Internal light path five-axis linkage laser cutting control method based on industrial control

Abstract

The invention relates to the technical field of industrial control, and discloses an internal light path five-axis linkage laser cutting control method based on industrial control. The method comprises the following steps: determining a time anchor point in a five-axis motion data stream according to a laser cutting start-stop event triggering moment, extracting frequency band energy distribution of each axial displacement sensor in a preset window before and after the anchor point, and constructing an inner light path feature set; inputting the feature set into a state comparison network, and generating a comparison representation of an actual path and a standard path through comparative learning; the machining frequency serves as a tool variable, analysis and comparison characterization are conducted through a causal reasoning model, and the contribution degree of each sensor channel to the cutting precision is obtained; fusing the dynamic optical path stability index and the path offset index to adjust a screening threshold, and screening out high-contribution pure optical path characterization; and pure light path representation time domain grouping is carried out according to a cutting waveform phase, frequency domain weighting is carried out by taking frequency band energy as an attention weight, and a final control instruction is generated through layered attention fusion.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of industrial control, in particular to an inner light path five-axis linkage laser cutting control method based on industrial control. BACKGROUND

[0002] In the field of industrial manufacturing, laser cutting technology is widely used in metal processing, aerospace parts manufacturing and other complex scenarios due to its high precision and high speed. With the continuous upgrading of processing needs, five-axis linkage laser cutting has become an important means to cope with high-precision manufacturing tasks because it can realize the processing of complex curved surfaces in space. However, in the actual processing process, the stability of the inner light path directly affects the precision and quality of laser cutting, which is affected by mechanical vibration, environmental temperature fluctuation, five-axis motion trajectory coupling and other factors, and is prone to light path deviation, resulting in cutting errors. Existing laser cutting control methods focus more on the planning and control of motion trajectories, and lack sufficient perception and adjustment of light path dynamic changes. Some methods try to achieve light path monitoring by setting sensors in the light path to collect light intensity, phase and other information, but due to the strong coupling relationship between the displacement of each axis and the light path state during five-axis motion, single sensor data cannot fully reflect the real state of the light path, and misjudgment or omission is easy to occur. At the same time, in traditional control methods, the distinction between cutting active and inactive states is not accurate enough, and there is a lack of comparative analysis of the processing path under different states, which makes it impossible to effectively identify the path deviation caused by changes in the light path. In terms of multi-sensor data fusion, existing technologies often use simple weighting or feature splicing methods, which do not fully consider the contribution differences of each sensor channel to cutting accuracy, and are prone to introduce redundant information, affecting the generation efficiency and accuracy of control commands. When the number of processing times accumulates, the stability of the light path system will show dynamic changes, and existing methods lack the ability to analyze the causes of such dynamic changes, making it difficult to trace the error source, resulting in the adjustment of control parameters lagging behind the changes in the actual light path state, and further affecting the consistency of cutting quality. SUMMARY

[0003] The purpose of the present application is to provide an inner light path five-axis linkage laser cutting control method based on industrial control to solve the problems raised in the background.

[0004] To achieve the above purpose, the present application provides an inner light path five-axis linkage laser cutting control method based on industrial control, which comprises: Based on the triggering time of the laser cutting start-stop event, the time anchor point of the cutting active period is determined in the five-axis motion data stream, and the frequency band energy distribution of each axial displacement sensor in the pre-set window before and after the time anchor point is extracted, and an inner light path feature set is constructed; inputting the internal light path feature set into a state comparison network, calculating the similarity of the same machining path in the cutting active state and the cutting inactive state through comparative learning, and generating a comparison representation of the actual machining path and the standard path; using a causal inference model to perform counterfactual analysis on the comparison representation with the number of machining times as a tool variable, and obtaining the channel contribution degree of each sensor channel to the cutting accuracy; fusing a dynamic light path stability index and a path offset index, inputting a pre-trained model to adaptively adjust a filtering threshold, and filtering out a pure light path representation with a channel contribution degree higher than the adjusted threshold; grouping the pure light path representation in the time domain according to the phase of the cutting waveform, implementing frequency domain weighting with the frequency band energy as the attention weight, and generating a final control instruction through hierarchical attention fusion.

[0005] Preferably, the time anchor point of the cutting active period includes: analyzing the triggering time of the laser cutting start-stop event, combining the displacement rate in the five-axis motion data stream, and locating the cutting state switching time; extracting five-axis displacement data in a preset window before and after the time, performing fast Fourier transform to obtain a frequency domain signal and dividing it into multiple sub-bands; calculating the normalized energy value of each sub-band of each axis, and arranging it in the axial order to construct the internal light path feature set.

[0006] Preferably, the state comparison network includes a double-branch structure with parameter sharing, which processes the internal light path feature set in the active state and the inactive state respectively; Each branch extracts local frequency domain features through a convolution layer, and generates path potential representation after compression by a pooling layer; calculate the similarity score of the path potential representation in the active state and the inactive state, and generate the comparison representation of the quantitative cutting state influence through difference operation.

[0007] Preferably, the number of machining times is divided into three levels of low frequency, medium frequency and high frequency by counting the number of cutting start-stop events per unit time; The causal inference model takes the comparison representation as a feature vector, the cutting state as a processing variable, and the cutting accuracy as a result variable; Through counterfactual analysis, the prediction deviation of each sample in the cutting active state and the cutting inactive state is calculated, and the channel contribution degree of each sensor channel is aggregated.

