Light spot size self-adaptive control method for laser cladding additive manufacturing

Abstract

The invention relates to the technical field of laser cladding manufacturing, and discloses a light spot size self-adaptive control method for laser cladding additive manufacturing. The method comprises the following steps: according to real-time monitoring data in a laser cladding process, performing multi-dimensional fusion processing on dynamic characteristics of a molten pool to generate a spatial-temporal characteristic tensor containing temperature gradient distribution, molten pool morphology characteristics and spectral characteristic information; inputting the spatial-temporal feature tensor into a depth probabilistic reasoning model, carrying out probabilistic representation on a cladding abnormal state, and outputting a light spot regulation and control parameter vector with confidence; constructing a three-dimensional energy field distribution model according to the vector, and predicting an energy field propagation path and an evolution trend based on a dynamic graph convolutional network; and finally, according to a prediction result, a self-adaptive light spot adjustment mechanism based on the generative adversarial network is constructed, a multi-stage light spot adjustment threshold dynamically changing along with the cladding process is generated, self-adaptive control over the light spot size is achieved, and the cladding forming quality is improved.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of laser cladding manufacturing, in particular to a spot size self-adaptive control method for laser cladding additive manufacturing. BACKGROUND

[0002] As one of the core technologies in the field of advanced manufacturing, laser cladding additive manufacturing technology has been widely used in high-end manufacturing fields such as aerospace, rail transportation and energy equipment, due to its unique advantages in complex component forming, high-performance material preparation and waste parts repair. This technology uses a laser beam as an energy source to make metal powder or wire rapidly melt on the surface of the substrate to form a molten pool, and then accumulates layer by layer according to the preset path to realize near-net forming of the component. In this process, the laser spot size, as a key parameter that directly affects the energy input density, its control accuracy directly determines the forming quality, microstructure performance and mechanical reliability of the cladding layer. If the spot size is too large, it will lead to a decrease in energy density, and problems such as poor bonding of the cladding layer, increased porosity and excessive surface roughness will occur. If the spot size is too small, it will cause local energy concentration, leading to excessive melting of the substrate, increased thermal deformation and even cracks and other defects. Therefore, dynamic and accurate control of the spot size is the core requirement to ensure the stability of the laser cladding additive manufacturing process and the consistency of product quality.

[0003] The current spot size control technology in the laser cladding additive manufacturing process still has many technical bottlenecks to be solved. Traditional spot control methods mostly use an open-loop control mode based on preset process parameters, that is, fixed spot parameters determined by preliminary experiments are used for full-process processing, which cannot cope with complex and variable dynamic interference factors in the cladding process. For example, in the multi-layer cladding process, the temperature gradient between the substrate and the cladding layer will continue to change with the increase of the layer number, causing significant fluctuations in the flowability and solidification rate of the molten pool. At the same time, factors such as fluctuations in the powder feed rate of metal powder, changes in the flow rate of protective gas and attenuation of laser energy will all cause deviations between the actual molten pool state and the preset process parameters, and then cause quality problems such as uneven cladding layer thickness and out-of-tolerance forming size, making it difficult to meet the manufacturing requirements of high-precision and high-performance components.

[0004] To break through the limitations of open-loop control, the industry has gradually developed closed-loop control technology based on real-time monitoring. By introducing monitoring equipment such as visual sensors, infrared thermographs or spectral analyzers, the molten pool image, temperature distribution or spectral information during the cladding process is collected, and the size of the light spot is dynamically adjusted based on traditional control algorithms (such as PID control, fuzzy control). However, such technology still has obvious shortcomings in monitoring dimension fusion, state representation accuracy and adjustment mechanism adaptability: existing monitoring systems are mostly limited to single-dimensional information collection, such as only acquiring molten pool morphology characteristics through visual sensors or only monitoring molten pool temperature through infrared thermographs, failing to achieve deep fusion of multi-dimensional information such as temperature gradient distribution, molten pool morphology characteristics and spectral characteristics, resulting in incomplete characterization of the dynamic state of the molten pool and easy misjudgment of the state; traditional control algorithms rely on manually designed empirical models, making it difficult to accurately model the nonlinear, strongly coupled characteristics and uncertain factors in the cladding process, unable to probabilistically represent abnormal states, resulting in a lack of confidence support for control parameter output and insufficient control accuracy and robustness; existing adjustment mechanisms mostly use fixed thresholds or linear adjustment strategies, which cannot adaptively generate adjustment thresholds according to the dynamic evolution trend of the cladding process, and when the cladding working conditions change dramatically, adjustment lag or overshoot phenomena are easy to occur, making it difficult to ensure stable cladding quality throughout the process.

[0005] With the expansion of laser cladding additive manufacturing to high-end application scenarios such as large-scale complex components, multi-material composite forming and high-precision repair, higher requirements are placed on the dynamic response speed, adaptive ability and intelligent level of the light spot control technology. The existing technology lacks the ability to deeply mine multi-dimensional spatiotemporal characteristics, accurately assess the state based on probability reasoning, and adaptively adjust based on dynamic evolution prediction, and has failed to meet the technical needs of the above high-end application scenarios, becoming a key technical bottleneck restricting the development of laser cladding additive manufacturing technology to higher precision and higher reliability. Therefore, developing a laser cladding light spot size control method that can realize multi-dimensional information fusion, probabilistic state representation, dynamic trend prediction and adaptive threshold adjustment is of great significance for improving the stability of the laser cladding additive manufacturing process, product quality consistency and process adaptability, and promoting the widespread application of the technology in high-end manufacturing. SUMMARY

[0006] The purpose of the present application is to provide a laser cladding additive manufacturing light spot size adaptive control method to solve the problems raised in the background art.

[0007] To achieve the above purpose, the present application provides a laser cladding additive manufacturing light spot size adaptive control method, which comprises: According to the real-time monitoring data collected in the laser cladding process, the dynamic characteristics of the molten pool are multi-dimensionally fused and processed to generate a space-time feature tensor containing temperature gradient distribution, molten pool morphology characteristics and spectral feature information; The space-time feature tensor is input into a deep probabilistic inference model to probabilistically represent the abnormal state in the cladding process, and output a light spot regulation parameter vector with confidence; According to the light spot regulation parameter vector, a three-dimensional energy field distribution model is constructed, and the propagation path and evolution trend of the energy field distribution are predicted based on a dynamic graph convolution network; For the prediction results of the energy field distribution model, an adaptive light spot adjustment mechanism based on a generative adversarial network is constructed, and through the adversarial training of the standard working condition generator and the real-time state discriminator, a multi-level light spot adjustment threshold dynamically changing with the cladding process is generated.

