On-line identification method for roughness of side surface of thin-walled part based on molten pool time sequence image

Abstract

The application discloses a kind of based on directional energy deposition thin-walled part side surface roughness online identification method of molten pool time sequence image, including the following steps: obtaining molten pool time sequence image in situ in the manufacturing process of directional energy deposition thin-walled part, the corresponding relationship between image position and thin-walled part entity space position is constructed, and based on the surface roughness parameter of thin-walled part side surface, labeled training dataset is constructed, and pretreatment is carried out;Constructing spatiotemporal feature adaptive fusion classification model, including spatial feature extraction module, molten pool dynamic fluctuation feature extraction module, double-layer time sequence modeling module, mid-term feature fusion module and adaptive pooling classification module;The spatiotemporal feature adaptive fusion classification model is trained, and the roughness identification model is obtained, and based on roughness identification model output side surface roughness grade identification result.The application can effectively associate in-situ molten pool time sequence image and entity roughness representation result, realize the high-precision online identification of thin-walled part side surface roughness.

Description

Technical Field

[0001] This invention relates to the field of surface roughness detection technology, and specifically to an online identification method for the side surface roughness of directional energy deposition thin-walled parts based on molten pool time-series images. Background Technology

[0002] Directed Energy Deposition (DED) technology has significant advantages in the efficient manufacturing of complex thin-walled metal components and has been widely applied in high-performance parts manufacturing, structural repair, and remanufacturing. However, surface quality control of DED-formed parts remains a critical issue restricting its further engineering applications. For thin-walled parts, the overall surface morphology mainly consists of two parts: the variation in the top deposition height and the side surface roughness. The side surface roughness directly affects the dimensional accuracy, service performance, and subsequent processing costs of the part. If significant depressions, collapsed edges, or abnormal roughness appear in local areas of the side surface, it will not only reduce the appearance consistency of the part but may also weaken the mechanical properties of the component, and in severe cases, even lead to forming failure.

[0003] Existing online monitoring studies on surface quality in the DED process mainly focus on monitoring and analyzing macroscopic morphological features such as top deposition height, melt channel geometry, and thermal field distribution. However, research on the formation mechanism of side surface roughness in thin-walled parts and its online identification methods is still relatively insufficient. This is because the formation of side surface roughness is not determined by the molten pool morphology at a single moment, but is affected by a variety of factors such as powder feeding state, laser energy input, melt pool boundary fluctuations, multilayer thermal accumulation, and interlayer coupling evolution, exhibiting obvious spatiotemporal correlation and cross-layer accumulation.

[0004] In addition, existing methods generally have the following shortcomings: First, they lack a precise location mapping mechanism between in-situ monitoring images and surface roughness on the solid side, making it difficult to establish a reliable data annotation basis; second, they mostly use single-frame images or simple time-series aggregation methods for analysis, making it difficult to effectively characterize the dynamic evolution law of melt pool fluctuations; third, in the feature fusion and classification stages, their ability to express morphological features at different scales is limited, resulting in insufficient ability to distinguish complex roughness states.

[0005] Therefore, there is an urgent need for an online identification technology for the side surface roughness of directional energy deposition thin-walled parts that can effectively correlate in-situ time-series images of the molten pool with the results of solid roughness characterization and simultaneously model short-term dynamic features within the layer and cumulative evolution features between layers, so as to achieve high-precision online identification of the side surface roughness state of thin-walled parts. Summary of the Invention

[0006] To overcome the defects and shortcomings of existing technologies, this invention provides an online identification method for the side surface roughness of directional energy deposition thin-walled parts based on molten pool time-series images. By establishing a mapping relationship between the spatial position of the image and the entity, and constructing a spatiotemporal feature adaptive fusion classification model, the method achieves online identification of the side surface roughness level of the thin-walled part. This solves the problems in existing technologies, such as the difficulty in accurately registering the in-situ image with the entity's side surface roughness, insufficient mining of dynamic evolution features of the molten pool, and low accuracy in identifying complex roughness states.

