A Deep Learning-Based Method for Position Recognition of Automotive Welded Components

Abstract

This invention provides a deep learning-based method for position recognition of automotive welded parts, relating to the interdisciplinary fields of image recognition and optical measurement. Addressing the challenges of topological distortion and concurrent computational resource congestion caused by high-frequency, high-intensity light, this method obtains the area ratio and intersection-union ratio (IU / U) parameters of the target contour enclosing matrix between the current and historical frames, generates anomaly markers for morphological abrupt changes, and simultaneously extracts the feature matching oscillation frequency of the underlying inference unit. These parameters are then input into a collaborative scheduling and control module to perform cross-dimensional weighted calculations, generating a concurrent congestion assessment index. In response to this index breaking through the concurrent circuit breaker threshold, a pose control signal is generated and bypassed, driving the inference unit to suspend feature matching and triggering the external servo feed enable cutoff. This achieves deep decoupling between spatiotemporal distortion perception and underlying computational resource scheduling, eliminates computational queue congestion, and outputs highly robust target pose control signals.

Description

Technical Field

[0001] This invention relates to the field of image recognition and optical measurement technology, specifically a method for position recognition of automotive welded parts based on deep learning. Background Technology

[0002] With the deepening evolution of intelligent manufacturing, deep learning-based industrial vision recognition technology has been widely applied in automated production lines. In precision manufacturing, vision recognition not only needs to complete basic feature extraction, but also needs to combine optical measurement and position contour measurement rules to provide downstream actuators with continuous and high-precision spatial coordinate references. Under complex and dynamically changing physical conditions, ensuring the output stability of the visual feature calculation network under extreme environmental disturbances has become a key factor restricting the accuracy of overall control.

[0003] In on-site image processing terminals applied to automotive welding production lines, existing time-weighted smoothing techniques face significant challenges. For example, patent document CN121095194A proposes a deep learning-based intelligent weld quality diagnosis system, which demonstrates effectiveness in suppressing pseudo-alignment errors and multimodal fusion through a deformable convolution feature alignment mechanism with confidence factors.

[0004] However, the aforementioned and similar existing technologies still have limitations in dealing with the high-frequency, strong light and alternating sparks interference unique to automotive welding production lines: traditional static filtering and alignment mechanisms are difficult to handle transient, extreme local high-brightness pixel noise, which not only causes drastic shifts in single-frame feature coordinates but also leads to distorted expansion of the target bounding box in spatial topology; existing processing logic is usually divided into independent and uncoordinated stages, and the severe distortion pressure missed by the front end accumulates on the back end, causing the underlying state machine to frequently jump between "accepting the current frame" and "calling historical frames". This timing misalignment causes a large amount of memory read / write and feature parsing computation resources to be consumed on "pseudo-targets" for a long time, which in turn leads to the blocking of underlying concurrent scheduling, making it impossible to issue critical pose control signals within a very short operation time window. Summary of the Invention

[0005] The purpose of this invention is to provide a deep learning-based method for identifying the position of automotive welded components. This method extracts spatial topological distortion features (target contour matrix deformation) output by the visual feature computation network across dimensions, along with the temporal computation severity (feature matching oscillation frequency) of the underlying inference unit. Furthermore, it introduces the system's concurrent backlog queue depth as a load-sensing hub, constructing a closed-loop collaborative control center from the image domain to the system scheduling domain. This addresses the problems mentioned in the background section.

[0006] To achieve the above objectives, the present invention provides the following technical solution: A deep learning-based method for position recognition of automotive welded parts, comprising the following steps: Step S1: Obtain the target contour enclosing matrix that characterizes the light spot interference in the automotive welding production line, extract the area ratio and intersection-union ratio parameter of the target contour enclosing matrix of the current frame and the previous historical frame whose prediction confidence exceeded the set judgment threshold, so as to generate a morphological change anomaly identifier; and extract the feature matching oscillation frequency inside the bottom inference unit. Step S2: Input the morphological mutation anomaly identifier and the feature matching oscillation frequency into the collaborative scheduling and control module. Based on the spatial topological distortion features represented by the morphological mutation anomaly identifier and the computational blocking features represented by the feature matching oscillation frequency, perform a load-aware nonlinear penalty alignment operation to generate a concurrent blocking evaluation index for characterizing the abnormal load state of the system. Step S3: In response to determining that the concurrent blocking evaluation index breaks through the preset concurrent circuit breaker threshold, generate pose control signaling for scheduling the underlying inference unit and configuring the external automotive welding robot arm servo controller; Step S4: In response to the pose control signal, drive the underlying inference unit to perform the suspension action of the feature matching calculation process, and simultaneously output the pose control signal to the external automotive welding robot arm servo controller to trigger the external feed enable cut-off action, thereby realizing the recovery of host computing resources and the physical position locking of the external actuator.

[0007] Compared with the prior art, the beneficial effects of the present invention are: This method extracts the area ratio and intersection-union ratio (IU / U) parameters of the target contour enclosing matrix between the current and historical frames to generate morphological abrupt change anomaly markers. Simultaneously, it extracts the feature matching oscillation frequency of state machine flips in the underlying inference unit. In the collaborative scheduling and control module, it performs load-aware nonlinear penalty alignment operations to generate a concurrent blocking evaluation index. It can deeply fuse spatial topological distortions caused by the physical environment with logical jitter caused by the underlying algorithm getting stuck in "pseudo-target" matching, quantifying the visual solution network's resistance to extreme noise. This effectively avoids misjudgments caused by relying solely on single-frame confidence.

[0008] This method further extracts the centroid wandering deviation of the target contour enclosing matrix within the sliding time window, generates a dynamic compensation factor through normalized mapping, and uses this factor to perform nonlinear boundary compensation operations on the preset upper limit threshold for area expansion and the lower limit threshold for spatial coincidence. It can separate legitimate perspective distortion caused by normal physical movement of the robotic arm from pseudo-topological expansion caused by light spots, enabling the judgment boundary envelope to adapt to the kinematic evolution of the external actuator. This enhances the robustness of the position recognition algorithm in continuous spatial movement.

[0009] This method generates a penalty gain coefficient by performing a nonlinear exponential mapping based on the concurrent backlog queue depth and the maximum queue tolerance threshold. When the dynamically calculated concurrent blocking evaluation index breaks through the circuit breaker threshold, the inference unit is suspended from computation, and a pose control signal is bypassed to cut off the external servo feed enable. By introducing a resource load duty cycle, a surge in nonlinear penalty weights is triggered before the system's concurrent resource / processing load capacity approaches its exhaustion limit / physical critical point, instantly halting worthless deep convolution calculations and achieving synchronous emergency blocking of the host-side computation queue and the external servo axial displacement. This enables rapid recovery of host computing resources and secure physical locking of external actuators, ensuring deterministic business response within a very short time window. Attached Figure Description

[0010] Figure 1 This is a topology diagram of a deep learning-based method for identifying the position of automotive welded parts according to the present invention.

[0011] Figure 2 This is a schematic diagram illustrating the technical logic of the overall solution of the present invention. Detailed Implementation

[0012] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention.

[0013] It is understood that the terms “first,” “second,” etc., used in this application may be used herein to describe various elements, but unless otherwise stated, these elements are not limited by these terms. These terms are used only to distinguish one element from another.

[0014] Example 1: Please see Figure 1 and Figure 2 The present invention provides a technical solution: A deep learning-based method for position recognition of automotive welded parts includes: Step S1: Obtain the target contour enclosing matrix that characterizes the light spot interference in the automotive welding production line, extract the area ratio and intersection-union ratio parameter of the target contour enclosing matrix of the current frame and the previous historical frame whose prediction confidence exceeded the set judgment threshold, so as to generate a morphological change anomaly identifier; and extract the feature matching oscillation frequency inside the bottom inference unit. In this embodiment, the prediction confidence refers to the objective probability value output by the following visual feature calculation network at the end of the inference forward propagation process, which represents whether a pixel region or candidate bounding box belongs to a real welded component.

