A flaky ore double-mode fusion recognition method and system in a high-dust environment

Abstract

The present application belongs to the technical field of computer vision and data fusion, and relates to a high-dust environment under a sheet ore dual-mode fusion recognition method and system, comprising the following steps: calculating a second-order inertia tensor; outputting a phase-encoding predicted state vector flow through phase modulation; enabling a data reading link to analyze digital instructions containing two-dimensional positions and predicted dip angles; triggering a micro resonant cavity to emit a pulse flow to build a transient low-dust airflow area according to the matrix; realizing space-time coordinate alignment by mapping with a two-dimensional mass density distribution map; weighting and splicing structured features of a two-dimensional optical image data flow to output a category decision signal; and feeding back the category decision signal to a Kalman filter model for self-calibration. The present application solves the problem of alignment deviation of multi-modal sensor data caused by object surface shielding and uncertain motion posture in a high-dust environment, which further affects the fusion recognition accuracy.

Description

Technical Field

[0001] This invention belongs to the technical field of computer vision and data fusion, and relates to a dual-modal fusion recognition method and system for sheet-like ore in a high-dust environment. Background Technology

[0002] In the field of mineral processing and sorting, the identification and classification of bulk materials is a crucial step in improving resource utilization. During the free-fall transport of ores, their motion is highly variable, and the large amount of dust generated can adhere to the ore surface or disperse in the optical path, interfering with and obscuring information from identification methods that rely on optical features. Furthermore, the tumbling and rotation of non-spherical materials such as flaky ores during their descent presents technical challenges for the synchronization and spatial alignment of multi-sensor data.

[0003] A prior art Chinese invention patent application (application number 202210405128.0) discloses an image processing method, apparatus, device, and storage medium. The method acquires multiple images including a target object and determines the environment type in which the target object exists. It then obtains a preset machine learning model corresponding to that environment type and processes the images using this model to obtain target attribute information of the target object. This solution aims to improve the accuracy of target attribute prediction by employing specific models for different environment types.

[0004] However, the aforementioned existing technologies have limitations when facing complex scenarios with high dust levels and high-speed object movement. Even with conventional dust removal devices, single-modal optical recognition systems have room for energy optimization in their wide-area purging methods, and it is difficult to guarantee a clean observation window for high-speed moving objects at the moment of camera exposure. If multimodal information fusion is attempted by combining X-ray detection technology, the lack of a precise prediction mechanism for the complex motion posture of sheet-like ore will lead to spatial misalignment when matching X-ray data and optical data collected at different times and locations. This misalignment of multi-sensor data will reduce the overall performance of subsequent fusion recognition algorithms.

[0005] Therefore, how to overcome the problem of alignment deviation in multimodal sensor data caused by surface occlusion and uncertainty of motion posture in high dust environments, which affects the accuracy of fusion recognition, is a technical problem that needs to be solved by those skilled in the art. Summary of the Invention

[0006] In a first aspect, the present invention provides a method for dual-modal fusion identification of flaky ore in a high-dust environment, comprising the following steps: S1. Call the detector to collect the original X-ray transmission signal stream of the target object, deconvolve to reconstruct a two-dimensional mass density distribution map, and calculate the second-order inertial tensor based on this. S2. Combine the initial conditions of the transport with the second-order inertial tensor to initialize the Kalman filter model to generate a continuous state vector, and then output the phase-coded predicted state vector stream through phase modulation. S3. Use a hardware phase-locked loop circuit to lock the phase-encoded prediction state vector stream, and enable the data reading link to parse out the digital command containing the two-dimensional position and prediction tilt angle. S4. Map the two-dimensional position to the predicted tilt angle to generate an array drive control matrix, and trigger the micro resonant cavity to emit pulse flow based on this matrix to construct a transient low-dust airflow zone; S5. Capture the two-dimensional optical image data stream in the transient low-dust airflow area and assign a unified timestamp, and map it with the two-dimensional mass density distribution map to achieve spatial and temporal coordinate alignment. S6. Extract the structured features of the two-dimensional mass density distribution map after spatial and temporal coordinate alignment to generate a dynamic feature weight matrix. After weighting the two-dimensional optical image data stream, the structured features are spliced ​​together to output the category decision signal. S7. Record the actual arrival time of the target object and compare it with the predicted time to generate a time prediction deviation vector, and feed it back to the Kalman filter model for self-calibration.

[0007] A further aspect of this invention involves calling a detector to acquire the original X-ray transmission signal stream of the target object, deconvolving it to reconstruct a two-dimensional mass density distribution map, and calculating the second-order inertial tensor based on this map, including the following steps: The one-dimensional transmission attenuation signal of a freely falling target object is acquired using an X-ray detector array to generate the original X-ray transmission signal stream. The original X-ray transmission signal stream was subjected to logarithmic inverse transformation and deconvolution operation by physical attenuation model to reconstruct a two-dimensional surface density matrix as a two-dimensional mass density distribution map. Calculate the position of the centroid of the two-dimensional mass density distribution map, as well as the moment of inertia and product of inertia around the centroid, and combine them to generate a second-order inertial tensor for quantifying attitude stability.

[0008] A further aspect of the present invention, which outputs a phase-coded predicted state vector stream via phase modulation, includes the following steps: The belt speed parameters of the conveying equipment are extracted to form the initial conditions for conveying. Combined with the rotational dynamic process noise covariance matrix set by the second-order inertia tensor, the Kalman filter model is initialized together. The driving Kalman filter model performs iterative prediction based on a preset spatial quantization error threshold, generating a continuous state vector with future prediction times. The predicted time in the continuous state vector is embedded in the data packet synchronization header as a phase reference, and the output is a phase-coded predicted state vector stream.

[0009] A further aspect of this invention involves enabling the data reading link to parse a digital command containing two-dimensional position and predicted tilt angle, comprising the following steps: The phase reference is extracted from the predicted state vector stream of the phase code. When the phase difference between the phase reference and the local clock signal is less than a preset threshold and the duration reaches the preset lock time, the lock state of the hardware phase-locked loop circuit is triggered. In the locked state, output a high-level enable signal to activate the data read enable link; The data read enable link extracts the two-dimensional position and predicted tilt angle of the target object from the data frame and encapsulates them into digital instructions.

[0010] A further aspect of this invention involves extracting the two-dimensional position and predicted tilt angle of the target object from the data frame via a data read enable link and encapsulating them into digital instructions, including the following steps: Activate the data parsing logic unit coupled to the hardware phase-locked loop circuit; With the data read enable link kept open, the position coordinates and rotation angle word length contained in the serial data are extracted according to the preset frame format. The extracted position coordinates and rotation angle word length are loaded into the output buffer queue of the hardware control module and repackaged into digital instructions for downstream concurrent scheduling.

[0011] A further aspect of the present invention maps two-dimensional position to predicted tilt angle to generate an array drive control matrix, and triggers a micro-resonant cavity to emit pulse streams based on this matrix to construct a transient low-dust airflow region, comprising the following steps: The two-dimensional position and predicted tilt angle are mapped to the virtual coordinate system of the preset micro-resonant cavity array through affine transformation; A subset of micro-resonant cavities covering the projected contour of the target object is calculated to generate a binary high-level on-state vector, which is then used as the array drive control matrix. According to the array drive control matrix, a drive pulse is applied to the matched high-level micro resonant cavity, thereby exciting the sound and air coupled pulse flow to strip away dust and construct a transient low-dust airflow zone that envelops the object.

[0012] A further aspect of this invention involves capturing a two-dimensional optical image data stream within a transient low-dust airflow region and assigning it a unified timestamp, then mapping it to a two-dimensional mass density distribution map to achieve spatial-temporal coordinate alignment. This includes the following steps: When the target object arrives at the preset optical acquisition area along with the transient low-dust airflow zone, the industrial camera is triggered to acquire data to generate a two-dimensional optical image data stream. By using pre-calibrated sensor physical position offset parameters and predicted coherent motion trajectories for time interpolation, each row of the two-dimensional optical image data stream is given a precise acquisition time as a unified timestamp. The initial physical coordinates of each pixel in the two-dimensional optical image data stream are deduced from the kinematics based on the unified timestamp, and then matched with the corresponding points in the two-dimensional mass density distribution map at the pixel level to achieve spatial and temporal coordinate alignment.

