Mineral detection and sorting method based on deep learning

Abstract

The application discloses a mineral detection and sorting method based on deep learning, and relates to the technical fields of mineral detection and sorting and image data processing. The method obtains multi-modal detection data and background reference data of ore particles, generates a mode-independent drift state vector from the background reference data, and obtains a drift trend code of each mode through a three-branch time sequence coding network. A conditional correction parameter generation network generates a position-by-position scaling factor map and a position-by-position offset factor map corresponding to the spatial position of the intermediate feature map for each correction level of the encoder, and performs position-by-position affine correction on the intermediate feature map of each mode encoder before pixel-level cross-attention fusion between modes, to obtain a mineral segmentation result, a target mineral content estimate value and a sorting execution instruction. When the confidence of the mineral phase boundary pixel degenerates, the main segmentation network is frozen and only the conditional correction parameter generation network is updated. The method can reduce the influence of the differential drift of multi-modal sensors on the segmentation and sorting instructions.

Description

Technical Field

[0001] This invention relates to the field of mineral detection, sorting, and image data processing technology, specifically a mineral detection and sorting method based on deep learning. Background Technology

[0002] Existing mineral detection and sorting production lines typically install visible light imaging, hyperspectral detection, near-infrared reflectance spectroscopy, or X-ray transmission detection components on conveyor belts to collect surface spectral characteristics, near-infrared absorption characteristics, or internal density correlation characteristics of ore particles. The collected results are then input into classification models, segmentation models, or multimodal fusion models to output ore category, mineral regional distribution, or grade estimation results. For fine-grained intergrowths of mixed minerals, pixel-level semantic segmentation can reflect the spatial distribution of target minerals within the particles more precisely than particle-level classification, and is therefore often used for pre-judgment in subsequent sorting actions such as concentrate retention, tailings removal, or intermediate product buffering. While these conventional methods can achieve basic identification when the sensor conditions are close to baseline, they often focus on the fusion of different modal features, neglecting to address the differential drift issues that arise from multiple sensors during long-term continuous operation.

[0003] In actual mineral sorting operations, the drift sources of the visible light hyperspectral detection unit, the near-infrared reflectance spectral detection unit, and the X-ray dual-energy transmission detection unit are not the same.

[0004] Among these, the visible light hyperspectral image is more susceptible to light source aging, dust adhering to the optical window, and changes in band response; the near-infrared reflectance spectroscopy detection unit is more prone to background baseline tilt, peak shift, and changes in detection response; while the X-ray dual-energy transmission detection unit may experience transmission intensity shifts due to changes in X-ray source response, detector sensitivity, and the ratio of high- and low-energy channels. If only the input image or input spectrum is normalized as a whole, or if only uniform affine correction is performed on the granular feature space, then the drift errors of different modes will still enter the pixel-by-pixel feature layer of the deep learning network encoder, causing feature mismatches for multiple modes corresponding to the same spatial location.

[0005] Furthermore, the overall normalization of the image, spectral, or transmission data in the input layer only changes the amplitude or baseline state of the original detection data, making it difficult to handle spatial location-level feature mismatch caused by differences in receptive fields, downsampling scales, and modal branch responses during deep encoding; the category probability correction, content threshold correction, or sorting instruction correction in the output layer occurs after cross-modal fusion, and cannot prevent intermediate features with drift errors from participating in the cross-attention weight calculation.

[0006] Therefore, the current technical problem to be solved is: During long-term continuous operation of mineral detection and sorting production lines, how can we address the differential progressive drift caused by multimodal sensors such as visible light hyperspectral, near-infrared reflectance spectroscopy, and X-ray dual-energy transmission, instead of simply normalizing the input layer or correcting the output layer afterward? Instead, we can apply drift compensation to each modal branch according to its spatial location in the intermediate feature layer of the multimodal semantic segmentation network encoder. This ensures that the input features received by the pixel-level cross-attention layer between modalities are already in a compensated state before the association weights are formed, thereby reducing the flipping of mineral phase boundary pixel categories, the shift in the estimated content of target minerals, and the fluctuation of sorting execution instructions. Summary of the Invention

[0007] (a) Technical problems to be solved To address the shortcomings of existing technologies, this invention provides a deep learning-based mineral detection and sorting method. In each detection cycle, it simultaneously acquires multimodal detection data of ore particles and background reference data for each modality. Modal-independent drift state vectors are generated from the background reference data for each modality. The drift state vectors of the current cycle and historical cycles are input into a three-branch temporal coding network. A conditional correction parameter generation network generates a position-wise scaling factor map and a position-wise offset factor map that match the spatial position of the intermediate feature map of the corresponding modal encoder. Before pixel-level cross-attention fusion between modalities, position-wise affine normalization correction is performed on the intermediate feature maps of each modal branch of the encoder. Subsequently, a mineral category probability map, a mineral spatial distribution segmentation mask map, an estimated value of the target mineral content, and sorting execution instructions are output. When the pixel confidence in the mineral phase boundary region degrades, the main parameters of the multimodal semantic segmentation network are frozen, and online updates are performed only on the conditional correction parameter generation network. Thus, this solves the technical problems described in the background art.

[0008] (II) Technical Solution To achieve the above objectives, the present invention is implemented through the following technical solution: a mineral detection and sorting method based on deep learning, executed by a mineral detection and sorting controller, acquires multimodal detection data of ore particles and background reference data for each modality; preprocesses the multimodal detection data and generates modality-independent drift state vectors from the background reference data; inputs the drift state vectors of the current and historical periods into a three-branch temporal coding network to obtain drift trend codes for each modality; The conditional correction parameter generation network encodes each modality's drift trend and generates a position-wise scaling factor map and a position-wise offset factor map corresponding to the spatial location of the intermediate feature map of the multimodal semantic segmentation network encoder for at least one encoding layer to be corrected. Before the intermodal pixel-level cross-attention layer, the scaling factor map and offset factor map are used in each modal encoder branch to perform position-wise affine normalization correction on the encoder's intermediate feature map according to its spatial location. The corrected feature map is then input into the intermodal pixel-level cross-attention layer to obtain the mineral segmentation result, the target mineral content estimate, and the sorting execution instruction. When the confidence of mineral phase boundary pixels degrades, the main parameters of the multimodal semantic segmentation network are frozen, and only the conditional correction parameter generation network is updated online; otherwise, the current correction parameters are used.

[0009] Furthermore, the visible light hyperspectral image data is spectrally normalized according to spatial pixel position to obtain a pixel-by-pixel spectral reflectance tensor; baseline correction and spatial interpolation are performed on the near-infrared reflectance spectral data to obtain a pixel-by-pixel near-infrared absorbance tensor. Logarithmic transformation and energy ratio calculation were performed on the X-ray dual-energy transmission image data to obtain pixel-by-pixel equivalent atomic number map and pixel-by-pixel density map.

[0010] Furthermore, generating modally independent drift state vectors from background reference data includes: The visible light hyperspectral mode calculates the mean shift of background reflectance in each band and the change in correlation coefficient between bands; the near-infrared mode calculates the background baseline drift slope and characteristic peak shift; and the X-ray dual-energy transmission mode calculates the shift of transmission intensity ratio in the blank area and the deformation of the energy response curve.

[0011] A mineral detection and sorting system based on deep learning includes a data acquisition interface that acquires multimodal detection data of ore particles and background reference data for each modality, and transmits the data to the mineral detection and sorting controller. The mineral detection and sorting controller includes a data preprocessing module, a drift calculation module, a drift trend encoding module, a condition correction parameter generation module, a multimodal semantic segmentation module, a content estimation module, a sorting instruction determination module, a boundary confidence monitoring module, and an online update module. The data preprocessing module preprocesses the multimodal detection data, the drift calculation module generates modal-independent drift state vectors from the background reference data, and the drift trend encoding module inputs the current and historical period drift state vectors into the three-branch temporal coding network to obtain the drift trend encoding of each modality. The conditional correction parameter generation module generates position-wise scaling factor maps and position-wise offset factor maps corresponding to the spatial positions of the intermediate feature maps of the encoder based on the drift trend encoding of each modality. Before the intermodal pixel-level cross-attention layer, the multimodal semantic segmentation module performs position-wise affine normalization correction on the intermediate feature maps in each modal branch of the encoder using position-wise scaling factor maps and position-wise offset factor maps, and inputs the corrected feature maps into the intermodal pixel-level cross-attention layer. The content estimation module generates an estimated value of the target mineral content based on the mineral segmentation results, and the sorting instruction determination module generates a sorting execution instruction based on the estimated value of the target mineral content and outputs it through the sorting execution interface. When the confidence of the mineral phase boundary pixel degrades, the boundary confidence monitoring module triggers the online update module to update the condition correction parameter generation module. The online update module freezes the main parameters of the multimodal semantic segmentation module and only updates the condition correction parameter generation module; otherwise, the current correction parameters are used.

