## A-weighting equalisation: Designing and deploying to Arm Cortex-M devices

Modern embedded processors, software frameworks and design tooling now allow engineers to apply advanced measurement concepts to smart factories as part of the I4.0 revolution.

In recent years, PM (predictive maintenance) of machines has received great attention, as factories look to maximise their production efficiency while at the same time retaining the invaluable skills of experienced foremen and production workers.

Traditionally, a foreman would walk around the shop floor and listen to the sounds a machine would make to get an idea of impending failure. With the advent of I4.0 technology, microphones, edge DSP algorithms and ML may now be employed in order to ‘listen’ to the sounds a machine makes and then make a classification and prediction.

One of the major challenges is how to make a computer hear like a human.

## Physics of the human ear

An illustration of the human ear shown on the right. As seen, the basic task of the ear is to translate sound (air vibration) into electrical nerve impulses for the brain to interpret.

### How does it work?

The ear achieves this via three bones (Stapes, Incus and Malleus) that act as a mechanical amplifier for vibrations received at the eardrum. These amplified sounds are then passed onto the Cochlea via the Oval window (not shown). The Cochlea (shown in purple) is filled with a fluid that moves in response to the vibrations from the oval window. As the fluid moves, thousands of nerve endings are set into motion. These nerve endings transform sound vibrations into electrical impulses that travel along the auditory nerve fibres to the brain for analysis.

### Modelling perceived sound

Due to complexity of the fluidic mechanical construction of the human auditory system, low and high frequencies are typically not discernible. Researchers over the years have found that humans are most perceptive to sounds in the 1-6kHz range, although this range varies according to the subject’s physical health.

This research led to the definition of a set of weighting curves: the so-called A, B, C and D weighting curves, which equalises a microphone’s frequency response. These weighting curves aim to bring the digital and physical worlds closer together by allowing a computerised microphone-based system to hear like a human.

The A-weighing curve is the most widely used as it is mandated by IEC-61672 to be fitted to all sound level meters. The B and D curves are hardly ever used, but C-weighting may be used for testing the impact of noise in telecoms systems.

The frequency response of the A-weighting curve is shown above, where it can be seen that sounds entering our ears are de-emphasised below 500Hz and are most perceptible between 0.5-6kHz. Notice that the curve is unspecified above 20kHz, as this exceeds the human hearing range.

## ASN FilterScript

ASN’s FilterScript symbolic math scripting language offers designers the ability to take an analog filter transfer function and transform it to its digital equivalent with just a few lines of code.

The analog transfer functions of the A and C-weighting curves are given below:

$$H_A(s) \approx \displaystyle{7.39705×10^9 \cdot s^4 \over (s + 129.4)^2\quad(s + 676.7)\quad (s + 4636)\quad (s + 76655)^2}$$

$$H_C(s) \approx \displaystyle{5.91797×10^9 \cdot s^2\over(s + 129.4)^2\quad (s + 76655)^2}$$

These analog transfer functions may be transformed into their digital equivalents via the bilinear() function. However, notice that $$H_A(s)$$ requires a significant amount of algebracic manipulation in order to extract the denominator cofficients in powers of $$s$$.

### Convolution

A simple trick to perform polynomial multiplication is to use linear convolution, which is the same algebraic operation as multiplying two polynomials together. This may be easily performed via Filterscript’s conv() function, as follows:

y=conv(a,b);


As a simple example, the multiplication of $$(s^2+2s+10)$$ with $$(s+5)$$, would be defined as the following three lines of FilterScript code:

a={1,2,10};
b={1,5};
y=conv(a,b);


which yields, 1 7 20 50  or $$(s^3+7s^2+20s+50)$$

For the A-weighting curve Laplace transfer function, the complete FilterScript code is given below:

ClearH1;  // clear primary filter from cascade

Main() // main loop

a={1, 129.4};
b={1, 676.7};
c={1, 4636};
d={1, 76655};

aa=conv(a,a); // polynomial multiplication
dd=conv(d,d);

aab=conv(aa,b);
aabc=conv(aab,c);