[0008] Preferably, the real-time monitoring data collected in the laser cladding process is multi-dimensionally fused and processed to generate a space-time feature tensor containing temperature gradient distribution, molten pool morphology characteristics and spectral feature information, including: The cladding process monitoring data collected by the multi-source sensor is subjected to time sequence alignment and noise filtering processing, and an adaptive time regularization algorithm is used to eliminate the sampling frequency difference to obtain synchronous multi-dimensional monitoring data stream; The multi-dimensional monitoring data stream is input into a space-time feature extraction network, wherein a three-dimensional convolution kernel is used to extract the molten pool morphology characteristics, and a long short-term memory network is used to capture the dynamic change law, to obtain a preliminary space-time feature mapping; Cross attention mechanism is applied to the space-time feature mapping to calculate the correlation weight between different monitoring features, and the features are dynamically fused according to the weight to obtain an optimized space-time feature representation; The optimized space-time feature is input into a sparse encoder for dimension reduction processing to generate a low-dimensional space-time feature tensor containing temperature gradient distribution, molten pool morphology characteristics and spectral feature information.

[0009] Preferably, the space-time feature tensor is input into a deep probabilistic inference model to probabilistically represent the abnormal state in the cladding process, and output a light spot regulation parameter vector with confidence, including: The space-time feature tensor is input into a deep feature transformation network to obtain a high-dimensional latent feature representation through nonlinear mapping; A hybrid kernel density estimation model is constructed based on the high-dimensional latent feature representation, and a multi-scale kernel function combination is used to represent the nonlinear features of the cladding process; The probability density estimation value of the feature point is calculated, the model uncertainty is quantified by a variational Bayesian method, and a probability feature distribution with confidence interval is obtained; An abnormal feature point deviating from the normal distribution is screened by using an abnormality detection algorithm based on a statistical distance to generate a confidence spot regulation parameter vector.

[0010] Preferably, a three-dimensional energy field distribution model is constructed according to the spot regulation parameter vector, and a dynamic graph convolution network is used to predict the propagation path and evolution trend of the energy field distribution, including: The spot regulation parameter vector is mapped to a three-dimensional machining space coordinate system, and a radial basis function interpolation is used to construct an initial energy field distribution model; The time dimension information is combined, and a time-space interpolation algorithm is used to dynamically correct the energy field distribution to obtain a time-varying energy field distribution model; The time-varying energy field distribution model is input into a dynamic graph convolution network, a graph structure model is constructed based on a cladding path, and a graph attention mechanism is used to capture the energy transfer relationship between nodes; The historical energy distribution data is combined, and a sequence prediction algorithm is used to generate a prediction result of the propagation path and evolution trend of the energy field distribution.

[0011] Preferably, the prediction result of the energy field distribution model is used to construct an adaptive spot regulation mechanism based on a generative adversarial network, including: A standard working condition generator of a conditional generative adversarial network is trained to generate simulation data conforming to the standard cladding features; A deep convolutional discriminator network is constructed, the energy field distribution prediction result and the generated data are input, and the state boundary features are learned through adversarial training; An adaptive clustering algorithm is used to divide the discrimination result into multiple levels to generate a dynamic spot regulation threshold set; The real-time cladding features are compared with the threshold set to dynamically adjust the spot control parameters.

[0012] Preferably, the standard working condition generator of the conditional generative adversarial network generates simulation data conforming to the standard cladding features, including: Multi-scale feature extraction is performed on historical standard cladding data to obtain hierarchical feature representation; A deep convolutional generative network with conditional constraints is constructed to ensure the spatial continuity of the generated data; The generated network parameters are optimized through adversarial training to make the generated data consistent with the real data in the feature space distribution.

[0013] Preferably, the deep convolutional discriminator network is constructed, the energy field distribution prediction result and the generated data are input, and the state boundary features are learned through adversarial training, including: A multi-scale convolution kernel structure is designed to capture energy distribution features of different scales, a feature pyramid network is used to fuse multi-level feature representation, and a dynamic weight adjustment mechanism is introduced to balance the generation and discrimination process.

[0014] Preferably, the adaptive clustering algorithm is used to divide the discrimination results into multiple levels to generate a dynamic spot adjustment threshold set, comprising: The number of clustering centers is automatically determined based on density peak detection, the similarity measurement method is adjusted according to the dynamic characteristics of the energy field, and the multi-level threshold boundary division is determined through iterative optimization.

[0015] Preferably, the real-time cladding features are compared with the threshold set, and the spot control parameters are dynamically adjusted, comprising: The multi-dimensional distance measurement of real-time features and each level threshold is calculated, the weighted distance evaluation is performed according to the feature confidence, and the control parameter adjustment instruction is generated based on the evaluation result.

[0016] Preferably, the control parameter adjustment instruction is generated based on the evaluation result, comprising: The self-learning fuzzy rule base is constructed to dynamically update the rule weight, the multi-input fuzzy reasoning system is designed to process multi-dimensional evaluation data, and the accurate control instruction is output through the defuzzification process.

[0017] Compared with the prior art, the beneficial effects of the present application are: The spot size adaptive control method for laser cladding additive manufacturing first performs multi-dimensional fusion processing on the real-time monitoring data collected during the laser cladding process to generate a space-time feature tensor containing temperature gradient distribution, molten pool morphology characteristics and spectral feature information. The fusion of such multi-dimensional information can fully capture the dynamic changes of the molten pool, compared with the traditional monitoring method which only relies on a single parameter, the real state of the molten pool during the cladding process can be more accurately reflected, avoiding the problem of inaccurate molten pool state judgment caused by one-sided information, and providing a more reliable basis for subsequent spot regulation.

[0018] The space-time feature tensor is input into a deep probability reasoning model to probabilistically represent the abnormal state during the cladding process and output a spot regulation parameter vector with confidence. The model can predict and quantitatively analyze the abnormal state that may occur during the cladding process, and through probabilistic representation, the staff can clearly understand the possibility and degree of abnormal state. The regulation parameter vector with confidence provides a clear and reference direction for spot adjustment, avoiding the blindness of traditional regulation process due to the lack of quantitative basis, and helping to take effective regulation measures before or at the initial stage of abnormal state, reducing the probability of defects.

