Abstract: In example embodiments, improved techniques are provided for classifying elements of an infrastructure model that represents linear infrastructure (e.g., roads). The techniques may extract a set of cross sections perpendicular to a centerline of the linear infrastructure from the infrastructure model, generate a graph representation of each cross section to produce a set of graphs having nodes that represent elements and edges that represent contextual relationships, provide the set of graphs to a trained graph neural network (GNN) model, and produce therefrom class predictions for the elements. The class predictions may include one or more predicted classes for each element with a respective confidence. A best predicted class for each element may be selected and assigned to the element, thereby creating a new version of the infrastructure model. For elements that extend through multiple cross sections, the selection may involve aggregating predicted classes originating from the different graphs.

Abstract: In example embodiments, a dataset labeling software process combines diversity selection and with one or more uncertainty selection techniques in a multi-stage process that produces a labeled training dataset for training an image-based ML model. In an initial stage, diversity selection alone may be used to seed a labeled training dataset. In a subsequent stage, diversity selection may be combined with a first procedure for computing uncertainty (e.g., uncertainty selection based on ensemble learning) to build the labeled training dataset until a first stopping condition is met. Optionally, in a still further stage, diversity selection may be combined with a second, different procedure for computing uncertainty (e.g., uncertainty selection based on Monte Carlo (MC) dropout) to further build the labeled training dataset until a second stopping condition is met.

Abstract: In an example embodiment, a structural health monitoring software application provides non-contact structural health monitoring using a video of a structure captured by a video camera. The application selects an area of interest and divides the area of interest into a grid of cells. One or more machine vision algorithms are selected from a set of multiple machine vision algorithms provided by the application, wherein the set of multiple machine vision algorithms includes at least one phase-based algorithm and at least one template matching algorithm. The application applies the one or more machine vision algorithms to the video of the structure to determining a displacement of each cell, detects defects or damage based on differences in the displacement of cells, and displays an indicator of the detected defects or damage in a user interface.

Abstract: In example embodiments, techniques are provided for customizing a ML model for a specific user absent user-coding. The techniques may provide a “black box” service to the user where the intricacies of ML model training are abstracted, and the user simply provides source data and makes high level selections. The techniques may be used with a variety of types of ML model architectures and ML Pipelines.

Abstract: In example embodiments, a semantic multimodal search function is provided in an engineering application for searching engineering documents. In an indexing phase, the application indexes a library of engineering documents to build an embedding database. For each of the engineering documents, a region detector extracts regions that each correspond to a different mode of technical or engineering data, and a set of ML models is used to generate embeddings that represent the semantic significance of technical or engineering data in each of the regions. In a query phase, the application receives search input and uses the region detector to extract regions that each correspond to a different mode of input. The application uses the ML models to generate embeddings that represent the semantic significance of input in each of the regions, and then performs a vector search between these embeddings and those in the embedding database.

Abstract: In various example embodiments, model calibration techniques are used to identify one or more external structural loads on a structure. Initially, material or geometry attributes of elements of a model are adjusted to minimize discrepancy between measured responses and modeled responses. Thereafter, the model is used to search for external structural loads that minimize discrepancy between the measured responses and the modeled responses. Discrepancy may be measured by an error function that looks to dynamic responses and/or static responses. A solution that minimizes discrepancy may be determined using a genetic algorithm that generates a set of proposed external structural loads, applies the set to the model to produce the modeled responses, computes an error function that measures the difference between the measured responses and the modeled responses, and evolves the solutions to minimize an error produced by the error function, with such operation proceeding until a stopping criteria is reached.

Abstract: In example embodiments, a visualization application uses geographically relevant style images as guidance to automatically generate realistic 2D renders of a 3D infrastructure model. The application generates a synthetic render of the 3D infrastructure model and retrieves a set of style images that correspond to a geographics position associated with the model. The synthetic 2D render, the set of style images and, optionally, one or more user-provided text guidance phrases and/or mask images are applied to a realistic 2D render generator of the application. The realistic 2D render generator performs image translation (guided by the optional text guidance phrases and/or mask images) to adjust visual appearance of the infrastructure in the synthetic 2D render based on the visual appearance of the set of style images and to generate realistic context based on what appears in the set of style images, thereby producing a realistic 2D render.