[0015] Prediction confidence is a specific tensor element in the two-dimensional coordinate probability set output by the underlying deep learning model. Its generation mechanism is a normalized value obtained after non-linear mapping through the activation function (Sigmoid or Softmax layer) of the network output layer. Its absolute value range is limited to the closed interval [0,1].

[0016] The higher the prediction confidence value, the higher the mathematical certainty that the extracted target contour enclosing matrix is ​​not contaminated by high-frequency splash noise and belongs to a real physical target; the aforementioned judgment threshold actually corresponds to the lower limit parameter of the bottom model filter used by the output layer of the visual feature calculation network to distinguish the normalized interval of the effective target and the background noise, which is excerpted from the following paragraph "Obtaining the filter threshold set to represent the lower limit of confidence by pre-experimental calibration"; Step S2: Input the morphological mutation anomaly identifier and the feature matching oscillation frequency into the collaborative scheduling and control module. Based on the spatial topological distortion features represented by the morphological mutation anomaly identifier and the computational blocking features represented by the feature matching oscillation frequency, perform a load-aware nonlinear penalty alignment operation to generate a concurrent blocking evaluation index for characterizing the abnormal load state of the system. Step S3: In response to determining that the concurrent blocking evaluation index breaks through the preset concurrent circuit breaker threshold, generate pose control signaling for scheduling the underlying inference unit and configuring the external automotive welding robot arm servo controller; Step S4: In response to the pose control signaling, the event dispatcher inside the collaborative scheduling and management module drives the underlying inference unit to perform the suspension action of the feature matching calculation process, and simultaneously outputs the pose control signaling to the external automotive welding robot arm servo controller via the industrial network communication stack bypass to trigger the external feed enable cutting action, thereby realizing the recovery of host computing resources and the physical position locking of the external actuator.

[0017] The steps of obtaining the target contour enclosing matrix and generating the morphological abrupt change anomaly identifier specifically include: The underlying inference unit of the on-site image processing terminal extracts the two-dimensional coordinate probability set that represents the predicted position of the welded component from the output of the pre-trained visual feature solving network, and then parsely maps it into a target contour enclosing matrix. Calculate the real-time area ratio (hereinafter referred to as the real-time ratio) of the area of ​​the target contour bounding matrix and the area of ​​the target contour bounding matrix of the historical frame whose prediction confidence exceeded the set judgment threshold, and calculate the intersection-union ratio parameter of the overlapping area of ​​the two spatial coordinates. In response to the detection of at least one of the following distortion triggering conditions: the real-time area ratio is greater than a preset upper limit threshold for area expansion, and the intersection-union ratio parameter is less than a preset lower limit threshold for spatial overlap; Generate the morphological abrupt change anomaly identifier representing spatial topological distortion. The step of extracting feature matching oscillation frequencies specifically includes: polling and listening to the visual feature prediction callback record representing the prediction completion status notification / asynchronous execution result response in the business operation context of the underlying inference unit; The cumulative number of flips that occur between the feature output state of the visual feature calculation network in the current frame and the state of calling the historical sliding window queue to perform coordinate weighted compensation within a continuously set underlying system clock cycle; The accumulated number of flips is used as the feature matching oscillation frequency to characterize the severity of time-series calculation.

[0018] The steps for calculating the concurrent blocking evaluation index specifically include: setting a sliding observation window in the observation data circular queue of the collaborative scheduling and control module; extracting the first cumulative number of times the morphological mutation anomaly flag is triggered as true within the sliding observation window; and extracting the cumulative sum of the feature matching oscillation frequencies within the sliding observation window during the same period. The first cumulative occurrence count is multiplied by a preset spatial distortion penalty factor, and the cumulative sum is multiplied by a preset temporal severity penalty factor. The two product results are then summed to generate the concurrent blocking evaluation index.

[0019] The pose control signaling is configured to be synchronously triggered: updating the operation control status word for the underlying inference unit to the service suspension status; and encapsulating and generating the pose control signaling carrying the priority forwarding status word in the independent emergency communication queue of the industrial network communication stack.

[0020] Before detecting the distortion triggering condition, the steps of obtaining the target contour enclosing matrix and generating the morphological abrupt change anomaly identifier further include: extracting the centroid wandering deviation of the target contour enclosing matrix of the historical frame in which the previous prediction confidence exceeded the set judgment threshold within a set sliding window. Based on the preset normalization mapping rule, the centroid wandering deviation is transformed into a dimensionless dynamic compensation factor, and the dynamic compensation factor is used to perform nonlinear boundary compensation operations on the preset area expansion upper limit threshold and the preset spatial coincidence lower limit threshold, respectively, and output adaptive expansion upper limit threshold and adaptive coincidence lower limit threshold. In response to detecting that the real-time area ratio is greater than the adaptive expansion upper limit threshold, or that the intersection-union ratio parameter is less than the adaptive overlap lower limit threshold, the morphological mutation anomaly identifier is generated.

[0021] The steps of performing multiplication operations on the first cumulative occurrence count with a preset spatial distortion penalty factor and on the cumulative sum value with a preset temporal severity penalty factor, respectively, specifically include: Obtain the concurrent backlog queue depth of the underlying inference unit performing the feature comparison task; obtain the maximum queue tolerance threshold preset by the system; perform a nonlinear exponential mapping operation based on the resource load duty cycle of the concurrent backlog queue depth and the maximum queue tolerance threshold; and construct a penalty gain coefficient for the underlying processing load capacity dissipation state. The preset spatial distortion penalty factor and the preset temporal severity penalty factor are subjected to nonlinear mapping operations through the penalty gain coefficient to generate dynamically aligned target distortion penalty factor and target severity penalty factor. The first cumulative occurrence count and the target distortion penalty factor are multiplied together, and the cumulative sum value and the target severity penalty factor are multiplied together. The two product results are then summed to generate the concurrent blocking evaluation index.

[0022] Define the target contour enclosing matrix (denoted as ). This represents the coordinate lattice structure of the maximum physical projection range of the currently welded component. A morphological abrupt change anomaly identifier is defined (denoted as...). The binary state identifier parameter represents the distortion caused by strong light interference that exceeds the physical deformation limit in the spatial topology of the recognized contour within the current image frame. The feature matching oscillation frequency is defined as (denoted as...). This represents the cumulative number of state transitions within the underlying inference unit's internal state machine during a continuously set clock cycle, between "receiving the current frame prediction output" and "calling the historical compensation queue." The centroid wander deviation is defined as... This is used to quantify the perspective displacement of automotive welded parts in multiple consecutive frames of images caused by the normal physical movement of a robotic arm. The concurrent backlog queue depth is defined (denoted as...). This represents the queue length of unprocessed image frames or feature matching tasks awaiting allocation of processing load capacity within the current underlying system. A penalty gain coefficient (denoted as...) is defined. The concurrency blocking evaluation index is defined as the adjustment scale that performs nonlinear amplification on the static evaluation factor based on the current backlog state of the underlying system resources. The pure numerical result characterizes the abnormal state of the overall system load. Its generation logic has been upgraded from the original linear multiplication and addition to a dynamic nonlinear weighted summation controlled by the backlog state of the architecture.

[0023] This embodiment extracts the two-dimensional coordinate probability set output by a pre-trained visual feature calculation network; performs an inverse normalized physical coordinate system mapping operation, and outputs the target contour enclosing matrix. Specifically, in the system memory space, it polls and listens for completion status notifications / asynchronous execution result responses of visual feature prediction; it accumulates and counts the callback values ​​triggered by the conditional transition instructions of the state machine, and outputs the feature matching oscillation frequency. Read the system's First-In-First-Out (FIFO) task buffer queue status word; extract the total number of currently unqueued request nodes, and output the concurrent backlog queue depth. .

