Crane with stress analysis capability of boom structure

Abstract

The application belongs to the technical field of cranes, and specifically provides a crane with arm frame structure stress analysis capability, comprising: a collection module for collecting load data and working condition information of the automobile crane arm frame to obtain a characteristic vector for structure stress analysis; a prediction module for predicting based on an arm frame stress prediction model according to the characteristic vector to obtain an arm frame stress cloud map, a global maximum stress value and a plurality of regional maximum stress values; a mapping and labeling module for mapping the plurality of regional maximum stress values to the arm frame stress cloud map based on extreme value positioning and labeling; and an analysis module for performing stress analysis on the arm frame structure based on the labeled complete automobile crane arm frame stress cloud map and the global maximum stress value. The application can visualize the overall stress distribution while significantly improving the prediction accuracy of the stress concentration area, providing a reliable risk assessment basis for engineering decision-making, and has strong timeliness, which can effectively meet the response requirements of actual application scenarios.

Description

Technical Field

[0001] This invention relates to the field of crane technology, and more specifically to a crane with the capability of boom structure stress analysis. Background Technology

[0002] As truck crane boom structures continue to evolve towards larger, more complex, and highly nonlinear forms, their structural safety performance assessment faces increasingly severe challenges. While traditional structural stress analysis, centered on the finite element method (FEA), can guarantee high computational accuracy, it suffers from inherent limitations such as high computational costs and long analysis cycles. Especially under multiple load combinations, full-domain, full-condition stress simulation often requires tens of days or even longer, failing to meet the urgent demands of modern engineering for real-time analysis and efficient design iteration. Furthermore, traditional stress prediction methods typically output only a single stress value, lacking a visual representation of the overall stress distribution. Although local analysis accuracy is good, the overall control capability is insufficient, making it difficult to intuitively reflect stress concentration areas and distribution patterns, thus limiting its practical application value in engineering decision-making. Summary of the Invention

[0003] To address the aforementioned technical problems, this invention provides a crane with boom structure stress analysis capabilities. Based on a detail-enhanced dual-branch network-based boom stress prediction model for truck cranes, it predicts boom stress cloud maps, global maximum stress values, and maximum stress values ​​in multiple regions. The global maximum stress value and the maximum stress values ​​in multiple regions are mapped onto the boom stress cloud map and annotated for boom structure load-bearing capacity assessment. This avoids excessively large errors in the labeled data caused by directly reading values ​​from the stress cloud map. While visualizing the overall stress distribution, it significantly improves the prediction accuracy of stress concentration areas, providing a reliable risk assessment basis for engineering decisions. The prediction time is short, exhibiting strong timeliness and effectively meeting the response requirements of practical application scenarios.

[0004] The technical solution adopted in this invention is as follows:

[0005] This invention proposes a crane with boom structure stress analysis capabilities, comprising: a data acquisition module for acquiring load data and operating condition information of the truck crane boom to obtain a feature vector for structural stress analysis; a prediction module for predicting based on the feature vector using a boom stress prediction model to obtain a boom stress cloud map, a global maximum stress value, and multiple regional maximum stress values, wherein the boom stress prediction model is a detail-enhanced bi-branch network model, and the bi-branch of the boom stress prediction model includes a stress cloud map prediction branch and a stress value prediction branch; a mapping and annotation module for mapping multiple regional maximum stress values ​​to the boom stress cloud map based on extreme value localization and annotating it to obtain a fully annotated truck crane boom stress cloud map; and an analysis module for performing stress analysis on the boom structure based on the fully annotated truck crane boom stress cloud map and the global maximum stress value.

[0006] In addition, the crane with boom structure stress analysis capability proposed according to the present invention may also have the following additional technical features:

[0007] According to one embodiment of the present invention, the stress cloud map prediction network of the stress cloud map prediction branch is constructed based on a conditional generation mechanism; the stress value prediction network of the stress value prediction branch is constructed based on a multi-task learning network with an attention mechanism; and the stress prediction model of the boom of the truck crane is constructed by co-training the stress cloud map prediction network and the stress value prediction network.

[0008] According to one embodiment of the present invention, the stress cloud map prediction network includes a condition processing module, a feature fusion module, and a detail-preserving upsampling decoder. The condition processing module provides high-dimensional feature embeddings of load conditions to the feature fusion module and provides nonlinear mappings of conditional feature vectors to the Dropout layer of the detail-preserving upsampling decoder. The feature fusion module generates a high-dimensional feature map based on the high-dimensional feature embeddings provided by the condition processing module. The detail-preserving upsampling decoder includes multiple cascaded detail-preserving upsampling blocks, which reconstruct the boom stress cloud map based on the high-dimensional feature map using the detail-preserving upsampling blocks with the high-dimensional feature embeddings.

[0009] According to one embodiment of the present invention, the detail-preserving upsampling block includes a bilinear upsampling layer and a dual-path convolutional layer. The dual-path convolutional layer includes a main path convolutional layer and a detail path convolutional layer in parallel. After the feature maps extracted by the main path convolutional layer and the detail path convolutional layer are weighted and fused, they are residually connected with the feature map output by the bilinear upsampling layer of the detail-preserving upsampling block. During the process of reconstructing the boom stress cloud map by the detail-preserving upsampling decoder, a residual refinement mechanism is introduced to weightedly superimpose the initial boom stress cloud map generated by the independent residual network onto the boom stress cloud map.

[0010] According to one embodiment of the present invention, the input of the stress value prediction network includes a high-dimensional feature embedding output by the conditional processing module and a latent vector provided by the feature fusion module. The high-dimensional feature embedding and the latent vector are concatenated and then passed through a shared feature extraction layer to obtain a shared feature vector, so as to extract key features in different prediction tasks of the stress value prediction network. The prediction tasks of the stress value prediction network include stress value regression task, uncertainty estimation task, stress distribution prediction task, and confidence assessment task.

[0011] According to one embodiment of the present invention, a modeling module for constructing a boom stress prediction model is further included. Specifically, the modeling module is used to: collect historical load data, geometric structure, and historical working condition information of the truck crane boom; establish a mechanical model of the truck crane boom; and obtain historical stress cloud maps based on finite element analysis software to form a sample set; construct a stress cloud map prediction network for the stress cloud map prediction branch based on a conditional generation mechanism; construct a stress value prediction network for the stress value prediction branch using a multi-task learning network based on an attention mechanism; and perform collaborative training on the stress cloud map prediction network and the stress value prediction network based on the sample set and according to the joint loss function of multi-task learning to obtain the boom stress prediction model of the truck crane.

[0012] According to one embodiment of the present invention, a data augmentation module is further included, which is specifically used for: performing image gradient enhancement and random affine transformation on the historical stress cloud map to achieve image enhancement of the historical stress cloud map; sampling random noise on the historical load data and superimposing the random noise on the normalized historical load data to achieve conditional enhancement of the historical load data; pairing the image-enhanced and conditionally enhanced historical stress cloud map and historical load data to obtain an enhanced sample set, and the modeling module using the enhanced sample set to perform collaborative training on the stress cloud map prediction network and the stress value prediction network.

[0013] According to an embodiment of the present invention, the feature vector of the structural stress analysis includes a conditional feature vector and a potential vector. The acquisition module is specifically used to: acquire load data and working condition information of the truck crane boom; and obtain the conditional feature vector and the potential vector based on the load data and working condition information, using the condition processing module and the feature fusion module.

