A method for jet compensation in ore sorting machines based on AI video image dynamic capture

Abstract

This invention discloses a jetting compensation method for ore sorting machines based on AI video image dynamic capture, belonging to the field of intelligent ore sorting technology. The method includes the following steps: a feature extraction step: acquiring real-time video images recording moving target ore, and extracting two independent feature sets for each target ore in the real-time video image, the feature sets including a basic kinematic feature set and a secondary motion risk feature set; a parallel prediction step; a decision fusion step: fusing the baseline predicted landing point and the trajectory compensation vector to generate the final predicted landing point of the target ore; and an instruction generation step. This invention, by establishing a secondary motion compensation model, can accurately predict and quantify trajectory deviations caused by non-ideal factors such as material surface slippage and rotation. By fusing this compensation vector with the prediction results of the basic physical model, the accuracy of the final landing point prediction is significantly improved.

Description

Technical Field

[0001] This invention relates to the field of intelligent ore sorting technology, specifically to a jet compensation method for ore sorting machines based on AI video image dynamic capture. Background Technology

[0002] In modern mineral processing, machine vision-based intelligent sorting machines are widely used to replace manual sorting, improving efficiency and grade. Their core principle involves using a high-speed industrial camera to capture images of falling or conveyor belt ore, identifying target ores that do not meet quality, composition, or size requirements, and then driving a high-pressure air valve to inject airflow at precise times, altering their trajectory and thus achieving separation.

[0003] Existing technologies typically employ a "recognition-tracking-linear ballistic prediction" model. This model performs reasonably well when handling materials with regular shapes, dryness, and simple trajectories. However, its sorting accuracy drops drastically when faced with complex scenarios involving a mixture of "high-moisture muddy ore and large pieces of dry ore." Its fundamental flaw lies in:

[0004] Physical model failure: Fine, muddy ore adheres to the surface of larger ore pieces. When ejected at the end of the conveyor belt, the lubrication of the mud causes unexpected secondary rolling, sliding, and bouncing between the ore and the equipment, and between the ore pieces themselves. This secondary motion completely destroys the ideal parabolic model.

[0005] Prediction deviation exceeds limits: The secondary random motion caused by lubrication causes huge deviations in traditional linear or polynomial fitting prediction models. The difference between the predicted injection landing point and the actual landing point far exceeds the industrial accuracy requirements of millimeters, such as the ±5mm safety window, resulting in missed or incorrect selection, which seriously affects the concentrate recovery rate.

[0006] Therefore, how to effectively predict and compensate for trajectory deviations caused by secondary motion under harsh working conditions of high-moisture mud formation and dry block mixing, and dynamically maintain the prediction deviation of the injection execution point within an acceptable range of industrial precision, is a technical problem that urgently needs to be solved in this field.

[0007] The information disclosed in the background section above is only intended to enhance the understanding of the background of this disclosure, and therefore may include information that does not constitute prior art known to those skilled in the art. Summary of the Invention

[0008] This invention aims to solve the technical problem in existing ore separators where the accuracy of the injection prediction point is insufficient due to the secondary rolling and bouncing of the ore when processing mixtures of highly moist muddy ore and large dry ore pieces. Specifically, this solution strives to dynamically maintain the prediction deviation of the final injection execution point within a safe window of ±5mm in the aforementioned specific scenario.

[0009] The technical solution of this invention is: a method for jet compensation in an ore sorting machine based on AI video image dynamic capture, comprising the following steps:

[0010] S1. Feature extraction step: Acquire real-time video images of moving target ore, and extract two independent feature sets for each target ore in the real-time video image. The feature sets include: a basic kinematic feature set to characterize the basic physical trajectory of the target ore, and a secondary motion risk feature set to quantify the risk of the target ore undergoing secondary non-ideal motion.

[0011] S2. Parallel prediction step: Input the basic kinematic feature set into the preset benchmark ballistic prediction model to calculate the benchmark predicted landing point of the target ore; at the same time, input the secondary motion risk feature set into the preset secondary motion compensation model to calculate the trajectory compensation vector used to correct the benchmark predicted landing point.