[0024] Get the depth of the concurrent backlog queue and the preset maximum queue tolerance threshold Based on the resource load duty cycle of the concurrent backlog queue depth and the maximum queue tolerance threshold, the operation is performed according to the nonlinear exponential amplification rule, and the penalty gain coefficient is output. Its operational logic is represented as follows: ; Where k is a constant characterizing the system's stress sensitivity, when the concurrent backlog depth... Approaching the maximum queue tolerance threshold At that time, the penalty gain coefficient A step amplification occurs.

[0025] In this embodiment, the maximum queue tolerance threshold is serialized and configured in the global business registry; it is synchronously read by the main thread when the underlying calculation action of constructing the nonlinear penalty gain coefficient is triggered. Its specific value is limited by the logical cache space tolerance boundary allocated by the current operating system or the underlying computing runtime environment / business isolation sandbox for the visual task processing environment; for example, in a standard concurrent logical runtime node environment, to prevent business process-level buffer queue overflow and state deadlock, its absolute value is configured between [1024, 2048] frames based on the architecture resource pool quota. The maximum queue tolerance threshold represents the maximum tolerable abstract buffer node size of the underlying task scheduling stack before triggering business-level resource rejection allocation.

[0026] In the high-frequency, high-intensity light and spark-splashing environment of automotive welding production lines, traditional static thresholds and linear evaluation mechanisms exhibit severe lag and rigidity under extreme conditions. Static area ratios and cross-union ratios cannot distinguish between pseudo-topological expansion caused by light spots and perspective distortion caused by normal robotic arm movement, easily leading to false positives. Using fixed penalty factors for linear blocking evaluation, when on-site computing nodes encounter continuous, extremely strong interference leading to a massive backlog of concurrent tasks, the linearly increasing evaluation index cannot break through the circuit breaker threshold in time before the underlying computing architecture falls into concurrent deadlock / processing resources are completely exhausted, posing a fatal risk of silent system failure.

[0027] This embodiment constructs an adaptive cross-dimensional evaluation and circuit-breaking closed-loop mechanism. Constrained by the physical space movement patterns of the external automotive welding machine, the centroid movement deviation is reconstructed. By acquiring the topological displacement parameters of the contours of continuous historical frames, dynamic boundary compensation is performed on the rigid area expansion upper threshold and the spatial coincidence lower threshold based on these parameters. This processing enables the threshold envelope to adapt to the legitimate movement trajectory of the robotic arm, filtering out interference from physical perspective artifacts at the data input end.

[0028] By introducing deep awareness of underlying computing resources, the system extracts the depth of the concurrent backlog queue. When underlying computing resources are sufficient, the system maintains linear multiply-accumulate evaluation; when encountering sudden and continuous strong light causing a backlog of recalculation tasks and a surge in queue depth, the system immediately initiates exponential amplification logic for the penalty gain coefficient. Through this penalty gain coefficient, synchronous nonlinear scaling and weight amplification are performed on the spatial distortion penalty factor and the temporal adverse penalty factor represented by the original heterogeneous constant.

[0029] If the concurrent backlog queue depth is caused by extreme transient intense light... The overflow exceeds the maximum queue tolerance threshold within a single clock cycle. The underlying system will silently trigger an abnormal blocking branch, forcibly assigning the penalty gain coefficient to the arithmetic upper bound extreme value / numerical tolerance limit that the system architecture can bear. This ensures that the concurrent blocking evaluation index of the subsequent multiply-accumulate output unconditionally exceeds the concurrent circuit breaker threshold within microseconds, completely blocking redundant feature calculation overhead. In this embodiment, the arithmetic upper bound extreme value or numerical tolerance limit is essentially a pre-calibrated limit penalty constant, designed to ensure that the calculated concurrent blocking evaluation index exceeds the concurrent circuit breaker threshold within a single scheduling cycle.

[0030] The implementation method for step S1 is described as follows: Obtain the target contour enclosing matrix that characterizes the light spot interference in the automotive welding production line. Using the underlying inference unit of the on-site image processing terminal, extract the two-dimensional coordinate probability set representing the predicted position of the welded parts, output by the pre-trained visual feature solving network at the end of the inference forward propagation process. Analyze the two-dimensional coordinate probability set, perform an inverse normalization connection closure mapping action, and generate the target contour enclosing matrix.

[0031] To address the generation mechanism of the target contour enclosing matrix, this embodiment employs a visual feature calculation network based on a convolutional neural network (CNN) architecture. This network's topology consists of a cascaded feature extraction backbone layer and a coordinate regression output layer. Its application aims to extract the true edges of components from pixel domains heavily contaminated by high-frequency sparks. A training dataset of automotive welding images, containing a preset image size, is constructed by extracting frames from real-time video streams of physical welding nodes. Based on pixel-level semantic masking rules, independent channel separation annotations are performed on areas interfered by strong light sources and the true metal contours of components. During training, a focal loss function is introduced to enhance the network's discrimination weights between bright artifact edges and true metal edges. After iterative optimization, this network model is deployed in the bottom-level inference unit. The visual feature calculation network's data structure is a set of feature tensor matrix mappings containing multi-layer neuron weight parameters. In the current business, it takes upstream optical sensing data as input and outputs feature space results containing probability distributions to the downstream architecture.

[0032] In the inference execution layer, a two-dimensional coordinate probability set output by the visual feature computation network is obtained; this set is represented as a tensor structure containing multiple levels of feature channels. A pre-calibrated filtering threshold representing the lower limit of confidence is obtained, and the coordinate extrema points in the tensor structure whose probability values ​​exceed the filtering threshold are extracted; where: The filtering threshold (representing the set judgment threshold) is the lower bound of the probability of blocking high-frequency splashing background noise from being extracted as a valid boundary contour. In this embodiment, the preferred range is a floating-point real number of 0.70~0.95, with a preferred value of 0.85. The specific value of this parameter is based on conventional experimental calibration of the false detection / false detection balance rate in the field. Through this boundary constraint, the invalid occupation of system feature analysis computing resources by low-level artifacts is forcibly blocked. Perform inverse normalization of the physical pixel coordinate system on the extracted extreme points; obtain the inverse normalized set of physical coordinate points, and extract the coordinates of the extreme left, extreme right, extreme top, and extreme bottom boundaries of this set; construct a rectangular closed topological structure defined by the four vertices; and output the target contour enclosing matrix. The computational logic of the above inverse normalization and bounding rectangle generation is expressed as follows:

[0033] in, The bounding matrix of the target contour Let be the coordinates of the i-th pixel in the two-dimensional coordinate probability set that has exceeded the filtering threshold. This refers to finding the absolute minimum value of the vertical coordinate (Y-axis) among the set of valid pixels that have passed the filtering threshold in the two-dimensional coordinate probability set. This means finding the absolute maximum value of the horizontal coordinate (X-axis) among the valid pixel set that exceeds the filtering threshold. and Explanation and and The explanation is similar and will not be repeated; mapping scaling coefficient The determination steps are as follows: obtain the width of the native output physical matrix of the underlying optical sensor and the set width of the input tensor of the visual feature calculation network; obtain the pixel width and pixel height of the native output physical matrix of the underlying optical sensor and the set width and set height of the input tensor of the visual feature calculation network respectively; divide the pixel width by the set width to obtain the lateral mapping scale coefficient. The vertical mapping scale coefficient is obtained by dividing the pixel height by the set height. This eliminates the heterogeneous dimensional differences in two-dimensional space. The mapping scale coefficient ensures that the output boundary can be seamlessly mapped back to the real physical field of view. In this embodiment, the mapping scale coefficient logically represents the magnification / reduction factor of the mapping from the algorithm's tensor space to the objective physical field of view.

[0034] By decoupling the scaling of the horizontal and vertical coordinate systems, an inverse normalization recombination mapping based on two-dimensional independent coefficients is performed.

