skip to main content


Title: Design and Implementation of Experiential Learning Modules for Structural Analysis
Structural analysis is the foundation of a structural engineer’s education, which generally includes topics ranging from basic statics to solving complex indeterminate structures. Most courses focus on theoretical approaches to solving and understanding problems with a large focus on determining internal forces, external reactions, and displacements. In many cases, courses also include concurrent discussion about deflected shapes and actual behavior compared to theoretical assumptions. Courses may also use pictures, videos, simulations, and small, table-top models to illustrate such behavior. However, students still struggle with understanding structural behavior and the effects of as-built versus theoretical connections. Such differences are difficult to convey simply using photos, videos, simulations, and small, table-top models. Students never have the opportunity to physically feel the differences and experience large-scale models that illustrate such behavior mainly due to cost, fabrication complexity, and material stiffness. The authors from University A and University B designed large-scale, lightweight models for in-class use that allow students to experience structural behavior and feel the differences between various types of as-built connections. This paper provides a detailed overview of the design, fabrication, and implementation of four large-scale experiential learning modules for an undergraduate structural analysis course using lightweight and flexible fiberglass reinforced polymer (FRP) structural shapes. The first module focuses on the behavior of beam-to-column connections compared to theoretical assumptions; the second module focuses on load paths and tributary areas related to a typical floor system; the third module focuses on the deflected shapes of determinate and indeterminate beams; and the fourth module focuses on the behavior of a portal frame subjected to both vertical and horizontal loads with various support configurations. The four modules were used throughout the structural analysis course at University A and University B to illustrate structural behavior concurrent to the presentation of various structural analysis concepts.  more » « less
Award ID(s):
1726621
NSF-PAR ID:
10170890
Author(s) / Creator(s):
Date Published:
Journal Name:
ASEE Annual Conference proceedings
ISSN:
1524-4644
Format(s):
Medium: X
Sponsoring Org:
National Science Foundation
More Like this
  1. null (Ed.)
    Most undergraduate civil engineering programs include an introductory course in reinforced concrete design. The course generally includes an introduction to the fundamentals of reinforced concrete behavior, the design of simple beams and one-way slabs to resist shear and flexure, and the design of short columns. Because of the scale of typical civil engineering structures, students commonly do not get to experience large or full-scale structural behavior as a part of an undergraduate reinforced concrete course. Rather, students typically learn fundamental concepts through theoretical discussions, small demonstrations, or pictures and images. Without the interaction with full-scale structural members, students can struggle to develop a clear understanding of the fundamental behavior of these systems such as the differences in behavior of an over or under-reinforced beam. Additionally, students do not build an appreciation for the variations between as-built versus theoretical designs. Large-scale models can illustrate such behavior and enhance student understanding, but most civil engineering programs lack the physical equipment to perform testing at this scale. The authors from St. Louis University (SLU) and Rose-Hulman Institute of Technology (RHIT) have designed and implemented large-scale tests for in-class use that allow students to experience fundamental reinforced concrete behavior. Students design and test several reinforced concrete members using a modular strong-block testing system. This paper provides a detailed overview of the design, fabrication, and implementation of three large-scale experiential learning modules for an undergraduate reinforced concrete design course. The first module focuses on service load and deflections of a reinforced concrete beam. The first and second modules also focus on flexural failure modes and ductility. The third module focuses on shear design and failure modes. Each module uses a large scale reinforced concrete beam (Flexure specimens: 12 in. x 14 in. x 19 ft, Shear specimens: 12 in. x 14 in. x 10 ft.) that was tested on a modular strong-block testing system. The three modules were used throughout the reinforced concrete design course at SLU and RHIT to illustrate behavior concurrent to the presentation of various reinforced concrete design concepts. 
