A dynamic weld seam detection system for flexible production line robotic arms integrating laser tracker and dual photoelectric frequency comb.

Abstract

This invention discloses a dynamic inspection system for weld seams of a robotic arm on a flexible production line, integrating a laser tracker and dual photoelectric frequency combs. Based on numerical simulation, it establishes the discrete optimal intervals of the repetition frequency difference and carrier envelope offset frequency, constructing a precise time-domain measurement benchmark. Furthermore, it integrates fault saliency ranking and parallel convolutional neural networks, extracting signal morphology indicators of interference fringes and rearranging the sampling sequence to eliminate servo lag errors in high-speed tracking by the laser tracker. The inspection system provided by this invention can achieve real-time precise reconstruction of the weld seam trajectory of a robotic arm in complex environments on flexible production lines, effectively solving the problems of nonlinear interference and dynamic asynchrony in cross-scale measurements, significantly improving the inspection efficiency and positioning accuracy of automated welding, and possessing broad application prospects.

Description

Technical Field

[0001] This invention relates to the field of weld inspection technology, and in particular to a dynamic weld inspection system for a flexible production line robotic arm that integrates a laser tracker and a dual photoelectric frequency comb. Background Technology

[0002] High-precision dynamic sensing by robotic arms in flexible production lines is one of the core technologies for realizing intelligent manufacturing. Traditional ranging methods based on a single physical model are often limited by hysteresis errors caused by hardware servo bandwidth when facing high-speed dynamic tasks, while purely data-driven methods struggle to guarantee absolute measurement accuracy in the absence of physical constraints. Dual-optical frequency comb technology provides a high sampling rate and high-precision absolute distance measurement method, which is crucial for welding and assembly tasks requiring strict spatial positioning. Compared with traditional laser interferometry, the accuracy of dual-optical comb ranging is highly dependent on the matching of system frequency parameters and must strictly adhere to the bandpass sampling theorem to avoid spectral aliasing.

[0003] Hierarchical deep ranking networks (FSHSM-PCNN) effectively eliminate sensor hysteresis by explicitly introducing a ranking mechanism of signal morphology indicators during feature extraction. The fault saliency ranking module, as an effective data augmentation mechanism, can regroup key features scattered due to transmission delays over time, thereby naturally enhancing the synergistic effect of signals during feature learning.

[0004] Currently, dynamic ranging systems for robotic arms need to overcome the following technical bottlenecks:

[0005] 1) The dual-comb system needs to achieve global discrete optimization of parameters under bandpass sampling constraints to overcome spectral aliasing and phase distortion: Determining the repetition frequency difference and carrier envelope offset frequency that do not cause aliasing in the complex spectral structure of multi-heterodyne interferometry is a major challenge. A theoretical model and numerical simulation method incorporating the evolution of the ultrashort pulse electric field need to be established to ensure that the multi-heterodyne beat frequency signal is strictly within the Nyquist baseband and to overcome the measurement uncertainty caused by time jitter.

[0006] 2) Deep learning models need to eliminate the effects of sensor lag in dynamic time-series signals and extract submerged key features: Realizing the true motion state online in real time amidst data asynchrony and feature ambiguity caused by the high-speed movement of a robotic arm is extremely challenging. This requires introducing a Fault Significance Hierarchy and Ranking (FSHSM) module, which uses morphological indicators such as global maximum and mean values ​​to hierarchically rearrange the original data, thereby revealing the reinforcing synergistic effects masked by time lag.

[0007] 3) The parallel feature extraction architecture needs to balance temporal trends and collaborative features to avoid information loss and maintain high dynamic measurement accuracy: The architecture based on parallel convolutional neural networks (PCNN) needs to dynamically balance the original temporal information and the rearranged collaborative information. The core challenge lies in designing a dual-branch network that can simultaneously capture the kinematic trends of the robotic arm (temporal state features) and the sorted and enhanced detailed features (enhanced collaborative features). While ensuring the effectiveness of feature fusion, the weighting mechanism can be used to improve the absolute ranging performance of the system in complex noisy environments.

[0008] To address the aforementioned technical bottlenecks and reduce the trial-and-error costs of physical experiments while improving model robustness, this invention introduces a parameter optimization method based on numerical simulation. By constructing a pulse evolution model incorporating temporal jitter noise, it achieves discrete-range optimization of the parameters of the dual-comb system. This mechanism not only significantly improves the signal-to-noise ratio of the ranging system but also provides high-fidelity simulation training data for deep learning networks. This invention combines dual-comb numerical simulation parameter optimization, a deep hierarchical ranking mechanism, and a parallel convolutional neural network to propose a novel dynamic ranging framework for robotic arms, achieving a unified approach to physical accuracy benchmarks, dynamic error correction, and collaborative feature extraction within an intelligent perception framework. Summary of the Invention

[0009] To address the shortcomings of existing technologies, this invention provides a dynamic detection system for weld seams of a flexible production line robotic arm that integrates a laser tracker and dual photoelectric frequency combs. The system aims to solve the comprehensive sensing problems of robotic arm systems in flexible production line environments in terms of cross-scale spatial tracking, dynamic signal lag, and absolute ranging accuracy.

[0010] To achieve the above objectives, the technical solution of the present invention is: a dynamic inspection system for weld seams of a flexible production line robotic arm integrating a laser tracker and a dual photoelectric frequency comb, the system comprising:

[0011] The dual-comb ranging subsystem is used to establish an evolution model of the pulse electric field evolution process including the signal laser and the local oscillator laser. The influence of repetition frequency difference and carrier envelope offset frequency on ranging accuracy is analyzed by numerical simulation. By searching for the discrete optimal interval, it is ensured that the sampled spectrum is in the baseband range and does not undergo phase distortion, thereby generating a high-steady-state ultrashort pulse sequence that satisfies the bandpass sampling law.

[0012] The laser tracker and dual-comb common-path emission subsystem are used to divide the beam into a reference optical path and a measurement optical path using a beam splitter, and to establish a reference reference for eliminating drift. The measurement beam is injected into the collimated emission axis of the laser tracker through a coaxial coupling element, and the tracker's servo turntable guides the beam toward the target ball at the end of the robotic arm, constructing a reciprocating measurement channel containing spatial time-of-flight information, and realizing the spatial physical alignment of the measurement optical axis and the robotic arm's motion vector.