[0013] A further aspect of the present invention outputs a category decision signal, comprising the following steps: The two-dimensional mass density distribution map is input into a two-branch neural network to extract the internal density gradient to form structured features; A two-dimensional matrix is ​​generated based on the gray-level uniform confidence score calculation of structured features to produce a dynamic feature weight matrix; The weighted surface features are obtained by performing element-wise multiplication between the dynamic feature weight matrix and the surface feature map of the two-dimensional optical image data stream. The weighted surface features and structured features are fed into a fully connected classifier to calculate the score and determine the output category decision signal.

[0014] A further aspect of this invention involves negatively feeding the data back to the Kalman filter model for self-calibration, including the following steps: The falling edge pulse of the object blocking signal is captured by a grating sensor located at the end of the material falling channel to record the actual arrival time; Match the corresponding specified prediction time from the historical prediction records, calculate the microsecond-level difference between the actual arrival time and the specified prediction time to generate a time prediction deviation vector; The time prediction deviation vector is used as feedback error and the velocity-related process noise covariance matrix in the Kalman filter model is adjusted to eliminate system drift.

[0015] Secondly, the present invention provides a dual-modal fusion identification system for flaky ore in high-dust environments, comprising the following modules: The density map generation module calls the detector to collect the original X-ray transmission signal stream of the target object, deconvolve and reconstruct a two-dimensional mass density distribution map, and calculates the second-order inertial tensor based on this. The predicted state vector generation module combines the initial conditions and the second-order inertial tensor to initialize the Kalman filter model to generate a continuous state vector, and then outputs a phase-encoded predicted state vector stream through phase modulation. The prediction instruction parsing module uses a hardware phase-locked loop circuit to lock the phase-encoded prediction state vector stream and enables the data reading enable link to parse out the digital instruction containing two-dimensional position and prediction tilt angle. The transient low-dust airflow zone generation module maps the two-dimensional position to the predicted tilt angle to generate an array drive control matrix, and triggers the micro resonant cavity to emit pulse flow based on this matrix to construct the transient low-dust airflow zone; The spatiotemporal coordinate alignment module captures two-dimensional optical image data streams in transient low-dust airflow areas and assigns them a unified timestamp, which is then mapped to a two-dimensional mass density distribution map to achieve spatial and temporal coordinate alignment. The fusion classification module extracts the structured features of the two-dimensional quality density distribution map after spatial and temporal coordinate alignment to generate a dynamic feature weight matrix. After weighting the two-dimensional optical image data stream, the structured features are stitched together to output the category decision signal. The model self-calibration module records the actual arrival time of the target object and compares it with the predicted time to generate a time prediction deviation vector, which is then negatively fed back to the Kalman filter model for self-calibration.

[0016] In summary, the present invention has the following beneficial technical effects: 1. By constructing a motion prediction model based on Kalman filtering and combining it with a hardware phase-locked loop circuit for timing synchronization, high-frequency iterative prediction of the spatial position and attitude of sheet-like ore during free fall was achieved. The output is a timestamped, phase-encoded predicted state vector stream, which is used as a unified execution reference by downstream control and acquisition systems. This transforms the passive responses of different sensors into active presets, providing a unified spatiotemporal reference system for acquiring X-ray and optical data, effectively reducing spatial misalignment and alignment deviations caused by data acquisition at different times and locations.

[0017] 2. Utilizing predicted spatiotemporal information, a micro-resonant cavity array is driven to emit transient acoustic-gas coupled pulses before the target object reaches the optical acquisition area. A transient low-dust airflow region is constructed based on the predicted contour of the target object. The transient acoustic-gas coupled pulse flow concentrates on the object surface for a short period during the target's passage, stripping away dust particles and creating a brief low-dust-density observation window around it. Because the optical camera's triggering is coupled during the establishment period of this clean field, the system provides unobstructed observation conditions for optical imaging without requiring continuous wide-area purging, ensuring the quality of surface optical feature data acquisition in dusty environments.

[0018] 3. A two-dimensional mass density distribution map reflecting the internal structure is used as attention-guided information. An attention network is employed to evaluate its uniformity to generate a dynamic feature weight matrix, which is then multiplied with optical image features. This mechanism ensures that optical features in regions with uniform density distribution receive higher weights in classification decisions, while the weights of optical features in regions with uneven density are suppressed. This physical property-guided fusion method effectively reduces misjudgments caused by localized surface contamination or minor internal anomalies, improving the reliability of joint decisions based on multimodal data.

[0019] 4. A closed-loop self-calibration mechanism is introduced. At the end of the acquisition area, a grating sensor captures the actual arrival time of the target object and compares it with the model's predicted time to generate a time prediction deviation vector. This time prediction deviation vector is fed back as feedback error to the Kalman filter model, dynamically correcting its state covariance matrix. This allows the model to learn and correct itself according to changes in the actual environment, reducing system drift during long-term operation and ensuring the long-term prediction accuracy and operational stability of the entire multimodal fusion recognition system. Attached Figure Description

[0020] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the accompanying drawings used in the description of the embodiments or the prior art will be briefly introduced below. The drawings are used to provide a further understanding of the present invention.

[0021] Figure 1 A flowchart illustrating an embodiment of this application is disclosed.

[0022] Figure 2 A schematic diagram of the framework in the embodiments of this application is disclosed.

[0023] Figure 3 The online adaptive convergence curve of the Kalman filter model in the embodiments of this application is disclosed.

[0024] Figure 4 The locking trigger determination curve of the hardware phase-locked loop circuit in the embodiments of this application is disclosed. Detailed Implementation

[0025] The following is in conjunction with the appendix Figure 1 - Figure 4 A preferred description of the present invention is provided below.

[0026] See attached document Figure 1 This invention proposes a dual-modal fusion identification method for flaky ore in high-dust environments, comprising the following steps: S1. Call the detector to collect the original X-ray transmission signal stream of the target object, deconvolve to reconstruct a two-dimensional mass density distribution map, and calculate the second-order inertial tensor based on this. S2. Combine the initial conditions of the transport with the second-order inertial tensor to initialize the Kalman filter model to generate a continuous state vector, and then output the phase-coded predicted state vector stream through phase modulation. S3. Use a hardware phase-locked loop circuit to lock the phase-encoded prediction state vector stream, and enable the data reading link to parse out the digital command containing the two-dimensional position and prediction tilt angle. S4. Map the two-dimensional position to the predicted tilt angle to generate an array drive control matrix, and trigger the micro resonant cavity to emit pulse flow based on this matrix to construct a transient low-dust airflow zone; S5. Capture the two-dimensional optical image data stream in the transient low-dust airflow area and assign a unified timestamp, and map it with the two-dimensional mass density distribution map to achieve spatial and temporal coordinate alignment. S6. Extract the structured features of the two-dimensional mass density distribution map after spatial and temporal coordinate alignment to generate a dynamic feature weight matrix. After weighting the two-dimensional optical image data stream, the structured features are spliced ​​together to output the category decision signal. S7. Record the actual arrival time of the target object and compare it with the predicted time to generate a time prediction deviation vector, and feed it back to the Kalman filter model for self-calibration.

[0027] In one embodiment of the present invention, step S1 includes the following steps: The one-dimensional transmission attenuation signal of a freely falling target object is acquired using an X-ray detector array to generate the original X-ray transmission signal stream. The original X-ray transmission signal stream was subjected to logarithmic inverse transformation and deconvolution operation by physical attenuation model to reconstruct a two-dimensional surface density matrix as a two-dimensional mass density distribution map. Calculate the position of the centroid of the two-dimensional mass density distribution map, as well as the moment of inertia and product of inertia around the centroid, and combine them to generate a second-order inertial tensor for quantifying attitude stability.

