Sheet metal part deformation stress distribution prediction method based on multispectral neural network fusion

Abstract

The invention discloses a multispectral neural network fused sheet metal part deformation stress distribution prediction method, and belongs to the technical field of sheet metal part stress detection. In order to solve the problem that non-contact full-field prediction is difficult to realize by simultaneously utilizing a deformation field and a temperature field in existing sheet metal part stress distribution prediction, the method comprises the following steps: acquiring visible light, a first infrared multispectral image sequence and a second infrared multispectral image sequence when a sheet metal part bears an external load, and calculating deformation field data according to displacement of a mark point in a visible light image; calculating temperature field data according to a radiation brightness difference value of the two infrared band images, stacking a normalized deformation field and a normalized temperature field in a channel dimension to form a multi-channel feature tensor, inputting the multi-channel feature tensor into a three-dimensional convolutional neural network comprising at least three convolutional layers and two pooling layers, outputting a stress distribution probability graph corresponding to each stress value interval, and obtaining a stress distribution probability graph corresponding to each stress value interval; and taking a maximum probability corresponding interval to generate a final stress distribution prediction map. The method is mainly used for the non-contact prediction of the surface stress distribution of the sheet metal part under the static or dynamic load.

Description

Technical Field

[0001] This invention relates to the field of sheet metal stress detection technology. More specifically, this invention relates to a method for predicting the deformation stress distribution of sheet metal parts using a multispectral neural network fusion approach. Background Technology

[0002] Accurately obtaining the surface stress distribution under external loads is crucial for structural strength assessment and fatigue life prediction of sheet metal parts. Currently, commonly used stress measurement methods mainly include strain gauge electrical measurement and finite element simulation analysis. Strain gauge electrical measurement obtains local strain values ​​by attaching strain gauges to the surface of the sheet metal part, and then converts them into stress using Hooke's law. However, this method can only obtain stress data at discrete points and cannot reflect the continuous stress distribution across the entire surface of the sheet metal part. For sheet metal parts with geometric features such as openings, welds, and bends, stress concentration areas are often difficult to predict. If the strain gauge placement does not cover these areas, critical high-stress information will be missed. Furthermore, the strain gauge attachment process requires high precision and is prone to detachment or failure under dynamic loads or high-temperature environments.

[0003] Finite element method (FEM) simulation analysis requires known load boundary conditions and material constitutive models for calculation. In practice, sheet metal parts are subjected to complex loads, boundary conditions are difficult to determine precisely, and factors such as the plastic behavior of the material and residual stress can reduce the accuracy of simulation results. To calibrate the simulation model, actual strain gauge data is usually still required. Therefore, this method cannot independently achieve rapid, non-contact prediction of stress distribution.

[0004] In recent years, digital image correlation methods have been used to measure the deformation field on the surface of sheet metal parts and attempt to infer the stress field from the deformation field. However, due to the thermal effects generated by sheet metal parts during stress, especially under dynamic or cyclic loading, the thermal stress caused by local temperature rise can affect the final stress distribution. Relying solely on deformation field data cannot distinguish the contributions of mechanical stress and thermal stress, leading to significant deviations in the inference results. On the other hand, infrared thermal imaging technology can measure the surface temperature field, but single-band infrared thermography is easily affected by ambient reflected radiation, making it difficult to obtain absolute temperature values, which in turn affects the quantitative calculation of thermal stress. When fusing deformation field and temperature field analysis, the two types of data have different physical dimensions and spatial resolutions, lacking effective fusion methods to simultaneously utilize their complementary information. These difficulties make it difficult for existing technologies to achieve non-contact, full-field, and high-precision prediction of sheet metal deformation stress distribution. Summary of the Invention

[0005] One object of the present invention is to solve at least the above-mentioned problems and to provide at least the advantages that will be described later.