[0013] The fault significance-based hierarchical sorting module addresses the signal lag problem caused by servo response in dynamic measurements by using global max pooling and average pooling operators to extract signal morphology indicators of interference fringes. The original fringe sequence is then hierarchically rearranged using a fast non-dominated sorting algorithm to aggregate the scattered key features in the time dimension, thereby eliminating nonlinear lag and revealing the strengthening synergistic effect between data.

[0014] A parallel convolutional neural network feature extraction module is used to implement dynamic error correction based on a deep hierarchical sorting network. The parallel network extracts the temporal state features of the original sequence and the enhanced collaborative features of the rearranged sequence, respectively. After weighted fusion, these features are mapped to precise absolute distance prediction values ​​through a fully connected regression layer. The precise absolute distance prediction values ​​are combined with the instantaneous azimuth and pitch angles of the laser tracker to calculate the three-dimensional spatial coordinates. The spatial position deviation is extracted by comparing the actual trajectory of the robotic arm end effector with the theoretical weld trajectory, and the final dynamic detection result of the weld is output.

[0015] Furthermore, in the evolution model, a mathematical expression for the single-pulse electric field is established, and a time-domain superposition model of an infinite pulse sequence is defined to obtain the electric field representation of a laser pulse train with repetitive periods. The carrier envelope offset frequency is determined based on the carrier envelope phase slip between adjacent pulses.

[0016] Furthermore, in the evolution model, a coaxial multiheterodyne interference sampling model is constructed. First, the repetition frequency difference between the signal laser and the local oscillator laser is defined. Within the update period, the multiheterodyne interference signal contains two sets of cross-correlation fringes: a reference set and a measurement set. Its time-domain waveform is described as the cross-correlation convolution of the signal light field and the local oscillator light field. The flight time delay generated by the measurement optical path is linearly mapped to the radio frequency time domain, forming cross-correlation interference fringes.

[0017] Furthermore, multi-heterodyne interference signal acquisition is performed, and the coherent superposition signal of the local oscillation pulse and the return signal pulse is received through a photodetector. The local oscillation pulse is used to perform linear optical sampling on the measurement pulse carrying displacement characteristics. By utilizing the slip effect generated by different repetition frequencies, a multidimensional original cross-correlation interference fringe sequence containing the dynamic displacement information of the robotic arm is generated within the update cycle.

[0018] Furthermore, based on the multiheterodyne beat frequency signal set and the repetition frequency difference between the laser and the local oscillator laser, the optimal interval for signal discretization is searched through numerical simulation to generate the original interference fringe sequence that satisfies the constraints and has no phase distortion.

[0019] Furthermore, in the laser tracker and dual-comb common-path emission subsystem, an optical beam-splitting model of the reference arm and the measurement arm is constructed. The beam is divided into a reference beam and a measurement beam according to the energy splitting ratio using a beam splitter. The reference beam reaches the detector after traveling a fixed path length, and its electric field model is represented as the time shift of the signal source electric field. Using the servo turntable of the laser tracker, the measurement beam is injected into the tracker's emission optical axis through a coaxial coupling element, ensuring that the optical axis center of the measurement beam coincides with the aiming line of sight of the tracker. This allows the dual-comb measurement beam to follow the dynamic guidance of the tracker and point in real time at the target ball at the end of the robotic arm. Based on the instantaneous distance of the target ball at the end of the robotic arm relative to the optical center of the tracker, the round-trip propagation path length of the measurement beam in space is obtained, and the dynamic optical path difference between the measurement arm and the reference arm is calculated. This optical path difference is linearly mapped into the time delay of the interference fringes in the subsequent photoelectric detection stage.

[0020] Furthermore, based on the repetition frequency of the local oscillator and the signal laser, and the carrier envelope offset frequency, the frequency distribution of the multi-heterodyne beat frequency signal in the radio frequency domain is obtained. Through photoelectric conversion, the optical frequency difference in the terahertz band is linearly reduced in dimension and mapped to the radio frequency signal set in the megahertz band. The output signal of the photodetector is spectrally filtered based on a bandpass filter to ensure that the filtered multi-heterodyne beat frequency signal set is strictly located within the first Nyquist baseband interval and does not overlap with the zero-frequency or half-repetition frequency boundary, thereby ensuring the integrity of the time-domain waveform. Using an analog-to-digital converter with the repetition frequency of the local oscillator laser as the clock reference, the filtered analog signal is synchronously digitally acquired to generate a discretized multidimensional original cross-correlation interference fringe sequence. This sequence is directly used as the input data for the subsequent deep hierarchical sorting network to solve the flight time delay of the robotic arm end effector.

[0021] Furthermore, the global maximum and mean values ​​of the original sequence are extracted using a dual-channel pooling layer as signal morphology indicators, and the samples are rearranged using a fast non-dominated sorting algorithm to eliminate signal hysteresis caused by the tracker servo response.

[0022] Furthermore, the global maximum and global average values ​​are calculated based on the signal morphology index vector of the original sequence through a dual-channel pooling layer; an index sequence is generated using a fast non-dominated sorting algorithm, and the original sequence is rearranged into hierarchical samples; temporal state features and reinforcement synergy features are extracted based on a parallel convolutional neural network, and the absolute distance is output through a regression layer after weighted fusion.

[0023] Furthermore, a dual-channel pooling layer is used to extract the morphological features of the signal at each sampling time. Global max pooling is used to capture the signal abrupt change features, and global average pooling is used to characterize the overall signal intensity. A fast non-dominated sorting algorithm is used to perform multi-objective hierarchical division of the morphological feature indicators, generating an index sequence that reflects the significant differences in the signal. Based on the index sequence, the multidimensional original interference fringe sequence is rearranged in the time dimension to generate hierarchical samples that eliminate time lag and have enhanced synergistic effects. The input is a one-dimensional convolutional neural network with two parallel branches to process the time state features and synergistic features respectively. The two features are cascaded and fused, and mapped to the precise absolute distance prediction value of the robotic arm end effector through a fully connected regression layer.

[0024] Furthermore, the dual-comb ranging subsystem is specifically implemented by including the following steps:

[0025] S11: Establish an electric field evolution model for ultrashort pulse lasers, considering the single-pulse electric field. The product of the carrier signal and the Gaussian envelope function is expressed in the time domain as follows:

[0026]

[0027] in This represents the Gaussian envelope function determined by the laser locking model properties. The carrier angular frequency, This refers to the carrier envelope phase.