Na=conv(aabc,dd);
Nb = {0 ,0 , 1 ,0 ,0 , 0, 0}; // define numerator coefficients
G = 7.397e+09; // define gain

Ha = analogtf(Nb, Na, G, "symbolic");
Hd = bilinear(Ha,0, "symbolic");

Num = getnum(Hd);
Den = getden(Hd);
Gain = getgain(Hd)/computegain(Hd,1e3); // set gain to 0dB@1kHz



Frequency response of analog vs digital A-weighting filter for $$f_s=48kHz$$. As seen, the digital equivalent magnitude response matches the ideal analog magnitude response very closely until $$6kHz$$.

### The ITU-R 486–4 weighting curve

Another weighting curve of interest is the ITU-R 486–4 weighting curve, developed by the BBC. Unlike the A-weighting filter, the ITU-R 468–4 curve describes subjective loudness for broadband stimuli. The main disadvantage of the A-weighting curve is that it underestimates the loudness judgement of real-world stimuli particularly in the frequency band from about 1–9 kHz.

Due to the precise definition of the 486–4 weighting curve, there is no analog transfer function available. Instead the standard provides a table of amplitudes and frequencies – see here. This specification may be directly entered into Filterscript’s firarb() function for designing a suitable FIR filter, as shown below:

ClearH1;  // clear primary filter from cascade
ShowH2DM;

interface L = {10,400,10,250}; // filter order

Main()

// ITU-R 468 Weighting
A={-29.9,-23.9,-19.8,-13.8,-7.8,-1.9,0,5.6,9,10.5,11.7,12.2,12,11.4,10.1,8.1,0,-5.3,-11.7,-22.2};
F={63,100,200,400,800,1e3,2e3,3.15e3,4e3,5e3,6.3e3,7.1e3,8e3,9e3,1e4,1.25e4,1.4e4,1.6e4,2e4};

A={-30,A};  //  specify arb response
F={0,F,fs/2};   //

Hd=firarb(L,A,F,"blackman","numeric");

Num=getnum(Hd);
Den={1};
Gain=getgain(Hd);



Frequency response of an ITU-R 468-4 FIR filter designed with Filterscript’s firarb() function  for $$f_s=48kHz$$

As seen, FilterScript provides the designer with a very powerful symbolic scripting language for designing weighting curve filters. The following discussion now focuses on deployment of the A-weighting filter to an Arm based processor via the tool’s automatic code generator. The concepts and steps demonstrated below are equally valid for FIR filters.

### Automatic code generation to Arm processor cores via CMSIS-DSP

The ASN Filter Designer’s automatic code generation engine facilitates the export of a designed filter to Cortex-M Arm based processors via the CMSIS-DSP software framework. The tool’s built-in analytics and help functions assist the designer in successfully configuring the design for deployment.

Before generating the code, the H2 filter (i.e. the filter designed in FilterScript) needs to be firstly re-optimised (transformed) to an H1 filter (main filter) structure for deployment. The options menu can be found under the P-Z tab in the main UI.

All floating point IIR filters designs must be based on Single Precision arithmetic and either a Direct Form I or Direct Form II Transposed filter structure. The Direct Form II Transposed structure is advocated for floating point implementation by virtue of its higher numerically accuracy.

Quantisation and filter structure settings can be found under the Q tab (as shown on the left). Setting Arithmetic to Single Precision and Structure to Direct Form II Transposed and clicking on the Apply button configures the IIR considered herein for the CMSIS-DSP software framework.

Select the Arm CMSIS-DSP framework from the selection box in the filter summary window:

The automatically generated C code based on the CMSIS-DSP framework for direct implementation on an Arm based Cortex-M processor is shown below:

As seen, the ASN Filter Designer’s automatic code generator generates all initialisation code, scaling and data structures needed to implement the A-weighting filter IIR filter via Arm’s CMSIS-DSP library.

## ASN Filter Designer supports all Arm Cortex-M demo-boards

Did you know that there are 23 billion IoT embedded devices currently deployed around the world? This figure is expected to grow to a whopping 1 trillion devices by 2050!