[0019] A three-dimensional energy field distribution model is constructed based on the spot regulation parameter vector, and a dynamic graph convolution network is used to predict the propagation path and evolution trend of the energy field distribution. The three-dimensional energy field distribution model can intuitively show the distribution of laser energy in the cladding area, and the dynamic graph convolution network has strong spatiotemporal data processing capability and can accurately predict the change law of the energy field. Through the prediction of the propagation path and evolution trend of the energy field, the influence of laser energy on the cladding process can be grasped in advance, so that the timing and amplitude of spot adjustment can be better grasped, and the adjustment of the spot size can be adapted to the change of the energy field, ensuring that the cladding area obtains uniform and required energy supply, and further improving the quality stability of the cladding layer.

[0020] An adaptive spot adjustment mechanism based on a generative adversarial network is constructed for the prediction results of the energy field distribution model. Through the adversarial training of the standard working condition generator and the real-time state discriminator, multi-level spot adjustment thresholds dynamically changing with the cladding process are generated. This adversarial training method can continuously optimize the adjustment mechanism. The standard working condition generator can provide an ideal cladding working condition reference, and the real-time state discriminator can accurately identify the difference between the current cladding state and the standard working condition, and then generate multi-level spot adjustment thresholds that can be dynamically adjusted according to the real-time changes of the cladding process, realizing adaptive control of the spot size. This adjustment mechanism can keep the spot size always matched with the real-time cladding demand, whether the molten pool state changes slightly or fluctuates greatly. It can adjust the spot size in time to ensure that the cladding process is always in a stable and ideal state, effectively reducing defects such as cracks, pores, and poor bonding caused by mismatched spots, improving the surface quality and internal performance of the formed part, reducing the dependence on the experience of operators, improving the stability and consistency of the production process, and helping to promote the application of laser cladding additive manufacturing technology in more high-end manufacturing scenarios. BRIEF DESCRIPTION OF DRAWINGS

[0021] Figure 1 A working principle diagram of the spot size adaptive control method for laser cladding additive manufacturing described in the present application; Figure 2 A flowchart for generating a spatiotemporal feature tensor by multi-dimensional fusion processing of molten pool dynamic features; Figure 3 A flowchart for inputting a spatiotemporal feature tensor into a deep probability inference model to output a spot regulation parameter vector. DETAILED DESCRIPTION

[0022] With reference to the drawings of the embodiments of the present application, the technical solutions in the embodiments of the present application will be described clearly and completely. 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 are within the protection scope of the present application.

[0023] Referring to Figure 1 The present application provides a laser cladding additive manufacturing spot size adaptive control method, which integrates multi-source sensor data, deep learning and probabilistic inference technology to realize real-time dynamic adjustment of spot parameters. The overall implementation scheme of the method includes the following steps: during the laser cladding process, real-time monitoring data is collected by high-speed cameras, infrared thermographs and spectrometers, etc. These data cover molten pool temperature, morphology and spectral information. First, the collected data are subjected to multi-dimensional fusion processing to generate a spatiotemporal feature tensor containing temperature gradient distribution, molten pool morphology characteristics and spectral feature information. This tensor can comprehensively reflect the dynamic behavior of the molten pool. Next, the spatiotemporal feature tensor is input into a deep probabilistic inference model. This model uses Bayesian networks and variational inference technology to probabilistically represent abnormal states during the cladding process and outputs a spot regulation parameter vector with confidence, which includes spot size, power and scanning speed, etc. Then, based on the spot regulation parameter vector, a three-dimensional energy field distribution model is constructed. This model describes the distribution of energy in the processing area through radial basis function interpolation and spatiotemporal interpolation algorithms and predicts the propagation path and evolution trend of the energy field distribution using a dynamic graph convolution network to identify potential heat accumulation or defects in advance. Finally, an adaptive spot adjustment mechanism based on a generative adversarial network is constructed for the prediction results of the energy field distribution model. This mechanism generates multi-level spot adjustment thresholds that change dynamically with the cladding process by training standard working condition generators and real-time state discriminators to realize automatic optimization of spot size. The entire method ensures the stability and consistency of cladding quality through closed-loop control.

[0024] Embodiment 1: Referring to Figure 2The parallel data streams collected by high-speed cameras, infrared thermal imagers and spectrometers are coordinated, which naturally have different timestamps and signal noise problems due to differences in sensor physical characteristics and sampling mechanisms. An adaptive time warping algorithm is used to handle the timing alignment challenge, which dynamically analyzes the time series characteristics of each data stream, maps sampling points of different frequencies to a common time axis by finding the optimal warping path, and considers the dynamic time warping distance between data points instead of simple linear interpolation, thereby eliminating the differences in sampling frequency while maximizing the preservation of the dynamic characteristics of the original signal. The synchronous noise filtering process uses differentiated strategies for different types of sensor data. For thermal imager temperature data, a threshold denoising method based on wavelet transform is used to smooth random fluctuations. For high-speed camera image sequences, a median filter is used to suppress salt and pepper noise. For spectral signals, a moving average filter is used to eliminate high-frequency interference. Finally, the multi-dimensional monitoring data stream is strictly synchronized in time and the signal-to-noise ratio is significantly improved.

[0025] The synchronized multi-dimensional monitoring data stream is immediately fed into a specially designed spatiotemporal feature extraction network for deep feature mining. The network architecture is designed as a dual-channel structure for parallel processing of spatial and temporal information. In the spatial dimension, three-dimensional convolution kernels are used to perform convolution operations on consecutive molten pool image sequences. These convolution kernels slide in three-dimensional body data to capture local spatial patterns of molten pool morphology, such as molten pool boundary profile, depression depth, and surface ripple features. The temporal dimension channel is composed of long short-term memory network units. This network receives spatial feature sequences extracted from the three-dimensional convolution layer, remembers the long-term dependencies of molten pool dynamic changes through its internal gating mechanism, and learns the evolution laws of temperature gradient distribution and molten pool geometry over time, such as identifying abnormal patterns of rapid temperature rise accompanied by molten pool expansion. The output features of the spatial and temporal channels are spliced in the later stages of the network to form a preliminary spatiotemporal feature mapping. This mapping expresses the complete state information of the molten pool within a specific time window in the form of multiple layers of feature maps.