[0035] Lateral mapping scale coefficient The meaning is the magnification factor representing the mapping of the algorithm tensor space to the X-axis direction of the objective physical field of view; in this embodiment, the preferred value is a real number in the range of 1 to 10, and the preferred value is 4; Longitudinal mapping scale coefficient The meaning is the magnification factor of the mapping of the algorithm tensor space to the Y-axis direction of the objective physical field of view; in this embodiment, the preferred range is a real number from 1 to 10, and the preferred value is 4; specifically, it is dynamically aligned according to the physical resolution of the actual sensor.

[0036] The target contour enclosing matrix is ​​defined as the largest set of outermost pixels that define the physical boundary of the target within the current image frame. In the current system communication architecture, it serves as the basic input data for spatial topological distortion assessment. Before performing morphological distortion determination, the centroid wandering deviation of the target contour enclosing matrix within a set sliding window is extracted from historical frames whose previous prediction confidence exceeded a set determination threshold.

[0037] This embodiment constructs a feature extraction model based on time-series trajectories to eliminate normal perspective distortion caused by the physical movement of the robotic arm. A sliding time window queue of length W is allocated in memory, and the target contour enclosing matrix of consecutive historical frames is obtained and stored in this queue. The diagonal coordinate extrema of each frame's matrix are extracted, and the geometric centroid coordinates of each frame's enclosing matrix are calculated by taking the average of the coordinates. The geometric centroid coordinates of the first and last frames within the current time window are obtained, and Euclidean distance is performed on them to extract the intermediate vector representing the magnitude of the physical position translation. Differential / differential processing is performed on this intermediate vector based on the sliding time window step size to extract its rate of change feature, thereby outputting the centroid wandering deviation. The feature extraction operation law of the centroid displacement in this embodiment is expressed as follows:

[0038]

[0039] in Let be the geometric centroid coordinates of the t-th frame image. This is the centroid wandering deviation. and These represent the leftmost boundary coordinates (minimum X value) and rightmost boundary coordinates (maximum X value) of the effective target contour that crosses the confidence threshold in the current t-th frame image, respectively, in the physical field of view mapping system. and These represent the uppermost boundary coordinates (minimum Y value) and lowermost boundary coordinates (maximum Y value) of the effective target contour in the physical field of view mapping system in the current t-th frame image, respectively.

[0040] W represents the frame depth of the sliding window. The acquisition cycle time constant for a single image frame. The frame depth W of the sliding window represents the total number of historical image frames with step sizes cached in memory for motion difference extraction; in this embodiment, a constant in the range of 3 to 15 is preferred, with a preferred value of 5. and These correspond to the centroid coordinates of the latest frame (frame t). The X-axis component and the Y-axis component. and These are the coordinate components of the starting point of the sliding window's history. They correspond to the historical frames before W frames backwards (the first W frames). The X-axis and Y-axis components of the centroid coordinates of a frame.

[0041] Acquisition cycle time constant This means the time interval constraint between two consecutive independent prediction probability sets generated by the bottom-level optical sensor; in this embodiment, the preferred range is a real number of 5 to 25 milliseconds, and the preferred value is 10 milliseconds. The centroid wandering deviation represents the actual physical movement rate of the welded component in the camera's field of view. The logical constraint boundary is limited by the maximum physical feed rate limit of the servo motor.

[0042] The centroid wandering deviation is obtained and transformed into a dimensionless dynamic compensation factor based on a preset normalization mapping rule. By extracting the maximum physical feed rate limit preset in the system architecture as a benchmark parameter, the centroid wandering deviation is divided by this benchmark parameter for scale normalization to obtain the dynamic compensation factor; the dynamic compensation factor is then used to set the upper limit threshold for area expansion. Lower limit threshold for overlap with preset space Perform boundary compensation operations for nonlinear product scaling and scaling respectively, and output an adaptive expansion upper limit threshold. Adaptive overlap lower limit threshold .

[0043] Among them, the preset upper limit threshold for area expansion The area expansion ratio, calibrated to characterize the limit of normal perspective deformation, is specifically set to a real number within the range of [1.2, 1.5], with a preferred value of 1.3; a preset lower limit threshold for spatial overlap is also included. It is designated as the lower bound characterizing the objective overlap between consecutive frames of the same target, and its specific value range is set to a real number in the range of [0.5, 0.8], with a preferred value of 0.65.

[0044] The specific execution logic is as follows:

[0045]

[0046]

[0047] Furthermore, dynamic compensation factor The meaning is to characterize the relative load percentage of the currently identified mechanical physical walk rate within the global allowable physical limit; in this embodiment, the preferred range is a real number from 0 to 1, and the preferred value is 0.15; Maximum physical feed rate limit The meaning is the maximum absolute linear velocity boundary reference value of the controlled servo motor in the current production system environment; in this embodiment, the preferred range is a real number of 1000~3000, and the preferred value is 2000; Subsequently, the areas of the target contour bounding matrices of the current frame and the previous historical frame whose prediction confidence exceeded a set judgment threshold are extracted, and the real-time area ratio between the two is calculated. Based on the diagonal coordinate extreme values ​​of the two bounding matrices, the extreme values ​​of the overlapping rectangles are extracted using geometric boundary intersection derivation rules to calculate the intersection area and the system union area, and the quotient is used to calculate the intersection-union ratio parameter of the spatial coordinate overlap region. The geometric extraction logic of the above intersection-union ratio is expressed as follows:

[0048]

[0049] in The overlapping area of ​​the bottom layer. The area covered by the target contour bounding matrix in the current frame. The coverage area of ​​the target contour bounding matrix for historical frames. and These are the minimum and maximum diagonal coordinates of the target contour enclosing matrix in the current frame, respectively. and These are the minimum and maximum diagonal coordinates of the target contour enclosing matrix of the historical frame, respectively.

[0050] In response to detecting at least one of the two distortion triggering conditions, namely, the real-time area ratio being greater than the adaptive expansion upper limit threshold or the intersection-union ratio parameter being less than the adaptive overlap lower limit threshold, a trigger identifier / effective bit is written to the underlying data register structure to generate a morphological abrupt change anomaly identifier characterizing spatial topological distortion.

[0051] The logic for determining the morphological mutation anomaly identifier is as follows: Obtain the real-time area ratio and the adaptive expansion upper limit threshold, and perform a first threshold out-of-bounds check; simultaneously obtain the intersection-union ratio parameter and the adaptive overlap lower limit threshold, and perform a second threshold out-of-bounds check. In response to the first or second extreme value comparison result satisfying a preset trigger condition, the concurrent anomaly interception mechanism is activated, and an emergency response execution branch is initiated at the system scheduling layer. This branch calls the underlying write-operation instruction to directly update a specific flag bit in the business status table of the concurrent scheduling and control module, setting its logical status word to active / activated, thereby completing the export of the morphological mutation anomaly identifier. The flipping of this Boolean state will serve as the core digital messenger for subsequent blocking assessment.

[0052] The morphological mutation anomaly identifier data type is a logical state scalar, with state enumeration values ​​of {0: normal, 1: mutation occurred}, used to trigger the evaluation branch of system load anomalies.

[0053] Extract the feature matching oscillation frequency within the underlying inference unit. Within the business operation context of the underlying inference unit, a daemon process is initiated to poll and listen for visual feature prediction completion status notifications / asynchronous execution result responses. The number of cumulative flips between the feature output state of the visual feature calculation network in the current frame and the state of coordinate weighted compensation execution via the historical sliding window queue is counted within a continuously set underlying system clock cycle. This cumulative flip count is directly stored as the feature matching oscillation frequency characterizing the severity of the time-series calculation.