Abstract: In example embodiments, techniques are provided for classifying elements of infrastructure models using a convolutional graph neural network (GNN). Graph-structured data structures are generated from infrastructure models, in which nodes represent elements and edges represent contextual relationships among elements (e.g., based on proximity, functionality, parent-child relationships, etc.). During training, the GNN learns embeddings from the nodes and edges of the graph-structured data structures, the embeddings capturing contextual clues that distinguish between elements that may share similar geometry (e.g., cross section, volume, surface area, etc.), yet serve different purposes.

Abstract: In example embodiments, a component management service identifies related components using a combination of ML model geometry-based classification and text-based classification. The component management service accesses a plurality of components, wherein each component is associated with a geometry mesh and textual metadata. It classifies each component based on geometric similarity by providing the geometry mesh as input to one or more ML models. A geometry classification confidence score is produced for each classification. The component management service also classifies each component based on textual similarity by providing the textual metadata as input to the one or more ML models. A textual classification confidence score is produced for each classification. It calculates an overall confidence score for each classification by combining the geometry classification confidence score and the textual metadata classification confidence score.

Abstract: In example embodiments, a hybrid classification/anomaly machine learning architecture is provided that combines a classification model and an anomaly model to perform an engineering task. The classification model and the anomaly model may be used in parallel, and their inference results compared, with their consistency used to improve confidence, and their inconsistency used to detect when additional training or other improvement is required and to capture data useful in such additional training/improvement.

Abstract: In example embodiments, a proprietary implementation of tile-based and/or non-tile-based graphics processing is provided at run-time as a binary for use in an open-source infrastructure modeling platform. An open-source repository from which the infrastructure modeling platform is built includes a backend module and definitions of the interface of a visualization module from a private repository that implements the tile-based and/or non-tile-based processing. The open-source repository lacks code for the underlying implementation of the visualization module. The implementation of the visualization module is instead maintained in a private repository and built therefrom. In operation, the backend module calls an exposed function of a DLL that returns a pointer to a binary implementation of the visualization module.

Abstract: Techniques are provided for determining a stopping condition for 3D model decimation process. In an embodiment, a size of the 3D model may be determined based on a minimum oriented bounding box, wherein the 3D model includes vertices, edges, and faces that define a shape of a physical object. A smallest dimension value of the minimum oriented bounding box may be selected to represent the size of the 3D model. The selected dimensions value may be multiplied by a decimation factor to generate a stopping condition value. A decimation process may be performed on the 3D model until all remaining elements (e.g., edges) of the 3D model have an error value that is equal to or greater than the stopping condition value. As such, the level of detail of the 3D model is simplified while also preserving the shape of the 3D model.

Abstract: In example embodiments, techniques are provided for using a combination of multiple ML models and signal processing to extract links and connectivity from a schematic diagram in an image-only format. A first ML model (i.e. link segmenter) may produce a first set of predictions about the positions of link segments in the schematic diagram (e.g., in the form of a segmentation map). A second ML model (i.e. keypoint detector) may produce a second set of predictions about starting and stopping points of link segments in the schematic diagram (e.g., in the form of one or more heatmaps). A signal processing module may combine the first set of predictions and the second set of predictions to produce a description of links and connectivity they provide (e.g., combining the segmentation map with data from the one or more heatmaps). The results of the combining may be saved as a graph connectivity matrix.

Abstract: In example embodiments, anomaly and change detection software of a cloud-based design review service is provided for detecting anomalies and/or changes in 3D roadway models. The software analyzes the constituent meshes of the 3D roadway model that represent components and extracts template drops at locations along a horizontal alignment to produce an ordered list of template drops. The software then looks to differences in depths, widths, cross slopes and/or other geometric properties manifest in individual template drops, or between preceding/subsequent template drops of the ordered list, to detect anomalies and/or changes. Indications of the components associated with the detected anomalies and/or changes are displayed in a visualization of the 3D roadway model in a user interface.

