114 can determine the railway networks connected between the origin and the destination by: discovering any one or more sections of any one or more of the railway networks being interconnected to one another between the origin and the destination; and outlining any one or more routes along the any one or more sections connecting the origin to the destination.

At this stage of the clearance process 200, the predictive modelling module 108 obtains the historical clearance data 140 for the determined route(s) (Block 210). Some of the relevant historical clearance data 140 may be stored in the database 102, and some may be accessed from a remote system 118.

Using the one or more determined route(s), the dimensional parameters, and the historical clearance data, the AI model 110 of the predictive modelling module 108 performs predictive modeling (Block 212) and determines predictive results (e.g., probabilities, confidence intervals, pass-fail scores, tabulated numbers of critical points, etc.) for clearance of the load carried by the railcar along the determined route(s) (Block 214). The predictive results can in general include one or more metrics characterizing the transport envelope of the load carried on the railcar clearing the historical clearance data on the one or more railway routes. As its goal, predictive modelling module 108 seeks to calculate a probability value, a confidence interval, a pass-fail score, a tabulated number of critical points, or other numerical metric of securing suitable clearance along the one or more identified routes. For example, using the AI model 110, the predictive modelling module 108 determines probabilities or other metrics characterizing the transport envelope of the load carried on the railcar clearing obstacles in the historical clearance data on the determined route(s) of the railway network(s). The calculations can use dynamic adjustment factors, safety margins, and accuracy metrics. The analysis for the clearance may also use calculations for any speed restrictions in passing critical points along the railway route. Because the required processing may be intensive, the computer system 100 can use dedicated and/or remote resources to perform the analysis, such as provided by a remote system 118.

Using the AI model 110, the predictive modelling module 108 generates a clearance prediction (Block 216), which is assessed (Decision 218). If the clearance is appropriate (Yes), the predictive modelling module 108 determines, based on the determined probabilities or other metric(s), at least one route on one or more of the determined routes connecting the origin to the destination and generates a clearance report or recommendation (Block 226). The clearance report can include information indicative of the at least one route to transport the load carried on the railcar. This clearance report can be communicated with one or the input-output interfaces, such as a computer screen, a printer, an electronic communication, or the like.

In some instances, the clearance process 200 may produce clearance probabilities or metrics that are not adequate when the clearance process 200 determines whether clearance has been achieved within an appropriate threshold of probability or the like (Decision 218). In general, the AI model 110 can determine that the one or more metrics for the railway routes fail to meet a criterion for the transport envelope of the load carried on the railcar to clear the historical clearance data on the railway routes. For example, the AI model 110 may instead determine that the probabilities fall below a threshold. Should clearance issues arise, the computer system 100 can mark the schematic data, such as the image file, what the “clearance window” would be (Block 220). In this case, the AI model 110 can define at least a maximum clearance window along at least one of the routes having the one or more metrics closest to meeting the criterion, e.g., having at least a higher level of the probabilities. Then, the AI model 110 can suggest one or more alternatives to at least the railcar to match the maximum clearance window (Block 222) and can revise the proposal (Block 224) so the analysis can be repeated. Other alternatives can also be recommended. For example, an alternative railway route may be determined and suggested to the user.

At Block 222, the computer system 100 can also automatically select a railcar type based on characteristics of the load and the identified routes. For example, given the load's dimensional parameters, the historical clearance data, and the one or more identified routes, the computer system 100 may determine an appropriate type of railcar for the load if a specialized form of transport is necessary. For example, the computer system 100 may determine a specialized type of railcar to achieve the transport. To do this, the computer system can reference a database of specialized railcars, either stored in the database or obtained from an external system (e.g., from Kasgro Rail Corporation).

To obtain the input data in Block 202, the schematic data can be manually input by a user in a user interface by filling out form fields for different dimensions of interest. Although this represents one possibility, allowing schematic image data to be used greatly simplifies and streamlines the clearance process. Accordingly, to obtain the input data in Block 202, the schematic data can be schematic image data obtained using an image capture interface, such as a camera, a scanner, or other imaging devices. Alternatively, the schematic data may be an image file stored in the database 102 of the computer system 100, and the processor 106 can access the image file from storage. The image file can be received from a remote system (118) and can be downloaded to the database 102 for later retrieval. Likewise, the image file can be generated with the computer system 100 using appropriate software and stored in the database 102 for later access.

In fact, AI models at remote systems 118 can be accessed through application program interfaces (API) to generate AutoCAD drawings for the image files, including generating code snippets in various programming languages to automate and enhance AutoCAD design processes. Additionally, open-source applications can generate CAD files from text prompts, allowing models to be created and imported into CAD programs.