[0012] An electronic device includes a processor, a memory, a communication interface, and a bus, wherein the processor, the memory, and the communication interface are connected via the bus. The memory stores a computer program. When the computer program is executed by the processor, the processor acquires multimodal detection data of ore particles and background reference data of each modality through the communication interface. The multimodal detection data is preprocessed, and modal-independent drift state vectors are generated from the background reference data. The drift state vectors of the current and historical periods are input into a three-branch temporal coding network to obtain the drift trend codes for each modality. The conditional correction parameter generation network generates a position-by-position scaling factor map and a position-by-position offset factor map based on the drift trend codes. In at least one coding layer to be corrected in a multimodal semantic segmentation network, a position-wise scaling factor map and a position-wise offset factor map corresponding to the spatial position of the encoder intermediate feature map are generated. Before the intermodal pixel-level cross-attention layer, the position-wise scaling factor map and the position-wise offset factor map are used to perform position-wise affine normalization correction on the encoder intermediate feature map of each modal branch. The corrected feature map is then input into the intermodal pixel-level cross-attention layer to obtain the mineral segmentation result, the target mineral content estimate, and the sorting execution instruction. When the confidence of mineral phase boundary pixels degrades, the main parameters of the multimodal semantic segmentation network are frozen, and only the conditional correction parameter generation network is updated online; otherwise, the current correction parameters are used.

[0013] A computer-readable storage medium storing a computer program, which, when executed by a processor, performs the following steps: Acquire multimodal detection data of ore particles and background reference data for each mode; preprocess the multimodal detection data and generate modal-independent drift state vectors from the background reference data; input the drift state vectors of the current and historical periods into a three-branch temporal coding network to obtain the drift trend codes for each mode; The conditional correction parameter generation network generates a position-wise scaling factor map and a position-wise offset factor map corresponding to the spatial position of the encoder intermediate feature map for at least one coding level to be corrected in the multimodal semantic segmentation network based on drift trend encoding. Before the intermodal pixel-level cross-attention layer, the position-wise scaling factor map and the position-wise offset factor map are used to perform position-wise affine normalization correction on the encoder intermediate feature map of each modality branch. The corrected feature map is then input into the intermodal pixel-level cross-attention layer to obtain the mineral segmentation result, the target mineral content estimate, and the sorting execution instruction. When the confidence of mineral phase boundary pixels degrades, the main parameters of the multimodal semantic segmentation network are frozen, and only the conditional correction parameter generation network is updated online; otherwise, the current correction parameters are used.

[0014] (III) Beneficial Effects This invention provides a mineral detection and sorting method based on deep learning, which has the following beneficial effects: By simultaneously acquiring multimodal particle detection data and background reference data for each modality, the mineral identification input is separated from the sensor state reference input. Each modal background reference data generates a modally independent drift state vector, which avoids misinterpreting mineral composition changes as sensor drift and reduces misjudgments at the compensation source. By inputting the modally independent drift state vectors of the current and historical periods into a three-branch temporal coding network, the progressive drift trends of different sensing modalities are tracked separately. This result drives the conditional correction parameter generation network to produce position-by-position scaling factor maps and position-by-position offset factor maps corresponding to the spatial positions of the encoder's intermediate feature maps, reducing the impact of instantaneous disturbances on the correction parameters.

[0015] This invention does not merely normalize the input layer detection data as a whole, nor does it retrospectively correct the output layer classification probability, content threshold, or sorting instructions. Instead, it performs scaling and offset correction on the feature responses corresponding to spatial locations at the intermediate feature layers of each modal encoder branch. Because the correction occurs before the inter-modal pixel-level cross-attention layer, the drifted features do not directly participate in the calculation of cross-modal association weights, thereby reducing the probability that the cross-attention layer misinterprets sensor response drift as differences in mineral categories.

[0016] By generating a mineral spatial distribution segmentation mask map through a pixel-by-pixel mineral category probability map, and combining the pixel area ratio of the target mineral category with the pixel-by-pixel density map to generate an estimated value of the target mineral content, the pixel-level recognition results are transformed into granular-level sorting criteria, reducing the transmission of segmentation bias to sorting instructions.

[0017] The average confidence of boundary pixels is used to trigger conditional correction parameters to generate online network updates. During online updates, the main parameters of the multimodal semantic segmentation network are frozen, and pixels in high-confidence core regions are selected as update samples. This allows the correction branches to adapt to sensor drift without disrupting the mineral category discrimination boundaries of the main segmentation network. Simultaneously, when background reference data is missing, modal drift exceeds limits, spatial registration fails, or execution feedback is abnormal, updates are paused, stable correction parameters are reused, intermediate products are buffered, and conservative sorting is implemented. This creates a fallback loop between model correction, feedback updates, and sorting execution, reducing the risk of misclassification caused by uncorrectable drift. Attached Figure Description

[0018] Figure 1 This is a schematic diagram of the overall architecture of a mineral detection and sorting system; Figure 2 This is the main flowchart of the mineral detection and sorting method; Figure 3 A schematic diagram illustrating the synchronous acquisition of multimodal detection data and background reference data; Figure 4 A schematic diagram of modal independent drift state vector generation and three-branch timing coding; Figure 5 A schematic diagram of the adaptive feature correction parameter map of the network output space for generating conditional correction parameters; Figure 6 A schematic diagram of position-wise affine normalization correction of encoder feature layer and pixel-level cross-attention fusion between modalities; Figure 7 A schematic diagram illustrating mineral spatial distribution segmentation, target mineral content estimation, sorting instruction generation, and execution feedback. Figure 8 Flowchart for online update of boundary confidence monitoring and condition correction parameters generation network; Figure 9 This is a schematic diagram illustrating the global anomaly degradation state transition. Figure 10 A schematic diagram illustrating the deployment of virtual devices, electronic devices, and computer-readable storage media. Detailed Implementation

[0019] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0020] Please see Figures 1-10This invention provides a mineral detection and sorting method based on deep learning.

[0021] Please refer to Figure 1 As shown, the mineral detection and sorting system 10 of this embodiment is installed in the mineral detection and sorting production line. The conveyor belt 11 carries ore particles 12 through the detection area 13. The detection area 13 is equipped with a visible light hyperspectral detection unit 20, a near-infrared reflectance spectroscopy detection unit 30, and an X-ray dual-energy transmission detection unit 40. The three units are connected to the mineral detection and sorting controller 60 through a data acquisition interface 50.

[0022] The mineral detection and sorting controller 60 can be an edge computing unit, a local industrial control computer, or a control system composed of an edge terminal and a server terminal, located at the sorting site. Internally, it includes a data acquisition and control module 61, a data preprocessing module 62, a drift state calculation module 63, a drift trend encoding module 64, a condition correction parameter generation module 65, a multimodal semantic segmentation module 66, a content estimation module 67, a sorting instruction determination module 68, a boundary confidence monitoring module 69, an online update module 610, and an anomaly degradation module 611. The sorting instruction determination module 68 outputs execution instructions corresponding to concentrate retention, tailings removal, or intermediate product buffering to the sorting execution unit 71 through the sorting execution interface 70. The execution results of the sorting execution unit 71 are transmitted back to the mineral detection and sorting controller 60 via the execution feedback interface 72. This structure forms a closed loop between on-site data acquisition, deep learning inference, drift compensation, sorting execution, and feedback correction.

[0023] Figure 1 The short labels such as "high-spectrum," "near-infrared," "transmission unit," "acquisition interface," and "controller" represent simplified displays of the visible light hyperspectral detection unit 20, the near-infrared reflectance spectroscopy detection unit 30, the X-ray dual-energy transmission detection unit 40, the data acquisition interface 50, and the mineral detection and sorting controller 60, respectively. The following text uses the full terminology for description.

[0024] In this embodiment, the mineral detection and sorting system 10 refers to the overall system used to complete multimodal detection of ore particles 12, deep learning semantic segmentation, target mineral content estimation, and sorting execution control. This system includes both field sensing components and sorting execution components, as well as software modules and model parameters deployed in the mineral detection and sorting controller 60. The mineral detection and sorting system 10 is not necessarily limited to a specific mechanical rejection mechanism; its core lies in the drift compensation processing closed loop formed by multimodal sensing data and background reference data on the controller side.

[0025] In this embodiment, the detection cycle refers to the processing cycle in which the mineral detection and sorting controller 60 completes one multimodal detection data reception, data preprocessing, model inference, content estimation, and sorting command output after the ore particle 12 enters the detection area 13. The detection cycle can be determined by the position signal of the conveyor belt 11, the trigger signal of the particle entering the detection area 13, or the image frame arrival signal. The engineering significance of the detection cycle is to ensure that the visible light hyperspectral data, near-infrared reflectance spectral data, and X-ray dual-energy transmission data of the same ore particle 12 can be classified into the same processing batch, thereby avoiding the incorrect fusion of data from different particles.

[0026] In this embodiment, the monitoring cycle refers to the period during which the boundary confidence monitoring module 69 statistically analyzes the pixel confidence of the mineral phase boundary region and determines whether an online update is needed. The monitoring cycle can be equal to one detection cycle or can cover multiple consecutive detection cycles. When the dust on site changes rapidly, the light source attenuation changes significantly, or the boundary pixel category flipping occurs frequently, the monitoring cycle can be set shorter than the conventional calibration interval; when the production line is in a relatively stable operating condition, the monitoring cycle can cover multiple detection cycles, thereby reducing unnecessary updates.

[0027] In this embodiment, the background reference data for each modality refers to the background reference signals collected by the visible light hyperspectral detection unit 20, the near-infrared reflectance spectral detection unit 30, and the X-ray dual-energy transmission detection unit 40 under conditions without ore particle obstruction. The background reference data for each modality is used to characterize the current response state of the corresponding sensor, and not as a mineral category label. The background reference data can be obtained from… Figure 1 The blank background area 14 in the middle; the blank background area 14 can be a particle-free interval area between adjacent ore particles 12 on the conveyor belt 11, or a fixed reference plate that can be switched to the detection field of view, or a periodic empty sampling area. Figure 3 The blank background area 14 in the text represents the area or equivalent reference object used to collect background reference data for each mode when there are no mineral particles obscuring the background.

