Real-time collision prediction and early warning method for astronomical robot fiber positioner

Abstract

This invention belongs to the field of astronomical fiber optic positioning and robot safety control technology, specifically relating to a real-time collision prediction and early warning method for astronomical robot fiber optic positioners. More than one hundred robot fiber optic positioners are arranged on the focal plane of a multi-object fiber optic spectroscopic telescope. While retaining the focal plane measurement camera and image acquisition link, this invention uses a programmable hardware chip for optical point position recognition to obtain the position and motion state of the end-point fiber optic point. Position features, angle features, motion state features, timestamps, and identification information are used to construct a time-series input sequence, which is then input into a lightweight spatiotemporal trajectory prediction model to obtain the end-point trajectory in the future prediction time domain. The future mechanism configuration is then recovered based on the inverse kinematics of a double-rotating mechanism, and the geometric gap between adjacent robot fiber optic positioners is calculated. When the geometric gap is less than a preset safety threshold, a collision warning signal is output. The collision prediction accuracy of this invention on a real platform is 98.98%.

Description

Technical Field

[0001] This invention belongs to the field of astronomical fiber optic positioning and robot safety control technology, specifically relating to a real-time collision prediction and early warning method for high-density dual-rotation robot fiber optic positioner arrays, and more particularly to an online early warning method based on visual perception, motion state extraction, spatiotemporal trajectory prediction, and geometric collision assessment using a hardware programmable chip for light spot position recognition. Background Technology

[0002] Multi-object fiber optic spectroscopic telescopes (MEPS) are specialized telescopes capable of observing multiple celestial objects simultaneously. Their focal planes are equipped with numerous robotic fiber optic positioners, each responsible for moving a fiber optic cable to the corresponding star image position to receive light signals. To ensure that as many positions on the focal plane as possible are reachable by the fiber optic cable, the range of motion of adjacent positioners often overlaps. Therefore, when many positioners move simultaneously, adjacent robotic arms may approach or even collide with each other. Collisions can not only damage the positioners but also cause the fiber optic ends to deviate from their target positions, thus reducing observation efficiency and quality.

[0003] The Large Sky Multi-Object Fiber Spectroscopic Telescope (LAMOST) fiber positioning system is a typical high-density fiber positioning system. The robotic fiber positioners in this system typically employ a dual-rotation structure, with a central arm and an eccentric arm jointly driving the end fiber's movement within the focal plane. When multiple positioners are closely arranged, each positioner often has potential collision relationships with several adjacent positioners. As positioner size decreases and arrangement density increases, and the future array size may further expand, the collision risk will also increase, becoming a significant factor affecting the system's safe operation and observational efficiency.

[0004] In a closed-loop observation system, the measuring camera repeatedly captures images of the focal plane. The system identifies the position of the backlighting fiber optic end from these images, calculates the rotation angles of the central axis and eccentric axis based on the difference between the current position and the target position, and converts this into drive pulses to control the positioner's movement. After the positioner moves, the measuring camera captures images again and performs feedback corrections until the positioning error meets the requirements. In other words, the collision risk is not only related to target allocation and path planning before observation, but also affected by factors such as motor asynchrony, mechanism deviation, feedback errors, or temporary anomalies during actual operation.

[0005] In existing technologies, one type of method mainly performs path planning and conflict resolution before the observation begins, that is, calculates the movement path of each locator in advance, hoping that the locator will not collide when moving along the predetermined path. This type of method is suitable for handling planned conflicts, but it usually relies on a pre-established model and ideal operating conditions; once the movement begins, if there is asynchrony in motor receiving commands, deviation in initial parameters, hardware abnormalities, or external disturbances, the originally planned path may not fully reflect the actual movement process.

[0006] Another type of method mainly performs collision detection during operation, that is, judging whether the locator has obviously approached or made contact based on the current image or sensor information. Although this type of method can detect some risks, it is often an "alarm after seeing the danger" approach, with a relatively late response time. For a large number of closely arranged and overlapping dual rotary locator arrays, it is difficult to leave enough time for the upper control system to decelerate, pause or replan based solely on the detection at the current moment.

[0007] Furthermore, some trajectory prediction or collision detection methods are designed for general moving targets or other mechanical systems, and do not fully consider the central arm and eccentric arm dual-rotation structure of the astronomical robot fiber optic positioner, as well as the fixed adjacency relationship between multiple adjacent positioners. Therefore, when these methods are directly applied to high-density fiber optic positioning arrays, they often fail to simultaneously meet the requirements of prediction accuracy, response speed, and ease of engineering deployment.

[0008] Therefore, a real-time collision prediction and early warning method is needed for fiber optic locator arrays for high-density dual-rotation robots in astronomy. This method should predict the movement trend of the locator in the short term while preserving the existing imaging links and closed-loop observation process as much as possible, and determine whether a collision will occur in the future based on the geometric relationship of the mechanism, thereby buying time for system intervention. Summary of the Invention

[0009] To enable the acquisition of the position and motion state of the locator's end effector using real-time images from the focal plane, predict the short-term trajectory, and output collision warning signals in advance based on the geometric relationship of the dual-rotation mechanism, this invention provides a real-time collision prediction and warning method for an astronomical robot fiber optic locator.

[0010] A real-time collision prediction and early warning method for an astronomical robot fiber optic locator is provided. The real-time collision prediction and early warning method is applied to a multi-object fiber optic spectroscopic astronomical telescope, which refers to an astronomical observation device that simultaneously collects the spectra of more than one hundred celestial targets through more than one hundred optical fibers.

[0011] The multi-object fiber optic spectroscopic telescope includes a robotic fiber optic locator array consisting of more than one hundred robotic fiber optic locators distributed on the focal plane, and an online monitoring system.

[0012] The robot fiber optic positioner includes a central axis 1, a central arm, an eccentric axis 2, an eccentric arm, and an end fiber optic cable.

[0013] The central axis 1 is located at the center of the robot's fiber optic positioner base. One end of the central arm is coaxially connected to the central axis 1, and the other end of the central arm is connected to one end of the eccentric arm via an eccentric shaft 2. An end fiber is installed at the other end of the eccentric arm. The central arm and the eccentric arm are sequentially connected along the focal plane to form a two-stage series rotation structure, constituting a double-rotation actuator. The central arm rotation angle θ... i (t) The angle formed by rotating about the center axis 1 as the center of rotation and taking the counterclockwise direction when viewed from the focal plane downwards as the positive direction; the eccentric arm rotation angle φ i (t) The angle formed by rotating with the axis of eccentric shaft 2 as the center of rotation, with the direction of the extension of the central arm from the central axis 1 to the eccentric shaft 2 as the reference direction, and with the counterclockwise direction when viewed from the focal plane as the positive direction.

[0014] The end of each optical fiber is fixedly attached to the end of the eccentric arm of a corresponding robot fiber optic positioner.

[0015] The operational safety framework of the multi-object fiber optic spectroscopic telescope comprises three layers from top to bottom. This framework is a hierarchical safety control system established around observation mission planning, systematic error suppression, and online monitoring during execution. The online monitoring system is located on the third layer, the real-time monitoring layer. The first layer of the operational safety framework is the planning stage protection layer, namely the target allocation and motion planning layer during the mission planning stage, which is used to complete target allocation, motion planning, and static geometric feasibility checks before observation execution. The second layer is the dynamic correction layer, namely the upstream error suppression and stability control layer, which is used to suppress positioning deviations caused by upstream disturbances such as atmospheric refraction correction, wind field, thermal deformation, and active optical adjustment. The third layer is the real-time monitoring layer, which is used to predict the future short-term trajectory based on real-time images and output collision warnings during execution.

[0016] The online monitoring system includes a focal plane measurement camera, an image acquisition module, a light spot position recognition hardware programmable chip (FPGA), a trajectory prediction and processing module, a geometric collision evaluation module, and an early warning output module.

[0017] The focal plane measurement camera is located above the focal plane and is used to image the back-illuminated fiber optic point on the focal plane.

[0018] The function of the robot fiber optic locator is to move the end fiber to the target star image position, and the center of the back-illuminated spot formed by the end fiber in the focal plane image is the end fiber point.

[0019] The target star image position refers to the target coordinate position on the focal plane after the celestial body to be observed is imaged by the telescope optical system. The robot fiber optic locator moves the end fiber to this position to receive the light signal of the celestial body.

[0020] The operation steps of the real-time collision prediction and early warning method are as follows:

[0021] S1: Acquire real-time observation data of the robot fiber optic positioner array to be monitored, wherein the real-time observation data includes at least the image information of the end fiber optic point of each robot fiber optic positioner in the focal plane.

[0022] S2: Extract the end position of each robot fiber optic locator based on the image information. The end position is the end fiber optic point, which serves as a position feature.

[0023] The angle features are obtained by solving the corresponding central arm rotation angle and eccentric arm rotation angle by combining the position features with the mechanism geometric parameters;

[0024] Based on the grayscale distribution change of the end fiber optic points between frames of continuous images in the image information, an image motion estimation method based on local window grayscale consistency constraints is used to obtain the image planar motion amount. Based on the amplitude of the image planar motion amount and the preset state, a threshold is divided, and a static state or a preset motion stage is determined to obtain the motion state characteristics of each robot fiber optic positioner.

[0025] The local window grayscale consistency constraint refers to the constraint that the grayscale distribution of the same end fiber point is approximately consistent in the local neighborhood of adjacent image frames.

[0026] The motion state features refer to the state labels used to distinguish whether the robot fiber optic positioner is stationary or in different stages of motion.

[0027] The geometric parameters of the mechanism include the rotation center coordinates of the i-th robot fiber optic positioner, the length of the central arm, the length of the eccentric arm, the rotation angle range of the central arm, the rotation angle range of the eccentric arm, the adjacency relationship between adjacent robot fiber optic positioners, and the center distance between adjacent robot fiber optic positioners.

[0028] The adjacency relationship between adjacent robot fiber optic positioners is such that the center distance within the focal plane is less than the sum of the inspection radii of the two robot fiber optic positioners, and there is an overlapping adjacent arrangement in the workspace, preferably a six-neighbor adjacency relationship.

[0029] The adjacency relationship refers to the geometric adjacency relationship formed when the center distance between two robot fiber optic positioners is no greater than the sum of their inspection radii in a fixed arrangement on the focal plane, and their workspaces overlap.

[0030] The six-neighbor adjacency relationship refers to the fact that in a close arrangement of approximately hexagons, a robot fiber optic locator and up to six other robot fiber optic locators with the closest center distance and potentially overlapping workspaces form adjacent candidate pairs, and subsequent collision evaluation is only performed between these adjacent candidate pairs.

[0031] S3: The position features, angle features, motion state features, timestamps and identification information of each robot fiber optic locator at continuous historical moments are used to form a time-series input sequence, and input into a lightweight spatiotemporal trajectory prediction model to obtain the end trajectory of each robot fiber optic locator in the future prediction time domain.

[0032] The temporal input sequence refers to a feature set of multiple frames of images arranged in chronological order.

[0033] The lightweight spatiotemporal trajectory prediction model refers to a model with fewer than 1.50M parameters, used to predict future terminal positions based on historical positions and states.

[0034] The lightweight spatiotemporal trajectory prediction model includes a physically guided feature extraction branch, a local temporal feature extraction branch, and a global temporal dependency modeling branch.

[0035] The physical guidance feature extraction branch is a computational branch used to introduce the geometric and kinematic relationships of the double rotary mechanism. The physical guidance feature extraction branch constructs physical features based on the relationship between the length of the central arm, the length of the eccentric arm, the rotation angle of the central arm, the rotation angle of the eccentric arm, and the partial derivative relationship between the end position and the rotation angles of the central arm and the eccentric arm, so that the prediction model retains the geometric constraints of the double rotary mechanism while learning the changes in historical trajectory.

[0036] The partial derivative relationship between the end position and the rotation angles of the central arm and the eccentric arm is a local Jacobian relationship that causes the end position to change when the rotation angles of the central arm and the eccentric arm change slightly.

[0037] The local temporal feature extraction branch is a computational branch used to extract local motion changes within a short period of time. It uses temporal convolution operations to extract short-term local motion patterns. Temporal convolution refers to performing convolution operations on the feature set of consecutive frame images along the time dimension.

[0038] The global temporal dependency modeling branch is a computational branch used to extract the overall motion trend within a longer historical window. It uses an attention mechanism to extract long-term dependency features. The attention mechanism refers to a computational method that assigns weights based on the correlation between different temporal features.

[0039] After fusing the position features, angle features, and motion state features, the encoder and lightweight decoder respectively output the predicted trajectories of all robot fiber optic positioners in the future prediction time domain.