    more » « less
  2. Introductory steel design courses focus on the analysis and design of primary members, which typically include tension members and connections, compression members, flexural members, and beam-columns. Introducing structural steel design concepts to students presents its fair share of challenges. First, it is difficult for students to visualize and accurately predict the potential failure modes of a tension member: yielding of the gross section, rupture of the net section, and block shear. Second, it is also difficult for students to visualize the buckling modes of steel columns, which vary with shape and type of bracing. Students particularly struggle with the determination of buckling modes between strong and weak axes based on effective lengths. Third, flexural failure modes of steel beams are very difficult for students to visualize and understand when each mode controls. The failure modes are complex and fall into three categories for compact shapes: yielding of the cross section, inelastic lateral torsional buckling, and elastic lateral torsional buckling, which is dependent on the unbraced length of the compression flange. Non-compact sections also include local buckling of the flange or web, but identifying the relationship between the unbraced length and beam span and how the unbraced length affects the flexural capacity tends to be the most difficult concept for students to grasp. This paper provides a detailed overview of the design, fabrication, and implementation of three large-scale experiential learning modules for an undergraduate steel design course. The first module focuses on the tension connections by providing physical models of various failure types including yielding of the gross section, rupture of the net section, and block shear; the second module focuses on the capacity of columns with different amounts of lateral bracing about the weak axis; and the third module focuses on the flexural strength of a beam with different unbraced lengths to illustrate the difference between lateral torsional buckling and flange local buckling/yielding of the gross section. The three modules were used throughout the steel design course at Saint Louis University and Rose-Hulman Institute of Technology to illustrate the failure mechanisms associated with the design of steel structures. 
    more » « less
  3. null (Ed.)
    Geotechnical engineering undergraduate curriculum typically consist of courses in soil mechanics and foundation design that include a variety of topics that are difficult for students to understand and master. Behavior of the below grade natural and built geomaterials discussed in these courses can be difficult for students to visualize. Typically, the mechanisms of behavior are demonstrated using small-scale laboratory tests, two-dimensional sketches, simple table-top models, or video simulations in the classroom. Students rarely have the opportunity to observe large-scale behavior of foundations in the field or laboratory. The authors from Rose-Hulman Institute of Technology and Saint Louis University designed and implemented a large-scale foundation testing system to address several topics that students tend to struggle with the most, including 1) the difference in strength and service limit states in shallow foundation design, 2) soil-structure interaction associated with lateral behavior of deep foundations, and 3) the influence of near-surface soil on lateral behavior of deep foundations. This paper provides a detailed overview of the design, fabrication, and implementation of two large-scale experiential learning modules for undergraduate courses in soil mechanics and foundation engineering. The first module utilizes shallow foundations in varying configurations to demonstrate the differences in strength and service limit state behavior of shallow foundations. The second module utilizes a relatively flexible pile foundation embedded in sand to demonstrate the lateral behavior of deep foundations. The first module was used in the soil mechanics courses at Rose-Hulman Institute of Technology and Saint Louis University to compare theoretical and observed behavior of shallow foundations. The second module was used in the foundation engineering course at Rose-Hulman Institute of Technology to illustrate the concepts of soil-structure interaction and the influence of near-surface soil on lateral behavior of deep foundations. 
    more » « less
  4. Full-scale testing can be a powerful form of experiential learning in structural engineering courses. Most curricula focus on the proverbial “nuts and bolts” of structural engineering by teaching students to calculate forces and displacements along with member capacities. Pictures, videos, simulations, and small-scale projects are sometimes used to illustrate structural behavior. However, students regularly struggle to grasp structural behavior whether that is simply sketching a deflected shape or describing failure mechanisms. Rather than passively experiencing structural element or system behavior through pictures, videos, simulations, and small-scale projects, full-scale testing provides students with a first-hand, lasting understanding of fundamental behavior. Additionally, students also gain invaluable perspectives often difficult to glean from traditional classroom instruction such as constructability and tolerance issues. Full-scale testing is essential for student understanding of structural engineering concepts and there is a significant need for well-organized experiential learning opportunities with appropriate scales that successfully illustrate structural behavior. This paper provides the rationale and design approach for full-scale experiential learning opportunities in structural engineering. The rational of the project is based on faculty’s observations related to student understanding of structural behavior in Structural Analysis, Reinforced Concrete Design, Steel Design, and Foundation Design courses further reinforced by survey data regarding students’ perception of the most difficult topics to understand in each course. The design approach of the experiential learning modules highlights several factors including desired structural behavior, scale, testing capabilities, and implementation feasibility. The paper concludes with brief descriptions of thirteen experiential learning modules developed for the four courses to improve student understanding of structural behavior. 