[0026] Although the preliminary spatiotemporal feature map contains rich information, there may be redundancy or weak correlation between features, so it is necessary to introduce a cross-attention mechanism to optimize the feature representation. The working principle of this mechanism is to let the features from different sources talk to each other to evaluate the strength of their association. Specifically, the mechanism regards temperature features, topological features, and spectral features as query, key, and value sequences, respectively. By calculating the dot product attention score between the query and the key, a set of association weight matrices is obtained, which clearly indicates which features contribute more to the overall state at a specific cladding moment. According to the calculated dynamic weights, the preliminary feature map is weighted and fused. For example, when the temperature gradient is abnormally steep, the weight of the temperature feature will automatically increase, making it dominant in the fused feature. This dynamic weighting strategy ensures that the fused spatiotemporal feature representation highlights key information while suppressing non-key interference.

[0027] Although the optimized spatiotemporal feature representation has been refined, its dimension is still high, which is not conducive to real-time processing. Therefore, a sparse encoder is needed to reduce the dimension for compression to generate a more compact tensor representation. The sparse encoder is trained through unsupervised learning to obtain an over-complete dictionary basis. The goal is to reconstruct the input feature using as few linear combinations of dictionary bases as possible. This process naturally introduces a sparsity constraint, making most coefficients zero. In the encoding phase, the optimized high-dimensional spatiotemporal features are projected onto this dictionary. By solving an optimization problem with an L1 regularization term, a set of sparse coefficients is obtained, which constitute the low-dimensional spatiotemporal feature tensor. Although the final generated tensor has a significantly reduced dimension, it still retains the most discriminative temperature gradient distribution information, key geometric features of the molten pool topography, and spectral feature information reflecting material composition changes in the original data due to the characteristics of sparse coding.

[0028] The delay from data acquisition to the generation of the final tensor is strictly controlled within a few milliseconds to meet the stringent requirements of real-time control. The size of the three-dimensional convolution kernel is set to 5x5x5 according to the typical size of the molten pool image to ensure that the receptive field can cover a meaningful spatial region. The number of hidden layer units of the long short-term memory network is set to 128 to balance the model capacity and computational efficiency. The cross-attention mechanism uses 8 parallel attention heads to enable the model to simultaneously focus on information from different representation subspaces. The dictionary size of the sparse encoder is determined experimentally to be 1000 atoms to achieve a good trade-off between reconstruction accuracy and computational complexity. All these parameters are optimized and determined based on the statistical analysis results of a large amount of historical cladding data. The parameters of various filtering algorithms and alignment algorithms in the data processing process have adaptive adjustment capabilities. For example, the size of the adaptive time warping window will dynamically change according to the stability of the real-time data stream, ensuring that the system can adapt to different processing conditions and material properties. This high degree of adaptability and robustness makes the fusion processing method widely applicable in actual industrial scenarios.

[0029] Example 2: refer to Figure 3 After the spatio-temporal feature tensor enters the deep probabilistic inference model, it first undergoes feature transformation to extract higher-level abstract representations. The deep feature transformation network consists of three fully connected layers, each containing batch normalization and ReLU activation operations. The first layer increases the dimension of the input tensor to 256 to expand the feature representation capacity. The second layer compresses the dimension to 128 to focus on key information. The third layer further compresses it to 64 to form a high-dimensional latent feature representation. This step-by-step dimension reduction strategy effectively removes redundant information while preserving main features. The high-dimensional latent feature representation not only contains the static properties of the molten pool but also encodes its dynamic change trends. The hybrid kernel density estimation model based on the high-dimensional latent feature representation uses multi-scale kernel functions to capture the complex nonlinear characteristics of the cladding process. The model combines Gaussian kernel functions for local smoothing features, polynomial kernel functions for global trends, and Laplace kernel functions for sharp distributions. These kernel functions are linearly combined through adaptive weights, which are dynamically adjusted according to the scale range of the feature values, allowing the model to respond to both rapid fluctuations and slow drifts. The Parzen window method is used to calculate the probability density of each feature point, with the feature point as the center to construct the probability density contribution. To quantify the uncertainty of the model, a variational Bayesian inference framework is introduced, which assumes that the latent features follow a Gaussian prior distribution and finds the variational distribution closest to the true posterior distribution through iterative optimization. During the inference process, the lower bound of the variation is calculated and the variational parameters are continuously optimized by the gradient ascent algorithm until convergence. The final probability feature distribution not only contains the mean estimate of each feature but also gives its confidence interval.

[0030] Anomaly detection mechanism based on statistical distance is adopted to screen the probability feature distribution to identify abnormal points deviating from the normal mode. The Mahalanobis distance between each feature point and the normal feature distribution is calculated and cross-validated with the KL divergence. The Mahalanobis distance is standardized by the feature covariance matrix, which can effectively eliminate the dimensional influence between features. A dynamic threshold is set to automatically mark abnormal feature points. The threshold size is updated every five minutes according to the statistical characteristics of the historical cladding data to adapt to the changes in process conditions. The marked abnormal points will trigger the recalculation of the light spot regulation parameter vector. The generated light spot regulation parameter vector contains the adjustment amount of laser power, scanning speed and light spot diameter and their respective confidence, which is derived from the variance estimation of the variational Bayesian inference and mapped to the 0-1 interval through the sigmoid function. The final output parameter vector is stored in array form and attached with a timestamp to ensure synchronization with the control system.

[0031] When the light spot regulation parameter vector is mapped to the three-dimensional processing space coordinate system, an equal-interval grid division strategy is adopted. Each grid node corresponds to a spatial position and contains control information from the parameter vector. When constructing the initial energy field distribution model using the radial basis function interpolation method, the Gaussian function is selected as the basis function due to its good local approximation characteristics. The center point of the basis function is set at the grid node, and the shape parameter is determined through cross-validation. In the interpolation process, the energy value of each spatial point is obtained by weighted average of the parameter values of its adjacent nodes, and the weight is inversely proportional to the distance. The generated initial energy field can continuously and smoothly reflect the energy distribution profile of the entire processing area.