[0028] Define a time-domain superposition model for an infinite pulse sequence, for a repetition period of... The laser pulse train, its entire field Represented as:

[0029]

[0030] This can be simplified to the form of pulse sequence summation, where The pulse index is an integer. The carrier envelope phase shift is the amount of phase shift between adjacent pulses, which determines the carrier envelope offset frequency. .

[0031] S12: Construct a coaxial multiheterodyne interference sampling model, defining the repetition frequency difference between the signal laser and the local oscillator laser as... The time shift of optical sampling Represented as:

[0032]

[0033] During the update cycle Inside, the photodetector receives multiheterodyne interference signals. It contains two sets of cross-correlation fringes, one for reference and one for measurement. Its time-domain waveform can be described as the cross-correlation convolution of the signal light field and the local oscillator light field:

[0034]

[0035] in and The electric field distributions of the signal light and the local oscillator light are represented respectively, and the time-of-flight delay generated by the optical path is measured. The cross-correlation interference fringes are linearly mapped to the radio frequency time domain.

[0036] S13: Perform discrete-range optimization of system parameters, defining the comb modes of the signal laser and the local oscillator laser in the optical frequency domain as follows: Level and First Level. Define the set of multiheterodyne beat frequencies. Its frequency components are determined by the beat frequency between the signal tooth and the local oscillator tooth:

[0037]

[0038] in , These are the repetition frequencies of the local oscillator and the signal laser, respectively. , This corresponds to the carrier envelope offset frequency. Through photoelectric conversion, the optical frequency difference in the terahertz band is linearly reduced in dimension and mapped to the radio frequency signal set in the megahertz band. To avoid aliasing and ensure signal integrity, the bandpass sampling theorem must be satisfied. The optimization constraints are set as follows:

[0039] and

[0040] Search through numerical simulation and The discrete optimal interval ensures the multiheterodyne signal set Completely fall into the first Nyquist base zone It does not overlap with zero frequency or half-repetition frequency, thus ensuring the integrity of the time-domain waveform.

[0041] S14: A random time jitter noise model is introduced to simulate the stability of the laser under real physical conditions, within the repetition period of the signal laser. Superimposed random noise term :

[0042]

[0043] The original interference fringe sequence, including nonlinear errors, is generated using the optimized frequency parameters and the injected noise model. This serves as the physical benchmark training data for subsequent deep learning networks.

[0044] Furthermore, the specific implementation of the laser tracker and the dual-comb shared-optical-path emission subsystem includes the following steps:

[0045] S21: Constructing the optical beam splitting topology for the reference arm and the measuring arm

[0046] The output pulse of the parameter-optimized signal laser is defined as the source electric field. Using fiber optic bundle splitters (FBS) to achieve energy coupling ratios The optical path is divided into a reference channel and a measurement channel.

[0047] The reference channel introduces a fixed optical path delay. Its time-domain electric field model Represented as a time-delayed copy of the source electric field:

[0048]

[0049] in The speed of light in a vacuum. After being shaped by a collimator, the beam from the measurement channel is injected into the subsequent tracking and guidance unit as a "time-of-flight" carrier.

[0050] S22: Establish a spatial kinematic guidance model for coaxial coupling of the laser tracker.

[0051] Define the mechanical base coordinate system of the laser tracker as follows: Using coaxially coupled optical elements, the measurement beam is injected non-destructively into the center of the azimuth-elevation two-axis servo turntable of the tracker.

[0052] Let the instantaneous azimuth angle of the servo turntable be... The pitch angle is Then the measuring beam The outgoing unit direction vector It can be represented as:

[0053]

[0054] The vector is pointed in real time to the cooperative target (such as a corner cube prism) installed at the end of the robotic arm, forming a dynamic closed-loop optical link to ensure that the measurement beam always propagates along the center of the line of sight, eliminating geometric alignment errors caused by the high-speed movement of the robotic arm, and enabling the dual-comb measurement beam to follow the dynamic guidance of the tracker and point in real time to the target ball at the end of the robotic arm.

[0055] S23: Constructing a dynamic round-trip propagation and time-varying electric field model

[0056] Assume the target ball at the end of the robotic arm is The instantaneous distance relative to the optical center of the tracker at time is Then the round-trip propagation path length of the measured beam in space is Considering the Doppler effect and path delay introduced by the robotic arm's motion, the returned measured electric field... Described as:

[0057]

[0058] in The optical loss coefficient includes atmospheric attenuation and target sphere reflectivity. It is a single-pulse envelope function. For the first The carrier envelope phase evolution term of each pulse.

[0059] Therefore, a dynamic optical path difference model between the measuring arm and the reference arm is established:

[0060]

[0061] This optical path difference will be linearly mapped into a time delay of the interference fringes during the photoelectric detection process. .

[0062] S24: Perform multiheterodyne temporal convolution sampling and linear scaling transformation

[0063] The total optical field received by the photodetector is the superposition of the reference light and the measurement light reflected from the target. This is because the signal light and the local oscillator light have a different repetition frequency. The interference current output by the detector In the time domain, this manifests as the cross-correlation convolution of two pulsed electric fields:

[0064]

[0065] in For including round-trip flight time The electric field of the measurement signal, For the local oscillator electric field, This indicates taking the real part.

[0066] Based on the time stretching principle of dual optical combs, the actual flight time Measurement delay linearly amplified and mapped onto the RF time axis The mapping relationship strictly follows:

[0067]

[0068] This step occurs within a single measurement update cycle. The system completes the absolute ranging sampling of the instantaneous position of the robotic arm, generating a raw interference fringe signal containing distance features.

[0069] Furthermore, the specific implementation of the fault saliency-based hierarchical ranking module includes the following steps:

[0070] S31: Constructing a multidimensional signal morphology index extraction model

[0071] To address the sensor dynamic hysteresis effect caused by the high-speed movement of the robotic arm in a flexible production line, a discretized original cross-correlation interference fringe sequence is defined as follows: A dual pooling layer is introduced to perform morphological feature compression on time-series signals.

[0072] Extract each sampling time using the Global Max Pooling (GMP) operator. The signal peak abrupt change characteristics are used to capture local extrema caused by phase jitter:

[0073]

[0074] in This is the global maximum value.

[0075] The global average pooling (GAP) operator is used to extract the overall energy intensity features of the signal to characterize the average modulation depth of the interference fringes.

[0076]

[0077] in, This is the global average. For feature dimensions or number of sensor channels, Indicates the first Time of the first Signal amplitude in each dimension. This leads to the construction of a morphological index set containing both local and global temporal features. .