    more » « less
  5. Obeid, Iyad Selesnick (Ed.)
    Electroencephalography (EEG) is a popular clinical monitoring tool used for diagnosing brain-related disorders such as epilepsy [1]. As monitoring EEGs in a critical-care setting is an expensive and tedious task, there is a great interest in developing real-time EEG monitoring tools to improve patient care quality and efficiency [2]. However, clinicians require automatic seizure detection tools that provide decisions with at least 75% sensitivity and less than 1 false alarm (FA) per 24 hours [3]. Some commercial tools recently claim to reach such performance levels, including the Olympic Brainz Monitor [4] and Persyst 14 [5]. In this abstract, we describe our efforts to transform a high-performance offline seizure detection system [3] into a low latency real-time or online seizure detection system. An overview of the system is shown in Figure 1. The main difference between an online versus offline system is that an online system should always be causal and has minimum latency which is often defined by domain experts. The offline system, shown in Figure 2, uses two phases of deep learning models with postprocessing [3]. The channel-based long short term memory (LSTM) model (Phase 1 or P1) processes linear frequency cepstral coefficients (LFCC) [6] features from each EEG channel separately. We use the hypotheses generated by the P1 model and create additional features that carry information about the detected events and their confidence. The P2 model uses these additional features and the LFCC features to learn the temporal and spatial aspects of the EEG signals using a hybrid convolutional neural network (CNN) and LSTM model. Finally, Phase 3 aggregates the results from both P1 and P2 before applying a final postprocessing step. The online system implements Phase 1 by taking advantage of the Linux piping mechanism, multithreading techniques, and multi-core processors. To convert Phase 1 into an online system, we divide the system into five major modules: signal preprocessor, feature extractor, event decoder, postprocessor, and visualizer. The system reads 0.1-second frames from each EEG channel and sends them to the feature extractor and the visualizer. The feature extractor generates LFCC features in real time from the streaming EEG signal. Next, the system computes seizure and background probabilities using a channel-based LSTM model and applies a postprocessor to aggregate the detected events across channels. The system then displays the EEG signal and the decisions simultaneously using a visualization module. The online system uses C++, Python, TensorFlow, and PyQtGraph in its implementation. The online system accepts streamed EEG data sampled at 250 Hz as input. The system begins processing the EEG signal by applying a TCP montage [8]. Depending on the type of the montage, the EEG signal can have either 22 or 20 channels. To enable the online operation, we send 0.1-second (25 samples) length frames from each channel of the streamed EEG signal to the feature extractor and the visualizer. Feature extraction is performed sequentially on each channel. The signal preprocessor writes the sample frames into two streams to facilitate these modules. In the first stream, the feature extractor receives the signals using stdin. In parallel, as a second stream, the visualizer shares a user-defined file with the signal preprocessor. This user-defined file holds raw signal information as a buffer for the visualizer. The signal preprocessor writes into the file while the visualizer reads from it. Reading and writing into the same file poses a challenge. The visualizer can start reading while the signal preprocessor is writing into it. To resolve this issue, we utilize a file locking mechanism in the signal preprocessor and visualizer. Each of the processes temporarily locks the file, performs its operation, releases the lock, and tries to obtain the lock after a waiting period. The file locking mechanism ensures that only one process can access the file by prohibiting other processes from reading or writing while one process is modifying the file [9]. The feature extractor uses circular buffers to save 0.3 seconds or 75 samples from each channel for extracting 0.2-second or 50-sample long center-aligned windows. The