[0014] According to an embodiment of the present invention, the prediction module is specifically used to: input the conditional feature vector and the latent vector in parallel into the stress cloud map prediction network and the stress value prediction network to obtain the prediction result of the single inference output of the boom stress prediction model. The prediction result includes the boom stress cloud map tensor, normalized multi-region stress value, stress distribution probability, prediction uncertainty quantification index, and confidence score; introduce a hierarchical judgment mechanism to perform quality assessment and calibration on the prediction result; and process the prediction result obtained through the hierarchical judgment mechanism. The result processing includes: obtaining the boom stress cloud map based on the stress cloud map tensor, and obtaining the global maximum stress value and multiple region maximum stress values ​​based on the normalized multi-region stress value.

[0015] According to one embodiment of the present invention, the mapping and annotation module is specifically used to: convert the boom stress cloud map into a grayscale stress distribution map and extract stress intensity information; based on extreme value positioning, identify the maximum stress pixel points and their coordinates corresponding to the maximum stress values ​​of multiple regions in the boom stress cloud map according to the stress intensity information; map the maximum stress values ​​of multiple regions to the maximum stress pixel points of multiple regions in the boom stress cloud map and annotate them.

[0016] The beneficial effects of this invention are:

[0017] The crane of this invention, equipped with boom structure stress analysis capabilities, predicts boom stress cloud maps, global maximum stress values, and maximum stress values ​​in multiple regions based on a boom stress prediction model using an augmented detail bi-branch network. It then maps the maximum stress values ​​in multiple regions onto the boom stress cloud map and annotates them for boom structure stress capacity analysis. This avoids excessively large errors in the labeled data caused by directly reading values ​​from the stress cloud map. While visualizing the overall stress distribution, it significantly improves the prediction accuracy of stress concentration areas, providing a reliable risk assessment basis for engineering decisions. The acquisition module, prediction module, mapping and annotation module, and analysis module are easily encapsulated in the crane operator's cab platform through standardized interfaces. Furthermore, the boom stress prediction model based on an augmented detail bi-branch network obtains prediction results in a short time, exhibiting strong timeliness and effectively meeting the response requirements of practical application scenarios. Attached Figure Description

[0018] Figure 1 This is a block diagram of a crane with boom structure stress analysis capability according to an embodiment of the present invention;

[0019] Figure 2 This is a mechanical model diagram of a truck crane boom according to a specific embodiment of the present invention;

[0020] Figure 3 This is a schematic diagram of the technical principle of a stress cloud map prediction network according to an embodiment of the present invention;

[0021] Figure 4 This is a schematic diagram of the technical principle of a stress point prediction network according to an embodiment of the present invention;

[0022] Figure 5 This is a schematic diagram illustrating the working principle of the prediction module in the quality assessment and calibration stage according to an embodiment of the present invention.

[0023] Figure 6 The image shows the predicted results obtained from the crane and the fully annotated stress cloud diagram of the truck crane boom in a specific embodiment of the present invention. Detailed Implementation

[0024] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0025] like Figure 1 As shown, the crane with boom structure stress analysis capability according to an embodiment of the present invention includes: a data acquisition module 10, a prediction module 20, a mapping and annotation module 30, an analysis module 40, and an optional modeling module. The data acquisition module 10 is used to acquire load data and operating condition information of the truck crane boom to obtain feature vectors for structural stress analysis. The prediction module 20 predicts based on the feature vectors using a boom stress prediction model to obtain a boom stress cloud map, a global maximum stress value, and maximum stress values ​​in multiple regions. The mapping and annotation module 30 maps the maximum stress values ​​in multiple regions to the boom stress cloud map based on extreme value localization and annotates it to obtain a fully annotated truck crane boom stress cloud map. The analysis module 40 performs stress analysis on the boom structure based on the fully annotated truck crane boom stress cloud map and the global maximum stress value. The modeling module is used to construct a boom stress prediction model, which is a detail-enhanced dual-branch network model. The dual branches of the boom stress prediction model include a stress cloud map prediction branch and a stress value prediction branch.

[0026] Among them, the stress cloud map prediction network of the stress cloud map prediction branch can be constructed based on the conditional generation mechanism; the stress value prediction network of the stress value prediction branch can be constructed based on the multi-task learning network of the attention mechanism; when constructing the boom stress prediction model of the truck crane, the stress cloud map prediction network and the stress value prediction network can be trained together so that the boom stress prediction model can achieve balanced control under different task objectives, thereby coordinating and optimizing the performance of each task.

[0027] It is understood that the boom stress prediction model used by the crane with boom structure stress analysis capability in the embodiments of the present invention is constructed based on a detail-enhanced dual-branch network model. This network model adopts a modular neural network architecture, can be built based on the PyTouch architecture and implement standardized interface encapsulation, has excellent platform compatibility and deployment flexibility, and can run efficiently on various hardware environments such as embedded devices, industrial control computers and mobile computing platforms.

[0028] Therefore, the crane's acquisition module 10, prediction module 20, mapping and annotation module 30, and analysis module 40 can be seamlessly integrated into the crane operator's cab platform through standardized packaging. This allows for real-time output of the boom stress cloud map, the global maximum stress value, and the maximum stress values ​​for each region, automatically mapping and annotating the maximum stress values ​​for each region to the corresponding positions in the stress cloud map. This avoids the annotation errors caused by manual reading of the cloud map in traditional methods, and while maintaining a clear and intuitive view of the overall stress distribution, significantly improves the accuracy of identifying stress concentration areas. This provides crane operators with highly reliable real-time structural safety monitoring and risk assessment capabilities, effectively supporting engineering decision-making and operational safety under complex working conditions.

[0029] In one embodiment of the present invention, the modeling module can be set outside the crane to pre-build a boom stress prediction model for the boom structure of the truck crane, or it can be embedded in the crane operator's cab operating platform along with the acquisition module 10, prediction module 20, mapping and annotation module 30 and analysis module 40 to build a boom stress prediction model according to actual needs.

[0030] Specifically, the modeling module can be used to: collect historical load data, geometric structure, and historical working condition information of the truck crane boom; considering that deep learning model training depends on high-quality, representative datasets, first establish a mechanical model of the truck crane boom, and obtain historical stress cloud maps based on finite element analysis software to form a sample set; construct a stress cloud map prediction network for the stress cloud map prediction branch based on a conditional generation mechanism; construct a stress value prediction network for the stress value prediction branch using a multi-task learning network based on an attention mechanism; and, based on the sample set and according to the joint loss function of multi-task learning, collaboratively train the stress cloud map prediction network and the stress value prediction network to obtain the truck crane boom stress prediction model.

[0031] In one embodiment of the present invention, the mechanical model of the truck crane boom is based on the actual service conditions, geometry, and load characteristics of the crane, and is directly modeled parametrically using ANSYS APDL (a finite element analysis software language). During the modeling process, the key load-bearing parts of the truck crane (such as the boom root) are structurally refined and analyzed in detail, while non-critical features that have little impact on the overall mechanical performance are ignored, thereby improving simulation efficiency while ensuring accuracy.

