算法构成

这里简要介绍在各种算法中使用的个部分。

Delay

This is used to buffer a signal so you can time align it to some other operation. For example, if you average nine samples and you want to subtract the average from the original signal, you can use a delay of five samples of the original signal so you can use values that are itself plus the four samples before and four samples after.

Attenuate

Attenuation is a simple but useful operation that can scale a signal down before it is amplified in some fashion with filtering or some other operation that adds gain to the signal. Typically attenuation is measured in decibels (dB). You can attenuate power or amplitude depending on your application. If you cut the amplitude by half, you are reducing it by -6 dB. If you want to attenuate by other dB values, you can check the dB scale here. As it relates to the Speedbag algorithm, I’m basically trying to create clear gaps in the signal, for instance squelching or squishing smaller values closer to zero so that squaring values later can really push the peaks higher but not having as much effect on the values pushed down towards zero. I used this technique to help accentuate the bursts of acceleration versus background vibrations of the speed bag platform.

Sliding Window Average

Sliding Window Average is a technique of calculating a continuous average of the incoming signal over a given window of samples. The number of samples to be averaged is known as the window size. The way I like to implement a sliding window is to keep a running total of the samples and a ring buffer to keep track of the values. Once the ring buffer is full, the oldest value is removed and replaced with the next incoming value, and the value removed from the ring buffer is subtracted from the new value. That result is added to the running tally. Then simply divide the running total by the window size to get the current average whenever needed.

Rectify

This is a very simple concept which is to change the sign of the values to all positive or all negative so they are additive. In this case, I used rectification to change all values to positive. As with rectification, you can use a full wave or half wave method. You can easily do full wave by using the abs() math function that returns the value as positive. You can square values to turn them positive, but you are changing the amplitude. A simple rectify can turn them positive without any other effects. To perform half wave rectification, you can just set any value less than zero to zero.

Compression

In the DSP world Compression is typically defined as compressing the amplitudes to keep them in a close range. My compression technique here is to sum up the values in a window of samples. This is a form of down-sampling as you only get one sample out each time the window is filled, but no values are being thrown away. It’s a pure total of the window, or optionally an average of the window. This was employed in a few of the algorithms to try to identify bursts of acceleration from quieter times. I didn’t actually use it in the final algorithm.

FIR Filter

Finite Impulse Response (FIR) is a digital filter that is implemented via a number of taps, each with its assigned polynomial coefficient. The number of taps is known as the filter’s order. One strength of the FIR is that it does not use any feedback, so any rounding errors are not cumulative and will not grow larger over time. A finite impulse response simply means that if you input a stream of samples that consisted of a one followed by all zeros, the output of the filter would go to zero within at most the order +1 amount of 0 value samples being fed in. So, the response to that single sample of one lives for a finite amount of samples and is gone. This is essentially achieved by the fact there isn’t any feedback employed. I’ve seen DSP articles claim calculating filter tap size and coefficients is simple, but not to me. I ended up finding an online app called tFilter that saved me a lot of time and aggravation. You pick the type of filter (low, high, bandpass, bandstop, etc) and then setup your frequency ranges and sampling frequency of your input data. You can even pick your coefficients to be produced in fixed point to avoid using floating point math. If you’re not sure how to use fixed point or never heard of it, I’ll talk about that in the Embedded Optimization Techniques section.

基于嵌入式目标运行机进行优化

Magnitude Squared

Mag Square is a technique that can save computing power of calculating square roots. For example, if you want to calculate the vector for X and Z axis, normally you would do the following: val = sqr((X * X) + (Y * Y)). However, you can simply leave the value in (X * X) + (Y * Y), unless you really need the exact vector value, the Mag Square gives you a usable ratio compared to other vectors calculated on subsequent samples. The numbers will be much larger, and you may want to use attenuation to make them smaller to avoid overflow from additional computation downstream.