[0054] The feature matching oscillation frequency is extracted as follows: a bypass observation mechanism / module, independent of the main recognition logic, is configured at the system scheduling layer. This process does not directly read image pixels but maps probes to the instruction flow execution stack of the underlying inference unit. Within a set clock cycle observation window, the logical branch jump records within the traversal state machine are maintained. In response to the detection of an abnormal cross-branch jump callback between the memory address output by the deep network and the compensation address for calling historical cache frame data, a step-by-step accumulation operation is performed on the state counter. When the observation window time boundary is reached, the final accumulated value in the register is extracted and reset to zero. The extracted integer value is then discarded to obtain the feature matching oscillation frequency. The feature matching oscillation frequency represents the trigger density of the quantized state machine conditional transition instructions, with the preferred logic constraint boundary limited by the maximum number of branch prediction callbacks of the CPU within a single clock cycle.

[0055] The implementation method for step S2 is as follows: The morphological mutation anomaly identifier and the feature matching oscillation frequency are input into the collaborative scheduling and control module. Based on the spatial topological distortion features represented by the morphological mutation anomaly identifier and the computational blocking features represented by the feature matching oscillation frequency, a load-aware nonlinear penalty alignment operation is performed to generate a concurrent blocking evaluation index for characterizing the abnormal load state of the system.

[0056] Within the non-blocking / high-concurrency observation data ring bus of the collaborative scheduling and control module, a sliding observation window is set. The sliding observation window defines the buffer queue depth range for the system state machine to perform concurrent anomaly assessment. In this embodiment, the preferred range is a buffer depth of 10-50 frames of images, with a preferred value of 30 frames. The system iterates through and extracts the first cumulative occurrence (denoted as ) of the morphological mutation anomaly flag with a Boolean value of true within the sliding observation window. ), and simultaneously extract the cumulative sum of the feature matching oscillation frequencies within the time window (denoted as ), Obtain the concurrent backlog queue depth of the underlying inference unit performing the feature comparison task; This embodiment focuses on obtaining the concurrent backlog queue depth, which is essentially a lossless probe of the system application layer microservice scheduling stack. It locates the first-in-first-out (FIFO) task buffer queue bound to the underlying inference unit in system memory; extracts the current enqueue and dequeue status nodes of this task buffer queue; calculates the total number of image frame nodes currently pushed into the queue but not yet allocated processing resources based on the depth difference between the first and last nodes; and directly outputs this total number of nodes and assigns it to the concurrent backlog queue depth. This probe process is executed in an independent scheduling branch by calling an asynchronous state probe interface, avoiding observer effects on the main business identification pipeline. The concurrent backlog queue depth represents the real-time queue length characterizing the dissipation of the underlying processing load capacity and the task backlog status, with logical constraints limited to [0, maximum queue tolerance threshold]. ].

[0057] Penalty gain coefficient is constructed based on the depth of the concurrent backlog queue. The specific configuration inputs are the concurrent backlog queue depth extracted in the previous steps, the maximum queue tolerance threshold, and the system stress sensitivity constant k read from the configuration file. The concurrent backlog queue depth and the maximum queue tolerance threshold are obtained, and a division operation is performed between them to obtain the ratio representing the current concurrent load saturation. The natural constant base and the preset system stress sensitivity constant are obtained, and the system stress sensitivity constant is multiplied by the load saturation ratio. The product result is extracted as the power exponent for nonlinear amplification. Using the natural constant base as the base, an exponential mapping operation is performed based on the power exponent, and the penalty gain coefficient is finally output.

[0058] The system stress sensitivity constant k is a static adjustment scale that controls the steepness of the nonlinear mapping curve. It is extracted and invoked in response to the triggering of the concurrent blocking evaluation calculation branch. Its objective value range is converged within the range of [5,10]; specifically, by utilizing the natural exponential property of this range, it ensures that the gain coefficient of the evaluation system is extremely flat when the backlog load is low (queue depth occupancy is less than 50%), while it can produce a step-like numerical surge when the backlog load is extremely high (queue occupancy is close to 90%), thereby forcing the realization of a hard threshold breakdown without delay.

[0059] The system stress sensitivity constant k defines the curvature sensitivity of the system when it needs to penalize and amplify the backlog of concurrent tasks.

[0060] The logical formula for calculating the penalty gain coefficient Where k is the system's compressive sensitivity constant; when much smaller When the penalty gain coefficient approaches its steady-state low value, it will eventually reach its maximum value. Approaching The coefficient exhibits a steep, exponential step jump, thus triggering a circuit breaker. The optimal value of the penalty gain coefficient is close to 1 when there is no congestion, and dynamically approaches the system's set maximum cutoff value when there is extreme congestion.

[0061] The penalty gain coefficient is obtained as a dynamic weighting factor, and compared with the preset spatial distortion penalty factor read by the system. and preset timing-based adverse penalty factors Performing a product operation maps static heterogeneous parameters to unified baseline feature values, generating a dynamically aligned target distortion penalty factor. With the target severe punishment factor .

[0062] The first cumulative occurrence count and the target distortion penalty factor are obtained and multiplied across dimensions, respectively; the cumulative sum value and the target severity penalty factor are obtained and multiplied across dimensions, respectively; the output of these two products is extracted and summed and aggregated to finally generate a purely numerical concurrent blocking evaluation index (denoted as ). Its execution logic is as follows:

[0063]

[0064]

[0065] in and These represent the first cumulative occurrence count and the cumulative total value, respectively. The concurrent blocking assessment index is a single scalar threshold that characterizes the state of cross-dimensional spatial and temporal co-deterioration, and is used as the sole decision output of the system scheduling layer.

[0066] Regarding the aforementioned preset spatial distortion penalty factor In its initial state, this parameter is serialized and stored in the local static configuration file of the collaborative scheduling and management module. In response to the main scheduling system service startup process, this parameter is loaded once into the static global parameter set / benchmark reference rule table of the runtime architecture for high-frequency calls by the main process. During specific business deployment, its value boundary is determined by the historical load test baseline of the computational overhead for large-area pixel distortion by the field computing nodes. In this embodiment, its objectively configured empirical enumeration range is limited to [1.5, 3], used to establish the basic mapping base for the spatial dimension. The preset spatial distortion penalty factor means assigning a static initial weight benchmark to the spatial morphological distortion characteristics in the global evaluation.

[0067] Regarding the aforementioned preset timing-based severe penalty factor During the initial deployment phase, it reads the local environment variable space to fix the constant; in response to the wake-up action of the concurrent scheduling main thread, it attaches the constant to the high-frequency business access channel / front-end data exchange area of ​​the system scheduling layer for high-frequency reading. Based on the throughput overhead stress test baseline of the field computing node for high-frequency state machine flip instructions, its objectively configured empirical enumeration range is limited to [2,5], aiming to establish a fixed mapping base for the timing oscillation frequency. The preset timing degradation penalty factor means to give the state machine flip degradation characteristics a static initial weight benchmark in the global concurrency evaluation.

[0068] Regarding the implementation of steps S3 and S4: the concurrent circuit breaker threshold must be calibrated in the pre-processing stage. The specific execution flow is as follows: An independent test sandbox is built, configured with a system operating framework and communication network components consistent with the real production line; simulated image streams with progressively increasing frequencies are continuously injected into the sandbox to artificially create a high-concurrency business state; the underlying performance monitoring interface is called to monitor the logical cache resource occupancy rate at the system architecture level in real time; in response to the detection that the occupancy rate is approaching the preset safety limit boundary (preferably 95% of the cache capacity), indicating that the system is about to fall into deadlock due to memory overflow, the system immediately extracts the concurrent blocking evaluation index output through fusion calculation at this time, and directly solidifies this index value as the concurrent circuit breaker threshold. The concurrent circuit breaker threshold represents the limit of heterogeneous interference fusion index that the system can withstand before a complete shutdown.

[0069] In response to the determination that the concurrent blocking assessment index has fallen below the preset concurrent circuit breaker threshold, pose control signaling is generated for scheduling the underlying inference unit and configuring the external automotive welding robot arm servo controller. Specifically, multi-source service monitoring parameters, including spatial boundary morphological abrupt changes and underlying inference unit blocking status, are acquired; a system-level concurrent deadlock trend assessment is performed based on these multi-source monitoring parameters; and according to the assessment result of the concurrent deadlock trend, the corresponding emergency level pose control signaling is output to the lower-level communication network card and the underlying inference unit, where the underlying inference unit specifically loads and runs a pre-trained visual feature calculation network logic component.