[0078] S32: Perform hierarchical partitioning based on fast non-dominated sort.

[0079] Introducing the Fast Non-Dominated Sort Algorithm (FNSA) to integrate the set of morphological indicators Each element in the matrix is ​​considered a multi-objective optimization solution vector. All sampling times are divided into different non-domination levels based on dominance relationships.

[0080] Define dominance relationship: for two moments and ,like and If at least one inequality is strictly true, then the time of determination is... Dominant moment .

[0081] By iteratively stripping away the non-dominated layers, a hierarchical index sequence reflecting the dynamic significance differences of the signal is generated. This causes high-significance features to be prioritized:

[0082]

[0083] in This represents the original time index after rearrangement.

[0084] S33: Generate enhanced collaborative hierarchical sample sequences

[0085] According to the index sequence For the original interference fringe sequence Nonlinear mapping rearrangement is performed along the time dimension to generate new sample sequences that eliminate temporal lags and exhibit a reinforced synergistic effect. .

[0086] In this sequence, signal segments with similar dynamic saliency are spatially clustered, allowing key interference features that are submerged by robotic arm motion noise to emerge in the new hierarchical structure, providing a high signal-to-noise ratio input tensor for feature extraction by subsequent parallel networks.

[0087] Furthermore, the parallel convolutional neural network feature extraction module is specifically implemented by including the following steps:

[0088] S41: Constructing a Parallel Convolutional Neural Network (PCNN) Feature Extraction Topology: Establishing a one-dimensional convolutional neural network architecture containing two independent parallel branches.

[0089] The first branch is defined as the "temporal state-aware branch" to discretize the original interference fringe sequence. Using this as input, the aim is to capture time-varying state information during the robotic arm's motion. Multi-layer convolution and the ReLU activation function are used to extract temporal state features that preserve the robotic arm's kinematic trends. :

[0090]

[0091] The second branch is defined as the "enhanced collaborative perception branch", which is the hierarchical sample sequence rearranged after step S3. Using the input as input, we leverage convolutional kernels to aggregate features with similar saliency across discontinuous time points, extracting enhanced collaborative features masked by dynamic lag. (Reinforced Synergistic Features).

[0092]

[0093] in and These are the trainable network parameters for the two branches, respectively.

[0094] S42: Perform dual-channel feature convolution and non-linear mapping

[0095] Multi-layer one-dimensional convolutional layers (Conv1D), batch normalization layers (BN), and max pooling layers are configured in the two branches respectively.

[0096] set up For network layer indexes, the input feature tensor is... Then the first Layer output Represented as:

[0097]

[0098] in This represents a one-dimensional convolution operation. and These represent the kernel weights and biases, respectively. For ReLU activation function:

[0099]

[0100] go through After layer feature extraction, the first branch outputs a deep temporal state feature vector. The second branch outputs a deep-enhanced collaborative feature vector. .

[0101] S43: Implement adaptive feature weighted fusion and absolute ranging regression

[0102] Introduce a feature fusion layer and define adaptive weight coefficients. and For time series features With synergistic features Perform weighted concatenation to generate a comprehensive feature representation. :

[0103]

[0104] Construct a fully connected regression network (FCN) to integrate features Mapped to precise absolute distance predictions at the end of the robotic arm .

[0105] The physical calibration formula for precise absolute distance is:

[0106]

[0107] in The speed of light in a vacuum For group refractive index, This is the time delay of the stripe center after network correction.

[0108] Define the regression loss function as the distance between the predicted distance and the theoretical distance of the dual-comb physical model. The mean squared error (MSE) between the parameters is used for backpropagation optimization of the network parameters.

[0109]

[0110] in The number of training batch samples is denoted by . By minimizing this loss function, the network automatically learns to eliminate nonlinear errors caused by sensor hysteresis, outputting measurement results with high dynamic response.

[0111] The beneficial effects of this invention are as follows:

[0112] (1) In view of the problem of complex parameter matching and limited accuracy in absolute ranging of robotic arms in flexible production lines, this invention breaks through the limitations of traditional experimental trial and error methods, which are time-consuming and difficult to achieve global optimization. By establishing a numerical simulation model that includes the evolution of ultra-short pulse electric field and random time jitter noise, and combining the bandpass sampling law to perform discrete interval parameter scanning, the optimal configuration of the dual optical comb system in terms of repetition frequency difference and carrier envelope offset frequency is ensured, thereby effectively avoiding spectral aliasing and phase distortion in the multi-heterodyne sampling process, and providing the system with a measurement benchmark with high signal-to-noise ratio and conforming to physical laws.

[0113] (2) To solve the problem of signal lag and key feature submersion caused by sensor transmission delay in high-speed movement of robotic arms, this invention deeply integrates signal morphology analysis with non-dominated sorting theory; by designing a fault significance-based hierarchical sorting module (FSHSM), using dual-channel pooling operators to extract time-varying morphological indicators and perform hierarchical rearrangement, the influence of nonlinear time lag is innovatively eliminated at the data level; this mechanism can re-aggregate highly significant data points scattered at different time intervals in the time dimension, showing the deep reinforcement synergistic effect of dynamic signals;

[0114] (3) In order to achieve high-precision and robust error correction in dynamic and complex environments, this invention constructs a parallel perception framework that combines physical model and data-driven collaboration. The framework adopts a dual-branch parallel convolutional neural network (PCNN) structure to extract the original temporal state features that retain kinematic trends and the enhanced collaborative features that have been sorted and enhanced in parallel. It also achieves deep fusion of multi-dimensional information through an adaptive weighting mechanism. Experimental verification shows that the framework can significantly improve the regression accuracy of nonlinear dynamic errors. While ensuring computational efficiency, it effectively solves the engineering pain point that traditional methods cannot take into account both temporal information and local mutation features. Attached Figure Description

[0115] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0116] Figure 1 The trajectory tracking error diagram under the coarse measurement benchmark of the laser tracker shows the absolute error response curve of the laser tracker's independent ranging and demonstrates the micron-level random noise distribution generated by the servo bandwidth under the dynamic displacement of the robotic arm, reflecting its coarse measurement accuracy characteristics as a large space guidance benchmark.

[0117] Figure 2 The error distribution diagram for independent ranging using dual optical combs shows that, in the absence of absolute position prior, although the physical ranging accuracy meets the sub-micron level requirements, the measurement results exhibit periodic non-ambiguous range jumps due to the limitation of repetition frequency, reflecting the range limitations of a single type of sensor.