[0008] Preferably, the dynamic light path stability index is obtained by extracting the low frequency band energy of the internal light path feature set, and weighting the energy proportion of each layer after wavelet packet decomposition; The path offset index is obtained by calculating the product of the machining path alignment cost and the spatial correlation coefficient in the cutting active state and the cutting inactive state.

[0009] Preferably, the dynamic light path stability index, path offset index and channel contribution degree are combined into a comprehensive vector input into a Transformer model; The basic threshold value is dynamically adjusted based on the light path quality score output by the Transformer model; Channels with a channel contribution degree higher than the adjusted threshold value are screened, and their comparison representation is extracted to generate the pure light path representation.

[0010] Preferably, the time domain grouping according to the cutting waveform phase includes extracting the phase period from the laser waveform signal, and dividing the pure light path representation into multiple phase interval groups; Each group of pure light path representations is subjected to frequency domain transformation, and weighted processing is performed using normalized frequency band energy, and after inverse transformation, a weighted light path representation is formed.

[0011] Preferably, the on-site machining elements are extracted based on a machining feature recognition model, and the current machining scene is determined according to a preset machining scene knowledge graph; The process specification file corresponding to the scene is called to divide the light path large-scale region and the light path small-scale region.

[0012] Preferably, the texture complexity of the light path large-scale region is analyzed, and when the local region index exceeds the threshold value, the light path small-scale sub-region is stripped out; A real-time feedback mechanism of the large-scale analysis model and the small-scale analysis model is established; Based on the preliminary result of the large-scale analysis model, the small-scale analysis model is triggered to perform fine checking of the light path parameters.

[0013] Preferably, a light path feature warehouse is constructed by integrating multi-source process data, and a light path feature matrix is generated by extracting feature vectors in each dimension; The linkage degree between features in each dimension is calculated to construct a dimension linkage graph; A light path process linkage graph is formed by cross-dimension connection through key feature linkage degrees; Based on the light path process linkage graph, a core light path feature sequence is screened to drive the generation of the final control instruction.

[0014] Compared with the prior art, the present application has the following advantages: By determining the time anchor point based on the triggering moment of the laser cutting start-stop event, the frequency band energy distribution of each axial displacement sensor is extracted to construct an internal light path feature set, which can accurately capture the associated information of the light path and the motion state in the cutting process, and realize detailed perception of the dynamic change of the light path. This feature extraction method based on the time anchor point effectively associates the five-axis motion data stream with the light path state data, overcoming the problem of separation of motion and light path information in traditional methods, and making the information in the feature set more consistent with the light path change law in the actual machining scene. The similarity calculation of the same machining path in the active and inactive states is introduced into the state comparison network to cut, and the generated comparison representation can clearly reflect the differences of the path in different states, providing an intuitive basis for identifying the deviation caused by the change of the light path. This way of comparison learning avoids the one-sidedness that may occur in single-state analysis, making the judgment of path deviation more objective and accurate, which helps to discover light path abnormalities in complex machining processes in a timely manner. With the number of machining times as the tool variable, the counterfactual analysis of the on and off representations is performed using a causal reasoning model, which can deeply explore the internal relationship between each sensor channel and the cutting accuracy, and clearly determine the contribution of each channel. This process breaks free from the limitations of traditional multi-sensor fusion, which does not distinguish the importance of channels, highlights the role of key channels, reduces the interference of redundant information, makes subsequent data processing more targeted, and improves the accuracy of light path state judgment. The dynamic light path stability index and path deviation index are fused to adaptively adjust the filtering threshold, and the pure light path representation with high contribution degree is selected, further optimizing the quality of feature data. This dynamic threshold adjustment mechanism can flexibly adjust the filtering standard according to the real-time state of the light path, ensuring that the retained feature information is both comprehensive and core, laying a reliable foundation for the generation of subsequent control instructions. According to the time domain grouping based on the cutting waveform phase, and the frequency domain weighting is implemented with the frequency band energy as the attention weight, the final control instruction is generated through hierarchical attention fusion, which fully combines the feature advantages of time domain and frequency domain. Time domain grouping effectively distinguishes the light path features at different phases, and frequency domain weighting highlights the energy information of key frequency bands. The hierarchical attention fusion makes the integration of features more hierarchical and targeted, and the generated control instruction can more accurately adapt to the actual light path state, which helps to better cope with the dynamic changes of the light path in the five-axis linkage laser cutting process, and ensures the stability of the cutting process and the consistency of the cutting results. BRIEF DESCRIPTION OF DRAWINGS

[0015] Figure 1 The working principle diagram of the industrial control-based internal light path five-axis linkage laser cutting control method described in the present application; Figure 2 The flowchart for determining the time anchor point and constructing the feature set; Figure 3 The flowchart for calculating the channel contribution degree based on causal reasoning; Figure 4 The flowchart for screening pure light path representation; Figure 5 The flowchart for machining scene recognition and region division. DETAILED DESCRIPTION

[0016] With reference to the drawings of the embodiments of the present application, the technical solutions in the embodiments of the present application will be clearly and completely described. Obviously, the described embodiments are only a part of the embodiments of the present application, rather than all the embodiments of the present application. Based on the embodiments of the present application, all other embodiments obtained by a person of ordinary skill in the art without creative work fall within the protection scope of the present application.