[0007] To achieve the above objectives, the present invention adopts the following technical solution:

[0008] This invention provides an online method for identifying the side surface roughness of directional energy deposition thin-walled parts based on molten pool time-series images, comprising the following steps:

[0009] In-situ molten pool time-series images of the manufacturing process of directional energy deposition thin-walled parts are obtained, the correspondence between the image positions and the spatial positions of the thin-walled parts is constructed, and a labeled training dataset is constructed based on the surface roughness parameters of the side surfaces of the thin-walled parts.

[0010] Preprocess the training dataset;

[0011] A spatiotemporal feature adaptive fusion classification model is constructed, including a spatial feature extraction module, a melt pool dynamic fluctuation feature extraction module, a two-layer temporal modeling module, a mid-term feature fusion module, and an adaptive pooling classification module;

[0012] The spatial feature extraction module encodes the spatial features of multiple frames of molten pool images frame by frame to obtain a spatial feature map. The molten pool dynamic fluctuation feature extraction module extracts the fluctuation features of the spatial feature map and fuses them with the spatial features of the current frame to obtain enhanced frame-level features. The dual-layer temporal modeling module extracts intra-layer temporal features and inter-layer temporal features respectively. The intermediate feature fusion module performs channel compression and weight calibration on the intra-layer temporal features and inter-layer temporal features respectively, and splices them to obtain fused features. The adaptive pooling classification module generates the weights of each branch and performs weighted fusion on the output features of each branch to output the side surface roughness level prediction result.

[0013] The spatiotemporal feature adaptive fusion classification model is trained based on the preprocessed training set to obtain a roughness recognition model. The roughness recognition model outputs the side surface roughness level recognition result.

[0014] As a preferred technical solution, the correspondence between the image position and the spatial position of the thin-walled component is constructed and represented as follows:

[0015] ;

[0016] in, Indicates the number of times since the laser was turned on. Frame image, Indicates the first The length of the distance from the starting point to the acquisition point corresponding to the frame image. For printing speed, The frame rate is the image sampling rate.

[0017] As a preferred technical solution, the surface roughness parameter of the side surface of the thin-walled part is expressed as:

[0018] ;

[0019] in, Indicates the measured area. Point The surface height at that location.

[0020] As a preferred technical solution, the training dataset is preprocessed, including: image registration, sliding window sampling, region of interest cropping, image enhancement, and size unification.

[0021] As a preferred technical solution, sliding window sampling specifically includes:

[0022] A fixed number of consecutive images are extracted from a continuous melt pool image sequence according to a preset window length to form a sample sequence;

[0023] The number of frames corresponding to a single sample sequence is represented as:

[0024] ;

[0025] in, The number of frames in the sample sequence. For the corresponding deposition length, For printing speed, The frame rate is the image sampling rate.

[0026] As a preferred technical solution, the molten pool dynamic fluctuation feature extraction module extracts the fluctuation features of the spatial feature map, specifically as follows:

[0027] ;

[0028] in, Indicates the first Layer Spatial features corresponding to the frame image;

[0029] The fluctuation features are concatenated with the original spatial features of the current frame in the channel dimension to obtain the joint features, represented as:

[0030] ;

[0031] Then, through convolution, batch normalization, and activation function mapping, the enhanced frame-level features are obtained, represented as:

[0032] ;

[0033] in, Represents a 1×1 convolution. Indicates batch normalization, This represents the activation function. Indicates joint features, This represents the enhanced frame-level features.

[0034] As a preferred technical solution, the dual-layer temporal modeling module includes an intra-layer ConvLSTM and an inter-layer ConvLSTM. The intra-layer ConvLSTM performs temporal modeling on the melt pool image sequence within a single layer and extracts intra-layer temporal features. The inter-layer ConvLSTM performs cross-layer temporal modeling on the intra-layer temporal features of multiple deposition layers to obtain inter-layer temporal features.

[0035] As a preferred technical solution, the intermediate feature fusion module performs channel compression and weight calibration on the intra-layer temporal features and inter-layer temporal features respectively, and splices them to obtain the fused features, represented as:

[0036] ;

[0037] in, This represents the result of expanding the intra-layer temporal features in the channel dimension. and These represent mapping operations on intra-layer temporal features and inter-layer temporal features, respectively.

