium ore resources, refined and precise process control is crucial. Properly setting flotation cell control parameters, such as appropriate sensor control unit setpoints and PID parameters, can effectively improve titanium ore recovery efficiency. While numerous methods for adjusting froth flotation control parameters have emerged, most of these methods suffer from several challenges. For one thing, existing froth flotation control parameter adjustments rely primarily on regular inspections by field operators, who manually adjust the parameters of various control units during the flotation process based on on-site production conditions. However, this traditional adjustment method makes it difficult for field operators to respond promptly and accurately to changes in the production process. The lack of a real-time feedback mechanism prevents them from promptly identifying and addressing abnormal fluctuations in the flotation process. Furthermore, existing control parameter adjustment methods often rely on manual operator experience, requiring repeated experimentation and accumulated experience to adjust control parameters. However, this approach exhibits limited adaptability and robustness when dealing with the complex, dynamic, and nonlinear operating conditions of flotation processes. Furthermore, manual experience often overlooks the coupling relationships between multiple variables in the flotation process, making it difficult to precisely adjust control parameters, thereby limiting improvements in flotation efficiency and the stability of product quality. Summary of the Invention

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

[0004] In one aspect, a method for dynamic optimization of froth flotation control parameters driven by multi-source information is provided, the method comprising:

[0005] S1, collecting RGB images and corresponding 3D point clouds of the flotation work area foam for a period of time before the current moment, and constructing them into RGB image sequences and 3D point cloud sequences in chronological order;

[0006] S2. Inputting the RGB image sequence and the 3D point cloud sequence into a multimodal data fusion and digital representation module to generate a multivariate digital representation sequence;

[0007] S3, inputting the multivariable digital representation sequence and the multi-source sensor time series sequence involved in the flotation process into a multi-step time series prediction module to generate a multivariable digital representation prediction sequence and a multi-source sensor prediction sequence for a period of time after the current moment;

[0008] S4, inputting the multivariable digital representation prediction sequence and the multi-source sensor prediction sequence into a set value real-time prediction module, predicting and outputting the sensor control unit that needs to modify the set value at the current moment, and the specific modification value;

[0009] S5. Input the multi-source sensor prediction sequence and the modified value of the sensor control unit setting value at the current moment into the PID parameter real-time prediction module, predict and output the PID parameters that should be adjusted by the sensor control unit whose setting value needs to be modified at the current moment, and realize dynamic optimization of foam flotation control parameters driven by multi-source information.

[0010] Optionally, the multimodal data fusion and digital representation module is composed of a module encoder and a module decoder. The encoder and decoder are both composed of 4 processing units connected in series, which are used to extract and fuse the 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 the deep high-dimensional features of the foam in the RGB image through multiple layers of convolution. The cross-attention module uses the RGB image features as the query and the 3D point cloud features extracted by three-dimensional convolution at the same time as the key and value to perform interactive fusion between multimodal data. In the first processing unit and the fourth processing unit of the module encoder, jump connections are respectively introduced to integrate the fused high-dimensional features into the subsequent corresponding decoders to achieve the retention of semantic information and promote the fusion of multi-scale information. Finally, the multivariate digital representation sequence is generated through the module decoder.

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

[0012] First, embed the input sequence into words to achieve high-dimensional mapping of sequence features and obtain high-order feature vectors;

[0013] The high-dimensional feature vector is then fed into a Seq2Seq encoder, which is composed of a stack of multiple LSTM units. Each LSTM unit is responsible for receiving the high-dimensional feature vector output by the 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 transfer feature information between LSTM units. The sequence representation features are normalized and input into the 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.

[0014] The high-level semantic features are then output to a multi-layer perceptron, and the processed results are fed into a Seq2Seq decoder. The Seq2Seq decoder uses the same unit structure as the Seq2Seq encoder, but operates in the opposite direction, and is used to gradually 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.

[0015] Finally, after normalization again, the prediction output of the corresponding 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:

[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 enhancing the expressiveness and discriminability of the feature. Then, channel splicing is performed to form a high-dimensional feature for guiding the sensor control unit. The high-dimensional feature passes through the gated network and outputs the weight parameter W corresponding to each sensor control unit. i , the weight parameter W i The sensor control unit that simulates the visual perception of the on-site operator and guides the decision to modify the set value is used. i The value 0 or 1 represents the on or off of the corresponding sensor control unit, where on means that the setting value of the corresponding sensor control unit needs to be modified, and off means that the specified 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 setting value is modified through the gating decision unit.