As examples, FIGS. 3A and 3B illustrate representations of image files 230A-B for use by the disclosed computer system 100. These image files 230A-B can include any combination of depictions of the load with dimensional information, depictions of a railroad car with dimensional information, and tables and/or other textual information. For example, the image files 230A-B in FIGS. 3A-3B include an end view 232 of the load (e.g., transformer) with dimensional information when supported on a selected railroad car. The image files 230A-B also include a side elevational view 234 of the load (e.g., transformer) with dimensional information when supported on the selected railroad car. A table 236 is depicted and includes clearance dimensions in the form of widths of the load on the railroad car at different heights from the rail. Additional textual information 238 may also be provided as shown in the image file 230B. The AI model 110 can use transforms 111 and language models to extract text from any tables 236 and textual information 238 in the image files.

In the clearance process 200 of FIG. 2, dimensional parameters for the load and the railcar can be extracted from these types of image files by the image processing module 112. As an example, manufacturer drawings of the load to be transported can be uploaded to the computer system 100. The load can be equipment, such as a transformer, a boiler, a wind turbine component, or any type of heavy-lift and over-dimension cargo, which requires significant coordination and time to transport. The drawings may also include a selected railcar on which the load is to be transported.

The image file for these drawings can be in a suitable format, such as PDF, AutoCAD, or other formats. The computer system 100 processes the image file to determine dimensional parameters related to the load and the railcar (if present in the image file). These dimensional parameters can include one or more values for the load's vertical dimension (e.g., height), lateral dimension (e.g., width, diameter, etc.), and the longitudinal dimension (e.g., length).

Moreover, more than one image file can be uploaded to, retrieved from, or generated by the computer system 100 to be processed and combined, such as one image file for the load and another image file for the railcar. If the image files uploaded to the computer system do not include the railcar, for example, then one or more separate image files for the desired railcar may be uploaded to the computer system. Alternatively, a particular railcar can be separately selected in the computer system 100, and the dimensions for the railcar combined with the dimensional parameters of the load extracted from the image file. For example, user selections can be made in a user interface of the computer system to select a desired railcar. Typical rail cars include a flat car, a bulkhead flat car, a gondola car, a hopper car, and the like.

To determine the dimensional parameters associated with the load carried on the railcar, first dimensional parameters associated with the load can be determined by processing the schematic image data of the load without image data of a railcar to be used. Instead, a user-based selection of the railcar can be obtained using a graphical user interface, and the selection can be obtained from storage in the database. Alternatively, the computer system 100 may make the selection of the railcar automatically based on the first dimensional parameters associated with the load as well as any other details related to the load (e.g., name of the load, type of the load, etc.). Second dimensional parameters associated with the selection of the railcar are then added to the first dimensional parameters associated with the load to complete the combined dimensional parameters.

Selecting the railcar for the given load and route(s) by the computer system 100 may take into account one or more parameters, including weight of the load relative to the railcar's capacity, length of the load relative to the platform size of the railcar, height of the load compared to the platform height, number of axles for calculating and approving weight distribution per axle on the railcar, location of the center of gravity (COG) of the load, and availability and cost of the railcar. Not all railcars may be available for every route and departure location. If a specific car is found to be unavailable, the process can adjust the selection accordingly. Railcar data can be accessed directly from remote systems (118) to obtain technical specifications, technical diagrams, and official data from railcar manufacturers.

As an alternative, the schematic image data being processed may have both the load and the railcar included. In this case, the combined dimensional parameters associated with the load carried on the railcar can be completed based on processing the image data.

To define the transport envelope of the load carried on the railcar, the computer system 100 appropriately scales and combines the dimensional parameters. For better understanding of the transport envelope, FIG. 4 illustrates an example of a maximum clearance profile 240 for a train on a railway. As noted above, dimensional parameters can include one or more values for the load's vertical dimension (e.g., height), lateral dimension (e.g., width, diameter, etc.), and the longitudinal dimension (e.g., length). Additionally, the dimensional parameters of the computer system 100 can include one or more values of the railcar's vertical dimension (e.g., height), lateral dimension (e.g., width, diameter, etc.), and longitudinal dimension (e.g., length). Moreover, given that the load is to be carried on the railcar and may be held on various support structures, the dimensional parameters can include one or more values for the combined vertical dimension (e.g., combined height above the rail), lateral dimension (e.g., width, diameter, etc.), and the longitudinal dimension (e.g., length). Additional parameters, such as weight, horizontal center of gravity relative to the geometric center, combined vertical center of gravity above the rail, etc., related to the load and/or transport may be extracted from the image file or received through user inputs.