[0032] by Figure 2 In one specific embodiment, the crane boom body is a multi-section U-shaped telescopic box structure, and its working conditions are divided into several typical modes based on the combination of boom length and amplitude. To accurately assess its structural safety, the analysis focuses on several hazardous working conditions. Slider contacts are set between the telescopic boom sections, and fixed constraints are applied at the boom tail root and hydraulic cylinder support locations; the material parameters are set as Q690D high-strength steel (elastic modulus E=206GPa, Poisson's ratio ν=0.25, density ρ=7850kg / m³). The load data meshing uses a combination of shell and solid elements. Thin-walled components such as the boom use SHELL181 shell elements to accurately capture their mechanical response; solid structures such as the sliders use SOLID185 elements. Mesh quality is strictly controlled during the meshing process, prioritizing the use of mapping and sweeping methods to generate regular hexahedral meshes. Figure 2 In the middle, the left figure represents the mechanical model of the crane boom before the luffing plane decomposition in ANSYS APDL parametric modeling, and the right figure represents the mechanical model of the crane boom after the luffing plane decomposition in ANSYS APDL parametric modeling. Q P represents the rated lifting load. Gb P represents the weight of the boom. y P represents Q Vertical boom axis force decomposition, N represents P Q Force decomposition along the boom axis, F S M represents the rated lifting rope tension. Lx P represents Q and F S The resulting eccentric bending moment, e1 represents the eccentricity between the fixed pulley at the top of the boom and the boom axis, e2 represents the eccentricity between the guide pulley of the boom and the boom axis, and F1 represents the tension F of the hoisting rope. s The component of force generated along the boom axis, F2 represents the tension in the hoisting rope, F s The component of force in the direction of the vertical boom axis, θ represents the amplitude angle.

[0033] Based on this parametric mechanical model, n sets of historical stress cloud maps and corresponding historical stress data under different working conditions are generated in batches using APDL command streams. The historical stress data can include the maximum stress value of the entire machine and the maximum stress value of key load-bearing components, thus forming a sample set. Historical load data and historical working condition information can be stored in text format, while historical stress data is saved as image files. A strict one-to-one correspondence is maintained between working conditions, loads, and cloud maps to ensure data consistency and traceability.

[0034] To address the issues of feature distribution shift and gradient oscillation caused by inconsistent spatial scales and load dimensions in deep learning model training, data in the sample set can be standardized. Specifically, this can include dataset consistency verification, load data normalization, and one-hot encoding of operating condition information.

[0035] In the data consistency verification steps of the dataset: a multi-condition stress dataset class is constructed for the systematic organization and standardization of the dataset; during the data loading phase, this dataset class first iterates through and verifies each data source file; if historical load data, historical stress data, and historical stress clouds are detected... Figure 3 If the number of samples from the three classes does not match, the "minimum length" strategy is used to prune the data to ensure that the number of samples from the three classes is consistent, thus laying the foundation for stable training of the model in the future.

[0036] In the load data normalization process: all historical load data in the dataset are extracted to construct a load data list. Using the maximum and minimum values ​​of the historical load data in this list as a benchmark, linear normalization is performed on all historical load data and the load magnitudes of the corresponding working conditions, uniformly mapping historical load data of different magnitude ranges to the [0,1] interval. This eliminates the interference of load dimension differences on model training and avoids bias in model parameter updates caused by data magnitude imbalance.

[0037] In the one-hot encoding process for operating condition information: for discrete operating condition information, in order to realize its effective input in the model and eliminate the spurious order relationship that may be introduced by traditional sequence number encoding, one-hot encoding is used for numerical conversion.

[0038] In one specific embodiment of the present invention, each working condition category can be mapped to a binary vector of length equal to the total number of working condition categories. Only the positions corresponding to the working condition category in the vector are set to 1, while the remaining positions are set to 0. Specific encoding information is shown in Table 1 below. This process can transform working condition information into unbiased features that the model can directly learn, improving the learning accuracy of the correlation between working conditions and stress distribution.

[0039] Table 1 Operating Condition Information Coding

[0040]

[0041] Stress contour map standardization and preprocessing: A fixed-region cropping strategy is adopted for historical stress contour maps to eliminate the influence of inconsistent image spatial scales and redundant backgrounds. Specifically, array slicing operations using the OpenCV library are used to accurately extract the Region of Interest (ROI) containing the main structural features and key stress distribution information from the stress contour map, while redundant background regions unrelated to stress distribution are removed. This processing not only achieves complete uniformity of the input size of the contour map, but also effectively reduces the computational burden on the model by removing redundant information, while ensuring the consistency of the input region positions across different samples. This provides standardized input for the model to learn common stress distribution patterns across samples, ultimately improving the model's feature learning ability and generalization performance.

[0042] In one embodiment of the present invention, such as Figure 3 As shown, the stress cloud map prediction network includes a condition processing module, a feature fusion module, and a detail-preserving upsampling decoder.

[0043] The condition processing module is the core unit in the boom stress prediction model, specifically designed to handle input condition information. Through a collaborative design of multi-layer fully connected networks and activation functions, it provides high-dimensional feature embeddings of load conditions to the feature fusion module and nonlinear mappings of conditional feature vectors to the dropout layer of the detail-preserving upsampling decoder. This design effectively utilizes load and other conditional information to guide the stress contour map generation process, accurately capturing the complex correlations in stress distribution under different loads. Ultimately, it improves the fitting ability and nonlinear problem handling capabilities of the stress contour map prediction network, ensuring that the predicted stress contour map more closely reflects actual engineering scenarios. Compared to the simple embedding method used in CGAN (Conditional Generative Adversarial Network) models, the newly added intermediate layer in this module significantly enhances the expressive power of conditional features, especially in small-sample scenarios, making it more effective at capturing the nonlinear relationship between load and stress distribution.

[0044] In one specific embodiment of the present invention, the idea of ​​encoding conditional information in the Conditional Generative Adversarial Network (CGAN) model can be used as a reference to first process the payload data. L i Normalize the operating data c i One-hot encoding is performed to normalize the 1D load data. P L 'and the 4D working condition feature vector after one-hot encoding C i The mapped feature vector is obtained by passing it through a linear layer and then applying a non-linear activation function. LeakyReLU ( x(See the following formula) Different transformations are performed to introduce nonlinear features to enhance the stability of model training; by adding a Dropout layer, the output of some neurons is randomly ignored to alleviate the overfitting problem of the model in small sample scenarios; the feature vector is then subjected to the above steps again to ensure that the feature vector dimension is expanded to a 1024-dimensional conditional feature vector C, and gradually transformed into feature patterns corresponding to different stress concentration regions.

[0045] (1)

[0046] In the formula, x For eigenvalues; α This is the secondary slope.

[0047] The feature fusion module generates a high-dimensional feature map based on the high-dimensional feature embedding provided by the condition processing module.

[0048] In the original CGAN, the conditional information feature representation of the feature fusion module is weak and easily dominated by noise. In a specific embodiment of this invention, to enhance the conditional constraints and reasonable diversification capabilities of the feature fusion module, the concept of latent space modeling from the Variational Autoencoder (VAE) is introduced. A normal distribution is sampled from the constructed fixed-dimensional 256-dimensional latent space to obtain a latent vector Z of the same dimension. i The feature fusion module obtains the deeply processed conditional feature vector C and latent vector Z. i Then, the vectors are concatenated to obtain a 1280-dimensional vector, ensuring that the module is driven by conditional information and precisely controls the diversity of the generated results. After further processing by a multi-layer fully connected network and activation functions, the concatenated high-dimensional vector is mapped to a (1024×4×4) high-dimensional space, realizing the integration of different spatial features. This module balances the diversity and accuracy of the generated stress cloud maps, improving the model's practicality and generalization ability.

[0049] The detail-preserving upsampling decoder includes multiple cascaded detail-preserving upsampling blocks. The detail-preserving upsampling blocks based on high-dimensional feature embedding achieve progressive reconstruction of the boom stress cloud map based on the high-dimensional feature map.