[0012] S3. Decision Fusion Step: Fusion of the baseline predicted landing point and the trajectory compensation vector to generate the final predicted landing point of the target ore;

[0013] S4. Instruction generation step: Based on the final predicted landing point, determine the injection execution instruction to change the trajectory of the target ore, and drive the injection device to execute the injection execution instruction.

[0014] In a specific embodiment, in step S1:

[0015] The basic kinematic feature set consists of data describing the position and velocity of the target ore in the two-dimensional plane at the current moment;

[0016] The secondary motion risk feature set is composed of data used to quantify the surface slipperiness, rotation state, and visual size of the target ore.

[0017] In a specific embodiment, the extraction process of the basic kinematic feature set includes:

[0018] S1a: In the real-time video image, the independent target ore is identified by the instance segmentation algorithm;

[0019] S1b: Continuously track the geometric centroid of the target ore, calculate its current position data and instantaneous velocity data, and combine the position data and instantaneous velocity data to form the basic kinematic feature set.

[0020] In a specific embodiment, the extraction process of the secondary motion risk feature set includes:

[0021] S11: Analyze the optical reflection characteristics of the target ore in real-time video images and generate a mud-like gloss value to characterize the degree of slipperiness of its surface.

[0022] S12: Analyze the morphological changes of the target ore between consecutive image frames and calculate the equivalent rotational angular velocity value used to characterize its rotational state;

[0023] S13: The mud-like gloss value, the equivalent rotational angular velocity value, and the ore pixel area obtained from the real-time video image are used together to form the secondary motion risk feature set.

[0024] In a specific embodiment, in step S2:

[0025] The reference ballistic prediction model is a Kalman filter model that makes predictions based on gravity and the initial motion state.

[0026] The secondary motion compensation model is a pre-trained feedforward neural network model used to establish a nonlinear mapping relationship between the secondary motion risk feature set and the trajectory deviation.

[0027] In a specific embodiment, the fusion process in step S3 is as follows:

[0028] The trajectory compensation vector is summed with the baseline predicted landing point to obtain a corrected coordinate point, which is then determined as the final predicted landing point.

[0029] In a specific embodiment, in step S4, the blowing execution command includes:

[0030] The optimal blow valve number is used to specify the physical valve that performs the blow action.

[0031] The blow delay time is used to determine the trigger time of the blow action.

[0032] In a specific embodiment, the process of determining the blowing execution command includes:

[0033] S41: Determine the optimal injection valve number based on the horizontal coordinate components of the final predicted landing point;

[0034] S42: Calculate the flight time of the target ore from its current location to its final predicted landing point, and subtract the system processing time covering steps S1 to S3 from this time to determine the blasting delay time.

[0035] This invention provides an improved method for jet compensation in ore sorting machines based on AI video image dynamic capture, which has the following improvements and advantages compared with the prior art:

[0036] (1) The core advantage of this invention lies in its adaptability to changes in working conditions. By quantifying key environmental and material state characteristics such as "mud-like gloss" and "equivalent rotational angular velocity" and using them as inputs, the algorithm can proactively identify and adapt to different working conditions. Therefore, regardless of whether the processed material is dry, high-moisture, mud-like, or mixed, this invention can maintain consistently high-precision prediction performance. This robustness enhances the stability of the production line, reduces equipment downtime and parameter recalibration caused by changes in raw materials, and expands the application scope of intelligent sorting equipment.

[0037] By quantifying risk characteristics such as "mud-like gloss" and "equivalent rotational angular velocity", the disturbance source data is digitized, and its impact is calculated by a specialized compensation model.

[0038] (2) This invention, by establishing a secondary motion compensation model, can accurately predict and quantify trajectory deviations caused by non-ideal factors such as slippery material surfaces and rotation. By integrating this compensation vector with the prediction results of the basic physical model, the accuracy of the final landing point prediction is significantly improved. The practical effect is to reduce the prediction deviation from the centimeter level under complex working conditions using traditional methods and stabilize it within ±5mm, directly reducing the ore's missed and incorrect selection rates, thereby improving the concentrate recovery rate and final grade.