[0118] Figure 3 The error convergence curve of the fusion ranging system of the present invention shows that after "coarse and fine coupling" and dynamic correction, the ranging residual smoothly converges to near zero without any fuzzy breakpoints, verifying the system's full-range high-precision measurement capability across scales.

[0119] Figure 4 The flowchart illustrates the working principle of laser tracker and dual optical comb collaborative ranging, and elaborates on the physical implementation steps from dual optical comb light source parameter optimization, coaxial optical path guidance, multi-heterodyne beat frequency acquisition to integer ambiguity resolution, demonstrating the cross-scale collaborative logic of the hardware system.

[0120] Figure 5The flowchart of the data processing based on the deep hierarchical ranking network shows the calculation process of rearranging the original interference data using the fault saliency classification and ranking module and extracting the enhanced collaborative features through the parallel convolutional neural network, thus verifying the effectiveness of the algorithm in real-time compensation for dynamic hysteresis error.

[0121] Figure 6 The flowchart of numerical simulation and discrete optimization of dual optical comb system parameters is presented, which shows the calculation logic of full parameter spatial scanning based on the ultrashort pulse electric field evolution model, and establishes the discrete optimal interval of repetition frequency difference and carrier envelope offset frequency to construct a high signal-to-noise ratio physical measurement benchmark.

[0122] Figure 7 The diagram shows the logic for resolving integer ambiguity based on the prior of the laser tracker. It describes in detail the mathematical process of using coarse measurement data from the laser tracker to assist fine measurement data from the dual optical comb in calculating the number of integers, and clarifies the algorithm principle for achieving unambiguous absolute ranging across the entire range through the "coarse-fine coupling" mechanism.

[0123] Figure 8 The internal execution logic diagram of the Fault Significance Hierarchy and Ranking Module (FSHSM) is presented, demonstrating the specific steps of extracting morphological indicators using dual-channel pooling operators, performing fast non-dominated ranking, and generating reinforced collaborative hierarchical samples, revealing the core mechanism of deep learning networks in eliminating dynamic signal hysteresis. Detailed Implementation

[0124] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described below with reference to the accompanying drawings and examples. It should be understood that the specific examples described herein are merely illustrative and not intended to limit the invention.

[0125] This specific embodiment aims to elaborate on a flexible production line robotic arm weld seam dynamic detection system that integrates a laser tracker and dual photoelectric frequency combs. This system achieves high-precision absolute ranging in complex dynamic environments through cross-scale collaboration at the hardware level and deep learning error correction at the software level. The overall hardware collaboration logic of this embodiment is as follows: Figure 4 As shown, the data processing algorithm flow is as follows: Figure 5 As shown.

[0126] The system of this invention includes a dual-comb ranging subsystem, a laser tracker and a dual-comb shared optical path emission subsystem, a fault saliency-based hierarchical ranking module (FSHSM), and a parallel convolutional neural network (PCNN) feature extraction module. The specific implementation process is as follows:

[0127] 1. Physical evolution model and parameter discrete optimization based on dual-comb ranging subsystem

[0128] The dual-comb ranging subsystem is used to establish a physical evolution model, including the electric field equations for ultrashort pulses, carrier envelope phase drift, and the principle of multi-pulse time-domain superposition. A numerical simulation optimization strategy based on the bandpass sampling theorem is designed. Through full-parameter scanning, the discrete optimal interval between the repetition frequency difference and the carrier envelope offset frequency is searched to ensure that the multiheterodyne beat frequency signal is strictly within the Nyquist baseband and free from spectral aliasing, thus establishing a high signal-to-noise ratio physical measurement benchmark. Based on the electric field evolution theory of ultrashort pulse lasers, a numerical simulation model incorporating random time jitter noise is constructed. To address the frequency parameter sensitivity of the dual-comb ranging subsystem, discrete interval parameter scanning is performed to establish a high signal-to-noise ratio physical measurement benchmark.

[0129] Specifically, the following steps are included:

[0130] S11: Constructing an evolution model for ultrashort pulse electric fields

[0131] Define a single-pulse electric field The product of the carrier signal and the Gaussian envelope function is expressed in the time domain as follows:

[0132]

[0133] in The Gaussian envelope is determined by the characteristics of the laser-locked model. For amplitude, The pulse width. The carrier angular frequency, This refers to the carrier envelope phase.

[0134] Constructing a time-domain superposition model of infinite pulse sequences Assume the laser repetition period is For a repetition period of The laser pulse train is represented across the entire field as:

[0135]

[0136] This can be simplified to the form of pulse sequence summation, where The pulse index is an integer. The carrier envelope phase shift is the amount of phase shift between adjacent pulses, and the carrier envelope offset frequency is defined as... .

[0137] S12: Implement parameter discrete interval search based on bandpass sampling theorem

[0138] Figure 6 This is the algorithm flowchart for this process. To prevent spectral aliasing of multiple heterodyne interference signals in the radio frequency domain and to ensure signal integrity, the bandpass sampling theorem must be strictly followed, and the beat frequency signal set must be defined. The distribution constraints. Specifically, the bandwidth of the effective interference signal (i.e., The Nyquist condition must be satisfied. Numerical simulations were performed to study the repetition frequency difference across the entire parameter space. With carrier envelope offset frequency Perform a joint scan to identify discrete optimal zones that satisfy the following dual conditions:

[0139]

[0140] and

[0141] and

[0142] like Figure 4 As shown in the "Parameter Optimization Module", by... Locked at a specific frequency point obtained from the simulation (e.g., at) Time setting This ensures that the signal-to-noise ratio of the measurement system is maximized at the physical level and that phase distortion is eliminated.

[0143] Using an analog-to-digital converter (ADC) to control the repetition frequency of the local oscillator laser Using a clock reference, the filtered analog signal is synchronously digitized and acquired.

[0144] In each measurement update cycle Within this timeframe, a discretized multidimensional original cross-correlation interferometric fringe sequence is generated by extracting a time window containing both the complete reference pulse and the measured pulse response. :

[0145]

[0146] in The number of sampling points within a single measurement cycle is used as the input data for subsequent deep hierarchical sorting networks to calculate the flight time delay of the robotic arm's end effector.

[0147] 2. Dual ranging and data acquisition based on a laser tracker and a dual-comb common-path emission subsystem.