[0050] In one embodiment of the present invention, the detail-preserving upsampling block includes a bilinear upsampling layer and a dual-path convolutional layer. The dual-path convolutional layer includes a main path convolutional layer and a detail path convolutional layer in parallel, aiming to simultaneously maintain the macroscopic structural integrity of the stress distribution and the clarity of microscopic details in stress concentration areas. After weighted fusion of the feature maps extracted by the main path convolutional layer and the detail path convolutional layer, a residual connection can be made with the feature map output by the bilinear upsampling layer of the detail-preserving upsampling block.

[0051] Meanwhile, to further optimize gradient flow and enhance feature reuse, a residual refinement mechanism is introduced during the reconstruction of the boom stress cloud map by the detail-preserving upsampling decoder. The initial boom stress cloud map generated by the independent residual network is weighted and superimposed onto the boom stress cloud map to improve the representation accuracy of stress concentration areas.

[0052] The specific principle of the detail-preserving upsampling decoder is as follows:

[0053] (1) Input the high-dimensional feature map provided by the feature fusion module into a basic convolutional layer for preliminary feature adjustment. This layer can be processed using a 3×3 convolutional kernel, which aims to enhance feature representation and prepare for subsequent upsampling operations.

[0054] (2) The adjusted high-dimensional feature map is input into the DPUB module (Detail-Preserving Upsampling Block, DPUB), and its resolution is doubled by bilinear interpolation (see Equation (2)). The result is sent to the main path and the detail path at the same time. The main path ensures the coherence of the overall stress field structure through two 3×3 convolutional layers. The detail path extracts high-frequency details with the same structure (see Equation (3)). The feature maps output by the main path and the detail path are fused according to the experimentally optimized weights (see Equation (4)).

[0055] (2)

[0056] In the formula, Input a low-resolution feature map; This is the upsampled high-resolution feature map; The weights are for bilinear interpolation, and these values ​​are determined by the target pixel. With source pixel The spatial distance determines this.

[0057] (3)

[0058] In the formula, Feature maps output by the main path or detail path; The convolution kernel has a size of 3×3.

[0059] (4)

[0060] In the formula, Y o The output feature map after fusion. Y m Feature map extracted from the main path. Y d Feature maps extracted for detail paths, These are the weighting coefficients.

[0061] In this process, to alleviate the gradient vanishing problem in deep network training, a residual connection method is introduced. That is, after each DPUB module completes bilinear upsampling, main path and detail path feature extraction and weighted fusion operations, the current output feature map is residually connected with the feature map upsampled by that module, resulting in equation (5). The feature map is used as input for the next module and iterated n times to gradually increase its resolution.

[0062] (5)

[0063] In the formula, This is the output feature map of the DPUB module.

[0064] (3) High-resolution, high-dimensional features are fused using an output network consisting of multiple convolutional layers. These convolutional layers gradually reduce the number of feature channels from 16 to a final 3 channels, generating a preliminary RGB stress cloud map. F r Meanwhile, to improve the characterization accuracy of stress concentration regions, a residual refinement mechanism is introduced. The initially generated contour map is input into a lightweight residual network consisting of 16-channel 3×3 convolutions and 1×1 convolutions. Its output is then superimposed onto the original image with small weights to obtain a refined contour map. :

[0065] (6)

[0066] In the formula, Insignificant weight values; This is the output of the residual network.

[0067] (4) The refined cloud map The nonlinear transformation is performed using the Tanh activation function (see Equation (7)), mapping each pixel value to the standard range of [-1, 1]. The mapped pixel values ​​are then fine-tuned, and the values ​​are strictly limited to the range of [0.0, 1.0] through a cropping operation, generating a high-quality boom stress cloud map that can be directly used for visualization and analysis.

[0068] (7)

[0069] In the formula, For a specific pixel location in the final feature map Pixel values; It is a natural exponential function; This is the final pixel value at that pixel location after mapping.

[0070] In summary, the entire decoder architecture gradually restores the spatial details and resolution of the predicted image through cascaded detail-preserving upsampling blocks (DPUBs), aiming to generate a stress distribution cloud map with high visual fidelity and conforming to physical logic. The output is then fine-tuned through residual refinement layers, achieving a visualization result of stress distribution conforming to physical laws under specific scenarios, from input conditions of load and working conditions.

[0071] In one embodiment of the present invention, such as Figure 4 As shown, the input to the stress value prediction network includes the high-dimensional feature embedding output from the conditional processing module. f 512 The latent vectors provided by the feature fusion module z i High-dimensional feature embedding f 512 and latent vectors z i The data is concatenated and then passed through a shared feature extraction layer to obtain a shared feature vector of dimension 1024. This vector is used for key feature extraction in different prediction tasks of the stress value prediction network. The prediction tasks of the stress value prediction network include stress value regression, uncertainty estimation, stress distribution prediction, and confidence assessment. Specifically, the stress value regression task predicts the maximum stress value in multiple regions and the global maximum stress value; the uncertainty estimation task quantifies the error fluctuation range of the prediction results; the stress distribution prediction task expands the single stress point value to a spatial distribution representation; and the confidence assessment task quantifies the reliability of the model's prediction results. (Figure shows high-dimensional feature embedding.) f 512 The first 512 dimensions of the conditional feature vector C provided by the conditional processing module. f 768 Represents high-dimensional feature embedding f 512 and latent vectors z i The concatenated feature vectors a 1~ a 1024 Represents the first layer of neurons. b 1~ b 1024 This represents the second layer of neurons.

[0072] In one specific embodiment of the present invention, the workflow of the stress value prediction network is as follows:

[0073] The multi-task learning network receives high-dimensional feature embeddings from the output of the conditional processing module. f 512 , and the latent vector of the feature fusion module z iThe features are concatenated and then passed through a shared feature extraction layer to obtain a shared feature vector with a dimension of 1024. f This facilitates the extraction of key features for different subsequent prediction tasks.

[0074] For shared feature vectors f Dimensional adjustments are made to align the number of channels, which are then input into a multi-head self-attention mechanism for computation. The formula for calculating the scaled dot product attention independently for each head is as follows:

[0075] (8)

[0076] In the formula, , n The number of heads representing attention. head i Representing the i Each attention head calculation branch, softmax This represents a mathematical function that transforms attention values ​​into a probability distribution. For the first i The query, key, and value projection matrix of the head. T Represents transpose. This is a dimension value, which can be 128.

[0077] The outputs of all attention heads are concatenated along the feature dimension and then linearly transformed to output the attention-weighted feature vector. It restores its dimensions to their original form and shares feature vectors with the original. f After performing residual joins, the data enters the LayerNorm layer for normalization.

[0078] (9)

[0079] In the formula, C H This represents the multi-head attention characteristics after splicing. concat This represents a splicing operation. This is the corresponding learnable weight matrix; The feature vector after attention weighting; and These are the mean and standard deviation of the eigenvectors after residual concatenation, respectively. It is the numerical stability constant; and For learnable affine transformation parameters; The optimized feature vector, Residual This represents a residual connection.

[0080] To avoid multi-task conflicts, the optimized feature vectors are used. The neural network substructures, specifically designed for different tasks, are used to perform multiple prediction tasks in parallel. The stress value regression task is implemented using a multilayer perceptron (MLP), which consists of multiple fully connected layers and nonlinear activation functions. This network progressively compresses the feature dimensions, ultimately outputting a normalized stress value. The calculation process is as follows:

[0081] (10)

[0082] In the formula, Sigmoid For activation functions; LeakyReLU For activation functions; The weight matrix is ​​a learnable weight matrix; f ' is the optimized feature vector; These are normalized stress prediction values.