I used this technique in the final algorithm to help accentuate the bursts of acceleration from the background vibrations. I only used Z * Z in my calculation, but I then attenuated all the values by half or -6dB to bring them back down to reasonable levels for further processing. For example, after removing the bias if I had some samples around 2 and then some around 10, when I squared those values I now have 4 and 100, a 25 to 1 ratio. Now, if I attenuate by .5, I have 2 and 50, still a 25 to 1 ratio but now with smaller numbers to work with.

Fixed Point

Using fixed point numbers is another way to stretch performance, especially on microcontrollers. Fixed point is basically integer math, but it can keep precision via an implied fixed decimal point at a particular bit position in all integers. In the case of my FIR filter, I instructed tFilter to generate polynomial values in 16-bit fixed point values. My motivation for this was to ensure I don’t use more than 32-bit integers, which would especially hurt performance on an 8-bit microcontroller.

Rather than go into the FIR filter code to explain how fixed point works, let me first use a simple example. While the FIR filter algorithm does complex filtering with many polynomials, we could implement a simple filter that outputs the same input signal but -6dB down or half its amplitude. In floating point terms, this would be a simple one tap filter to multiply each incoming sample by 0.5. To do this in fixed point with 16 bit precision, we would need to convert 0.5 into its 16-bit fixed point representation. A value of 1.0 is represented by 1 * (216) or 65,536. Anything less than 65536 is a value less than 1. To create a fixed point integer of 0.5, we simply use the same formula 0.5 * (216), which equals 32,768. Now we can use that value to lower the amplitude by .5 of every sample input. For example, say we input into our simple filter a sample with the value of 10. The filter would calculate 10 * 32768 = 327,680, which is the fixed point representation. If we no longer care about preserving the precision after the calculations are performed, it can easily be turned back into a non-fixed point integer by simply right shifting by the number of bits of precision being used. Thus, 327680 >> 16 = 5. As you can see, our filter changed 10 into 5 which of course is the one half or -6dB we wanted out. I know 0.5 was pretty simple, but if you had wanted 1/8 the amplitude, the same process would be used, 65536 * .125 = 8192. If we input a sample of 16, then 16 * 8192 = 131072, now change it back to an integer 131072 >> 16 = 2. Just to demonstrate how you lose the precision when turning back to integer (the same as going float to integer) if we input 10 into the 1/8th filter it would yield the following, 10 * 8192 = 81920 and then turning it back to integer would be 81920 >> 16 = 1, notice it was 1.25 in fixed point representation.

Getting back to the FIR filters, I picked 16 bits of precision, so I could have a fair amount of precision but balanced with a reasonable amount of whole numbers. Normally, a signed 32-bit integer can have a range of - 2,147,483,648 to +2,147,483,647, however there now are only 16 bits of whole numbers allowed which is a range of -32,768 to +32,767. Since you are now limited in the range of numbers you can use, you need to be cognizant of the values being fed in. If you look at the FEPFilter_get function, you will see there is an accumulator variable accZ which sums the values from each of the taps. Usually if your tap history values are 32 bit, you make your accumulator 64-bit to be sure you can hold the sum of all tap values. However, you can use a 32 bit value if you ensure that your input values are all less than some maximum. One way to calculate your maximum input value is to sum up the absolute values of the coefficients and divide by the maximum integer portion of the fixed point scheme. In the case of the FEP FIR filter, the sum of coefficients was 131646, so if the numbers can be 15 bits of positive whole numbers + 16 bits of fractional numbers, I can use the formula (231)/131646 which gives the FEP maximum input value of + or - 16,312. In this case, another optimization can be realized which is not to have a microcontroller do 64-bit calculations.

Walking the Signal Processing Chain

Delays Due to Filtering

Before walking through the processing chain, we should discuss delays caused by filtering. Many types of filtering add delays to the signal being processed. If you do a lot of filtering work, you are probably well aware of this fact, but, if you are not all that experienced with filtering signals, it’s something of which you should be aware. What do I mean by delay? This simply means that if I put in a value X and I get out a value Y, how long it takes for the most impact of X to show up in Y is the delay. In the case of a FIR filter, it can be easily seen by the filter’s Impulse response plot, which, if you remember from my description of FIR filters, is a stream of 0’s with a single 1 inserted. T-Filter shows the impulse response, so you can see how X impacts Y’s output. Below is an image of the FEP’s high pass filter Impulse Response taken from the T-Filter website. Notice in the image that the maximum impact on X is exactly in the middle, and there is a point for each tap in the filter.