[0006] To achieve these objectives and other advantages according to the present invention, a method for predicting the deformation stress distribution of sheet metal parts by multispectral neural network fusion is provided, comprising the following steps: A multispectral camera is used to acquire a multispectral image sequence of sheet metal parts under external load. The multispectral image sequence includes a visible light band image, a first infrared band image, and a second infrared band image. The visible light band image is used to identify the displacement of the marker points on the surface of the sheet metal parts. The first infrared band image and the second infrared band image correspond to two different thermal radiation wavelength ranges, respectively. The deformation field data of the sheet metal part surface is calculated based on the displacement of the marked points in the visible light band image, and the temperature field data of the sheet metal part surface is calculated based on the difference in radiance between the first infrared band image and the second infrared band image. The temperature field data is used to characterize the thermal stress related distribution. The deformation field data and the temperature field data are normalized respectively, and the normalized deformation field data and temperature field data are stacked in the time dimension and the channel dimension to form a multi-channel feature tensor. The multi-channel feature tensor is input into a pre-trained three-dimensional convolutional neural network, which contains at least three convolutional layers and at least two pooling layers. The three-dimensional convolutional neural network outputs a stress distribution probability map that matches the geometric dimensions of the sheet metal part. The value of each pixel in the stress distribution probability map represents the probability that the position belongs to a different stress value range. Based on the probability values ​​of each pixel in the stress distribution probability map, the stress value range corresponding to the highest probability is selected to generate the final stress distribution prediction map.

[0007] Preferably, the method further includes: pre-setting marking points on the surface of the cabinet sheet metal parts before acquiring the multispectral image sequence; the marking points are circular high reflectivity spots with a diameter of 2-5mm, evenly distributed in a grid pattern, with a grid spacing of 50-100mm; the marking points avoid the weld seams, rounded corners and opening areas of the sheet metal parts.

[0008] Preferably, the step of calculating the deformation field data of the sheet metal surface based on the displacement of marked points in the visible light band image specifically includes: The visible light band image sequence is matched frame-to-frame using digital image correlation to identify the sub-pixel coordinates of the marker points in each frame image. Based on the reference coordinates of the marked points in the initial frame image and the displacement coordinates of the current frame, the actual physical displacement is converted using camera calibration parameters; The discrete displacement data is processed into continuous data using a thin plate spline interpolation algorithm to generate full-surface continuous deformation field data of the sheet metal part.

[0009] Preferably, the step of calculating the temperature field data of the sheet metal surface based on the difference in radiance between the first infrared band image and the second infrared band image specifically includes: Spatial registration and Gaussian filtering noise reduction are performed on the first infrared band image and the second infrared band image; Based on Planck's radiation law, the difference in radiance between the two bands is used to eliminate environmental reflected radiation interference, and the absolute temperature value of each pixel is calculated. Spatially smooth the temperature data through interpolation to generate single-channel continuous temperature field data that characterizes the thermal stress distribution on the surface of sheet metal parts.

[0010] Preferably, the normalization process uses the min-max normalization method, with the specific formula as follows: Where x represents the deformation field data or temperature field data before normalization, x min x is the minimum value of the corresponding data. max The maximum value of the corresponding data is [0,1]. The normalized data range is [0,1].

[0011] Preferably, the normalized deformation field data and temperature field data are stacked in both the time and channel dimensions, specifically: The normalized deformation field data corresponding to N consecutive frames of images are arranged in order of acquisition time to obtain the deformation field time series tensor. The normalized temperature field data corresponding to N consecutive frames of images are arranged in the order of acquisition time to obtain the temperature field time series tensor. The deformation field time series tensor is used as the first channel group, and the temperature field time series tensor is used as the second channel group. They are spliced ​​together in the channel dimension to form a four-dimensional feature tensor with shape (N, H, W, 2), where H and W are the number of pixels in the height direction and the number of pixels in the width direction of the sheet metal surface image, respectively.

[0012] Preferably, the three-dimensional convolutional neural network comprises an input layer, a first convolutional layer, a first pooling layer, a second convolutional layer, a second pooling layer, a third convolutional layer, a transposed convolutional upsampling layer, and a softmax output layer connected in sequence. The input layer receives the four-dimensional feature tensor; The first to third convolutional layers all use 3×3×3 convolutional kernels, with 32, 64 and 128 output channels respectively, which are used to extract spatiotemporal joint features; Both pooling layers use a 2×2×2 max pooling kernel with a stride of 2 for downsampling and compressing feature maps. The transposed convolutional upsampling layer restores the feature map to a size consistent with the input image spatial dimensions. The softmax output layer classifies each pixel and outputs the probability distribution of each stress interval, thereby generating the stress distribution probability map.