[0038] As a preferred technical solution, the adaptive pooling classification module includes a GeM branch, an attention pooling branch, a global statistics branch, and a dynamic gating network;

[0039] The dynamic gating network adaptively generates the weights of each branch based on the fusion features of the input, and performs weighted fusion of the output features of each branch to output the predicted surface roughness level, which is represented as:

[0040] ;

[0041] ;

[0042] ;

[0043] in, This represents a dynamic gating network. , , These represent the feature vectors output by the GeM branch, the attention pooling branch, and the global statistics branch, respectively. , , This represents the corresponding weight value. Indicates a fully connected layer. This represents a regularization operation. This represents the final output roughness level prediction result.

[0044] The present invention also provides an online identification system for the side surface roughness of directional energy deposition thin-walled parts based on molten pool time-series images, for implementing the above-mentioned online identification method for the side surface roughness of directional energy deposition thin-walled parts based on molten pool time-series images, including: an image acquisition module, a position correspondence module, a training dataset construction module, a data preprocessing module, a model construction module, a model training module, and an identification result output module;

[0045] The image acquisition module is used to acquire in-situ molten pool timing images during the manufacturing process of directional energy deposition thin-walled parts;

[0046] The position correspondence module is used to construct the correspondence between the image position and the spatial position of the thin-walled component;

[0047] The training dataset construction module is used to construct a labeled training dataset based on the surface roughness parameters of the side surfaces of thin-walled parts;

[0048] The data preprocessing module is used to preprocess the training dataset;

[0049] The model building module is used to build a spatiotemporal feature adaptive fusion classification model, including a spatial feature extraction module, a melt pool dynamic fluctuation feature extraction module, a two-layer temporal modeling module, a mid-term feature fusion module, and an adaptive pooling classification module;

[0050] The spatial feature extraction module encodes the spatial features of multiple frames of molten pool images frame by frame to obtain a spatial feature map. The molten pool dynamic fluctuation feature extraction module extracts the fluctuation features of the spatial feature map and fuses them with the spatial features of the current frame to obtain enhanced frame-level features. The dual-layer temporal modeling module extracts intra-layer temporal features and inter-layer temporal features respectively. The intermediate feature fusion module performs channel compression and weight calibration on the intra-layer temporal features and inter-layer temporal features respectively, and splices them to obtain fused features. The adaptive pooling classification module generates the weights of each branch and performs weighted fusion on the output features of each branch to output the side surface roughness level prediction result.

[0051] The model training module is used to train the spatiotemporal feature adaptive fusion classification model based on the preprocessed training set to obtain a roughness recognition model.

[0052] The recognition result output module is used to output the side surface roughness level recognition result based on the roughness recognition model.

[0053] Compared with the prior art, the present invention has the following advantages and beneficial effects:

[0054] (1) By establishing the kinematic mapping relationship between the visual sampling frame rate and the deposition rate, this invention achieves accurate spatial registration between the time-series image of the molten pool and the surface roughness characterization results of the solid side, thus solving the problem that online image data and offline roughness labels are difficult to correspond.

[0055] (2) The present invention explicitly extracts the difference information between adjacent frames through the molten pool dynamic fluctuation feature extraction module, thereby enhancing the model’s sensitivity to dynamic phenomena such as local fluctuations in the molten pool, boundary abrupt changes and brightness migration.

[0056] (3) The present invention adopts a two-layer ConvLSTM structure to model the short-term dynamic evolution law within the layer and the thermal accumulation evolution law between the layers respectively. Compared with the method of using only static features or simple temporal aggregation, it can more fully explore the spatiotemporal dependence relationship contained in the roughness formation process.

[0057] (4) The present invention adopts a mid-term fusion strategy and an adaptive pooling classification module, which introduces inter-layer compact evolution features while maintaining high-dimensional dynamic information within the layer, and improves the discrimination ability between different roughness categories through multi-branch morphological representation and dynamic gating fusion mechanism.