[0070] In response to the pose control signal, the underlying inference unit is driven to perform the suspension action of the feature matching calculation process, and the pose control signal is simultaneously output to the external automotive welding robot arm servo controller to trigger the cutting action of the external feed enable.

[0071] The aforementioned closed-loop suspension and disconnection actions are specifically manifested as dual-track routing and state transition rules between the application-level scheduling and control layer and the network communication layer. In the application-layer state control branch, an application-layer state reset operation is performed: Actively intercept requests for computing resource allocation and data processing to the underlying inference unit; forcibly update the operation control status word of the underlying inference unit to the business suspension status; forcibly stop the feature comparison process of the current image, actively stop the system computing power from being consumed in deep network convolution of the image tensor that has been completely contaminated by extremely strong light spots, and instantly stop the system processing load at the application layer to reclaim CPU cycles and prevent invalid resource occupation under extreme interference.

[0072] In the network communication routing branch, within the same system clock cycle, the pose control signaling is directly routed to the network interface controller via the event dispatcher; scheduling actions are performed to bypass the blocked primary service first-in-first-out (FIFO) packet queue; in the bypass buffer of the network layer, an emergency network message carrying a high-priority routing identifier and blocking control word is encapsulated directly according to the predefined underlying data structure of the industrial network communication stack. This emergency network message is instantiated in the memory control stack as a fixed-length structured data payload, whose enumerated and defined data structure topology specifically includes: a frame start identifier for triggering the effective timing, a priority forwarding status word field (high-priority VLAN tag encapsulation area) occupying a bit width, a target routing field for identifying the logical identifier of the target peripheral servo node, a core service control payload segment carrying a static control word, and a transmission error control field for data integrity verification; and the emergency network message, which has completed the above topology formatting and encapsulation, is forcibly pushed into the priority transmission channel of the communication protocol stack.

[0073] The objective byte layout of emergency messages in the network protocol stack is fixed through static architecture enumeration. A logical flag indicating that the message has the highest latency preemption priority in the switch / industrial routing architecture is used; its priority status enumeration value is set to the first normal flow flag and the second priority preemption flag. The static control word means a hard machine instruction code that forces the servo controller at the external receiving end to interrupt and lock the axial displacement; its control word enumeration value is logically defined as a maintain execution scalar and a force block scalar.

[0074] Further scheduling of the independent emergency communication queue of the industrial network communication stack, directly encapsulating and generating pose control signaling carrying priority forwarding status words through a non-blocking transmission channel, and sending it out with priority; constrained by the objective system boundary of the external automotive welding robot arm servo motor, the pose control signaling is used to physically cut off the feed command axis (feed enable signal) of the robot arm servo motor, ensuring deterministic triggering of anti-collision protection actions in the event of data congestion. This achieves the reclamation of host computing resources and the physical position locking of the external actuator.

[0075] When the concurrent blocking evaluation index monotonically increases and eventually exceeds the preset concurrent circuit breaker threshold, the mapping system experiences a surge in invalid visual feature recognition tasks due to extreme light spot interference, leading to a state evolution where the queuing depth at the business layer approaches the collapse threshold. This surge in the index represents a quantitative convergence of the topological feature distortion and the deterioration in the frequency of state switching caused by continuous strong light interference.

[0076] The concurrent backlog queue depth and the penalty gain coefficient follow a non-linear exponential amplification relationship. At the business data flow level, when the queue depth is low, the penalty gain coefficient maintains a smooth, steady-state output; however, when the concurrent backlog queue depth approaches the maximum queue tolerance threshold, the penalty gain coefficient is exponentially driven by the system's stress sensitivity constant, resulting in a step-like numerical burst. This non-linear boundary mapping design prevents the linear penalty factor from triggering circuit breaker intervention in time before the business queuing channel becomes completely blocked. By introducing exponential amplification logic for the resource load duty cycle, it is ensured that the system can reach its limit value before the business logic channel's carrying capacity is exhausted. The exponential positive correlation between the concurrent backlog queue depth and the penalty gain coefficient is pointed out, which mathematically supports the anti-collision benefit of this method by decisively triggering external feed enable cutoff actions within a very short time window.

[0077] The centroid wandering deviation exhibits a positive proportional compensation relationship with the adaptive expansion upper limit threshold and a negative proportional reduction relationship with the adaptive coincidence lower limit threshold. The underlying business logic of this parameter combination lies in quantifying the legitimate relative motion displacement of the welding object within the field of view. By normalizing the extracted centroid wandering deviation relative to the maximum physical feed rate limit, a dynamic compensation factor is generated. This factor directly performs cross-dimensional boundary compensation operations on the static judgment envelope formed by the preset area expansion upper limit threshold and the preset spatial coincidence lower limit threshold. The nonlinear compensation mapping design between the centroid wandering deviation and the bidirectional deformation threshold is presented, and the geometric derivation mechanism of the coordinate tensor supports the method's ability to effectively filter out the risk of false detection of perspective artifacts caused by legitimate physical displacement.

[0078] The method in this embodiment is deployed in a concurrent recognition environment of a car welding production line where high-frequency strong light and target wandering perspective deformation are superimposed. During the continuous image frame input stage, the target contour enclosing matrix is ​​acquired, and real-time area ratio and intersection-union ratio parameters are extracted. When the visual processing network is continuously injected with high-brightness pixel noise, the state variables within this method begin to update dynamically. Based on the currently extracted real-time morphological mutation anomaly identifier and feature matching oscillation frequency, combined with the backlog depth features fed back by the task buffer queue, this method performs data alignment and penalty weighting. When the input end continuously experiences high-intensity transient interference, the change in queue depth directly drives the nonlinear mapping model to recalculate the gain weights, thereby updating the final output pose control signaling state. Based on the above model, when the system's maximum queue tolerance threshold is calibrated to 1024, the system's stress sensitivity constant is preferably 5, and the maximum physical feed rate limit is configured to 2000, the theoretical response calculated by this method is shown in the table below.

[0079] Table 1: Examples of theoretical state deduction based on this embodiment under extreme transient strong light interference and task backlog disturbance.

[0080] A horizontal comparison of the theoretically derived data reveals that under the first set of normal load inputs, based on a very small concurrent backlog queue depth, the penalty gain coefficient remains within the inefficient multiplier range, exhibiting low intrusion on effective computing resources under normal operating conditions. However, in the second and third sets of extreme data groups where high-frequency, strong light injection leads to task backlog, compared to existing technologies that rely solely on static, preset spatial distortion penalty factors and preset temporal adverse penalty factors for linear accumulation, this method represents a qualitative shift in the internal data flow path. This method utilizes a step-like feature amplification triggered by the resource load duty cycle, enabling the computational evolution of heterogeneous index fusion to overcome the sluggish linear slope. This theoretical state deduction confirms that this scheme eliminates computational redundancy caused by continuous idling on "pseudo-targets," shortens the business logic flow link from abnormal feature extraction to the generation of locking signaling, and ensures deterministic convergence of control decisions under extreme transient loads. This invention defines the following application range: Normal Feature Recognition and Adaptive Compensation Range: The filtering threshold used to distinguish effective targets from background noise is limited to a real number within the range of [0.7, 0.95] based on the aforementioned settings. In this embodiment, the preferred value of 0.85 is used as an example trigger boundary. The triggering mechanism of this boundary stems from the fact that, under normal physical walk conditions and small background noise, the probability distribution of the forward propagation output of the visual feature calculation network remains within the confidence domain. By extracting the centroid Euclidean distance of consecutive multiple frames to generate the centroid walk deviation, the topological distortion features caused by real physical translation and local noise expansion are effectively separated at the business flow level. If this constraint is deviated from or not followed, using a fixed area and cross-union ratio threshold for static comparison will lead to an engineering defect where a single static threshold misjudges normal robotic arm movement as light spot expansion distortion.