[0028] In this embodiment, the modal-independent drift state vector refers to the set of drift indices calculated by the drift state calculation module 63 based on the corresponding modal background reference data. The drift state vector for the visible hyperspectral mode may include the shift in the mean background reflectance of each band and the change in the correlation coefficient between bands; the drift state vector for the near-infrared mode may include the background baseline drift slope and the characteristic peak shift; the drift state vector for the X-ray dual-energy transmission mode may include the shift in the ratio of transmission intensity in blank areas and the deformation of the energy response curve. These indices are used to describe changes in the sensor's response state, rather than directly describing the mineral type of the ore particles 12.

[0029] In this embodiment, the three-branch temporal coding network refers to the temporal modeling structure set in the drift trend coding module 64, which includes three coding branches corresponding to the visible hyperspectral mode, near-infrared mode, and X-ray dual-energy transmission mode, respectively. Each branch receives the drift state vector sequence of the current period and historical periods of the corresponding mode and outputs the drift trend code of the corresponding mode. In the main embodiment, the three-branch temporal coding network can be a three-branch gated recurrent unit network; without changing the independent coding mechanism of each mode, a three-branch long short-term memory network, a three-branch temporal convolutional network, or a lightweight temporal encoder can also be used.

[0030] In this embodiment, the conditional correction parameter generation network refers to the network structure set in the conditional correction parameter generation module 65. Its input is the drift trend encoding of each modality, and its output is a position-by-position scaling factor map and a position-by-position offset factor map. Here, the conditional correction parameters are not fixed constants, but are generated driven by the current sensor drift trend conditions. This network can be implemented using a combination of a multilayer perceptron and an upsampling structure, a lightweight convolutional decoder, or a conditional normalization parameter generation network, as long as its output can establish a mapping relationship with the spatial position of the intermediate feature map of the corresponding modality encoder.

[0031] In this embodiment, position-by-position affine normalization correction is performed in the encoder of the multimodal semantic segmentation module 66, where the intermediate feature map of each modality branch is subjected to amplitude scaling and reference offset correction according to its spatial position. The correction relationship of a single spatial position of a modality at a certain coding level can be represented by the following formula: F′(m,l,x,y,c)=G(m,l,x,y,c)×F(m,l,x,y,c)+B(m,l,x,y,c) Where F(m,l,x,y,c) represents the pre-correction feature response of mode m at spatial position (x,y) and channel c in encoder level l; F′(m,l,x,y,c) represents the post-correction feature response at the same position and channel; G(m,l,x,y,c) represents the scaling factor of the position-by-position scaling factor map at that position and channel; and B(m,l,x,y,c) represents the offset factor of the position-by-position offset factor map at that position and channel.

[0032] If channel sharing is used in the implementation, G and B can share multiple channels in the same spatial location; if an independent approach is used, corresponding parameters can be generated for each channel separately.

[0033] The above statement means that drift compensation should occur in the intermediate feature layers of the encoder, rather than just on the input image, input spectrum, final classification probability, or sorting threshold. Figure 5The feature dimensions shown refer to the spatial size of the intermediate feature map of the modal encoder in the height and width directions. The positional scaling factor map and the positional offset factor map are matched with this spatial size in a way that includes direct output, upsampling, downsampling, or interpolation mapping, so that the feature response at each spatial location is drift compensated before entering the intermodal pixel-level cross-attention layer.

[0034] Compared to input layer correction, position-by-position correction of the encoder's intermediate feature maps can be applied to the semantic feature responses after convolution, downsampling, or feature extraction, and can compensate for spatial position-level feature biases that are difficult to eliminate by normalization of the original amplitude alone. Compared to output layer correction, its processing occurs before cross-modal fusion, which can prevent the cross-attention layer from forming incorrect association weights based on uncompensated features.

[0035] In this embodiment, the intermodal pixel-level cross-attention layer refers to the fusion layer in the multimodal semantic segmentation module 66. This layer receives the feature maps of each modality after position-by-position affine normalization correction and calculates the association weights for multimodal features located in the same spatial position or corresponding neighborhood. The main function of this layer is to fuse the complementary information of visible hyperspectral mode, near-infrared mode, and X-ray dual-energy transmission mode in the same mineral phase region. In this embodiment, it needs to be set after position-by-position feature correction to avoid uncompensated sensor drift directly entering the cross-modal weight calculation.

[0036] In this embodiment, the pixels in the mineral phase boundary region refer to the pixels adjacent to different mineral categories or their neighboring pixels in the mineral spatial distribution segmentation mask. The mineral phase boundary region is usually most sensitive to sensor drift because the spectral or transmission differences between adjacent minerals in this region are small, and even a small modal shift can cause the maximum category probability to flip.

[0037] In this embodiment, the average confidence level of boundary pixels refers to the average result obtained by the boundary confidence monitoring module 69 through statistical analysis of the maximum category probability values ​​of pixels in the mineral phase boundary region. This confidence level is not an abstract indicator used to evaluate the overall quality of the model, but rather serves as an engineering judgment metric to trigger online updates. Its baseline value can be determined by the calibration period, the historical baseline running period, or a baseline sample after manual review.

[0038] like Figure 1 As shown, the static deployment of the mineral detection and sorting system 10 includes a conveyor belt 11, ore particles 12, a detection area 13, a blank background area 14, a visible light hyperspectral detection unit 20, a near-infrared reflectance spectral detection unit 30, an X-ray dual-energy transmission detection unit 40, a data acquisition interface 50, a mineral detection and sorting controller 60, a sorting execution interface 70, a sorting execution unit 71, and an execution feedback interface 72.

[0039] The conveyor belt 11 sequentially feeds ore particles 12 into the detection area 13. Visible light hyperspectral detection unit 20, near-infrared reflectance spectroscopy detection unit 30, and X-ray dual-energy transmission detection unit 40 are arranged facing the detection area 13, and the collected data is pushed to the mineral detection and sorting controller 60 via the data acquisition interface 50. The sorting execution interface 70 connects the mineral detection and sorting controller 60 and the sorting execution unit 71, allowing the sorting execution commands output by the controller to be converted into actions such as blowing, baffle switching, mechanical tossing, or flow channel switching. The execution feedback interface 72 is used to return whether the execution action has been completed. If no execution feedback is returned, the abnormal degradation module 611 blocks the data within this cycle from entering the online update sample pool to avoid model drift caused by misjudging the execution mechanism failure.

[0040] like Figure 1 As shown, the static deployment of the mineral detection and sorting system 10 includes a conveyor belt 11, ore particles 12, detection area 13, blank background area 14, visible light hyperspectral detection unit 20, near-infrared reflectance spectroscopy detection unit 30, X-ray dual-energy transmission detection unit 40, data acquisition interface 50, mineral detection and sorting controller 60, sorting execution interface 70, sorting execution unit 71, and execution feedback interface 72.

[0041] Conveyor belt 11 sequentially feeds ore particles 12 into detection area 13. Visible light hyperspectral detection unit 20, near-infrared reflectance spectroscopy detection unit 30, and X-ray dual-energy transmission detection unit 40 are respectively arranged facing detection area 13, and send the measured data to mineral detection and sorting controller 60 through data acquisition interface 50. Sorting execution interface 70 connects mineral detection and sorting controller 60 and sorting execution unit 71, so that the sorting execution command output by mineral detection and sorting controller is converted into operations such as blowing, baffle switching, mechanical tossing, or flow channel switching. Execution feedback interface 72 is used to return whether the execution action is in place. If the execution feedback is not returned, the abnormal degradation module 611 blocks the data of that period from entering the online update sample pool, avoiding misjudgment of execution mechanism failure as model drift.

[0042] like Figure 1As shown, the visible light hyperspectral detection unit 20 can specifically be a linear scan hyperspectral imaging component or an area array hyperspectral imaging component, and its underlying components include a light source, a spectral imaging sensor, an acquisition drive circuit, and a data output interface; the near-infrared reflectance spectral detection unit 30 can specifically be a near-infrared linear scan spectral component or a short-wave infrared imaging component; the X-ray dual-energy transmission detection unit 40 can specifically include high and low energy X-ray channels, a transmission detector, and a data output interface within a shielding housing. All of the above components are electrically or communicatively connected to the mineral detection and sorting controller 60 via the data acquisition interface 50. The technical advantage of this connection is that it allows the three types of detection data corresponding to the same ore particle 12 to be uniformly numbered and aligned on the controller side, reducing interference from mismatched data of different particles in subsequent pixel-level fusion.

[0043] In one embodiment, the visible light hyperspectral detection unit 20 and the near-infrared reflectance spectral detection unit 30 can also be combined as a broadband imaging unit covering the visible to near-infrared bands. However, the mineral detection and sorting controller 60 still needs to split its output into visible light hyperspectral mode and near-infrared mode, and their corresponding feature inputs to ensure the independence of subsequent drift state branches.

[0044] like Figure 1 As shown, the data acquisition and control module 61 inside the mineral detection and sorting controller 60 can be a data access logic encapsulated by an industrial control program, or it can be an acquisition thread running on an edge computing unit. This module receives three types of sensor data from the data acquisition interface 50 and establishes a detection cycle according to the trigger signal of the conveyor belt 11 or the arrival signal of the image frame in the detection area 13.

