Academia.eduAcademia.edu

Figure 1 – uploaded by Carlton Ho

See full PDFdownloadDownload figure
Where TV, WV, AV, and SRV are the following: Table 1. Response actions based on 200-metre standard deviation values (via Network Rail)

Table 1 Where TV, WV, AV, and SRV are the following: Table 1. Response actions based on 200-metre standard deviation values (via Network Rail)

Related Figures (15)

Figure 2. Schematic track geometry quality course using an exponential model [25] Where Hyzinit (1)is the mean and of inj — the variance. The factors affecting the track ageing process have remained the same throughout the simulation. The results showed that the NLinit,, valu grew with each tamping activity until it eventually stabilized to one value. This is indicative of the lower efficiency of tamping over time, i.e., the degrading state of the ballast. This was further noticeable since the quality loss rate (b,,) continued to diverge with each tamping, meaning that the condition is degrading quicker (Fig. 3). The developed model was used to simulate the longitudinal level of the high-speed TGV line given two different maximum allowable velocities (250 km/h and 300 km/h). The simulation showed that the two cases followed a very similar pattern, with the only difference being that the simulation with the higher velocity exhibited quicker deterioration and will require more frequent interventions.
Figure 2. Schematic track geometry quality course using an exponential model [25] Where Hyzinit (1)is the mean and of inj — the variance. The factors affecting the track ageing process have remained the same throughout the simulation. The results showed that the NLinit,, valu grew with each tamping activity until it eventually stabilized to one value. This is indicative of the lower efficiency of tamping over time, i.e., the degrading state of the ballast. This was further noticeable since the quality loss rate (b,,) continued to diverge with each tamping, meaning that the condition is degrading quicker (Fig. 3). The developed model was used to simulate the longitudinal level of the high-speed TGV line given two different maximum allowable velocities (250 km/h and 300 km/h). The simulation showed that the two cases followed a very similar pattern, with the only difference being that the simulation with the higher velocity exhibited quicker deterioration and will require more frequent interventions.
Figure 3. NLinit (left) and b,, (right) after each tamping [25] A different study [26] has been carried out on UK railway lines regarding the whole lifecycle cost (WLCC) analysis. The work here mainly focuses on the economical aspect and analyses trade-offs between costs for transportation, construction, maintenance, and use. The track quality in this study was expressed as the standard deviation value of the horizontal and vertical geometry features over 200-metre segments. Monte-Carlo simulations were performed to generate a range of scenarios where the WLCC of an asset is estimated. One of the findings was that maintaining the quality of a high- speed passenger line at an average level was more beneficial economically. Intuitively, performing the minimum amount of tamping reduces the maintenance costs significantly, however that created a larger increase in track use costs, as well as increasing risk of derailment. The outcome of this study suggests that the maintenance scheduling of a railway track involves a trade-off between quality and cost and may be different for individual line classes. While this is an important aspect of quality assessment and maintenance scheduling, this study has not yet considered a lifecycle cost analysis and will focus mainly on assessing track quality.
Figure 3. NLinit (left) and b,, (right) after each tamping [25] A different study [26] has been carried out on UK railway lines regarding the whole lifecycle cost (WLCC) analysis. The work here mainly focuses on the economical aspect and analyses trade-offs between costs for transportation, construction, maintenance, and use. The track quality in this study was expressed as the standard deviation value of the horizontal and vertical geometry features over 200-metre segments. Monte-Carlo simulations were performed to generate a range of scenarios where the WLCC of an asset is estimated. One of the findings was that maintaining the quality of a high- speed passenger line at an average level was more beneficial economically. Intuitively, performing the minimum amount of tamping reduces the maintenance costs significantly, however that created a larger increase in track use costs, as well as increasing risk of derailment. The outcome of this study suggests that the maintenance scheduling of a railway track involves a trade-off between quality and cost and may be different for individual line classes. While this is an important aspect of quality assessment and maintenance scheduling, this study has not yet considered a lifecycle cost analysis and will focus mainly on assessing track quality.
