the whitepaper - Auckland
Transkript
the whitepaper - Auckland
ARDA Activity Detection The cornerstone of true virtual coaching A white paper by ARDA Activity Detection The cornerstone of true virtual coaching The underpinning feature of training devices on the market today is their ability to classify data into ‘training zones’. A training zone is effectively a range of values that denote a particular ‘mode’for a measured parameter, logged over time or distance and adapted for a biological factor such as age/height/weight etc. For example, a 30-year-old with a heart rate between 110 and 130 bpm might be referred to as being in the ‘fat burning zone’ due to the high rate of fat oxidation at this particular training intensity. Similar training zones can be calculated from data provided by any tracking device. Training zones are useful for providing athletes with intensity targets during a workout. However, there are too many environmental and physiological variables involved in exercise for training zone analysis to provide any more complex feedback. Assumptions about these variables can lead to inaccurate interpretation of exercise data, and consequently inaccurate coaching advice in an automated system. For example, asking a user to ‘go into the fast zone’ assumes they are training on the flat. If the user is running up a particularly large hill, attempting to comply with the training device’s instructions to speed up could be frustrating, or could even result in injury. This can cause the user to lose faith in their training device, or even to stop using it entirely. Further to this, these assumptions prevent any kind of useful post-hoc workout analysis. TRAINING ZONES HEART RATE Figure 1: heart rate training zones © 2012 Performance Lab Technologies Limited. All rights reserved. Contains confidential and proprietary information. Unauthorised use or publication is expressly prohibited. 48 Enterprise St, Birkenhead, Auckland 0626, New Zealand Phone +64 9 4801422 web pltech.co.nz Email info@performancelab.co.nz 1 Does the training zone classification in the above graph help us understand what actually happened during the workout? We can see the ‘high intensity’ segments, but what terrain was the user on during those segments? If they were performed on the flat, then the user may have been doing speed work. If it was a hill, then the user was doing hill training. This difference may not matter if the goal of a workout is to simply achieve a particular heart rate, but for true performance analysis, context is absolutely vital. If a coach knows what kind of training was being conducted when a set of statistics was gathered, they can compare training segments to similar segments from previous workouts. This allows them to see which aspects of performance are improving, and which are deteriorating. The move to activity detection The ARDA engine is designed to perform three functions that are missing from today’s training tools: 1. Provide bespoke coaching to a user in real-time; 2. Modify a user’s training plan based on performance improvement or deterioration; 3. Help a user achieve very specific performance goals. To achieve this, raw data that is obtained from a user engaged in exercise must be processed by breaking up the workout into its component parts using ARDA’s Activity Classification system. As the user exercises, the system trawls for patterns in multiple streams of data generated by linked sensors (such as off-the-shelf accelerometers and heart rate monitors), waiting for a match to a certain set of criteria. © 2012 Performance Lab Technologies Limited. All rights reserved. Contains confidential and proprietary information. Unauthorised use or publication is expressly prohibited. 48 Enterprise St, Birkenhead, Auckland 0626, New Zealand Phone +64 9 4801422 web pltech.co.nz Email info@performancelab.co.nz 2 An example This example follows a cyclist training for an upcoming Ironman competition. The cyclist is male, 29 years old, 74 kg (163lbs), 180cm (5'11") tall. The athlete is training with a smartphone app powered by the ARDA engine. As the user cycles, ARDA continually samples the raw data from the cyclist’s sensors. In this instance a heart rate monitor, a cadence (RPM) meter, and an embedded GPS receiver. Coordinates from the GPS are matched against an in-built terrain database to determine current incline. 15 minutes into the workout, ARDA notices the following conditions: Heart Rate 168 bpm Cadence 70 rpm Incline 1% In ARDA’s classification scheme, these data points are within the classification range for a ‘big-gear time trial’, or ‘BGTT’ (thresholds for BGTT: a cadence of 62–77 rpm and a heart rate of 162–178 bpm on an incline of between -1% and 1%, that occurs for longer than 90 seconds). If the data remains in this range for 90 seconds, ARDA flags the beginning of a big-gear time trial activity and starts logging data and providing feedback based on this activity. In order to detect an activity’s completion in a sensible way, the software employs ‘edge forgiveness’ criteria for that activity. For a BGTT, if any parameter falls out of zone for more than 10 seconds, the big-gear time trial is considered to have ended and the data is labeled accordingly. This segment can now be compared against historic training of the same activity. © 2012 Performance Lab Technologies Limited. All rights reserved. Contains confidential and proprietary information. Unauthorised use or publication is expressly prohibited. 48 Enterprise St, Birkenhead, Auckland 0626, New Zealand Phone +64 9 4801422 web pltech.co.nz Email info@performancelab.co.nz 3 HEART RATE CADENCE TERRAIN POWER EASY BGTT EASY UP TEMPO EASY UP TEMPO EASY AT E P E P EASY TRAINING TYPES WORKOUT Figure 2: Training Types, or ‘Activities’ The chart above shows classification of cycling data where heart rate, cadence, and altitude are measured over the duration of a workout. These parameters have been used to define classifications of different activity segments, which appear at the bottom of the graphic. (e.g. ‘EASY’, ‘BGTT’, ‘UP TEMPO’ etc). In ARDA’s view, a workout is made up of a series of activity segments which together form the complete workout. The activity detection process can be applied to any cyclic sport. The ARDA engine currently supports running, cycling, swimming, and rowing. © 2012 Performance Lab Technologies Limited. All rights reserved. Contains confidential and proprietary information. Unauthorised use or publication is expressly prohibited. 