[0013] Preferably, the training process of the three-dimensional convolutional neural network includes: A training dataset is constructed, which contains sample data of multiple sets of cabinet sheet metal parts. Each set of sample data includes: a multispectral image sequence, deformation field data and temperature field data calculated by the multispectral image sequence, and the actual stress data corresponding to the set of samples obtained by actual measurement by strain gauges. The four-dimensional feature tensor is used as the network input, and the one-hot encoded label corresponding to the real stress data is used as the supervision signal. The model is trained using the cross-entropy loss function, and the network weights are iteratively updated using gradient descent until the loss function converges.

[0014] Preferably, the stress value range is divided according to the material properties of the cabinet sheet metal parts, and includes at least four levels: 0~0.4 times the yield strength: low stress zone; 0.4~0.6 times the yield strength: medium to low stress zone; 0.6~0.8 times the yield strength: medium to high stress zone; 0.8 times the yield strength to the material's fracture strength: high stress zone.

[0015] Preferably, the external load is the type of load that the cabinet sheet metal parts actually bear during operation, including static load and dynamic load; The static load has a loading range of 0~5000N, the dynamic load has a loading frequency of 1~10Hz, and the relative position of the multispectral camera and the sheet metal part is kept fixed during the loading process, with a shooting distance of 1~2m.

[0016] The present invention has at least the following beneficial effects: The method provided by this invention simultaneously acquires images of visible light and two different infrared bands using a multispectral camera, obtaining deformation field data and temperature field data (after eliminating environmental reflection interference) respectively. These two data are then normalized and stacked into a multi-channel feature tensor, which is input into a three-dimensional convolutional neural network for stress interval classification and prediction. Compared to methods that rely solely on the deformation field to infer stress, this method introduces temperature field information, enabling the differentiation between the contributions of mechanical and thermal stresses, making the prediction results closer to the actual stress distribution. Utilizing the difference in radiance between the two infrared bands to invert absolute temperature effectively reduces the impact of environmental reflection radiation on temperature measurement accuracy, thereby improving the reliability of thermal stress calculation. Employing a three-dimensional convolutional neural network for joint feature extraction of the deformation and temperature fields avoids the tedious process of manually designing fusion rules. The stress distribution probability map output by the network can cover the entire surface of the sheet metal part, achieving full-field non-contact stress distribution prediction. Compared to strain gauge electrical measurement methods, this method eliminates the need to attach numerous sensors to the sheet metal surface, reducing wiring workload and the risk of missed sensor placement, making it particularly suitable for sheet metal parts with complex geometric features (such as openings, welds, and bends). Compared to finite element simulation analysis, this method directly extrapolates from measured deformation and temperature data, without requiring precise knowledge of load boundary conditions and material constitutive models. This allows for a more realistic reflection of the stress state of sheet metal parts under actual working conditions. Furthermore, this method can process image sequences under both static and dynamic loads, and through feature stacking in the time dimension, it can capture the evolution trend of stress over time.

[0017] Other advantages, objectives and features of the present invention will become apparent in part from the following description, and in part from those skilled in the art through study and practice of the invention. Detailed Implementation

[0018] The present invention will be further described in detail below with reference to embodiments, so that those skilled in the art can implement it based on the description.

[0019] This invention provides a method for predicting the deformation stress distribution of sheet metal parts using multispectral neural network fusion, comprising the following steps: S1. Use a multispectral camera to acquire a multispectral image sequence of sheet metal parts under external load. The multispectral image sequence includes a visible light band image, a first infrared band image, and a second infrared band image. The visible light band image is used to identify the displacement of the marker points on the surface of the sheet metal parts. The first infrared band image and the second infrared band image correspond to two different thermal radiation wavelength ranges, respectively. The multispectral camera can be a commercially available multispectral camera capable of simultaneously acquiring visible light and two infrared bands. The wavelength range corresponding to the visible light band is typically 400–700 nm. The first infrared band can be selected within the mid-wave infrared range of 3–5 μm, and the second infrared band can be selected within the long-wave infrared range of 8–12 μm. The external load can be a static load or a dynamic load. The loading range of the static load can be 0–5000 N, and the loading frequency of the dynamic load can be 1–10 Hz. During acquisition, the shooting distance of the multispectral camera can be set to 1–2 m, and the relative position of the camera and the sheet metal part should be kept fixed. The camera can be mounted on an adjustable-height tripod or a fixed bracket, with the lens facing the surface of the sheet metal part. Before acquiring the image sequence, marker points can be preset on the surface of the sheet metal part. The marker points can be circular high-reflectivity spots with a diameter of 2–5 mm, evenly distributed in a grid pattern, with a grid spacing of 50–100 mm, and the marker points should avoid weld seams, rounded corners, and opening areas.