Readings from inspections on the line in question are stored on a server and access to these data have been granted to the University of Edinburgh for research purposes. The data set explored is fron
Readings from inspections on the line in question are stored on a server and access to these data have been granted to the University of Edinburgh for research purposes. The data set explored is fron
he period September 2003 — April 2021. Usually, inspections are carried out every six weeks, which provides a large history of the behaviour of the track. During an inspection, the train sensors accelerometers) collect data in a raw format, which are immediately processed using the inertial method (double integration) for accelerometer data to obtain the final displacement readings. High- pass filters are applied to remove very long-wavelength forms (>100 metres) and design features. For vertical profiles, transducers are used to filter out the displacement of the suspension system and provide a measurement of the actual rail surface. As mentioned earlier, Network Rail employs two monitoring techniques as a means of scheduling interventions. The first one includes estimating standard deviation of geometry features for 200-mett track segments and studying them over time. Based on this value, maintenance activities may be required. An example of this is shown in Fig. 6. The large vertical drops are indicative of maintenan occurring at different points in time. The example shown includes both large-scale (red markers) anc local (green markers) maintenance works. It can be seen that for this segment, over the lifespan of th track the standard deviation for vertical profile readings does not exceed 1.4 mm, which is indicating a satisfactory state of the railway. This type of behaviour dominates most of the line, as maintenance thresholds are fairly low on high-speed tracks, and it is constantly held to a very high standard of quality. Table 1 shows typical actions by Network Rail given different ranges of standard deviation values for vertical profile based on line speed. Tracks with a lower maximum allowable velocity present a lower risk for passengers, and therefore have a higher maintenance threshold than high- speed tracks.
he period September 2003 — April 2021. Usually, inspections are carried out every six weeks, which provides a large history of the behaviour of the track. During an inspection, the train sensors accelerometers) collect data in a raw format, which are immediately processed using the inertial method (double integration) for accelerometer data to obtain the final displacement readings. High- pass filters are applied to remove very long-wavelength forms (>100 metres) and design features. For vertical profiles, transducers are used to filter out the displacement of the suspension system and provide a measurement of the actual rail surface. As mentioned earlier, Network Rail employs two monitoring techniques as a means of scheduling interventions. The first one includes estimating standard deviation of geometry features for 200-mett track segments and studying them over time. Based on this value, maintenance activities may be required. An example of this is shown in Fig. 6. The large vertical drops are indicative of maintenan occurring at different points in time. The example shown includes both large-scale (red markers) anc local (green markers) maintenance works. It can be seen that for this segment, over the lifespan of th track the standard deviation for vertical profile readings does not exceed 1.4 mm, which is indicating a satisfactory state of the railway. This type of behaviour dominates most of the line, as maintenance thresholds are fairly low on high-speed tracks, and it is constantly held to a very high standard of quality. Table 1 shows typical actions by Network Rail given different ranges of standard deviation values for vertical profile based on line speed. Tracks with a lower maximum allowable velocity present a lower risk for passengers, and therefore have a higher maintenance threshold than high- speed tracks.
Figure 6. Standard deviation values for a single 200-metre track segment (2003-2020) showing the quality of the left short-wave top profile (35m) and maintenance works done during the period. Local tamping includes manual packing and sprinter tamping.
Figure 6. Standard deviation values for a single 200-metre track segment (2003-2020) showing the quality of the left short-wave top profile (35m) and maintenance works done during the period. Local tamping includes manual packing and sprinter tamping.
Figure 7. Instances of track response after large-scale and local tamping. If ballast is renewed, this is expected to stabilize the track condition at that location and potentially reduce the rate of deterioration, assuming this was the root cause of the defect. However, if this procedure is performed multiple times at one location, it may lead to worn out material being inserted under the sleeper again, which may not be effective. These activities are usually performed by Network Rail in September/October each year. This is different from high output tamping, where the old ballast is simply compacted under the sleeper using a tamper, which will cause further rounding and wearing out of the particles. Figure 7 displays the difference in response of the standard deviation of the top left rail profile after both large-scale and local works within the same 50-metre track seoment.