Less known, is that 80% of IoT devices are based around Arm’s Cortex-M microcontroller technology. Sometimes clients ask us if we support their Arm Cortex-M based demo-board of choice. The answer is simply: yes!

## 200+ IC vendors supported

The ASN Filter Designer has an automatic code generator for Arm Cortex-M cores, which means that we support virtually every Arm based demo-board: ST, Cypress, NXP, Analog Devices, TI, Microchip/Atmel and over 200+ other manufacturers. Our compatibility with Arm’s free CMSIS-DSP software framework removes the frustration of implementing complicated digital filters in your IoT application – leaving you with code that is optimal for Cortex-M devices and that works 100% of the time.

The Arm Cortex-M family of microcontrollers are an excellent match for IoT applications. Some of the advantages include:

• Low power and cost – essential for IoT devices
• Microcontroller with DSP functionality all-in-one
• Embedded hardware security functionality
• Cortex-M4 and M7 cores with hardware floating support (enhanced microcontrollers)
• Freely available CMSIS-DSP C library: supporting over 60 signal processing functions

## Automatic code generation for Arm’s CMSIS-DSP software framework

Simply load your sensor data into the ASN Filter Designer signal analyser and perform a detailed analysis. After identifying the wanted and unwanted components of your signal, design a filter and test the performance in real-time on your test data. Export the designed design to Arm MDK, C/C++ or integrate the filter into your algorithm in another domain, such as in Matlab, Python, Scilab or Labview.

Use the tool in your RAD (rapid application development) process, by taking advantage of the automatic code generation to Arm’s CMSIS-DSP software framework, and quickly integrate the DSP filter code into your main application code.

Let the tool analyse your design, and automatically generate fully compliant code for either the M0, M0+, M3, M4 and the newer M23 and M33 Cortex cores. Deploy your design within minutes rather than hours.

## Proud Arm knowledge partner

We are proud that we are an Arm knowledge partner! As an Arm DSP knowledge partner, we will be kept informed of their product roadmap and progress for the coming years.

Try it for yourself and see the benefits that the ASN Filter Designer can offer your organisation by cutting your development costs by up to 75%!

## The sensor measurement challenge

Sensors come in all type of shapes and forms…There are sensors for audio, pressure, temperature, weight, strain, light, humidity…the list is almost endless.

The challenge for most, is that many sensors used in these IoT measurement applications require filtering in order to improve the performance of the sensor’s measurement data in order to make it useful for analysis.

Before jumping into the disussion, let’s first have a look at what sensor data really is….

All sensors produce measurement data. These measurement data contain two types of components:

• Wanted components, i.e. information what we want to know
• Unwanted components, measurement noise, 50/60Hz powerline interference, glitches etc – what we don’t want to know

Unwanted components degrade system performance and need to be removed.

So, the challenge for every designer is first to identify what aspects of the data we want to keep, i.e. ‘the wanted components’ and what we need to filter out, the so called ‘unwanted components’. After establishing what need to be filtered out, the challenge then which domain do we tackle this problem in, i.e. the analog domain or in the digital domain ? Each domain has its pros and cons, as we will now discuss for a practical classic sensor measurement challenge using a loadcell.

A classic sensor measurement challenge using a loadcell is shown below.

Looking at the hardware setup, we see that have a loadcell excited by a DC excitation voltage, and the general idea is that the sensor’s differential bridge voltage is amplifier by the instrumentation amplifier (IA) when strain is applied.

For those of you unfamiliar with this type of technology, a loadcell is a strain measurement sensor that is comprised of 4 strain gauges, it’s also referred to as a Wheatstone bridge, hence the terminology bridge sensor.

Analysing the signals in the schematic, we see that the differential voltage is passed through 2 filters in order to remove powerline interference and reduce measurement noise.

### What are the challenges?

The Instrumentation amplifier (IA) has high impedance inputs, which makes it easy to connect EMI (electromagnet interference) filters to the inputs. However, any mismatches with these filters will generally degrade the instrumentation amplifier’s common-mode rejection ratio, which is undesirable.