[0083] Uncertainty estimation aims to quantify the reliability of prediction results and provide a basis for risk assessment in engineering decisions. This task can employ heteroscedastic uncertainty modeling methods to directly learn the prediction variance from unknown data.

[0084] (11)

[0085] In the formula, To address the heteroscedasticity uncertainty of the prediction results under the current input conditions, Softplus The formula is shown in equation (12) below.

[0086] (12)

[0087] In the formula, x represent Softplus The input variables of the formula.

[0088] The stress distribution prediction task transforms the continuous stress value regression problem into a discrete probability distribution estimation, providing a richer representation of stress information. This task employs a two-layer fully connected network probability output layer and regularization techniques to discretize the stress value domain into multiple intervals, outputting the probability for each interval:

[0089] (13)

[0090] In the formula, z These are the feature vectors of the intermediate hidden layer; Softmax The activation function is shown in equation (14).

[0091] (14)

[0092] The confidence assessment task measures the degree of confidence in the current stress prediction results from the model's own perspective, complementing uncertainty estimation and evaluating the reliability of the prediction results from different angles. This task comprehensively considers feature quality and the model's internal state, outputting an overall prediction confidence score. The input is transformed into 256 dimensions through a linear layer, then processed by the LeakyReLU activation function and compressed to 1 dimension by another linear layer, and finally output through the Sigmoid function. Confidence level within c .

[0093] In summary, by employing a multi-task learning network based on an attention mechanism in the stress value prediction network, four types of output results will be simultaneously generated, namely the normalized global maximum stress value. or maximum stress value in the region Uncertainty measure Stress distribution probability p and confidence level u .

[0094] In one embodiment of the present invention, to achieve coordinated optimization of the two branches of stress cloud map generation and stress value prediction, the present invention also designs a joint loss function based on multi-task learning. This function needs to consider both image reconstruction quality and numerical prediction accuracy simultaneously, and fuses the losses of each task through a weighted combination mechanism, enabling the model to achieve balanced adjustment among different task objectives, thereby synergistically optimizing the performance of each task. As shown in Equation 15 below, the joint loss function... L t Using weighted summation as the overall framework, and by organically integrating loss terms of different properties, we can provide effective guidance for the collaborative optimization of the model.

[0095] (15)

[0096] In the formula, L i Represents the image reconstruction loss function. L s The loss function representing stress value prediction. L un Represents the loss function for uncertainty estimation. L dist The loss function represents the stress distribution prediction. L conf The loss function represents the confidence level assessment. Predict branch weight coefficients for stress cloud diagrams; Estimating weighting coefficients for uncertainty; Weighting coefficients for stress distribution prediction; The confidence level is assessed using weighting coefficients.

[0097] The process of constructing the image reconstruction loss function for the stress cloud prediction branch is shown in equations (16) to (17) below:

[0098] Image reconstruction loss aims to ensure the visual fidelity and physical plausibility of the generated stress cloud map. As one of the important outputs of the model, its loss function is designed to measure the difference between the predicted cloud map and the real cloud map from multiple dimensions, so as to ensure that the generated stress cloud map is as close as possible to the real situation in terms of visual effect, structural features and detailed information.

[0099] In one embodiment of the present invention, pixel-level L1, SSIM (structural similarity index), and edge are selected as loss functions, as shown in formula (16). Among them, pixel-level L1 loss serves as the basic reconstruction loss, which can intuitively measure the pixel-level differences of the image. It calculates the absolute difference between the corresponding pixels of the predicted cloud map and the real cloud map and then averages the results to obtain the overall pixel-level difference. SSIM is a structural similarity loss designed based on the characteristics of the human visual system. By comparing the brightness, contrast, and structural information of the image, it better meets the perceptual needs of engineers for stress cloud maps. The sliding window operation involved in the calculation process can capture the local structural features of the image. Edge is an edge-preserving loss. For stress concentration areas, the gradient of the predicted cloud map and the real cloud map is calculated using the Sobel operator. The difference between the two gradients is compared, which forces the model not only to learn the accurate distribution of stress but also to learn the stress variation law, ensuring the physical rationality of the generated cloud map in terms of mechanical response.

[0100] (16)

[0101] In the formula, L L1 Represents pixel-level L1 loss, L SSIM Represents structural similarity loss. L edge To preserve loss at the edge, N Batch size; H Image height; W Image width; C This refers to the number of image channels; To predict the image in ( H , W , C The pixel value of ) The pixel values ​​of the real image at (h, w, c); To predict the gradient field of the contour map; The gradient field of the real cloud map is shown in equation (17).

[0102] (17)

[0103] In the formula, These represent the horizontal and vertical convolution kernels of the Sobel operator, respectively.

[0104] The above three loss functions are linearly combined according to certain weight ratios to obtain the image reconstruction loss function of the stress cloud map prediction branch:

[0105] (18)

[0106] In the formula, , The weighting coefficients can be set based on experiments or experience.

[0107] The process of constructing the stress value prediction branch loss function is shown in equations (19) to (22) below:

[0108] The core task of stress value prediction is to accurately predict the maximum stress value in multiple regions (such as critical load-bearing structures) and globally. To this end, mean squared error (MSE) can be used as a loss function to measure the deviation between the predicted and actual values. The stress value prediction loss function... L s The formula is:

[0109] (19)

[0110] In the formula, for i Predicted maximum stress values ​​for key regions; for i The true maximum stress value of each key area; and These are the globally predicted maximum stress value and the globally actual maximum stress value, respectively.

[0111] To quantify the cognitive uncertainty of stress value prediction branch prediction, an uncertainty regularization term is introduced into the uncertainty estimation loss function. This regularization term is defined as the mean of the uncertainty values ​​in the model output, and its mathematical expression is the prediction uncertainty. β The expectation is that by minimizing this term, the model is guided to learn to provide a moderate and reasonable estimate of uncertainty without sacrificing prediction accuracy; the uncertainty estimation loss function is... L un The specific formula is as follows:

[0112] (20)

[0113] In the formula, This is an uncertain value.

[0114] The stress distribution prediction task transforms the problem of regressing continuous stress values ​​into estimation of discrete probability distributions. This task divides the stress range into... K Given mutually exclusive intervals, the model is required to output the probability that the stress value falls within each interval. To measure the information difference between the predicted and true distributions, KL (Kullback-Leibler divergence) is used as the loss function. By minimizing the KL divergence, the model is forced to learn the stress probability patterns hidden behind the data, with the goal of making the predicted stress probability distribution... Infinitely approximating the target distribution of the construction based on real stress Thus, the stress distribution prediction loss function is obtained. L dist The formula is as follows:

[0115] (twenty one)

[0116] In the formula, To predict the stress probability distribution, To construct the target distribution, For the target distribution at the th j The probability values ​​for each interval. To predict the distribution at the th j The probability values ​​for each interval.

[0117] The confidence assessment task aims to quantify the model's degree of confidence in point predictions. To achieve this, the task is optimized using a specialized supervised loss, the core of which lies in utilizing a target confidence level constructed based on normalized prediction error as a supervisory signal. This loss function guides the model to learn a fundamental rule: assigning high confidence to accurately predicted samples and low confidence to inaccurately predicted samples. By minimizing this loss, the model's confidence output is effectively calibrated, thus becoming an intrinsic evaluator that reliably reflects prediction quality. Confidence Assessment Loss Function L conf The specific formula is as follows:

[0118] (twenty two)

[0119] In the formula, The normalized stress value predicted by the model; For true normalized stress labels; To predict confidence levels; The target confidence level.