[0058] (5) The present invention can realize online identification of the surface roughness level of the directional energy deposition thin-walled part, which can provide support for quality anomaly early warning, process parameter optimization and closed-loop control of the manufacturing process. Attached Figure Description

[0059] Figure 1 This is a schematic flowchart of the online identification method for side surface roughness of directional energy deposition thin-walled parts based on molten pool time-series images according to the present invention;

[0060] Figure 2 This is a schematic diagram of the overall architecture of the directional energy deposition manufacturing platform and the paraxial vision monitoring system of the present invention.

[0061] Figure 3 This is a schematic diagram illustrating the process of constructing a labeled training dataset for this invention.

[0062] Figure 4 This is a schematic diagram of the overall network architecture of the spatiotemporal feature adaptive fusion classification model of the present invention;

[0063] Figure 5 This is a schematic diagram of the network architecture of the molten pool dynamic fluctuation feature extraction module of the present invention;

[0064] Figure 6 This is a schematic diagram of the network architecture of the two-layer timing modeling module of the present invention;

[0065] Figure 7 This is a schematic diagram of the network architecture of the intermediate feature fusion module of the present invention;

[0066] Figure 8 This is a schematic diagram of the network architecture of the adaptive pooling classification module of the present invention. Detailed Implementation

[0067] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention.

[0068] Example 1

[0069] like Figure 1 As shown, this embodiment provides an online method for identifying the side surface roughness of a directional energy deposition thin-walled part based on a molten pool time-series image, including the following steps:

[0070] S1: Obtain in-situ molten pool time-series images during the manufacturing process of directional energy deposition thin-walled parts; establish the correspondence between image position and the spatial position of the thin-walled part based on visual sampling frame rate and deposition speed; and construct a labeled training dataset based on the surface roughness parameters of the side surface of the thin-walled part.

[0071] like Figure 2 As shown, a directional energy deposition (DOD) manufacturing platform and a rangefinder vision monitoring system are constructed. The DOD manufacturing platform includes a robot, a laser, a powder feeding system, and a cladding head. Preferably, a FANUC Robot M-20iD robot, an RFL-C10000 fiber-coupled semiconductor laser, an RC / PGF / D dual-cylinder parallel powder feeder, and an RC52-CN9 laser cladding head can be used to complete the deposition manufacturing of thin-walled parts. The rangefinder vision monitoring system includes a camera gimbal and an industrial camera mounted on the end bracket of the robotic arm. Preferably, a Mecaweld-M2 welding monitoring camera can be used to acquire grayscale image sequences of the molten pool in situ during the manufacturing process.

[0072] In this embodiment, the correspondence between the in-situ molten pool timing image and the spatial position of the thin-walled part is achieved based on the kinematic mapping relationship established by the image sampling frame rate and the printing speed. The kinematic mapping relationship satisfies:

[0073] ;

[0074] in, Indicates the number of times since the laser was turned on. Frame image, Indicates the first The length of the distance from the starting point to the acquisition point corresponding to the frame image. For printing speed, The image sampling frame rate;

[0075] like Figure 3 As shown, the labels of the training dataset are determined by the surface roughness parameters of the side surfaces of the thin-walled parts. The surface roughness parameters can be obtained using an optical microscope or a laser confocal microscopy measurement device. Specifically, after deposition, the fabricated thin-walled part is placed under a surface morphology measurement device to perform three-dimensional morphology characterization of its side surface, and the surface roughness parameters are extracted as label values, defined as follows:

[0076] ;

[0077] in, Indicates the measured area. Point Surface height at that location;

[0078] S2: To improve the quality of the model input and reduce irrelevant background interference, the training dataset is preprocessed. The preprocessing includes image registration, sliding window sampling, region of interest cropping, image enhancement and size unification.

[0079] In this embodiment, sliding window sampling is performed by extracting a fixed number of consecutive images from a continuous melt pool image sequence according to a preset window length, which is then used as a sample sequence. The number of frames corresponding to a single sample sequence can be determined by the following formula:

[0080] ;

[0081] in, The number of frames in the sample sequence. For the corresponding deposition length, For printing speed, The frame rate is the image sampling rate.