[0032] The time dimension information is introduced to modify the initial energy field to capture the dynamic evolution behavior of the energy field. The spatio-temporal interpolation algorithm uses time as the fourth dimension and constructs a four-dimensional interpolation function. This function uses radial basis interpolation in the spatial domain and linear interpolation in the time domain to balance the calculation efficiency and accuracy. By superimposing the energy field data of multiple consecutive time slices to form a four-dimensional hyper-volume data, the energy field value at each time point is calculated through spatio-temporal interpolation to obtain the time-varying energy field distribution model. This model can predict the energy accumulation at any spatial point in the future few seconds. The time-varying energy field distribution model is converted into a graph structure to input the dynamic graph convolution network for processing. The nodes of the graph are composed of discrete points in the processing space, and the edges are defined by the adjacency relationship between the nodes. The edge weights are set according to the physical law of energy transfer, such as inversely proportional to the square of the distance. The graph convolution network uses a graph attention mechanism to capture the nonlinear interaction between nodes. This mechanism calculates an attention coefficient for each edge to represent the strength of energy transfer. The attention coefficient is calculated by the similarity of node features and normalized by softmax. The skip connection is used between network layers to avoid gradient disappearance and maintain multi-scale features. The final output is an energy flow vector for each node.

[0033] By combining historical energy distribution data, a gated recurrent unit network (GRU) predicts the propagation path of the energy field. The network takes the energy field sequence of the past ten time steps as input and memorizes long-term dependencies through its hidden state, outputting predicted energy field values ​​for the next five time steps. The propagation path is visualized as a three-dimensional vector field, showing the direction and intensity of energy flow. The evolution trend is identified by calculating the divergence of the energy gradient field to determine energy accumulation or dissipation regions. The entire prediction result is transmitted to the downstream spot adjustment module in a structured data format. The deep probabilistic inference model and the energy field prediction module run in parallel on an industrial computer, with each processing cycle not exceeding 50 milliseconds. The weights of the deep feature transformation network are fine-tuned every two hours using the latest data to maintain model adaptability, and the kernel function parameters of the hybrid kernel density estimation are recalibrated every ten minutes. The graph structure of the dynamic graph convolutional network is updated every five seconds based on the real-time cladding path, and the hidden state of the GRU is reset after each processing layer to avoid cross-layer interference. These designs ensure that the system can continuously adapt to changing processing conditions. The energy field prediction results are output every 0.5 seconds and transmitted to the motion control card via gigabit Ethernet, achieving millisecond-level synchronization with the laser modulation signal to ensure real-time and accurate control.

[0034] Example 3: Dynamic optimization of spot parameters is achieved by simulating standard operating conditions and learning state boundaries. The core of this method lies in training a conditional generative adversarial network (GAN) to generate simulated data conforming to cladding specifications and extracting key features using a discriminator. The training process begins with multi-scale feature extraction of historical standard cladding data to obtain hierarchical feature representations. This multi-scale feature extraction employs a convolutional pyramid network structure, containing multiple parallel convolutional layers. Each layer uses convolutional kernels of different sizes to capture cladding features from local details to global trends. For example, small-sized convolutional kernels focus on the microscopic morphological changes of the molten pool, while large-sized kernels identify the overall temperature distribution pattern. The extracted hierarchical features are then input into a deep convolutional generative network with conditional constraints. This network uses an encoder-decoder architecture, where the encoder compresses input features through convolutional and pooling layers, and the decoder reconstructs data through deconvolutional and upsampling layers. Conditional constraints ensure the spatial continuity and physical plausibility of the generated data by embedding cladding process parameters such as laser power and scanning speed as additional input channels into the network.

[0035] The goal of the generative network is to output simulated cladding feature data whose distribution should be consistent with the real standard data. During adversarial training, the generator and discriminator alternately optimize. The generator attempts to generate indistinguishable data, while the discriminator strives to distinguish between real and generated data. This game is balanced by minimizing the adversarial loss function, whose expression is: ; Where: symbol This represents the overall loss value of the generative adversarial network. Represents the expectation operator, From the distribution of real data Samples sampled from the middle, The discriminator function outputs a scalar representing the probability that the input is the true data. From the distribution of potential noise Random vectors sampled from the middle, The generator function maps noise into the data space. Training uses the gradient descent algorithm to iteratively update the network parameters until the generated data is indistinguishable from real data in the feature space.

[0036] When constructing the deep convolutional discriminator network, emphasis is placed on its ability to capture multi-scale energy distribution features. The network design includes multiple sets of convolutional layers with kernel sizes such as 3x3, 5x5, and 7x7 to process input data within different receptive fields. Smaller kernels capture local gradient changes in the energy field, while larger kernels perceive global distribution trends. Activation functions and batch normalization layers follow the convolutional layers to enhance non-linear representation and training stability. The output feature maps are fed into a feature pyramid network for multi-level feature fusion. The feature pyramid integrates low-resolution high-level semantic features with high-resolution low-level detail features through a top-down path and lateral connections, thereby generating a unified multi-scale feature representation. A dynamic weight adjustment mechanism is introduced during discriminator training. This mechanism automatically adjusts the learning rates of both the generator and discriminator based on their current loss ratio. For example, when the discriminator's accuracy is too high, its weight is reduced to prevent generator training from stalling. The weight adjustment follows a dynamic update rule, with values ​​calculated for each batch to ensure balance during training.

[0037] An adaptive clustering algorithm is employed to divide the discriminator's output into multiple levels to generate a dynamic set of light spot adjustment thresholds. Based on the density peak detection principle, the clustering algorithm automatically determines the number of natural categories in the data. Density peaks are identified by calculating the local density of each data point and the minimum distance to high-density points. Local density is calculated using truncated kernel density estimation, while the minimum distance identifies the nearest high-density point. The similarity measurement method is adjusted according to the dynamic characteristics of the energy field; for example, Euclidean distance is used in energy accumulation regions to emphasize numerical differences, while cosine similarity is used in energy diffusion regions to focus on directional consistency. The clustering process is iteratively optimized, recalculating cluster centers and data point membership degrees in each iteration until the objective function converges. The final threshold set includes multiple levels such as normal, warning, and abnormal states, each corresponding to a light spot adjustment range. The threshold boundaries are defined by the statistical characteristics of the cluster centers, such as the mean and standard deviation.