[0081] Within this range, the method continuously drives the underlying logic to parse the two-dimensional coordinate probability set to generate a target contour enclosing matrix containing four integer coordinate values. Based on the dynamic compensation factor generated by the centroid wandering deviation, it performs nonlinear boundary compensation operations on the preset area expansion upper limit threshold and the preset spatial coincidence lower limit threshold, and outputs the adaptive expansion upper limit threshold and the adaptive coincidence lower limit threshold, so that the threshold envelope can adapt to the legal movement trajectory of the robotic arm.

[0082] High-load feature blocking and cross-dimensional weighted evaluation interval: The interval determination condition is triggered by performing a nonlinear exponential mapping operation based on the resource load duty cycle of the concurrent backlog queue depth and the maximum queue tolerance threshold obtained in real time. The triggering mechanism of this boundary originates from the fact that when encountering continuous extreme strong light input, the extracted morphological mutation anomaly markers frequently become true, and the cross-branch jump between the state machine accepting the current frame prediction and calling the historical frame to perform compensation causes a surge in the frequency of feature matching oscillations, resulting in a sharp accumulation of feature comparison tasks in the first-in-first-out buffer queue. If the dynamic exponential adjustment constraint based on resource dissipation state is deviated from or not followed, and a static fixed penalty factor is continued to be used for linear blocking evaluation when encountering a massive backlog of concurrent tasks, the linearly increasing evaluation exponent will not be able to break through the circuit breaker in time before the system crashes, thus causing a large amount of memory read / write and feature parsing computing resources to be stuck on processing pseudo-targets for a long time.

[0083] When the data congestion development falls into this range, this method initiates the step amplification logic of the penalty gain coefficient based on the aforementioned judgment conditions. The penalty gain coefficient is used to perform dynamic nonlinear scale alignment on the preset spatial distortion penalty factor and the preset temporal adverse penalty factor, respectively. Then, the first cumulative occurrence number and cumulative sum value are obtained and cross-dimensional product operation and summation are performed with the aligned target distortion penalty factor and target adverse penalty factor, respectively, to generate a concurrent blocking evaluation index to characterize the abnormal state of the overall business load.

[0084] Extreme overflow and concurrent deadlock circuit breaker interception range: The maximum queue tolerance threshold is limited to the range of [1024, 2048] frames based on the aforementioned settings, in order to perform forced intervention before the underlying task scheduling stack triggers business-level resource rejection allocation. The triggering mechanism of this boundary stems from the extreme transient and extremely strong light spot, which directly causes the total number of request nodes that have not been dequeued within a single clock cycle to overflow and exceed the set objective tolerance limit. It marks the absolute convergence bottleneck of the lossless identification algorithm and the anti-crash red line of data queue congestion at the business architecture flow layer.

[0085] If this hard-blocking constraint is deviated from or not followed, critical pose control signaling will be unable to be issued within the extremely short operation time window, leading to irreversible concurrent scheduling blockage and the risk of silent failure due to complete loss of the physical locking capability of external actuators. In response to triggering this extreme deadlock condition, the abnormal blocking branch of this method forcibly assigns the penalty gain coefficient to the arithmetic upper bound extreme value / numerical tolerance limit that the system architecture can bear, so that the concurrent blocking evaluation index of the cross-dimensional multiply-accumulate output unconditionally exceeds the pre-calibrated concurrent circuit breaker threshold within microseconds; this method actively updates the operation control status word for the underlying inference unit to the service suspension state to suspend the feature comparison action, and drives the industrial network communication stack to bypass and encapsulate pose control signaling carrying the priority forwarding status word 0xFF and the static control word 0x00, which is preferentially issued to the external servo controller to instantly cut off the external feed enable, thereby recovering the host computing resources and forcibly converging the unsolvable underlying algorithm distortion to the safest physical static state.

[0086] for Figure 1 Further explanation: BIU stands for Low-level Inference Unit, which is deployed on the on-site image processing terminal, runs a pre-trained visual feature calculation network, and is responsible for outputting a set of two-dimensional coordinate probabilities representing the predicted position of the welded parts. It is the physical perception frontier that receives environmental noise interference.

[0087] TBM stands for Target Contour Enclosure Matrix, representing the geometric tensor data after physically hierarchically mapping the two-dimensional coordinate probability set. It is used to extract the area ratio and intersection-union ratio (IU / U) parameters of the current frame and historical frames. .

[0088] MMF stands for Morphological Abnormality Identifier. It indicates that when the real-time area ratio or intersection-union ratio parameter of the target contour enclosing matrix exceeds the expansion or overlap threshold of the dynamic adaptive compensation, the generated logical state scalar represents the distortion of the spatial topology caused by transient arc light interference.

[0089] FOF represents the feature matching oscillation frequency, which characterizes the cumulative number of flips that occur between receiving the current frame and calling the historical sliding window to perform coordinate weighting. It quantitatively reflects the severity of time-series computation caused by the algorithm getting stuck in "pseudo-target" matching.

[0090] CQD represents the concurrent backlog queue depth, characterizing the queue length of unprocessed image frames or alignment tasks in the underlying task scheduling stack. A non-linear exponential mapping is performed based on this parameter and the duty cycle of the maximum tolerance threshold. The penalty gain coefficient of the generation blocking system is generated.

[0091] CSM stands for Coordinated Scheduling and Control Module, which represents the core hub containing the circular queue of observation data. It is responsible for fusing morphological mutation anomaly identifiers with feature matching oscillation frequencies and performing cross-dimensional weighted summation operations in conjunction with penalty gain coefficients.

[0092] BEI stands for Concurrency Blocking Assessment Index, which represents a pure numerical scalar result output by the Cooperative Scheduling and Control Module. It is used to characterize the current load anomaly and the approaching state of concurrent scheduling deadlock in the software and hardware system across the entire domain.

[0093] PCS stands for Pose Control Signaling, representing an emergency command payload generated in response to the concurrent blocking evaluation index breaking through a preset concurrent circuit breaker threshold. This signaling carries a priority forwarding status word and forcibly seizes the highest scheduling control of the system using a bypass queue.

[0094] WSC stands for Servo Controller for Automotive Welding Robotic Arm, representing the control center of the external physical actuator. Upon receiving the pose control signal, it instantly executes the feed enable cut-off action, transforming the unsolvable algorithmic disorder into an absolutely safe physical static locked state.

[0095] Figure 1 The data flow and physical response mechanism is as follows: The on-site vision system acquires the target contour enclosing matrix representing the light spot interference in the automotive welding production line. It extracts the area ratio and intersection-union ratio (IU / I) parameters of the target contour enclosing matrix between the current frame and the previous historical frame whose prediction confidence exceeded a set threshold, generating a morphological abrupt change anomaly identifier. Simultaneously, it extracts the feature matching oscillation frequency within the underlying inference unit. The aforementioned morphological abrupt change anomaly identifier and feature matching oscillation frequency are input into the collaborative scheduling and control module. Based on the spatial topological distortion features and computational blocking features represented by these two identifiers, a load-aware nonlinear penalty alignment operation is performed to generate a concurrent blocking evaluation index representing the abnormal load state of the system. When the concurrent blocking evaluation index is determined to have broken through a preset concurrent melting threshold, pose control signals are immediately generated for scheduling the underlying inference unit and configuring the external automotive welding robotic arm servo controller. In response to the pose control signaling, the underlying inference unit is driven to perform the suspension action of the feature matching calculation process, and the pose control signaling is simultaneously output to the external automotive welding robot arm servo controller via the industrial network communication stack bypass, so as to trigger the external feed enable cutting action, thereby realizing the rapid recovery of host computing resources and the safe physical position locking of the external actuator.