[0120] In few-shot learning scenarios, the quality and diversity of samples are among the core factors determining the final performance of the model. In one embodiment of the present invention, the crane further includes a data augmentation module, which is specifically used for: performing image gradient enhancement and random affine transformation on the historical stress cloud map to achieve image enhancement of the historical stress cloud map; sampling random noise from the historical load data and superimposing the random noise onto the normalized historical load data to achieve conditional enhancement of the historical load data; pairing the image-enhanced and conditionally enhanced historical stress cloud map with the historical load data to obtain an enhanced sample set, and using the enhanced sample set to co-train the stress cloud map prediction network and the stress value prediction network.

[0121] Specifically, when enhancing historical stress cloud maps, edge detail sharpening can be performed on each original stress cloud map using unsharp masking filtering to enhance gradient information in stress concentration areas. When performing random affine transformations, a random affine transformation for the edge detail sharpened image can be defined using PyTorch's RandomAffine class, including a rotation angle of ±0.3° and a horizontal and vertical translation of ±0.3%. Then, the image after each transformation enhancement is iterated multiple times.

[0122] When performing conditional enhancement on historical load data, a noise value within a specific range can be randomly sampled from a small normal distribution N with a mean of zero. This value is then added to the normalized historical load data from the data processing to obtain the enhanced load sample. For each load sample, multiple enhancement iterations are performed according to the specified enhancement factor. The number of iterations of the load data should be consistent with the number of iterations of the stress contour plot.

[0123] In this embodiment, in response to the data characteristics of crane stress cloud maps (such as the continuity of stress distribution, the saliency of key areas, and the mapping correlation between grayscale / color and stress values) and the challenges faced by small sample learning, such as sample scarcity and model overfitting, a hybrid enhancement strategy that integrates image enhancement and conditional enhancement is proposed. Through the collaborative design of multi-dimensional enhancement operations, the diversity and quality of the sample set are systematically improved, providing better data support for model training in the modeling module.

[0124] In one embodiment of the present invention, the feature vector of structural stress analysis includes a conditional feature vector and a potential vector. The acquisition module 10 is specifically used to: acquire load data and working condition information of the boom of the truck crane; and obtain the conditional feature vector and the potential vector based on the load data and working condition information, using the condition processing module and the feature fusion module.

[0125] Specifically, the global load extrema (maximum and minimum global load values) statistically obtained during the construction of the training boom stress prediction model can be used to map the collected load data to the standard interval [0,1]. The working condition information can be converted into a four-dimensional one-hot encoded vector by querying a predefined encoding mapping table. Simultaneously, a 256-dimensional latent vector is generated by sampling from a standard normal distribution, serving as a random seed for stress cloud map generation. Based on this, conditional feature vectors and latent vectors are formed through feature extraction and feature fusion. The encoding mapping table can refer to Table 1 disclosed above, or can be adjusted according to actual needs. The conditional feature vectors and latent vectors and their generation process can be referred to the relevant descriptions of the condition processing module and feature fusion module of the stress cloud map prediction network above, and will not be repeated here.

[0126] In one embodiment of the present invention, the prediction module 20 is specifically used for: in the model forward propagation stage, inputting the conditional feature vector and the latent vector in parallel into the stress cloud map prediction network and the stress value prediction network to obtain the prediction result of the boom stress prediction model's single inference output. The prediction result includes the boom stress cloud map tensor, normalized multi-region stress value, stress distribution probability, prediction uncertainty quantification index, and confidence score; in the quality assessment and calibration stage, introducing a hierarchical judgment mechanism to assess and calibrate the prediction result; and processing the prediction result obtained through the hierarchical judgment mechanism. Specifically, the result processing may include: based on the stress cloud map tensor, performing inverse normalization processing to transform the value range from [-1,1] to the standard image range of [0,255], and obtaining a visualized boom stress cloud map through color mapping; and restoring the normalized multi-region stress value to the actual global maximum stress value and multiple regional maximum stress values ​​in MPa through linear inverse transformation.

[0127] Specifically, such as Figure 5 As shown, during the quality assessment and calibration phase, the distribution entropy value can be judged based on the stress distribution probability of the predicted results. For example, the effective range of the distribution entropy value can be set to 0.3-2.0 for judgment. If it fails, a hybrid resampling process is triggered, for example, using the latent vector z output by the feature fusion module at 50%. i The original vector and the 256-dimensional feature vector (new vector) after 50% resampling are mixed and resampled; otherwise, confidence judgment is performed based on the confidence score of the prediction result, for example, the confidence threshold is set to 0.97 for judgment. If it fails, the model returns to the forward propagation stage to perform full resampling; otherwise, uncertainty judgment is performed based on the prediction uncertainty quantification index of the prediction result, for example, the uncertainty threshold is set to 0.25 for judgment. If it fails, latent vector mixing sampling is triggered again; after all judgments pass, the prediction results that pass the hierarchical judgment mechanism are processed.

[0128] In one embodiment of the present invention, the mapping and annotation module 30 is specifically used to: convert the boom stress cloud map into a grayscale stress distribution map and extract stress intensity information; identify the maximum stress pixel points and their coordinates in multiple regions corresponding to the maximum stress values ​​in the boom stress cloud map based on extreme value positioning and stress intensity information; map the maximum stress values ​​in multiple regions to the maximum stress pixel points in multiple regions of the boom stress cloud map and annotate them.

[0129] Before converting the boom stress cloud map into a grayscale stress distribution map, the mapping and annotation module 30 can also preprocess the stress cloud map tensor output by the model using PyTorch tensor operations. By adjusting the dimensions and transmitting data between devices, the shape is converted from [1,3,512,150] to a NumPy array of [512,150,3]. Inverse normalization is then performed to linearly map the value range from [-1,1] to [0,1]. The inverse normalized boom stress cloud map is then converted into a grayscale stress distribution map, which is convenient for extracting stress intensity information using OpenCV's cvtColor function.

[0130] When identifying the maximum stress pixel points and their coordinates corresponding to the maximum stress values ​​in multiple regions of the boom stress cloud map, the minMaxLoc algorithm of the OpenCV library can be called to locate extreme values ​​based on stress intensity information for each predefined key load-bearing geometric region (such as the root of each boom section of the crane) to quickly identify the maximum stress pixel points and their coordinates corresponding to the maximum stress values ​​in multiple regions.

[0131] When performing visual annotation, solid dots can be used to mark extreme value locations to enhance visual salience. At the same time, text labels in the format "region X: YY.Y MPa" for the maximum stress pixels in each region, combined with background filling technology, can ensure text readability.

[0132] In one embodiment of the present invention, when the analysis module 40 performs stress analysis on the boom structure based on the fully annotated truck crane boom stress cloud map and the global maximum stress value, it can analyze the quality of the generated stress cloud map based on the PSNR of the truck crane boom stress cloud map, analyze the bearing capacity margin based on the structural bearing capacity limit value of the boom steel and the global maximum stress value, and give a safety level evaluation based on the size of the margin.

[0133] Specifically, the structural bearing capacity limit value of the boom steel can be obtained based on the yield strength of the steel. For example, the yield strength can be divided by a safety factor of 1.1 to obtain the structural bearing capacity limit value. Thus, the bearing capacity margin can be obtained as follows: bearing capacity margin = (structural bearing capacity limit value - predicted global maximum stress) / structural bearing capacity limit value × 100%. When the margin is ≥30%, the safety level is rated as "excellent"; when the margin is 20%-30%, the safety level is rated as "good"; when the margin is 10%-20%, the safety level is rated as "medium"; and when the margin is <10%, the safety level is rated as "poor".

