Industrial IoT has created a bottleneck that threatens to overwhelm data analysis tools. Part one of “Industrial IoT’s Smart Shift to the Edge” examines how equipping smart sensors at the edge of the IoT makes sensing, measuring and analyzing data more manageable and leads to more timely, valuable insights.
Reduced time to insights at the edge node allow critical decisions to be made in near real-time. With theoretically unlimited processing power and communications data, the full bandwidth from all edge node sensed information could be sent to a distant compute station in the cloud. Then, vast computations analyze the data to mine valuable details that can be leveraged to make better informed decisions. However, the limitations of battery power, communications bandwidth, and compute cycle intensive algorithms had previously made this scenario a concept rather than a practical implementation.
The pioneer days of the industrial IoT and its precursor, machine-to-machine (M2M) communication, were largely defined by the role of cloud platforms as the primary application enablers. Intelligent systems have historically relied only upon cloud level capability for their insight. The actual edge sensor devices had been relatively unsophisticated. However, this old premise is currently being shaken up as low power computing capabilities at the edge node advance at a faster rate than those at the cloud1. Edge nodes are now a realistic option and offer the capability to sense, measure, interpret and connect.
There is a smart partition paradigm shift underway from the connected sensor model to the intelligent device model. This is providing more available architecture choices and allowing organizations deploying the industrial IoT to enhance their physical assets and processes in unique ways. Edge computing analytics, also known as edge intelligence or interpretation, is driving this shift. Mass industrial IoT deployments rely on the availability of a diverse set of intelligent nodes that are secure, highly energy efficient and easy to manage.
The highest quality sensed data can still be marginalized without careful attention to an application’s requirements within edge node analytics. Edge sensor devices may be constrained by energy, bandwidth, or raw computational power. These constraints propagate to protocol choices that can cut IP stacks down to minimal flash memory or RAM. This can make it challenging to program and there can be some sacrifice of the IP benefit.
Edge processing can be an analytic proposition as an approach to analyze data close to its source in addition to sending it to a remote server for cloud-level analysis. Moving the real-time analytics edge processing as early as possible in the signal chain reduces the payload burden down-stream and shortens latency. With the initial data processing done at the edge node, the required data formatting, communications bandwidth, and eventual aggregation at the gateway to the cloud is simplified. Time sensitive feedback loops through close coupling to the sensor provide immediate processing that provides valuable insights for informed decision-making2.
However, this requires advance intelligence about what specific information is valuable in the sensed and measured data. It may also vary from edge node to edge node due to spatial separation or application differences. Event alerts, triggers and interrupt detection can ignore a majority of the data and transmit only what is necessary.
The time value of money states that a dollar today is worth more than a dollar tomorrow. Analogously, there is a time constant for data. The time value of data means that the data you have sensed in this fractional second will not mean as much in the future – even an hour from now. Excellent mission critical IoT examples of this are heat surge sensing, gas leak detection or sensing catastrophic machinery failure that requires immediate action. Time sensitive data value decay starts at the point of interpretation. The longer the latency to effectively interpret your data and take action, the resulting decision will be less valuable. In order to solve the temporal depreciation riddle within the industrial IoT, insights must gain insights earlier in the signal chain.
Processing algorithms within the edge sensor node can be used to filter, decimate, tune, and refine the sampled data down to the minimum required subset. This requires first defining the narrow data of interest. Adjustable bandwidth, sample rate and dynamic range help establish this baseline in the analog domain of the hardware at the onset. By using the required analog settings, the sensor will target only necessary information and provide a shorter time constant to quality interpreted data.
Digital post processing at the edge can refine the relevant data further. Frequency analysis of the data at the edge sensor can make early decisions about signal content before the information leaves the node. Performing fast Fourier transforms (FFTs), finite impulse response (FIR) filtering, and using intelligent decimation are some high order computational blocks that narrow the scope of the sampled data. In some cases, only an incremental breadcrumb of pass or fail information is needed to be transmitted out of the edge sensor node after dramatically reducing the full bandwidth of data.
Figure 1 illustrates that without a front-end analog filter or a digital post-processing filter, a simple signal with decimation by 8 (left) will alias new unwanted signals (center) to frequency fold into the new desired signal band (right). Digital post processing, with a digital signal processor (DSP) or microcontroller unit (MCU), using a half-band FIR low-pass filter as a companion to decimation, will help prevent this issue by filtering the interfering aliased signals.
Processing for Edge Node Insights—Smart Factory
A leading industrial IoT application is a solution for factory machine condition monitoring. The solution identifies and predicts machine performance issues in advance of failure. At the edge sensor node, a multi-axis high dynamic range accelerometer monitors vibration displacement at various locations on industrial machines. Raw data is filtered and decimated for frequency domain interpretation within a microcontroller unit. An FFT compared against known performance limits can be processed for testing against pass, fail and warning alerts downstream. Processing gain within the FFT can be achieved through FIR filtering to remove wideband noise that is otherwise outside the bandwidth of interest.
The edge node processing is an important component in machine condition monitoring. The full bandwidth of sampled data creates a significant bottleneck for the aggregation at the wireless gateway. A single machine may have many sensors, and hundreds of machines may be simultaneously monitored. The filtering and intelligent decision making within the microcontroller unit offers a low bandwidth output to the wireless transceiver eliminating intensive filter processing at the cloud.
Figure 2 shows a signal chain for machine condition monitoring where an accelerometer sensor measures a displacement vibration signature. With post processing at the edge sensor node, frequency analysis can be done within a narrow bandwidth of interest by filtering and decimating the sampled data ahead of FFT computation.
During FFT computation, similar to a real-time oscilloscope, the processing can be blind to new time domain activity until the FFT is complete. An alternate time domain path in a second thread may prevent gaps in the data analysis.
If mechanical signature frequencies of interest are known precisely, the sample rate of the ADC and FFT size within the microcontroller unit can be planned such that the maximum amount of energy falls within the width of a single histogram bin. This will prevent the signal power from leaking across multiple bins, diluting the precision of the amplitude measurement.
Figure 3 provides an example of an FFT where specific predetermined zones are interpreted within the edge node MCU for more than one observed mechanical component. Bin energy that peaks within the required green zone represents satisfactory operation, while the yellow and red zones indicate warning and critical alarms respectively. Instead of transmitting the full sensor bandwidth, a lower data rate alarm or trigger breadcrumb can alert the system of an excursion event within the zones of interest.
In the past, concepts like the ones described above were restricted from implementations due to limitations with computational horsepower, processing, security, power management and sensor technology. However, advances in these disciplines have built the fundamental framework for industrial IoT and made sensing, measuring, interpreting data at the edge node possible with additional consideration for secure, energy efficient ultralow power processing. Part two of this article will take a closer look at these technologies and how they make the industrial IoT shift to the edge possible.
2 Grainne Murphy and Colm Prendergast. “Precision Counts in IoT.” Analog Devices, Inc., August 2016.
About the author
Ian Beavers is a product engineering manager for the Automation Energy and Sensors team located at Analog Devices, Greensboro, NC. He has worked for the company since 1999. Ian has over 19 years of experience in the semiconductor industry. Ian earned a bachelor’s degree in electrical engineering from North Carolina State University and an MBA from the University of North Carolina at Greensboro.