[0039] (3) The trajectory prediction problem is decoupled into two sub-tasks: "ideal trajectory prediction" and "non-ideal motion compensation". The most suitable models (Kalman filter, neural network) are used to process them efficiently, and then the results are fused. Through accurate compensation for non-ideal motion, even under harsh working conditions, the prediction deviation can be stably controlled within ±5mm, ensuring high sorting accuracy and high material recovery rate, and has a strong adaptability to changes in working conditions. Attached Figure Description

[0040] Figure 1 This is a block diagram illustrating a method for jet compensation in an ore sorting machine based on AI video image dynamic capture, according to the present invention. Detailed Implementation

[0041] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to specific embodiments.

[0042] Example 1:

[0043] Please see Figure 1 This invention provides a technical solution for a jet compensation method for ore sorting machines based on AI video image dynamic capture: a jet compensation method for ore sorting machines based on AI video image dynamic capture, the specific steps of which include:

[0044] S1. Feature extraction step: Acquire real-time video images of moving target ore, and extract two independent feature sets for each target ore in the real-time video image. The feature sets include: a basic kinematic feature set to characterize the basic physical trajectory of the target ore, and a secondary motion risk feature set to quantify the risk of the target ore undergoing secondary non-ideal motion.

[0045] The feature extraction step is the data source for the entire algorithm. Its core function is to extract two sets of independent but complementary features from the real-time video images for each target ore, providing complete input for subsequent parallel prediction.

[0046] The fundamental kinematic feature set is used to characterize the basic physical trajectory of the target ore. This set of features follows classical Newtonian mechanics and is the basis for predicting ideal parabolic motion.

[0047] The secondary motion risk feature set is used to quantify the risk of the target ore undergoing secondary non-ideal motion. This is the key innovation of the present invention, which aims to capture and quantify the non-ideal factors that cause traditional physical models to fail.

[0048] By dividing the features into two categories, this invention fundamentally decomposes the problem into an "ideal physical part" and a "non-ideal risk part," namely, the basic kinematic feature set and the secondary motion risk feature set. This allows for the use of different optimization models for targeted processing, avoiding the limitation of a single model being unable to take into account both motion modes.

[0049] S2. Parallel prediction step: Input the basic kinematic feature set into the preset benchmark ballistic prediction model to calculate the benchmark predicted landing point of the target ore; at the same time, input the secondary motion risk feature set into the preset secondary motion compensation model to calculate the trajectory compensation vector used to correct the benchmark predicted landing point.

[0050] The baseline ballistic prediction model is a Kalman filter model, which is well known to those skilled in the art. In this embodiment, the input is the basic kinematic feature set, and the output is the baseline predicted landing point, which is used to predict the ideal physical trajectory.

[0051] The secondary motion compensation model is a pre-trained feedforward neural network model used to establish a nonlinear mapping relationship between the secondary motion risk feature set and the trajectory deviation. The input of the secondary motion compensation model is the secondary motion risk feature set, namely mud gloss, equivalent rotational angular velocity, and pixel area, and the output is a trajectory compensation vector, which is used to compensate for the deviation caused by non-ideal factors.

[0052] The baseline predicted impact point output by the baseline ballistic prediction model is combined with the trajectory compensation vector output by the secondary motion compensation model to obtain a corrected final predicted impact point that is closer to the actual situation.

[0053] The basic kinematic feature set extracted from S1 is input into the preset baseline ballistic prediction model. This model is specifically designed to handle ideal motion. Its task is to calculate the baseline predicted landing point based on the current position and velocity of the ore, under the ideal condition of only considering gravity.

[0054] Meanwhile, the secondary motion risk feature set extracted from S1 is input into the preset secondary motion compensation model. This model is specifically designed to handle non-ideal motion. Its task is to learn and predict the trajectory deviation caused by these risks based on quantified risk indicators, and finally output a trajectory compensation vector to correct the baseline predicted landing point.