[0134] In some other embodiments of the present invention, tensile strength or fatigue strength can also be used as a benchmark to calculate the structural bearing capacity limit value of the boom steel. In this case, it is only necessary to replace the yield strength and safety factor with the corresponding values. For example, tensile strength can be used as the calculation criterion, and the tensile strength can be divided by the tensile strength coefficient to obtain the structural bearing capacity limit value.

[0135] In a specific embodiment of the present invention, taking a company's truck crane modified into a crane with boom structure stress analysis capabilities as an example, based on its actual working condition information and geometric structural characteristics, an original sample dataset containing four typical working conditions was constructed using ANSYS APDL simulation, collecting a total of 240 valid samples. Each working condition had 60 samples. Through the data augmentation steps disclosed above (including geometric transformation and noise injection, etc.), the dataset was expanded to 1200 augmented samples, forming the final augmented sample set O.

[0136] The augmented sample set O was randomly divided into a training set D (840 sets), a validation set M (240 sets), and a test set T (120 sets) in a 7:2:1 ratio to ensure that the various working conditions and load levels were evenly distributed across the three data subsets. The training set D was used for model parameter learning, the validation set M for hyperparameter tuning and training process monitoring, and the test set T, as unseen data, was used for the final evaluation of the model's generalization ability. The key hyperparameter configurations for model training are shown in Table 2 below, including core parameters such as learning rate scheduling strategy, batch size, and multi-task loss weight allocation. These parameters were determined through grid search and Bayesian optimization to achieve the best performance balance under limited sample conditions.

[0137] Table 2 Model Training Parameter Settings

[0138]

[0139] Based on the hyperparameter configuration in Table 2, and through monitoring the training process over 200 rounds, the boom stress prediction model constructed in this embodiment exhibits excellent convergence characteristics: the total training loss decreases rapidly in the first 50 rounds, verifying that the total loss converges synchronously and the difference between the two is always controlled within 5%, indicating that the optimizer parameters are set reasonably and there is no obvious overfitting tendency; dynamic learning rate scheduling triggers two learning rate decays in the middle of training (approximately 80-120 rounds), helping the model break through local optima by reducing the update step size; finally, the model reaches a multi-task loss balance state on the validation set: the total loss is 0.0234, of which image loss is 0.0162, stress loss is 0.0072, uncertainty regularization is 0.0001, distribution loss is 0.0002, and confidence loss is 0.0003. Each loss term achieves stable convergence under weighted combination.

[0140] To verify the effectiveness of the boom stress prediction model, the actual output effect of the complete model is demonstrated through typical working condition cases. Figure 6 The complete model output of a truck crane boom under working condition 1 and 400,000 N load is shown, demonstrating the comprehensive performance of the method from multiple dimensions, including stress cloud diagram generation, multi-region stress analysis and engineering application output.

[0141] like Figure 6 As shown, the fully annotated stress cloud map of the truck crane boom provided by the mapping and annotation module 30 performs excellently in both visual quality and physical accuracy. The generated cloud map clearly shows the stress distribution pattern of the structure and successfully captures the stress concentration phenomenon in key areas such as the root of the first, second, and third boom sections. The color transition of the cloud map is natural, the edge details are intact, and it highly matches the actual physical laws. In terms of quantitative evaluation, the PSNR of the generated cloud map reaches 28.6 dB, and the SSIM is 0.924, indicating that the model maintains excellent performance in both pixel-level accuracy and structural similarity. Especially in areas where the stress gradient changes sharply, the model can still maintain clear boundary features, avoiding the over-smoothing phenomenon common in traditional methods. In terms of stress value prediction, the model accurately predicts the global maximum stress value of 930.0 MPa, and the maximum stress values ​​of each key load-bearing area of ​​the structure are 403.2 MPa, 356.6 MPa, and 311.3 MPa, respectively. The probability of the stress distribution in the 900-950 MPa range is 0.38, the prediction confidence is 0.97, and the uncertainty value is 0.12. The relative errors between the predicted stress values ​​and the actual finite element analysis results in each region are all controlled within 5%, meeting the engineering accuracy requirements and providing a reliable risk assessment basis for engineering decisions. Moreover, in the GPU environment, the boom stress prediction model built based on RTX 4060 Ti can achieve efficient real-time prediction, with the time taken for a single prediction controlled within 100 milliseconds. Compared with the long prediction time of the traditional finite element method, it has better timeliness and can effectively meet the response requirements of actual application scenarios.

[0142] In the description of this invention, the terms "first" and "second" are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of indicated technical features. Thus, a feature defined as "first" or "second" may explicitly or implicitly include one or more of that feature. "A plurality of" means two or more, unless otherwise explicitly specified.

[0143] In this invention, unless otherwise explicitly specified and limited, the terms "installation," "connection," "linking," and "fixing," etc., should be interpreted broadly. For example, they can refer to a fixed connection, a detachable connection, or an integral part; they can refer to a mechanical connection or an electrical connection; they can refer to a direct connection or an indirect connection through an intermediate medium; they can refer to the internal communication of two components or the interaction between two components. Those skilled in the art can understand the specific meaning of the above terms in this invention according to the specific circumstances.

[0144] In this invention, unless otherwise explicitly specified and limited, "above" or "below" the second feature can mean that the first feature is in direct contact with the second feature, or that the first feature is in indirect contact with the second feature through an intermediate medium. Furthermore, "above," "over," and "on top" of the second feature can mean that the first feature is directly above or diagonally above the second feature, or simply that the first feature is at a higher horizontal level than the second feature. "Below," "below," and "under" the second feature can mean that the first feature is directly below or diagonally below the second feature, or simply that the first feature is at a lower horizontal level than the second feature.

[0145] In the description of this specification, the references to terms such as "one embodiment," "some embodiments," "example," "specific example," or "some examples," etc., indicate that a specific feature, structure, material, or characteristic described in connection with that embodiment or example is included in at least one embodiment or example of the present invention. In this specification, the illustrative expressions of the above terms do not necessarily refer to the same embodiment or example. Moreover, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in one or more embodiments or examples. Furthermore, without contradiction, those skilled in the art can combine and integrate the different embodiments or examples described in this specification, as well as the features of different embodiments or examples.

[0146] Any process or method description in the flowchart or otherwise herein can be understood as representing a module, segment, or portion of code comprising one or more executable instructions for implementing a particular logical function or process, and the scope of the preferred embodiments of the invention includes additional implementations in which functions may be performed not in the order shown or discussed, including substantially simultaneously or in reverse order depending on the functions involved, as will be understood by those skilled in the art to which embodiments of the invention pertain.

[0147] The logic and / or steps represented in the flowchart or otherwise described herein, for example, can be considered as a ordered list of executable instructions for implementing logical functions, and can be embodied in any computer-readable medium for use by, or in conjunction with, an instruction execution system, apparatus, or device (such as a computer-based system, a processor-included system, or other system that can fetch and execute instructions from, an instruction execution system, apparatus, or device). For the purposes of this specification, "computer-readable medium" can be any means that can contain, store, communicate, propagate, or transmit programs for use by, or in conjunction with, an instruction execution system, apparatus, or device. More specific examples (a non-exhaustive list) of computer-readable media include: an electrical connection having one or more wires (electronic device), a portable computer disk drive (magnetic device), random access memory (RAM), read-only memory (ROM), erasable and editable read-only memory (EPROM or flash memory), fiber optic devices, and portable optical disc read-only memory (CDROM). Alternatively, the computer-readable medium may be paper or other suitable media on which the program can be printed, since the program can be obtained electronically, for example, by optically scanning the paper or other medium, followed by editing, interpreting, or otherwise processing as necessary, and then stored in a computer memory.