[0040] The fusion refers to concatenating position features, angle features, and motion state features according to their correspondence at the same time into a unified feature vector, and then converting them into a latent feature representation of the same dimension through a linear mapping or encoding network;

[0041] The encoder is a network structure that converts input feature vectors into latent feature representations.

[0042] The lightweight decoder is an output network structure that converts latent features into future terminal positions with a small number of parameters.

[0043] The input feature vector at each moment includes the robot fiber optic locator identifier, timestamp, motion state label, end-effector position coordinates, central arm rotation angle, and eccentric arm rotation angle;

[0044] The input feature vector at each moment is defined as a set of model input data obtained by combining the identification information, timestamp, motion state label, end position coordinates, center arm rotation angle and eccentric arm rotation angle of the same robot fiber optic locator in a fixed order at a single sampling moment; the input feature vectors of m consecutive frames are stacked in time order to form a temporal input sequence, which serves as the common input of the three branches of the lightweight spatiotemporal trajectory prediction model.

[0045] S4: Based on the predicted trajectory of each fiber optic positioning robot in the future time domain, combine the inverse kinematics of the double rotary mechanism to recover the future mechanism configuration of each fiber optic positioning robot and the fiber optic positioner of the adjacent robot, and calculate the geometric gap between each fiber optic positioning robot and the fiber optic positioner of the adjacent robot.

[0046] The inverse kinematics refers to the geometric calculation process of inversely determining the rotation angles of the central arm and the eccentric arm from the end position;

[0047] The geometric gap refers to the distance between adjacent robot fiber optic positioners used to determine whether they are too close.

[0048] The future mechanism configuration includes the central axis rotation center, eccentric axis rotation center, end fiber point position, central arm position, and eccentric arm position of the adjacent robot fiber optic positioner at the future predicted time.

[0049] The rotation center of the eccentric shaft is determined by the rotation center of the central shaft, the length of the central arm, and the rotation angle of the central arm obtained by inverse kinematics recovery.

[0050] The location of the terminal fiber point is determined by the terminal trajectory in the future prediction time domain.

[0051] S5: When the geometric gap between any adjacent robot fiber optic positioners is less than the preset safety threshold, it is determined that there is a collision risk and an early warning signal is output.

[0052] The preset security threshold is determined by the physical dimensions of the optical fiber structure and the security margin.

[0053] The safety margin is used to cover trajectory prediction error, image measurement error, and performance tolerance;

[0054] The trajectory prediction error refers to the deviation between the predicted end position and the actual end position.

[0055] The image measurement error refers to the deviation in end-position measurement caused by camera imaging, light spot extraction, and centroid positioning.

[0056] The execution tolerance refers to the deviation between the actual motion and the commanded motion caused by differences in motor drive, mechanism assembly, transmission clearance, and control response.

[0057] The trajectory prediction endpoint displacement error of a single robot fiber optic locator is less than 28 micrometers, the forward inference time of the reference single locator is less than 3 milliseconds, the collision prediction accuracy is 98.98%, and the recall rate is 99.85%.

[0058] The trajectory prediction endpoint displacement error refers to the Euclidean distance error between the predicted endpoint and the corresponding actual endpoint, which is the square root of the sum of the squares of the coordinate errors of the final predicted frame in the x and y directions; wherein, the smaller the distance between the coordinates of the predicted endpoint and the coordinates of the actual endpoint, the smaller the trajectory prediction endpoint displacement error.

[0059] When implemented in array-level batch parallelism, the total processing latency is less than 6 seconds in the prediction time domain.

[0060] The array-level batch parallel implementation refers to using more than one hundred robot fiber optic positioners or more than one hundred adjacent candidate pairs as the same batch of inputs to complete trajectory prediction and collision assessment in parallel.

[0061] Further technical solutions are as follows:

[0062] The focal plane measurement camera of the online monitoring system is used to capture focal plane images. The image acquisition module is used to receive and transmit focal plane images. The light spot position recognition hardware programmable chip (FPGA) is a programmable hardware chip used to process the position and motion information of optical fiber points in the image in parallel. The trajectory prediction processing module is used to output the future end trajectory. The geometric collision evaluation module is used to calculate the spatial proximity between adjacent robot optical fiber positioners. The early warning output module is used to output collision risk warnings to the upper control system.

[0063] In step S2, the extraction of the end position of each robot fiber optic locator adopts the adaptive threshold T centroid localization method based on global statistical features. The global statistical features refer to the mean and standard deviation of the overall gray level of a focal plane image frame. The centroid localization method refers to the method of calculating the center coordinates of the end fiber optic point using the pixel gray level of the end fiber optic point as the weight. The adaptive threshold T formula (1) is as follows:

[0064]

[0065] In formula (1), μ is the mean gray value of the image, σ is the standard deviation of the gray value of the image, and α is the threshold coefficient, preferably 3;

[0066] According to formula (1), pixels with gray values ​​greater than the adaptive threshold T are selected as candidate pixels for the end fiber point, and binary threshold segmentation is performed. The binary threshold segmentation refers to an image segmentation method that divides pixels that meet the threshold condition into the end fiber point region and the remaining pixels into the background region. After threshold segmentation, the center coordinates of the end fiber point are obtained by connecting component labeling and centroid calculation. The connecting component labeling refers to a processing method that merges adjacent candidate fiber point pixels into the same pixel region. The formula (2) for the center coordinates of the end fiber point is as follows:

[0067]

[0068] In formula (2), R is the pixel region corresponding to the end fiber point, I(q) is the gray value of pixel q, and x q and y q Let x be the coordinates of pixel q in the focal plane image. c and y c These are the coordinates of the center of the end fiber optic point.

[0069] In step S2, the central arm rotation angle θ is calculated based on the end position of each fiber optic robot positioner. i (t) and eccentric arm rotation angle φ i (t), which satisfies the positive kinematic relationship of the double rotary mechanism. The positive kinematic relationship refers to the geometric relationship of the end position calculated from the rotation angle of the central arm and the rotation angle of the eccentric arm. The positive kinematic relationship formula (3) is as follows:

[0070]

[0071] In formula (3), p i t =(x i t ,y i t Let be the coordinates of the end position of the i-th robot fiber optic positioner at time t, and c be the coordinates of the end position of the i-th robot fiber optic positioner at time t. i =(x i 0 ,y i 0 Let l1 and l2 be the coordinates of the rotation center of the central axis of the i-th robot fiber optic positioner, and l1 and l2 be the lengths of the central arm and the eccentric arm, respectively; and the corresponding rotation angles of the central arm and the eccentric arm are recovered from the end position using inverse kinematics;

[0072] The Euclidean distance r between the end fiber point of the i-th robot fiber optic positioner and its central axis (1) rotation center in the focal plane is the radial distance r.i t The radial distance formula (4) is as follows:

[0073]

[0074] In formula (4), r i t Let x be the radial distance from the end position of the i-th robot fiber optic positioner to the center of its central axis rotation. i t and y i t x represents the coordinates of the end position. i 0 and y i 0 The coordinates of the rotation center are the central axis.

[0075] Center arm rotation angle θ i t and eccentric arm rotation angle φ i t The calculation formula (5) is as follows:

[0076]

[0077] In formula (5), θ i t Let φ be the rotation angle of the central arm of the i-th robot fiber optic positioner at time t. i t Let r be the eccentric arm rotation angle of the i-th robot fiber optic positioner at time t, l1 and l2 be the lengths of the central arm and eccentric arm, respectively, and r be the eccentric arm rotation angle. i t The radial distance is obtained from formula (4).

[0078] In step S2, the motion state features are obtained by an image motion estimation method based on local window grayscale consistency constraints. The image motion estimation method is used to estimate the image plane motion vector based on the local grayscale changes of the end fiber points in consecutive image frames. Under the condition of constant brightness, the optical flow constraint is satisfied. The optical flow refers to the apparent motion of the grayscale pattern of the fiber points in the image between adjacent frames. The optical flow constraint formula (6) is as follows:

[0079]

[0080] In formula (6), I x I y and I t Let represent the gradients of the image grayscale in the x, y, and time directions, respectively, and u and v represent the image planar motion vectors v=(u,v). TThe components in the x and y directions; based on the estimated optical flow amplitude, the robot fiber optic positioner is divided into a stationary state or multiple preset motion stages, and the optical flow amplitude and the discrete motion state label satisfy the following formula (7):

[0081]

[0082] In formula (7), M is the optical flow amplitude and S is the discrete motion state label; preferably, the discrete motion state label S takes the values ​​of 0, 1, 2, and 3, where 0 represents the stationary state, 1 represents the eccentric arm retraction stage, 2 represents the central arm alignment stage, and 3 represents the eccentric arm alignment stage.

[0083] In step S3, the input feature vector formula (8) at each time step is as follows:

[0084]

[0085] In formula (8), f i t Let ID be the input feature vector at time t. i This is the identifier for the robot's fiber optic locator, where t is the timestamp and S is the time stamp. i t For motion status labels, x i t and y i t For the end position, θ i t and φ i t These are the center arm rotation angle and the eccentric arm rotation angle, respectively.

[0086] The temporal input sequence consists of m consecutive historical observations; the m consecutive historical observations refer to the m consecutive images of the same robot fiber optic locator obtained at the sampling time before the current prediction time, and their corresponding position, angle and motion state features. The lightweight spatiotemporal trajectory prediction model outputs the predicted end trajectory for the next n frames, and the predicted end trajectory formula (9) is as follows:

[0087]

[0088] In formula (9), X i Let p be the temporal input sequence formed by the i-th robot fiber optic positioner within m consecutive historical observation frames. i t+k To predict the end position in the future k-th frame, Θ represents the learnable model parameters, and F... Θ is the mapping function corresponding to the lightweight spatiotemporal trajectory prediction model, preferably m equals n, and both are 15.

[0089] In step S3, the physical guidance feature extraction branch, local temporal feature extraction branch, and global temporal dependency modeling branch of the lightweight spatiotemporal trajectory prediction model are described as follows:

[0090] The physical guidance feature extraction branch constructs physical features based on the local Jacobian information of the forward kinematics of the double rotary mechanism. The local Jacobian information refers to the partial derivative relationship between the end position and the rotation angles of the central arm and the eccentric arm. The Jacobian component formula (10) is as follows:

[0091]

[0092] In formula (10), x and y are the abscissa and ordinate of the end fiber point in the focal plane, respectively; θ is the rotation angle of the central arm; φ is the rotation angle of the eccentric arm; l1 is the length of the central arm; l2 is the length of the eccentric arm; ∂x / ∂θ, ∂x / ∂φ, ∂y / ∂θ and ∂y / ∂φ are the partial derivatives of the end coordinates with respect to the rotation angles of the central arm and the eccentric arm, respectively, which are used to represent the local sensitivity of the end position to the changes in the two-stage rotation angles;

[0093] The four Jacobian components in formula (10) are stacked into a physical feature tensor J, where tensor J is an array organized in multiple dimensions, and physical guidance features are obtained through linear mapping; where Gaussian error linear unit activation function (GELU) is an activation function used to introduce nonlinear transformation, and the physical guidance feature formula (11) is as follows:

[0094]

[0095] In formula (11), F p Physical guidance features; X p The input is the feature to the physical guided feature extraction branch; J is the physical feature tensor formed by stacking the four Jacobian components of formula (10); W p and W J For linear mapping weights; b p and b J Here are the bias parameters; GELU is the activation function of the Gaussian error linear unit.

[0096] The local temporal feature extraction branch uses depthwise separable temporal convolution to extract short-term local motion patterns. The depthwise separable temporal convolution refers to the operation of first performing temporal convolution by channel and then reconstructing features through pointwise convolution. Among them, random deactivation regularization (Dropout) refers to the operation of randomly masking some features during training to reduce the risk of overfitting. The linear rectified activation function (ReLU) refers to the activation function that sets the negative input to zero and retains the positive input. The local temporal feature formula (12) is as follows:

[0097]

[0098] In formula (12), F l X represents local temporal features; K represents the temporal input feature tensor. d K is a depthwise convolution kernel that performs in-channel convolution along the time direction; p is the pointwise convolution kernel used to reconstruct channel features; * indicates convolution operation; ReLU is the linear rectified activation function; Dropout is the random deactivation regularization operation;

[0099] The global temporal dependency modeling branch uses a multi-head attention mechanism to extract long-term dependency features. The multi-head attention mechanism refers to the operation of mapping query Q, key K, and value V to multiple subspaces and performing weighted summation. Query Q, key K, and value V are the features to be matched, matching index features, and value features in the attention calculation, respectively. Among them, tensor concatenation operation (Concat) refers to connecting multiple feature tensors according to a specified dimension, and the normalized exponential function (Softmax) refers to the function that converts real number vectors into normalized weights. The global temporal dependency feature formulas (13) and (14) are as follows:

[0100]

[0101] In formula (13), F g The global time-series dependency features; Q, K, and V are the query feature, key feature, and value feature, respectively; head1 to head h For the output of h attention heads; W O These are the mapping parameters for the multi-head attention output; Concat is the operation that concatenates the outputs of multiple attention heads according to the feature dimension.