module generates 8 absolute LFCC features where the zeroth cepstral coefficient is replaced by a temporal domain energy term. For extracting the rest of the features, three pipelines are used. The differential energy feature is calculated in a 0.9-second absolute feature window with a frame size of 0.1 seconds. The difference between the maximum and minimum temporal energy terms is calculated in this range. Then, the first derivative or the delta features are calculated using another 0.9-second window. Finally, the second derivative or delta-delta features are calculated using a 0.3-second window [6]. The differential energy for the delta-delta features is not included. In total, we extract 26 features from the raw sample windows which add 1.1 seconds of delay to the system. We used the Temple University Hospital Seizure Database (TUSZ) v1.2.1 for developing the online system [10]. The statistics for this dataset are shown in Table 1. A channel-based LSTM model was trained using the features derived from the train set using the online feature extractor module. A window-based normalization technique was applied to those features. In the offline model, we scale features by normalizing using the maximum absolute value of a channel [11] before applying a sliding window approach. Since the online system has access to a limited amount of data, we normalize based on the observed window. The model uses the feature vectors with a frame size of 1 second and a window size of 7 seconds. We evaluated the model using the offline P1 postprocessor to determine the efficacy of the delayed features and the window-based normalization technique. As shown by the results of experiments 1 and 4 in Table 2, these changes give us a comparable performance to the offline model. The online event decoder module utilizes this trained model for computing probabilities for the seizure and background classes. These posteriors are then postprocessed to remove spurious detections. The online postprocessor receives and saves 8 seconds of class posteriors in a buffer for further processing. It applies multiple heuristic filters (e.g., probability threshold) to make an overall decision by combining events across the channels. These filters evaluate the average confidence, the duration of a seizure, and the channels where the seizures were observed. The postprocessor delivers the label and confidence to the visualizer. The visualizer starts to display the signal as soon as it gets access to the signal file, as shown in Figure 1 using the “Signal File” and “Visualizer” blocks. Once the visualizer receives the label and confidence for the latest epoch from the postprocessor, it overlays the decision and color codes that epoch. The visualizer uses red for seizure with the label SEIZ and green for the background class with the label BCKG. Once the streaming finishes, the system saves three files: a signal file in which the sample frames are saved in the order they were streamed, a time segmented event (TSE) file with the overall decisions and confidences, and a hypotheses (HYP) file that saves the label and confidence for each epoch. The user can plot the signal and decisions using the signal and HYP files with only the visualizer by enabling appropriate options. For comparing the performance of different stages of development, we used the test set of TUSZ v1.2.1 database. It contains 1015 EEG records of varying duration. The any-overlap performance [12] of the overall system shown in Figure 2 is 40.29% sensitivity with 5.77 FAs per 24 hours. For comparison, the previous state-of-the-art model developed on this database performed at 30.71% sensitivity with 6.77 FAs per 24 hours [3]. The individual performances of the deep learning phases are as follows: Phase 1’s (P1) performance is 39.46% sensitivity and 11.62 FAs per 24 hours, and Phase 2 detects seizures with 41.16% sensitivity and 11.69 FAs per 24 hours. We trained an LSTM model with the delayed features and the window-based normalization technique for developing the online system. Using the offline decoder and postprocessor, the model performed at 36.23% sensitivity with 9.52 FAs per 24 hours. The trained model was then evaluated with the online modules. The current performance of the overall online system is 45.80% sensitivity with 28.14 FAs per 