[0020] The multispectral image sequence acquisition process is as follows: While applying an external load to the sheet metal part, a multispectral camera is activated to continuously capture images, obtaining multiple frames of visible light images, a first infrared image, and a second infrared image containing a time series. The spatial resolution of each frame remains consistent, for example, set to 1024×1024 pixels. For the two infrared images, radiometric calibration is required beforehand to convert the image grayscale values ​​into radiance values. Parameter setting method: Based on the ambient temperature and the emissivity of the sheet metal part material, infrared bands with the following parameters can be selected: the first infrared band has a center wavelength of 4μm and a bandwidth of 1μm, and the second infrared band has a center wavelength of 10μm and a bandwidth of 2μm.

[0021] S2. Calculate the deformation field data of the sheet metal surface based on the displacement of the marked points in the visible light band image, and at the same time calculate the temperature field data of the sheet metal surface based on the difference in radiance between the first infrared band image and the second infrared band image. The temperature field data is used to characterize the thermal stress related distribution. The deformation field data is calculated based on the displacement of marker points in visible light images. The specific process is as follows: First, digital image correlation is used to perform inter-frame matching on the visible light image sequence to identify the sub-pixel coordinates of marker points in each frame. The initial frame coordinates of the marker points are used as the reference coordinates, and the difference between the coordinates of the marker points in subsequent frames and the reference coordinates is the pixel displacement. Then, the pixel displacement is converted into actual physical displacement in mm using camera calibration parameters (including focal length, pixel size, and working distance). Next, a thin-plate spline interpolation algorithm is used to process the discrete marker point displacement data into continuous data, generating continuous deformation field data for the entire surface of the sheet metal part. This data is a two-dimensional matrix, where each element represents the displacement of the corresponding pixel position. Possible interpolation algorithm parameters: The smoothing factor for thin-plate spline interpolation can be set to 0.01~0.1. The spatial resolution of the deformation field data is consistent with that of the visible light image. The marker material can be high-reflectivity aluminum foil or silver-plated plastic sheet, which is adhered to the surface of the sheet metal part. Assembly position: The marking points are pasted on the outer surface of the load-bearing area of ​​the sheet metal part, avoiding edges and holes.

[0022] Temperature field data is calculated based on the difference in radiance between the first and second infrared band images. The process is as follows: First, the first and second infrared images are spatially registered to ensure that the same physical point in both images corresponds to the same pixel coordinates. Then, Gaussian filtering is applied to the registered images for noise reduction. The size of the Gaussian kernel can be set to 3×3 pixels, and the standard deviation σ can be set to 1.0. Based on Planck's radiation law, the difference in radiance between the two bands is used to eliminate interference from environmental reflected radiation, and the absolute temperature value of each pixel is calculated. Specifically, the equivalent temperature value of environmental reflected radiation can be set to 25℃, and the emissivity of the sheet metal surface can be set to 0.85~0.95. After obtaining the temperature of each pixel, spatial smoothing interpolation (e.g., using bicubic interpolation) is performed on the temperature data to generate single-channel continuous temperature field data with a spatial resolution consistent with the visible light image. Each pixel value in this temperature field data represents the surface temperature at that location (in °C), thus characterizing the trend of thermal stress distribution.

[0023] S3. Normalize the deformation field data and the temperature field data respectively, and stack the normalized deformation field data and temperature field data in the time dimension and the channel dimension to form a multi-channel feature tensor. Specifically, the normalization process employs the min-max normalization method, with the following formula: Where x represents the deformation field data or temperature field data before normalization, x min x is the minimum value of the corresponding data. max The maximum value of the corresponding data is [0,1]. The normalized data range is [0,1].

[0024] For deformation field data, x min It can be set to 0 mm, x max It can be set to 1.2 times the maximum measured deformation of the sheet metal part. For example, when the maximum deformation is 5mm, x max Take 6mm.

[0025] For temperature field data, x min It can be set to the ambient temperature (e.g., 25℃), x max This can be set to the highest temperature of the sheet metal part under ultimate load (e.g., 80℃). During normalization, the entire data matrix is ​​traversed, and a linear transformation is performed on each element. This step requires no additional equipment and is implemented in software on a computer.