[0045] For example, when ore particles 12 enter the detection area 13, the data acquisition and control module 61 will receive the same detection cycle, which includes visible light hyperspectral image frames, near-infrared reflectance spectral frames, and X-ray dual-energy transmission frames; and each of these will be assigned the same particle index. This index is sent to the data preprocessing module 62 along with the data, ensuring that subsequent multimodal input data have a spatial and temporal correspondence. This avoids the mismerging of data from different ore particles 12 due to the continuous movement of the conveyor belt 11, and reduces spatial mismatch during cross-modal fusion from the source.

[0046] like Figure 2 As shown, the method flow of this embodiment begins with step S101. Figure 2 The flowcharts shown as numbers 101-108 in the figure correspond to steps S101-S108 in this embodiment. The difference between the two is only in the simplified way shown in the figure, but the order of steps and the technical meaning are the same.

[0047] In step S101, the data acquisition and control module 61 acquires multimodal detection data of the ore particles 12 within the detection cycle, and simultaneously acquires background reference data for each modality. The background reference data for each modality comes from... Figure 1 The blank background area 14 can be set as a particle-free interval zone on the conveyor belt 11. When a particle-free obstruction zone is detected passing through the detection area 13, the visible light hyperspectral detection unit 20, the near-infrared reflectance spectral detection unit 30, and the X-ray dual-energy transmission detection unit 40 respectively collect data from this area as background reference data. This background reference data is not sent to the mineral category training label system, but is used by the drift state calculation module 63 to calculate the sensor state change. Thus, the mineral composition change of the ore particles 12 themselves is separated from the sensor state change in terms of data source, reducing the possibility of mistaking particle composition differences for sensor drift.

[0048] like Figure 3 As shown, in one field operation, ore particles 12 are arranged on conveyor belt 11 in a continuous but not completely overlapping manner. When an unobstructed area appears between adjacent ore particles 12, the data acquisition and control module 61 identifies this area as a blank background area 14 and controls each detection unit to collect background reference data. If the coverage density of ore particles 12 on site is high and the blank interval is difficult to meet the background sampling requirements, a fixed background reference plate that can be cut into the detection field of view can be set at the edge of the detection area 13, or short-term idle sampling can be performed after batch change, shift change, or sensor preheating. Although the background reference sources of the above embodiments are different, they all convert the current state of the corresponding modal sensor into a reference observation that can be read by the drift state calculation module 63, so that subsequent drift compensation has an independent input basis. Figure 3 The arrows pointing from the visible light hyperspectral detection unit 20, the near-infrared reflectance spectral detection unit 30, and the X-ray dual-energy transmission detection unit 40 to the particle data and background data respectively represent the particle multimodal detection data stream and the background reference data stream for each modality. Both types of data are entered into subsequent processing after the data acquisition and control module 61 establishes the detection cycle or background sampling cycle index.

[0049] like Figure 2As shown, in step S102, the data preprocessing module 62 performs spatial alignment and physical quantization on the particle multimodal detection data. For visible hyperspectral image data, the data preprocessing module 62 performs spectral normalization according to the spatial pixel position, converting the response of each band into reflectance; for near-infrared reflectance spectral data, the data preprocessing module 62 performs baseline correction and spatial interpolation, mapping the near-infrared data to the same spatial position as the visible hyperspectral data; for X-ray dual-energy transmission image data, the data preprocessing module 62 calculates the equivalent atomic number correlation map and density correlation map through logarithmic transformation and energy ratio calculation. In this way, the different physical quantities output by these three types of sensors can be converted into pixel-level inputs, which can enter the same semantic segmentation network, and the subsequent pixel-level cross-attention layers between modalities can also calculate the correlation weights at the same spatial position.

[0050] like Figure 2 As shown, in step S102, if the data preprocessing module 62 detects that the spatial registration errors of multiple modalities are greater than those of the calibration board or the registration error threshold determined by the historical reference registration results, then the detection period is considered an anomaly in spatial registration. Combined with... Figure 9 As shown, the abnormal degradation module 611 responds to the spatial registration abnormality by blocking the periodic data from entering the online update module 610, and causes the sorting instruction determination module 68 to output an intermediate product buffer instruction for the ore particles 12 near the threshold, so as to avoid the feature that causes pixel misalignment from entering the cross attention layer and forming incorrect cross-modal association weights, and to prevent incorrect samples from contaminating the condition correction parameter generation network.

[0051] like Figure 4 As shown, the background reference data of each mode is obtained by using the drift state calculation module 63 in step S103, and an independent drift state vector for each mode in the current period is generated respectively.

[0052] For the visible hyperspectral modes, the drift state calculation module 63 calculates the mean shift of background reflectance in each band and the change in correlation coefficient between bands. The mean shift of background reflectance in each band refers to the light source intensity, optical window contamination, or detector response attenuation. The change in correlation coefficient between bands characterizes whether the response morphology of different bands changes relative to the reference state. For the near-infrared modes, the drift state calculation module 63 analyzes and obtains the background baseline drift slope and characteristic peak shift. The background baseline drift slope characterizes the overall near-infrared background curve tilt, while the characteristic peak shift characterizes the shift in spectral peak position.

[0053] For the X-ray dual-energy transmission mode, the drift state calculation module 63 analyzes and obtains the transmission intensity ratio offset and energy response curve deformation in the blank region. These two data represent the response differences between the high and low channels and the shape changes of the curve, respectively. In further calculations, instead of using a single global drift quantity to describe the combined state of the three sensors, the drift state vectors associated with the physical mechanism of each mode are solved separately.

[0054] like Figure 4 As shown, each component of the modal-independent drift state vector comes from the background reference data of the corresponding mode, rather than the classification result of ore particle 12.

[0055] Specifically, if the background reflectance of the visible hyperspectral detection unit 20 shows a uniform shift across multiple bands, but the background baseline of the near-infrared reflectance spectral detection unit 30 does not exhibit a significant tilt, the drift state calculation module 63 will only generate a strong reflectance shift expression in the visible hyperspectral mode drift state vector, without performing equal-intensity drift compensation for the near-infrared mode. This reduces the risk of one mode drift being erroneously propagated to other modes, thereby reducing overcompensation or undercompensation caused by subsequent uniform correction.

[0056] like Figure 4 As shown, in one embodiment, the drift state vector of the visible hyperspectral mode can also incorporate the first-order difference morphology of the background spectrum; the drift state vector of the near-infrared mode can also incorporate the peak spacing variation between multiple reference peaks; and the drift state vector of the X-ray dual-energy transmission mode can also incorporate the noise variance variation in the blank region. These substitutions do not change the core mechanism of generating independent drift state vectors for each mode using background reference data; their function remains the same: converting sensor state changes into engineering inputs that can be read by the subsequent calibration parameter generation network.

[0057] like Figure 4 As shown, in step S104, the drift trend encoding module 64 reads the historical drift state vectors of the previous several consecutive detection cycles from the historical state storage unit, arranges them and the current cycle drift state vector in chronological order, and inputs them into the three-branch temporal coding network respectively.

[0058] The historical window length is determined by the detection cycle, sensor calibration cycle, production line cycle time, or the rate of change of the average confidence of boundary pixels. It is used to distinguish between short-term noise and long-term drift. If the historical window is too short, the drift trend encoding is easily affected by a single dust blockage or instantaneous noise. If the historical window is too long, the drift trend encoding will lag in its response to sudden changes in sensor state.

[0059] Therefore, the mineral detection and sorting controller 60 can set the historical window length according to the calibration cycle or historical benchmark data, and the abnormal degradation module 611 will suspend online updates when the parameters exceed the limits. The historical state storage unit can be implemented by the internal storage area of ​​the mineral detection and sorting controller 60, or it can be implemented by... Figure 10 The historical state storage area in the memory 92 shown is implemented; it is used to save the historical drift state vector, the correction parameter diagram of the previous stable period, the network version of the condition correction parameter generation network, and the abnormal state record. It is not necessary to use separate labels in the attached figure.

[0060] like Figure 4 As shown, the first branch of the three-branch temporal coding network receives the drift state vector sequence of the visible hyperspectral mode, the second branch receives the drift state vector sequence of the near-infrared mode, and the third branch receives the drift state vector sequence of the X-ray dual-energy transmission mode. In the main implementation, each branch adopts a gated cyclic unit structure to retain historical drift trends and suppress occasional noise in a gated manner. The output of each branch is the drift trend code for the corresponding mode. Figure 4 The three trend encoding output boxes on the right correspond to visible hyperspectral mode drift trend encoding, near-infrared mode drift trend encoding, and X-ray dual-energy transmission mode drift trend encoding, respectively. These three are used as inputs to the conditional correction parameter generation module 65 in subsequent step S105. This ensures that the conditional correction parameter generation module 65 receives drift expressions with temporal trend meaning, rather than instantaneous deviations from a single detection cycle, thereby reducing correction parameter oscillations caused by short-term dust disturbances.

[0061] like Figure 4 As shown, in one implementation, the three-branch temporal coding network can employ a three-branch long short-term memory network to adapt to larger sensor drift spans; it can also employ a three-branch temporal convolutional network to reduce inference latency of edge computing units within the detection cycle; or it can employ a lightweight temporal encoder to extract drift trends using correlations during historical cycles. The common thread among these alternatives is the need to ensure that the drift sequences of the three modalities are independently input, independently encoded, and still retain their corresponding modal identities at the output, avoiding regression to global drift coding.