The instrumentation amplifier usually has a large gain (100 is quite typical), so any unwanted differential voltage on the inputs will be amplified. Looking at the filters, the notch depth of the powerline cancellation (50Hz/60Hz) filter will be dependent on component tolerances, and will vary over time and with temperature…This is problematic as we’ll discuss in the following section.

Finally, any analog filter or filters will require careful PCB layout and eat up precious board space, which is undesirable for many modern devices.

## Loadcell digital – is digital any better ?

Replacing the instrumentation amplifier with a 24bit sigma-delta ADC (analog-to-digital converter), we simplify the circuitry – although many ADCs don’t tolerate high impedance at their inputs, which may be problematic for good RFI (radio frequency interference) filter design.

Nevertheless, some sigma-delta devices have an in-built 50/60Hz notch filter which simplifies the filtering requirement. Although these devices are more expensive, and the choice of sampling frequency is limited, they may be good enough for some applications.

## ASP vs DSP

So, which domain is best for solving our measurement challenge, i.e. do we use analog signal processing (ASP) or digital signal processing (DSP)? In order to answer this objectively, we need to first breakdown the pros and cons of each domain.

### Analog filters

Let’s first look at an implementation using ASP.

The most obvious advantage is that analog filters have excellent resolution, as there are no ‘number of bits’ to consider. Analog filters have good EMC properties as there is no clock generating noise. There are no effects of aliasing, which is certainly true for the simpler op-amps, which don’t have any fancy chopping or auto-calibration circuitry built into them, and analog designs can be cheap which is great for cost sensitive applications.

### Sound great, but what’s the bad news?

Analog filters have several significant disadvantages that affect filter performance, such as component aging, temperature drift and component tolerance. Also, good performance requires good analog design skills and good PCB layout, which is hard to find in the contemporary skills market.

One big minus point is that filter’s frequency response remains fixed, i.e. a Butterworth filter will always be a Butterworth filter – any changes the frequency response would require physically changing components on the PCB – not ideal!

### Digital filters

Let’s now look at an implementation using DSP.

The first impression is that a digital solution is more complicated, as seen above with the five building blocks. However, digital filters have high repeatability of characteristics, and as an example, let’s say that you want to manufacture 1000 measurement modules after optimising your filter design. With a digital solution you can be sure that the performance of your filter will be identical in all modules. This is certainly not the case with analog, as component tolerance, component aging and temperature drift mean that each module’s filter will have its own characteristics.

Digital filters are adaptive and flexible, we can design and implement a filter with any frequency response that we want, deploy it and then update the filter coefficients without changing anything on the PCB!

It’s also easy to design filters with linear phase and at very low sampling frequencies – two things that are tricky with analog.

### Sound great, but what’s the bad news?

The effect of aliasing and if designing in fixed point, finite word length issues must be taken into account, including the limitation of the ADC and DAC. As there is clock source, digital designs will produce more EMI than analog filters.

## Conclusion

When designing modern IoT sensor measurement applications, digital filters offer a greater degree of design flexibility and high repeatability of characteristics over their analog counterparts.

With the advent of modern processor technology and design tooling, it is estimated that about 80% of IoT smart sensor devices are currently deployed using digital devices, such as Arm’s Cortex-M family. The Arm Cortex-M4 is a very popular choice with hundreds of silicon vendors, as it offers DSP functionality traditionally found in more expensive DSPs. Implementation is further simplified by virtue of ASN’s strong partnership with Arm who together provide a rich offering of easy to use filter design tooling and a free DSP software framework (CMSIS-DSP). These tools and well documented software framework allow you to get your IoT application up and running within minutes.

## Internet of Things (IoT) smart sensors: bridging the technology gap….

The internet of things (IoT) has gained tremendous popularity over the last few years, as many organisations strive to add IoT smart sensor technologies to their product portfolios. The basic paradigm centres around connecting everything to everything, and exchanging all data. This could be house hold appliances to more blue sky applications, such as smart cities. But what does this particularly mean for you?