[0026] The normalized deformation field data and temperature field data are stacked in both the time and channel dimensions to form a multi-channel feature tensor. Specifically, the normalized deformation field data corresponding to N consecutive frames are arranged in chronological order of acquisition time to obtain a deformation field time-series tensor; the normalized temperature field data corresponding to N consecutive frames are arranged in chronological order of acquisition time to obtain a temperature field time-series tensor; the deformation field time-series tensor is used as the first channel group, and the temperature field time-series tensor is used as the second channel group, and they are concatenated in the channel dimension to form a four-dimensional feature tensor of shape (N, H, W, 2), where H and W are the number of pixels in the height and width directions of the sheet metal surface image, respectively. The spatial dimension of this tensor is consistent with the resolution of the sheet metal surface image. N is the number of time frames, which can be set to 5-10 frames. This step is performed in computer memory and requires no hardware modification.

[0027] S4. Input the multi-channel feature tensor into a pre-trained three-dimensional convolutional neural network. The three-dimensional convolutional neural network contains at least three convolutional layers and at least two pooling layers. The three-dimensional convolutional neural network outputs a stress distribution probability map that matches the geometric dimensions of the sheet metal part. The value of each pixel in the stress distribution probability map represents the probability that the position belongs to different stress value ranges. Specifically, the three-dimensional convolutional neural network includes an input layer, a first convolutional layer, a first pooling layer, a second convolutional layer, a second pooling layer, a third convolutional layer, a transposed convolutional upsampling layer, and a softmax output layer connected in sequence. The input layer receives a four-dimensional feature tensor; the first convolutional layer uses a 3×3×3 convolutional kernel, with 32 output channels and a stride of 1, and uses ReLU as the activation function; the first pooling layer uses a 2×2×2 max pooling kernel with a stride of 2; the second convolutional layer uses a 3×3×3 convolutional kernel, with 64 output channels; the second pooling layer uses a 2×2×2 max pooling kernel with a stride of 2; the third convolutional layer uses a 3×3×3 convolutional kernel, with 128 output channels; then a transposed convolutional upsampling layer is connected to restore the feature map to the same spatial size as the input image; finally, a softmax output layer is connected to classify each pixel and output the probability distribution of each stress value interval, thereby generating a stress distribution probability map.

[0028] The stress range can be divided according to the yield strength of the sheet metal material. For example, it can be divided into four ranges: 0 to 0.4 times the yield strength is the low stress range, 0.4 to 0.6 times the yield strength is the medium-low stress range, 0.6 to 0.8 times the yield strength is the medium-high stress range, and 0.8 times the yield strength to the material's fracture strength is the high stress range. The yield strength can be taken as 235 MPa (for ordinary carbon steel).

[0029] Each pixel in the probability graph outputs a probability vector of length 4, with the four components corresponding to the probabilities of the four stress intervals, and the sum of the probabilities is 1.

[0030] The training process of the 3D convolutional neural network includes: constructing a training dataset, where each sample contains a multispectral image sequence, calculated deformation and temperature fields, and actual stress data measured using strain gauges (as labels). Labels are encoded using one-hot encoding. The loss function is cross-entropy loss, the optimizer can be Adam, the learning rate can be set to 0.001, the batch size can be set to 8, and the number of training epochs can be set to 100, until the loss function converges. The network can be deployed on a computer equipped with a GPU. S5. Select the stress value range corresponding to the highest probability based on the probability value of each pixel in the stress distribution probability map, and generate the final stress distribution prediction map.

[0031] For each pixel location on the stress distribution probability map, obtain the probability values ​​of the four stress intervals corresponding to that location (e.g., p1, p2, p3, p4). Find the interval index corresponding to the maximum value among them, and assign the representative value of that interval (e.g., the midpoint of the interval) or interval label to that pixel. After traversing all pixels, a stress distribution prediction map with the same size as the input image is obtained. The value or color of each pixel in the map represents the predicted stress level at that location. For example, low stress areas can be displayed in blue, medium-low stress areas in green, medium-high stress areas in yellow, and high stress areas in red. Finally, this prediction map is output to visually evaluate the stress distribution of sheet metal parts under different loads. The prediction results can be compared with the finite element simulation calculation results, or with the stress values ​​measured at strain gauge points, to calculate the prediction accuracy (without specifying the percentage). A confusion matrix can be used to evaluate classification performance, and the Kappa coefficient can be calculated to verify the consistency between prediction and measurement.