[0102]

[0103] In formula (14), head a W is the output of the a-th attention head; a Q W a K and W a V These are the query mapping matrix, key mapping matrix, and value mapping matrix for the a-th attention head, respectively; d k is the dimension of the key vector; Softmax is the normalization exponential function; the superscript T indicates matrix transpose;

[0104] The physical guidance feature extraction branch, the local temporal feature extraction branch, and the global temporal dependency modeling branch are concatenated and fused through linear projection. The fused feature formula (15) is as follows:

[0105]

[0106] In formula (15), F represents the fusion feature, Ffusion ... g For global temporal dependency features, F l For local temporal features, F p For physical guidance features, [·] represents feature concatenation operation, W f For linear projection parameters, R d This represents the d-dimensional feature space where the fused features reside.

[0107] In step S4, the geometric gap includes at least one of the following: the distance between the end fiber points of two adjacent robot fiber optic positioners, the vertical distance from the end fiber point of one robot fiber optic positioner to the eccentric arm of the adjacent robot fiber optic positioner, and the distance from the end fiber point of one robot fiber optic positioner to the rotation center of the eccentric arm of the adjacent robot fiber optic positioner; wherein, the position of the end fiber point of the i-th robot fiber optic positioner at the predicted future time is p. i The center of rotation of the eccentric shaft is e. i The eccentric arm direction vector of the adjacent robot fiber optic positioner j is d. j The eccentric arm direction vector d j It refers to the vector pointing from the center of rotation of the eccentric axis to the end fiber optic point; the Euclidean norm refers to the vector length; and the cross product is the vector product operation used to calculate the distance from a point to a line.

[0108] The geometric clearance formulas (16) to (18) are as follows:

[0109]

[0110] In formula (16), D ee ij p is the distance between the two end fiber optic points of adjacent robot fiber optic positioners i and j; i and p j Let be the positions of the end fiber points of the i-th and j-th robot fiber optic positioners at the predicted future time, respectively; ||·|| is the Euclidean norm, used to represent the Euclidean distance between the two points;

[0111]

[0112] In formula (17), D el ij p is the vertical distance from the end fiber optic point of a robot fiber optic positioner to the eccentric arm of an adjacent robot fiber optic positioner. i Let e ​​be the position of the end fiber point of the i-th robot fiber optic positioner at the predicted future time; j Let d be the position of the rotation center of the eccentric axis of the j-th robot fiber optic positioner; jLet be the eccentric arm direction vector of the j-th robot fiber optic positioner, and d j =p j -e j × represents the cross product operation;

[0113]

[0114] In formula (18), D ec ij p is the distance from the end fiber optic point of a robot fiber optic positioner to the eccentric axis rotation center of the adjacent robot fiber optic positioner. i Let e ​​be the position of the end fiber point of the i-th robot fiber optic positioner at the predicted future time; j Let be the position of the eccentric axis rotation center of the j-th robot fiber optic positioner at the predicted future time; ||·|| is the Euclidean norm;

[0115] The formula for the minimum geometric gap between adjacent robot fiber optic positioners (19) is as follows:

[0116]

[0117] In formula (19), D min ij is the minimum geometric gap between adjacent robot fiber optic positioners i and j; min represents the minimum value among the distance between end points, the distance from the point to the eccentric arm, and the distance from the point to the rotation center of the eccentric shaft.

[0118] The collision risk determination formula (20) for adjacent robot fiber optic positioners posing a collision risk is as follows:

[0119]

[0120] In formula (20), C is the set of adjacent robot fiber optic locators that are determined to have a collision risk; D safe This is a preset safety threshold; when D min ij Less than D safe When the collision risk is determined, add the adjacent pair (i,j) to set C and output the collision risk judgment.

[0121] The preset security threshold D safe Formula (21) is as follows:

[0122]

[0123] In formula (21), D safe For preset safety thresholds; D fiber Δ represents the physical dimensions of the fiber optic structure or the effective mechanical envelope dimensions; Δ is the safety margin used to cover trajectory prediction errors, image measurement errors, and performance tolerances.

[0124] In step S5, the prediction time-domain formula (22) for early warning triggering is as follows:

[0125]

[0126] In formula (22), Δt is the sampling interval, t pred For the time taken for trajectory prediction, t eval The time required for collision assessment, t comm To delay communication and execution, so as to ensure that the warning signal is output before physical contact occurs.

[0127] The beneficial technical effects of this invention are reflected in the following aspects:

[0128] 1. This invention, while retaining the existing charge-coupled device imaging link of the telescope, accelerates the extraction of end position and motion state through a light spot position recognition hardware programmable chip. It does not require additional independent sensors for each robot fiber optic locator, reducing the complexity of system modification and front-end sensing and processing latency, and is suitable for online deployment of high-density astronomical robot fiber optic locator arrays.

[0129] The method of this invention is applicable to high-density dual-rotation robot fiber optic positioner arrays in a six-neighbor overlapping workspace. It can achieve low-latency online collision risk monitoring of large-scale arrays by combining front-end perception accelerated by a light spot position recognition hardware programmable chip (FPGA) with back-end trajectory prediction and geometric collision assessment, while retaining the existing charge-coupled device imaging link (CCD imaging link).

[0130] 2. This invention does not only perform collision detection on the current position, but uses image information from continuous historical moments to predict the short-term trajectory in the future. Therefore, it can detect the risk of future collisions before actual contact occurs, allowing time for the upper control system to perform deceleration, pause, replanning or other safety measures, thus improving the advance warning and practicality.

[0131] 3. This invention incorporates the motion relationship of the dual-rotation mechanism into trajectory prediction and combines it with multiple geometric clearance indices for collision risk assessment, enabling it to better adapt to the complex geometric constraints formed by the joint motion of the central arm and the eccentric arm. Experimental results show that the trajectory prediction endpoint displacement error of this invention can be reduced to 27.09 micrometers, an improvement of 48.19% compared to the standard Transformer baseline, thus providing more reliable future position information for subsequent collision assessment.

[0132] 4. The real-time monitoring layer in this invention, within a three-layer security framework, provides real-time monitoring and early warning of residual deviations, asynchronous errors, and abnormal proximity events during the execution phase without altering the upstream target allocation and motion planning results. Experimental results show that this invention achieves a collision prediction accuracy of 98.98% and a recall rate of 99.85% on a real platform, with an average inference time of 2.6 milliseconds for a single robot fiber optic locator. This is beneficial for improving the operational safety and observation efficiency of large-scale astronomical robot fiber optic positioning systems. Attached Figure Description

[0133] Figure 1 This is a schematic diagram of the focal panel of the LAMOST robot fiber optic positioner in an embodiment of the present invention to which the present invention applies;

[0134] Figure 2 This is a schematic diagram of the structure of the robot fiber optic positioner to which this invention is applicable and the adjacent collision relationship;

[0135] Figure 3 This is a schematic diagram of the three-layer security framework in which the present invention is situated;

[0136] Figure 4 This is a schematic diagram of the area division and the coordinated orientation of adjacent units during the planning phase of this invention;

[0137] Figure 5 This is a schematic diagram of the overall process of the real-time collision prediction and early warning method based on the light spot position recognition hardware programmable chip perception of the present invention;

[0138] Figure 6 This is a schematic diagram illustrating the extraction of fiber optic point distribution and optical flow motion state under back illumination conditions using a programmable hardware chip for optical point position recognition based on the present invention.

[0139] Figure 7 This is a schematic diagram of the collision criterion for adjacent robot fiber optic positioners in this invention. Detailed Implementation

[0140] The present invention will be further described below with reference to the accompanying drawings and embodiments. It should be understood that the following embodiments are for illustrative purposes only and are not intended to limit the scope of protection of the present invention.

[0141] Example 1

[0142] This invention provides a real-time collision prediction and early warning method for an astronomical robot fiber optic locator, applicable to multi-object fiber optic spectroscopic telescopes.

[0143] See Figure 1A multi-object fiber optic spectroscopic telescope is an astronomical observation device that simultaneously acquires the spectra of more than one hundred celestial objects using more than one hundred optical fibers. The focal plane instrument of a multi-object fiber optic spectroscopic telescope is its data receiving instrument, which includes an array of more than one hundred robotic fiber optic positioners distributed across the focal plane and an online monitoring system. This array of more than one hundred robotic fiber optic positioners enables real-time acquisition of astronomical observation data.

[0144] See Figure 2 The robot fiber optic positioner includes a central axis 1, a central arm, an eccentric axis 2, an eccentric arm, and an end fiber optic cable.

[0145] See Figure 2 The central axis 1 is located at the center of the robot's fiber optic positioner base. One end of the central arm is coaxially connected to the central axis 1, and the other end of the central arm is connected to one end of the eccentric arm via an eccentric axis 2. The other end of the eccentric arm is equipped with a terminal fiber optic cable. The central arm and eccentric arm are sequentially connected along the focal plane to form a two-stage series rotation structure, constituting a double-rotation actuator. The central arm rotation angle θ... i (t) The angle formed by rotating about the axis of central axis 1 as the center of rotation, with the counterclockwise direction when viewed from the focal plane as the positive direction; the eccentric arm rotation angle φ i (t) The angle formed by rotating with the axis of eccentric shaft 2 as the center of rotation, with the direction of the extension of the central arm from the central axis 1 to the eccentric shaft 2 as the reference direction, and with the counterclockwise direction when viewed from the focal plane as the positive direction.

[0146] The end of each optical fiber is fixedly attached to the end of the eccentric arm of a corresponding robot fiber optic positioner.

[0147] The operational safety framework of the multi-object fiber optic spectroscopic telescope comprises three layers from top to bottom. This operational safety framework is a hierarchical safety control system established around observation mission planning, systematic error suppression, and online monitoring of the execution process.

[0148] The online monitoring system is permanently installed on the third-layer real-time monitoring layer. (See also...) Figure 3 The first layer of the operational safety framework is the planning phase protection layer, namely the target allocation and motion planning layer in the mission planning phase, which is used to complete target allocation, motion planning and static geometric feasibility checks before observation execution; the second layer is the dynamic correction layer, namely the upstream error suppression and stability control layer, which is used to suppress positioning deviations caused by upstream disturbances such as atmospheric refraction correction, wind field, thermal deformation and active optical adjustment; the third layer is the real-time monitoring layer, which is used to predict the future short-term trajectory based on real-time images and output collision warnings during execution.

[0149] See Figure 3 , Figure 3The three-layer safety framework of the robot fiber optic positioner is shown. The first layer corresponds to the target allocation and motion planning in the task planning stage, the second layer corresponds to the upstream error suppression and stability control, and the third layer corresponds to the online monitoring system in claim 1, which consists of a focal plane measurement camera, an image acquisition module, a light spot position recognition hardware programmable chip, a trajectory prediction and processing module, a geometric collision evaluation module, and an early warning output module.

[0150] See Figure 4 This demonstrates the regional division and the directional relationship between adjacent units during the planning phase. Figure 4 In the planning stage, 'a' represents the regional division based on the overlapping relationship of the workspaces of adjacent robot fiber optic locators. Region 1, Region 2, and Region 3 correspond to the potential overlapping regions formed by the target unit and different adjacent locators, respectively. Figure 4 In this context, 'b' represents a schematic diagram showing that after planning the cooperative orientation of adjacent units based on the regional division, the central arm and eccentric arm of each robot fiber optic positioner move in a direction that avoids each other as much as possible; it is used to illustrate that before the online monitoring layer of this invention is executed, the first two layers of safety mechanisms have minimized planned mechanical interference; this invention further provides real-time early warning for residual deviations, asynchronous errors, or abnormal proximity events during unplanned execution.

[0151] The online monitoring system includes a focal plane measurement camera, an image acquisition module, a light spot position recognition hardware programmable chip (FPGA), a trajectory prediction and processing module, a geometric collision assessment module, and an early warning output module.

[0152] A focal plane measurement camera is located above the focal plane and is used to image the backlit end fiber optic point on the focal plane. An image acquisition module receives and transmits the focal plane image; a programmable field-of-sight (FPGA) chip for spot position recognition is a programmable hardware chip used for parallel processing of the fiber optic point position and motion information in the image; a trajectory prediction processing module outputs the predicted future trajectory of the fiber optic end; a geometric collision assessment module calculates the spatial proximity between adjacent robot fiber optic positioners; and a warning output module outputs collision risk warnings to the upper-level control system.