[0055] The above-mentioned design, which uses a secondary motion compensation model and a baseline ballistic prediction model for parallel processing, greatly improves computational efficiency. It achieves model decoupling, allowing the baseline ballistic prediction model to maintain its physical simplicity and accuracy, while utilizing a powerful nonlinear model to fit and predict complex random disturbances that are difficult to describe with physical formulas.

[0056] S3. Decision Fusion Step: Fusion of the baseline predicted landing point and the trajectory compensation vector to generate the final predicted landing point of the target ore;

[0057] The decision fusion step is the integration stage of the computational results, fusing the results of the two parallel computation branches to generate the final decision basis. Specifically, it involves performing vector addition on the baseline predicted impact point output by the baseline ballistic prediction model and the trajectory compensation vector output by the secondary motion compensation model, resulting in a corrected final predicted impact point that more closely approximates reality. This step ensures that the final prediction result simultaneously adheres to ideal physical laws and compensates for complex real-world disturbances. The fusion process is simple and efficient, combining the advantages of both models to generate a prediction result that is far more accurate and robust than a single model.

[0058] S4. Instruction generation step: Based on the final predicted landing point, determine the injection execution instruction to change the trajectory of the target ore, and drive the injection device to execute the injection execution instruction.

[0059] This step serves as the interface between the algorithm and the physical actuators, translating the high-precision prediction results into specific hardware control instructions. Based on the final predicted landing point generated by S3, the system determines two key execution parameters:

[0060] Pulse valve: Select the corresponding physical pulse valve based on the horizontal coordinates of the final predicted landing point.

[0061] Injection time: Calculate the flight time from the current moment to when the ore arrives directly below the valve, and subtract the system processing delay to obtain the precise moment when the injection is triggered.

[0062] The instruction generation step realizes a closed loop from "prediction" to "execution", ensuring that the high-precision calculation results can be transformed into precise strike actions in the physical world without deviation, thus achieving the ultimate goal of sorting.

[0063] In a specific embodiment, in step S1:

[0064] The basic kinematic feature set consists of data describing the position and velocity of the target ore in the two-dimensional plane at the current moment;

[0065] The basic kinematic feature set consists of data describing the position and velocity of the target ore in a two-dimensional plane at the current moment. These two physical quantities are the basic parameters necessary to describe any projectile motion.

[0066] The secondary motion risk feature set is composed of data used to quantify the surface slipperiness, rotation state, and visual size of the target ore.

[0067] The secondary motion risk feature set consists of data used to quantify the surface slipperiness, rotational state, and visual size of the target ore. These three features are designed based on the root causes of secondary motion.

[0068] Surface slipperiness: directly relates to the friction between the ore and the conveyor belt or other ores, and is the main cause of unexpected slippage.

[0069] Rotational state: This is the most direct representation of secondary motion. The way energy and momentum are exchanged in a rotating ore is completely different from that in a translational ore.

[0070] Visual dimensions: To a certain extent, these can be related to the quality and windward area of ​​the ore, and are an important factor affecting its trajectory and air resistance.

[0071] This step concretizes the abstract concept of "risk" into three measurable and quantifiable physical characteristics, providing a solid and effective data foundation for the establishment of the secondary motion compensation model.

[0072] In a specific embodiment, the extraction process of the basic kinematic feature set includes:

[0073] S1a: In the real-time video image, the independent target ore is identified by the instance segmentation algorithm;

[0074] By employing advanced instance segmentation networks such as Mask R-CNN, each frame of image captured by a high-speed camera is processed. This algorithm can not only outline each ore, but also generate its precise pixel-level mask, thereby separating ore that are stuck together or overlapping at the data level.

[0075] S1b: Continuously track the geometric centroid of the target ore, calculate its current position data and instantaneous velocity data, and combine the position data and instantaneous velocity data to form the basic kinematic feature set.

[0076] Specific calculations:

[0077] Location data P i (t): Mask for the i-th ore at time t i (t), whose geometric centroid coordinates (x) i (t),y i (t) is calculated and used as its location data.