[0038] The historical standard cladding data from the database contains thousands of standard processing cases, each case has complete sensor records, and the network parameters of multi-scale feature extraction are determined by cross-validation. The number of convolution kernels is set to 32, 64, and 128 corresponding to different scales. The training of the generative adversarial network uses the Adam optimizer with a learning rate of 0.0001 and a batch size of 32. The training period is at least 1000 rounds. The multi-scale convolution output of the discriminator is compressed into a feature vector through a global average pooling layer and then input into a fully connected layer for classification. The density calculation window size of the adaptive clustering algorithm is dynamically adjusted according to the data dimension, and the upper limit of the iteration number is set to 100 to ensure the convergence of the algorithm. The generated threshold set is stored in the form of a query table for real-time comparison. When the mechanism is running, energy field prediction data is received every 0.1 seconds and spot adjustment suggestions are output. These suggestions are transmitted to the laser controller through a digital interface to realize closed-loop control.

[0039] The standard working condition generator of the conditional generative adversarial network is specifically implemented by an encoder part composed of four convolutional layers, each followed by a LeakyReLU activation function and a max pooling layer, and a decoder part composed of four deconvolutional layers, each followed by a ReLU activation function and an up-sampling layer. The conditional input is combined with the noise vector through splicing and then input into the generation network. The discriminator network adopts a five-layer convolutional structure with the number of convolution kernels increasing from 16 to 128. The feature pyramid network includes three levels corresponding to input sizes of 1 / 4, 1 / 2, and the original scale feature map. When the feature maps are fused, a 1x1 convolution is used to adjust the number of channels. Dynamic weight adjustment is based on loss ratio calculation. When the discriminator loss is lower than the generator loss by a certain percentage, the weight coefficient decreases. In the density peak detection of the clustering algorithm, the local density calculation uses a Gaussian kernel function with a bandwidth parameter that is adaptively selected by the nearest neighbor distance. After threshold division, each clustering category is assigned a state label and its boundary value is calculated. The boundary is determined by the percentile of the data points within the category. For example, the normal state takes 5% to 95% percentile, and the warning state takes the data outside the boundary. When the system is integrated, all components are deployed on an industrial computer, and CUDA is used to accelerate the training and inference process. The data flow uses pipeline processing to ensure real-time performance delay control within 50 milliseconds.

[0040] In the case of laser cladding repair of the blade edge of a water turbine, the processing material is cobalt-based alloy powder, and the base material is 13Cr4Ni stainless steel. The design of the deep convolutional discriminator network needs to process time-varying energy field distribution data from the previous module, which presents each data point containing spatial coordinates (X, Y, Z) and energy intensity value (E) in the form of a three-dimensional matrix. The size of the network input layer needs to adapt to the spatial dimension of the energy field data, for example, 50x50x20 grid points. When constructing the multi-scale convolution kernel structure, the first group of convolution kernels adopts a small size kernel of 3x3x3, with a step size of 1 and a padding of 1, which is specifically used to capture the local gradient changes between adjacent grid points in the energy field, such as small fluctuations in energy intensity. The second group of convolution kernels adopts a medium size kernel of 5x5x5, with a step size of 2 and a padding of 2, which is used to identify energy distribution patterns in larger regions, such as hot spot areas formed by energy accumulation. The third group of convolution kernels adopts a large size kernel of 7x7x7, with a step size of 3 and a padding of 3, which aims to perceive the global energy distribution trend of the entire processing area, such as the overall symmetry of the energy field. After each group of convolution layers, a LeakyReLU activation function is connected, with a negative slope parameter of 0.01, and a 3D batch normalization layer is connected to stabilize the training process. The multi-scale feature maps output by the convolution operation are concatenated in the channel dimension to form a comprehensive feature cube.

[0041] When the feature pyramid network is used to fuse these multi-level feature representations, the network constructs two paths from top to bottom and horizontally. The top-down path gradually upsamples the high-level low-resolution feature maps to the original input size through 3D deconvolution operations, and the horizontal connection element-wise adds the upsampled features to the same scale high-resolution feature maps from the convolution layer. The fusion process is divided into three levels. After the first level fusion, the feature map size is 25x25x10, the second level is 12x12x5, and the third level is 6x6x3. After each fusion level, a 1x1 convolution kernel is used to adjust the number of channels to reduce the computational load. The introduced dynamic weight adjustment mechanism updates the network weight in real time according to the loss value ratio of the generator and the discriminator. This mechanism calculates the ratio of the generator loss and the discriminator loss after each training batch, and automatically reduces the learning rate of the discriminator when the ratio exceeds the set threshold. The weight update formula is based on the exponential moving average to calculate the new weight to avoid sharp fluctuations. The weight coefficient is initially set to 0.5 and dynamically adjusted according to the training stability.

[0042] The adaptive clustering algorithm processes the energy field feature vectors extracted by the discriminator network, each feature vector corresponds to an energy field state at a processing time, and its dimension is 256. The algorithm first automatically determines the number of cluster centers based on density peak detection, calculates the local density of each feature vector The value is calculated using a Gaussian kernel function, which depends on the distance between the vector and the adjacent vector, and the distance between each vector and the nearest vector with higher density Vectors with higher local density and larger distance are selected as cluster centers, which is assisted by decision map visualization. Operators can observe the distribution of and Similarity measurement is adjusted according to the dynamic characteristics of energy field. Mahalanobis distance is used in areas with dramatic energy changes to consider the covariance structure of features, and Euclidean distance is used in areas with smooth energy distribution to simplify calculations. Multi-level threshold boundary division is determined by iterative optimization. The iterative process uses the K-medoids algorithm to reselect center points and assign data points each iteration until the objective function converges. After clustering, each cluster is assigned a state label and its boundary value is calculated to form a set of dynamic spot adjustment thresholds. Referring to Table 1, the intermediate results obtained by the discriminator network when repairing the leading edge of a water turbine blade at a specific location are shown. The results are used for subsequent threshold division: Table 1: Energy field feature vector clustering analysis

[0043] The batch size is set to 16 during network training, the training period is 500 rounds, the initial learning rate is 0.001, and the cosine annealing strategy is used for gradual decay. The number of channels for multi-scale convolution kernels is configured as 32 channels for small-scale, 64 channels for medium-scale, and 128 channels for large-scale. The fusion operation of the feature pyramid network uses bilinear interpolation for upsampling. The threshold value for dynamic weight adjustment is set to 1.5. When the loss ratio exceeds this value, the discriminator learning rate is multiplied by a decay factor of 0.8. The iteration stop condition for the adaptive clustering algorithm is set to three consecutive iterations with a cluster center change of less than 0.001. The entire system runs in the case of processing water turbine blades, completing a complete discrimination and clustering analysis cycle every 0.2 seconds. The generated threshold set is dynamically updated at a frequency of once per second to adapt to changes in processing conditions.