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

[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 the original feature information is retained through skip connection superposition;

[0020] Then, the feature distribution is stabilized by normalization operation, and the channel dimension is compressed by 1×1 convolution to output the weight parameters W corresponding to each sensor control unit. i ;

[0021] The gate decision unit is composed of four stacked structural units to achieve mapping modeling from sensor prediction sequence to modified set values. The processing process of each structural unit is as follows:

[0022] Taking the prediction sequence of each sensor as input, the original data is projected into a unified semantic space through the word embedding layer. After layer normalization to stabilize the feature distribution, it is input into the self-attention layer to capture the sequence's own dependencies. In this process, skip connections are introduced to preserve the original feature information.

[0023] After that, layer normalization, multi-layer perceptron and skip connection are performed to achieve further feature compression and reconstruction.

[0024] Optionally, the PID parameter real-time prediction module predicts the PID parameters that should be adjusted by the sensor control unit that needs to modify the set value at the current moment by fusing the deviation sequence between the sensor prediction sequence and the modified value of the sensor control unit set value at the current moment, the difference between the previous and next moments, and the accumulated error information. Through deep learning modeling of these information, adaptive optimization and dynamic regulation of the control strategy are achieved, thereby improving the stability and responsiveness of the system under complex working conditions. Among them, the PID parameter K P , K I , K D They correspond to the current error, historical cumulative error and future prediction error respectively. The specific processing process is:

[0025] Calculate the difference between each predicted value in the sensor prediction sequence and the modified value of the sensor control unit setting value at the current moment to generate the sensor prediction deviation sequence e(t), and further calculate Δe(t) and ∫e(t)dt based on this;

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

[0027] The multi-source information consisting of the modified value of the sensor control unit setting value at the current moment, e(t), Δe(t), ∫e(t)dt, and knowledge covariates is uniformly input into the PID parameter prediction module. By introducing the cross-attention mechanism, effective fusion between features is achieved to enhance the model's perception of key information. The fused features are further subjected to nonlinear mapping and parameter extraction by the fully connected layer to predict and output the PID parameters that the sensor control unit should adjust at the current moment.

[0028] Optionally, the predicted loss function of the PID parameters that should be adjusted by the sensor control unit that needs to modify the set value at the current moment is calculated by an explicit modeling method, specifically:

[0029] By taking the current value of the sensor sequence as the starting value and combining the modified value of the sensor control unit setting value at the current moment predicted by the model as the target end point, the predicted PID parameters are used to generate the system response curve, and then the system response curve is sampled to obtain the output change sequence of the system, and then the output change sequence is compared with the preset self-built sequence label, and the loss function L is calculated by the error between the two sequences. total , to achieve more accurate prediction of PID parameters with more physical meaning and dynamic characteristics;

[0030] The loss function L total The calculation of the loss function is significantly expanded by introducing the system response curve, from the original K P , K I , K D The three parameter values ​​are extended to multiple response curve sampling points, which not only increases the amount of calculation data, making the loss calculation and subsequent back propagation more robust and accurate; it also incorporates the process information of the system dynamic response, achieving higher accuracy in model feature characterization. The loss function L total Fully integrating the control-related indicator parameters, the formula is as follows:

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

[0032] Among them, α1, α2, and α3 are the weight coefficients of the corresponding sub-loss functions, and their specific values ​​are determined by adaptive learning during model training;

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

[0034] L overshoot The over-tuning loss is used to suppress the part of the output change sequence that exceeds the modified value of the set value, thereby improving the stability of the system. set is the modified value of the set value, the formula is:

[0035] L response In order to respond to the speed loss, the system is encouraged to reach the target value quickly to improve the dynamic performance, t reach The time when the modified value of the set value is first reached is:

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

[0037] The acquisition and construction module is used to collect the RGB image and corresponding 3D point cloud of the flotation work area foam in a period of time before the current moment, and construct them into RGB image sequence and 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 multimodal data fusion and digital representation module to generate a multivariate digital representation sequence;

[0039] The second generation module is used to input the multivariable 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 multivariable digital representation prediction sequence and a multi-source sensor prediction sequence for a period of time after the current moment;

[0040] A first prediction module is used to input the multivariable digital representation prediction sequence and the multi-source sensor prediction sequence into a set value real-time prediction module, and predict and output the sensor control unit that needs to modify the set value at the current moment, as well as the specific modification value;

[0041] The second prediction module is used to input the multi-source sensor prediction sequence and the modified value of the sensor control unit setting value at the current moment into the PID parameter real-time prediction module, predict and output the PID parameters that should be adjusted by the sensor control unit that needs to modify the setting value at the current moment, and realize dynamic optimization of foam flotation control parameters driven by multi-source information.