[0017] With reference to Figure 1 The present application provides an internal optical path five-axis linkage laser cutting control method based on industrial control, which comprises the following steps: The time anchor point of the cutting activation period is determined by the triggering time of the laser cutting start-stop event. The frequency band energy distribution of each axial displacement sensor within the preset window before and after the time anchor point is extracted from the five-axis motion data stream. An internal optical path feature set containing multi-axis multi-frequency band features is constructed. The state comparison network adopts a double-branch structure to process the feature sets of the activated and non-activated states. Local frequency domain features are extracted through a convolution layer and similarity is calculated to generate a comparison representation of the actual processing path and the standard path. The processing frequency is used as a tool variable. The influence degree of each sensor channel on the cutting accuracy in the comparison representation is analyzed by a causal reasoning model to obtain the channel contribution degree. The dynamic optical path stability index and the path deviation index are input into a pre-trained model to dynamically adjust the screening threshold. The high-contribution-degree channels are selected to form a pure optical path representation. The pure optical path representation is grouped in the time domain according to the phase of the laser waveform. The frequency band energy is used as the attention weight to implement frequency domain weighting. Finally, the control instruction is generated through hierarchical attention fusion.

[0018] Embodiment 1: With reference to Figure 2 The present application provides an internal optical path five-axis linkage laser cutting control method based on industrial control, which comprises the following steps:

[0019] In the positioning process of the time anchor point of the cutting activation period, the system first monitors the power output signal of the laser. When the power value is detected to jump from the standby threshold to the working threshold, the synchronous data acquisition of the five-axis displacement sensor is triggered. After the acceleration data of the displacement sensor is preprocessed, the second derivative of the displacement change rate of each axis is calculated by using a sliding window. The window width is set to 50 milliseconds and the step is 10 milliseconds. By finding the local extreme points of the second derivatives of each axis, different weight coefficients are given according to the inertia differences of each axis. Finally, the weighted median is calculated as the initial anchor point position. After the initial anchor point is determined, the system expands a time window of 200 milliseconds before and after it. The window range can cover the stable rising stage of the laser power and the response delay of the mechanical system.

[0020] The displacement data within the time window is resampled, raising the original sampling rate to a uniform 1 kHz to ensure time alignment across all axis data. The resampling process utilizes a cubic spline interpolation algorithm to maintain the signal waveform characteristics while eliminating phase deviations caused by sampling rate differences. A 4096-point fast Fourier transform is applied to the resampled data, and the spectral analysis range is limited to 0-500 Hz, which encompasses the primary mechanical vibration characteristics of the five-axis linkage system. The frequency domain signal is divided into 10 equal-width subbands, each with a width of 50 Hz. When calculating the energy of each subband, a modified Hanning window function is used to suppress spectral leakage. The window function parameters are dynamically adjusted based on the mechanical characteristics of each axis. A 5% overlap is established between adjacent subbands, and the energy in the overlapping region is distributed using a cosine function to ensure continuity in energy calculation across bands.

[0021] The frequency band energy values ​​for each axis are normalized, and a dynamic range compression algorithm is used to map the energy values ​​to the 0-1 range. This normalization process preserves the relative proportions of the energy in each frequency band while suppressing the influence of outliers. The resulting internal optical path feature set is arranged in a fixed order along the X / Y / Z / A / C axes, containing 10 frequency band energy values ​​for each axis, forming a 50-dimensional feature vector. The feature vector is supplemented with a timestamp and processing status marker for subsequent status comparison and analysis.

[0022] The state comparison network adopts a parameter-sharing dual-branch architecture. The two branches have identical network structures but process input data in different states. The first convolutional layer of the network is configured with 64 filters of width 5. The convolution operation is performed along the frequency band dimension, focusing on extracting the fundamental frequency component and its harmonic features. The second convolutional layer uses 128 filters of width 3 to capture the coupling relationship between frequency bands at a higher level. The third convolutional layer uses 256 filters of width 1 to achieve global fusion of cross-band features. Each convolutional layer is followed by a maximum pooling operation with a pooling window width of 4 and a stride of 4, effectively compressing the feature dimension while retaining significant features.

[0023] A contrastive loss function is used during network training to enable the network to learn to distinguish the differences between the features of the active and inactive states. The outputs of the two branches are similarly calculated in the feature fusion layer, using a modified Euclidean distance metric that assigns higher weights to high-frequency features. The similarity score is converted to a probability value in the range of 0-1 using a sigmoid function, indicating the degree of consistency of the path features in the two states. The comparison representation is generated using a feature difference calculation method, where the feature vectors of the active and inactive states are element-by-element subtracted. The difference result is then weighted and fused with the similarity score. The weighting coefficient is dynamically adjusted based on the signal-to-noise ratio of each frequency band to ensure that high-reliability features dominate the final comparison representation.

[0024] Network parameters are updated using an adaptive moment estimation optimization algorithm, with the learning rate dynamically decaying based on training progress. The batch size is set to 256, and the number of training epochs is determined using early stopping based on validation set performance. To prevent overfitting, a random dropout mechanism is introduced into the network, with a dropout rate of 20% for hidden layers. The training data undergoes rigorous augmentation processing, including the addition of Gaussian noise, random band masking, and timing jittering, to improve the model's generalization capabilities.