Figure 7. Instances of track response after large-scale and local tamping. If ballast is renewed, this is expected to stabilize the track condition at that location and potentially reduce the rate of deterioration, assuming this was the root cause of the defect. However, if this procedure is performed multiple times at one location, it may lead to worn out material being inserted under the sleeper again, which may not be effective. These activities are usually performed by Network Rail in September/October each year. This is different from high output tamping, where the old ballast is simply compacted under the sleeper using a tamper, which will cause further rounding and wearing out of the particles. Figure 7 displays the difference in response of the standard deviation of the top left rail profile after both large-scale and local works within the same 50-metre track seoment.
others. Results show that on average, deterioration rates after large-scale works are approximately 42% higher than after sprinter tamping for vertical profiles and 63% higher for twist measurements.
others. Results show that on average, deterioration rates after large-scale works are approximately 42% higher than after sprinter tamping for vertical profiles and 63% higher for twist measurements.
number of other functions available, each of which may be more suitable for a specific problem and data set. Initially, the model will not perform well on training data set, at which point a backpropagation process begins to minimize this loss function [28]. This is done using the stochastic gradient descent method [29]. Each time the loss function is calculated for a given set of examples, gradient descent works its way back to the input layer and corrects the trainable parameters (weights and biases) to reduce the loss. The passage of the entire training set in smaller batches of examples (or as a whole) is known as one epoch. Minimization works by estimating the sensitivity of the loss function with respect to the weights and biases at each layer, i.e., the gradient. Selecting the optimal activation and loss functions for the specific problem helps the convergence. To estimate the gradient, the chain rule is applied.
number of other functions available, each of which may be more suitable for a specific problem and data set. Initially, the model will not perform well on training data set, at which point a backpropagation process begins to minimize this loss function [28]. This is done using the stochastic gradient descent method [29]. Each time the loss function is calculated for a given set of examples, gradient descent works its way back to the input layer and corrects the trainable parameters (weights and biases) to reduce the loss. The passage of the entire training set in smaller batches of examples (or as a whole) is known as one epoch. Minimization works by estimating the sensitivity of the loss function with respect to the weights and biases at each layer, i.e., the gradient. Selecting the optimal activation and loss functions for the specific problem helps the convergence. To estimate the gradient, the chain rule is applied.
Figure 10. Signal alignment using cross-correlation Firstly, readings from the two inspections must be aligned spatially, so the peaks and troughs of the two signals match and allow for a comparison to be made. To do this, cross-correlation between the signals was used — this takes one of the signals and slides it across the other one, until a good match is found. This process is also known as a sliding dot product, which can indicate by how many steps the first set of measurements is lagging behind the second and the two can be aligned. Figure 10 displays the results from the operation. This process has been made automatic for all measurements. Estimation of deterioration rate
Figure 10. Signal alignment using cross-correlation Firstly, readings from the two inspections must be aligned spatially, so the peaks and troughs of the two signals match and allow for a comparison to be made. To do this, cross-correlation between the signals was used — this takes one of the signals and slides it across the other one, until a good match is found. This process is also known as a sliding dot product, which can indicate by how many steps the first set of measurements is lagging behind the second and the two can be aligned. Figure 10 displays the results from the operation. This process has been made automatic for all measurements. Estimation of deterioration rate
plot below it, which can be automatically detected. The bottom half of the figure is showing the same phenomenon, although for deterioration. Each segment on the track will produce a different set of rates, which are used to help classify its condition. The number of values in each set will be defined by the number of measurements in the segment (typically 250). Two features related to these sets are used as input to the neural network algorithm — the standard deviation and skewness. Figure 11. Deterioration rates indicating improvement (top) and worsening (bottom) in quality. The blue and orange lines in the inspection comparison plots show the two consecutive inspections and the regions which have experienced a change. The deterioration plots show a rolling average for a smoother and more interpretable line.
plot below it, which can be automatically detected. The bottom half of the figure is showing the same phenomenon, although for deterioration. Each segment on the track will produce a different set of rates, which are used to help classify its condition. The number of values in each set will be defined by the number of measurements in the segment (typically 250). Two features related to these sets are used as input to the neural network algorithm — the standard deviation and skewness. Figure 11. Deterioration rates indicating improvement (top) and worsening (bottom) in quality. The blue and orange lines in the inspection comparison plots show the two consecutive inspections and the regions which have experienced a change. The deterioration plots show a rolling average for a smoother and more interpretable line.