Almost all IoT applications involve the use of sensors. But how do SME and even multi-national organisations transform their legacy product offering into a 21st century IoT application? One the first challenges that many organisations face is how to migrate to an IoT application while balancing design time, time to market, budget and risk.

We recently completed a project for a client who manufactured their own sensors, but wanted to improve their sensor measurement accuracy from ±10% to better than ±0.5% without going down the road of a massive re-design project.

The question that they asked us was simply: “Is it possible to get high measurement accuracy performance from a signal that is corrupted with all kinds of interference components without a hardware re-design?”

Our answer: “Yes, but the winning recipe centres around knowing what architectural building blocks to use”.

Traditionally, many design bureaus will evaluate the sensor performance and try and improve the measurement accuracy performance by designing new hardware and adding a few standard basic filtering algorithms to the software. This sort of intuitive approach can lead to very high development costs for only a modest increase in sensor performance. For many SMEs these costs can’t be justified, but perhaps there’s a better way?

## Algorithms: the winning recipe

Algorithms and mathematics are usually regarded by many organisations as ‘academic black magic’ and are generally overlooked as a solution for a robust IoT commercial application. As a consequence very few organisations actually take the time to analytically analyse a sensor measurement problem, and those who do invent something tend to come up with something that’s only useable in the lab. There has been a trend over the years to turn to Universities or research institutes, but once again the results are generally too  academic and are based more on getting journal publications, rather than a robust solution suitable for the market.

Our experience has been that the winning recipe centres around the balance of knowing what architectural blocks to use, and having the experience to assess what components to filter out and what components to enhance.  In some cases, this may even involve some minor modifications to the hardware in order to simplify the algorithmic solution. Unfortunately, due to the lack of investment in commercially experienced, academically strong (Masters, PhD) algorithm developers and the pressure of getting a project to the finish line, many solutions (even from reputable multi-national organisations) that we’ve seen over the years only result in a moderate increase in performance.

Despite the plethora of commercially available data analysis software, many organisations opt to do basic data analysis in Microsoft Excel, and tend to stay away from any detailed data analysis as it’s considered an unnecessary academic step that doesn’t really add any value.   This missed opportunity generally leads to problems in the future, where products need to be recalled for a ‘round of patchwork’ in order solve the so called ‘unforeseen problems’. A second disadvantage is that performance of the sensors may be only satisfactory, whereas a more detailed look may have yielded clues on how make the sensor performance good or in some cases even excellent.

## Algorithms can save the day!

“Although many organisations regard data analysis as a waste of money, our experience and customers prove otherwise.”

Investing in detailed data analysis at the beginning of a project usually results in some good clues as to what needs to be filtered out and what needs to be enhanced in order to achieve the desired performance.   In many cases,  these valuable clues allow  experienced algorithm developers to concoct a combination of signal processing building blocks without re-designing any hardware – which is very desirable for many organisations! Our experience has shown that this fundamental first step can cut project development costs by as much as 75%, while at the same time achieving the desired smart sensor measurement performance demanded by the market.

## So what does this all mean in the real world?

Returning the story of our customer, after undertaking a detailed data analysis of their sensor data, our developers were able design a suitable algorithm achieving a ±0.1% measurement accuracy from the original ±10% with only minor modifications to the hardware. This enabled the customer to present his IoT application at a trade show and go into production on time, and yes, we stayed within budget!

It’s estimated that the global smart sensor market will have over 50 billion smart devices in 2020. At least 80% of these IoT/IIoT smart sensors (temperature, pressure, gas, image, motion, loadcells) will use Arm’s Cortex-M technology – where the largest growth is in smart Image sensors (ADAS) & smart Temperature sensors (HVAC).

## IoT sensor measurement challenge

The challenge for most, is that many sensors used in these applications require a little bit of filtering in order to clean the measurement data in order to make it useful for analysis.