[0025] During actual deployment, the system monitors changes in laser power in real time, triggering the time-anchor positioning process. The collected displacement data is preprocessed and a fixed process is used to construct an internal optical path feature set. This feature set is input into a pretrained state comparison network to generate a real-time comparison representation. This comparison representation is then fed into a subsequent processing module for cutting accuracy analysis and control command generation. The latency of the entire processing chain is kept below 10 milliseconds, meeting the real-time requirements of industrial control systems.

[0026] This implementation achieves precise extraction and state comparison of internal optical path features during five-axis laser cutting through a carefully designed signal processing process and deep neural network architecture. The system can sensitively capture changes in mechanical motion states, providing reliable feature representation for subsequent machining quality analysis and control decisions. The network's parameter sharing mechanism effectively reduces model complexity, enabling the system to operate efficiently in industrial control scenarios. Multi-level frequency domain analysis and feature fusion strategies ensure the system's adaptability to diverse machining conditions.

[0027] Example 2: See Figure 3 This paper describes a hierarchical statistical mechanism for processing times and the application of a causal inference model in an industrial control-based internal optical path five-axis laser cutting control method, focusing on the calculation of channel contributions. This implementation establishes a framework combining statistical analysis and causal inference to quantitatively evaluate the actual impact of each sensor channel on cutting accuracy.

[0028] The hierarchical statistics of the number of processing times are implemented using a dynamic sliding window method. The system maintains a time window of 300 seconds in length and counts the number of laser cutting start and stop events that occur within the window in real time. Event detection is based on the rising and falling edge triggers of the laser power signal, and a 5-millisecond debounce delay is set to prevent false counting. The statistical results are updated once a minute and divided into three levels according to the preset threshold: the low-frequency level corresponds to 0 to 2 cutting operations per minute, the medium-frequency level corresponds to 2 to 5 times per minute, and the high-frequency level corresponds to more than 5 times per minute. The grading information is stored in the form of unique hot encoding, and a timestamp is attached for subsequent time alignment processing. In the data preprocessing stage, the system performs continuity verification on the start and stop events in the continuous processing cycle to exclude isolated events caused by system abnormalities.

[0029] The causal inference model is constructed as a fully connected neural network with three hidden layers. The input layer receives a 256-dimensional alignment representation vector generated by the previous state alignment network. The first hidden layer contains 128 neurons with a leaky rectified linear unit activation function; the second hidden layer contains 64 neurons with a hyperbolic tangent activation function; and the third hidden layer is configured with 32 neurons with a linear activation function. The output layer is a single neuron structure that predicts the cutting accuracy deviation value. The model training uses the mean square error loss function, and the optimization process uses the stochastic gradient descent algorithm with Nesterov momentum. The training data set includes alignment representations, cutting state labels, and measured accuracy deviation values from historical processing records.

[0030] The encoding of the processing variables uses a binary vector representation, with the cutting activation state labeled as [1, 0] and the non-activation state labeled as [0, 1]. During the counterfactual analysis, for each sensor channel, the feature values of other channels remain unchanged, and only the state label of the target channel is changed. The model predicts the cutting accuracy deviation of the target channel under two hypothetical states, and calculates the absolute difference between the predicted results as the initial contribution of the channel. The analysis process is repeated at multiple sampling points in the feature space to improve the stability of the estimate through the Monte Carlo method.

[0031] The calculation of channel contribution includes time dimension smoothing processing. The system maintains a sliding window with a length of 30 processing periods, and performs cubic spline interpolation on the initial contribution sequence of each channel in the window. The interpolation node interval is set to 100 milliseconds, generating a continuous time function curve. The curve is numerically integrated within a predetermined time interval, and the integral result is divided by the length of the time interval to obtain the average contribution. The final channel contribution is standardized by Z-score, converted to a distribution form with mean zero and standard deviation one, facilitating cross-channel comparison.

[0032] The model deployment stage uses an online learning mechanism. Every time 100 cutting operations are completed, the system automatically collects new processing data and performs incremental training on the causal inference model. The selection of training data uses an importance sampling strategy, preferentially retaining processing period samples with significant changes in channel contribution. To prevent model drift, a model version management mechanism is set up, and when the prediction error of three consecutive versions exceeds the threshold, the model rollback operation is triggered.

[0033] During real-time operation, the system continuously receives the alignment representation stream output by the state alignment network. At the end of each processing period, the causal inference process is triggered: first, determine the classification label according to the number of processing times in the current time window; then input the alignment representation and the classification label into the causal inference model; finally, perform counterfactual analysis to calculate the contribution of each channel. The calculation results are stored in a ring buffer for use by the subsequent pure light path representation filtering module.

[0034] This embodiment realizes the fine evaluation of the contribution of each component in the multi-axis system by establishing a correlation model between the processing frequency and the influence of the sensor channel. The hierarchical statistical mechanism considers the differentiated influence of different processing rhythms on the system state, making the analysis results more consistent with the actual working conditions. The structure design of the neural network model balances the feature extraction capability and the computing efficiency, adapting to the real-time requirements of industrial control scenarios. The counterfactual analysis framework effectively separates the independent influence of the sensor channel, avoiding the confusion bias in traditional correlation analysis methods. The smoothing processing in the time dimension enhances the stability of the contribution index, providing reliable decision basis for the light path control.