Below is a diagram of a few of the FEP’s high pass filter signals. The red signal is the input from the accelerometer or the newest sample going into the filter, the blue signal is the oldest sample in the filter’s ring buffer. There are 19 taps in the FIR filter so they represent a plot of the first and last samples in the filter window. The green signal is the value coming out of the high pass filter. So to relate to my X and Y analogy above, the red signal is X and the green signal is Y. The blue signal is delayed by 36 milliseconds in relation to the red input signal which is exactly 18 samples at 2 milliseconds, this is the window of data that the filter works on and is the Finite amount of time X affects Y.

Notice the output of the high pass filter (green signal) seems to track changes from the input at a delay of 18 milliseconds, which is 9 samples at 2 milliseconds each. So, the most impact from the input signal is seen in the middle of the filter window, which also coincides with the Impulse Response plot where the strongest effects of the 1 value input are seen at the center of the filter window.

It’s not only a FIR that adds delay. Usually, any filtering that is done on a window of samples will cause a delay, and, typically, it will be half the window length. Depending on your application, this delay may or may not have to be accounted for in your design. However, if you want to line this signal up with another unfiltered or less filtered signal, you are going to have to account for it and align it with the use of a delay component.

Front End Processor

I’ve talked at length about how to get to a final solution and all the components that made up the solution, so now let’s walk through the processing chain and see how the signal is transformed into one that reveals the punches. The FEP’s main goal is to remove bias and create an output signal that smears across the bursts of acceleration to create a wave that is higher in amplitude during increased acceleration and lower amplitude during times of less acceleration. There are four serial components to the FEP: a High Pass FIR, Attenuator, Rectifier and Smoothing via Sliding Window Average.

The first image is the input and output of the High Pass FIR. Since they are offset by the amount of bias, they don’t overlay very much. The red signal is the input from the accelerometer, and the blue is the output from the FIR. Notice the 1g of acceleration due to gravity is removed and slower changes in the signal are filtered out. If you look between 24,750 and 25,000 milliseconds, you can see the blue signal is more like a straight line with spikes and a slight ringing on it, while the original input has those spikes but meandering on some slow ripple.

Next is the output of the attenuator. While this component works on the entire signal, it lowers the peak values of the signal, but its most important job is to squish the quieter parts of the signal closer to zero values. The image below shows the output of the attenuator, and the input was the output of the High Pass FIR. As expected, peaks are much lower but so is the quieter time. This makes it a little easier to see the acceleration bursts.

Next is the rectifier component. Its job is to turn all the acceleration energy in the positive direction so that it can be used in averaging. For example, an acceleration causing a positive spike of 1000 followed by a negative spike of 990 would yield an average of 5, while a 1000 followed by a positive of 990 would yield an average of 995, a huge difference. Below is an image of the Rectifier output. The bursts of acceleration are slightly more visually apparent, but not easily discernable. In fact, this image shows exactly why this problem is such a tough one to solve; you can clearly see how resonant shaking of the base causes the pattern to change during punch energy being added. The left side is lower and more frequent peaks, the right side has higher but less frequent peaks.

The 49 value sliding window is the final step in the FEP. While we have done subtle changes to the signal that haven’t exactly made the punches jump out in the images, this final stage makes it visually apparent that the signal is well on its way of yielding the hidden punch information. The fruits of the previous signal processing magically show up at this stage. Below is an image of the Sliding Window average. The blue signal is its input or the output of the Rectifier, and the red signal is the output of the sliding window. The red signal is also the final output of the FEP stage of processing. Since it is a window, it has a delay associated with it. Its approximately 22 samples or 44 milliseconds on average. It doesn’t always look that way because sometimes the input signal spikes are suddenly tall with smaller ringing afterwards. Other times there are some small spikes leading up to the tall spikes and that makes the sliding window average output appear inconsistent in its delay based on where the peak of the output shows up. Although these bumps are small, they are now representing where new acceleration energy is being introduced due to punches.