[0042] On the other hand, an electronic device is provided, comprising a processor and a memory, wherein the memory stores at least one instruction, and the at least one instruction is loaded and executed by the processor to implement the above-mentioned multi-source information-driven dynamic optimization method for froth flotation control parameters.

[0043] On the other hand, a computer-readable storage medium is provided, wherein the storage medium stores at least one instruction, and the at least one instruction is loaded and executed by a processor to implement the above-mentioned multi-source information-driven dynamic optimization method for froth flotation control parameters.

[0044] The beneficial effects brought about by the technical solution provided by the present invention include at least:

[0045] 1) This invention breaks through the limitations of traditional single sensors or RGB image analysis. By integrating multi-source information such as RGB and 3D point cloud data, it realizes multi-dimensional perception of froth flotation and provides more comprehensive and detailed control parameter adjustment.

[0046] 2) The present invention simulates the visual perception of on-site operators with a digitally represented prediction sequence, and uses this as a guide, combined with the prediction sequence trend of multi-source sensors, to intelligently decide the range and amplitude of the setting value of the sensor control unit, thereby dynamically selecting the sensor control unit to be adjusted and the modification value of its corresponding setting value, realizing real-time optimization and control based on the working condition characterization information.

[0047] 3) The present invention simulates the input mechanism of the classic PID controller by fusing characteristic information such as the deviation sequence between the prediction sequence and the modified value of the sensor control unit set value, the difference between the previous and next moments, and the accumulated error. Through deep learning modeling of these characteristics, the adaptive optimization and dynamic regulation of the control strategy are achieved, thereby improving the stability and responsiveness of the system under complex working conditions.

[0048] 4) The present invention proposes a control system training optimization scheme for display modeling and its corresponding loss function design. By mapping the predicted PID parameters to the corresponding response curve, the calculation dimension of the loss function is expanded, and the robustness of the model feature characterization is improved, thereby achieving more accurate parameter prediction. At the same time, a customized loss function is constructed by combining the PID control law and the dynamic response characteristics, so that the model training process is more in line with the control system performance index requirements. BRIEF DESCRIPTION OF THE DRAWINGS

[0049] In order to more clearly illustrate the technical solutions in the embodiments of the present invention, the following briefly introduces the drawings required for use in the description of the embodiments. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without creative work.

[0050] Figure 1 This is a flow chart of a multi-source information-driven dynamic optimization method for froth flotation control parameters provided by an embodiment of the present invention;

[0051] Figure 2 This is a structural block diagram of a multimodal data fusion and digital representation module provided by an embodiment of the present invention;

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

[0053] Figure 4 This is a block diagram of the Seq2Seq codec structure provided by an embodiment of the present invention;

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

[0055] Figure 6 This is a block diagram of the gate control network structure provided by an embodiment of the present invention;

[0056] Figure 7 This is a structural block diagram of a gate control decision unit provided by an embodiment of the present invention;

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

[0058] Figure 9 This is a display modeling flow chart provided by an embodiment of the present invention;

[0059] Figure 10 This is a block diagram of a multi-source information-driven froth flotation control parameter dynamic optimization system provided by an embodiment of the present invention;

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

[0061] In order to make the technical problems, technical solutions and advantages to be solved by the present invention clearer, a detailed description will be given below with reference to the accompanying drawings and specific embodiments.