[0028] Specifically, step S1 of this invention is executed on a data processing server connected to the X-ray detector array. The X-ray detector array is invoked to acquire, in real time, the one-dimensional transmission attenuation signal of the freely falling target object, with the detector array sampling at a preset frequency. The X-ray photon flux passing through the target object is continuously captured, forming a time-series data set consisting of a one-dimensional intensity vector that varies over time. This data set is defined as the original X-ray transmission signal stream.

[0029] The original X-ray transmission signal stream is a two-dimensional matrix, where rows represent time sampling points and columns represent different pixel units of the X-ray linear array detector. Each value in the matrix is ​​the X-ray photon count value received by the corresponding pixel at the corresponding time.

[0030] The data processing server employs a deconvolution algorithm based on a physical attenuation model to process the raw X-ray transmission signal stream. The physical attenuation model, based on the Beer-Lambert law, describes the exponential attenuation of X-ray intensity as it passes through matter. Specifically, for each one-dimensional intensity vector in the raw X-ray transmission signal stream, the server uses a pre-calibrated initial intensity value without a target object. Normalization is performed, and then the inverse logarithmic transform of the Beer-Lambert law is applied to solve the intensity attenuation value as the integral value of the mass density line along the X-ray path.

[0031] By mapping continuous-time sampling points to vertical coordinates in two-dimensional space and stacking a series of solved one-dimensional line integral vectors, a two-dimensional mass density distribution map of the target object on a two-dimensional projection plane can be reconstructed. Two-dimensional mass density distribution map A two-dimensional digital image matrix, its pixel values It is proportional to the surface density of the projected object at that location, and the unit is kg / m².

[0032] Based on two-dimensional mass density distribution map Calculate the second-order inertia tensor used to quantify the attitude stability of a falling target object. Second-order inertial tensor This describes the mass distribution of an object in a two-dimensional plane and its response to rotational motion. The calculation process involves first integrating and summing all pixel values ​​of the two-dimensional mass density distribution map to obtain the total mass of the target object. Let the physical area of ​​a single pixel after physical size calibration be... The unit is m², and the formula for calculating the total mass is as follows:

[0033] Secondly, the position of the centroid of the target object is obtained by weighted averaging. The calculation formula is as follows:

[0034]

[0035] The moment of inertia and product of inertia of the two-dimensional mass density distribution map around the center of mass are calculated, and these components are combined into a 2x2 symmetric matrix, which is the second-order inertial tensor. The specific calculation formula is as follows:

[0036]

[0037] In the formula, These are the row and column pixel indices of the two-dimensional mass density distribution map, respectively. In the first Line number The pixel values ​​of the column, in kg / m²; and It is the first Column and number The physical space coordinates corresponding to the row, in meters; and These represent the objects orbiting their centers of mass. shaft and Moment of inertia of the shaft; and It is the product of inertia, and since it is a symmetric matrix, it satisfies... Once the calculation is complete, pass it as the final output of this step to the downstream steps.

[0038] Furthermore, the parameters set in this step are based on the following: assuming the target object is ore with an average particle size of 20 mm and a free fall velocity of approximately 2 m / s, and to ensure a vertical resolution of no less than 40 pixels, the sampling frequency of the X-ray detector array is... The setting is 4 kHz, based on the fact that the sampling frequency is limited by the falling speed of the target object and the required spatial resolution. The physical vertical spacing of a single row of pixels needs to reach [a certain value]. According to the formula The required sampling rate is calculated to be at least 4 kHz. Assuming the application scenario is the separation of calcite and dolomite, where the density difference is small, the tube voltage of the X-ray source is set in the range of 70-90 kV to improve the signal-to-noise ratio.

[0039] For example, suppose an L-shaped sheet of ore falls freely through the detection area. After processing in step S1, a discretized 3x3 pixel two-dimensional mass density distribution map is obtained. Its origin is located at the top left corner, and its row index is... The column index is Assuming the physical area of ​​each pixel after calibration... Assuming the mass density value of this graph, to facilitate subsequent calculations of minute physical quantities, a specific scaling factor is introduced to perform dimensionless processing. The following examples are all equivalent values ​​after scaling:

[0040] First, the data processing server calculates the total mass of the target object. This is obtained by summing all elements of the matrix and multiplying by the pixel area. Calculate the position of the centroid. center of mass coordinate Center of mass coordinate .

[0041] Calculate the second-order inertia tensor based on the centroid position. , quantity .

[0042] Quantity .

[0043] Inertial product The second-order inertial tensor output in step S1. This tensor will be passed to step S2 to initialize the Kalman filter model.

[0044] In one embodiment of the present invention, step S2 includes the following steps: The belt speed parameters of the conveying equipment are extracted to form the initial conditions for conveying. Combined with the rotational dynamic process noise covariance matrix set by the second-order inertia tensor, the Kalman filter model is initialized together. The driving Kalman filter model performs iterative prediction based on a preset spatial quantization error threshold, generating a continuous state vector with future prediction times. The predicted time in the continuous state vector is embedded in the data packet synchronization header as a phase reference, and the output is a phase-coded predicted state vector stream.

[0045] Specifically, step S2 continues on the data processing server. This step aggregates upstream data and initializes the prediction model to generate timing control signals. The server retrieves and receives the second-order inertial tensor generated in step S1. Simultaneously, it extracts the velocity parameters representing the conveying equipment at the material release point from the pre-configured system parameter repository. These velocity parameters are defined as initial motion conditions, i.e., a two-dimensional vector representing the translational velocity of the target object at the instant it leaves the conveying equipment, with units of m / s.

[0046] A Kalman filter model is initialized using initial motion conditions and a second-order inertia tensor. The Kalman filter model is a recursive Bayesian filter used to process time-varying signals, suitable for estimating and predicting the state of dynamic systems. In this embodiment, its state vector... Defined as , representing the two-dimensional position, two-dimensional velocity, tilt angle, and angular velocity of the target object's center of mass, respectively.

[0047] The initialization process specifically includes setting the translational velocity and the initial position determined by the position of the centroid of the second-order inertial tensor, which are both in the initial motion conditions, as the initial values ​​of the state vector of the Kalman filter model. Simultaneously, the second-order inertia tensor is used as a key input to define the process noise covariance matrix. Submatrices related to rotational dynamics.

[0048] Specifically, the process noise covariance matrix The diagonal elements represent the uncertainty in the model prediction of each state variable, and their rotation-related diagonal terms... and The value of is inversely proportional to the trace of the second-order inertia tensor. That is, the greater the inertia, the more stable the object's attitude, and the smaller the random perturbation assumed by the model, so that the model's prediction process can take into account the inherent attitude stability of the target object.

[0049] After initialization, the Kalman filter model performs iterative predictions in a high-frequency loop. The frequency of these predictions is set at a relatively high level, such as 100-200 kHz. This range is determined based on spatial quantization error control. When the ore's falling speed reaches 2 m / s, using an iteration frequency of 100 kHz results in a physical spatial displacement of only [value missing] between two adjacent prediction states. This displacement is much smaller than the physical resolution of 1mm for a micro-resonant cavity array, ensuring that the output phase-encoded command does not produce a spatial step effect when driving the downstream airflow jet, i.e., a jet blind zone caused by spatial quantization error, thus guaranteeing the smoothness and real-time performance of the predicted trajectory.

[0050] The loop operates in a fixed microsecond time step. Advance. In each iteration step The Kalman filter model only performs calculations in the prediction phase, updating the state vector based on the built-in kinematic equations of free fall. To further illustrate the prediction process of the Kalman filter model, its core mathematical operations follow the following state-space equations at each time step. The state prediction equation and the covariance prediction equation are as follows:

[0051]

[0052] in, At any moment Based on the time The state prediction vector made from the information provided so far. It is the final state estimate from the previous moment; and It is the corresponding state covariance matrix; superscript Represents the transpose of a matrix or vector; These are control inputs used to describe the effects of deterministic external forces such as gravity. In this example, the control input vector is the acceleration due to gravity. ,Right now Control input matrix This is used to map two-dimensional control inputs onto a six-dimensional state vector, and its form is:

[0053] It is a state transition matrix built based on Newton's laws of motion, describing how the state changes from time step [1] according to the physical model. Evolution to the moment For a time step Its form is:

[0054] In the formula, a series of future moments are generated through the above equation. Predicted two-dimensional position Predicting the dip angle and predicted angular velocity A continuous state vector, the continuous state vector is derived from the internal state vector. The extracted and organized data structure has the following format: .