The clearance in the maximum clearance profile 240 is defined as a distance from an outer edge of the load to structures on the railroad right-of-way. Many railroads have different minimum clearance distances to be met. The clearance required can also be related to the speed of the railcar passing the structure. For example, a smaller clearance distance would relate to a slower speed, whereas a greater clearance distance would relate to a higher speed. Some obstacles may require the train to pass at walking speed to pass the obstacle. A predefined clearance distance may be needed for the railcar to travel at track speed past the structure. Because trains on adjacent rails may pass by the load, clearance requirements also account for the track centers (i.e., the distance from the centerline of one track to the centerline of adjacent track(s)).

In general, a “loading gauge” can refer to a maximum physical size of a railcar and its load. By measuring various dimensions along the length of the railcar and carried load, the processor can determine the loading gauge for the specific configuration. Although the loading gauge describes the outer dimensions of configuration of the railcar and its load, the railcar with the load can occupy a more dynamic envelope representing a larger volume that rolling stock can occupy as it travels along a railway track at speed.

For example, FIG. 5A illustrates an example of a calculated clearance envelope 250 for a railcar and a load. The envelope is generally defined by the combination of height, width, and the edge chamfers from both the top and bottom corners of the load profile (front view) on the railcar relative to the top of the rail and the centerline. Lateral swaying, vertical bouncing, track canting around corners, and the like can produce the larger dynamic clearance envelope, which is also shown in FIG. 5A.

For further explanation, FIG. 5B illustrates a more detailed example of a calculated clearance envelope 255 for a railcar and a load. In this example, the load is a transformer. The railcar (not shown) in this example is 9′ 4″ wide and 2′ 5″ high. Multiple coordinates (i.e., widths relative to the railcar's centerline at specific heights from the top of rail) for the geometric data are detailed for the load, and the profile for the calculated clearance envelope 255 is defined by these coordinates. The load's center of gravity is also given.

As noted previously, the predictive modelling module 108 in FIG. 2 calculates a distance between obstacles (structures, etc.) and the clearance envelope of the railcar and the load. The calculation may account for the speeds at which the envelope can pass. Additionally, the calculation can account for appropriate track tolerances and the accuracy of measurements. The clearance envelope developed by the computer system 100 may be displayed in a graphical user interface, and the user may be able to make adjustments and refinements in the interface.

In the predictive analysis, the clearance envelope of the load on the railcar is checked for fit and clearance issues along the route(s). Clearance data along the route is accessed from a clearance database to perform the comparative fit. Weight restrictions (e.g., weight restriction data) along the route are also checked in the comparison. Therefore, the criteria for the clearance prediction may account for the overall size and gross weight of the load, the maximum allowable dimensions (“envelope”), and the combined weight and size of load and railcar for the route.

For height and width of the load on the railcar, the calculations add together the load's and railcar's height and width, and the calculations check that the combined heights and widths constructing the transport envelope do not exceed the clearance limitations defined in the maximum allowable dimensions (“envelope”) for a given route. To calculate the weight per axle, the weight of the load is combined with the weight of the empty car, and the total weight is divided by the number of axles on the car. The aim is to keep the center of gravity (COG) of the load as close as possible to the center of the railcar, using counterweights if necessary. Additionally, the COG of the load is preferably as close as possible to the longitudinal center of the railcar to avoid creating an imbalance in the weight distribution that exceeds certain thresholds, such as 10% on either axle group (front or rear).

As noted above, the computer system 100 can receive user inputs for the transportation of the load. To obtain the input data, for example, the computer system 100 can obtain the origin and the destination using a graphical user interface. For example, FIG. 6 illustrates an example graphical user interface 260 for configuring, selecting, and entering information. As will be appreciated, any number of formats can be used for the graphical user interface 260 of the computer system (100).

In this example, upload of a schematic image file can be selected for the load and railcar (262). A drop-down selection (264) of available railcars can be used. Geographical regions (266) for mapping can be selected, and an origin and a destination for the transportation of a load can be selected or entered. Within the graphical user interface 260, the mapping information may be selectable as stored locations in the computer system, may be selectable locations on a visual map, or may be form fields for entering physical addresses, locations, or other geographical information.