[0153] The function of the robot fiber optic positioner is to move the end fiber to the target star image position. The center of the back-illuminated spot formed by the end fiber in the focal plane image is the end fiber point.

[0154] The target star image position refers to the target coordinate position on the focal plane after the celestial body to be observed is imaged by the telescope's optical system. The robot fiber optic locator moves the end fiber to this position to receive the light signal of the celestial body.

[0155] This invention retains the existing charge-coupled device (CCD) imaging link of the telescope for focal plane image acquisition. Subsequent fiber centroid extraction and optical flow estimation are performed by a hardware programmable chip for spot position recognition, reducing latency on the sensing side. For each frame of the acquired backlit image, the mean grayscale value μ and the standard deviation σ are first calculated, and an adaptive threshold is determined according to T = μ + ασ, where α is preferably 3. Then, the centroid of the connected regions after threshold segmentation is calculated to obtain the center coordinates (x, y) of the fiber end. c ,y c The value is taken as the end-effector position measurement of the robot's fiber optic locator at that moment. This processing method is suitable for online implementation using a programmable chip for light spot position recognition, enabling real-time position input with low computational overhead. Preferably, the camera exposure time for the focal plane image is 50 milliseconds, and the image sampling period is 0.4 seconds / frame, i.e., 2.5 frames / second. According to the statistical results of 26,334 detected light spots in 399 experimental images, the average half-width and full height of the light spots is 5.81 ± 0.19 pixels, and the average extracted area is 74.65 pixels, indicating that the backlit light spots have a stable medium expansion size. In offline verification, the average difference between the centroid method and the two-dimensional Gaussian fitting center is 0.0112 pixels, the root mean square error is 0.0124 pixels, and the 95th percentile difference is 0.0204 pixels, which can meet the accuracy requirements of end-effector position extraction in this embodiment.

[0156] See Figure 5 , Figure 5 The overall flow of the real-time collision prediction and early warning method of the present invention is shown, which consists of five operation steps: acquiring real-time observation data, extracting position features and motion state features, constructing a time-series input sequence and predicting the future end trajectory, restoring the future mechanism configuration and calculating the geometric gap, and outputting an early warning signal according to a preset safety threshold.

[0157] The operational steps for real-time collision prediction and early warning using the astronomical robot fiber optic locator are as follows:

[0158] S1: Acquire real-time observation data of the robot fiber optic positioner array to be monitored, wherein the real-time observation data includes at least the image information of the end fiber optic point of each robot fiber optic positioner in the focal plane.

[0159] S2: Extract the end position of each robot fiber optic locator based on the image information. The end position is the end fiber optic point, which serves as a position feature.

[0160] By combining the positional features with the mechanism's geometric parameters, the corresponding rotation angles of the central arm and eccentric arm are solved to obtain the angle features.

[0161] Based on the grayscale distribution changes of the end fiber points between frames of continuous images in the image information, an image motion estimation method based on local window grayscale consistency constraints is used to obtain the image planar motion amount. According to the amplitude of the image planar motion amount and the preset state, a threshold is divided, and a static state or a preset motion stage is determined to obtain the motion state characteristics of each robot fiber optic positioner.

[0162] The local window grayscale consistency constraint refers to the constraint that the grayscale distribution of the same end fiber point should be approximately consistent in the local neighborhood of adjacent image frames.

[0163] The motion state characteristics refer to the state labels used to distinguish whether the robot fiber optic positioner is stationary or in different stages of motion.

[0164] The geometric parameters of the mechanism include the rotation center coordinates of the i-th robot fiber optic positioner, the length of the central arm, the length of the eccentric arm, the rotation angle range of the central arm, the rotation angle range of the eccentric arm, the adjacency relationship between adjacent robot fiber optic positioners, and the center distance between adjacent robot fiber optic positioners.

[0165] The adjacency relationship between adjacent robot fiber optic positioners is such that the center distance within the focal plane is less than the sum of the inspection radii of the two robot fiber optic positioners, and there is an overlapping adjacent arrangement in the workspace, preferably a six-neighbor adjacency relationship.

[0166] The adjacency relationship refers to the geometric adjacency formed when, in a fixed arrangement on the focal plane, the center distance between two robot fiber optic positioners is no greater than the sum of their inspection radii, and their workspaces overlap.

[0167] The six-neighbor adjacency relationship refers to the fact that in a close arrangement of approximately hexagons, a robot fiber optic locator and up to six other robot fiber optic locators with the closest center distance and potentially overlapping workspaces form adjacent candidate pairs, and subsequent collision evaluation is only performed between these adjacent candidate pairs.

[0168] The specific steps for extracting the end position of each robot fiber optic positioner are as follows:

[0169] See Figure 6 This demonstrates the distribution of light spots and the extraction process of optical flow motion state after the optical fiber is lit under back illumination conditions. Figure 6 In this context, 'a' represents the original spot image of a single end fiber point under back illumination conditions. Figure 6 In this context, b represents the three-dimensional distribution of grayscale values ​​within a local window of the fiber optic point at the end. Figure 6 In this context, 'c' represents the extracted results of the arc-shaped motion trajectory and optical flow motion state of the end fiber point of a single robot fiber optic positioner in consecutive frames. Figure 6In the diagram, 'd' represents the trajectory distribution and motion state extraction results of the end fiber points of multiple robot fiber optic positioners in the same focal plane image. This corresponds to the adaptive threshold centroid localization method and the image motion estimation method based on local window grayscale consistency constraints; where the center of the end fiber point is used to form position features, and the grayscale changes in consecutive frames are used to form motion state features.

[0170] An adaptive threshold T centroid localization method based on global statistical features is adopted. The global statistical features refer to the mean and standard deviation of the overall gray level of a frame of focal plane image. The centroid localization method refers to the method of calculating the center coordinates of the end fiber point using the pixel gray level of the end fiber point as the weight. The adaptive threshold T formula (1) is as follows:

[0171]

[0172] In formula (1), μ is the mean gray value of the image, σ is the standard deviation of the gray value of the image, and α is the threshold coefficient, preferably 3.

[0173] According to formula (1), pixels with gray values ​​greater than the adaptive threshold T are selected as candidate pixels for the end fiber point, and binary threshold segmentation is performed. The binary threshold segmentation refers to an image segmentation method that divides pixels that meet the threshold condition into the end fiber point region and the remaining pixels into the background region. After threshold segmentation, the center coordinates of the end fiber point are obtained by connecting component labeling and centroid calculation. The connecting component labeling refers to a processing method that merges adjacent candidate fiber point pixels into the same pixel region. The formula (2) for the center coordinates of the end fiber point is as follows:

[0174]

[0175] In formula (2), R is the pixel region corresponding to the end fiber point, I(q) is the gray value of pixel q, and x q and y q Let x be the coordinates of pixel q in the focal plane image. c and y c These are the coordinates of the center of the end fiber optic point.

[0176] The preferred center coordinates of the end fiber optic point are calculated according to formula (2), in which the gray value of each candidate pixel is used as a weight in the center coordinate calculation. According to the statistical results of 26,334 detected light points in 399 experimental images, the average half-height and full width of the light points is 5.81±0.19 pixels, and the average extracted area is 74.65 pixels. In offline verification, the average difference between the centroid method and the two-dimensional Gaussian fitting center is 0.0112 pixels, the root mean square error is 0.0124 pixels, and the 95th percentile difference is 0.0204 pixels, indicating that the centroid localization method can provide a stable position input for subsequent trajectory prediction.

[0177] For the i-th robot fiber optic positioner, its end effector position satisfies the positive kinematics relationship of the double rotary mechanism, that is, the end effector focal plane coordinates are determined by the position of the rotation center, the length of the central arm, the length of the eccentric arm, and the rotation angle θ of the central arm. i t and eccentric arm rotation angle φ i t The decision is made jointly. Based on this mechanism, after obtaining the end-effector position, the corresponding central arm rotation angle θ can be recovered by further combining known geometric parameters. i t and eccentric arm rotation angle φ i t This allows us to obtain the mechanism state variables required for subsequent trajectory prediction and collision assessment. Simultaneously, optical flow estimation based on a local window grayscale consistency constraint is performed on consecutive image frames, satisfying I under constant brightness conditions. x u+I y v+I t =0; Based on the optical flow amplitude and state division threshold, the motion state of the robot fiber optic locator can be discretized into a stationary state or several preset motion stages, thereby forming a motion state label S. Preferably, the label can be {0,1,2,3}.

[0178] In solving the angular features, the relationship between the end position, the rotation center of the central axis, the length of the central arm, and the length of the eccentric arm is preferably determined according to formulas (3) to (5); where formula (3) represents the forward kinematic relationship of the double rotary mechanism, formula (4) represents the radial distance from the end fiber point to the rotation center of the central axis, and formula (5) represents the inverse kinematic calculation relationship for recovering the rotation angle of the central arm and the rotation angle of the eccentric arm from the end position. In the nominal physical parameters of the upgraded robot fiber optic positioner, the length of the central arm and the length of the eccentric arm are both 3 mm, and the center distance between adjacent robot fiber optic positioners is 10.6 mm.

[0179] The central arm rotation angle θ is calculated based on the end position of each fiber optic robot positioner. i (t) and eccentric arm rotation angle φ i (t), which satisfies the positive kinematic relationship of the double rotary mechanism. The positive kinematic relationship refers to the geometric relationship of the end position calculated from the rotation angle of the central arm and the rotation angle of the eccentric arm. The positive kinematic relationship formula (3) is as follows:

[0180]

[0181] In formula (3), p i t =(x i t ,y i t Let be the coordinates of the end position of the i-th robot fiber optic positioner at time t, and c be the coordinates of the end position of the i-th robot fiber optic positioner at time t.i =(x i 0 ,y i 0 Let l1 and l2 be the coordinates of the rotation center of the central axis of the i-th robot fiber optic positioner, and l1 and l2 be the lengths of the central arm and the eccentric arm, respectively; and the corresponding rotation angles of the central arm and the eccentric arm are recovered from the end position using inverse kinematics;

[0182] The Euclidean distance r between the end fiber point of the i-th robot fiber optic positioner and its central axis (1) rotation center in the focal plane is the radial distance r. i t The radial distance formula (4) is as follows:

[0183]

[0184] In formula (4), r i t Let x be the radial distance from the end position of the i-th robot fiber optic positioner to the center of its central axis rotation. i t and y i t x represents the coordinates of the end position. i 0 and y i 0 The coordinates of the center of rotation are the central axis.

[0185] Center arm rotation angle θ i t and eccentric arm rotation angle φ i t The calculation formula (5) is as follows:

[0186]

[0187] In formula (5), θ i t Let φ be the rotation angle of the central arm of the i-th robot fiber optic positioner at time t. i t Let r be the eccentric arm rotation angle of the i-th robot fiber optic positioner at time t, l1 and l2 be the lengths of the central arm and eccentric arm, respectively, and r be the eccentric arm rotation angle. i t The radial distance is obtained from formula (4).

[0188] In motion state feature extraction, the image motion estimation method preferably adopts optical flow estimation based on local window grayscale consistency constraints, which satisfies the optical flow constraint of formula (6) under constant brightness conditions; then, the optical flow amplitude is mapped to discrete motion state labels according to formula (7). Preferably, the discrete motion state label S takes the values ​​0, 1, 2, and 3, where 0 represents the stationary state, 1 represents the eccentric arm retraction stage, 2 represents the central arm alignment stage, and 3 represents the eccentric arm alignment stage.

[0189] The motion state features are obtained by an image motion estimation method based on local window grayscale consistency constraints. The image motion estimation method is used to estimate the image plane motion vector based on the local grayscale changes of the end fiber points in consecutive image frames. Under the condition of constant brightness, the optical flow constraint is satisfied. The optical flow refers to the apparent motion of the grayscale pattern of the fiber points in the image between adjacent frames. The optical flow constraint formula (6) is as follows:

[0190]

[0191] In formula (6), I x I y and I t Let represent the gradients of the image grayscale in the x, y, and time directions, respectively, and u and v represent the image planar motion vectors v=(u,v). T The components in the x and y directions; based on the estimated optical flow amplitude, the robot fiber optic positioner is divided into a stationary state or multiple preset motion stages, and the optical flow amplitude and the discrete motion state label satisfy the following formula (7):

[0192]

[0193] In formula (7), M is the optical flow amplitude and S is the discrete motion state label; preferably, the discrete motion state label S takes the values ​​of 0, 1, 2, and 3, where 0 represents the stationary state, 1 represents the eccentric arm retraction stage, 2 represents the central arm alignment stage, and 3 represents the eccentric arm alignment stage.