[0148] The laser tracker and dual-comb shared-path emission subsystem utilizes the tracker's servo turntable to guide the laser beam in real-time towards the target sphere at the end of the robotic arm, achieving coarse guidance and optical path closure within a large space. It performs multi-heterodyne interferometry signal acquisition, using local oscillation pulses to perform linear optical sampling of measurement pulses carrying time-of-flight information, generating a raw cross-correlation interference fringe sequence in the radio frequency domain that includes the dynamic displacement characteristics of the robotic arm. This subsystem leverages the laser tracker's large-range tracking capability to overcome the unambiguous range limitation of the dual-comb, achieving full-range measurement through a "coarse-fine coupling" mechanism. The noise characteristics of the laser tracker's independent ranging are as follows: Figure 1 As shown (due to limitations in electronic bandwidth, micrometer-level random noise exists), the periodic jump characteristics of the dual optical comb in the unresolved ambiguity state are as follows: Figure 2 As shown (limited to the non-fuzzy range) (It is serrated).

[0149] Specifically, the following steps are included:

[0150] S21: Constructing the optical beam splitting topology for the reference arm and the measuring arm

[0151] The optimized signal laser output pulse is defined as the source electric field. An optical fiber beam splitter divides the optical path into a reference channel and a measurement channel according to the energy coupling ratio. A fixed optical path delay is introduced into the reference channel, while the beam from the measurement channel is collimated and then injected into the subsequent tracking and guiding unit.

[0152] S22: Establish a spatial kinematic guidance model for coaxial coupling of the laser tracker.

[0153] Establish base coordinate system Real-time reading of the azimuth angle of the servo turntable With pitch angle Control the unit direction vector of the measurement beam's output. Always pointing towards the target ball at the end of the robotic arm (CR2):

[0154]

[0155] This establishes a closed-loop optical link, ensuring that the measurement beam from the dual optical comb can continuously return to the photodetector.

[0156] S23: Perform multiheterodyne temporal convolutional sampling, linear scaling transformation, and integer ambiguity resolution using coarse distance measurements obtained from a laser tracker. As prior information, the precise distance measurement of the dual optical comb Perform the solution.

[0157] S22 is the core step in calculating the number of cycles of the dual optical comb using coarse measurements from a laser tracker. Figure 7 This demonstrates the mathematical logic.

[0158] First, based on the repetition frequency of the dual optical comb Calculate the unambiguous range of a single pulse :

[0159]

[0160] in The speed of light in a vacuum The group refractive index.

[0161] Next, calculate integer multiples of the pulse period. (Full week ambiguity):

[0162]

[0163] Finally, the full-range absolute distance is synthesized. :

[0164]

[0165] This step completes the leap from micrometer-level fuzzy measurement to meter-level absolute measurement, with the data flow as follows: Figure 4 As shown in the "Data Solving" section.

[0166] 3. Data preprocessing based on the Fault Significance Ranking Module (FSHSM)

[0167] The Fault Significance-Based Hierarchical Ranking Module (FSHSM) addresses signal hysteresis caused by servo response in dynamic measurements. It extracts signal morphology indicators of interference fringes using global max pooling and average pooling operators. A fast non-dominated ranking algorithm is then used to hierarchically rearrange the original fringe sequence, clustering scattered key features along the time dimension to eliminate nonlinear hysteresis and reveal reinforcing synergistic effects between data. The data processing flow is as follows: Figure 5 As shown in the first part, to address the sensor sampling hysteresis and nonlinear feature overload problems caused by the high-speed movement of the robotic arm, a deep hierarchical sorting network (FSHSM-PCNN) is introduced to rearrange the original signal.

[0168] Specifically, the following steps are included:

[0169] S31: Constructing a multidimensional signal morphology index extraction model

[0170] Define the discretized original cross-correlation interference fringe sequence as A dual-channel pooling layer is used to extract the data at each sampling time. Signal morphology characteristics.

[0171] Extracting local peak abrupt change features using the Global Max Pooling (GMP) operator :

[0172]

[0173] Global average pooling (GAP) operator is used to extract global energy intensity features. :

[0174]

[0175] Constructing a set of morphological indicators .

[0176] S32: Perform Fast Non-Dominated Sort (FNSA) and Hierarchical Partitioning

[0177] Set of morphological indicators The elements in the vector are considered as multi-objective optimization solutions. Dominance relations are defined as follows: for two time points... and ,like and If at least one inequality is strictly true, then the time of determination is... Dominant moment ( ).

[0178] By iteratively stripping away the non-dominated layer, all time points are divided into... Level Generate an index sequence that reflects the dynamic significance of signal differences. :

[0179]

[0180] The signal segments corresponding to the top index have higher dynamic significance (FaultSignificance).

[0181] S33: Generate enhanced collaborative hierarchical sample sequences

[0182] According to the index sequence For the original interference fringes A nonlinear mapping rearrangement is performed in the time dimension; the complete internal transformation process is as follows: Figure 8 As shown, a new sample sequence is generated. :

[0183]

[0184] exist In this process, key features that were scattered due to transmission delay are re-aggregated, revealing a masked "reinforced synergistic effect," which provides high feature density input for subsequent networks.

[0185] 4. The Parallel Convolutional Neural Network (PCNN) feature extraction module performs dynamic error correction and regression.

[0186] The Parallel Convolutional Neural Network (PCNN) feature extraction module extracts the temporal state features of the original sequence and the enhanced collaborative features of the rearranged sequence through a dual-branch network. It then fuses the dual-branch features using an adaptive weighting mechanism and maps the combined features to a precise absolute distance through a fully connected regression layer, achieving real-time correction and high-precision restoration of the high-speed trajectory of the robotic arm. This module establishes a dual-branch deep learning architecture, fusing temporal and collaborative features to achieve accurate compensation for dynamic measurement errors. The convergence curve of the corrected system's ranging error is shown below. Figure 3 As shown (smoothly converges to near zero with no fuzzy breakpoints).

[0187] Specifically, the following steps are included:

[0188] S41: Establishing a Parallel Feature Extraction Topology

[0189] like Figure 5 As shown in the latter part, the network contains two independent parallel branches:

[0190] First branch (time-aware branch): based on the original sequence As input, a one-dimensional convolutional layer (Conv1D) is used to extract the time-varying state feature vector of the robotic arm's kinematics. .

[0191] The second branch (cooperative sensing branch): using sequence rearrangement Using the same structure as input, convolutional layers are used to extract enhanced collaborative feature vectors across time intervals. .