[0078] P i (t)=(x i (t),y i (t))

[0079] Instantaneous velocity data V i (t): Calculated by comparing the centroid positions of two consecutive frames (or multiple frames after smoothing). Let the camera frame rate be f, and the frame interval be Δt = 1 / f.

[0080]

[0081] Among them, v ix (t) and v iy (t) represents the horizontal and vertical velocity components, respectively.

[0082] This method provides a robust and accurate kinematic data extraction process. Instance segmentation ensures accurate analysis of individual ore, while centroid tracking and differential calculation are standard and effective methods for obtaining motion states.

[0083] In a specific embodiment, the extraction process of the secondary motion risk feature set includes:

[0084] S11: Analyze the optical reflection characteristics of the target ore in real-time video images and generate a mud-like gloss value to characterize the degree of slipperiness of its surface.

[0085] By analyzing optical features such as pixel grayscale distribution, highlight area proportion, or texture complexity within the ore mask region—for example, a wet, slippery muddy surface typically reflects light more evenly and has softer highlights, while a dry rock surface has sharper textures—these optical features can be mapped to a quantified "muddy gloss" G by training a classification or regression model. m value.

[0086] S12: Analyze the morphological changes of the target ore between consecutive image frames and calculate the equivalent rotational angular velocity value used to characterize its rotational state;

[0087] Detailed explanation: By comparing the mineral mask in consecutive frames... i The equivalent rotational angular velocity ω is calculated by the change in the principal axis direction angle. If the principal axis angle of the mask in frame t-Δt is θ(t-Δt) and in frame t is θ(t), then the equivalent rotational angular velocity ω is... e It can be approximated as:

[0088]

[0089] Where: θ(t) refers to the ore mask in the image frame at the current time t. i The principal axis direction angle is a geometric parameter that describes the stretching direction of an object's shape.

[0090] Δt: refers to the time interval between two consecutive frames;

[0091] ω e Equivalent rotational angular velocity: used to quantify how fast the target ore rotates.

[0092] S13: The mud-like gloss value, the equivalent rotational angular velocity value, and the ore pixel area obtained from the real-time video image are used together to form the secondary motion risk feature set.

[0093] Specific explanation: This step aggregates risk information from multiple dimensions. The final risk feature set vector can be F. risk ={G m ,ω e A i ,|V i (t)|}, where A i It is the pixel area of ​​the mineral, |V i (t)| represents the magnitude of the instantaneous velocity. Velocity is included because slippery ore moving at high speeds may exhibit more intense secondary motion.

[0094] This method transforms visual information into physically meaningful risk indicators, ensuring the accuracy of the compensation model.

[0095] In a specific embodiment, in step S2:

[0096] The reference ballistic prediction model is a Kalman filter model that makes predictions based on gravity and the initial motion state.

[0097] Specific explanation: Kalman filters are used to track dynamic systems with well-defined physical models, whose state vectors can be defined as X = [x, y, v]. x ,v y ] T Its state transition matrix will only include the effect of gravitational acceleration g. This model can effectively filter measurement noise and provide smooth, reliable predictions of ideal trajectories.

[0098] X: Represents the state vector of the Kalman filter at any time t.

[0099] x, y: These two components represent the horizontal and vertical coordinates of the target ore's centroid in a two-dimensional plane.

[0100] v x ,v y These two components represent the horizontal and vertical velocities of the target ore's center of mass in a two-dimensional plane.

[0101] The secondary motion compensation model is a pre-trained feedforward neural network model used to establish a nonlinear mapping relationship between the secondary motion risk feature set and the trajectory deviation.

[0102] Specific details: A multilayer perceptron is used as the compensation model. This network is trained offline on a large amount of experimental data, learning a complex mapping from the input (secondary motion risk feature set) to the output (the deviation vector between the actual landing point and the ideal landing point).

[0103] Formula expression:

[0104] ΔP compensation =SMC-Net(G m ,ω e A i V i (t))

[0105] in:

[0106] ΔP compensation =(Δx) c ,Δy c ) is a two-dimensional trajectory compensation vector.

[0107] SMC-Net(·) represents the feedforward neural network model.