Figure 12. Variation of deterioration rate SD and standard deviations ratio (SD1/SD2) in a worsened portion of the track (left) and in an improved portion of the track (right) Model training and performance The network consists of 3 layers in total — an input layer, hidden, and an output layer (see Fig. 8). The number of neurons in each layer were varied, but it was found that the model performed best when the input layer consisted of 50 neurons, the hidden layer of 15, and 3 neurons in the output layer (one for each class). The initial weights of the model were generated using a random normal distribution:
Figure 12. Variation of deterioration rate SD and standard deviations ratio (SD1/SD2) in a worsened portion of the track (left) and in an improved portion of the track (right) Model training and performance The network consists of 3 layers in total — an input layer, hidden, and an output layer (see Fig. 8). The number of neurons in each layer were varied, but it was found that the model performed best when the input layer consisted of 50 neurons, the hidden layer of 15, and 3 neurons in the output layer (one for each class). The initial weights of the model were generated using a random normal distribution:
Figure 14. Model comparison versus original measurements Following post-processing, the overall accuracy of the model on an unseen data set was 99.6%, although the data used for this evaluation was still highly imbalanced. The majority of the examples were class 3, which the network does not struggle to identify. The detec (maintenance) on the track was 88.9% and for class 1 (development of d ion accuracy for class 2 efects) — 77.5%. It is suspected that the better performance for detecting class 2 is partly due to the fact that maintenance results in a large absolute change in readings. On the other hand, deterioration is much more gradual and therefore not as statistically significant, making class | more difficu t to detect. Upon closer examination, it was found that the weaker performance in identifying de erioration is mainly due to small variations which are very close to the decision boundary of the model. Figure 14 provides examples of the classification abilities of the developed model.
Figure 14. Model comparison versus original measurements Following post-processing, the overall accuracy of the model on an unseen data set was 99.6%, although the data used for this evaluation was still highly imbalanced. The majority of the examples were class 3, which the network does not struggle to identify. The detec (maintenance) on the track was 88.9% and for class 1 (development of d ion accuracy for class 2 efects) — 77.5%. It is suspected that the better performance for detecting class 2 is partly due to the fact that maintenance results in a large absolute change in readings. On the other hand, deterioration is much more gradual and therefore not as statistically significant, making class | more difficu t to detect. Upon closer examination, it was found that the weaker performance in identifying de erioration is mainly due to small variations which are very close to the decision boundary of the model. Figure 14 provides examples of the classification abilities of the developed model.
Table 3. Estimated approximate maintenance efficiency The top vertical profiles (short-wave) and the twist measurements have been examined separately, as well as together. Similarly, long-wave profile was also checked individually. All features were then combined to come up with an overall efficiency value. Not all of the tamping records indicate the exact locations of tamping, but a specific asset where several smaller segments have been tamped. The impact of this has been removed as much as possible, however it may still have a residual effect and reduce these values (mainly for high output tamping). In addition, while most of the maintenance activities are known, there are also some missing records of both high output tamping and sprinter tamping/manual packing. If the full records are available, the efficiency may be slightly higher.
Table 3. Estimated approximate maintenance efficiency The top vertical profiles (short-wave) and the twist measurements have been examined separately, as well as together. Similarly, long-wave profile was also checked individually. All features were then combined to come up with an overall efficiency value. Not all of the tamping records indicate the exact locations of tamping, but a specific asset where several smaller segments have been tamped. The impact of this has been removed as much as possible, however it may still have a residual effect and reduce these values (mainly for high output tamping). In addition, while most of the maintenance activities are known, there are also some missing records of both high output tamping and sprinter tamping/manual packing. If the full records are available, the efficiency may be slightly higher.
Related topics:
Civil EngineeringBuilding and Construction Materials
580 California St., Suite 400
San Francisco, CA, 94104
© 2025 Academia. All rights reserved