[0035] Example 3: refer to Figure 4 , which relates to the calculation process of the dynamic light path stability index and the path offset index in the internal light path five-axis linkage laser cutting control method based on industrial control, as well as the threshold adjustment and pure light path representation screening mechanism based on the Transformer model. This embodiment realizes the dynamic evaluation and optimal selection of the light path quality in the laser cutting process through multi-dimensional feature fusion and adaptive decision-making.

[0036] The calculation of the dynamic light path stability index starts with the extraction of the low-frequency energy in the internal light path feature set. The system selects 5 sub-band energy values in the 0-50Hz range from the 50-dimensional internal light path feature vector to form a low-frequency feature vector. A 5-layer wavelet packet decomposition is performed on this vector, and a db4 wavelet basis function is used for time-frequency analysis. During the decomposition process, the wavelet coefficient information entropy of each node is calculated by the following formula: Where: represents the entropy value of the th node, is the normalized energy proportion of the th wavelet coefficient of the th node, is the total number of wavelet coefficients. The weight coefficients of each decomposition layer are the inverses of the corresponding node entropy values, which are normalized by softmax to form the hierarchical weights. The final dynamic light path stability index is calculated as the logarithmic transformation value of the sum of the weighted energies of each layer, which reflects the energy distribution uniformity of the light path system in the low-frequency band.

[0037] The generation of path deviation index contains two parallel computing processes. The alignment cost calculation adopts an improved dynamic time warping algorithm, which represents the machining path in the active state and the inactive state as a sequence of three-dimensional space trajectory points. The algorithm introduces an axial weight matrix, which gives different alignment cost coefficients according to the motion characteristics of different mechanical axes. The spatial correlation coefficient calculation adopts a multi-scale analysis method, which calculates the Pearson correlation coefficient at three spatial scales of 1mm, 5mm and 10mm respectively, and takes the weighted average value as the final similarity measure. The path deviation index is finally represented as the product of the alignment cost and the spatial correlation coefficient, and is compressed to the range of 0-1 by the hyperbolic tangent function.

[0038] The input sequence of the Transformer model is constructed as a comprehensive vector of 20 historical time steps. Each comprehensive vector contains three components: dynamic light path stability index, path deviation index, and channel contribution degree. Position encoding is implemented using a trainable parameter matrix, with dimensions matching the input feature dimensions. The self-attention mechanism is configured with 4 independent heads, and the key, query, and value matrix dimensions of each head are set to 64. The feedforward network contains two fully connected layers with a hidden layer dimension of 256, using Gaussian error linear unit activation function. The light path quality score output layer adopts a double-channel structure, which respectively predicts stability score and deviation score , and the final score is calculated as: where: and are trainable weights, is the bias term, represents the S-shaped function. The score output range is between 0-1, which is directly converted into a threshold adjustment coefficient acting on the basic threshold value.

[0039] The basic threshold value is initialized as the median of all channel contribution degrees. The adjustment coefficient acting range is limited within ±30% to prevent extreme threshold jumps. The selection of pure light path representation implements an iterative optimization process: the first round of selection selects channels with contribution degree higher than the current threshold value; calculates the statistical quantity of the contribution degree distribution of the remaining channels; if the new median exceeds the original median by 5%, update the threshold value for the next round of selection. The iterative process is executed at most three times, and the final reserved channels constitute the pure light path representation.

[0040] In the real-time running phase, the system maintains a length-20 comprehensive vector ring buffer. Whenever a new machining cycle is completed, the dynamic light path stability index and path offset index at the current time are calculated, together with the channel contribution degree, to form a new comprehensive vector update buffer. The Transformer model performs forward calculation every 5 machining cycles, outputs the latest light path quality score, and updates the screening threshold. After the threshold adjustment signal is triggered, the pure light path representation screening module immediately performs channel selection operations to generate a new feature subset for subsequent processing.

[0041] This embodiment constructs an index system for comprehensive evaluation of light path stability by combining wavelet packet analysis with multi-scale space correlation calculation. The dynamic weight distribution mechanism enables the system to adaptively focus on key feature components under different working conditions. The time series processing capability of the Transformer model effectively captures the evolution law of the light path state, and the score output and threshold adjustment form a closed-loop control. The iterative screening strategy ensures feature quality while avoiding excessive removal of useful information, maintaining the robustness of the control system. The modular design of the entire processing flow supports parallel computing, meeting the real-time requirements of high-throughput industrial scenarios.

[0042] Embodiment 4: relates to the extraction of cutting waveform phase and the time-domain grouping processing mechanism of pure light path representation in the industrial control-based internal light path five-axis linkage laser cutting control method. This embodiment realizes fine modeling of dynamic characteristics in the laser cutting process through phase synchronization analysis and frequency domain weighting fusion. The following combines waveform data examples in specific machining scenarios to explain the specific implementation process of this embodiment in detail.

