Abstract Summary
The Author considers the dynamic analysis of the pedestrian crowd on the bridge. The key assumptions in this project are random values of the walking velocities, no interaction between pedestrians, and no traffic effects. The recurrent neural networks superimpose the weighted responses generated by N pedestrians walking individually across the bridge at different points in time. The approach of the algorithm is to evaluate each pedestrian, and their GRF and simulate the response of the structure. After all, responses have been calculated, the recurrent neural networks superimpose the responses to determine the overall structural response during the simulation window. The hypothesis is that N pedestrians cross the bridge in a 30-minute window. The arrival times (onto the bridge) are distributed according to Poisson distribution. The pedestrian movements or the distribution can be validated by GPS traces This is something you can look into if developing a more rigorous simulation. The modal force and response for each pedestrian result shifted along the time axis to match their random start time. and the Poisson distribution of the arrival time. In the successive step, the algorithm can calculate the combined crowd-induced loading and response simply by superimposing the data for each pedestrian individually according to the weighted sum method. Finally, the algorithm can plot the crowd loading and total modal response. This is the modal response so it corresponds to the response at the mid-span location of the beam. The response at any other point along the beam can be obtained using the mode shape. Obtained the crowd--induced dynamic displacement, the algorithm can numerically differentiate this vector of numbers twice to obtain the vertical acceleration of the bridge/beam. This result represents a good step as vibration serviceability limits are usually reported in terms of acceleration. The advantages of Recurrent Neural Networks are mainly three. The first is that they can process inputs of any length. Therefore the number of people can be an independent variable. This is in contrast to other networks that can only process fixed-length inputs. As a result, we can use RNN with both short and very long sequences without changing the architecture of our network. The second advantage is that the hidden state acts as some kind of memory. As the network processes the elements of a sequence one by one, the hidden state stores and combines information about the entire sequence. The algorithm can consider the mass effect of Pedestrians. The mass effect can be an innovative aspect of the relationship between modal frequency and pacing frequency. Finally, the third advantage of Recurrent Neural Networks is that weights can be variable through time phases. This allows the network to maintain the same size (with the same number of parameters) for variable-length sequences.