[0062] like Figure 5As shown, in step S105, the conditional correction parameter generation module 65 calls the drift trend encoding of each modality into the conditional correction parameter generation network to produce a position-wise scaling factor map and a position-wise offset factor map. The conditional correction parameter generation network can consist of a multilayer perceptron and upsampling units, or it can be composed of a lightweight convolutional decoder. The input is the drift trend encoding of the corresponding modality, and the output spatial size matches the spatial size of the intermediate feature map of the encoder for that modality. If the feature map sizes of different levels in the encoder are inconsistent, the conditional correction parameter generation network can output parameter maps of corresponding sizes for each level to be corrected, or it can first output a reference-size parameter map and then map it to the corresponding level through upsampling, downsampling, or interpolation.

[0063] like Figure 5 As shown, the position-by-position scaling factor map is used to adjust the feature response amplitude at each spatial position of the encoder's intermediate feature map, and the position-by-position offset factor map is used to adjust the feature response baseline offset at each spatial position of the encoder's intermediate feature map. Unlike channel weights or granular weights, the position-by-position scaling factor map and the position-by-position offset factor map will form different compensation intensities in the mineral phase boundary region and the mineral core region.

[0064] For example, if the reflectivity of the visible hyperspectral mode decreases due to dust, while the X-ray dual-energy transmission mode remains relatively constant, the conditional correction parameter generation module 65 can perform greater scaling or offset correction on the spatial position near the boundary of the visible hyperspectral encoder branch, without applying the same level of compensation to the X-ray dual-energy transmission encoder branch. This allows the correction action to vary with the mode and spatial position, rather than applying the same compensation to the entire particle.

[0065] like Figure 5 As shown, in one implementation, the conditional correction parameter generation network can also employ a conditional normalization parameter generation structure. That is, the drift trend encoding is mapped to the scaling and offset parameters required by the encoder normalization layer; then, it is extended to the feature map spatial location through the spatial mapping unit. If edge computing resources are limited, the conditional correction parameter generation module 65 can use multiple channels at the same spatial location, sharing the same scaling and offset factors. If higher detail is required for the boundaries, a channel-independent parameter map generation method can be used. These variations all utilize the mechanism of generating spatial location-level correction parameters from drift trend encoding.

[0066] like Figure 6As shown, in step S106, the encoder of the multimodal semantic segmentation module 66 includes a visible light hyperspectral encoder branch, a near-infrared encoder branch, and an X-ray dual-energy transmission encoder branch. The modal input data output by the data preprocessing module 62 is fed into the corresponding encoder branch, and each encoder branch generates an intermediate feature map at at least one coding level. The mineral detection and sorting controller 60 sends the position-by-position scaling factor map and position-by-position offset factor map output by the conditional correction parameter generation module 65 into the corresponding encoder branch to perform position-by-position affine normalization correction on the intermediate feature map. This correction occurs before the intermodal pixel-level cross-attention layer, so only the drift-compensated features participate in the subsequent cross-modal association weight calculation.

[0067] like Figure 6 As shown, the main body for position-by-position affine normalization correction is the encoder software logic in the multimodal semantic segmentation module 66. It receives three types of inputs: the encoder intermediate feature map, the position-by-position scaling factor map, and the position-by-position offset factor map. This logic outputs the corrected feature map and sends it to the pixel-level cross-attention layer between modalities. If the correction parameter map is inconsistent with the spatial size of the current encoding layer, the encoder first calls the size mapping logic to upsample, downsample, or interpolate the correction parameter map so that its spatial position corresponds to the encoder intermediate feature map. This avoids leaving the drift error to be absorbed by the cross-attention layer, reducing the risk that the cross-attention layer will misinterpret the sensor drift as a difference in mineral category.

[0068] like Figure 6 As shown, the intermodal pixel-level cross-attention layer inputs the corrected feature maps of the three modalities into the corresponding spatial locations or neighborhoods, which then calculates the cross-modal association weights, resulting in fused feature maps at different scales. Instead of directly fusing with the uncompensated original encoded features, the association weights are calculated using the drift-corrected features.

[0069] For example, when the visible hyperspectral and near-infrared modes are significantly affected by dust at mineral phase boundaries, while the X-ray dual-energy transmission mode response is closer to the baseline state, the cross-attention layer can also calculate corrected feature results to reduce the probability that the boundary pixel category is dominated by a single mode with drift influence. This microscopic effect reduces the probability of boundary region category flipping being propagated to segmentation mask area deviation.

[0070] like Figure 6As shown, in one implementation, the inter-modal pixel-level cross-attention layer can employ single-head cross-attention, multi-head cross-attention, or a spatial neighborhood-based modal association weight calculation module. Regardless of the structure adopted, its input should be the feature maps of each modality after position-by-position affine normalization correction, and the output of this layer is used by the subsequent decoder to generate pixel-by-pixel mineral category probability maps. If multi-head cross-attention is selected, different attention heads can focus on surface spectral differences, near-infrared absorption differences, and X-ray transmission differences respectively. If a spatial neighborhood weight calculation module is adopted, neighborhood features near the boundary of the same mineral phase can be included in the association weight calculation, thereby reducing the impact of single-pixel noise.

[0071] like Figure 2 and Figure 6 As shown, if the multimodal semantic segmentation module 66 detects a missing modality or an abnormal feature map size during the encoder calibration stage, the anomaly downgrade module 611 can lock the online update action for that cycle and call the calibration parameter map generated in the previous reference cycle for conservative inference. For the missing modality X-ray dual-energy transmission mode, and when pixel-by-pixel density-related information cannot be generated, the content estimation module 67 can downgrade the target mineral content estimation to a relative abundance estimation, so that the sorting instruction determination module 68 will output an intermediate product buffer instruction through particles near the threshold. Thus, the sensor's absence does not directly lead to irreversible missorting of concentrate or tailings.

[0072] like Figure 7 As shown, in step S107, the decoder of the multi-scale fused feature map is upsampled layer by layer by the multi-modal semantic segmentation module 66 to obtain a pixel-by-pixel mineral category probability map. At this time, the content estimation module 67 selects the mineral category corresponding to the highest probability at each pixel position to generate a mineral spatial distribution segmentation mask map; after statistically analyzing the pixel area ratio of the target mineral category, it adds the pixel-by-pixel density-related information obtained from the preprocessing of X-ray dual-energy transmission data to generate the content estimate of the target mineral.

[0073] exist Figure 7 In this process, density information is a pixel-wise density correlation map formed from X-ray dual-energy transmission image data in data preprocessing module 62. This map is used as a weighted input to content estimation module 67. Simultaneously, it is used in conjunction with the mineral spatial distribution segmentation mask to generate a content estimate of the target mineral content. This content estimate is not a simple classification label but a granular physical quantity representation used for subsequent sorting. This transforms the pixel-level segmentation results into inputs that can drive on-site sorting, avoiding a situation where the model output remains at an abstract probability level.

[0074] like Figure 7As shown, the sorting instruction determination module 68 compares the estimated value of the target mineral content with the content sorting threshold. The content sorting threshold can be determined by mineral processing requirements, historical benchmark samples, or manual testing results. When the threshold is within a fluctuating range, it represents the buffer range caused by testing errors, sensor fluctuations, and model output uncertainties. If the current estimated target mineral content is greater than the retention threshold, the sorting instruction determination module 68 outputs a concentrate retention instruction to the sorting execution unit 71 through the sorting execution interface 70. If the current target mineral content is lower than the rejection threshold, a tailings rejection instruction is output. If the current target mineral content falls within the fluctuating range of the threshold, an intermediate product buffer instruction is output.

[0075] Furthermore, particles with uncertainties at the boundary are transferred to the buffer flow direction instead of being forcibly divided, in order to reduce irreversible misclassification of particles near the threshold.

[0076] like Figure 7 As shown, in one embodiment, the sorting execution unit 71 can be a jetting assembly, a baffle switching assembly, a mechanical actuation assembly, or a material flow channel switching assembly. The jetting assembly changes the falling trajectory of the ore particles 12 through controlled airflow; the baffle switching assembly changes the material channel into which the particles enter by switching the baffle position; the mechanical actuation assembly changes the particle position through pushing or actuating actions; and the material flow channel switching assembly changes the diversion path through the state of the flow channel gate. All of the above components are driven by the same logic instructions received by the sorting execution interface 70, and therefore do not change the core of data processing and drift compensation in this embodiment.

[0077] like Figure 7 and combined Figure 1 As shown, after the sorting execution unit 71 completes its action, the execution feedback interface 72 sends an action completion status back to the mineral detection and sorting controller 60. If the execution feedback interface 72 fails to return confirmation within the preset confirmation time, the anomaly degradation module 611 records the instruction confirmation timeout status and blocks the ore particle sample corresponding to the sorting execution instruction from entering the online update sample pool. If the execution feedback is inconsistent with the instruction, for example, the sorting instruction determination module 68 outputs a concentrate retention instruction but the execution feedback shows that the particles have entered the buffer channel, the anomaly degradation module 611 records the status inconsistency and requires the sorting instruction determination module 68 to output an intermediate product buffer instruction for particles near the threshold in the subsequent detection cycles. This prevents the execution mechanism from abnormally contaminating the online update sample of the model.

[0078] like Figure 8As shown, in step S108, the boundary confidence monitoring module 69 receives the pixel-by-pixel mineral category probability map and the mineral spatial distribution segmentation mask map. Based on the neighborhood relationships between adjacent pixels of different mineral categories, it determines the pixels in the mineral phase boundary region and extracts the maximum category probability value of these boundary region pixels, calculating the average confidence of the boundary pixels. The reference period can be the period after sensor calibration or the period corresponding to historical reference samples. The degradation threshold is determined by historical reference samples, calibration samples, or manually reviewed samples; its engineering significance is to distinguish between normal probability fluctuations and boundary segmentation degradation. When the average confidence of the boundary pixels decreases more than the degradation threshold relative to the reference period, the boundary confidence monitoring module 69 sends an update trigger signal to the online update module 610.