[0055] The continuous state vector sequence is phase-coded to generate a phase-coded prediction state vector stream, a serial digital signal stream in which each data frame contains a continuous state vector. This encoding process is performed at each prediction time step. As a digital clock reference, this time value is placed in the synchronization header field of each state vector data packet, thereby embedding timing information as a phase reference into the output data stream, ensuring that downstream hardware circuits can achieve accurate data synchronization and parsing by locking this phase reference.

[0056] For example, the second-order inertial tensor output in step S1 is received. and its associated initial position of the centroid The belt speed of the conveyor is retrieved from the parameter library as the initial motion condition, and is set to 1.5 m / s in the horizontal direction. Initial tilt angle and angular velocity All are assumed to be 0. Therefore, the initial state vector of the Kalman filter model is... Set the prediction time step. gravitational acceleration .

[0057] Now we will perform the first iteration of prediction to calculate the time. The state. Based on the state prediction equation. The new x-coordinate is calculated as follows: New y-coordinate: ; New x-axis velocity: New y-axis velocity: ; New tilt angle: New angular velocity: .

[0058] After the prediction is completed, the server combines the required fields into a continuous state vector. This vector is serialized and its timestamp of 0.01 s is encoded into the synchronization header of the data packet. It is then sent out as part of the phase-coded prediction state vector stream for tracking and locking by the hardware phase-locked loop circuit in the downstream step S3.

[0059] In one embodiment of the present invention, step S3 includes the following steps: The phase reference is extracted from the predicted state vector stream of the phase code. When the phase difference between the phase reference and the local clock signal is less than a preset threshold and the duration reaches the preset lock time, the lock state of the hardware phase-locked loop circuit is triggered. In the locked state, output a high-level enable signal to activate the data read enable link; The data read enable link extracts the two-dimensional position and predicted tilt angle of the target object from the data frame and encapsulates them into digital instructions.

[0060] Specifically, step S3 of the method of the present invention is executed on the hardware control module configured on the execution end. This module integrates a hardware phase-locked loop circuit. The hardware phase-locked loop circuit is an integrated circuit, which mainly consists of a phase detector, a loop filter and a voltage-controlled oscillator. Its function is to synchronize the output signal with the input reference signal in terms of frequency and phase.

[0061] The hardware control module captures the phase-encoded predicted state vector stream output from step S2 in real time via its high-speed serial interface. The clock data recovery unit inside the hardware phase-locked loop circuit extracts the prediction time from the data frame synchronization header of the phase-encoded predicted state vector stream. This signal serves as an external reference clock signal, constituting a phase reference, i.e., composed of continuous predicted times. The rhythm information implied in the sequence.

[0062] The phase detector of the hardware phase-locked loop circuit continuously compares the phase of the external reference clock signal with the phase of the local clock signal generated by the internal voltage-controlled oscillator, and generates an error voltage signal proportional to the phase difference.

[0063] After the error voltage signal is smoothed by the loop filter, its amplitude remains continuously below the preset lock-in determination threshold. This lock-in determination threshold is equivalently mapped to the lock-in time error threshold (described later) in the time phase domain. After a preset locking time, the locking detection logic of the hardware phase-locked loop circuit determines that the system enters the locking state. This locking state is the stable operating mode of the hardware phase-locked loop circuit. In this mode, the output of its internal oscillator maintains a constant phase relationship with the external reference signal.

[0064] To describe the triggering condition of the locked state, let the reference clock pulse edge time corresponding to the phase reference of the phase-encoded predicted state vector stream be . The edge timing of the local clock pulse generated by the voltage-controlled oscillator inside the hardware phase-locked loop circuit is... If the system enters a locked state, the condition for enabling the data read link to be opened can be expressed as:

[0065] In the formula, System time; The edge time of the external reference clock signal; This represents the edge time of the local clock signal; This is the preset locking time error threshold, which is the time dimension expression of the aforementioned locking determination threshold. It is the starting moment when the judgment condition begins to be met; The preset lock time, lock time To prevent false locking caused by transient noise, the clock recovery circuit is typically set to a length of 10 to 50 reference clock cycles. This cycle is determined based on the typical characteristics of electromagnetic interference in industrial environments. In ore conveying, transient electromagnetic spikes caused by the start and stop of high-power belt motors usually last for tens to hundreds of microseconds. At a data rate of 100 kHz, i.e., a single cycle of 10 μs, 10 to 50 cycles correspond to a time window of 0.1 ms to 0.5 ms. This window length is just long enough to span and filter out most of the transient high-frequency interference in industrial environments, while ensuring that the clock recovery circuit can quickly and stably lock before the target ore arrives, achieving a balance between anti-interference and response speed.

[0066] The above formula shows that when the absolute value of the edge time difference between the reference clock and the local clock is within a certain duration... The error value within the time limit is always no greater than the lock time error threshold. When this time occurs, it is determined to be in a locked state. The locking time error threshold is included. It is the equivalent phase error threshold mapped to the time dimension, and its value is usually between 0.05 and 0.2 clock cycles.

[0067] For example, for a data rate of 100 kHz, the lock-time error threshold It can be set to 0.5-2.0. The numerical basis lies in the camera's exposure synchronization accuracy requirements. When an industrial camera captures an object moving at 2 m / s, to avoid image blurring or spatial misalignment exceeding one pixel, assuming a pixel accuracy of 0.1 mm, the trigger time jitter must be controlled within... μ Within 2.0 μs. The phase error of the phase-locked loop is controlled within 2.0 μs, which occupies the overall allowable jitter range. The proportion is less than 5%, leaving sufficient timing margin for the subsequent hardware execution circuits, ensuring that the optical data stream and X-ray density map can achieve sub-pixel spatial alignment. This value is the result of a trade-off between fast locking and jitter suppression capabilities.

[0068] Upon entering the locked state, the lock detection logic outputs a high-level enable signal. This signal activates the data read enable link, which refers to an enable gate controlled by the lock signal of the hardware phase-locked loop (PLL) circuit. Activating this gate activates a data parsing logic unit coupled to the PLL circuit, controlling the downstream data parsing unit to operate on the input data stream. With the data read enable link remaining active, the data parsing logic unit parses the data payload portion of the phase-encoded prediction state vector stream according to a predefined data frame format.

[0069] The data parsing unit selectively extracts two-dimensional position data from each state vector. With predicted dip angle These two fields are then repackaged into a new data structure, which forms a digital instruction. This instruction is a data vector containing all the spatial attitude information required to drive the downstream actuators, and its structure is as follows: It is then cached in the output register for use in the downstream step S4.

[0070] For example, the hardware control module receives a predicted state vector stream containing the phase-encoded first continuous state vector [0.01 s, (16.75 mm, 2.74 mm), 0 rad, 0 rad / s] generated in step S2. The hardware phase-locked loop circuit extracts the phase reference representing 0.01 s from the synchronization header of this data packet. Assume the system's locking time error threshold... It was set to 0.5 Lock time It is set to 10 clock cycles.

[0071] Upon receiving the data packet, the phase detector calculated a phase error of 0.1. This value is less than the lock time error threshold of 0.5. Since this condition has been met for more than 10 clock cycles during the reception of preceding data packets, the hardware phase-locked loop circuit confirms and maintains the locked state.

[0072] Therefore, the lock detection logic continuously outputs a high-level signal to keep the data read enable link open. The enabled data parsing logic unit then reads the payload portion of the data packet and extracts the two-dimensional position based on the preset offset and length. The predicted dip angle is The data parsing process combines these two data fields to form a digital instruction [16.75 mm, 2.74 mm, 0 rad]. This digital instruction is then loaded into the output FIFO buffer of the hardware control module, awaiting reading and execution by the control logic in step S4.