Once schematic image file(s) and other inputs are entered, analysis (268) can be initiated in the graphical user interface 260 to determine the clearance envelope and provide a result (270) for the given inputs. As disclosed herein, the result (270) can take the form of a predictive clearance value, e.g., an estimate of how the given load on the selected railcar can clear the various clearance limits that are known on the one or more selected routes. The results (270) can be calculated as a percentage, a confidence level, or another numerical value. The results (270) can be arranged in a hierarchy or other type of comparison. The results (270) can be output in another graphical user interface, in tables, and in any other suitable format.

Should the predicted clearance fall below a predefined threshold or other metric, the computer system (100) can highlight any issues in the results (270), such as pinpointing any critical points where clearance is restricted or obstructed. The computer system (100) can also determine and provide a suggested alteration within the results (270). For example, the computer system (100) can provide one or more alternative railroad cars for transporting the load. Also, the computer system (100) can provide one or more alternative routes for transporting the load.

As noted above, the disclosed systems and methods use AI techniques, such as a convolutional neural network (CNN) for image-based evaluations. The CNN is trained directly with graphical representations to evaluate and classify the quality of the threaded tubular connections.

FIG. 7 schematically illustrates a convolutional neural network (CNN) 300 used for automated evaluation and analysis of graphical representations in a computing environment. (Reference numerals to elements in other figures are provided in the discussion below.)

Again, the computing environment (50) can include the computer system (100), remote system (118), and processes (200) discussed above. The CNN 300 is a type of deep neural network (DNN) having three additional features: local receptive fields, shared weights, and pooling. An input layer 310 and an output layer 350 of the CNN 300 function similar to the input and output layers of a DNN. However, the CNN 300 is distinguished from a DNN in that hidden layers of the DNN are replaced with one or more convolutional hidden layers 320, pooling hidden layers 330, and fully connected hidden layers 340.

Using localized receptive fields, nodes in the convolutional hidden layers 320 receive inputs from localized regions in the previous layer. Meanwhile, using shared weights, each node in a convolutional hidden layer 320 assigns the same set of weights to the relative positions of a localized region.

The input layer 310 of the CNN 300 includes data representing an image (e.g., a graphical representation, graphical user interface, graphs, curves, tables, etc. uploaded to the image processing module 112). For example, the data can include an array of numbers representing the pixels of the image, with each number in the array including a value from 0 to 255 describing the pixel intensity at that position in the array. The image can be passed through a convolutional hidden layer 320, an optional non-linear activation layer (not shown), a pooling hidden layer 330, and fully connected hidden layers 340 to get an output at the output layer 350. While only one of each hidden layer is shown in the present example, it is appreciated that multiple convolutional hidden layers 320, non-linear layers, pooling hidden layers 330, and/or fully connected hidden layers 340 can be included in the CNN 300.

The first layer of the CNN 300 is the convolutional hidden layer 320, which analyzes the image data of the input layer 310. Each node of the convolutional hidden layer 320 is connected to a region of nodes (pixels) of the input image called a receptive field. The convolutional hidden layer 320 can be considered as one or more filters (each filter corresponding to a different activation or feature map), and each convolutional iteration of a filter can be considered a node or neuron of the convolutional hidden layer 320. For example, the region of the input image that a filter covers at each convolutional iteration would be the receptive field for the filter. Each connection between a node and a receptive field for that node learns a weight and, in some cases, an overall bias such that each node learns to analyze its particular local receptive field in the input image. Each node of the convolutional hidden layer 320 will have the same weights and bias (called a shared weight and a shared bias). For example, the filter has an array of weights (numbers) and the same depth as the input.

The convolutional nature of the convolutional hidden layer 320 is due to each node of the convolutional layer being applied to its corresponding receptive field. At each convolutional iteration, the filter's values are multiplied by a corresponding number of the original pixel values of the image data. The multiplications from each convolutional iteration can be summed together to obtain a total sum for that iteration or node. The process is continued at a next location in the input image according to the receptive field of the next node in the convolutional hidden layer 320. For example, a filter can be moved by a step amount to the next receptive field. Processing the filter at each unique location of the input volume produces a number representing the filter results for that location, resulting in a total sum value being determined for each node of the convolutional hidden layer 320.

The mapping from the input layer 310 to the convolutional hidden layer 320 is referred to as an activation map (or feature map). The activation map includes a value for each node representing the filter results at each location of the input volume. The activation map can include an array containing the various total sum values resulting from each iteration of the filter on the input volume. The convolutional hidden layer 320 can include several activation maps to identify multiple features in an image.