[0032] The above technical solution, by fusing visible light and dual-band infrared images to obtain deformation field and temperature field information respectively, and then using a three-dimensional convolutional neural network for joint feature extraction and stress interval classification, can predict the stress distribution in the entire field without contacting the sheet metal surface. This avoids the limitation of sparse measurement points of traditional strain gauges. Furthermore, dual-band infrared thermometry can eliminate environmental reflection interference, improve the accuracy of temperature field data, and thus enhance the reliability of stress prediction.

[0033] <Example> Taking the side panel sheet metal of a certain type of communication cabinet as an example, the sheet metal material is DC01 cold-rolled steel plate with a thickness of 1.5mm, a yield strength of 210MPa, and a fracture strength of 300MPa. The sheet metal part has external dimensions of 800mm × 600mm, and its surface has two circular openings with a diameter of 10mm and a weld seam with a length of 200mm. Before the experiment, the surface of the sheet metal part was first cleaned, and then marking points were affixed to the surface. The marking points were circular high-reflectivity aluminum foil patches with a diameter of 3mm, evenly arranged in a grid pattern with a grid spacing of 80mm. The marking points avoided the weld seam, rounded corners, and opening areas, and a total of 48 marking points were affixed.

[0034] The sheet metal part was fixed on the loading platform of the universal testing machine, and a static load was applied using a four-point bending loading method. A commercially available multispectral camera capable of simultaneously acquiring visible light, mid-wave infrared (3μm~5μm), and long-wave infrared (8μm~12μm) wavelengths was used. The camera was mounted on a tripod 1.5m directly in front of the loading platform, with the lens optical axis perpendicular to the sheet metal surface. The camera's image resolution was set to 1280×1024 pixels, and the frame rate was 5 frames / second. An initial frame was captured as a baseline before loading. Then, a static load was applied at a rate of 100N / s, gradually increasing from 0N to 3000N. The load was held steady for 5 seconds at each of the following holding periods: 500N, 1000N, 1500N, 2000N, 2500N, and 3000N. Ten frames of multispectral images were continuously acquired during each holding period. Each frame sequence contained one visible light image, one mid-wave infrared image, and one long-wave infrared image.

[0035] For each set of image data under load, the following steps are performed. First, the sub-pixel coordinates of the marker points are extracted from the visible light image. Using digital image correlation, the current frame is matched with the initial frame to obtain the pixel displacement of each marker point in the horizontal and vertical directions. The pixel displacement is converted into actual physical displacement in mm using camera calibration parameters (focal length 25mm, pixel size 5.5μm, working distance 1.5m). After obtaining the displacement data of 48 discrete points, continuous deformation field data for the entire sheet metal surface is generated using a thin-plate spline interpolation algorithm. The interpolated grid spacing is 1mm, and the deformation field matrix size is 1280×1024. For infrared images, the mid-wave infrared image and the long-wave infrared image are spatially registered first, and then noise reduction is performed using a 3×3 pixel Gaussian filter kernel (σ=1.0). According to Planck's radiation law, the emissivity of the sheet metal surface is set to 0.90, and the equivalent ambient reflection temperature is set to 25℃. The absolute temperature value of each pixel is retrieved using the difference in radiance between the two bands. After obtaining the temperature matrix, spatial smoothing is performed using bicubic interpolation to generate single-channel temperature field data with the same resolution as the deformation field, in °C.

[0036] The deformation field and temperature field data are both normalized using a min-max method. For the deformation field, the minimum value is 0 mm, and the maximum value is 1.2 times the measured maximum deformation under the load (e.g., if the maximum deformation at 3000 N is 4.2 mm, then 5.0 mm is used). For the temperature field, the minimum value is the ambient temperature of 25℃, and the maximum value is 1.1 times the measured highest temperature under the load (e.g., if the highest temperature is 58℃, then 64℃ is used). After normalization, the feature tensors corresponding to five consecutive frames are stacked in the time dimension. Then, the deformation field is used as the first channel, and the temperature field is used as the second channel, and the tensors are stacked in the channel dimension to obtain a 5×1280×1024×2 four-dimensional feature tensor.