[0073] See Figure 4 The horizontal axis represents the system running time, and the vertical axis represents the equivalent time difference between the edges of the reference clock pulse and the local clock pulse. During the initial acquisition phase of the system, the phase difference fluctuates significantly, exhibiting a damped oscillation curve. As the phase-locked loop adjusts with negative feedback, the phase difference gradually decreases. The two horizontal dashed lines in the figure define the preset locking threshold. ， The example is set to 0.5. .

[0074] When the error time difference curve first falls into ±0.5 When the time difference amplitude remains within the threshold range, the system does not respond immediately; instead, it continuously monitors and only responds if and only if the time difference amplitude remains continuously within the threshold range to reach the preset threshold value. After the cycle, at the timing node indicated by the arrow in the figure, the hardware lock detection logic determines that the system has reached a truly stable lock state.

[0075] At this point, as shown in the gray shaded area on the right side of the diagram, the locking state is activated, and a high-level enable signal is output, meaning the authorized channel is open. This triggers the downstream circuitry to read and parse the phase-encoded data without error. This delayed confirmation mechanism effectively solves the false locking phenomenon caused by transient electromagnetic interference in industrial environments, ensuring the timing accuracy required for multimodal data alignment.

[0076] In one embodiment of the present invention, step S4 includes the following steps: The two-dimensional position and predicted tilt angle are mapped to the virtual coordinate system of the preset micro-resonant cavity array through affine transformation; A subset of micro-resonant cavities covering the projected contour of the target object is calculated to generate a binary high-level on-state vector, which is then used as the array drive control matrix. According to the array drive control matrix, a drive pulse is applied to the matched high-level micro resonant cavity, thereby exciting the sound and air coupled pulse flow to strip away dust and construct a transient low-dust airflow zone that envelops the object.

[0077] Specifically, step S4 is executed on the hardware control module at the execution end. The hardware control module is directly coupled to a micro resonant cavity array, wherein the micro resonant cavity array is a two-dimensional arrangement of microelectromechanical systems (MEMS) devices, and each unit on it is a Helmholtz resonant cavity that can be excited by an electrical signal to generate a high-speed airflow pulse.

[0078] The controller retrieves the latest digital instruction generated in step S3 from its internal cache. By performing a coordinate mapping operation, the two-dimensional position contained in the digital instruction is mapped... With predicted dip angle The system transforms from the global coordinate system to the preset virtual coordinate system of the micro-resonant cavity array. The virtual coordinate system is a two-dimensional Cartesian coordinate system with the geometric center or a corner point of the micro-resonant cavity array as the origin and the axis parallel to the row and column directions of the array. It is fixed to the physical surface of the micro-resonant cavity array, and the origin and axis are aligned with the physical layout of the array.

[0079] This mapping process is achieved through a pre-calibrated affine transformation matrix. To further clarify this coordinate mapping process, let the global coordinates in the digital command be... The corresponding coordinates in the virtual coordinate system are This mapping relationship is defined by the following affine transformation formula:

[0080] in, It is an affine transformation matrix, which includes translation and rotation transformations, used to compensate for the physical installation position offset and angular deviation of the micro-resonant cavity array relative to the origin of the global coordinate system; It is the installation rotation angle of the micro-resonant cavity array relative to the global coordinate system; and Its origin is shaft and Translation vector along the axis. Parameters The calibration is determined by optical calibration methods and stored in the controller's non-volatile memory during the initial installation or calibration of the device.

[0081] After completing the coordinate mapping, the controller calculates the activation vector to drive the micro-resonant cavity array based on the mapped position and tilt angle. Specifically, this calculation involves determining a subset of resonant cavities that matches the projected contour of the target object in the virtual coordinate system, setting the index positions of all resonant cavities in this subset to a high level "1" and the remaining positions to a low level "0", thereby generating a binary activation vector. This vector is a one-dimensional binary array whose length is equal to the total number of resonant cavities in the micro-resonant cavity array. The value of each element in the array determines whether the corresponding resonant cavity is triggered.

[0082] The enable vector is loaded into a parallel output port to form an array drive control matrix that corresponds one-to-one with the drive pins of the entire resonant cavity array. This is the physical implementation of the enable vector, representing the level state applied to each drive pin of the entire array.

[0083] The hardware control module, following the instructions of the array drive control matrix, operates at the times specified by the phase-encoded predicted state vector flow. A high-voltage pulse is applied to the excitation source of the corresponding micro-resonator calibrated as high level in the array drive control matrix. This pulse excites the corresponding micro-resonator to generate a transient high-voltage acoustic-air coupled pulse flow. This is a special fluid pulse generated by the Helmholtz resonator when it is excited. It has the high-speed propagation characteristics of sound waves and the macroscopic momentum of airflow, which can achieve efficient surface cleaning.

[0084] These pulse streams, emitted simultaneously from different locations, converge in the space surrounding the target object. With their powerful shearing force, they strip away and disperse the dust particles attached to the surface of the target object, thereby forming a transient low-dust airflow zone that envelops the moving target object. This zone is a short-lived local space created by the acoustic-air coupled pulse streams around the target object, with a dust density far lower than the ambient level, thus creating a low-dust window for subsequent optical acquisition.

[0085] For example, the controller reads the digital instruction [16.75 mm, 2.74 mm, 0 rad] generated in step S3 from the cache. It then invokes the preset affine transformation matrix. Perform coordinate mapping on the two-dimensional position specified in the command. Assume the calibration parameter is the rotation angle. Translation vector The mapped virtual coordinates are then... , The controller uses the virtual coordinates and predicted tilt angle. Based on the pre-stored approximate size of the target object, such as a circular region with a diameter of 5 mm, the subset of resonant cavities that need to be activated is calculated.

[0086] Assuming the resolution of the micro-resonant cavity array is 1 mm x 1 mm, the controller calculates that the center is located at... All resonant cavities are located in a circular region with a diameter of 5 mm. This may include approximately 20 resonant cavities indexed from (15, -9) to (19, -5).

[0087] The controller generates an activation vector with a length equal to the total number of array elements, setting the positions corresponding to the 20 resonant cavities to 1 and the rest to 0. This activation vector is applied in parallel to form the array drive control matrix. At the moment associated with this digital command... At a certain point in time, 0.01 s later, the controller simultaneously applies driving pulses to these 20 selected micro resonant cavities, thereby generating an acoustic-air coupled pulse flow whose shape roughly matches the projected outline of the target object, and constructing a transient low-dust airflow zone that follows the body's motion.

[0088] In one embodiment of the present invention, step S5 includes the following steps: When the target object arrives at the preset optical acquisition area along with the transient low-dust airflow zone, the industrial camera is triggered to acquire data to generate a two-dimensional optical image data stream. By using pre-calibrated sensor physical position offset parameters and predicted coherent motion trajectories for time interpolation, each row of the two-dimensional optical image data stream is given a precise acquisition time as a unified timestamp. The initial physical coordinates of each pixel in the two-dimensional optical image data stream are deduced from the kinematics based on the unified timestamp, and then matched with the corresponding points in the two-dimensional mass density distribution map at the pixel level to achieve spatial and temporal coordinate alignment.

[0089] Specifically, step S5 is executed on a synchronization control unit connected to the industrial camera and the data processing server. This synchronization control unit continuously monitors the phase-encoded predicted state vector stream generated in step S2, and the predicted centroid longitudinal coordinates within the stream... When the camera is about to enter the preset optical acquisition area, it is triggered to start acquiring data. The preset optical acquisition area is a region whose spatial location is fixed during equipment installation, and its coordinates are calibrated and stored in the system as the reference position for camera triggering.