[0079] like Figure 8 As shown, after receiving the update trigger signal, the online update module 610 does not use the pixels in the boundary region as pseudo-label samples, but instead selects the pixels in the confidence core region of the mineral spatial distribution segmentation mask map.

[0080] Pixels in a high-confidence core region should simultaneously meet the following conditions: the maximum class probability reaches a high-confidence threshold, a certain pixel distance is maintained from the mineral phase boundary region, and the area of ​​its connected region satisfies the sample validity condition. Specifically, the boundary region is used to trigger updates, while the core region provides relatively reliable online update samples, reducing the risk of incorrect boundary labels directly entering the parameter update process.

[0081] like Figure 8 As shown, the online update module 610 pairs the segmentation results of the high-confidence core region pixels with their corresponding original multimodal features before correction to form online update samples, and performs online updates on the condition correction parameter generation network in the condition correction parameter generation module 65.

[0082] In the main implementation, the online update module 610 freezes the main parameters of the multimodal semantic segmentation module 66 and only updates the parameters of the conditional correction parameter generation module 65. Thus, the online update primarily alters the drift compensation branch, without directly changing the main segmentation network's basic ability to distinguish mineral categories. The technical advantage of this feature is that it reduces the risk of erroneous pseudo-labels causing overall boundary drift in the main segmentation network.

[0083] like Figure 8 As shown, in one implementation, the online update module 610 can be used through small-step updates, parameter rollback caching, or version verification. If, after an online update, the average confidence level of boundary pixels does not recover to the acceptable range determined by the baseline period within a continuous monitoring period, the abnormal degradation module 611 will roll back to the conditional correction parameter generation network of the previous version, combined with... Figure 9As shown, the abnormal degradation module 611 will also cause the sorting instruction determination module 68 to output intermediate product buffer instructions for particles near the threshold, so as to prevent the single erroneous online update from being amplified in subsequent detection cycles.

[0084] In one verification method, four sets of comparisons can be set up: input layer global normalization, output layer confidence or content threshold correction, encoder global affine correction, and position-by-position affine correction of the encoder intermediate feature map in this embodiment. These are compared under visible hyperspectral dust drift, near-infrared baseline drift, X-ray energy response drift, and combined drift conditions, comparing the average crossover ratio (CLOSE) of mineral segmentation, boundary F1 value, average confidence of boundary pixels, target mineral content estimation error, and sorting command volatility. This verification method is used to confirm the contribution of position-by-position feature correction to drift error mitigation without requiring changes to the aforementioned model structure and execution flow.

[0085] like Figure 1 and Figure 2 As shown, and in combination Figure 10 The hardware deployment relationship of the electronic device 90 shown is as follows: In the first parallel embodiment, the mineral detection and sorting controller 60 adopts a local edge computing unit deployment method. The data acquisition control module 61, data preprocessing module 62, drift state calculation module 63, drift trend encoding module 64, condition correction parameter generation module 65, multimodal semantic segmentation module 66, content estimation module 67, sorting instruction determination module 68, boundary confidence monitoring module 69, online update module 610, and anomaly degradation module 611 are all deployed in the local controller at the sorting site. In this implementation, all acquisition, compensation, segmentation, content estimation, instruction output, and acknowledgment processing are completed in a closed loop at the site. The multimodal data within the detection cycle can generate sorting execution instructions without leaving the site, which is suitable for working conditions where the conveyor belt 11 has a high speed and the sorting execution unit 71 requires a low-latency response.

[0086] In this embodiment, the historical state storage unit can be a combination of local non-volatile memory and cache, used to store historical drift state vectors and the correction parameter map of the previous stable cycle. If the local controller finds that the historical state is missing during power-on initialization, the abnormal degradation module 611 first enters the background reference sampling and reference cycle establishment process, completes the reference parameter writing, and then starts the online update. In this local deployment method, the historical state storage unit can also be... Figure 10 The storage area in the memory 92 shown is implemented, and at this time the processor 91 calls the program in the memory 92 to read and write the historical drift state vector, the correction parameter map and the abnormal state record.

[0087] In the second parallel embodiment, the mineral detection and sorting controller 60 adopts a collaborative deployment method between the edge and server sides. This deployment method can be used as... Figure 1The diagram illustrates a distributed implementation of the mineral detection and sorting controller 60. The server-side entity structure is not shown separately in the figure. The field edge terminal retains a data acquisition and control module 61, a data preprocessing module 62, a sorting instruction determination module 68, an anomaly degradation module 611, and a sorting execution interface 70. The server-side deploys a drift trend encoding module 64, a condition correction parameter generation module 65, a multimodal semantic segmentation module 66, a content estimation module 67, and an online update module 610. The edge terminal sends multimodal input data and background reference data indexed according to the detection cycle to the server via a communication channel. The server returns the segmentation results, content estimates, and correction parameter versions. The edge terminal then outputs the sorting execution instruction.

[0088] The core mechanism of this embodiment remains encoder position-by-position feature correction driven by background reference data, rather than ordinary remote classification. The server can handle large models and historical state caches, while the edge still has the ability to handle abnormal degradation. In the event of an interruption, the correction parameter map and sorting threshold of the previous stable version can be called for conservative sorting to prevent the production line from stopping directly due to the server being unreachable.

[0089] like Figure 3 , Figure 4 and Figure 5 As shown, in the third parallel embodiment, the system adopts a dual-modal downsizing detection path. This path can retain only the visible light hyperspectral detection unit 20 and the X-ray dual-energy transmission detection unit 40, or only the near-infrared reflectance spectroscopy detection unit 30 and the X-ray dual-energy transmission detection unit 40. In this case, the drift state calculation module 63 only generates the drift state vector corresponding to the retained mode, and the three-branch temporal coding network can be configured as a dual-branch temporal coding network accordingly. However, its core mechanism is still that each mode independently encodes the drift trend and generates the position-by-position scaling factor map and position-by-position offset factor map of the corresponding modal encoder branch. If the X-ray dual-energy transmission detection unit 40 is omitted and pixel-by-pixel density-related information cannot be formed, the content estimation module 67 can downgrade the target mineral content estimation to a relative content estimation based on the pixel area ratio, and make the sorting instruction determination module 68 expand the threshold fluctuation range, allowing more uncertain particles to enter the intermediate product buffer. Thus, while reducing the number of sensing modes, the drift state modeling and encoder position-by-position correction chain are still maintained, avoiding the degradation of the downsizing scheme into ordinary single-modal classification.

[0090] like Figure 4 and Figure 5As shown, in the fourth parallel embodiment, the drift trend encoding module 64 does not use a gated recurrent unit network, but instead adopts a three-branch temporal convolutional network. Each temporal convolutional branch performs convolution extraction on the drift state vector of the corresponding mode along the dimension of the detection period, and outputs the drift trend code. This structure is suitable for edge computing resources with limited resources and a fixed inference time. If there are long-term temperature rises, dust accumulation, or light source attenuation in the production line, a three-branch long short-term memory network can be considered to retain a longer drift memory. Regardless of which temporal network is selected, the conditional correction parameter generation module 65 receives the drift trend code of the corresponding mode and generates a position-by-position scaling factor map and a position-by-position offset factor map.

[0091] like Figure 9 As shown, if the data acquisition interface 50 does not receive a data frame from any of the detection units, such as the visible light hyperspectral detection unit 20, the near-infrared reflectance spectral detection unit 30, or the X-ray dual-energy transmission detection unit 40, within the detection cycle, the abnormal degradation module 611 can determine that there is missing modal data entering the detection cycle and generate a communication interruption.

[0092] In this case, the triggering conditions could be that the modal data under the detection cycle index has not arrived, the data frame length does not conform to the preset format, the data frame verification fails, or the communication receipt times out. The anomaly degradation module 611 performs three operations for this cycle: First, it prevents the ore particle sample corresponding to the current detection cycle from entering the online update sample pool; second, it selects an existing available modality for conservative inference or uses the previous stable cycle for parameter map correction; finally, it instructs the sorting instruction determination module 68 to output an intermediate product buffer instruction for particles near the threshold. The recovery condition is that in the continuous detection cycle, all three types of modal data frames are recovered well and the spatial registration error is restored to an acceptable range. This avoids missing modal data being mistakenly treated as sensor drift trends and incorrectly updated into subsequent detection cycles.

[0093] like Figure 9 As shown, when the background reference data for each modality is insufficient, obscured by ore particles 12, or the blank background area 14 is obscured by dust clumps, the drift state calculation module 63 cannot generate a reliable modal-independent drift state vector. At this time, the anomaly degradation module 611 stops writing and updating the drift state for the current cycle and calls the correction parameter map of the previous stable cycle from the historical state storage unit. If some background reference data is unavailable for several consecutive monitoring cycles, the anomaly degradation module 611 performs a recalibration prompt and outputs particles near the threshold as intermediate product buffers. This process does not simply involve alarms, but simultaneously uses parameters, blocks updates, and performs sorting degradation to prevent the passive drift vector from driving erroneous spatial correction parameters.