[0037] The feature tensor is input into a pre-trained 3D convolutional neural network. The network structure is as follows: the input layer receives a 5×1280×1024×2 tensor; the first convolutional layer uses a 3×3×3 convolutional kernel with 32 output channels; the first pooling layer uses 2×2×2 max pooling with a stride of 2; the second convolutional layer outputs 64 channels; the second pooling layer also uses 2×2×2 max pooling; the third convolutional layer outputs 128 channels; then a transposed convolutional upsampling layer is applied to restore the feature map size to 1280×1024; finally, a softmax output layer is used to obtain the probability of each pixel belonging to one of the four stress ranges. The four stress ranges are defined according to the yield strength of DC01 steel: 0~84MPa is the low stress range, 84~126MPa is the low-to-medium stress range, 126~168MPa is the medium-to-high stress range, and 168~300MPa is the high stress range. The network output probability map size is 1280×1024×4.

[0038] For each pixel, the stress range with the highest probability is selected as the prediction result for that point, generating the final stress distribution prediction map. In the prediction map, low-stress areas are displayed in blue, medium-low stress areas in green, medium-high stress areas in yellow, and high-stress areas in red. The prediction map is compared with the measured stress values ​​of actual strain gauges (a total of 12 strain gauges located in different stress regions) on the surface of the sheet metal part after the load is applied. The proportion of predicted stress ranges matching the measured ranges is within an acceptable range. This embodiment shows that the method of the present invention can achieve comprehensive prediction of the stress distribution on the surface of sheet metal parts under non-contact conditions, avoiding the problem of insufficient information in traditional point-based measurements.

[0039] Although embodiments of the present invention have been disclosed above, they are not limited to the applications listed in the specification and embodiments. They can be applied to various fields suitable for the present invention. For those skilled in the art, other modifications can be easily made. Therefore, without departing from the general concept defined by the claims and their equivalents, the present invention is not limited to the specific details and embodiments shown and described herein.

Claims

1. A method for predicting the deformation stress distribution of sheet metal parts using multispectral neural network fusion, characterized in that, Includes the following steps: A multispectral camera is used to acquire a multispectral image sequence of sheet metal parts under external load. The multispectral image sequence includes a visible light band image, a first infrared band image, and a second infrared band image. The visible light band image is used to identify the displacement of the marker points on the surface of the sheet metal parts. The first infrared band image and the second infrared band image correspond to two different thermal radiation wavelength ranges, respectively. The deformation field data of the sheet metal part surface is calculated based on the displacement of the marked points in the visible light band image, and the temperature field data of the sheet metal part surface is calculated based on the difference in radiance between the first infrared band image and the second infrared band image. The temperature field data is used to characterize the thermal stress related distribution. The deformation field data and the temperature field data are normalized respectively, and the normalized deformation field data and temperature field data are stacked in the time dimension and the channel dimension to form a multi-channel feature tensor. The multi-channel feature tensor is input into a pre-trained three-dimensional convolutional neural network, which contains at least three convolutional layers and at least two pooling layers. The three-dimensional convolutional neural network outputs a stress distribution probability map that matches the geometric dimensions of the sheet metal part. The value of each pixel in the stress distribution probability map represents the probability that the position belongs to a different stress value range. Based on the probability values ​​of each pixel in the stress distribution probability map, the stress value range corresponding to the highest probability is selected to generate the final stress distribution prediction map.

2. The method for predicting the deformation stress distribution of sheet metal parts using multispectral neural network fusion as described in claim 1, characterized in that, Also includes: Before acquiring the multispectral image sequence, pre-marked points are set on the surface of the sheet metal part; the marked points are circular high reflectivity spots with a diameter of 2-5 mm, evenly distributed in a grid pattern, with a grid spacing of 50-100 mm; the marked points avoid the weld seams, rounded corners and opening areas of the sheet metal part.

3. The method for predicting the deformation stress distribution of sheet metal parts using multispectral neural network fusion as described in claim 1, characterized in that, The calculation of the deformation field data of the sheet metal surface based on the displacement of marked points in the visible light band image specifically includes: The visible light band image sequence is matched frame-to-frame using digital image correlation to identify the sub-pixel coordinates of the marker points in each frame image. Based on the reference coordinates of the marked points in the initial frame image and the displacement coordinates of the current frame, the actual physical displacement is converted using camera calibration parameters; The discrete displacement data is processed into continuous data using a thin plate spline interpolation algorithm to generate full-surface continuous deformation field data of the sheet metal part.