[0090] Camera triggering is based on the instant the transient low-dust airflow zone constructed in step S4 reaches the optical acquisition area, ensuring that optical imaging is completed within the transient low-dust window. For example, when an industrial camera, such as a high-speed linear scan camera, is triggered, it captures the optical reflection signals from the target object's surface line by line as the target object crosses its field of view, thereby generating a two-dimensional optical image data stream. This two-dimensional optical image data stream is a two-dimensional matrix generated by the camera, where the values ​​represent the reflectivity or color information of the target object's surface in a specific wavelength band, such as visible light.

[0091] A timestamp operation is performed on each pixel in the two-dimensional optical image data stream. This operation incorporates pre-calibrated sensor physical position offset parameters, i.e., a transformation matrix containing translation and rotation components, describing the geometric relationship between the X-ray detector coordinate system and the industrial camera coordinate system, as well as the motion trajectory extracted from the phase-encoded predicted state vector stream. Specifically, for the first pixel in the two-dimensional optical image data stream... For each row pixel, its vertical coordinate in the global coordinate system is calculated based on its physical position on the camera sensor. Then, in the predicted motion trajectory data The search and interpolation are performed to calculate the centroid of the target object through the vertical coordinate. The time corresponding to the time.

[0092] To clarify the calculation method for time points, it is assumed that there exist two adjacent time points in the predicted state vector sequence. and Its corresponding predicted vertical coordinate and ,satisfy Then the unified timestamp of the corresponding time for that row of pixels. It can be calculated using linear interpolation:

[0093] This formula ensures that each line of optical image data is assigned a high-precision timestamp corresponding to its physical acquisition time. This timestamp forms a common time reference connecting the two-dimensional optical image data stream and the two-dimensional quality density distribution map, serving as a unified timestamp. Assign all pixels to that row.

[0094] According to the unified timestamp The two-dimensional optical image data stream is mapped pixel-level to the two-dimensional mass density distribution map obtained in step S1. This mapping is achieved through kinematic inverse calculation, assigning a uniform timestamp to any element in the optical image. and coordinates The pixels are used to calculate the physical point at the initial moment based on the predicted motion trajectory. Coordinates of time And correlate it with the corresponding two-dimensional mass density distribution map. The pixels at each location are associated. By performing this operation on all pixels, the spatial and temporal coordinates of the two heterogeneous data streams are ultimately aligned.

[0095] Spatial-temporal coordinate alignment is a data registration process that generates two images of the same size and with the same coordinate system: one is a density map, and the other is an optical map, with both images located at the same coordinate system. Each pixel corresponds to the same physical element on the target object.

[0096] For example, the synchronization control unit receives the predicted trajectory points generated in step S2, such as... Predicted location at time ,and Predicted location at time Assume the preset optical acquisition area starts at the vertical coordinate. Place.

[0097] When it is predicted that the target is about to arrive in the area, the industrial camera is triggered to capture data, and the camera captures the corresponding physical location. When the image line is displayed, a unified timestamp is calculated for it.

[0098] According to the interpolation formula above, the timestamp is This needs to be included. Time and position The optical pixels captured at that location, compared with the original two-dimensional mass density distribution map, as shown in the image taken at [location missing]. Alignment is performed at all times. Based on the motion trajectory model, the position of the physical point can be calculated in reverse. The position at that time.

[0099] Assuming there is no acceleration in the horizontal direction, its horizontal position is... For simplicity, we will use , The approximate velocity at 0.01 s. Its vertical position. It is possible Inverse solution, i.e. Thus obtain .

[0100] Therefore, in optical images The pixels collected at this location have coordinates approximately equal to those in the two-dimensional mass density distribution map. The pixels at each location are paired. By performing this mapping on each pixel of the optical image, a two-dimensional optical image data stream is output that is spatially and temporally perfectly aligned with the two-dimensional quality density distribution map.

[0101] In one embodiment of the present invention, step S6 includes the following steps: The two-dimensional mass density distribution map is input into a two-branch neural network to extract the internal density gradient to form structured features; A two-dimensional matrix is ​​generated based on the gray-level uniform confidence score calculation of structured features to produce a dynamic feature weight matrix; The weighted surface features are obtained by performing element-wise multiplication between the dynamic feature weight matrix and the surface feature map of the two-dimensional optical image data stream. The weighted surface features and structured features are fed into a fully connected classifier to calculate the score and determine the output category decision signal.

[0102] Specifically, step S6 is executed on the data processing server. Its core is to build and run a two-branch attention fusion neural network to classify target objects. Here, the two-branch neural network is a deep learning model with two parallel input paths. The two branches can have the same or different network structures to process data of different modalities respectively.

[0103] The server inputs the two-dimensional mass density distribution map, which is aligned with the spatial and temporal coordinates generated in step S5, into the first branch of the dual-branch neural network. This branch is a convolutional neural network, which extracts a series of multi-scale feature maps through multi-layer convolution and pooling operations. These feature maps encode information such as the density mean, variance, gradient direction, and texture of the two-dimensional mass density distribution map at a deep level, which together constitute structured features, containing deep information reflecting the internal physical properties of the object.

[0104] Simultaneously, the structured features are fed into an attention-generating network, which generates a dynamic feature weight matrix of the same size as the input image by evaluating the confidence level of gray-level uniformity embodied in the structured features. The core of the gray-level uniformity confidence level here is that for pure minerals, their internal density distribution should be relatively uniform; therefore, the variance or gradient magnitude of the gray-level values ​​in the density map is small, resulting in high confidence. Conversely, if cracks, impurities, or mixtures are present, the density is uneven, resulting in low confidence. The generated dynamic feature weight matrix is ​​a two-dimensional matrix, where each element has a value between 0 and 1, intuitively representing the confidence or importance of the optical features at the corresponding pixel location in the final classification decision.

[0105] Next, the two-dimensional optical image data stream, synchronized and aligned with the two-dimensional mass density distribution map, is input into the second branch of the dual-branch neural network. This branch is also a convolutional neural network, used to extract the original surface feature map. Subsequently, the server will use the dynamic feature weight matrix. Surface feature map output by the second branch For clarity, element-wise multiplication is performed, and surface features are weighted. The generation process is represented by the following formula:

[0106] in, This represents the Hadamard product, an element-wise multiplication operation. This operation dynamically adjusts the corresponding features in the surface feature map based on the weight values ​​at each position in the dynamic feature weight matrix, thereby obtaining a weighted surface feature. The effect is to suppress the contribution of optical features corresponding to low-confidence regions, which are non-uniform in the density map, while enhancing the influence of optical features corresponding to high-confidence regions, which are uniform in density. The weighted surface features and the structured features output from the first branch are concatenated along the channel dimension to form a high-dimensional feature vector that integrates internal structure and weighted surface information.

[0107] The high-dimensional feature vector is fed into the fully connected classifier at the end of the network. After several fully connected layers and a Softmax activation function, the output is the material classification score corresponding to each preset material category. The category with the highest score is selected as the decision result, and a category decision signal representing the category is output. This signal is a digital signal, such as an integer ID, used to uniquely identify the category of the target object being judged. For example, 0 represents calcite and 1 represents dolomite.

[0108] For example, suppose that after alignment in step S5, the system obtains a 2x2 pixel two-dimensional mass density distribution map. A 2x2 pixel two-dimensional optical image data stream . The unit is adu. The unit is gsv.

[0109] Two-dimensional mass density distribution map The signal is fed into the first branch of the CNN and simultaneously into the attention generation network. The attention network is analyzed using a two-dimensional quality density distribution diagram. The grayscale uniformity was examined, and it was found that the pixel value of 5.5 adu in the lower right corner differed significantly from other pixel values, indicating uneven density and low confidence in this area, while other areas had high confidence. Therefore, a dynamic feature weight matrix was generated. .at the same time, It is fed into the second branch of the CNN, assuming that its output is a surface feature map. Proportional to the input value, .

[0110] Perform element-wise multiplication to obtain weighted surface features . .

[0111] It can be seen that the optical eigenvalues ​​corresponding to the density anomaly region were suppressed from 18.0 to 1.80.