[0194] S3: After extracting position features, angle features, and motion state features, construct the input feature vector f for each time step. i t =[ID i ,t,S i t ,x i t ,y i t ,θ i t ,φ i t ], where ID iThis represents the identifier of the robot's fiber optic locator, where t represents the timestamp, (x i t ,y i t ) indicates the end position, θ i t and φ i t S represents the rotation angle of the center arm and the rotation angle of the eccentric arm, respectively. i t This represents a discrete motion state label. The position characteristics, angle characteristics, motion state characteristics, timestamps, and identification information of each robot fiber optic locator at continuous historical moments are used to construct a time-series input sequence, which is then input into the lightweight spatiotemporal trajectory prediction model PLST-Former to obtain the end-point trajectory of each robot fiber optic locator in the future prediction time domain.

[0195] In this embodiment, 15 consecutive historical observations are preferably used as input to predict the end motion trajectory of the subsequent 15 frames. The length of the historical window and the prediction window can also be adjusted according to the system sampling period and delay budget. In this embodiment, the image sampling period is 0.4 seconds, the 15 consecutive historical observations correspond to a 6-second observation window, and the predicted trajectory of the next 15 frames corresponds to a 6-second prediction time domain. The window length matches the warning trigger time constraint in claim 9.

[0196] The temporal input sequence refers to a set of features from multiple frames of images arranged in chronological order.

[0197] The lightweight spatiotemporal trajectory prediction model has fewer than 1.50M parameters and is used to predict future terminal positions based on historical positions and states.

[0198] The lightweight spatiotemporal trajectory prediction model includes a physical-guided feature extraction branch, a local temporal feature extraction branch, and a global temporal dependency modeling branch.

[0199] The physical guidance feature extraction branch is a computational branch used to introduce the geometric and kinematic relationships of the double rotary mechanism to enhance the physical rationality of the predicted trajectory. The physical guidance feature extraction branch constructs physical features based on the relationship between the length of the central arm, the length of the eccentric arm, the rotation angle of the central arm, the rotation angle of the eccentric arm, and the partial derivative relationship between the end position and the rotation angles of the central arm and the eccentric arm, so that the prediction model retains the geometric constraints of the double rotary mechanism while learning the changes in historical trajectories.

[0200] The partial derivative relationship between the end position and the rotation angles of the central arm and the eccentric arm is a local Jacobian relationship that causes the end position to change when the rotation angles of the central arm and the eccentric arm change slightly.

[0201] The local temporal feature extraction branch is a computational branch used to extract local motion changes within a short period of time. It uses temporal convolution operations to extract short-term local motion patterns. Temporal convolution refers to performing convolution operations on the feature set of consecutive frame images along the time dimension.

[0202] The global temporal dependency modeling branch is a computational branch used to extract the overall motion trend within a longer historical window. It uses an attention mechanism to extract long-term dependency features. The attention mechanism refers to a computational method that assigns weights based on the correlation between different temporal features.

[0203] By fusing the positional, angular, and motion state features, the encoder and lightweight decoder output the predicted trajectories of all robot fiber optic positioners in the future time domain. Compared to general pure data-driven time series models, this lightweight spatiotemporal trajectory prediction model is more suitable for high-density dual-rotation robot fiber optic positioner scenarios.

[0204] The fusion refers to concatenating position features, angle features, and motion state features according to their correspondence at the same time into a unified feature vector, and then converting it into a latent feature representation of the same dimension through a linear mapping or encoding network.

[0205] The encoder is a network structure that converts input feature vectors into latent feature representations.

[0206] The lightweight decoder is an output network structure that converts latent features into future terminal positions with a small number of parameters.

[0207] The input feature vector at each moment includes the robot fiber optic locator identifier, timestamp, motion status label, end-effector position coordinates, center arm rotation angle, and eccentric arm rotation angle.

[0208] The input feature vector at each moment is defined as a set of model input data obtained by combining the identification information, timestamp, motion state label, end position coordinates, center arm rotation angle and eccentric arm rotation angle of the same robot fiber optic locator in a fixed order at a single sampling moment; the input feature vectors of m consecutive frames are stacked in time order to form a temporal input sequence, which serves as the common input of the three branches of the lightweight spatiotemporal trajectory prediction model.

[0209] The formula (8) for the input feature vector at each time step is as follows:

[0210]

[0211] In formula (8), f i t Let ID be the input feature vector at time t. i This is the identifier for the robot's fiber optic locator, where t is the timestamp and S is the time stamp. it For motion status labels, x i t and y i t For the end position, θ i t and φ i t These are the center arm rotation angle and the eccentric arm rotation angle, respectively.

[0212] The temporal input sequence consists of m consecutive historical observations; the m consecutive historical observations refer to the m consecutive images of the same robot fiber optic locator obtained at the sampling time before the current prediction time, and their corresponding position, angle and motion state features. The lightweight spatiotemporal trajectory prediction model outputs the predicted end trajectory for the next n frames, and the predicted end trajectory formula (9) is as follows:

[0213]

[0214] In formula (9), X i Let p be the temporal input sequence formed by the i-th robot fiber optic positioner within m consecutive historical observation frames. i t+k To predict the end position in the future k-th frame, Θ represents the learnable model parameters, and F... Θ is the mapping function corresponding to the lightweight spatiotemporal trajectory prediction model, preferably m equals n, and both are 15.

[0215] When the image sampling period is 0.4 seconds / frame, 15 consecutive historical observations correspond to a 6-second observation window, and the predicted trajectory for the next 15 frames corresponds to a 6-second prediction time domain.

[0216] To eliminate input inconsistencies between different robot fiber optic locators, different frame numbers, and different feature quantities, before inputting the lightweight spatiotemporal trajectory prediction model, the robot fiber optic locator identifier, timestamp, end-effector position, center arm rotation angle, eccentric arm rotation angle, and motion state label are aligned according to the same focal plane coordinate system, the same sampling order, and a fixed feature order. The aligned input feature vectors of consecutive m frames are then stacked into a temporal input sequence.

[0217] The physical guidance feature extraction branch, local temporal feature extraction branch, and global temporal dependency modeling branch of the lightweight spatiotemporal trajectory prediction model are described below:

[0218] The physical guidance features are constructed using formulas (10) and (11), as detailed below:

[0219] The physical guidance feature extraction branch constructs physical features based on the local Jacobian information of the forward kinematics of the double rotary mechanism. The local Jacobian information refers to the partial derivative relationship between the end position and the rotation angles of the central arm and the eccentric arm. The Jacobian component formula (10) is as follows:

[0220]

[0221] In formula (10), x and y are the abscissa and ordinate of the end fiber point in the focal plane, respectively; θ is the rotation angle of the central arm; φ is the rotation angle of the eccentric arm; l1 is the length of the central arm; l2 is the length of the eccentric arm; ∂x / ∂θ, ∂x / ∂φ, ∂y / ∂θ and ∂y / ∂φ are the partial derivatives of the end coordinates with respect to the rotation angles of the central arm and the eccentric arm, respectively, which are used to represent the local sensitivity of the end position to small changes in the two-stage rotation angles.

[0222] The four Jacobian components in formula (10) are stacked into a physical feature tensor J, which is an array organized in multiple dimensions. The physical guided features are obtained through linear mapping. Among them, the Gaussian error linear unit activation function (GELU) is an activation function used to introduce nonlinear transformation. The physical guided feature formula (11) is as follows:

[0223]

[0224] In formula (11), F p Physical guidance features; X p The input is the feature to the physical guided feature extraction branch; J is the physical feature tensor formed by stacking the four Jacobian components of formula (10); W p and W J For linear mapping weights; b p and b J is the bias parameter; GELU is the Gaussian error linear unit activation function.

[0225] Formula (11) performs linear mapping and nonlinear transformation between local Jacobian information and input features to obtain physical guidance features.

[0226] The process of extracting short-time local motion patterns using depthwise separable temporal convolution according to formula (12) is explained in detail below:

[0227] The local temporal feature extraction branch uses depthwise separable temporal convolution to extract short-term local motion patterns. The depthwise separable temporal convolution refers to the operation of first performing temporal convolution by channel and then reconstructing features through pointwise convolution. Among them, random deactivation regularization (Dropout) refers to the operation of randomly masking some features during training to reduce the risk of overfitting. The linear rectified activation function (ReLU) refers to the activation function that sets the negative input to zero and retains the positive input. The local temporal feature formula (12) is as follows:

[0228]

[0229] In formula (12), F l X represents local temporal features; K represents the temporal input feature tensor. d K is a depthwise convolution kernel that performs in-channel convolution along the time direction; p is the pointwise convolution kernel used to reconstruct channel features; * indicates convolution operation; ReLU is the linear rectified activation function; Dropout is the random deactivation regularization operation.

[0230] The process of extracting long-term dependent features using the multi-head attention mechanism based on formulas (13) and (14) is as follows:

[0231] The global temporal dependency modeling branch uses a multi-head attention mechanism to extract long-term dependency features. The multi-head attention mechanism refers to the operation of mapping query Q, key K, and value V to multiple subspaces and performing weighted summation. Query Q, key K, and value V are the features to be matched, matching index features, and value features in the attention calculation, respectively. Among them, tensor concatenation operation (Concat) refers to connecting multiple feature tensors according to a specified dimension, and the normalized exponential function (Softmax) refers to the function that converts real number vectors into normalized weights. The global temporal dependency feature formulas (13) and (14) are as follows:

[0232]

[0233] In formula (13), F g The global time-series dependency features; Q, K, and V are the query feature, key feature, and value feature, respectively; head1 to head h For the output of h attention heads; W O These are the mapping parameters for the multi-head attention output; Concat is the operation that concatenates the outputs of multiple attention heads according to the feature dimension.

[0234]

[0235] In formula (14), head a W is the output of the a-th attention head; aQ W a K and W a V These are the query mapping matrix, key mapping matrix, and value mapping matrix for the a-th attention head, respectively; d k is the dimension of the key vector; Softmax is the normalization exponential function; the superscript T indicates matrix transpose.

[0236] The splicing and linear projection fusion are performed according to formula (15), and the specific operation is as follows:

[0237] The physical guidance feature extraction branch, the local temporal feature extraction branch, and the global temporal dependency modeling branch are concatenated and fused through linear projection. The fused feature formula (15) is as follows:

[0238]

[0239] In formula (15), F represents the fusion feature, Ffusion ... g For global temporal dependency features, F l For local temporal features, F p For physical guidance features, [·] represents feature concatenation operation, W f For linear projection parameters, R d This represents the d-dimensional feature space where the fused features reside.

[0240] S4: Based on the predicted trajectory of each fiber optic positioning robot in the future time domain, combine the inverse kinematics of the double rotary mechanism to recover the future mechanism configuration of each fiber optic positioning robot and the fiber optic positioner of the adjacent robot, and calculate the geometric gap between each fiber optic positioning robot and the fiber optic positioner of the adjacent robot.

[0241] The inverse kinematics refers to the geometric calculation process of inversely determining the rotation angles of the central arm and the eccentric arm from the end position.

[0242] The geometric gap refers to the distance between adjacent robot fiber optic positioners used to determine whether they are too close.

[0243] The future mechanism configuration includes the central axis rotation center, eccentric axis rotation center, end fiber point position, central arm position, and eccentric arm position of the adjacent robot fiber optic positioner at the predicted future time.

[0244] The rotation center of the eccentric shaft is determined by the rotation center of the central shaft, the length of the central arm, and the rotation angle of the central arm obtained by inverse kinematics recovery.

[0245] The location of the terminal fiber point is determined by the terminal trajectory in the future prediction time domain.

[0246] See Figure 7 , Figure 7 The collision criteria of adjacent robot fiber optic positioners are shown, where the distance between the end fiber optic points, the vertical distance from the end fiber optic point to the adjacent eccentric arm, and the distance from the end fiber optic point to the rotation center of the adjacent eccentric shaft correspond to the geometric gap calculation and collision risk determination defined by formulas (16) to (21).

[0247] During the collision risk assessment phase, for each pair of adjacent robot fiber optic positioners i and j, based on the predicted end-effector positions at future predicted times, and combined with the inverse kinematics of the double rotary mechanism, the corresponding future mechanism configuration is recovered, and the following types of geometric gaps are calculated: the distance D between the two end-effector fiber optic points. ee ij The vertical distance D from the end fiber optic point of a robot fiber optic positioner to the eccentric arm of an adjacent robot fiber optic positioner. el ij ; and the distance D from the end fiber optic point of a robot fiber optic positioner to the rotation center of the eccentric arm of the adjacent robot fiber optic positioner. ec ij When D is satisfied min ij <D safe When the adjacent pair is determined to have a collision risk, D safe =D fiber +Δ,D fiber Δ represents the physical dimensions of the fiber optic structure, and Δ is a safety margin used to cover prediction errors and performance tolerances.