Abstract: In example embodiments, a proprietary implementation of solid modeling is provided at run-time as a binary to an open-source infrastructure modeling platform. The binary includes functionality of a solid modeling module that manipulates and uses BReps that represent geometry of elements of an infrastructure model. When an application that utilizes the open-source infrastructure modeling platform requires solid modeling, it may have a backend module call an exposed function of a DLL that returns a pointer to the binary. The backend module uses the pointer to create a session, which may be divided into a number of individual partitions that each correspond to one of its individual threads. BReps may be assigned to individual partitions. When a thread requires BReps to be manipulated and/or used, the corresponding partition may be used to acquire the needed BReps, perform the solid modeling operations, and either return results or an error.

Abstract: Systems and methods are described for modeling and analyzing utility structures according to applied loads. Particularly, a model engine can utilize inputs related to a utility structure, environmental conditions to which the utility structure is subjected, and engineering standards expected of the utility structure, and analyze the structure's loading and performance based on analysis configuration inputs. An engine or multiple engines can be run locally or can be instantiated in a cloud to assist with multiple or complex calculations. Hybrid and geometric non-linear analyses and outputs can be performed or provided.

Abstract: In various example embodiments, techniques are provided for crack detection, assessment and visualization that utilize deep learning in combination with a 3D mesh model. Deep learning is applied to a set of 2D images of infrastructure to identify and segment surface cracks. For example, a Faster region-based convolutional neural network (Faster-RCNN) may identify surface cracks and a structured random forest edge detection (SFRED) technique may segment the identified surface cracks. Alternatively, a Mask region-based convolutional neural network (Mask-RCNN) may identify and segment surface cracks in parallel. Photogrammetry is used to generate a textured three-dimensional (3D) mesh model of the infrastructure from the 2D images. A texture cover of the 3D mesh model is analyzed to determine quantitative measures of identified surface cracks. The 3D mesh model is displayed to provide a visualization of identified surface cracks and facilitate inspection of the infrastructure.

Abstract: In example embodiments, techniques are described for implementing serverless property stores to hold properties that persist application customization data, such as settings. A serverless property store employs an “edge base” paradigm, wherein an edge computing device executes a property store service that maintains a local, periodically-synchronized copy of a portion of a database that stores properties (i.e., a local property cache”). A cloud container of a blob storage service of a cloud datacenter maintains a master copy of the database (i.e., a “property store database”). Read operations on a client computing device may be performed against the local property cache. Write operations may likewise be performed against the local property cache, however, they may be serialized via a write lock maintained in the cloud container. Multiple serverless property stores may be employed to store different properties each having different scopes.

Abstract: A sample may be generated for each point of a plurality of point clouds that represent a scene. A visibility ray may be created between each point of the plurality of point clouds and the one or more sources that generated the point. One or more sample, if any, that intersect a visibility ray may be identified. Each point corresponding to an intersecting sample may be determined to represent or likely represent an unwanted object if the visibility ray is from a different source that did not generate the point and the point is not coherent with any points generated by the different source. A visibility score for each point determined to represent or likely represent an unwanted object may be adjusted. A model may be generated, wherein the model does not include the unwanted object in the scene but includes the permanent object with see-through characteristics in the scene.

Abstract: In example embodiments, techniques are provided that use two different ML models (a symbol association ML model and a link association ML model), one to extract associations between text labels and one to extract associations between symbols and links, in a schematic diagram (e.g., P&ID) in an image-only format. The two models may use different ML architectures. For example, the symbol association ML model may use a deep learning neural network architecture that receives for each possible text label and symbol pair both a context and a request, and produces a score indicating confidence the pair is associated. The link association ML model may use a gradient boosting tree architecture that receives for each possible text label and link pair a set of multiple features describing at least the geometric relationship between the possible text label and link pair and produces a score indicating confidence the pair is associated.