[0094] like Figure 9As shown, if the current modal drift state vector exceeds the correctable range, the anomaly degradation module 611 determines that the sensor drift has exceeded the limit. The correctable drift range can be established by sensor calibration records, historical reference periods, or samples validated by the model, distinguishing between slow drifts that do not require compensation by the conditional calibration parameter generation network and hardware anomalies generated during recalibration.

[0095] If the visible hyperspectral mode experiences excessive drift while other modes do not, the anomaly degradation module 611 can reduce the participation intensity of the visible hyperspectral mode in cross-attention fusion, or stop using the visible hyperspectral background reference data of that cycle to update the conditional correction parameter generation network. If multiple modes exceed the limits simultaneously, the sorting instruction determination module 68 enters a conservative sorting mode. This prevents the model from continuing to perform forced binary sorting within a range far exceeding the compensation capability.

[0096] like Figure 9 As shown, when the sorting execution unit 71 fails to complete the action according to the sorting execution instruction, or when the feedback returned by the execution feedback interface 72 is inconsistent with the sorting execution instruction, the anomaly degradation module 611 determines that there is no response from the actuator or an inconsistent state. Specifically, after the sorting execution interface 70 outputs a tailings removal instruction, if the execution feedback interface 72 does not return an execution completion signal, or if the returned flow channel state is inconsistent with the tailings removal flow channel, the anomaly degradation module 611 records this cycle as an execution anomaly cycle. In this cycle, neither the model output nor the execution feedback enters the online update sample pool. The sorting instruction determination module 68 outputs intermediate product buffer instructions for particles near the subsequent threshold until the execution feedback interface 72 continuously returns a consistent state. The goal of this approach is to distinguish between actuator anomalies and model drift, preventing the misuse of mechanical non-response for updating model parameters.

[0097] like Figure 9 As shown, when the historical drift state vector corresponding to the historical window is insufficient, the degradation threshold is not successfully read, the high confidence threshold is missing, the content sorting threshold is out of bounds, or the drift correction range is missing, the abnormal degradation module 611 determines that the parameter is out of bounds or the parameter is unavailable. In this state, the mineral detection and sorting controller 60 does not perform online updates or generate new conditional correction parameter versions, but instead calls the correction parameter map of the previous stable cycle and outputs a parameter reset or recalibration prompt. For ore particles 12 whose target mineral content estimate is near the sorting threshold, the sorting instruction determination module 68 outputs an intermediate product buffer instruction. This avoids passive thresholds or erroneous parameters from entering the sorting closed loop, reducing irreversible missorting caused by parameter anomalies.

[0098] like Figure 9As shown, when the average confidence of boundary pixels decreases more than the degradation threshold relative to the baseline period, and the online update module 610 fails to select a sufficient number of high-confidence core region pixels, the model is considered to have poor confidence. In this case, the online update module 610 will not update the parameters and will continue to collect samples for subsequent monitoring periods. If high-confidence core region pixels cannot be selected for multiple consecutive monitoring periods, the abnormal degradation module 611 will revert to the previous stable version of the conditional correction parameter generation network and prompt for recalibration, avoiding the use of low-confidence pseudo-labels to update the model and reducing the risk of erroneous label propagation.

[0099] like Figure 9 As shown, when ore particles 12 exhibit stacking, abrupt changes in particle size distribution, abnormal surface moisture, short-term dust obstruction, or sudden changes in illumination, the data preprocessing module 62 may detect an increase in spatial registration error, and the boundary confidence monitoring module 69 may detect a sudden drop in the average confidence of boundary pixels. Here, the anomaly degradation module 611 needs to mark this period as a working condition deviation period and block its entry into the online update sample pool. If the working condition deviation only lasts for one period or a few periods, the system will resume normal drift compensation after recovery; conversely, if the working condition deviation persists, background reference resampling and conservative sorting will be triggered, which can distinguish between instantaneous on-site working condition disturbances and progressive sensor drift to reduce erroneous updates.

[0100] like Figure 9 As shown, the abnormal inputs of the anomaly degradation module 611 include at least communication interruption, data loss, background loss, drift exceeding limits, feedback anomaly, parameter out-of-bounds, insufficient confidence, and power-on anomaly; its output actions include at least blocking updates, using stable parameters, conservative sorting, buffered sorting, parameter rollback, and recalibration prompt. The anomaly degradation module 611 also forms a control connection with the online update module 610, the condition correction parameter generation module 65, and the sorting instruction determination module 68, so that abnormal states can respectively affect update blocking, parameter rollback, and switching sorting strategies.

[0101] like Figure 10As shown, this embodiment also provides a deep learning-based virtual device for mineral detection and sorting. This virtual device can be deployed in the mineral detection and sorting controller 60 and corresponds one-to-one with the aforementioned method steps. The virtual device includes a data acquisition module, a preprocessing module, a drift calculation module, a drift trend encoding module, a conditional correction parameter generation module, a feature correction fusion module, a segmentation content estimation module, a sorting instruction determination module, a boundary confidence monitoring module, an online update module, and an anomaly degradation module. These modules can be encapsulated code logic executed by a processor, processes, threads, function libraries, or model inference services running in an edge computing unit, or algorithm components called by the field control program. Their common feature is that there are explicit input-output relationships between the modules, rather than loosely stacked abstract function names.

[0102] like Figure 10 As shown, the data acquisition module corresponds to Figure 1 The data acquisition control module 61 in the middle can be an image acquisition driver, an industrial communication interface processing program, or a sensor data caching thread at its underlying layer. This module receives data input from the visible light hyperspectral detection unit 20, the near-infrared reflectance spectroscopy detection unit 30, and the X-ray dual-energy transmission detection unit 40 through the data acquisition interface 50, and outputs particle multimodal detection data organized according to the detection cycle index and background reference data for each modality. If this module implements data frame verification, it sends abnormal states to the abnormal degradation module, and does not directly discard data to continue updating the model, thus reducing the need for subsequent drift state calculations based on abnormal data.

[0103] like Figure 10 As shown, the preprocessing module corresponds to Figure 1 The data preprocessing module 62 in the middle. Specifically, its underlying layer can be a spectral normalization function, a near-infrared baseline correction function, a spatial interpolation function, an X-ray logarithmic transform function, and an energy ratio calculation function. That is, after obtaining the multimodal particle detection data, it outputs multimodal input data aligned according to the spatial pixel positions; the drift calculation module corresponds to Figure 1 The drift state calculation module 63 in the middle can have a background statistical function, a spectrum change extraction function and an X-ray transmission response comparison function at its bottom layer, which is used to convert the background reference data into a modally independent drift state vector.

[0104] like Figure 10 As shown, the drift trend encoding module corresponds to... Figure 1The drift trend encoding module 64, whose underlying layer can be a three-branch gated recurrent unit network, a three-branch long short-term memory network, or a three-branch temporal convolutional network, reads the historical drift state vector sequence from the historical state storage unit and outputs the drift trend encoding for each modality. The conditional correction parameter generation module corresponds to the conditional correction parameter generation module 65, whose underlying layer can be a multilayer perceptron with an upsampling structure, a lightweight convolutional decoder, or a conditional normalization parameter generation network, used to output position-wise scaling factor maps and position-wise offset factor maps. The output of this module flows to the feature correction fusion module.

[0105] like Figure 10 As shown, the feature correction and fusion module corresponds to the encoder and inter-modal pixel-level cross-attention layer in the multimodal semantic segmentation module 66. It receives input data and correction parameter maps from different modalities, performs position-by-position affine normalization correction on the intermediate feature map of the encoder before the inter-modal pixel-level cross-attention layer, and sends the corrected multimodal feature map into the cross-attention fusion logic; the segmentation content estimation module corresponds to the decoder and content estimation module 67 in the multimodal semantic segmentation module 66, and is used to provide a pixel-by-pixel mineral category probability map, a mineral spatial distribution segmentation mask map, and an estimated value of the target mineral content; the sorting instruction determination module corresponds to... Figure 1 The sorting instruction determination module 68 is used to output sorting execution instructions based on the content estimate and sorting threshold and send them to the sorting execution interface 70.

[0106] like Figure 10 As shown, the boundary confidence monitoring module corresponds to Figure 1 The boundary confidence monitoring module 69 is used to determine the pixels of the mineral phase boundary region based on the probability map and the segmentation mask, and to calculate the average confidence of the boundary pixels; the online update module corresponds to Figure 1 The online update module 610 is used to filter out high-confidence core region pixels when the degradation triggering condition is met, and simultaneously update the conditional correction parameter generation network. The abnormal degradation module corresponds to... Figure 1 The exception degradation module 611 is used to receive communication interruptions, data loss, drift exceeding limits, parameter out-of-bounds errors, execution feedback exceptions, and continuous confidence degradation states, and output actions such as pausing updates, parameter rollback, or intermediate product buffering or conservative sorting. This allows method steps to be clearly mapped to deployable software modules or code boundaries, supporting system class and device class object protection.

[0107] like Figure 10As shown, this embodiment also provides an electronic device 90. The electronic device 90 includes a processor 91, a memory 92, a communication interface 93, and a bus 94. The processor 91 can be a general-purpose processor, a graphics processor, an edge computing processor, or an acceleration processing unit for neural network inference; the memory 92 can include volatile and non-volatile memory for storing computer programs, three-branch timing coding network parameters, conditional correction parameter generation network parameters, multimodal semantic segmentation network parameters, historical drift state vectors, previous stable cycle correction parameter maps, and abnormal state records. The communication interface 93 is used to connect the data acquisition interface 50, the sorting execution interface 70, and the execution feedback interface 72. The bus 94 is used to transmit instructions and data between the processor 91, the memory 92, and the communication interface 93. Figure 10 The communication port in this embodiment has the same meaning as the communication interface 93 in this embodiment. It receives data from the three types of detection units through the data acquisition interface 50, outputs sorting execution instructions through the sorting execution interface 70, and receives feedback from the sorting execution unit 71 through the execution feedback interface 72.