[0062] An embodiment of the present invention provides a multi-source information-driven method for dynamic optimization of froth flotation control parameters. The method can be implemented by an electronic device, which can be a terminal or a server. Figure 1 The flowchart of the method is shown, and the processing flow may include the following steps:

[0063] S1, collecting RGB images and corresponding 3D point clouds of the flotation work area foam for a period of time before the current moment, and constructing them into RGB image sequences and 3D point cloud sequences in chronological order;

[0064] In an embodiment of the present invention, different camera devices at the same time point collect RGB images and corresponding 3D point clouds of the foam in the flotation work area for a period of time before the current time tn, and construct them into a corresponding modal sequence in chronological order. Since the foam in the flotation process has obvious three-dimensional characteristics, it is difficult to obtain sufficient depth information by relying solely on RGB images. Therefore, in an embodiment of the present invention, the RGB image sequence and the three-dimensional point cloud sequence are interactively fused.

[0065] S2. Inputting the RGB image sequence and the 3D point cloud sequence into a multimodal data fusion and digital representation module to generate a multivariate digital representation sequence;

[0066] Alternatively, as Figure 2 As shown in the figure, the multimodal data fusion and digital representation module is composed of a module encoder and a module decoder. The encoder and decoder are both composed of 4 processing units in series, which are used to extract and fuse the 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 the deep high-dimensional features of the foam in the RGB image through multi-layer convolution. The cross-attention module uses the RGB image features as the query and the 3D point cloud features extracted by three-dimensional convolution at the same time as the key and value to perform interactive fusion between multimodal data. In the first processing unit and the fourth processing unit of the module encoder, jump connections are introduced respectively to integrate the fused high-dimensional features into the subsequent corresponding decoders to achieve the retention of semantic information and promote the fusion of multi-scale information. Finally, the multivariate digital representation sequence (the multivariate digital representation sequence includes multidimensional sequences such as the bubble number representation sequence and the maximum bubble area sequence) is generated through the module decoder.

[0067] S3, inputting the multivariable digital representation sequence and the multi-source sensor time series sequence involved in the flotation process into a multi-step time series prediction module to generate a multivariable digital representation prediction sequence and a multi-source sensor prediction sequence for a period of time after the current moment;

[0068] Alternatively, as Figure 3 As shown, the multi-step time series prediction module extracts and captures time-dependent features of the multivariate digital representation sequence and the multi-source sensor time series sequence (such as time series data collected by the temperature sensor). The processing process is as follows:

[0069] First, embed the input sequence into words to achieve high-dimensional mapping of sequence features and obtain high-order feature vectors;

[0070] The high-dimensional feature vector is then fed into the Seq2Seq encoder, as Figure 4As shown in the figure, the Seq2Seq encoder is composed of multiple stacked LSTM units. Each LSTM unit is responsible for receiving the high-dimensional feature vector output by the 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 transfer feature information between LSTM units. The sequence representation features are normalized and input into the 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 digital representation sequence and the sensor sequence, enhancing the information interaction capability between sequences, and obtaining high-level semantic features.

[0071] The high-level semantic features are then output to a multi-layer perceptron, and the processed results are fed into a Seq2Seq decoder. The Seq2Seq decoder uses the same unit structure as the Seq2Seq encoder, but operates in the opposite direction, and is used to gradually 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 again, the prediction output of the corresponding 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.

[0073] S4, inputting the multivariable digital representation prediction sequence and the multi-source sensor prediction sequence into a set value real-time prediction module, predicting and outputting the sensor control unit that needs to modify the set value at the current moment, and the specific modification value;

[0074] Alternatively, as Figure 5 As shown, the processing process of the set value real-time prediction module is as follows:

[0075] 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 expressiveness and discriminability of the feature, and then channel splicing is performed to form a high-dimensional feature for guiding the sensor control unit (the sensor control unit consists of a sensor and a corresponding actuator (such as a motor, valve, etc.). The sensor is responsible for the real-time perception of the process state, and the actuator implements the corresponding dynamic adjustment according to the control instructions output by the algorithm). The high-dimensional feature passes through the gated network and outputs the weight parameter W corresponding to each sensor control unit. i , the weight parameter W i The sensor control unit that simulates the visual perception of the on-site operator and guides the decision to modify the set value is used. iThe value 0 or 1 represents the on or off of the corresponding sensor control unit, where on means that the setting value of the corresponding sensor control unit needs to be modified, and off means that the specified 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 setting value is modified through the gating decision unit.