[0082] In this embodiment, the region of interest of the original molten pool image is cropped to remove invalid background areas. Contrast-limited adaptive histogram equalization is used to enhance the image local contrast, so as to highlight the dark solidified areas and spatter particles and suppress overexposed areas. The enhanced image is then uniformly scaled to a preset size of 224×224.

[0083] S3: As Figure 4 As shown, a spatiotemporal feature adaptive fusion classification model is constructed. The spatiotemporal feature adaptive fusion classification model includes a spatial feature extraction module, a melt pool dynamic fluctuation feature extraction module, a two-layer temporal modeling module, a mid-term feature fusion module, and an adaptive pooling classification module.

[0084] In this embodiment, the spatial feature extraction module is used to perform frame-by-frame spatial feature encoding on multiple frames of molten pool images to obtain a spatial feature map characterizing the molten pool morphology information. Specifically, EfficientNetV2-S can be used as the backbone network to complete spatial feature extraction. The spatial feature extraction module encodes each frame of the molten pool image to obtain a frame-level spatial feature representation. Let the th frame be the spatial feature representation. Layer The spatial features corresponding to the frame image are ;

[0085] like Figure 5 As shown, the molten pool dynamic fluctuation feature extraction module is used to extract fluctuation features, calculate the difference information between spatial features of adjacent frames, and fuse the difference information with the spatial features of the current frame to enhance the model's ability to represent short-term dynamic fluctuations of the molten pool.

[0086] In this embodiment, the fluctuation features extracted by the molten pool dynamic fluctuation feature extraction module are represented as follows:

[0087] ;

[0088] The original spatial features and fluctuation features of the current frame are then concatenated along the channel dimension to obtain the joint features, represented as:

[0089] ;

[0090] Then, through 1×1 convolution, batch normalization, and activation function mapping, the enhanced frame-level features are obtained, represented as:

[0091] ;

[0092] like Figure 6 As shown, the dual-layer temporal modeling module includes intra-layer ConvLSTM and inter-layer ConvLSTM, which extract intra-layer dynamic evolution information and inter-layer cumulative evolution information, respectively. Intra-layer ConvLSTM is used to perform temporal modeling on the melt pool image sequence within a single layer to extract the dynamic evolution features during the single-layer deposition process; inter-layer ConvLSTM is used to perform cross-layer temporal modeling on the intra-layer temporal features of multiple depositional layers to characterize the cumulative effects and evolutionary relationships between different depositional layers.

[0093] Specifically, the intra-layer ConvLSTM is used to recursively model the continuous frame sequence within a single layer to obtain five layers of intra-layer temporal features. The five layers of intra-layer temporal features are then input into the inter-layer ConvLSTM in order to obtain inter-layer evolution features. In this embodiment, ConvLSTM can simultaneously extract spatial structure features and temporal evolution features from the melt pool image sequence.

[0094] like Figure 7As shown, the intermediate feature fusion module is used to perform channel compression and weight calibration on intra-layer and inter-layer temporal features respectively, and then concatenates the processed intra-layer and inter-layer temporal features along the channel dimension to form a unified fused feature. The intermediate fusion adopts a channel concatenation method instead of element-wise addition. The fused feature is represented as follows:

[0095] ;

[0096] in, This represents the result of expanding the intra-layer temporal features in the channel dimension. and These represent mapping operations on intra-layer temporal features and inter-layer temporal features, respectively.

[0097] Specifically, the intra-layer features are expanded from 320 channels to 128 channels, and the inter-layer features are compressed from 16 channels to 32 channels. Finally, they are spliced ​​together to form a 160-channel fused feature representation. In this way, the dynamic information within the layers can be preserved while introducing the cumulative evolution information between layers, forming a unified fused feature representation.

[0098] like Figure 8 As shown, the adaptive pooling classification module includes a GeM branch, an attention pooling branch, a global statistics branch, and a dynamic gating network. The dynamic gating network adaptively generates the weights of each branch based on the input fusion features and performs weighted fusion on the output features of each branch to output the predicted result of the surface roughness level.