[0112] Next, the structured features extracted from the first branch are concatenated, assuming that... This represents the non-uniformity score and weighted surface features, forming an 8-dimensional fused feature vector. This vector is input into a fully connected classifier, which calculates a score of 0.92 for calcite and 0.08 for dolomite. The highest score is then selected and output as the class decision signal representing calcite.

[0113] In one embodiment of the present invention, step S7 includes the following steps: The falling edge pulse of the object blocking signal is captured by a grating sensor located at the end of the material falling channel to record the actual arrival time; Match the corresponding specified prediction time from the historical prediction records, calculate the microsecond-level difference between the actual arrival time and the specified prediction time to generate a time prediction deviation vector; The time prediction deviation vector is used as feedback error and the velocity-related process noise covariance matrix in the Kalman filter model is adjusted to eliminate system drift.

[0114] Specifically, step S7 is executed on a timing controller located at the end of the equipment's data acquisition area and communicating with the data processing server. This timing controller uses a set of grating sensors connected laterally at the end of the material falling channel to monitor the passage of the target object in real time.

[0115] When a target object completely blocks and subsequently completely leaves the beam of light from the grating sensor, the grating sensor generates a falling edge pulse signal. The timing controller captures the moment of this pulse signal and records it as the actual arrival time. ,time It is a timestamp recorded by a high-precision hardware timer, marking the exact moment when the target object physically passes completely through the acquisition end control line.

[0116] The data processing server retrieves the phase-encoded prediction state vector stream corresponding to the currently passing target object from its internal prediction data stream cache, and reads the predicted moment of passing the longitudinal position of the grating sensor from it. This moment is defined as the prediction moment. The server then performs a subtraction operation to calculate the actual arrival time. With the predicted time The difference between them generates a time prediction bias vector with microsecond-level accuracy. To illustrate the calculation of the time prediction deviation vector, it is defined as follows:

[0117] in, This is the time prediction bias vector, a scalar whose sign indicates whether the prediction is ahead or behind. This bias vector quantifies the time difference of the prediction model error, and its unit is typically 1200 kilobytes per second. Its value is used as an error signal in a closed-loop feedback control system.

[0118] Finally, the time prediction bias vector is used as a calibration input and fed back to the Kalman filter model running in step S2 as negative feedback. Upon receiving the calibration input, the Kalman filter model performs a state update or parameter correction operation; specifically, it uses this bias to dynamically correct its process noise covariance matrix. The dynamic correction process of the Kalman filter model is an online learning process that feeds back the deviation between actual measurements and model predictions to the model, allowing it to adjust its internal parameters to better fit the real physical environment.

[0119] For example, if the time prediction deviation vector remains positive, it indicates that the actual arrival time is later than predicted, which may indicate that the model underestimates the actual air resistance. Based on this, the state covariance matrix will be fine-tuned, especially by increasing the noise term related to speed. This will allow the model to take into account greater uncertainty or implicit damping terms in future predictions, making its subsequent predictions more conservative and closer to the actual descent situation. Ultimately, this will achieve online adaptive convergence and improve the long-term accuracy of predictions.

[0120] For example, the target object identified as calcite continues to fall and is about to pass through the grating sensor located at the end of the acquisition area. The data processing server queries its predicted data stream to learn that the predicted centroid position of the object will be... The sensing line of the grating sensor passes through the time, therefore When the calcite object actually falls and completely passes through the grating, the timing controller captures a falling edge pulse and records that moment as... .

[0121] Server calculates time prediction deviation vector This value is 23. The time prediction bias vector is used as calibration input and negatively fed back to the Kalman filter model in step S2.

[0122] Upon receiving this positive deviation, the self-calibration logic unit within the Kalman filter model interprets it as the current model's prediction of the object's falling velocity being slightly too high. Therefore, a parameter fine-tuning is performed to adjust the process noise covariance matrix. Diagonal elements related to y-axis velocity The value is increased by a small percentage, for example, by 0.5%.

[0123] This approach causes the model to assume greater uncertainty in its prediction of vertical velocity during the next iteration, resulting in a wider confidence interval in the propagation of state covariance. This indirectly makes its long-term trajectory prediction more conservative, compensating for the influence of factors such as air resistance that are not explicitly modeled, thereby reducing time bias in subsequent predictions.

[0124] See appendix Figure 3 This section provides supplementary explanations of the online adaptive convergence process of the Kalman filter model. The horizontal axis represents the number of closed-loop iterations during system operation, i.e., the batches of objects that have been identified and compared. The vertical axis represents the time prediction deviation vector obtained from the measurement. The scatter plot in the graph visually illustrates the dynamic evolution trend of the error in the closed-loop feedback system. In the initial stage of system operation, due to deviations in initial state settings or environmental parameters, such as unmodeled air resistance, the time prediction error can reach [amount missing]. The above; as step S7 continuously captures the actual arrival time, As feedback error propagation, and by dynamically fine-tuning the velocity-related process noise covariance matrix in the Kalman filter model, the model's predicted trajectory gradually approximates the actual physical fall. The figure clearly shows that after approximately 30 closed-loop iterations, it exhibits online adaptive convergence characteristics, and the time prediction deviation vector... Ultimately, it stabilizes at the ideal zero-deviation reference, i.e., with slight fluctuations near the horizontal dashed line in the figure. This indicates that the self-calibration mechanism of this invention successfully eliminates system drift caused by long-term operation, ensuring high-precision spatiotemporal coordinate alignment between X-rays and optical images throughout the entire lifespan of the device.

[0125] See appendix Figure 2 The present invention also proposes a dual-modal fusion recognition system for flaky ore in high-dust environments, comprising the following modules: The density map generation module calls the detector to collect the original X-ray transmission signal stream of the target object, deconvolve and reconstruct a two-dimensional mass density distribution map, and calculates the second-order inertial tensor based on this. The predicted state vector generation module combines the initial conditions and the second-order inertial tensor to initialize the Kalman filter model to generate a continuous state vector, and then outputs a phase-encoded predicted state vector stream through phase modulation. The prediction instruction parsing module uses a hardware phase-locked loop circuit to lock the phase-encoded prediction state vector stream and enables the data reading enable link to parse out the digital instruction containing two-dimensional position and prediction tilt angle. The transient low-dust airflow zone generation module maps the two-dimensional position to the predicted tilt angle to generate an array drive control matrix, and triggers the micro resonant cavity to emit pulse flow based on this matrix to construct the transient low-dust airflow zone; The spatiotemporal coordinate alignment module captures two-dimensional optical image data streams in transient low-dust airflow areas and assigns them a unified timestamp, which is then mapped to a two-dimensional mass density distribution map to achieve spatial and temporal coordinate alignment. The fusion classification module extracts the structured features of the two-dimensional quality density distribution map after spatial and temporal coordinate alignment to generate a dynamic feature weight matrix. After weighting the two-dimensional optical image data stream, the structured features are stitched together to output the category decision signal. The model self-calibration module records the actual arrival time of the target object and compares it with the predicted time to generate a time prediction deviation vector, which is then negatively fed back to the Kalman filter model for self-calibration.

[0126] Each of the modules can be implemented in whole or in part through software, hardware, or a combination thereof. It supports hardware embedded in or independent of the processor in the computer device, and also supports software stored in the memory of the computer device, so that the processor can call and execute the operations corresponding to each of the above modules.

[0127] 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 the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention, and should all be included within the protection scope of the present invention.