[0044] The forward propagation calculation of the deep convolutional discriminator network takes about 15 milliseconds, and the execution time of the adaptive clustering algorithm is about 10 milliseconds. The entire processing chain is completed within 50 milliseconds, meeting the real-time control requirements. The initial value of the network parameters is He normal initialization method, and the training data comes from 200 historical processing cases of water turbine blade repair, each case containing energy field records of 1000 consecutive time points. The cutoff distance dc in the density calculation of the clustering algorithm is automatically determined by analyzing the statistical distribution of the distance between all feature vectors, usually taking the 2nd percentile of the distance distribution. After threshold boundary division, the boundary of each cluster is determined by the covariance matrix of the feature vectors in the cluster. The normal state corresponds to the region with Mahalanobis distance less than 2.5, the warning state corresponds to the region with Mahalanobis distance between 2.5 and 3.5, and the abnormal state corresponds to the region with Mahalanobis distance greater than 3.5. These thresholds are recalculated continuously as the processing proceeds to achieve true adaptive adjustment.

[0045] In example 5, a laser cladding system for repairing turbine blades of an aero-engine is taken as a specific scenario, and the repair material is a nickel-based high-temperature alloy powder, and the substrate is a directional solidification nickel-based high-temperature alloy. The real-time cladding feature data comes from the multi-dimensional feature vector output by the previous module, which is updated every 100 milliseconds and contains 12 dimensional features such as temperature gradient, molten pool area, and spectral intensity ratio. The threshold set comes from the three-level dynamic thresholds generated in example 4, corresponding to normal, warning and abnormal states respectively. When calculating the multi-dimensional distance measure of real-time features and thresholds, the system calculates three different types of distances in parallel. Euclidean distance is used to quantify the straight-line deviation of feature vectors and each threshold cluster center point in the numerical space, which involves the sum of squared differences of all feature dimensions and then taking the square root. Its sensitivity to abnormal dimensions is lower than that of Euclidean distance, and it is more suitable for industrial environments with noise interference. Cosine distance focuses on the consistency of the direction of the feature vector, and evaluates the similarity of the change trend by calculating the cosine value of the vector angle. When the vector length changes greatly but the direction is consistent, the distance remains small.

[0046] The confidence of each feature dimension in the weighted distance evaluation according to the feature confidence is from the variance estimate value output by the deep probabilistic inference model, and the confidence value ranges from 0 to 1. The higher the value, the higher the reliability of the dimension feature. The weighted process assigns a basic weight coefficient to each distance metric, where the Euclidean distance weight is set to 0.5, the Manhattan distance weight is 0.3, and the cosine distance weight is 0.2. These weights are multiplied by the confidence values of each feature dimension to obtain the final weighted weight. For example, when the confidence of the temperature gradient feature is low, its weight in all distance calculations will be correspondingly reduced, avoiding the excessive influence of low-quality data on the overall evaluation. The weighted distance evaluation result is combined linearly to form a comprehensive deviation score, which is mapped to a 0-100 scale for intuitive judgment of the deviation degree of the current state from the threshold.

[0047] The process of generating control parameter adjustment instructions based on the evaluation result relies on a self-learning fuzzy rule base, which initially contains 20 fuzzy rules defined by domain experts. Each rule is in the form of "if the deviation score is A and the confidence is B, then the adjustment action is C". The weight parameters in the rule base are dynamically updated through a reinforcement learning mechanism. After the system completes a processing cycle, the rules used during the period are adjusted based on the final cladding quality evaluation results. The quality evaluation is based on porosity, crack number and other indicators completed by metallographic detection. The designed multi-input fuzzy inference system receives two input variables, deviation score and average confidence level. Each input variable is divided into three fuzzy sets: deviation score is divided into "low, medium, high", and confidence is divided into "weak, medium, strong". The output variable spot adjustment amount is divided into three fuzzy sets: "reduce, maintain, increase". The inference process uses the Mamdani fuzzy model and uses the MIN-MAX inference method to calculate the activation strength of each rule. The outputs of all activated rules are aggregated to form a fuzzy output set.

[0048] When the fuzzy output is converted into precise control instruction by defuzzification, the gravity method is used to calculate the position of the fuzzy set centroid, the output universe of discourse is discretized into 100 levels, the membership degree value of each level is calculated, and then the weighted average value of all level points and their membership degrees is calculated to obtain the precise spot diameter adjustment amount. Then, the adjustment instruction is sent to the laser control system at a millisecond level interval. The instruction contains the spot size, laser power and scanning speed adjustment amount. These parameters are converted into voltage signals by a digital analog converter to drive the laser optical components and motion mechanisms. The system runs in real time in the repair of turbine blade tip. When the comprehensive deviation score of the feature vector and the normal state threshold is detected to be more than 25, the adjustment mechanism is automatically triggered to ensure that the cladding layer quality meets the aviation engine maintenance standards. The update period of the self-learning fuzzy rule base is synchronized with the repair period of each turbine blade, usually updated once after completing one blade. The rule weight adjustment uses the Q-learning algorithm with a learning rate of 0.1 and a discount factor of 0.9. The membership function of the multi-input fuzzy reasoning system uses a triangular function, the parameters of which are obtained by training historical data. The universe of discourse of the input variable deviation score is set to 0-100, the confidence level is 0-1, and the output variable adjustment amount is-10% to +10%. The defuzzification process is executed every 50 milliseconds to ensure the timeliness of the control instruction. The entire system is deployed on a real-time operating machine and communicates with the laser processing equipment through the Ethernet CAT protocol.

[0049] It should be noted that the relational terms herein such as first and second and the like are used solely 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 "comprises", "comprising", or any other variations thereof, are intended to cover a non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements does not include only those elements but can include other elements not expressly listed or inherent to such process, method, article, or apparatus.