Detection Processor

Now it’s time to move on to the Detection Processor (DET). The FEP outputs a signal that is starting to show where the bursts of acceleration are occurring. The DET’s job will be to enhance this signal and employ an algorithm to detect where the punches are occurring.

The first stage of the DET is an attenuator. Eventually, I want to add exponential gain to the signal to really pull up the peaks, but, before doing that, it is important to once again squish down the lower values towards zero and lower the peaks to keep from generating values too large to process in the rest of the DET chain. Below is an image of the output from the attenuator stage, it looks just like the signal output from the FEP, however notice the signal level peaks were above 100 from the FEP, and now peaks are barely over 50. The vertical scale is zoomed in with the max amplitude set to 500 so you can see that there is a viable signal with punch information.

With the signal sufficiently attenuated, it’s time to create the magic. The Magnitude Square function is where it all comes together. The attenuated signal carries the tiny seeds from which I’ll grow towering Redwoods. Below is an image of the Mag Square output, the red signal is the attenuated input, and the blue signal is the mag square output. I’ve had to zoom out to a 3,000 max vertical, and, as you can see, the input signal almost looks flat, yet the mag square was able to pull out unmistakable peaks that will aid the detection algorithm to pick out punches. You might ask why not just use these giant peaks to detect punches. One of the reasons I’ve picked this area of the signal to analyze is to show you how the amount of acceleration can vary greatly as you can see the peak between 25,000 and 25,250 is much smaller than the surrounding peaks, which makes pure thresholding a tough chore.

Next, I decided to put a Low Pass filter to try to remove any fast changing parts of the signal since I’m looking for events that occur in the 2 to 4 Hz range. It was tough on T-Filter to create a tight low pass filter with a 0 to 5 Hz band pass as it was generating filters with over 100 taps, and I didn’t want to take that processing hit, not to mention I would then need a 64-bit accumulator to hold the sum. I relaxed the band pass with a 0 to 19 Hz range and the band stop at 100 to 250 Hz. Below is an image of the low pass filter output. The blue signal is the input, and the red signal is the delayed output. I used this image because it allows the input and output signal to be seen without interfering with each other. The delay is due to 6 sample delay of the low pass FIR, but I have also introduced a 49 sample delay to this signal so that it is aligned in the center of the 99 sample sliding window average that follows in the processing chain. So it is delayed by a total of 55 samples or 110 milliseconds. In this image, you can see the slight amplification of the slow peaks by their height and how it is smoothed as the faster changing elements are attenuated. Not a lot going on here but the signal is a little cleaner, Earl Muntz might suggest I cut the low pass filter out of the circuit, and it might very well work without it.

The final stage of the signal processing is a 99 sample sliding window average. I built into the sliding window average the ability to return the sample in the middle of the window each time a new value is added and that is how I produced the 49 sample delayed signal in the previous image. This is important because the detection algorithm is going to have 2 parallel signals passed into it, the output of the 99 sliding window average and the 49 sample delayed input into the sliding window average. This will perfectly align the un-averaged signal in the middle of the sliding window average. The averaged signal is used as a dynamic threshold for the detection algorithm to use in its detection processing. Here, once again, is the image of the final output from the DET.