Claims

1. A method for dual-modal fusion recognition of flaky ore in a high-dust environment, characterized in that, Includes the following steps: S1. Call the detector to collect the original X-ray transmission signal stream of the target object, deconvolve to reconstruct a two-dimensional mass density distribution map, and calculate the second-order inertial tensor based on this. S2. Combine the initial conditions of the transport with the second-order inertial tensor to initialize the Kalman filter model to generate a continuous state vector, and then output the phase-coded predicted state vector stream through phase modulation. S3. Use a hardware phase-locked loop circuit to lock the phase-encoded prediction state vector stream, and enable the data reading link to parse out the digital command containing the two-dimensional position and prediction tilt angle. S4. Map the two-dimensional position to the predicted tilt angle to generate an array drive control matrix, and trigger the micro resonant cavity to emit pulse flow based on this matrix to construct a transient low-dust airflow zone; S5. Capture the two-dimensional optical image data stream in the transient low-dust airflow area and assign a unified timestamp, and map it with the two-dimensional mass density distribution map to achieve spatial and temporal coordinate alignment. S6. Extract the structured features of the two-dimensional mass density distribution map after spatial and temporal coordinate alignment to generate a dynamic feature weight matrix. After weighting the two-dimensional optical image data stream, the structured features are spliced ​​together to output the category decision signal. S7. Record the actual arrival time of the target object and compare it with the predicted time to generate a time prediction deviation vector, and feed it back to the Kalman filter model for self-calibration.

2. The method for dual-modal fusion recognition of flaky ore in a high-dust environment according to claim 1, characterized in that, The process involves using a detector to acquire the raw X-ray transmission signal stream of the target object, deconvolving it to reconstruct a two-dimensional mass density distribution map, and calculating the second-order inertial tensor based on this map. This includes the following steps: The one-dimensional transmission attenuation signal of a freely falling target object is acquired using an X-ray detector array to generate the original X-ray transmission signal stream. The original X-ray transmission signal stream was subjected to logarithmic inverse transformation and deconvolution operation by physical attenuation model to reconstruct a two-dimensional surface density matrix as a two-dimensional mass density distribution map. Calculate the position of the centroid of the two-dimensional mass density distribution map, as well as the moment of inertia and product of inertia around the centroid, and combine them to generate a second-order inertial tensor for quantifying attitude stability.

3. The method for dual-modal fusion recognition of flaky ore in a high-dust environment according to claim 1, characterized in that, The phase-modulated output phase-encoded predicted state vector stream includes the following steps: The belt speed parameters of the conveying equipment are extracted to form the initial conditions for conveying. Combined with the rotational dynamic process noise covariance matrix set by the second-order inertia tensor, the Kalman filter model is initialized together. The driving Kalman filter model performs iterative prediction based on a preset spatial quantization error threshold, generating a continuous state vector with future prediction times. The predicted time in the continuous state vector is embedded in the data packet synchronization header as a phase reference, and the output is a phase-coded predicted state vector stream.

4. The method for dual-modal fusion recognition of flaky ore in a high-dust environment according to claim 1, characterized in that, Enabling the data read enable link to parse the digital command containing the two-dimensional position and predicted tilt angle includes the following steps: The phase reference is extracted from the predicted state vector stream of the phase code. When the phase difference between the phase reference and the local clock signal is less than a preset threshold and the duration reaches the preset lock time, the lock state of the hardware phase-locked loop circuit is triggered. In the locked state, output a high-level enable signal to activate the data read enable link; The data read enable link extracts the two-dimensional position and predicted tilt angle of the target object from the data frame and encapsulates them into digital instructions.

5. The method for dual-modal fusion recognition of flaky ore in a high-dust environment according to claim 4, characterized in that, The data read enable link extracts the two-dimensional position and predicted tilt angle of the target object from the data frame and encapsulates them into digital instructions, including the following steps: Activate the data parsing logic unit coupled to the hardware phase-locked loop circuit; With the data read enable link kept open, the position coordinates and rotation angle word length contained in the serial data are extracted according to the preset frame format. The extracted position coordinates and rotation angle word length are loaded into the output buffer queue of the hardware control module and repackaged into digital instructions for downstream concurrent scheduling.

6. The method for dual-modal fusion recognition of flaky ore in a high-dust environment according to claim 1, characterized in that, The process involves mapping two-dimensional positions to predicted tilt angles to generate an array-driven control matrix, and then triggering a micro-resonant cavity to emit pulsed streams based on this matrix to construct a transient low-dust airflow region. This includes the following steps: The two-dimensional position and predicted tilt angle are mapped to the virtual coordinate system of the preset micro-resonant cavity array through affine transformation; A subset of micro-resonant cavities covering the projected contour of the target object is calculated to generate a binary high-level on-state vector, which is then used as the array drive control matrix. According to the array drive control matrix, a drive pulse is applied to the matched high-level micro resonant cavity, thereby exciting the sound and air coupled pulse flow to strip away dust and construct a transient low-dust airflow zone that envelops the object.

7. The method for dual-modal fusion recognition of flaky ore in a high-dust environment according to claim 1, characterized in that, Capturing a two-dimensional optical image data stream within a transient low-dust airflow region and assigning it a unified timestamp, then mapping it to a two-dimensional mass density distribution map to achieve spatial-temporal coordinate alignment, includes the following steps: When the target object arrives at the preset optical acquisition area along with the transient low-dust airflow zone, the industrial camera is triggered to acquire data to generate a two-dimensional optical image data stream. By using pre-calibrated sensor physical position offset parameters and predicted coherent motion trajectories for time interpolation, each row of the two-dimensional optical image data stream is given a precise acquisition time as a unified timestamp. The initial physical coordinates of each pixel in the two-dimensional optical image data stream are deduced from the kinematics based on the unified timestamp, and then matched with the corresponding points in the two-dimensional mass density distribution map at the pixel level to achieve spatial and temporal coordinate alignment.

8. The method for dual-modal fusion recognition of flaky ore in a high-dust environment according to claim 1, characterized in that, The output category decision signal includes the following steps: The two-dimensional mass density distribution map is input into a two-branch neural network to extract the internal density gradient to form structured features; A two-dimensional matrix is ​​generated based on the gray-level uniform confidence score calculation of structured features to produce a dynamic feature weight matrix; The weighted surface features are obtained by performing element-wise multiplication between the dynamic feature weight matrix and the surface feature map of the two-dimensional optical image data stream. The weighted surface features and structured features are fed into a fully connected classifier to calculate the score and determine the output category decision signal.

9. The method for dual-modal fusion recognition of flaky ore in a high-dust environment according to claim 1, characterized in that, The negative feedback is fed back to the Kalman filter model for self-calibration, including the following steps: The falling edge pulse of the object blocking signal is captured by a grating sensor located at the end of the material falling channel to record the actual arrival time; Match the corresponding specified prediction time from the historical prediction records, calculate the microsecond-level difference between the actual arrival time and the specified prediction time to generate a time prediction deviation vector; The time prediction deviation vector is used as feedback error and the velocity-related process noise covariance matrix in the Kalman filter model is adjusted to eliminate system drift.

10. A dual-modal fusion recognition system for flaky ore in a high-dust environment, characterized in that, Includes the following modules: The density map generation module calls the detector to collect the original X-ray transmission signal stream of the target object, deconvolve and reconstruct a two-dimensional mass density distribution map, and calculates the second-order inertial tensor based on this. The predicted state vector generation module combines the initial conditions and the second-order inertial tensor to initialize the Kalman filter model to generate a continuous state vector, and then outputs a phase-encoded predicted state vector stream through phase modulation. The prediction instruction parsing module uses a hardware phase-locked loop circuit to lock the phase-encoded prediction state vector stream and enables the data reading enable link to parse out the digital instruction containing two-dimensional position and prediction tilt angle. The transient low-dust airflow zone generation module maps the two-dimensional position to the predicted tilt angle to generate an array drive control matrix, and triggers the micro resonant cavity to emit pulse flow based on this matrix to construct the transient low-dust airflow zone; The spatiotemporal coordinate alignment module captures two-dimensional optical image data streams in transient low-dust airflow areas and assigns them a unified timestamp, which is then mapped to a two-dimensional mass density distribution map to achieve spatial and temporal coordinate alignment. The fusion classification module extracts the structured features of the two-dimensional quality density distribution map after spatial and temporal coordinate alignment to generate a dynamic feature weight matrix. After weighting the two-dimensional optical image data stream, the structured features are stitched together to output the category decision signal. The model self-calibration module records the actual arrival time of the target object and compares it with the predicted time to generate a time prediction deviation vector, which is then negatively fed back to the Kalman filter model for self-calibration.

Images