[0050] Although embodiments of the present application have been shown and described, it is to be understood that various modifications, substitutions, replacements and changes can be made to these embodiments without departing from the principles and spirit of the present application, and the scope of the present application is defined by the appended claims and their equivalents.

Claims

1. A method for adaptive control of spot size in laser cladding additive manufacturing, characterized in that, include: Based on real-time monitoring data collected during laser cladding, the dynamic characteristics of the molten pool are fused in multiple dimensions to generate a spatiotemporal feature tensor containing information on temperature gradient distribution, molten pool morphology, and spectral characteristics. The spatiotemporal feature tensor is input into the deep probabilistic inference model to probabilistically represent the abnormal state during the cladding process and output a light spot control parameter vector with confidence. Based on the light spot control parameter vector, a three-dimensional energy field distribution model is constructed, and the propagation path and evolution trend of the energy field distribution are predicted based on a dynamic graph convolutional network. Based on the prediction results of the energy field distribution model, an adaptive spot adjustment mechanism based on generative adversarial network is constructed. Through adversarial training between a standard working condition generator and a real-time state discriminator, a multi-level spot adjustment threshold that dynamically changes with the cladding process is generated.

2. The method of claim 1, wherein, The process involves multi-dimensional fusion processing of the dynamic characteristics of the molten pool based on real-time monitoring data collected during laser cladding, generating a spatiotemporal feature tensor that includes information on temperature gradient distribution, molten pool morphology, and spectral characteristics. The monitoring data of the cladding process collected by multi-source sensors are subjected to time alignment and noise filtering. An adaptive time warping algorithm is used to eliminate sampling frequency differences and obtain a synchronous multi-dimensional monitoring data stream. The multidimensional monitoring data stream is input into a spatiotemporal feature extraction network, where the spatial dimension uses a three-dimensional convolutional kernel to extract the morphological features of the melt pool, and the temporal dimension uses a long short-term memory network to capture the dynamic change patterns, thus obtaining a preliminary spatiotemporal feature mapping. A cross-attention mechanism is applied to the spatiotemporal feature mapping to calculate the correlation weights between different monitoring features. Based on the weights, the features are dynamically fused to obtain the optimized spatiotemporal feature representation. The optimized spatiotemporal features are input into a sparse encoder for dimensionality reduction, generating a low-dimensional spatiotemporal feature tensor containing information on temperature gradient distribution, melt pool morphology, and spectral features.

3. The method of claim 2, wherein, The spatiotemporal feature tensor is input into a deep probabilistic inference model to probabilistically represent abnormal states during the cladding process, and outputs a beam control parameter vector with confidence, including: The spatiotemporal feature tensor is input into a deep feature transformation network, and a high-dimensional latent feature representation is obtained through nonlinear mapping. Based on the high-dimensional latent feature representation, a hybrid kernel density estimation model is constructed, and a combination of multi-scale kernel functions is used to characterize the nonlinear features of the cladding process; Calculate the probability density estimate of feature points, quantify the model uncertainty using the variational Bayesian method, and obtain the probability feature distribution with confidence intervals; An anomaly detection algorithm based on statistical distance is used to screen out anomalous feature points that deviate from the normal distribution, and generate a light spot control parameter vector with confidence level.

4. The method of claim 3, wherein, Based on the light spot control parameter vector, a three-dimensional energy field distribution model is constructed, and the propagation path and evolution trend of the energy field distribution are predicted based on a dynamic graph convolutional network, including: The light spot control parameter vector is mapped to a three-dimensional processing space coordinate system, and an initial energy field distribution model is constructed using radial basis function interpolation. By combining time dimension information, the energy field distribution is dynamically corrected through spatiotemporal interpolation algorithm to obtain a time-varying energy field distribution model; The time-varying energy field distribution model is input into a dynamic graph convolutional network, a graph structure model is constructed based on the cladding path, and a graph attention mechanism is used to capture the energy transfer relationship between nodes. By combining historical energy distribution data, a sequence prediction algorithm is used to generate predictions of the propagation path and evolution trend of energy field distribution.

5. The method of claim 4, wherein, Based on the prediction results of the energy field distribution model, an adaptive spot adjustment mechanism based on a generative adversarial network is constructed, including: The standard working condition generator for training conditional generative adversarial networks generates simulated data that conforms to the cladding characteristics of the specifications; Construct a deep convolutional discriminator network, input the energy field distribution prediction results and generated data, and learn state boundary features through adversarial training; An adaptive clustering algorithm is used to divide the discrimination results into multiple levels to generate a dynamic spot adjustment threshold set. The real-time cladding characteristics are compared with the threshold set to dynamically adjust the spot control parameters.

6. The method of claim 5, wherein, The standard working condition generator of the training condition generative adversarial network generates simulated data that conforms to the standard cladding characteristics, including: Multi-scale feature extraction is performed on historical standard cladding data to obtain hierarchical feature representations; Construct a deep convolutional generative network with conditional constraints to ensure the spatial continuity of the generated data; By optimizing the parameters of the generator network through adversarial training, the generated data can be made to be consistent with the distribution of real data in the feature space.

7. The method of claim 6, wherein, The construction of the deep convolutional discriminator network, taking the energy field distribution prediction results and generated data as input, and learning state boundary features through adversarial training, includes: We design a multi-scale convolutional kernel structure to capture energy distribution features at different scales, use a feature pyramid network to fuse multi-level feature representations, and introduce a dynamic weight adjustment mechanism to balance the generation and discrimination processes.

8. The method of claim 7, wherein, The adaptive clustering algorithm is used to divide the discrimination results into multiple levels to generate a dynamic spot adjustment threshold set, including: The number of cluster centers is automatically determined based on density peak detection, the similarity measurement method is adjusted according to the dynamic characteristics of the energy field, and multi-level threshold boundary division is determined through iterative optimization.

9. The method of claim 8, wherein, The step of comparing real-time cladding features with a threshold set and dynamically adjusting the spot control parameters includes: Calculate the multidimensional distance metric between real-time features and thresholds at all levels, perform weighted distance evaluation based on feature confidence, and generate control parameter adjustment instructions based on the evaluation results.

10. The method of claim 9, wherein, The generation of control parameter adjustment instructions based on the evaluation results includes: A self-learning fuzzy rule base is constructed to dynamically update rule weights. A multi-input fuzzy inference system is designed to process multi-dimensional evaluation data and output precise control commands through the defuzzification process.

Images