[0248] The position of the end fiber point of the i-th robot fiber optic positioner at the predicted future time is p. i The center of rotation of the eccentric shaft is e. i The eccentric arm direction vector of the adjacent robot fiber optic positioner j is d. j The eccentric arm direction vector d j It refers to the vector pointing from the center of rotation of the eccentric axis to the end fiber point. The Euclidean norm refers to the vector length. The cross product refers to the vector product operation used to calculate the distance from a point to a line.

[0249] The geometric clearance formulas (16) to (18) are explained in detail below:

[0250] The distance D between the two end fiber points ee ij Formula (16) is as follows:

[0251]

[0252] In formula (16), D ee ij p is the distance between the two end fiber optic points of adjacent robot fiber optic positioners i and j; i and pj Let be the positions of the end fiber points of the i-th and j-th robot fiber optic positioners at the predicted future time, respectively; ||·|| is the Euclidean norm, used to represent the Euclidean distance between the two points;

[0253] The vertical distance D from the end fiber optic point of a robot fiber optic positioner to the eccentric arm of an adjacent robot fiber optic positioner el ij Formula (17) is as follows:

[0254]

[0255] In formula (17), D el ij p is the vertical distance from the end fiber optic point of a robot fiber optic positioner to the eccentric arm of an adjacent robot fiber optic positioner. i Let e ​​be the position of the end fiber point of the i-th robot fiber optic positioner at the predicted future time; j Let d be the position of the rotation center of the eccentric axis of the j-th robot fiber optic positioner; j Let be the eccentric arm direction vector of the j-th robot fiber optic positioner, and d j =p j -e j ; × represents the cross product operation.

[0256] The distance D from the end fiber optic point of a robot fiber optic positioner to the eccentric axis rotation center of an adjacent robot fiber optic positioner ec ij Formula (18) is as follows:

[0257]

[0258] In formula (18), D ec ij p is the distance from the end fiber optic point of a robot fiber optic positioner to the eccentric axis rotation center of the adjacent robot fiber optic positioner. i Let e ​​be the position of the end fiber point of the i-th robot fiber optic positioner at the predicted future time; j Let be the position of the eccentric axis rotation center of the j-th robot fiber optic positioner at the predicted future time; ||·|| is the Euclidean norm.

[0259] The formula for the minimum geometric gap between adjacent robot fiber optic positioners (19) is as follows:

[0260]

[0261] In formula (19), D min ijis the minimum geometric gap between adjacent robot fiber optic positioners i and j; min represents the minimum value among the distance between end points, the distance from the point to the eccentric arm, and the distance from the point to the rotation center of the eccentric shaft.

[0262] The collision risk determination formula (20) for adjacent robot fiber optic positioners posing a collision risk is as follows:

[0263]

[0264] In formula (20), C is the set of adjacent robot fiber optic locators that are determined to have a collision risk; D safe This is a preset safety threshold; when D min ij Less than D safe When the collision risk is determined, add the adjacent pair (i,j) to set C and output the collision risk judgment.

[0265] The preset security threshold D safe Formula (21) is as follows:

[0266]

[0267] In formula (21), D safe For preset safety thresholds; D fiber Δ represents the physical dimensions of the fiber optic structure or the effective mechanical envelope dimensions; Δ is the safety margin used to cover trajectory prediction errors, image measurement errors, and performance tolerances.

[0268] Preset safety threshold D safe The physical dimensions or effective mechanical envelope size D of the optical fiber structure fiber It is determined together with the safety margin Δ.

[0269] S5: When the geometric gap between any adjacent robot fiber optic positioners is less than a preset safety threshold, a collision risk is determined and a warning signal is output.

[0270] The preset safety threshold is determined by the physical dimensions of the optical fiber structure and the safety margin.

[0271] The safety margin is used to cover trajectory prediction errors, image measurement errors, and performance tolerances.

[0272] The trajectory prediction error refers to the deviation between the predicted end position and the actual end position.

[0273] The image measurement error refers to the deviation in end-position measurement caused by camera imaging, light spot extraction, and centroid positioning.

[0274] The execution tolerance refers to the deviation between the actual motion and the commanded motion caused by differences in motor drive, mechanism assembly, transmission clearance, and control response.

[0275] The prediction time domain of the warning trigger satisfies formula (22), that is, the product of the number of future prediction frames and the sampling interval is not less than the trajectory prediction time t. pred Collision assessment time t eval and communication and execution delay t comm sum.

[0276] The prediction time-domain formula (22) for early warning triggering is as follows:

[0277]

[0278] In formula (22), Δt is the sampling interval, t pred For the time taken for trajectory prediction, t eval The time required for collision assessment, t comm To delay communication and execution, so as to ensure that the warning signal is output before physical contact occurs.

[0279] Taking the preferred n=15 and Δt=0.4 seconds in this embodiment as an example, the future prediction time domain is 6 seconds, which can cover the time required for trajectory prediction, geometric collision assessment, early warning communication and execution response.

[0280] To ensure that the early warning signal can be output in time before physical contact occurs, the prediction time domain satisfies nΔt≥t. pred +t eval +t comm When the above conditions are met and a collision risk is detected, the system outputs a warning signal to the upper-level control or intervention module, thereby reserving time for subsequent deceleration, pausing, replanning, or other safety measures. Thus, this embodiment achieves online collision risk prediction and early warning for high-density dual-rotation robot fiber optic positioner arrays.

[0281] The trajectory prediction endpoint displacement error of a single robot fiber optic locator is less than 28 micrometers, the forward inference time of the reference single locator is less than 3 milliseconds, the collision prediction accuracy is 98.98%, and the recall rate is 99.85%.

[0282] The trajectory prediction endpoint displacement error refers to the Euclidean distance error between the predicted endpoint and the corresponding actual endpoint, which is the square root of the sum of the squares of the coordinate errors of the final predicted frame in the x and y directions; the smaller the distance between the predicted endpoint coordinates and the actual endpoint coordinates, the smaller the trajectory prediction endpoint displacement error.

[0283] In array-level batch parallel implementation, the total processing latency of the real platform with 65 upgraded robot fiber optic locators is 158.630 milliseconds, of which the spot front-end processing is 152.359 milliseconds, the batch trajectory prediction is 2.872 milliseconds, and the adjacent candidate pair collision screening is 3.399 milliseconds. The total processing latency of the simulation benchmark with 4000 robot fiber optic locators is 563.696 milliseconds, and can be extrapolated to about 1.20 seconds for about 10000 robot fiber optic locators based on the current prototype throughput, all of which are lower than the 6-second prediction time domain.

[0284] It should be noted that 2.6 milliseconds is the baseline forward inference time for a single robot fiber optic locator trajectory prediction model; the total array-level processing latency includes front-end light spot processing, batch trajectory prediction, and collision screening of adjacent candidate pairs, which are implemented in batch parallelism and are not serially accumulated according to the number of robot fiber optic locators.

[0285] Example 2

[0286] The operation steps of this embodiment 2 are the same as those of embodiment 1. The five operation steps—front-end image acquisition, position and motion state extraction, temporal trajectory prediction, geometric collision assessment, and early warning output—are all the same as those in embodiment 1. The difference from embodiment 1 is that this embodiment 2 is used to verify the feasibility and effectiveness of the above five steps on both a simulation platform and a real platform, and provides the input data, mechanism parameters, sample size, and early warning results for different platforms.

[0287] To facilitate differentiation, this embodiment 2, without changing the technical steps S1 to S5 in embodiment 1, only further explains the differences from embodiment 1 in the verification platform, array size, input data, sample composition, security threshold basis, and early warning output results.

[0288] S1: Acquire real-time observation data of the robot fiber optic locator array to be monitored. The basic operation of step (S1) in this embodiment 2 is the same as step (S1) in embodiment 1, that is, acquire the image information of the end fiber optic point of each robot fiber optic locator in the focal plane. Unlike embodiment 1, the real-time observation data in this embodiment 2 includes two types: one is the motion data of the simulation platform containing 4000 robot fiber optic locators, and the other is the focal plane image and motion data collected by the real platform of 65 upgraded robot fiber optic locators; in the real platform, the focal plane image is acquired by an industrial camera and processed by a light spot position recognition hardware programmable chip. A total of 110 target arrival experiments were conducted to verify the online early warning effect under the real hardware link.

[0289] S2: Extract position features, angle features, and motion state features. The basic operation of step (S2) in this embodiment 2 is the same as step (S2) in embodiment 1, namely, adaptive threshold centroid localization of the backlighting points, recovery of the central arm rotation angle and eccentric arm rotation angle based on the geometric parameters of the dual-rotation mechanism, and extraction of optical flow motion state based on the grayscale consistency constraint of the local window in continuous frames. The difference from embodiment 1 is that the backlighting images of the real platform in this embodiment 2 are statistically derived from 399 experimental images, containing a total of 26,334 detection points; the average half-height and full width of the points is 5.81 ± 0.19 pixels, and the average extracted area is 74.65 pixels; the average difference between the centroid method and the two-dimensional Gaussian fitting center is 0.0112 pixels, the root mean square error is 0.0124 pixels, and the 95th percentile difference is 0.0204 pixels; the length of the central arm and eccentric arm of the upgraded robot fiber optic positioner is 3 mm, and the center distance between adjacent robot fiber optic positioners is 10.6 mm.

[0290] S3: Construct a temporal input sequence and predict the future end-effector trajectory. The basic operation of step (S3) in this embodiment 2 is the same as step (S3) in embodiment 1, that is, constructing a temporal input sequence from position features, angle features, motion state features, timestamps, and identification information, and inputting it into the lightweight spatiotemporal trajectory prediction model PLST-Former to obtain the end-effector trajectory in the future prediction time domain. The difference from embodiment 1 is that this embodiment 2 uses the robot fiber optic locator motion sequence obtained under real observation conditions as evaluation data, comparing the lightweight spatiotemporal trajectory prediction model PLST-Former with the standard Transformer baseline; under the setting of predicting the next 15 frames from 15 consecutive historical observations, the trajectory prediction endpoint displacement error of PLST-Former on the test set is 27.09 micrometers, an improvement of 48.19% compared to the standard Transformer, and the baseline single-locator forward inference time for a single robot fiber optic locator is 2.6 milliseconds (i.e., 0.0026 seconds).

[0291] S4: Restore the future mechanism configuration and calculate the geometric gap between adjacent robot fiber optic positioners. The basic operation of step (S4) in this embodiment 2 is the same as step (S4) in embodiment 1, namely, based on the predicted end-effector trajectory, the future mechanism configuration is restored using the inverse kinematics of the double-rotating mechanism, and the distance between the end-effector fiber points, the vertical distance from the end-effector fiber point to the adjacent eccentric arm, and the distance from the end-effector fiber point to the rotation center of the adjacent eccentric shaft are calculated. Unlike embodiment 1, in this embodiment 2, collision assessment is performed simultaneously on both the simulation platform and the real platform; the simulation platform constructs 4000 robot fiber optic positioners, established based on nominal physical parameters of 3 mm for both arms and 10.6 mm for adjacent center distance, and obtains 10,967 positive collision samples (real collision pairs) after 6 rounds of target arrival experiments; the real platform records 677 real collision pairs in 110 target arrival experiments; the preset safety threshold is determined based on the effective mechanical envelope of the positioner, which includes protruding structures such as the eccentric shaft assembly and the eccentric support boss, and a conservative safety margin is introduced to cover prediction uncertainties and execution tolerances.

[0292] S5: Output a warning signal based on a preset safety threshold. The basic operation of step (S5) in this embodiment 2 is the same as step (S5) in embodiment 1, that is, when the minimum geometric gap between any adjacent robot fiber optic positioners is less than the preset safety threshold, a collision risk is determined and a warning signal is output. Unlike embodiment 1, this embodiment 2 further provides the warning output results: In simulation testing, the collision prediction accuracy of the method of this invention is 99.68%, and the recall rate is 99.76%; in real platform testing, the collision prediction accuracy is 98.98%, and the recall rate is 99.85%; the real platform generated 7 false positive warnings (false alarms) and 228,116 pairs of true negative collision-free samples, corresponding to an adjacent pair level false alarm rate of 3.07 × 10⁻⁵, or 0.0031%; in 110 reconstructions, the average number of false alarms per reconstruction was 0.064, of which 103 reconstructions had no false positive warnings, 7 reconstructions each had 1 false positive warning, and no reconstruction had more than 1 false positive warning.

[0293] When implemented in array-level batch parallelism, the total processing latency of the real platform with 65 upgraded robot fiber optic locators is 158.630 milliseconds, including 152.359 milliseconds for spot front-end processing, 2.872 milliseconds for batch trajectory prediction, and 3.399 milliseconds for adjacent candidate pair collision screening. The total processing latency of the simulation benchmark with 4,000 robot fiber optic locators is 563.696 milliseconds, which can be extrapolated to approximately 1.20 seconds for about 10,000 robot fiber optic locators based on the current prototype throughput, both of which are lower than the 6-second prediction time domain.