24 hours. Table 2 summarizes the performances of these systems. The performance of the online system deviates from the offline P1 model because the online postprocessor fails to combine the events as the seizure probability fluctuates during an event. The modules in the online system add a total of 11.1 seconds of delay for processing each second of the data, as shown in Figure 3. In practice, we also count the time for loading the model and starting the visualizer block. When we consider these facts, the system consumes 15 seconds to display the first hypothesis. The system detects seizure onsets with an average latency of 15 seconds. Implementing an automatic seizure detection model in real time is not trivial. We used a variety of techniques such as the file locking mechanism, multithreading, circular buffers, real-time event decoding, and signal-decision plotting to realize the system. A video demonstrating the system is available at: https://www.isip.piconepress.com/projects/nsf_pfi_tt/resources/videos/realtime_eeg_analysis/v2.5.1/video_2.5.1.mp4. The final conference submission will include a more detailed analysis of the online performance of each module. ACKNOWLEDGMENTS Research reported in this publication was most recently supported by the National Science Foundation Partnership for Innovation award number IIP-1827565 and the Pennsylvania Commonwealth Universal Research Enhancement Program (PA CURE). Any opinions, findings, and conclusions or recommendations expressed in this material are those of the author(s) and do not necessarily reflect the official views of any of these organizations. REFERENCES [1] A. Craik, Y. He, and J. L. Contreras-Vidal, “Deep learning for electroencephalogram (EEG) classification tasks: a review,” J. Neural Eng., vol. 16, no. 3, p. 031001, 2019. https://doi.org/10.1088/1741-2552/ab0ab5. [2] A. C. Bridi, T. Q. Louro, and R. C. L. Da Silva, “Clinical Alarms in intensive care: implications of alarm fatigue for the safety of patients,” Rev. Lat. Am. Enfermagem, vol. 22, no. 6, p. 1034, 2014. https://doi.org/10.1590/0104-1169.3488.2513. [3] M. Golmohammadi, V. Shah, I. Obeid, and J. Picone, “Deep Learning Approaches for Automatic Seizure Detection from Scalp Electroencephalograms,” in Signal Processing in Medicine and Biology: Emerging Trends in Research and Applications, 1st ed., I. Obeid, I. Selesnick, and J. Picone, Eds. New York, New York, USA: Springer, 2020, pp. 233–274. https://doi.org/10.1007/978-3-030-36844-9_8. [4] “CFM Olympic Brainz Monitor.” [Online]. Available: https://newborncare.natus.com/products-services/newborn-care-products/newborn-brain-injury/cfm-olympic-brainz-monitor. [Accessed: 17-Jul-2020]. [5] M. L. Scheuer, S. B. Wilson, A. Antony, G. Ghearing, A. Urban, and A. I. Bagic, “Seizure Detection: Interreader Agreement and Detection Algorithm Assessments Using a Large Dataset,” J. Clin. Neurophysiol., 2020. https://doi.org/10.1097/WNP.0000000000000709. [6] A. Harati, M. Golmohammadi, S. Lopez, I. Obeid, and J. Picone, “Improved EEG Event Classification Using Differential Energy,” in Proceedings of the IEEE Signal Processing in Medicine and Biology Symposium, 2015, pp. 1–4. https://doi.org/10.1109/SPMB.2015.7405421. [7] V. Shah, C. Campbell, I. Obeid, and J. Picone, “Improved Spatio-Temporal Modeling in Automated Seizure Detection using Channel-Dependent Posteriors,” Neurocomputing, 2021. [8] W. Tatum, A. Husain, S. Benbadis, and P. Kaplan, Handbook of EEG Interpretation. New York City, New York, USA: Demos Medical Publishing, 2007. [9] D. P. Bovet and C. Marco, Understanding the Linux Kernel, 3rd ed. O’Reilly Media, Inc., 2005. https://www.oreilly.com/library/view/understanding-the-linux/0596005652/. [10] V. Shah et al., “The Temple University Hospital Seizure Detection Corpus,” Front. Neuroinform., vol. 12, pp. 1–6, 2018. https://doi.org/10.3389/fninf.2018.00083. [11] F. Pedregosa et al., “Scikit-learn: Machine Learning in Python,” J. Mach. Learn. Res., vol. 12, pp. 2825–2830, 2011. https://dl.acm.org/doi/10.5555/1953048.2078195. [12] J. Gotman, D. Flanagan, J. Zhang, and B. Rosenblatt, “Automatic seizure detection in the newborn: Methods and initial evaluation,” Electroencephalogr. Clin. Neurophysiol., vol. 103, no. 3, pp. 356–362, 1997. https://doi.org/10.1016/S0013-4694(97)00003-9. 
    more » « less