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

[0077] 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 the original feature information is retained through skip connection superposition;

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

[0079] like Figure 7 As shown, the gate decision unit is composed of structural units stacked four times to achieve mapping modeling from sensor prediction sequence to modified set value. The processing process of each structural unit is:

[0080] Taking the prediction sequence of each sensor as input, the original data is projected into a unified semantic space through the word embedding layer. After layer normalization to stabilize the feature distribution, it is input into the self-attention layer to capture the sequence's own dependencies. In this process, skip connections are introduced to preserve the original feature information.

[0081] After that, layer normalization, multi-layer perceptron and skip connection are performed to achieve further feature compression and reconstruction.

[0082] S5. Input the multi-source sensor prediction sequence and the modified value of the sensor control unit setting value at the current moment into the PID parameter real-time prediction module, predict and output the PID parameters that should be adjusted by the sensor control unit whose setting value needs to be modified at the current moment, and realize dynamic optimization of foam flotation control parameters driven by multi-source information.

[0083] 1. Optionally, Figure 8 As shown, the PID parameter real-time prediction module predicts the PID parameters that should be adjusted by the sensor control unit that needs to modify the set value at the current moment by fusing the deviation sequence between the sensor prediction sequence and the modified value of the sensor control unit set value at the current moment, the difference between the previous and next moments, and the accumulated error information. Through deep learning modeling of these information, adaptive optimization and dynamic regulation of the control strategy are achieved, and the stability and responsiveness of the system under complex working conditions are improved. Among them, the PID parameter K P , K I , K DThey correspond to the current error, historical accumulated error, and future predicted error respectively (the embodiment of the present invention predicts the PID parameters commonly used in industrial control, which is not only compatible with the existing control framework but also can achieve a more precise control effect based on the algorithm drive). The specific processing process is as follows:

[0084] Calculate the difference between each predicted value in the sensor prediction sequence and the modified value of the sensor control unit setting value at the current moment to generate the sensor prediction deviation sequence e(t). On this basis, further calculate Δe(t) and ∫e(t)dt (where Δe(t) can be discretely quantized as e(t)-e(t-1) and ∫e(t)dt can be discretely quantized as the superposition of e(t)).

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

[0086] The multi-source information consisting of the modified value of the sensor control unit setting value at the current moment, e(t), Δe(t), ∫e(t)dt, and knowledge covariates is uniformly input into the PID parameter prediction module. By introducing the cross-attention mechanism, effective fusion between features is achieved to enhance the model's perception of key information. The fused features are further subjected to nonlinear mapping and parameter extraction by the fully connected layer to predict and output the PID parameters that the sensor control unit should adjust at the current moment.

[0087] Since the overall solution of the embodiment of the present invention has two prediction goals, one is the prediction of the sensor control unit setting value, and the other is the prediction of the sensor PID parameters. The prediction of the sensor control unit setting value itself has a clear numerical target and a relatively stable mapping relationship. Therefore, the mean square error (MSE) loss function can be directly used for training. The loss function form is as follows:

[0088]

[0089] Among them, y i With f(x i ) represent the label and predicted value of the set value respectively, and m is the number of samples.

[0090] In contrast, the prediction of sensor PID parameters is more complex. These parameters have a significant impact on the stability and response speed of the system. In addition, each parameter has a wide control range and a high degree of mutual coupling. Directly using PID parameters as labels for mean square error calculation faces great challenges in terms of model convergence speed and prediction accuracy. Therefore, an embodiment of the present invention designs a display modeling method to calculate the loss function of the PID parameter prediction.

[0091] Alternatively, as Figure 9 As shown, the predicted loss function of the PID parameters that the sensor control unit that needs to modify the set value at the current moment should adjust is calculated by a display modeling method, specifically:

[0092] By taking the current value of the sensor sequence as the starting value and combining the modified value of the sensor control unit setting value at the current moment predicted by the model as the target end point, the predicted PID parameters are used to generate the system response curve, and then the system response curve is sampled to obtain the output change sequence of the system, and then the output change sequence is compared with the preset self-built sequence label, and the loss function L is calculated by the error between the two sequences. total , to achieve more accurate prediction of PID parameters with more physical meaning and dynamic characteristics;

[0093] The loss function L total The calculation of the loss function is significantly expanded by introducing the system response curve (compared to the direct calculation of the loss function between the predicted PID parameters and the actual PID parameters). P , K I , K D The three parameter values ​​are extended to multiple response curve sampling points, which not only increases the amount of calculation data, making the loss calculation and subsequent back propagation more robust and accurate; it also incorporates the process information of the system dynamic response, achieving higher accuracy in model feature characterization. The loss function L total Fully integrating the control-related indicator parameters, the formula is as follows:

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

[0095] Among them, α1, α2, and α3 are the weight coefficients of the corresponding sub-loss functions, and their specific values ​​are determined by adaptive learning during model training;

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

[0097] L overshoot The over-tuning loss is used to suppress the part of the output change sequence that exceeds the modified value of the set value, thereby improving the stability of the system. set is the modified value of the set value, the formula is:

[0098] L response In order to respond to the speed loss, the system is encouraged to reach the target value quickly to improve the dynamic performance, t reach The time when the modified value of the set value is first reached is:

[0099] The overall training process of the overall model of the embodiment of the present invention, which is composed of a multimodal 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, is as follows:

[0100] At any current moment, the model first models, analyzes, and predicts the multivariable digital representation sequence and multi-source sensor time series for a period of time before the current moment, and generates a multivariable digital representation prediction sequence and a multi-source sensor prediction sequence for a period of time after the current moment. Based on this, the model predicts and outputs the sensor control units that need to modify the set values ​​at the current moment, as well as the specific modification values.

[0101] Afterwards, the model further uses the predicted sensor sequence to predict and output the PID parameters that should be adjusted for the sensor control unit that needs to modify the set value at the current moment.

[0102] Therefore, when the model generates the current set value and PID parameters based on the future prediction sequence, a differentiated strategy is adopted in the loss function design: the loss of the sensor set value adopts the mean square error (MSE) function, and the error is directly fed back to its label; while the loss of the PID parameter is realized through explicit modeling. The current value of the sensor sequence is used as the starting point, combined with the modified value of the set value, the system response curve is simulated and sampled, and the obtained output change sequence is compared with the custom sequence label to calculate the loss L total , to achieve effective prediction of PID parameters with higher accuracy.

[0103] like Figure 10 As shown, an embodiment of the present invention further provides a multi-source information-driven froth flotation control parameter dynamic optimization system, the system comprising:

[0104] The acquisition and construction module 1010 is used to acquire the RGB image and the corresponding 3D point cloud of the flotation work area foam in a period of time before the current moment, and construct them into an RGB image sequence and a 3D point cloud sequence in chronological order;

[0105] A first generating module 1020 is configured to input the RGB image sequence and the 3D point cloud sequence into a multimodal data fusion and digital representation module to generate a multivariate digital representation sequence;

[0106] The second generation module 1030 is configured to input the multivariate digital representation sequence and the multi-source sensor time series sequence involved in the flotation process into a 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;

[0107] The first prediction module 1040 is configured to input the multivariable digital representation prediction sequence and the multi-source sensor prediction sequence into a set value real-time prediction module, and predict and output the sensor control unit that needs to modify the set value at the current moment, as well as the specific modification value;

[0108] The second prediction module 1050 is used to input the multi-source sensor prediction sequence and the modified value of the sensor control unit setting value at the current moment into the PID parameter real-time prediction module, predict and output the PID parameters that should be adjusted by the sensor control unit that needs to modify the setting value at the current moment, and realize dynamic optimization of foam flotation control parameters driven by multi-source information.

[0109] The functional structure of a multi-source information driven froth flotation control parameter dynamic optimization system provided in an embodiment of the present invention corresponds to the multi-source information driven froth flotation control parameter dynamic optimization method provided in an embodiment of the present invention, which will not be described in detail here.

[0110] Figure 11 1 is a schematic structural diagram of an electronic device 1100 provided in an embodiment of the present invention. The electronic device 1100 may have relatively large differences due to different configurations or performances, and may include one or more processors (central processing units, CPU) 1101 and one or more memories 1102, wherein the memories 1102 store at least one instruction, and the at least one instruction is loaded and executed by the processor 1101 to implement the steps of the above-mentioned multi-source information-driven dynamic optimization method for froth flotation control parameters.

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

[0112] Those skilled in the art will understand that all or part of the steps to implement the above embodiments may be accomplished by hardware, or by a program to instruct the relevant hardware, and the program may be stored in a computer-readable storage medium, which may be a read-only memory, a disk, or an optical disk, etc.

[0113] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principles of the present invention should be included in the scope of protection of the present invention.