[0099] In this embodiment, the dynamic gating network adaptively generates the weights of each branch based on the input features, and performs weighted fusion on the output features of each branch to produce the predicted surface roughness level of the output side. The weighted fusion is expressed as follows:

[0100] ;

[0101] ;

[0102] ;

[0103] in, This represents a dynamic gating network. , , These represent the feature vectors output by the GeM branch, the attention pooling branch, and the global statistics branch, respectively. , , This represents the corresponding weight value. Indicates a fully connected layer. This represents a regularization operation. This represents the final output roughness level prediction result.

[0104] S4: Use the training dataset to train the spatiotemporal feature adaptive fusion classification model to obtain the trained roughness recognition model;

[0105] S5: In the online identification stage, the in-situ molten pool time sequence image of the manufacturing process to be identified is acquired. After preprocessing in the same way as in the training stage, it is input into the trained roughness identification model and the side surface roughness level identification result corresponding to the current deposition position is output.

[0106] In this embodiment, the spatiotemporal feature adaptive fusion classification model is trained using the constructed labeled training dataset to obtain a trained roughness recognition model. The training device is an NVIDIA GeForce RTX 4070 SUPER, the batch size is 2, the number of training epochs is 200, the optimizer is AdamW, and the initial learning rate is... The weight decays to The minimum learning rate is The loss function used is CrossEntropyLoss, and an automatic mixed-precision training strategy is employed.

[0107] ;

[0108] ;

[0109] in, Indicates category weight, The label smoothing coefficient is represented by [1.0, 0.5, 1.0], and the label smoothing coefficient is preferably set to 0.05.

[0110] In the online identification phase, the in-situ molten pool time-series image of the manufacturing process to be identified is acquired in real time. After preprocessing in the same way as in the training phase, it is input into the trained roughness identification model and outputs the side surface roughness level identification result corresponding to the current deposition position. The output result can be a discretized roughness level, such as class0, class1, class2. The identification result can be further used for early warning of abnormal surface quality areas, adjustment of process parameters, or closed-loop control of the manufacturing process.

[0111] This invention can effectively mine the dynamic evolution characteristics of the molten pool, improve the accuracy and real-time performance of online identification of the side surface roughness of directional energy deposition thin-walled parts, and provide support for quality early warning and closed-loop control of the forming process.

[0112] Example 2

[0113] This embodiment provides an online identification system for the side surface roughness of directional energy deposition thin-walled parts based on molten pool time-series images, used to implement the online identification method for the side surface roughness of directional energy deposition thin-walled parts based on molten pool time-series images in Embodiment 1 above. The system includes: an image acquisition module, a position correspondence module, a training dataset construction module, a data preprocessing module, a model construction module, a model training module, and an identification result output module.

[0114] In this embodiment, the image acquisition module is used to acquire in-situ molten pool timing images during the manufacturing process of directional energy deposition thin-walled parts;

[0115] Specifically, the image acquisition module includes a camera gimbal and an industrial camera mounted on the end effector of the robotic arm, used to acquire in-situ molten pool timing images.

[0116] In this embodiment, the position correspondence module is used to construct the correspondence between the image position and the spatial position of the thin-walled component;

[0117] In this embodiment, the training dataset construction module is used to construct a labeled training dataset based on the surface roughness parameters of the side surface of the thin-walled part;

[0118] In this embodiment, the data preprocessing module is used to preprocess the training dataset;

[0119] In this embodiment, the model building module is used to build a spatiotemporal feature adaptive fusion classification model, including a spatial feature extraction module, a melt pool dynamic fluctuation feature extraction module, a two-layer temporal modeling module, a mid-term feature fusion module, and an adaptive pooling classification module;

[0120] In this embodiment, the spatial feature extraction module performs frame-by-frame spatial feature encoding on multiple frames of molten pool images to obtain a spatial feature map. The molten pool dynamic fluctuation feature extraction module extracts the fluctuation features of the spatial feature map and fuses them with the spatial features of the current frame to obtain enhanced frame-level features. The dual-layer temporal modeling module extracts intra-layer temporal features and inter-layer temporal features respectively. The mid-term feature fusion module performs channel compression and weight calibration on the intra-layer temporal features and inter-layer temporal features respectively, and splices them to obtain fused features. The adaptive pooling classification module generates the weights of each branch and performs weighted fusion on the output features of each branch to output the side surface roughness level prediction result.