Applied after the convolutional hidden layer 320, the pooling hidden layer 330 simplifies the information in the output from the convolutional hidden layer 320. The pooling hidden layer 330 takes each activation map output from the convolutional hidden layer 320 and generates a condensed activation map using a pooling function. Max-pooling is one example of a pooling function that can be performed by the pooling hidden layer 330. The pooling hidden layer 330 may also use other known forms of pooling functions. The pooling function is applied to each activation map in the convolutional hidden layer 320.

In the final layer of connections in the CNN 300, the fully connected hidden layer 340 connects every node from the pooling hidden layer 330 to every one of the output nodes in the output layer 350. The fully connected hidden layer 340 obtains the output of the previous pooling hidden layer 330 (which represents the activation maps of high-level features) and determines the features that best correlate to a particular class. For example, the fully connected hidden layer 340 can determine the high-level features that strongly correlate to a particular class and can include weights (nodes) for the high-level features. A product can be computed between the weights of the fully connected hidden layer 340 and the pooling hidden layer 330 to obtain probabilities for the different classes. For example, if the CNN 300 is being used to predict that an object is a torque-turns curve, high values will be present in the activation maps that represent high-level features of a torque-turns curve.

As noted previously, the modules (e.g., the predictive machine learning module 108, the image processing module 112, and the mapping module 114) can be implemented using appropriate forms of artificial intelligence, such as a deep neural network or other AI models. In the ML pipeline, the AI models of the present disclosure can be trained by a software framework of a machine learning engine that manages, trains, deploys, and serves the AI model 110 according to the present disclosure. Existing engines, such as TensorFlow Serving or AWS SageMaker, can provide the infrastructure to deploy the AI model 110 of the present disclosure.

As an example, FIG. 7 illustrates a training process 400 for a training framework to train and deploy a neural network 418 in which configurations of the present technology may be implemented. The neural network 418 includes an input layer, a plurality of hidden layers, and an output layer. The neural network 418 can be a deep neural network (DNN), a deep auto-encoder neural network (deep ANN), a convolutional neural network (CNN), a recurrent neural network (RNN), or any other suitable neural network.

For the purposes of the present disclosure, the neural network 418 for the image processing module (112) can use a CNN as discussed herein to process schematic data for image file(s). The CNN is specifically designed for working with grid-like data, such as images, and can effectively perform tasks, such as image classification, object detection, image segmentation, and the like. The CNN uses convolutional layers to automatically detect patterns, such as edges, textures, and shapes, within the schematic image data to capture spatial hierarchies. Pooling layers within CNN reduce the spatial size of the representation, making the model less sensitive to small shifts or distortions in the image.

For the purposes of the present disclosure, the predictive modeling module (108) can use a DNN because the input data may be complex and unstructured. The DNN can handle the complex data structures to learn intricate patterns. The deep learning models, including CNNs, RNNs, and transformers, can be employed. For example, two or three-dimensional models of the railcar/load and obstacle clearances can train CNNs to predict whether the railcar can pass through specific obstacles on the railways. The model can analyze the shape and clearance of obstacles and compares them with the load/railcar's dimensions.

For the purposes of the present disclosure, the mapping module (114) can use any number of algorithms to determine routes, pathways, and logistics between the original and destination along the railways. For example, the mapping module (114) can use a Graph Neural Network (GNN) for graph-based route planning. Other algorithms include Genetic Algorithm (GA), Ant Colony Optimization (ACO), Mixed Integer Linear Programming (MILP), and Heuristic Search Algorithm.

In the end, each function of the AI models (110, 113, 115) in the various modules (108, 112, 114) could be combined together in a neural network.

As illustrated, the neural network 418 may have any number of two or more hidden layers. Each layer may have one or more nodes (represented by circles in the diagrammatic network). As depicted by the connecting lines, each node in a current layer is connected to every other node in a previous layer and a next layer. This is referred to as a fully connected neural network. Other neural network structures are also possible in alternative arrangements of the neural network 418, in which not every node in each layer is connected to every node in the previous and next layers.

Each node in the input layer can be assigned a value and output that value to every node in the next layer (e.g., hidden layer). The nodes in the input layer can represent features about a particular image. For example, a DNN used for classifying whether an object is a rectangle may have an input node representing whether the object has flat edges. In this example, assigning a value of 1 to the node may represent that the object does have flat edges and assigning a value of 0 to the node may represent that the object does not have flat edges. In another example, for the DNN taking an image as input, the input nodes may each represent a pixel of the image, such as a pixel of a training image, where the assigned value may represent the intensity of the pixel. Following this example, an assigned value of 1 may indicate that the pixel is completely black and an assigned value of 0 may indicate that the pixel is completely white.