[0294] It should be noted that 2.6 milliseconds is the baseline forward inference time for a single robot fiber optic locator trajectory prediction model; the total array-level processing latency includes front-end light spot processing, batch trajectory prediction, and collision screening of adjacent candidate pairs, which are implemented in batch parallelism and are not serially accumulated according to the number of robot fiber optic locators.

[0295] The above results show that, based on the same five operation steps as in Example 1, this Example 2 completes the verification through different data inputs from the simulation platform and the real platform, and proves that the image position extraction, angle recovery, motion state recognition, temporal trajectory prediction, geometric gap calculation and early warning triggering conditions defined by formulas (1) to (22) can jointly form an implementable online collision prediction and early warning process.

[0296] It should also be noted that the experimental data in Example 2 are only used to illustrate the feasibility and effectiveness of the method of the present invention, and do not constitute a limitation on the scope of protection of the present invention. For robot fiber optic positioner arrays of different sizes, quantities, sampling rates, or mechanism parameters, those skilled in the art can make equivalent adjustments to the historical window length, prediction window length, safety margin, state division threshold, and lightweight model structure without departing from the concept of the present invention.

Claims

1. A real-time collision prediction and early warning method for an astronomical robot fiber optic locator, wherein the real-time collision prediction and early warning method is applied to a multi-object fiber optic spectroscopic astronomical telescope, wherein the multi-object fiber optic spectroscopic astronomical telescope refers to an astronomical observation device that simultaneously collects the spectra of more than one hundred celestial targets through more than one hundred optical fibers. The multi-object fiber optic spectroscopic telescope includes a robotic fiber optic locator array consisting of more than one hundred robotic fiber optic locators distributed on the focal plane and an online monitoring system. The robot fiber optic positioner includes a central axis (1), a central arm, an eccentric axis (2), an eccentric arm, and an end fiber optic cable; The center shaft (1) is located in the center of the base of the robot fiber positioner, one end of the center arm is coaxially connected with the center shaft (1), the other end of the center arm is connected with one end of the eccentric arm through the eccentric shaft (2), and the other end of the eccentric arm is provided with the terminal optical fiber; the center arm and the eccentric arm are sequentially connected along the focal plane to form a two-stage series rotation structure, thereby constituting a double-rotation execution mechanism; the center arm rotation angle θ i (t) the included angle formed by taking the center axis (1) as the rotation center and rotating in the counterclockwise direction when viewed downward from the focal plane; the eccentric arm rotation angle φ i (t) the included angle formed by taking the eccentric shaft (2) as the rotation center, taking the extension direction of the center arm from the center shaft (1) to the eccentric shaft (2) as the reference direction, and rotating in the counterclockwise direction when viewed downward from the focal plane. The end of each optical fiber is fixed to the end of the eccentric arm of a corresponding robot fiber optic positioner. The operational safety framework of the multi-object fiber optic spectroscopic telescope comprises three layers from top to bottom. This operational safety framework is a hierarchical safety control system established around observation mission planning, systematic error suppression, and online monitoring of the execution process. The online monitoring system is located on the third layer, the real-time monitoring layer. The first layer of the operational safety framework is the planning phase protection layer, namely the target allocation and motion planning layer during the mission planning phase, which is used to complete target allocation, motion planning, and static geometric feasibility checks before observation execution; the second layer is the dynamic correction layer, namely the upstream error suppression and stability control layer, which is used to suppress positioning deviations caused by upstream disturbances such as atmospheric refraction correction, wind field, thermal deformation, and active optical adjustment; the third layer is the real-time monitoring layer, which is used to predict future short-term trajectories based on real-time images and output collision warnings during execution. The online monitoring system includes a focal plane measurement camera, an image acquisition module, a light spot position recognition hardware programmable chip (FPGA), a trajectory prediction and processing module, a geometric collision evaluation module, and an early warning output module; The focal plane measurement camera is located above the focal plane and is used to image the back-illuminated fiber optic point on the focal plane. The function of the robot fiber optic locator is to move the end fiber to the target star image position, and the center of the back illumination spot formed by the end fiber in the focal plane image is the end fiber point. The target star image position refers to the target coordinate position on the focal plane after the celestial body to be observed is imaged by the telescope optical system. The robot fiber optic locator moves the end fiber optic cable to this position to receive the light signal of the celestial body. The real-time collision prediction and early warning method is characterized by the following steps: S1: Acquire real-time observation data of the robot fiber optic positioner array to be monitored, wherein the real-time observation data includes at least the image information of the end fiber optic point of each robot fiber optic positioner in the focal plane. S2: Extract the end position of each robot fiber optic locator based on the image information. The end position is the end fiber optic point, which serves as a position feature. The angle features are obtained by solving the corresponding central arm rotation angle and eccentric arm rotation angle by combining the position features with the mechanism geometric parameters; Based on the grayscale distribution change of the end fiber optic points between frames of continuous images in the image information, an image motion estimation method based on local window grayscale consistency constraints is used to obtain the image planar motion amount. Based on the amplitude of the image planar motion amount and the preset state, a threshold is divided, and a static state or a preset motion stage is determined to obtain the motion state characteristics of each robot fiber optic positioner. The local window grayscale consistency constraint refers to the constraint that the grayscale distribution of the same end fiber point is approximately consistent in the local neighborhood of adjacent image frames. The motion state features refer to the state labels used to distinguish whether the robot fiber optic positioner is stationary or in different stages of motion. The geometric parameters of the mechanism include the rotation center coordinates of the i-th robot fiber optic positioner, the length of the central arm, the length of the eccentric arm, the rotation angle range of the central arm, the rotation angle range of the eccentric arm, the adjacency relationship between adjacent robot fiber optic positioners, and the center distance between adjacent robot fiber optic positioners. The adjacency relationship between adjacent robot fiber optic positioners is such that the center distance within the focal plane is less than the sum of the inspection radii of the two robot fiber optic positioners, and there is an overlapping adjacent arrangement in the workspace, preferably a six-neighbor adjacency relationship. The adjacency relationship refers to the geometric adjacency relationship formed when the center distance between two robot fiber optic positioners is no greater than the sum of their inspection radii in a fixed arrangement on the focal plane, and their workspaces overlap. The six-neighbor adjacency relationship refers to the fact that in an approximately hexagonal close arrangement, a robot fiber optic locator and its six closest robot fiber optic locators that may overlap in the workspace form an adjacent candidate pair, and subsequent collision evaluation is only performed between the adjacent candidate pairs. S3: The position features, angle features, motion state features, timestamps and identification information of each robot fiber optic locator at continuous historical moments are used to form a time-series input sequence, and input into a lightweight spatiotemporal trajectory prediction model to obtain the end trajectory of each robot fiber optic locator in the future prediction time domain. The temporal input sequence refers to a feature set of multiple frames of images arranged in chronological order. The lightweight spatiotemporal trajectory prediction model refers to a model with fewer than 1.50M parameters, used to predict future terminal positions based on historical positions and states. The lightweight spatiotemporal trajectory prediction model includes a physically guided feature extraction branch, a local temporal feature extraction branch, and a global temporal dependency modeling branch. The physical guidance feature extraction branch is a computational branch used to introduce the geometric and kinematic relationships of the double rotary mechanism. The physical guidance feature extraction branch constructs physical features based on the relationship between the length of the central arm, the length of the eccentric arm, the rotation angle of the central arm, the rotation angle of the eccentric arm, and the partial derivative relationship between the end position and the rotation angles of the central arm and the eccentric arm, so that the prediction model retains the geometric constraints of the double rotary mechanism while learning the changes in historical trajectory. The partial derivative relationship between the end position and the rotation angles of the central arm and the eccentric arm is a local Jacobian relationship that causes the end position to change when the rotation angles of the central arm and the eccentric arm change slightly. The local temporal feature extraction branch is a computational branch used to extract local motion changes within a short period of time. It uses temporal convolution operations to extract short-term local motion patterns. Temporal convolution refers to performing convolution operations on the feature set of consecutive frame images along the time dimension. The global temporal dependency modeling branch is a computational branch used to extract the overall motion trend within a longer historical window. It uses an attention mechanism to extract long-term dependency features. The attention mechanism refers to a computational method that assigns weights based on the correlation between different temporal features. After fusing the position features, angle features, and motion state features, the encoder and lightweight decoder respectively output the predicted trajectories of all robot fiber optic positioners in the future prediction time domain. The fusion refers to concatenating position features, angle features, and motion state features according to their correspondence at the same time into a unified feature vector, and then converting them into a latent feature representation of the same dimension through a linear mapping or encoding network; The encoder is a network structure that converts input feature vectors into latent feature representations. The lightweight decoder is an output network structure that converts latent features into future terminal positions with a small number of parameters. The input feature vector at each moment includes the robot fiber optic locator identifier, timestamp, motion state label, end-effector position coordinates, central arm rotation angle, and eccentric arm rotation angle; The input feature vector at each moment is defined as a set of model input data obtained by combining the identification information, timestamp, motion state label, end position coordinates, center arm rotation angle and eccentric arm rotation angle of the same robot fiber optic locator in a fixed order at a single sampling moment; the input feature vectors of m consecutive frames are stacked in time order to form a temporal input sequence, which is used as the common input of the three branches of the lightweight spatiotemporal trajectory prediction model; S4: Based on the predicted trajectory of each fiber optic positioning robot in the future time domain, combine the inverse kinematics of the double rotary mechanism to recover the future mechanism configuration of each fiber optic positioning robot and the fiber optic positioner of the adjacent robot, and calculate the geometric gap between each fiber optic positioning robot and the fiber optic positioner of the adjacent robot. The inverse kinematics refers to the geometric calculation process of inversely determining the rotation angles of the central arm and the eccentric arm from the end position; The geometric gap refers to the distance between adjacent robot fiber optic positioners used to determine whether they are too close. The future mechanism configuration includes the central axis rotation center, eccentric axis rotation center, end fiber point position, central arm position, and eccentric arm position of the adjacent robot fiber optic positioner at the future predicted time. The rotation center of the eccentric shaft is determined by the rotation center of the central shaft, the length of the central arm, and the rotation angle of the central arm obtained by inverse kinematics recovery. The location of the terminal fiber optic point is determined by the terminal trajectory in the future prediction time domain; S5: When the geometric gap between any adjacent robot fiber optic positioners is less than the preset safety threshold, it is determined that there is a collision risk and an early warning signal is output. The preset security threshold is determined by the physical dimensions of the optical fiber structure and the security margin. The safety margin is used to cover trajectory prediction error, image measurement error, and performance tolerance; The trajectory prediction error refers to the deviation between the predicted end position and the actual end position. The image measurement error refers to the deviation in end-position measurement caused by camera imaging, light spot extraction, and centroid positioning. The execution tolerance refers to the deviation between the actual motion and the commanded motion caused by differences in motor drive, mechanism assembly, transmission clearance, and control response. The trajectory prediction endpoint displacement error of a single robot fiber optic locator is less than 28 micrometers, the forward inference time of the reference single locator is less than 3 milliseconds, the collision prediction accuracy is 98.98%, and the recall rate is 99.85%. The trajectory prediction endpoint displacement error refers to the Euclidean distance error between the predicted endpoint and the corresponding actual endpoint, which is the square root of the sum of the squares of the coordinate errors of the final predicted frame in the x and y directions; wherein, the smaller the distance between the coordinates of the predicted endpoint and the coordinates of the actual endpoint, the smaller the trajectory prediction endpoint displacement error. When implemented in array-level batch parallelism, the total processing latency is less than 6 seconds in the prediction time domain. The array-level batch parallel implementation refers to using more than one hundred robot fiber optic positioners or more than one hundred adjacent candidate pairs as the same batch of inputs to complete trajectory prediction and collision assessment in parallel.

2. The real-time collision prediction and early warning method for an astronomical robot fiber optic locator according to claim 1, characterized in that: The focal plane measurement camera of the online monitoring system is used to capture focal plane images. The image acquisition module is used to receive and transmit focal plane images. The light spot position recognition hardware programmable chip (FPGA) is a programmable hardware chip used to process the position and motion information of optical fiber points in the image in parallel. The trajectory prediction processing module is used to output the future end trajectory. The geometric collision evaluation module is used to calculate the spatial proximity between adjacent robot optical fiber positioners. The early warning output module is used to output collision risk warnings to the upper control system.