[0108] (G m ,ω e Ai V i (t) represents the risk characteristics of the input;

[0109] G m : Mud-like gloss value;

[0110] ω e : Equivalent rotational angular velocity value.

[0111] A i : Ore pixel area, which is obtained by calculating the number of pixels contained in the ore mask Maski.

[0112] V i (t): Instantaneous velocity magnitude.

[0113] Dimensional validation: The model's input is dimensionless or features with different dimensions (glossiness value, angular velocity, number of pixels, velocity), and the output is a displacement vector with components Δx. c ,Δy c The dimension of is length. Neural networks learn weights to achieve nonlinear regression from multidimensional heterogeneous inputs to outputs with specific physical dimensions.

[0114] Kalman filtering solves the physical prediction problem under linear and Gaussian noise in an optimal way; feedforward neural networks solve the problem of random perturbation compensation for nonlinear and non-physical models with their strong fitting ability.

[0115] In a specific embodiment, the fusion process in step S3 is as follows:

[0116] The trajectory compensation vector is summed with the baseline predicted landing point to obtain a corrected coordinate point, which is then determined as the final predicted landing point.

[0117] Formula expression:

[0118] P final =P base +ΔP compensation

[0119] in:

[0120] P final =(x final ,y final ) is the final predicted landing point coordinate.

[0121] P base =(x base ,y base ) are the coordinates of the predicted landing point.

[0122] ΔP compensation =(Δx) c ,Δyc ) is the trajectory compensation vector.

[0123] In a specific embodiment, in step S4, the blowing execution command includes:

[0124] The optimal blow valve number is used to specify the physical valve that performs the blow action.

[0125] The blow delay time is used to determine the trigger time of the blow action.

[0126] In a specific embodiment, the process of determining the blowing execution command includes:

[0127] S41: Determine the optimal injection valve number based on the horizontal coordinate components of the final predicted landing point;

[0128] Specific explanation: The injection valves are typically arranged linearly along a horizontal direction. Each valve is responsible for a segment of the horizontal area. The final predicted landing point is determined by the horizontal coordinate x. final By comparing the valve's coordinate range, the optimal valve number can be uniquely determined.

[0129] S42: Calculate the flight time of the target ore from its current location to its final predicted landing point, and subtract the system processing time covering steps S1 to S3 from this time to determine the blasting delay time.

[0130] Formula expression:

[0131] Delay t =Δt fly -t proc

[0132] in:

[0133] Delay t It is the pulse delay time, which is how long it will take before the valve is triggered from the current moment.

[0134] Δt fly The required flight time is calculated based on the current motion state of the ore and its final predicted landing point.

[0135] t proc It is the inherent processing delay of the system, that is, the time consumed from image acquisition S1 to the completion of fusion decision S3.

[0136] This method not only identifies the valve number but also calculates the delay time more precisely. Specifically, it subtracts the system processing time t... proc It compensates for the time overhead of the calculation process itself and avoids lag-induced errors caused by calculation delays, making it a necessary link to achieve high-precision real-time control.

[0137] It should be noted that all calculation formulas in this application employ, but are not limited to, regression analysis from machine learning algorithms to deeply analyze the collected parameters and identify their natural trends and interrelationships. Specialized software, such as Python's Scikit-learn library or the R language, is used to automatically generate mathematical models that match the data. Then, cross-validation and other methods are used to objectively evaluate the model performance, and continuous feedback and optimization are combined to ensure that the created formulas truly reflect the inherent laws of the data, thereby guaranteeing their validity and accuracy, and ensuring that the calculation process conforms to the constraints of natural laws, rather than being based on artificially set rules.

[0138] The technical solution of this invention, or the part that contributes to the prior art, can be embodied in the form of a software product. This computer software product can be stored in a computer-readable storage medium, such as a computer floppy disk, read-only memory (ROM), random access memory (RAM), flash memory, hard disk, or optical disk, etc., including several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute the methods of the various embodiments of this invention.