48 Enterprise St, Birkenhead, Auckland 0626, New Zealand Phone +64 9 4801422 web pltech.co.nz Email info@performancelab.co.nz 4 Reliability from Research The sheer number of possibilities – overlapping data ranges, edge cases, and unexpected scenarios – makes activity detection by data analysis a very difficult problem to solve. If a virtual coach determines training context with a low level of reliability, any feedback given to the user or analysis performed on the data would be very difficult to trust. A virtual coach cannot in good conscience tell an athlete that they are suffering from muscular fatigue unless it knows what other athletes’ biometric data looks like on a similar incline, in similar weather conditions, with a similar training history, weight, height, and gender etc. The ARDA engine activity detection system is based on 20 years of research conducted during the professional coaching of over 3,000 men and women of all ages, from elite athletes training to win world championships, to sedentary individuals training for health and wellbeing. The techniques forming the basis of the ARDA software have been in use by PL Tech for many years, with a long list of successes in the field of remote performance training for elite athletes. Activity Detection revolutionises workout analysis Accurately and reliably classifying workout segments into training types allows a training device to start doing much more with the data it receives. The main areas of difference are true coaching advice, smart plan adjustment, and user-guided workouts. True coaching advice The difficulty of generating coaching advice using an automated training device is in determining the reasons why changes to the athlete’s biometric data have occurred. A particular change, or combination of changes, in heart rate, stride rate, or pace, might be entirely expected during a hill workout, but on the flat the same changes might indicate a problem that needs correction. Contextualised activity data allows ARDA to compare a runner’s data to a ‘baseline’ for a particular activity. This baseline corresponds to what an athlete ‘should’ be accomplishing based on their personal goals, their biometrics, and their training history. © 2012 Performance Lab Technologies Limited. All rights reserved. Contains confidential and proprietary information. Unauthorised use or publication is expressly prohibited. 48 Enterprise St, Birkenhead, Auckland 0626, New Zealand Phone +64 9 4801422 web pltech.co.nz Email info@performancelab.co.nz 5 In addition, contextualised raw data allows ARDA to keep track of performance trends over multiple workout sessions, augmenting a device’s existing ability to track general trends such as pace or general fitness. ARDA’s analysis engine detects the gap between a user’s ideal performance for an activity (based on their goals), as well as any trending areas of weakness based on prior workouts, and generates advice designed to close those gaps. Generated advice is prioritized, favouring more significant events over less significant events, and favouring broader trend summaries over more specific single issues. This prioritized coaching direction can then be given to the user audibly or visually. Without activity detection, today’s devices are forced to limit their feedback to effort correction, i.e. ‘go faster’, or ‘go into the easy zone’. Activity Detection allows a device to make much more detailed corrections. ARDA can suggest corrections to a user’s stride length, or make subtle pace corrections based on comparative performance on different inclines. Activity Detection shifts the focus of a training device from enforcing generic training plans to analysing how well an athlete is performing a particular task. This feature allows for a change in the paradigm of workout prescription. Users of training devices commonly complain that they are forced to enter speedwork sessions at inappropriate times. ARDA switches this around, and simply informs the user that they have to do a speedwork session at some point during their run. ARDA will detect when speedwork is taking place, and check it off the list. Additionally, a workout plan can now call for more detailed behaviours that depend on a user finding the correct environment, such as ‘5 steep hill climbs’. Smart Plan Adjustment Because ARDA’s activity detection engine takes into account a user’s biometric data (such as height and weight) and training history when a person starts using the device, it has the ability to generate highly specific training plans for users. In fact, there are dozens of questions that can be asked to refine an initial workout plan for a user. ARDA views these training plans in a unique way. Instead of a static, © 2012 Performance Lab Technologies Limited. All rights reserved. Contains confidential and proprietary information. Unauthorised use or publication is expressly prohibited. 48 Enterprise St, Birkenhead, Auckland 0626, New Zealand Phone +64 9 4801422 web pltech.co.nz Email info@performancelab.co.nz 6 cookie-cutter programme, an ARDA training plan will adapt over time based on the user’s performance trends. For example, if a user is showing excellent performance improvement on steep hill climbs, but is regularly demonstrating poor speedwork, the engine can shift the focus of future workouts onto race-pace training or pacing techniques. Furthermore, if an athlete is showing high levels of fatigue, ARDA can modify or postpone workouts to aid recovery. ARDA is a software engine capable of providing true performance analysis, coaching advice, and adaptive plans to amateur users and pro athletes by integrating with existing training devices. Performance Lab is currently approaching hardware vendors to license the technology. For more information, or to schedule a tech demo of the ARDA engine, please contact PLTech: Email info@performancelab.co.nz Phone Kerri McMaster +64 9 480 1422 © 2012 Performance Lab Technologies Limited. All rights reserved. Contains confidential and proprietary information. Unauthorised use or publication is expressly prohibited. 48 Enterprise St, Birkenhead, Auckland 0626, New Zealand Phone +64 9 4801422 web pltech.co.nz Email info@performancelab.co.nz 7