Let’s have a look at what sensor data really is…. All sensors produce measurement data. These measurement data contain two types of components:

• Wanted components, i.e. information what we want to know
• Unwanted components, measurement noise, 50/60Hz powerline interference, glitches etc – what we don’t want to know

Unwanted components degrade system performance and need to be removed.

## So, how do we do it?

DSP means Digital Signal Processing and is a mathematical recipe (algorithm) that can be applied to IoT sensor measurement data in order to clean it and make it useful for analysis.

But that’s not all! DSP algorithms can also help in analysing data, producing more accurate results for decision making with ML (machine learning). They can also improve overall system performance with existing hardware (no need to redesign your hardware – a massive cost saving!), and can reduce the data sent off to the cloud by pre-analysing data and only sending what is necessary.

Nevertheless, DSP has been considered by most to be a black art, limited only to those with a strong academic mathematical background. However, for many IoT/IIoT applications, DSP has been become a must in order to remain competitive and obtain high performance with relatively low cost hardware.

## Do you have an example?

Consider the following application for gas sensor measurement (see the figure below). The requirement is to determine the amplitude of the sinusoid in order to get an estimate of gas concentration (bigger amplitude, more gas concentration etc). Analysing the figure, it is seen that the sinusoid is corrupted with measurement noise (shown in blue), and any estimate based on the blue signal will have a high degree of uncertainty about it – which is not very useful if getting an accurate reading of gas concentration!

Algorithms clean the sensor data

After ‘cleaning’ the sinusoid (red line) with a DSP filtering algorithm, we obtain a much more accurate and usable signal which helps us in estimating the amplitude/gas concentration. Notice how easy it is to determine the amplitude of red line.

This is only a snippet of what is possible with DSP algorithms for IoT/IIoT applications, but it should give you a good idea as to the possibilities of DSP.

How do I use this in my IoT application?

As mentioned at the beginning of this article, 80% of IoT smart sensor devices are deployed on Arm’s Cortex-M technology. The Arm Cortex-M4 is a very popular choice with hundreds of silicon vendors, as it offers DSP functionality traditionally found in more expensive DSPs. Arm and its partners provide developers with easy to use tooling and a free software framework (CMSIS-DSP) in order to get you up and running within minutes.

## 2018: The death of big data for IoT

With the advent of smart cities, and society’s obsession of ‘being connected’, data networks have been overloaded with thousands of IoT sensors sending their data to the cloud, needing massive and very expensive computing resources to crunch the data.

### Is it really a problem?

The collection of all these smaller IoT data streams (from smart sensors), has ironically resulted in a big data challenge for IT infrastructures in the cloud which need to process

massive datasets – as such there is no more room for scalability. The situation is further complicated with the fact, that a majority of sensor data is coming from remote locations, which also presents a massive security risk.

It’s estimated that the global smart sensor market will have over 50 billion smart devices in 2020. At least 80% of these IoT/IIoT smart sensors (temperature, pressure, gas, image, motion, loadcells) will use Arm’s Cortex-M technology, but have little or no smart data reduction or security implemented.

### The current state of play

The modern IoT eco system problem is three-fold:

• Endpoint security
• Data reduction
• Data quality

Namely, how do we reduce our data that we send to the cloud, ensure that the data is genuine and how do ensure that our Endpoint (i.e. the IoT sensor) hasn’t been hacked?

### The cloud is not infallible!

Traditionally, many system designers have thrown the problem over to the cloud. Data is sent from IoT sensors via a data network (Wifi, Bluetooth, LoRa etc) and is then encrypted in the cloud. Extra services in the cloud then perform data analysis in order to extract useful data.

So, what’s the problem then?

This model doesn’t take into account invalid sensor data. A simple example of this, could be glue failing on a temperature sensor, such that it’s not bonded to the motor or casing that it’s monitoring. The sensor will still give out temperature data, but it’s not valid for the application.

As for data reduction – the current model is ok for a few sensors, but when the network grows (as is the case with smart cities), the solution becomes untenable, as the cloud is overloaded with data that it needs to process.