[0108] like Figure 10 As shown, when the electronic device 90 runs the program corresponding to this embodiment, the processor 91 reads program instructions from the memory 92 and sequentially executes data acquisition, spatial alignment, background drift state generation, three-branch temporal coding, conditional correction parameter generation, encoder position-by-position affine normalization correction, cross-attention fusion, mineral segmentation, content estimation, sorting instruction output, boundary confidence monitoring, online update, and anomaly degradation. When performing online updates, the processor 91 can set the main semantic segmentation network parameters to a read-only state, allowing only the conditional correction parameter generation network parameters to be written into the update results; if the update fails, the processor 91 reads the parameters of the previous stable version from the memory 92 and performs a rollback. The technical effect of this electronic device embodiment is that it connects the algorithm flow of this method with the executable hardware architecture, avoiding the solution remaining at the abstract model level.

[0109] like Figure 10As shown, this embodiment also provides a computer-readable storage medium 95. This computer-readable storage medium 95 can be a Flash memory, a solid-state drive, an EEPROM, an industrial controller local memory, a server storage medium, or other media capable of storing program instructions. The storage medium stores a computer program, which, when executed by the processor 91, implements the aforementioned mineral detection and sorting method. The program includes at least instructions for acquiring and indexing multimodal data, instructions for generating modally independent drift state vectors, instructions for performing three-branch temporal coding, instructions for generating position-by-position scaling factor maps and position-by-position offset factor maps, instructions for performing encoder position-by-position affine normalization correction, instructions for performing cross-attention fusion and mineral segmentation, instructions for generating sorting execution instructions, and instructions for updating the conditional correction parameter generation network when boundary confidence degrades.

[0110] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

Claims

1. A mineral detection and sorting method based on deep learning, executed by a mineral detection and sorting controller, characterized in that: Acquire multimodal detection data of ore particles and background reference data for each mode. The multimodal detection data includes visible light hyperspectral image data, near-infrared reflectance spectral data and X-ray dual-energy transmission image data. Preprocess the multimodal detection data and generate modally independent drift state vectors from the background reference data; The generation of modally independent drift state vectors from background reference data includes: calculating the mean offset of background reflectance in each band and the change in correlation coefficient between bands for the visible hyperspectral mode; calculating the background baseline drift slope and characteristic peak shift for the near-infrared mode; and calculating the offset of the transmission intensity ratio in the blank area and the deformation of the energy response curve for the X-ray dual-energy transmission mode. Each component of the modally independent drift state vector comes from the background reference data of the corresponding mode. Input the drift state vectors of the current and historical periods into a three-branch temporal coding network to obtain the drift trend codes for each mode; The three-branch temporal coding network includes at least one of a three-branch gated recurrent unit network, a three-branch long short-term memory network, a three-branch temporal convolutional network, and a lightweight temporal encoder. The three branches of the three-branch temporal coding network correspond to the visible light hyperspectral mode, the near-infrared mode, and the X-ray dual-energy transmission mode, respectively. Obtaining the drift trend code for each mode includes: arranging the drift state vector of the current period and the historical drift state vectors of the previous multiple consecutive detection periods in chronological order according to the mode to obtain the drift state vector sequence of each mode; inputting the drift state vector sequence of each mode into the branch corresponding to that mode; each branch independently outputs the drift trend code of the corresponding mode; the output retains the corresponding mode identity; and the drift state vector of the current period is written into the historical state storage area. The conditional correction parameter generation network encodes based on the drift trend of each modality, and generates a position-wise scaling factor map and a position-wise offset factor map corresponding to the spatial position of the intermediate feature map of the multimodal semantic segmentation network encoder for at least one coding level to be corrected. Before the intermodal pixel-level cross-attention layer, the intermediate feature map of the encoder is corrected by position-by-position affine normalization according to spatial position using the position-by-position scaling factor map and position-by-position offset factor map in each modal encoder branch. For modality, encoding level, spatial position and channel, the corrected feature response is determined by the scaling factor map, the feature response before correction and the offset factor map. The corrected feature map is then input into the intermodal pixel-level cross-attention layer to obtain the mineral segmentation result, the target mineral content estimate and the sorting execution instruction. When the confidence of mineral phase boundary pixels degrades, the main parameters of the multimodal semantic segmentation network are frozen, and only the conditional correction parameter generation network is updated online; otherwise, the current correction parameters are used.

2. The mineral detection and sorting method according to claim 1, characterized in that: The visible light hyperspectral image data is spectrally normalized according to spatial pixel position to obtain the pixel-by-pixel spectral reflectance tensor; the near-infrared reflectance spectral data is baseline corrected and spatially interpolated to obtain the pixel-by-pixel near-infrared absorbance tensor. Logarithmic transformation and energy ratio calculation were performed on the X-ray dual-energy transmission image data to obtain pixel-by-pixel equivalent atomic number map and pixel-by-pixel density map.

3. The mineral detection and sorting method according to claim 1, characterized in that: The position-by-position scaling factor map and the position-by-position offset factor map correspond one-to-one with the intermediate feature map of the corresponding coding level to be corrected in the height and width directions, and are shared for multiple channels in the same spatial location, or generated separately for each channel; when the spatial size of the correction parameter map is inconsistent with the spatial size of the intermediate feature map of the current coding level to be corrected, it is mapped to the spatial size of the current coding level to be corrected by at least one of upsampling, downsampling and interpolation, and then position-by-position affine normalization correction is performed.

4. The mineral detection and sorting method according to claim 1, characterized in that: The conditional correction parameter generation network employs at least one of the following: multilayer perceptron with upsampling unit, lightweight convolutional decoder, and conditional normalization parameter generation structure; The conditional normalization parameter generation structure encodes and maps the drift trend into the scaling and offset parameters of the encoder normalization layer, and extends it to the corresponding intermediate feature map spatial position of the encoder through the spatial mapping unit.

5. The mineral detection and sorting method according to claim 1, characterized in that: The background reference data for each modality comes from a blank background area, which is at least one of the following: a particle-free interval area on the conveyor belt, a fixed background reference plate, and a periodic unloaded sampling area. When a blank background area is available, establish a background sampling period index; when a blank background area is not available, pause online updates and continue using the current correction parameters.

6. The mineral detection and sorting method according to claim 5, characterized in that: When at least one of the following abnormal states is detected: missing background reference data, excessive modal drift, failed spatial registration, continuously degraded boundary confidence, abnormal execution feedback, and out-of-bounds parameters, the abnormal degradation module blocks online updates, uses the correction parameters from the previous stable cycle, and outputs at least one of the following: intermediate product buffer instruction, conservative sorting instruction, and recalibration prompt; when no abnormal state is detected, the online update judgment and sorting instruction output for the current detection cycle continue to be executed.

7. The mineral detection and sorting method according to claim 1, characterized in that: The mineral detection and sorting controller generates an operation record corresponding to each detection cycle. The operation record includes a particle index, a background sampling cycle index, a drift state vector for each mode, version identifiers for the position-by-position scaling factor map and the position-by-position offset factor map, boundary pixel confidence degradation judgment results, online update trigger status, and sorting execution instructions.

8. The mineral detection and sorting method according to claim 7, characterized in that: The operation log also includes the current version of the conditional correction parameter generation network, the previous stable version, update failure rollback flags, abnormal status records, high-confidence core region pixel indexes, and online update sample screening results; When the confidence of the mineral phase boundary pixel degrades, the online update module does not use the mineral phase boundary region pixels as pseudo-label samples. Instead, it selects core region pixels that maintain a preset pixel distance from the mineral phase boundary region, have a maximum class probability that reaches a high confidence threshold, and whose connected region area meets the sample validity conditions to form online update samples. The condition correction parameter generation network is updated while the main parameters of the multimodal semantic segmentation network are frozen. If the average confidence level of the boundary pixels after the online update does not recover to the corresponding interval of the baseline period, the mineral detection and sorting controller will roll back the current version to the previous stable version; otherwise, it will write the current version to the previous stable version.

9. The mineral detection and sorting method according to claim 1, characterized in that: After comparing the estimated content of the target mineral with the content sorting threshold, ore particles with a content higher than the retention threshold generate a concentrate retention instruction, ore particles with a content lower than the rejection threshold generate a tailings rejection instruction, and ore particles within the fluctuation range of the threshold generate an intermediate product buffer instruction. After the sorting execution unit executes the sorting execution instruction and completes the action, it sends back the action status through the execution feedback interface.

10. The mineral detection and sorting method according to claim 9, characterized in that: When the execution feedback interface fails to return the action status corresponding to the sorting execution instruction within the preset receipt time, or when the action status returned by the execution feedback interface is inconsistent with the sorting execution instruction, the mineral detection and sorting controller records at least one of the instruction receipt timeout flag and status inconsistency flag, blocks the ore particle sample corresponding to the sorting execution instruction from entering the online update sample pool, and outputs intermediate product buffer instructions for subsequent ore particles that are within the threshold fluctuation range. When the action status returned by the execution feedback interface is consistent with the sorting execution instruction, the ore particle sample corresponding to the sorting execution instruction will be retained as an optional update sample.

Images