[0139] The logic and / or steps represented in the flowchart or otherwise described herein, for example, can be considered as a sequenced list of executable instructions for implementing logical functions, and can be embodied in any computer-readable medium for use by, or in conjunction with, an instruction execution system, apparatus, or device (such as a computer-based system, a processor-included system, or other system that can fetch and execute instructions from, an instruction execution system, apparatus, or device). For the purposes of this specification, "computer-readable medium" can be any means that can contain, store, communicate, propagate, or transmit programs for use by, or in conjunction with, an instruction execution system, apparatus, or device.

[0140] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention, and all such modifications or substitutions should be covered within the scope of the claims of the present invention.

Claims

1. A method for compensation of pulse jetting in an ore sorting machine based on AI video image dynamic capture, characterized in that: Includes the following steps: S1. Feature extraction step: Acquire real-time video images of moving target ore, and extract two independent feature sets for each target ore in the real-time video image. The feature sets include: a basic kinematic feature set to characterize the basic physical trajectory of the target ore, and a secondary motion risk feature set to quantify the risk of the target ore undergoing secondary non-ideal motion. S2. Parallel prediction step: Input the basic kinematic feature set into the preset benchmark ballistic prediction model to calculate the benchmark predicted landing point of the target ore; at the same time, input the secondary motion risk feature set into the preset secondary motion compensation model to calculate the trajectory compensation vector used to correct the benchmark predicted landing point. S3. Decision Fusion Step: Fusion of the baseline predicted landing point and the trajectory compensation vector to generate the final predicted landing point of the target ore; S4. Instruction generation step: Based on the final predicted landing point, determine the injection execution instruction to change the trajectory of the target ore, and drive the injection device to execute the injection execution instruction. In step S1: The basic kinematic feature set consists of data describing the position and velocity of the target ore in the two-dimensional plane at the current moment; The secondary motion risk feature set is composed of data used to quantify the surface slipperiness, rotation state, and visual size of the target ore. The extraction process of the basic kinematic feature set includes: S1a: In the real-time video image, the independent target ore is identified by the instance segmentation algorithm; S1b: Continuously track the geometric centroid of the target ore, calculate its current position data and instantaneous velocity data, and combine the position data and instantaneous velocity data to form the basic kinematic feature set; The extraction process of the secondary motion risk feature set includes: S11: Analyze the optical reflection characteristics of the target ore in real-time video images and generate a mud-like gloss value to characterize the degree of slipperiness of its surface. S12: Analyze the morphological changes of the target ore between consecutive image frames and calculate the equivalent rotational angular velocity value used to characterize its rotational state; S13: The mud-like gloss value, the equivalent rotational angular velocity value, and the ore pixel area obtained from the real-time video image are used together to form the secondary motion risk feature set.

2. The method for compensation of jet blowing in an ore sorting machine based on AI video image dynamic capture according to claim 1, characterized in that, In step S2: The reference ballistic prediction model is a Kalman filter model that makes predictions based on gravity and the initial motion state. The secondary motion compensation model is a pre-trained feedforward neural network model used to establish a nonlinear mapping relationship between the secondary motion risk feature set and the trajectory deviation.

3. The method for jet compensation in an ore sorting machine based on AI video image dynamic capture according to claim 1, characterized in that, The fusion process in step S3 is as follows: The trajectory compensation vector is summed with the baseline predicted landing point to obtain a corrected coordinate point, which is then determined as the final predicted landing point.

4. The method for compensation of jet blowing in an ore sorting machine based on AI video image dynamic capture according to claim 1, characterized in that, The blowing execution instructions in step S4 include: The optimal blow valve number is used to specify the physical valve that performs the blow action. The blow delay time is used to determine the trigger time of the blow action.

5. The method for compensation of jet blowing in an ore sorting machine based on AI video image dynamic capture according to claim 4, characterized in that, The process of determining the blowing execution command includes: S41: Determine the optimal injection valve number based on the horizontal coordinate components of the final predicted landing point; S42: Calculate the flight time of the target ore from its current location to its final predicted landing point, and subtract the system processing time covering steps S1 to S3 from this time to determine the blasting delay time.

Images