In the image, the green and yellow signals are inputs to the detection algorithm, and the blue and red are outputs. As you can see, the green signal, which is a 49 samples delayed, is aligned perfectly with the yellow 99 sliding window average peaks. The detection algorithm monitors the crossing of the yellow by the green signal. This is accomplished by both maximum and minimum start guard state that verifies the signal has moved enough in the minimum or maximum direction in relation to the yellow signal and then switches to a state that monitors the green signal for enough change in direction to declare a maximum or minimum. When the peak start occurs and it’s been at least 260ms since the last detected peak, the state switches to monitor for a new peak in the green signal and also makes the blue spike seen in the image. This is when a punch count is registered. Once a new peak has been detected, the state changes to look for the start of a new minimum. Now, if the green signal falls below the yellow by a delta of 50, the state changes to look for a new minimum of the green signal. Once the green signal minimum is declared, the state changes to start looking for the start of a new peak of the green signal, and a red spike is shown on the image when this occurs.

Again, I’ve picked this time in the recorded data because it shows how the algorithm can track the punches even during big swings in peak amplitude. What’s interesting here is if you look between the 24,750 and 25,000 time frame, you can see the red spike detected a minimum due to the little spike upward of the green signal, which means the state machine started to look for the next start of peak at that point. However, the green signal never crossed the yellow line, so the start of peak state rode the signal all the way down to the floor and waited until the cross of the yellow line just before the 25,250 mark to declare the next start of peak. Additionally, the peak at the 25,250 mark is much lower than the surrounding peaks, but it was still easily detected. Thus, the dynamic thresholding and the state machine logic allows the speed bag punch detector algorithm to “Roll with the Punches”, so to speak.

最后感言

总而言之，我们在本文中讨论了很多这次技术众包背后的细节。首先，充分分析问题并确定需求的重要性，因为它涉及到所需项目的努力方向以及最终顺利完成它所需的各领域知识。第二，对于这种性质的问题，创建验证环境来构建算法是势在必行的，并且在这种情况下，它是具有基于Java技术搭建的数据可视化处理的原型。第三，是在最终的目标机器和环境中实现，在一台PC上你有优秀的经过优化的编译器和强大的CPU及缓存，但对于微控制器，针对目标机器的优化是留给你的工作。使用您知道的每个优化技巧，以保证微控制器能尽可能快地处理数据。第四，迭代开发可以帮助你解决这样的问题：在尝试中摸着石头过河，边学边做边验证。

当我回顾这个众包项目，让我来探讨成功的终极奥秘，我想主要是两点。为这项工作创造正确的工具是无价的。能够直观看到算法作用后的信号结果的能力是无价的。不仅绘制输出信号，而且它实时绘制，使我完全理解生成的加速度信息背后的因素。就好像Nate在练习拳击，而我正在看着我的屏幕上的波形信息。然而，根本的因素是，我意识到这是一个每秒发生2到4次的事情。我坚定信念并不懈地追求如何将原始的输入信号转换成显示这些事件的东西。我并没有完全指望让Google找到答案。请谨记：知识不是来自书中，它只是被记录在书中。首先，总要有人抛下书本去发现一些新东西，然后它才成为知识。使用你已掌握或者可以获取的知识，但不要害怕使用你的想象力来尝试之前没有尝试过的未解问题。所以记住，比方说，当你来到铺砌的道路的尽头，你是转过头找一条别人已经铺好的路，还是说在岔路口，继续前进，发现属于自己的新路。我不能只是谷歌如何用加速度计记录拳击速度沙袋被击打次数，但现在你们就可以了。

原始文章采用CC BY-SA 4.0，您可以自由地：

• 分享 — 在任何媒介以任何形式复制、发行本作品
• 演绎 — 修改、转换或以本作品为基础进行创作
• 在任何用途下，甚至商业目的。

• 只要你遵守许可协议条款，许可人就无法收回你的这些权利。

本文由翻译美国开源硬件厂商Sparkfun（火花快乐）的相关教程翻译，原始教程采用同样的CC BY-SA 4.0协议，为便于理解和方便读者学习使用，部分内容为适应国内使用场景稍有删改或整合，这些行为都是协议允许并鼓励的。

原始文章及相关素材链接：

https://www.sparkfun.com/news/2115?_ga=1.105187979.946766378.1445226389#requirements

https://learn.sparkfun.com/tutorials/lessons-in-algorithms