[0043] In a curve cutting operation of stainless steel plate, the laser power sensor collected a 12-second waveform signal. The signal sampling rate is 10 kHz, containing 3 complete cutting cycles. The system first performs Hilbert transform on the analog signal, constructs the analytic signal, and then calculates the instantaneous phase. The phase zero-crossing detection algorithm identifies 6 key time points, dividing each cutting cycle into 8 equal phase intervals, each interval corresponding to a 45-degree phase span. The phase interval division results and corresponding time-domain features of the first cutting cycle are shown in Table 1.

[0044] Table 1: Phase interval division example of the first cutting cycle Interval number Start time (ms) End time (ms) Phase range (degrees) Average power (W) Energy percentage (%) 1 0 31 0-45 852 10.2 2 31 62 45-90 1265 15.8 3 62 93 90-135 1540 19.3 4 93 124 135-180 1620 20.1 5 124 155 180-225 1580 19.7 6 155 186 225-270 1420 17.5 7 186 217 270-315 980 12.4 8 217 248 315-360 720 8.6

[0045] The time-domain packet of the pure light path characterization is implemented by the phase synchronization window method. Taking the 12 sensor channels screened by the previous stage module as an example, the system expands a window range of ±22.5 degrees to intercept the feature data with the midpoint of each phase interval as the center. For interval 3 (90-135 degrees), the actual processing window is 78.75-146.25 degrees, corresponding to a time window of 67-101 ms. After the displacement sensor data of each channel in the window is resampled and aligned, a 12-channel x 50-dimensional phase feature matrix is constructed.

[0046] The frequency domain weighting processing is independently implemented within the phase group. Taking the channel 7 data of interval 3 as an example, the system first performs a short-time Fourier transform on the 50-dimensional feature vector of this channel to obtain the complex spectrum of 20 sub-bands. The energy calculation uses the modulus square operation, and the energy values of the 20 sub-bands constitute the reference vector. The global average energy is calculated by counting the energy of the same frequency band of all phase intervals and all channels. The energy value of the 12th sub-band (corresponding to 275-300 Hz) of interval 3 channel 7 is 0.85, while the global average energy of this frequency band is 0.62, so its weight coefficient is calculated as 1.37. The weighting process preserves the phase information of the spectrum and only scales the amplitude.

[0047] The hierarchical attention fusion adopts a two-level structure. The first level processing is performed within a single phase group, and the weighted features of the above 12 channels are used as input. The feature vector of each channel generates a query, key, and value triple through linear transformation, and the attention weight between channels is calculated. The second level processing is implemented across phase groups, and the fusion results of the 8 phase groups are arranged in time sequence to capture the dynamic pattern within the period through the time attention mechanism. The attention calculation uses the scaled dot product method, and the temperature coefficient is set to the square root of the feature dimension.

[0048] In terms of real-time implementation, the system maintains a phase tracking state machine. When a new laser period is detected, the phase counter is initialized and the timing interrupt is started. The interrupt service routine is triggered every 5.6 ms (corresponding to a 45-degree phase change), which performs the following operations: reads the instantaneous value of each sensor at the current time; linearly interpolates the previous time data to obtain accurate phase-aligned sampling; updates the phase group data in the ring buffer. When 360-degree phase accumulation is completed, the frequency domain weighting and attention fusion processes are automatically triggered.

[0049] This implementation shows specific advantages in a certain automobile parts cutting production line. When dealing with door beam parts containing complex curves, the phase grouping mechanism effectively distinguishes the cutting characteristics of straight segments and circular arc segments. It is found in the monitoring data that interval 4 (135-180 degrees) usually corresponds to the curve cutting stage, and the frequency band energy weight of channel 3 (Z axis) in this interval is continuously higher than that in other intervals by more than 15%. This phase-related feature change pattern is accurately captured by the attention mechanism and is specifically processed in the subsequent control instruction generation.

[0050] When the system is implemented, a three-stage pipeline architecture is adopted to improve processing efficiency. The first stage of the pipeline is responsible for real-time phase tracking and data acquisition, the second stage performs frequency domain transformation and weighting calculation, and the third stage processes attention fusion and result output. The pipeline stages exchange data through a double buffering mechanism to avoid processing delay accumulation. Experimental observations show that this architecture can maintain a control frequency of 1 kHz while stabilizing the processing delay within 3 ms.

[0051] The parameterized design of the phase grouping strategy supports flexible adjustment. Users can customize the number of phase intervals (4 / 8 / 16 partitions are available) and the window expansion coefficient (the default ±22.5 degrees can be adjusted by ±5 degrees) according to the characteristics of the processed materials. In the aluminum cutting scenario, using a 16-partition mode can better capture high-frequency vibration characteristics; for thick steel plate cutting, reducing the window expansion coefficient helps to reduce the interference between adjacent phase intervals. This adaptability enables the system to optimize feature extraction for different processing needs.

[0052] Embodiment 5: see Figure 5 , which relates to the whole process implementation of processing scene recognition, light path region division, and multi-dimensional feature map construction in the internal light path five-axis linkage laser cutting control method based on industrial control. This implementation realizes adaptive modeling and control of light path characteristics in complex processing environments through a multi-modal data processing and hierarchical analysis architecture.