3. The real-time collision prediction and early warning method for an astronomical robot fiber optic locator according to claim 1, characterized in that: In step S2, the extraction of the end position of each robot fiber optic locator adopts the adaptive threshold T centroid localization method based on global statistical features. The global statistical features refer to the mean and standard deviation of the overall gray level of a focal plane image frame. The centroid localization method refers to the method of calculating the center coordinates of the end fiber optic point using the pixel gray level of the end fiber optic point as the weight. The adaptive threshold T formula (1) is as follows: In formula (1), μ is the mean gray value of the image, σ is the standard deviation of the gray value of the image, and α is the threshold coefficient, preferably 3; According to formula (1), pixels with gray values ​​greater than the adaptive threshold T are selected as candidate pixels for the end fiber point, and binary threshold segmentation is performed. The binary threshold segmentation refers to an image segmentation method that divides pixels that meet the threshold condition into the end fiber point region and the remaining pixels into the background region. After threshold segmentation, the center coordinates of the end fiber point are obtained by connecting component labeling and centroid calculation. The connecting component labeling refers to a processing method that merges adjacent candidate fiber point pixels into the same pixel region. The formula (2) for the center coordinates of the end fiber point is as follows: In formula (2), R is the pixel region corresponding to the end fiber point, I(q) is the gray value of pixel q, x q and y q are the coordinates of pixel q in the focal plane image, and x c and y c are the center coordinates of the end fiber point.

4. The real-time collision prediction and early warning method according to claim 1, characterized in that: In step S2, the center arm rotation angle θ is solved according to the end position of each optical fiber robot positioner i (t) and the eccentric arm rotation angle φ i (t) that satisfy the forward kinematics relationship of the double rotation mechanism, the forward kinematics relationship being a geometric relationship for calculating the end position from the center arm rotation angle and the eccentric arm rotation angle, and the forward kinematics relationship formula (3) being as follows: In formula (3), p i t =(x i t ,y i t Let be the coordinates of the end position of the i-th robot fiber optic positioner at time t, and c be the coordinates of the end position of the i-th robot fiber optic positioner at time t. i =(x i 0 ,y i 0 Let l1 and l2 be the coordinates of the rotation center of the central axis of the i-th robot fiber optic positioner, and l1 and l2 be the lengths of the central arm and the eccentric arm, respectively; and the corresponding rotation angles of the central arm and the eccentric arm are recovered from the end position using inverse kinematics; The Euclidean distance r between the end fiber point of the i-th robot fiber optic positioner and its central axis (1) rotation center in the focal plane is the radial distance r. i t The radial distance formula (4) is as follows: In formula (4), r i t Let x be the radial distance from the end position of the i-th robot fiber optic positioner to the center of its central axis rotation. i t and y i t x represents the coordinates of the end position. i 0 and y i 0 The coordinates of the rotation center are the central axis. Center arm rotation angle θ i t and eccentric arm rotation angle φ i t The calculation formula (5) is as follows: In formula (5), θ i t Let φ be the rotation angle of the central arm of the i-th robot fiber optic positioner at time t. i t Let r be the eccentric arm rotation angle of the i-th robot fiber optic positioner at time t, l1 and l2 be the lengths of the central arm and eccentric arm, respectively, and r be the eccentric arm rotation angle. i t The radial distance is obtained from formula (4).

5. The real-time collision prediction and early warning method according to claim 1, characterized in that: In step S2, the motion state features are obtained by an image motion estimation method based on local window grayscale consistency constraints. The image motion estimation method is used to estimate the image plane motion vector based on the local grayscale changes of the end fiber points in consecutive image frames. Under the condition of constant brightness, the optical flow constraint is satisfied. The optical flow refers to the apparent motion of the grayscale pattern of the fiber points in the image between adjacent frames. The optical flow constraint formula (6) is as follows: In formula (6), I x I y and I t Let represent the gradients of the image grayscale in the x, y, and time directions, respectively, and u and v represent the image planar motion vectors v=(u,v). T The components in the x and y directions; based on the estimated optical flow amplitude, the robot fiber optic positioner is divided into a stationary state or multiple preset motion stages, and the optical flow amplitude and the discrete motion state label satisfy the following formula (7): In formula (7), M is the optical flow amplitude and S is the discrete motion state label; preferably, the discrete motion state label S takes the values ​​of 0, 1, 2, and 3, where 0 represents the stationary state, 1 represents the eccentric arm retraction stage, 2 represents the central arm alignment stage, and 3 represents the eccentric arm alignment stage.

6. The real-time collision prediction and early warning method according to claim 1, characterized in that: In step S3, the input feature vector formula (8) at each time step is as follows: In formula (8), f i t Let ID be the input feature vector at time t. i This is the identifier for the robot's fiber optic locator, where t is the timestamp and S is the time stamp. i t For motion status labels, x i t and y i t For the end position, θ i t and φ i t These are the center arm rotation angle and the eccentric arm rotation angle, respectively. The temporal input sequence consists of m consecutive historical observations; the m consecutive historical observations refer to the m consecutive images of the same robot fiber optic locator obtained at the sampling time before the current prediction time, and their corresponding position, angle and motion state features. The lightweight spatiotemporal trajectory prediction model outputs the predicted end trajectory for the next n frames, and the predicted end trajectory formula (9) is as follows: In formula (9), X i Let p be the temporal input sequence formed by the i-th robot fiber optic positioner within m consecutive historical observation frames. i t+k To predict the end position in the future k-th frame, Θ represents the learnable model parameters, and F... Θ is the mapping function corresponding to the lightweight spatiotemporal trajectory prediction model, preferably m equals n, and both are 15.

7. The real-time collision prediction and early warning method according to claim 1, characterized in that: In step S3, the physical guidance feature extraction branch, local temporal feature extraction branch, and global temporal dependency modeling branch of the lightweight spatiotemporal trajectory prediction model are described as follows: The physical guidance feature extraction branch constructs physical features based on the local Jacobian information of the forward kinematics of the double rotary mechanism. The local Jacobian information refers to the partial derivative relationship between the end position and the rotation angles of the central arm and the eccentric arm. The Jacobian component formula (10) is as follows: In formula (10), x and y are the abscissa and ordinate of the end fiber point in the focal plane, respectively; θ is the rotation angle of the central arm; φ is the rotation angle of the eccentric arm; l1 is the length of the central arm; l2 is the length of the eccentric arm; ∂x / ∂θ, ∂x / ∂φ, ∂y / ∂θ and ∂y / ∂φ are the partial derivatives of the end coordinates with respect to the rotation angles of the central arm and the eccentric arm, respectively, which are used to represent the local sensitivity of the end position to the changes in the two-stage rotation angles; The four Jacobian components in formula (10) are stacked into a physical feature tensor J, where tensor J is an array organized in multiple dimensions, and physical guidance features are obtained through linear mapping; where Gaussian error linear unit activation function (GELU) is an activation function used to introduce nonlinear transformation, and the physical guidance feature formula (11) is as follows: In formula (11), F p Physical guidance features; X p The input is the feature to the physical guided feature extraction branch; J is the physical feature tensor formed by stacking the four Jacobian components of formula (10); W p and W J For linear mapping weights; b p and b J Here are the bias parameters; GELU is the activation function of the Gaussian error linear unit. The local temporal feature extraction branch uses depthwise separable temporal convolution to extract short-term local motion patterns. The depthwise separable temporal convolution refers to the operation of first performing temporal convolution by channel and then reconstructing features through pointwise convolution. Among them, random deactivation regularization (Dropout) refers to the operation of randomly masking some features during training to reduce the risk of overfitting. The linear rectified activation function (ReLU) refers to the activation function that sets the negative input to zero and retains the positive input. The local temporal feature formula (12) is as follows: In formula (12), F l X represents local temporal features; K represents the temporal input feature tensor. d K is a depthwise convolution kernel that performs in-channel convolution along the time direction; p is the pointwise convolution kernel used to reconstruct channel features; * indicates convolution operation; ReLU is the linear rectified activation function; Dropout is the random deactivation regularization operation; The global temporal dependency modeling branch uses a multi-head attention mechanism to extract long-term dependency features. The multi-head attention mechanism refers to the operation of mapping query Q, key K, and value V to multiple subspaces and performing weighted summation. Query Q, key K, and value V are the features to be matched, matching index features, and value features in the attention calculation, respectively. Among them, tensor concatenation operation (Concat) refers to connecting multiple feature tensors according to a specified dimension, and the normalized exponential function (Softmax) refers to the function that converts real number vectors into normalized weights. The global temporal dependency feature formulas (13) and (14) are as follows: In formula (13), F g The global time-series dependency features; Q, K, and V are the query feature, key feature, and value feature, respectively; head1 to head h For the output of h attention heads; W O These are the mapping parameters for the multi-head attention output; Concat is the operation that concatenates the outputs of multiple attention heads according to the feature dimension. In formula (14), head a W is the output of the a-th attention head; a Q W a K and W a V These are the query mapping matrix, key mapping matrix, and value mapping matrix for the a-th attention head, respectively; d k is the dimension of the key vector; Softmax is the normalization exponential function; the superscript T indicates matrix transpose; The physical guidance feature extraction branch, the local temporal feature extraction branch, and the global temporal dependency modeling branch are concatenated and fused through linear projection. The fused feature formula (15) is as follows: In formula (15), F represents the fusion feature, Ffusion ... g For global temporal dependency features, F l For local temporal features, F p For physical guidance features, [·] represents feature concatenation operation, W f For linear projection parameters, R d This represents the d-dimensional feature space where the fused features reside.

8. The real-time collision prediction and early warning method according to claim 1, characterized in that: In step S4, the geometric gap includes at least one of the following: the distance between the end fiber points of two adjacent robot fiber optic positioners, the vertical distance from the end fiber point of one robot fiber optic positioner to the eccentric arm of the adjacent robot fiber optic positioner, and the distance from the end fiber point of one robot fiber optic positioner to the rotation center of the eccentric arm of the adjacent robot fiber optic positioner; wherein, the position of the end fiber point of the i-th robot fiber optic positioner at the predicted future time is p. i The center of rotation of the eccentric shaft is e. i The eccentric arm direction vector of the adjacent robot fiber optic positioner j is d. j The eccentric arm direction vector d j It refers to the vector pointing from the center of rotation of the eccentric axis to the end fiber optic point; the Euclidean norm refers to the vector length; and the cross product is the vector product operation used to calculate the distance from a point to a line. The geometric clearance formulas (16) to (18) are as follows: In formula (16), D ee ij p is the distance between the two end fiber optic points of adjacent robot fiber optic positioners i and j; i and p j Let be the positions of the end fiber points of the i-th and j-th robot fiber optic positioners at the predicted future time, respectively; ||·|| is the Euclidean norm, used to represent the Euclidean distance between the two points; In formula (17), D el ij p is the vertical distance from the end fiber optic point of a robot fiber optic positioner to the eccentric arm of an adjacent robot fiber optic positioner. i Let e ​​be the position of the end fiber point of the i-th robot fiber optic positioner at the predicted future time; j Let d be the position of the rotation center of the eccentric axis of the j-th robot fiber optic positioner; j Let be the eccentric arm direction vector of the j-th robot fiber optic positioner, and d j =p j -e j × represents the cross product operation; In formula (18), D ec ij p is the distance from the end fiber optic point of a robot fiber optic positioner to the eccentric axis rotation center of the adjacent robot fiber optic positioner. i Let e ​​be the position of the end fiber point of the i-th robot fiber optic positioner at the predicted future time; j Let be the position of the eccentric axis rotation center of the j-th robot fiber optic positioner at the predicted future time; ||·|| is the Euclidean norm; The formula for the minimum geometric gap between adjacent robot fiber optic positioners (19) is as follows: In formula (19), D min ij is the minimum geometric gap between adjacent robot fiber optic positioners i and j; min represents the minimum value among the distance between end points, the distance from the point to the eccentric arm, and the distance from the point to the rotation center of the eccentric shaft. The collision risk determination formula (20) for adjacent robot fiber optic positioners posing a collision risk is as follows: In formula (20), C is the set of adjacent robot fiber optic locators that are determined to have a collision risk; D safe This is a preset safety threshold; when D min ij Less than D safe When the collision risk is determined, add the adjacent pair (i,j) to set C and output the collision risk judgment. The preset security threshold D safe Formula (21) is as follows: In formula (21), D safe For preset safety thresholds; D fiber Δ represents the physical dimensions of the fiber optic structure or the effective mechanical envelope dimensions; Δ is the safety margin used to cover trajectory prediction errors, image measurement errors, and performance tolerances.

9. The real-time collision prediction and early warning method according to claim 1, characterized in that: In step S5, the prediction time-domain formula (22) for early warning triggering is as follows: In formula (22), Δt is the sampling interval, t pred For the time taken for trajectory prediction, t eval The time required for collision assessment, t comm To delay communication and execution, so as to ensure that the warning signal is output before physical contact occurs.

Images