4. The method for predicting the deformation stress distribution of sheet metal parts using multispectral neural network fusion as described in claim 1, characterized in that, The calculation of the temperature field data of the sheet metal surface based on the difference in radiance between the first infrared band image and the second infrared band image specifically includes: Spatial registration and Gaussian filtering noise reduction are performed on the first infrared band image and the second infrared band image; Based on Planck's radiation law, the difference in radiance between the two bands is used to eliminate environmental reflected radiation interference, and the absolute temperature value of each pixel is calculated. Spatially smooth the temperature data through interpolation to generate single-channel continuous temperature field data that characterizes the thermal stress distribution on the surface of sheet metal parts.

5. The method for predicting the deformation stress distribution of sheet metal parts using multispectral neural network fusion as described in claim 1, characterized in that, The normalization process uses the min-max normalization method, and the specific formula is as follows: Where x represents the deformation field data or temperature field data before normalization, x min x is the minimum value of the corresponding data. max The maximum value of the corresponding data is [0,1]. The normalized data range is [0,1].

6. The method for predicting the deformation stress distribution of sheet metal parts using multispectral neural network fusion as described in claim 1, characterized in that, The normalized deformation field data and temperature field data are stacked in the time dimension and the channel dimension, specifically as follows: The normalized deformation field data corresponding to N consecutive frames of images are arranged in order of acquisition time to obtain the deformation field time series tensor. The normalized temperature field data corresponding to N consecutive frames of images are arranged in the order of acquisition time to obtain the temperature field time series tensor. The deformation field time series tensor is used as the first channel group, and the temperature field time series tensor is used as the second channel group. They are spliced ​​together in the channel dimension to form a four-dimensional feature tensor with shape (N, H, W, 2), where H and W are the number of pixels in the height direction and the number of pixels in the width direction of the sheet metal surface image, respectively.

7. The method for predicting the deformation stress distribution of sheet metal parts using multispectral neural network fusion as described in claim 6, characterized in that, The three-dimensional convolutional neural network comprises an input layer, a first convolutional layer, a first pooling layer, a second convolutional layer, a second pooling layer, a third convolutional layer, a transposed convolutional upsampling layer, and a softmax output layer connected in sequence. The input layer receives the four-dimensional feature tensor; The first to third convolutional layers all use 3×3×3 convolutional kernels, with 32, 64 and 128 output channels respectively, which are used to extract spatiotemporal joint features; Both pooling layers use a 2×2×2 max pooling kernel with a stride of 2 for downsampling and compressing feature maps. The transposed convolutional upsampling layer restores the feature map to a size consistent with the input image spatial dimensions. The softmax output layer classifies each pixel and outputs the probability distribution of each stress interval, thereby generating the stress distribution probability map.

8. The method for predicting the deformation stress distribution of sheet metal parts using multispectral neural network fusion as described in claim 6, characterized in that, The training process of the three-dimensional convolutional neural network includes: A training dataset is constructed, which contains multiple sets of sample data of sheet metal parts. Each set of sample data includes: a multispectral image sequence, deformation field data and temperature field data calculated from the multispectral image sequence, and the actual stress data corresponding to the set of samples obtained by actual measurement using strain gauges. The four-dimensional feature tensor is used as the network input, and the one-hot encoded label corresponding to the real stress data is used as the supervision signal. The model is trained using the cross-entropy loss function, and the network weights are iteratively updated using gradient descent until the loss function converges.

9. The method for predicting the deformation stress distribution of sheet metal parts using multispectral neural network fusion as described in claim 1, characterized in that, The stress value range is divided according to the material properties of the sheet metal part, and includes at least four levels: 0~0.4 times the yield strength: low stress zone; 0.4~0.6 times the yield strength: medium to low stress zone; 0.6~0.8 times the yield strength: medium to high stress zone; 0.8 times the yield strength to the material's fracture strength: high stress zone.

10. The method for predicting the deformation stress distribution of sheet metal parts using multispectral neural network fusion as described in claim 1, characterized in that, The external load refers to the type of load that the sheet metal part actually bears during operation, including static load and dynamic load; The static load has a loading range of 0~5000N, the dynamic load has a loading frequency of 1~10Hz, and the relative position of the multispectral camera and the sheet metal part is kept fixed during the loading process, with a shooting distance of 1~2m.

Images