[0053] The improved ResNet-50 architecture is used as the basic framework for the processing feature recognition model. The input system is equipped with an industrial camera to collect RGB-D image data streams, with an image resolution of 1920x1080 pixels and a depth information accuracy of 0.1 millimeters. The first 49 layers of the network retain the original ResNet-50 structure, and the 50th layer is replaced with a fully connected layer with 128 outputs. The output vector is processed by L2 normalization. The pre-set processing scene knowledge graph is stored in a graph database, containing template vectors of twelve typical processing scenes. During real-time recognition, the normalized feature vector is calculated with the templates in the knowledge graph for cosine similarity, and when the maximum similarity exceeds 0.85, it is determined as a matching scene. For unmatched scenes, the system starts the online learning process, adds the new scene feature vector to the knowledge graph and associates it with the default process specification.

[0054] The retrieval of process specification files adopts a semantic-based vector matching mechanism. Each file in the specification library is converted into a 256-dimensional semantic vector by a word embedding model, and the Euclidean distance is calculated with the output vector of the machining feature recognition model in the same latent space. The three closest specification files enter the weighted voting stage, and the weights are dynamically adjusted by the historical call success rate of the files. The selected process specification file contains the parameter set of the optical path region division, where the large-scale region is defined as the geometric unit with a continuous area of more than 10% of the total machining area on the machining surface, and the small-scale region is the fine structure with a local feature size less than 5 mm in the large-scale region.

[0055] The optical path region division algorithm is implemented on the three-dimensional point cloud data of the machining surface. The point cloud data is generated by a line laser scanner, and the point spacing is set to 0.5 mm. The improved watershed algorithm first calculates the normal vector field of the point cloud, and uses the normal vector discontinuity point as the initial watershed. The curvature constraint condition is introduced in the region growing process, and the growth is terminated when the curvature difference between adjacent points exceeds 0.05. The large-scale region merging strategy adopts a hierarchical clustering method, and the spatial distance threshold is set to 3 mm and the normal vector angle threshold is set to 15 degrees. The texture complexity analysis is performed on the divided regions, and the local binary pattern histogram and the gray level co-occurrence matrix contrast in the 9x9 neighborhood window are calculated. When the texture entropy value of the local window exceeds the overall distribution mean value by 1.5 standard deviations, the region is marked as a high complexity region and triggers sub-region segmentation.

[0056] The large-scale analysis model adopts a three-dimensional convolutional neural network architecture, and the input data is a voxelized machining region spatial grid. The grid resolution is set to 0.2 mm, and each voxel contains attributes such as material type and surface roughness. The network contains four convolutional blocks, each consisting of a 3x3x3 convolutional layer, a batch normalization layer, and a leaky rectifier unit. The small-scale analysis model is designed based on point cloud Transformer, which processes high-complexity sub-regions stripped from large-scale regions. The point feature extraction layer uses dynamic graph convolution, and the attention mechanism uses local sparse attention mode. The two models are connected in real time through a shared 128-dimensional latent space. The bottleneck layer output of the large-scale model is used as the query vector of the small-scale model, and the output features of the small-scale model are injected back into the jump connection layer of the large-scale model.

[0057] The light path feature warehouse is constructed as a multi-modal graph database system. The time series dimension stores sensor sampling data with a sampling frequency of 1 kHz; the frequency spectrum dimension saves frequency domain features after fast Fourier transform; and the spatial coordinate dimension records three-dimensional position information. The construction of the dimension link graph first calculates the mutual information value between features in each dimension, and the features with mutual information exceeding 0.3 establish a connection edge. The key feature link degree is calculated by the node centrality score through the random walk algorithm, and the multi-dimensional scaling technology is used to project different dimensional features to a public two-dimensional space when connecting across dimensions, and the feature nodes with a projection distance less than 0.1 establish a cross-dimensional link.

[0058] The update of the light path process link graph adopts an incremental graph learning strategy. After completing 100 processing cycles, the system extracts the feature change pattern in the newly added processing data. The new node calculates the initial embedding through the graph convolution network, and the existing node embedding is updated in the neighborhood aggregation mode. The graph topology optimization is based on the modularity detection, and the community reorganization is triggered when the subgraph modularity exceeds 0.7. The selection of the core light path feature sequence implements a dynamic path search algorithm, and the node sequence corresponding to the maximum weight path is searched on the link graph with the channel contribution degree as the edge weight. The sequence drives the parameter configuration of the control instruction generation module, realizing the adaptive optimization of the processing process.

[0059] The embodiment shows adaptability in the scene of cutting the blade of an aero-engine. When processing a complex curved surface containing cooling holes, the system automatically identifies the hole area as a high-complexity sub-area and starts a small-scale analysis model to finely check the light path parameters. The hole processing history data recorded by the feature warehouse is associated to the current operation through the process link graph, and the core feature sequence guides the dynamic adjustment of the laser focal point position. The whole processing process is completed within the industrial control cycle, maintaining the balance between processing precision and efficiency.

[0060] It should be noted that, in the present text, relational terms such as first and second are used merely to distinguish one entity or action from another, without necessarily requiring or implying any such actual relationship or order between such entities or actions. Moreover, the terms "comprising", "containing" or any other variant thereof are intended to cover non-exclusive inclusion, so that a process, method, article or apparatus including a series of elements includes not only those elements, but also other elements not explicitly listed or inherent to such process, method, article or apparatus.