[0121] The model training module is used to train the spatiotemporal feature adaptive fusion classification model based on the preprocessed training set to obtain the roughness recognition model;

[0122] The recognition result output module is used to output the side surface roughness level recognition result based on the roughness recognition model.

[0123] The above embodiments are preferred embodiments of the present invention, but the embodiments of the present invention are not limited to the above embodiments. Any changes, modifications, substitutions, combinations, or simplifications made without departing from the spirit and principle of the present invention shall be considered equivalent substitutions and shall be included within the protection scope of the present invention.

Claims

1. A method for online identification of side surface roughness of directional energy deposition thin-walled parts based on molten pool time-series images, characterized in that, Includes the following steps: In-situ molten pool time-series images of the manufacturing process of directional energy deposition thin-walled parts are obtained, the correspondence between the image positions and the spatial positions of the thin-walled parts is constructed, and a labeled training dataset is constructed based on the surface roughness parameters of the side surfaces of the thin-walled parts. Preprocess the training dataset; A spatiotemporal feature adaptive fusion classification model is constructed, including a spatial feature extraction module, a melt pool dynamic fluctuation feature extraction module, a two-layer temporal modeling module, a mid-term feature fusion module, and an adaptive pooling classification module; The spatial feature extraction module encodes the spatial features of multiple frames of molten pool images frame by frame to obtain a spatial feature map. The molten pool dynamic fluctuation feature extraction module extracts the fluctuation features of the spatial feature map and fuses them with the spatial features of the current frame to obtain enhanced frame-level features. The dual-layer temporal modeling module extracts intra-layer temporal features and inter-layer temporal features respectively. The intermediate feature fusion module performs channel compression and weight calibration on the intra-layer temporal features and inter-layer temporal features respectively, and splices them to obtain fused features. The adaptive pooling classification module generates the weights of each branch and performs weighted fusion on the output features of each branch to output the side surface roughness level prediction result. The spatiotemporal feature adaptive fusion classification model is trained based on the preprocessed training set to obtain a roughness recognition model. The roughness recognition model outputs the side surface roughness level recognition result.

2. The method for online identification of side surface roughness of directional energy deposition thin-walled parts based on molten pool time-series images according to claim 1, characterized in that, The correspondence between image positions and the spatial positions of thin-walled components is established and represented as follows: ； in, Indicates the number of times since the laser was turned on. Frame image, Indicates the first The length of the distance from the starting point to the acquisition point corresponding to the frame image. For printing speed, The frame rate is the image sampling rate.

3. The method for online identification of side surface roughness of directional energy deposition thin-walled parts based on molten pool time-series images according to claim 1, characterized in that, The surface roughness parameter of the side surface of a thin-walled part is expressed as: ； in, Indicates the measured area. Point The surface height at that location.

4. The method for online identification of side surface roughness of directional energy deposition thin-walled parts based on molten pool time-series images according to claim 1, characterized in that, The training dataset is preprocessed, including image registration, sliding window sampling, region of interest cropping, image enhancement, and size uniformity.

5. The method for online identification of side surface roughness of directional energy deposition thin-walled parts based on molten pool time-series images according to claim 4, characterized in that, Sliding window sampling specifically includes: A fixed number of consecutive images are extracted from a continuous melt pool image sequence according to a preset window length to form a sample sequence; The number of frames corresponding to a single sample sequence is represented as: ； in, The number of frames in the sample sequence. For the corresponding deposition length, For printing speed, The frame rate is the image sampling rate.