Claims

1. A multi-source information driven froth flotation control parameter dynamic optimization method, characterized in that: The method comprises: S1, collecting RGB images and corresponding 3D point clouds of the flotation work area foam for a period of time before the current moment, and constructing them into RGB image sequences and 3D point cloud sequences in chronological order; S2. Inputting the RGB image sequence and the 3D point cloud sequence into a multimodal data fusion and digital representation module to generate a multivariate digital representation sequence; S3, inputting the multivariable digital representation sequence and the multi-source sensor time series sequence involved in the flotation process into a multi-step time series prediction module to generate a multivariable digital representation prediction sequence and a multi-source sensor prediction sequence for a period of time after the current moment; S4, inputting the multivariable digital representation prediction sequence and the multi-source sensor prediction sequence into a set value real-time prediction module, predicting and outputting the sensor control unit that needs to modify the set value 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 sensor control unit setting value at the current moment into the PID parameter real-time prediction module, predict and output the PID parameters that should be adjusted by the sensor control unit whose setting value needs to be modified at the current moment, and realize dynamic optimization of froth flotation control parameters driven by multi-source information.

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. The encoder and decoder are both composed of 4 processing units connected in series, which are used to extract and fuse the 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 the deep high-dimensional features of foam in the RGB image through multi-layer convolution. The cross-attention module uses the RGB image features as the query and the 3D point cloud features extracted by three-dimensional convolution at the same time as the key and value to perform interactive fusion between multimodal data. In the first processing unit and the fourth processing unit of the module encoder, jump connections are introduced respectively to integrate the fused high-dimensional features into the subsequent corresponding decoders to achieve the retention of semantic information and promote the fusion of multi-scale information. Finally, the multivariate digital representation sequence is generated through the module decoder.

3. The method according to claim 1, characterized in that The multi-step time series prediction module extracts and captures time-dependent features of the multivariate digital representation sequence and the multi-source sensor time series, respectively. The processing process is as follows: First, embed the input sequence into words to achieve high-dimensional mapping of sequence features and obtain high-order feature vectors; The high-dimensional feature vector is then fed into a Seq2Seq encoder, which is composed of a stack of multiple LSTM units. Each LSTM unit is responsible for receiving the high-dimensional feature vector output by the 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 transfer feature information between LSTM units. The sequence representation features are normalized and input into the 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. The high-level semantic features are then output to a multi-layer perceptron, and the processed results are fed into a Seq2Seq decoder. The Seq2Seq decoder uses the same unit structure as the Seq2Seq encoder, but operates in the opposite direction, and is used to gradually 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 again, the prediction output of the corresponding 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.

4. The method according to claim 1, wherein The processing process of the set value real-time prediction module is as follows: 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 enhancing the expressiveness and discriminability of the feature. Then, channel splicing is performed to form a high-dimensional feature for guiding the sensor control unit. The high-dimensional feature passes through the gated network and outputs the weight parameter W corresponding to each sensor control unit. i , the weight parameter W i The sensor control unit that simulates the visual perception of the on-site operator and guides the decision to modify the set value is used. i The value 0 or 1 represents the on or off of the corresponding sensor control unit, where on means that the setting value of the corresponding sensor control unit needs to be modified, and off means that the specified 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 setting value is modified through the gating decision unit.

5. The method according to claim 4, characterized in that The processing process of the gating network is: 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 the original feature information is retained through skip connection superposition; Then, the feature distribution is stabilized by normalization operation, and the channel dimension is compressed by 1×1 convolution to output the weight parameters W corresponding to each sensor control unit. i ; The gate decision unit is composed of four stacked structural units to achieve mapping modeling from sensor prediction sequence to modified set values. The processing process of each structural unit is as follows: Taking the prediction sequence of each sensor as input, the original data is projected into a unified semantic space through the word embedding layer. After layer normalization to stabilize the feature distribution, it is input into the self-attention layer to capture the sequence's own dependencies. In this process, skip connections are introduced to preserve the original feature information. After that, layer normalization, multi-layer perceptron and skip connection are performed to achieve further feature compression and reconstruction.