[0061] Although embodiments of the present application have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made therein without departing from the principles and spirit of the application, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A five-axis linkage laser cutting control method based on industrial control, characterized in that: include: Based on the triggering moment of the laser cutting start and stop event, the time anchor point of the cutting activation period is determined in the five-axis motion data stream, and the frequency band energy distribution of each axial displacement sensor in the preset window before and after the time anchor point is extracted to construct the internal optical path feature set; Inputting the internal optical path feature set into a state comparison network, calculating the similarity of the same processing path in the cutting activation state and the non-activation state through comparative learning, and generating a comparison representation of the actual processing path and the standard path; Taking the number of processing times as the instrumental variable, a causal reasoning model is used to conduct a counterfactual analysis on the comparison representation to obtain the channel contribution of each sensor channel to the cutting accuracy; The dynamic optical path stability index and path deviation index are integrated and input into the pre-trained model to adaptively adjust the screening threshold, screening out pure optical path representations with channel contributions higher than the adjusted threshold. The pure optical path representation is grouped in the time domain according to the phase of the cut waveform, and frequency domain weighting is performed using the frequency band energy as the attention weight, and the final control instruction is generated through hierarchical attention fusion.

2. The method for controlling an internal optical path five-axis linkage laser cutting based on industrial control according to claim 1, characterized in that: Determining the time anchor point of the cutting activation period includes: analyzing the triggering moment of the laser cutting start and stop event, combining the displacement change rate in the five-axis motion data stream, and locating the cutting state switching moment; Extracting the five-axis displacement data within a preset window before and after the moment, performing fast Fourier transform to obtain a frequency domain signal and dividing it into multiple sub-bands; The normalized energy value of each sub-band of each axis is calculated, and the inner optical path feature set is constructed by arranging them in axial order.

3. The method for controlling an internal optical path five-axis linkage laser cutting based on industrial control according to claim 2, characterized in that: The state comparison network includes a parameter-sharing dual-branch structure, which processes the inner optical path feature sets in the activation state and the inactivation state respectively; Each branch extracts local frequency domain features through the convolution layer and generates path potential representation after compression by the pooling layer; The similarity scores of the potential representations of the pathways in the activated and inactivated states are calculated, and the comparison representations quantifying the impact of the cleavage state are generated by difference operation.

4. The method for controlling an internal optical path five-axis linkage laser cutting based on industrial control according to claim 3 is characterized in that: The processing times are divided into three levels: low frequency, medium frequency and high frequency by counting the number of cutting start and stop events in a unit time; The causal inference model uses the comparison representation as a feature vector, the cutting state as a processing variable, and the cutting accuracy as an outcome variable; The predicted deviation of each sample in the cutting activation and inactivation states is calculated through counterfactual analysis, and the channel contribution of each sensor channel is obtained by aggregation.

5. The method for controlling internal optical path five-axis linkage laser cutting based on industrial control according to claim 4, characterized in that: The dynamic optical path stability index is obtained by extracting the low-frequency energy of the internal optical path feature set, performing wavelet packet decomposition, and then weighting the energy proportion of each layer; The path deviation index is obtained by calculating the product of the machining path alignment cost and the spatial correlation coefficient in the cutting activation and inactivation states.

6. The method for controlling internal optical path five-axis linkage laser cutting based on industrial control according to claim 5, characterized in that: Combining the dynamic optical path stability index, the path deviation index and the channel contribution into a comprehensive vector and inputting it into a Transformer model; Dynamically adjust the basic threshold based on the optical path quality score output by the Transformer model; Channels whose channel contributions are higher than the adjusted threshold are screened, and their comparison representations are extracted to generate the pure light path representation.

7. The method for controlling internal optical path five-axis linkage laser cutting based on industrial control according to claim 6, characterized in that: The time domain grouping according to the cutting waveform phase includes: extracting the phase period from the laser waveform signal and dividing the pure optical path representation into a plurality of phase interval groups; The frequency domain transformation is performed on each group of pure optical path representations, and weighted processing is performed using normalized frequency band energy. After inverse transformation, a weighted optical path representation is formed.

8. The method for controlling internal optical path five-axis linkage laser cutting based on industrial control according to claim 1, characterized in that: Extract on-site processing elements based on the processing feature recognition model and determine the current processing scene based on the preset processing scene knowledge graph; The process specification file of the corresponding scene is called to divide the optical path into large-scale area and small-scale area.

9. The method for controlling internal optical path five-axis linkage laser cutting based on industrial control according to claim 8, characterized in that: Perform texture complexity analysis on the large-scale area of ​​the light path, and remove the small-scale sub-area of ​​the light path when the local area index exceeds the threshold; Establish a real-time feedback mechanism between large-scale analysis models and small-scale analysis models; Based on the preliminary results of the large-scale analysis model, the small-scale analysis model is triggered to perform fine verification of the optical path parameters.

10. The method for controlling internal optical path five-axis linkage laser cutting based on industrial control according to claim 9, characterized in that: Integrate multi-source process data to build a light path feature warehouse, extract feature vectors of each dimension to generate a light path feature matrix; Calculate the link degree between features in each dimension to build a dimension link graph; Through cross-dimensional connection of key feature link degrees, a light path process link map is formed; The core light path feature sequence is screened based on the light path process link map to drive the generation of the final control instruction.

Images