[0192] S42: Implement adaptive feature weighted fusion

[0193] Introducing adaptive weight coefficients and The dual-path features are cascaded and fused. To enhance the capture of dynamic mutation features, a [specific method / mechanism] is set... Generate a comprehensive feature tensor :

[0194]

[0195] This fusion mechanism ensures that the system can both... Keep track of movement trends, and be able to... Corrects nonlinear errors caused by hysteresis.

[0196] S43: Fully Connected Regression and Loss Function Optimization

[0197] Construct a fully connected regression network (FCN) to integrate features Mapped to precise absolute distance predictions at the end of the robotic arm .

[0198] The loss function is defined as the sum of the predicted value and the physical truth value of the dual optical comb. Mean squared error (MSE) between:

[0199]

[0200] By minimizing this loss function using the backpropagation algorithm, the system automatically learns to eliminate sensor hysteresis and environmental noise. The final output is as follows: Figure 3 As shown, this invention demonstrates that it combines the large range of a laser tracker with the sub-micron precision of a dual optical comb in dynamic environments.

[0201] This invention achieves optimal global parameter configuration at the physical level for a dual-comb ranging system, and the optimization results do not rely on costly experimental trial and error. By establishing a numerical simulation model that includes the evolution of the ultrashort pulse electric field and random time jitter noise, and combining it with the bandpass sampling theorem to perform discrete interval scanning, the precise locking of the repetition frequency difference and the carrier envelope offset frequency is ensured. Compared with traditional empirical adjustment methods, this method avoids spectral aliasing and phase distortion of multiple heterodyne signals at the source, improving ranging accuracy by two orders of magnitude, and providing a high signal-to-noise ratio and physically sound measurement benchmark for high-precision measurement tasks.

[0202] This invention significantly expands the absolute measurement range of the system while maintaining sub-micron level accuracy by integrating laser tracker guidance and dual-optical-comb precision measurement technology. The designed "coarse-fine coupling" mechanism utilizes the tracker's large-range data as prior information, effectively solving the problem of the dual-optical-comb being limited by the non-ambiguity range of repetition frequency, and achieving real-time resolution of pulse integer ambiguity. This design enables the system to achieve both flexible guidance using the tracker and microscopic precision measurement using the dual-optical-comb during cross-scale spatial tracking, overcoming the inherent contradiction between range and accuracy in single sensors, and achieving a unification of macroscopic tracking and microscopic sensing.

[0203] This invention constructs a parallel sensing framework based on a deep hierarchical ranking network, which significantly improves the system's error correction capability in high-speed dynamic environments. Addressing the signal lag problem caused by robotic arm motion, the time series is rearranged using a Fault Saliency Hierarchical Ranking Module (FSHSM), revealing the masked reinforcement synergy effect. Combining a dual-branch parallel convolutional neural network to simultaneously extract temporal state features and reinforcement synergy features makes the policy learning process more comprehensive and efficient. The entire solution, from physical parameter optimization and hardware collaborative acquisition to deep learning compensation, forms a complete sensing closed loop, which has been rigorously validated in terms of dynamic lag correction, feature extraction efficiency, and absolute ranging accuracy. This provides a complete solution for intelligent sensing in flexible production lines under dynamic uncertainties and high-precision requirements, combining high physical fidelity, high dynamic response, and data-driven robustness.

[0204] The simulation and flowchart results of this invention are as follows: Figures 1-5 As shown. In Figure 1 The absolute error response curve of the laser tracker's independent ranging is displayed, showing the distribution of micron-level random noise generated by the servo bandwidth under the dynamic displacement of the robotic arm, reflecting its coarse measurement accuracy characteristics as a large space guidance reference. Figure 2 The error distribution diagram for independent ranging using dual optical combs shows that, in the absence of absolute position prior, although the physical ranging accuracy meets the sub-micron level requirements, the measurement results exhibit periodic non-ambiguous range jumps due to the limitation of repetition frequency, reflecting the range limitations of a single type of sensor. Figure 3 The error convergence curve of the fusion ranging system of the present invention shows that after "coarse and fine coupling" and dynamic correction, the ranging residual smoothly converges to near zero without any fuzzy breakpoints, verifying the system's full-range high-precision measurement capability across scales. Figure 4 The flowchart illustrates the working principle of laser tracker and dual optical comb collaborative ranging, and elaborates on the physical implementation steps from dual optical comb light source parameter optimization, coaxial optical path guidance, multi-heterodyne beat frequency acquisition to integer ambiguity resolution, demonstrating the cross-scale collaborative logic of the hardware system. Figure 5 This is a flowchart of data processing based on a deep hierarchical ranking network, which shows the process of rearranging the original interferometric data using the fault saliency classification and ranking module, and extracting enhanced collaborative features through a parallel convolutional neural network, thus verifying the effectiveness of the algorithm in real-time compensation for dynamic hysteresis errors.

[0205] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some or all of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. A dynamic weld seam detection system for a flexible production line robotic arm integrating a laser tracker and a dual photoelectric frequency comb, characterized in that, The system specifically includes: The dual-comb ranging subsystem is used to establish an evolution model of the pulse electric field evolution process including the signal laser and the local oscillator laser. The influence of repetition frequency difference and carrier envelope offset frequency on ranging accuracy is analyzed by numerical simulation. By searching for the discrete optimal interval, it is ensured that the sampled spectrum is in the baseband range and does not undergo phase distortion, thereby generating a high-steady-state ultrashort pulse sequence that satisfies the bandpass sampling law. The laser tracker and dual-comb common-path emission subsystem are used to divide the beam into a reference optical path and a measurement optical path using a beam splitter, and to establish a reference reference for eliminating drift. The measurement beam is injected into the collimated emission axis of the laser tracker through a coaxial coupling element, and the tracker's servo turntable guides the beam toward the target ball at the end of the robotic arm, constructing a reciprocating measurement channel containing spatial time-of-flight information, and realizing the spatial physical alignment of the measurement optical axis and the robotic arm's motion vector. The fault significance-based hierarchical sorting module extracts signal morphology indicators of interference fringes using global max pooling and average pooling operators; the original fringe sequence is hierarchically rearranged using a fast non-dominated sorting algorithm to aggregate scattered key features in the time dimension, thereby eliminating nonlinear lag and revealing the reinforcing synergistic effect between data. A parallel convolutional neural network feature extraction module is used to implement dynamic error correction based on a deep hierarchical sorting network. The parallel network extracts the temporal state features of the original sequence and the enhanced collaborative features of the rearranged sequence, respectively. After weighted fusion, these features are mapped to precise absolute distance prediction values ​​through a fully connected regression layer. The precise absolute distance prediction values ​​are combined with the instantaneous azimuth and pitch angles of the laser tracker to calculate the three-dimensional spatial coordinates. The spatial position deviation is extracted by comparing the actual trajectory of the robotic arm end effector with the theoretical weld trajectory, and the final dynamic detection result of the weld is output.