[0096] Figure 1The TBM (Target Contour Enclosing Matrix) and MMF (Morphological Abruptness Matrix) in the middle form the noise perception interface in the spatial dimension; FOF (Feature Matching Oscillation Frequency) maps the temporal computation severity of the underlying state machine. These heterogeneous metrics, along with CQD (Concurrent Queue Depth), are integrated into the architecture's computational core CSM (Cooperative Scheduling and Control Module) for data alignment and weight amplification. The highest decision metric generated after module aggregation is the BEI (Concurrent Blocking Assessment Index). Once this index triggers a critical condition, it acts as the logical source to issue a high-priority PCS (Position Control Signal). This signal ultimately forces its way through the hardware and software architecture layers, reaching the underlying execution carriers WSC (Automotive Welding Robotic Arm Servo Controller) and BIU (Bottom Inference Unit), thus constructing a seamless feedback loop from microscopic pixel-level anomaly capture to macroscopic mechanical-level hardware interlocking encompassing the interaction of automotive welding component features.

[0097] The computational logic involved in this application can be constructed using algorithms such as regression analysis in machine learning, establishing a mathematical model by analyzing the inherent trends and interrelationships of the collected parameters. This process can be implemented using specialized computational tools (such as Python's Scikit-learn library or the R language environment). Throughout all calculations, to eliminate the influence of different physical dimensions and ensure that data is compared and analyzed on the same scale, the input parameters in each formula are dimensionless. The dimensionless techniques used include, but are not limited to, max-min normalization or Z-score standardization.

[0098] To decouple the core algorithm from specific application strategies and ensure the configurability and ease of debugging of the technical solution, all configurable operating parameters in the specific implementation path of this invention are read through a standardized "configuration interface". The data source of this configuration interface is a "data storage module" (e.g., a non-transitory computer-readable storage medium, such as a configuration file, database entry, or cloud configuration service), which is configured to store configuration data in key-value pair format.

[0099] It should be emphasized that the foregoing embodiments are merely illustrative of preferred implementations of the present invention and are not intended to limit the scope of protection of the present invention. This application also provides a computer-readable storage medium having computer program instructions stored thereon.

Claims

1. A method for position recognition of automotive welded parts based on deep learning, characterized in that, The specific steps include: Step S1: Obtain the target contour enclosing matrix that characterizes the light spot interference of the automotive welding production line, extract the area ratio and intersection-union ratio parameter of the target contour enclosing matrix of the current frame and the previous historical frame whose prediction confidence exceeded the set judgment threshold, so as to generate a morphological mutation anomaly label; the morphological mutation anomaly label characterizes the binary state label parameter, representing that the recognized contour in the current image frame has encountered strong light interference distortion that exceeds the physical deformation limit in the spatial topology. And extract the feature matching oscillation frequency within the underlying inference unit; The acquisition of the target contour enclosing matrix includes: extracting the set of two-dimensional coordinate probabilities representing the predicted position of the welded parts from the output of the pre-trained visual feature solving network through the low-level inference unit of the on-site image processing terminal, and parsing and mapping it into the target contour enclosing matrix; The steps for extracting features to match oscillation frequencies specifically include: In the business operation context of the underlying inference unit, the visual feature prediction callback records are polled and monitored. The cumulative number of flips occurs within a continuously set underlying system clock cycle between the feature output state of the visual feature calculation network receiving the current frame and the state of calling the historical sliding window queue to perform coordinate weighted compensation; the cumulative number of flips is used as the feature matching oscillation frequency characterizing the severity of the time-series calculation. Step S2: Input the morphological mutation anomaly identifier and the feature matching oscillation frequency into the collaborative scheduling and control module. Based on the spatial topological distortion features represented by the morphological mutation anomaly identifier and the computational blocking features represented by the feature matching oscillation frequency, perform a load-aware nonlinear penalty alignment operation to generate a concurrent blocking evaluation index for characterizing the abnormal load state of the system. The steps for calculating the concurrent blocking evaluation index specifically include: Within the observation data circular queue of the collaborative scheduling and control module, a sliding observation window is set; Extract the first cumulative number of times the morphological mutation anomaly marker is triggered as true within the sliding observation window, and extract the cumulative sum of the feature matching oscillation frequencies within the sliding observation window during the same period; The first cumulative occurrence count is multiplied by a preset spatial distortion penalty factor, and the cumulative sum is multiplied by a preset temporal severity penalty factor. The two product results are then summed to generate the concurrent blocking evaluation index. Step S3: In response to determining that the concurrent blocking evaluation index breaks through the preset concurrent circuit breaker threshold, generate pose control signaling for scheduling the underlying inference unit and configuring the external automotive welding robot arm servo controller; Step S4: In response to the pose control signal, drive the underlying inference unit to perform the suspension action of the feature matching calculation process, and simultaneously output the pose control signal to the external automotive welding robot arm servo controller to trigger the external feed enable cut-off action, thereby realizing the recovery of host computing resources and the physical position locking of the external actuator.

2. The method for position recognition of automotive welded parts based on deep learning according to claim 1, characterized in that: The steps for obtaining the generated morphological mutation anomaly identifiers specifically include: Calculate the real-time area ratio of the area of ​​the target contour bounding matrix to the area of ​​the target contour bounding matrix of the historical frame whose prediction confidence exceeded the set judgment threshold, and calculate the intersection-union ratio parameter of the overlapping area of ​​the two spatial coordinates. In response to the detection of at least one of the following distortion trigger conditions: The real-time area ratio is greater than the preset upper limit threshold for area expansion, and the intersection-union ratio parameter is less than the preset lower limit threshold for spatial overlap. Generate the morphological mutation anomaly identifier that characterizes spatial topological distortion.

3. The method for position recognition of automotive welded parts based on deep learning according to claim 2, characterized in that: The pose control signaling is configured to be triggered synchronously: Update the operation control status word for the underlying inference unit to the service suspension status; The industrial network communication stack encapsulates and generates the pose control signaling carrying the priority forwarding status word in the independent emergency communication queue.

4. The method for position recognition of automotive welded parts based on deep learning according to claim 3, characterized in that: Before detecting the distortion triggering condition, the steps of obtaining the target contour enclosing matrix and generating the morphological abrupt change anomaly identifier further include: Extract the centroid wandering deviation of the target contour bounding matrix within a set sliding window for the historical frame where the previous prediction confidence exceeded the set judgment threshold. Based on the preset normalization mapping rule, the centroid wandering deviation is transformed into a dimensionless dynamic compensation factor, and the dynamic compensation factor is used to perform nonlinear boundary compensation operations on the preset area expansion upper limit threshold and the preset spatial coincidence lower limit threshold, respectively, and output adaptive expansion upper limit threshold and adaptive coincidence lower limit threshold. In response to detecting that the real-time area ratio is greater than the adaptive expansion upper limit threshold, or that the intersection-union ratio parameter is less than the adaptive overlap lower limit threshold, the morphological mutation anomaly identifier is generated.

5. The method for position recognition of automotive welded parts based on deep learning according to claim 4, characterized in that: The steps of performing multiplication operations on the first cumulative occurrence count with a preset spatial distortion penalty factor and on the cumulative sum value with a preset temporal severity penalty factor, respectively, specifically include: Obtain the concurrent backlog queue depth of the underlying inference unit performing the feature comparison task; Obtain the system's preset maximum queue tolerance threshold, and perform a nonlinear exponential mapping operation based on the resource load duty cycle of the concurrent backlog queue depth and the maximum queue tolerance threshold to construct a penalty gain coefficient for the underlying processing load capacity dissipation state. The preset spatial distortion penalty factor and the preset temporal severity penalty factor are subjected to nonlinear mapping operations through the penalty gain coefficient to generate dynamically aligned target distortion penalty factor and target severity penalty factor. The first cumulative occurrence count and the target distortion penalty factor are multiplied together, and the cumulative sum value and the target severity penalty factor are multiplied together. The two product results are then summed to generate the concurrent blocking evaluation index.

Images