[0148] It should be understood that various parts of the present invention can be implemented in hardware, software, firmware, or a combination thereof. In the above embodiments, multiple steps or methods can be implemented in software or firmware stored in memory and executed by a suitable instruction execution system. For example, if implemented in hardware, as in another embodiment, it can be implemented using any one or a combination of the following techniques known in the art: discrete logic circuits having logic gates for implementing logical functions on data signals, application-specific integrated circuits (ASICs) having suitable combinational logic gates, programmable gate arrays (PGAs), field-programmable gate arrays (FPGAs), etc.

[0149] Those skilled in the art will understand that all or part of the steps of the methods in the above embodiments can be implemented by a program instructing related hardware. The program can be stored in a computer-readable storage medium, and when executed, the program includes one or a combination of the steps of the method embodiments.

[0150] Furthermore, the functional units in the various embodiments of the present invention can be integrated into a processing module, or each unit can exist physically separately, or two or more units can be integrated into a module. The integrated module can be implemented in hardware or as a software functional module. If the integrated module is implemented as a software functional module and sold or used as an independent product, it can also be stored in a computer-readable storage medium.

[0151] Although embodiments of the present invention have been shown and described above, it is understood that the above embodiments are exemplary and should not be construed as limiting the present invention. Those skilled in the art can make changes, modifications, substitutions and variations to the above embodiments within the scope of the present invention.

Claims

1. A crane with the capability of boom structure stress analysis, characterized in that, include: The data acquisition module is used to collect load data and working condition information of the boom of the truck crane to obtain the feature vector of structural stress analysis. The prediction module predicts based on the feature vector using the boom stress prediction model to obtain the boom stress cloud map, the global maximum stress value, and the maximum stress values ​​of multiple regions. The boom stress prediction model is a dual-branch network model based on detail enhancement, and the dual branches include a stress cloud map prediction branch and a stress value prediction branch. The mapping and annotation module maps the maximum stress values ​​of multiple regions to the stress cloud map of the boom based on extreme value positioning and annotates them to obtain a fully annotated stress cloud map of the truck crane boom. The analysis module performs stress analysis on the boom structure based on a fully annotated stress cloud map of the truck crane boom and the global maximum stress value. The stress cloud map prediction network of the stress cloud map prediction branch is constructed based on a conditional generation mechanism; the stress value prediction network of the stress value prediction branch is constructed based on a multi-task learning network with an attention mechanism; during the construction of the boom stress prediction model of the truck crane, the stress cloud map prediction network and the stress value prediction network are trained collaboratively. The stress cloud map prediction network includes a condition processing module, a feature fusion module, and a detail-preserving upsampling decoder. The condition processing module provides high-dimensional feature embeddings of load conditions to the feature fusion module and provides nonlinear mappings of conditional feature vectors to the Dropout layer of the detail-preserving upsampling decoder. The feature fusion module generates a high-dimensional feature map based on the high-dimensional feature embeddings provided by the condition processing module. The detail-preserving upsampling decoder includes multiple cascaded detail-preserving upsampling blocks, which reconstruct the boom stress cloud map based on the high-dimensional feature map using the detail-preserving upsampling blocks with the high-dimensional feature embeddings.

2. The crane with boom structure stress analysis capability according to claim 1, characterized in that, The detail-preserving upsampling block includes a bilinear upsampling layer and a dual-path convolutional layer. The dual-path convolutional layer includes a main path convolutional layer and a detail path convolutional layer in parallel. The feature maps extracted by the main path convolutional layer and the detail path convolutional layer are weighted and fused, and then residually connected with the feature map output by the bilinear upsampling layer of the detail-preserving upsampling block. During the reconstruction of the boom stress cloud map by the detail-preserving upsampling decoder, a residual refinement mechanism is introduced to weightedly superimpose the initial boom stress cloud map generated by the independent residual network onto the boom stress cloud map.

3. The crane with boom structure stress analysis capability according to claim 1, characterized in that, The input to the stress value prediction network includes the high-dimensional feature embedding output by the conditional processing module and the latent vector provided by the feature fusion module. The high-dimensional feature embedding and the latent vector are concatenated and then passed through a shared feature extraction layer to obtain a shared feature vector, so as to extract key features in different prediction tasks of the stress value prediction network. The prediction tasks of the stress value prediction network include stress value regression task, uncertainty estimation task, stress distribution prediction task, and confidence assessment task.

4. The crane with boom structure stress analysis capability according to claim 1, characterized in that, It also includes a modeling module for constructing a boom stress prediction model, the modeling module being specifically used for: Historical load data, geometric structure, and historical working condition information of the truck crane boom are collected to establish a mechanical model of the truck crane boom. Historical stress cloud diagrams are obtained based on finite element analysis software to form a sample set. The stress contour map prediction network of the stress contour map prediction branch is constructed based on the conditional generation mechanism; The stress value prediction network of the stress value prediction branch is constructed using a multi-task learning network based on an attention mechanism. Based on the sample set, the stress cloud map prediction network and the stress value prediction network are trained collaboratively according to the joint loss function of multi-task learning to obtain the boom stress prediction model of the truck crane.

5. The crane with boom structure stress analysis capability according to claim 4, characterized in that, It also includes a data enhancement module, which is specifically used for: Image enhancement of the historical stress cloud map is achieved by performing image gradient enhancement and random affine transformation on the historical stress cloud map. Random noise is sampled from the historical load data, and the random noise is superimposed on the normalized historical load data to enhance the conditions of the historical load data. By pairing historical stress contour maps and historical load data after image enhancement and conditional enhancement, an enhanced sample set is obtained. The modeling module uses the enhanced sample set to collaboratively train the stress cloud map prediction network and the stress value prediction network.

6. The crane with boom structure stress analysis capability according to claim 4 or 5, characterized in that, The feature vectors for structural stress analysis include conditional feature vectors and latent vectors. The acquisition module is specifically used for: Collect load data and operating condition information of the truck crane boom; Based on the load data and operating condition information, the condition feature vector and the potential vector are obtained using the condition processing module and the feature fusion module.

7. The crane with boom structure stress analysis capability according to claim 6, characterized in that, The prediction module is specifically used for: The conditional feature vector and the latent vector are input in parallel into the stress cloud map prediction network and the stress value prediction network to obtain the prediction result of the boom stress prediction model in a single inference. The prediction result includes the boom stress cloud map tensor, normalized multi-region stress value, stress distribution probability, prediction uncertainty quantification index and confidence score. A hierarchical judgment mechanism is introduced to assess and calibrate the prediction results. The prediction results obtained through the graded judgment mechanism are processed, including: obtaining the boom stress cloud map based on the stress cloud map tensor, and obtaining the global maximum stress value and the maximum stress values ​​of multiple regions based on the normalized multi-region stress values.

8. The crane with boom structure stress analysis capability according to claim 1, characterized in that, The mapping and annotation module is specifically used for: The stress cloud map of the boom is converted into a grayscale stress distribution map, and the stress intensity information is extracted; Based on extreme value localization, the maximum stress pixel point and its coordinates corresponding to the maximum stress value in multiple regions of the boom stress cloud map are identified according to the stress intensity information. The maximum stress values ​​of multiple regions are mapped to the maximum stress pixels of multiple regions in the stress cloud map of the boom, and then labeled.

Images