No endpoint security, i.e. the sensor could be hacked, and the hacker could send fake data to the cloud, which will then be encrypted and passed onto the ML (machine learning) algorithm as genuine data.

### What’s the solution?

Algorithms, algorithms….. and in built security blocks.

Over the last few years, hundreds of silicon vendors have been placing security IP blocks into their silicon together with a high performance Arm Cortex-M4 core. These so called enhanced micro-controllers offer designers a low cost and efficient solution for IoT systems for the foreseeable future.

A lot can be achieved by pre-filtering sensor data, checking it and only sending what is neccessary to the cloud. However, as with so many things, knowledge of security and algorithms are paramount for success.

## Some points to bear in mind when designing and implementing algorithms

I recently attended a seminar on advanced instrumentation, where algorithms were heavily featured. The project pitches heavily emphasised implementation rather than analysis and design, which started an interesting discussion, and led me to think about providing some hints that we’ve successfully used over the years:

1. What do we want to achieve? This is perhaps obvious, but I’ve seen that many people do over look this step and jump into Matlab or C in order to try something out. I would urge some caution here, and suggest that you think very carefully about what you’re about to undertake before writing a single line of code. Don’t be afraid to ask your colleagues/network for advice, as their suggestions may save you months of development time. Also consider using established techniques such as, MoSCoW.

2. The specifications: After establishing the ‘big picture’, split up the specifications into ‘must haves’ and ‘nice to haves’. This may take some time to work out, but undertaking this step saves a considerable amount of time in the development process, and keeps the client in the loop. The specifications don’t need to be 100% complete at this stage (they’re always minor details to be worked out), but make sure that you’re clear about what you’re about undertake, and don’t be afraid to do some analysis or short experiments if required.

3. Algorithm design: Sketch out the algorithm’s building blocks (Visio is a good tool), and for each idea produce a short list of bullets (pros and cons) and computational complexity. This will allow you easily review each concept with your peers.

4. Test data: arrange for some test vectors data (from clients or design some of your own synthetic signals), and sketch out a simple test plan of test vectors that you aim to use in order to validate your concept.

5. Development: Depending on your programming ability, you may decide to implement in C/C++, but Matlab/Octave are very good starting points, as the dynamic data types, vector math and toolboxes give you maximum flexibility. Use the testplan and vectors that you’ve designed in step 4. However, in the case of how to best design your algorithm for streaming applications, I would say that many aspects of the algorithm can be tested with an offline (data file) approach. For a majority of our radar and audio work, we always begin with data file comprised of 10-30seconds worth of data in order to prove that the algorithm functions as expected. Subsequent implementation steps can be used to make the algorithm streaming, but bear in mind that this may take a considerable amount of time!

6. Avoid a quick fix! Depending on the complexity of your algorithm, there will be certain testvectors that degrade the performance of your algorithm or even cause it to completely fail. Allocate sometime to investigate this behaviour, but remember to prioritise the importance, and don’t spend months looking for a minor bug. Try and avoid looking for a quick fix or a patch, as they generally re-appear in the future and kick you up the backside.

7. Implementation: after verifying that your concept is correct, you can finally consider target implementation. This step couples back to the previous steps, as the algorithm complexity will have direct influence on the implementation platform and development time. Some good questions to ask yourself: Is the target platform embedded? In which case, do I need an FPGA, DSP or microcontroller? Will it be fixed point or floating point? Perhaps it will be PC based, in which case is it for Windows, Linux or Mac or for a tablet? What tools do you need in order to develop and test the algorithm?

8. Validation: Verify that your implemented algorithm works with your test vectors and that look for any difficult cases that you can find – remembering point 6.

9. Documentation: In all of the aforementioned steps, documentation is essential. Make sure that you document your results, and provide a paper trail such that a colleague can continue with your work if you get hit by a bus.

About the author: Sanjeev Sarpal is director of algorithms and analytics at Advanced Solutions Nederland BV. He holds a PhD in signal processing and has over 20 years commercial experience with the design and deployment of algorithms for smart sensor applications.