6. The method for online identification of side surface roughness of directional energy deposition thin-walled parts based on molten pool time-series images according to claim 1, characterized in that, The dynamic fluctuation feature extraction module of the molten pool extracts the fluctuation features of the spatial feature map, specifically represented as follows: ； in, Indicates the first Layer Spatial features corresponding to the frame image; The fluctuation features are concatenated with the original spatial features of the current frame in the channel dimension to obtain the joint features, represented as: ； Then, through convolution, batch normalization, and activation function mapping, the enhanced frame-level features are obtained, represented as: ； in, Represents a 1×1 convolution. Indicates batch normalization, This represents the activation function. Indicates joint features, This represents the enhanced frame-level features.

7. The method for online identification of side surface roughness of directional energy deposition thin-walled parts based on molten pool time-series images according to claim 1, characterized in that, The two-layer temporal modeling module includes intra-layer ConvLSTM and inter-layer ConvLSTM. Intra-layer ConvLSTM performs temporal modeling on the melt pool image sequence within a single layer and extracts intra-layer temporal features. Inter-layer ConvLSTM performs cross-layer temporal modeling on the intra-layer temporal features of multiple deposition layers to obtain inter-layer temporal features.

8. The method for online identification of side surface roughness of directional energy deposition thin-walled parts based on molten pool time-series images according to claim 1, characterized in that, The intermediate feature fusion module performs channel compression and weight calibration on intra-layer and inter-layer temporal features respectively, and concatenates them to obtain the fused feature, which is represented as follows: ； in, This represents the result of expanding the intra-layer temporal features in the channel dimension. and These represent mapping operations on intra-layer temporal features and inter-layer temporal features, respectively.

9. The method for online identification of side surface roughness of directional energy deposition thin-walled parts based on molten pool time-series images according to claim 1, characterized in that, The adaptive pooling classification module includes the GeM branch, the attention pooling branch, the global statistics branch, and the dynamic gating network; The dynamic gating network adaptively generates the weights of each branch based on the fusion features of the input, and performs weighted fusion of the output features of each branch to output the predicted surface roughness level, which is represented as: ； ； ； in, This represents a dynamic gating network. , , These represent the feature vectors output by the GeM branch, the attention pooling branch, and the global statistics branch, respectively. , , This represents the corresponding weight value. Indicates a fully connected layer. This represents a regularization operation. This represents the final output roughness level prediction result.

10. An online identification system for side surface roughness of directional energy deposition thin-walled parts based on molten pool time-series images, characterized in that, The method for online identification of side surface roughness of directional energy deposition thin-walled parts based on molten pool time-series images as described in any one of claims 1-9 includes: an image acquisition module, a position correspondence module, a training dataset construction module, a data preprocessing module, a model construction module, a model training module, and an identification result output module. The image acquisition module is used to acquire in-situ molten pool timing images during the manufacturing process of directional energy deposition thin-walled parts; The position correspondence module is used to construct the correspondence between the image position and the spatial position of the thin-walled component; The training dataset construction module is used to construct a labeled training dataset based on the surface roughness parameters of the side surfaces of thin-walled parts; The data preprocessing module is used to preprocess the training dataset; The model building module is used to build a spatiotemporal feature adaptive fusion classification model, including a spatial feature extraction module, a melt pool dynamic fluctuation feature extraction module, a two-layer temporal modeling module, a mid-term feature fusion module, and an adaptive pooling classification module; The spatial feature extraction module encodes the spatial features of multiple frames of molten pool images frame by frame to obtain a spatial feature map. The molten pool dynamic fluctuation feature extraction module extracts the fluctuation features of the spatial feature map and fuses them with the spatial features of the current frame to obtain enhanced frame-level features. The dual-layer temporal modeling module extracts intra-layer temporal features and inter-layer temporal features respectively. The intermediate feature fusion module performs channel compression and weight calibration on the intra-layer temporal features and inter-layer temporal features respectively, and splices them to obtain fused features. The adaptive pooling classification module generates the weights of each branch and performs weighted fusion on the output features of each branch to output the side surface roughness level prediction result. The model training module is used to train the spatiotemporal feature adaptive fusion classification model based on the preprocessed training set to obtain a roughness recognition model. The recognition result output module is used to output the side surface roughness level recognition result based on the roughness recognition model.

Images