2. The flexible production line robotic arm weld dynamic detection system integrating a laser tracker and a dual photoelectric frequency comb as described in claim 1, characterized in that, In the evolution model, a mathematical expression for the electric field of a single pulse is established, and a time-domain superposition model of an infinite pulse sequence is defined to obtain the electric field representation of a laser pulse train with a repetitive period. The carrier envelope offset frequency is determined based on the carrier envelope phase slip between adjacent pulses.

3. The flexible production line robotic arm weld dynamic detection system integrating a laser tracker and a dual photoelectric frequency comb as described in claim 1, characterized in that, In the evolution model, a coaxial multiheterodyne interference sampling model is constructed. First, the repetition frequency difference between the signal laser and the local oscillator laser is defined. Within the update period, the multiheterodyne interference signal contains two sets of cross-correlation fringes: a reference set and a measurement set. Its time-domain waveform is described as the cross-correlation convolution of the signal light field and the local oscillator light field. The flight time delay generated by the measurement optical path is linearly mapped to the radio frequency time domain, forming cross-correlation interference fringes.

4. The flexible production line robotic arm weld dynamic detection system integrating a laser tracker and a dual photoelectric frequency comb as described in claim 3, characterized in that, The system performs multi-heterodyne interferometric signal acquisition, receiving the coherent superposition signal of the local oscillation pulse and the return signal pulse through a photodetector; it uses the local oscillation pulse to perform linear optical sampling on the measurement pulse carrying displacement characteristics, and utilizes the slip effect generated by different repetition frequencies to generate a multidimensional original cross-correlation interferometric fringe sequence containing the dynamic displacement information of the robotic arm within the update cycle.

5. The flexible production line robotic arm weld dynamic detection system integrating a laser tracker and a dual photoelectric frequency comb as described in claim 1, characterized in that, Based on the multiheterodyne beat frequency signal set and the repetition frequency difference between the laser and the local oscillator laser, the optimal interval of signal discretization is searched through numerical simulation to generate the original interference fringe sequence that satisfies the constraints and has no phase distortion.

6. The flexible production line robotic arm weld dynamic detection system integrating a laser tracker and a dual photoelectric frequency comb according to claim 1, characterized in that, In the laser tracker and dual-comb common-path emission subsystem, an optical beam-splitting model of the reference arm and the measurement arm is constructed. The beam is divided into a reference beam and a measurement beam according to the energy splitting ratio using a beam splitter. The reference beam reaches the detector after traveling a fixed path length, and its electric field model is represented as the time shift of the signal source electric field. Using the servo turntable of the laser tracker, the measurement beam is injected into the tracker's emission optical axis through a coaxial coupling element, ensuring that the optical axis center of the measurement beam coincides with the aiming line of sight of the tracker. This allows the dual-comb measurement beam to follow the dynamic guidance of the tracker and point in real time at the target ball at the end of the robotic arm. Based on the instantaneous distance between the target ball at the end of the robotic arm and the optical center of the tracker, the round-trip propagation path length of the measurement beam in space is obtained, and the dynamic optical path difference between the measurement arm and the reference arm is calculated. This optical path difference is linearly mapped into the time delay of the interference fringes in the subsequent photoelectric detection stage.

7. The flexible production line robotic arm weld dynamic detection system integrating a laser tracker and a dual photoelectric frequency comb as described in claim 1, characterized in that, Based on the repetition frequency and carrier envelope offset frequency of the local oscillator and signal laser, the frequency distribution of the multiheterodyne beat frequency signal in the radio frequency domain is obtained. Through photoelectric conversion, the optical frequency difference in the terahertz band is linearly reduced and mapped to the radio frequency signal set in the megahertz band. The output signal of the photodetector is spectrally filtered based on a bandpass filter to ensure that the filtered multiheterodyne beat frequency signal set is strictly located within the first Nyquist baseband interval and does not overlap with the zero frequency or half-repetition frequency boundary, thereby ensuring the integrity of the time domain waveform. Using an analog-to-digital converter with the repetition frequency of a local oscillating laser as the clock reference, the filtered analog signal is synchronously digitally acquired to generate a discretized multidimensional original cross-correlation interference fringe sequence. This sequence is directly used as the input data for the subsequent deep hierarchical sorting network to solve the flight time delay of the robotic arm end effector.

8. The flexible production line robotic arm weld dynamic detection system integrating a laser tracker and a dual photoelectric frequency comb according to claim 1, characterized in that, The global maximum and mean of the original sequence are extracted using a dual-channel pooling layer as signal morphology indicators, and the samples are rearranged using a fast non-dominated sorting algorithm to eliminate signal hysteresis caused by the tracker servo response.

9. The flexible production line robotic arm weld dynamic detection system integrating a laser tracker and a dual photoelectric frequency comb according to claim 1, characterized in that, The global maximum and global average values ​​are calculated based on the signal morphology index vector of the original sequence through a dual-channel pooling layer; an index sequence is generated using a fast non-dominated sorting algorithm, and the original sequence is rearranged into hierarchical samples; temporal state features and enhanced collaborative features are extracted based on a parallel convolutional neural network, and the absolute distance is output through a regression layer after weighted fusion.

10. The flexible production line robotic arm weld dynamic detection system integrating a laser tracker and a dual photoelectric frequency comb according to claim 1, characterized in that, A dual-channel pooling layer is used to extract the morphological features of the signal at each sampling time. Global max pooling is used to capture the signal abrupt change features, and global average pooling is used to characterize the overall signal intensity. A fast non-dominated sorting algorithm is used to perform multi-objective hierarchical division of the morphological feature indicators to generate an index sequence that reflects the significant differences in the signal. Based on the index sequence, the multidimensional original interference fringe sequence is rearranged in the time dimension to generate hierarchical samples that eliminate time lag and have enhanced synergistic effects. The input is a one-dimensional convolutional neural network with two parallel branches to process the time state features and synergistic features respectively. The two features are cascaded and fused, and mapped to the precise absolute distance prediction value of the robotic arm end effector through a fully connected regression layer.

Images