6. The method according to claim 1, characterized in that The PID parameter real-time prediction module predicts the PID parameters that should be adjusted by the sensor control unit that needs to modify the set value at the current moment by fusing the deviation sequence between the sensor prediction sequence and the modified value of the sensor control unit set value at the current moment, the difference between the previous and next moments, and the accumulated error information. Through deep learning modeling of these information, adaptive optimization and dynamic regulation of the control strategy are achieved, thereby improving the stability and responsiveness of the system under complex working conditions. Among them, the PID parameter K P , K I , K D They correspond to the current error, historical cumulative error and future prediction error respectively. The specific processing process is: Calculate the difference between each predicted value in the sensor prediction sequence and the modified value of the sensor control unit setting value at the current moment to 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 to obtain additional knowledge covariates through word embedding. The knowledge covariates accurately guide the generation of PID parameters according to the characteristics of each sensor control unit. The multi-source information consisting of the modified value of the sensor control unit setting value at the current moment, e(t), Δe(t), ∫e(t)dt, and knowledge covariates is uniformly input into the PID parameter prediction module. By introducing the cross-attention mechanism, effective fusion between features is achieved to enhance the model's perception of key information. The fused features are further subjected to nonlinear mapping and parameter extraction by the fully connected layer to predict and output the PID parameter K that the sensor control unit should adjust at the current moment. P *、K I *、K D *.

7. The method according to claim 1, characterized in that The predicted loss function of the PID parameters that should be adjusted by the sensor control unit that needs to modify the set value at the current moment is calculated by a display modeling method, specifically: By taking the current value of the sensor sequence as the starting value and combining the modified value of the sensor control unit setting value at the current moment predicted by the model as the target end point, the predicted PID parameters are used to generate the system response curve, and then the system response curve is sampled to obtain the output change sequence of the system, and then the output change sequence is compared with the preset self-built sequence label, and the loss function L is calculated by the error between the two sequences. total , to achieve more accurate prediction of PID parameters with more physical meaning and dynamic characteristics; The loss function L total The calculation of the loss function is significantly expanded by introducing the system response curve, from the original K P , K I , K D The three parameter values ​​are extended to multiple response curve sampling points, which not only increases the amount of calculation data, making the loss calculation and subsequent back propagation more robust and accurate; it also incorporates the process information of the system dynamic response, achieving higher accuracy in model feature characterization. The loss function L total Fully integrating the control-related indicator parameters, the formula is as follows: L total =α1L tracking +α2L overshoot +α3L response Among them, α1, α2, and α3 are the weight coefficients of the corresponding sub-loss functions, and their specific values ​​are determined by adaptive learning during model training; L tracking Tracking error sub-loss, used to reduce the output change sequence and self-built sequence label y ref (t) to improve the accuracy of the system, T is the length of the time series, and the formula is: L overshoot The over-tuning loss is used to suppress the part of the output change sequence that exceeds the modified value of the set value, thereby improving the stability of the system. set is the modified value of the set value, the formula is: L response In order to respond to the speed loss, the system is encouraged to reach the target value quickly to improve the dynamic performance, t reach The time when the modified value of the set value is first reached is:

8. A multi-source information driven froth flotation control parameter dynamic optimization system, characterized in that: The system comprises: The acquisition and construction module is used to collect the RGB image and corresponding 3D point cloud of the flotation work area foam in a period of time before the current moment, and construct them into RGB image sequence and 3D point cloud sequence in chronological order; A first generation module is configured to input the RGB image sequence and the 3D point cloud sequence into a multimodal data fusion and digital representation module to generate a multivariate digital representation sequence; The second generation module is used to input the multivariable 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 multivariable digital representation prediction sequence and a multi-source sensor prediction sequence for a period of time after the current moment; A first prediction module is configured to input the multivariable digital representation prediction sequence and the multi-source sensor prediction sequence into a set value real-time prediction module, and predict and output the sensor control unit that needs to modify the set value at the current moment, as well as the specific modification value; The second prediction module is used to input the multi-source sensor prediction sequence and the modified value of the sensor control unit setting value at the current moment into the PID parameter real-time prediction module, predict and output the PID parameters that should be adjusted by the sensor control unit that needs to modify the setting value at the current moment, and realize dynamic optimization of foam flotation control parameters driven by multi-source information.

9. An electronic device comprising a processor and a memory, wherein the memory stores 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 froth flotation control parameters as described in any one of claims 1-7.

10. A computer-readable storage medium, wherein at least one instruction is stored in the storage medium, 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 froth flotation control parameters as described in any one of claims 1-7.

Images