The fourth industrial revolution, or Industry 4.0, brings digitalization and intelligence to the entire factory, promising to transform companies to operate smarter and more efficiently. Enterprises that can upgrade machinery with intelligence and automation and integrate them into a system-wide smart infrastructure are leading this revolution. But, according to a Deloitte survey, only 20.7 percent of manufacturing organizations consider themselves ready to engage the business models that Industry 4.0 brings.[i]
One of the biggest challenges manufacturing operations face in adopting Industry 4.0 is the cost and complexity to instrument existing machines, connect them to a network, and deploy relevant software. This is especially costly with legacy equipment that is not enabled with the necessary sensors, intelligence, or ability to communicate with plant infrastructure. For small to medium enterprises, Industry 4.0 can require a significant investment that might be difficult to justify. However, enabling Industrial IoT (IIoT) doesn’t have to be so complex and costly.
Bringing Smart Listening to the Factory
Intel® AI Builder member A2IoT (A squared IoT) is a company focused on applying state of the art machine learning (ML) and deep learning (DL) to explore data from manufacturing operations. Instead of adding sensors and intelligence to equipment, A2IoT listens to them. Machines and manufacturing processes make characteristic noises that are rich with information about their health and the quality of the process.
There is a nominal sound to optimally running plant operations at each work site. A good welder has tuned his or her hearing to the proper crackle of a metal inert gas (MIG) welding process, the machine operator to the smooth running of bearings, and the mechanic to the subtle rhythm of an engine. When the sounds change, they immediately detect something is not right. Analyzing sounds, though, is a complex operation—regardless of whether it is done by the human brain or computationally.
A2IoT uses these unique sound characteristics of machines along with visual cues and deep learning techniques to not only detect anomalies in real time, but also measure overall equipment effectiveness (OEE) and provide insight into manufacturing operations right from the site of measurement. They have developed ML algorithms and trained them on data sets that can quickly and accurately analyze and classify a wide range of machine sounds, such as welding, motors, drilling machines, hammering, and others.
For example, welding is a task where sound and imagery are good indicators of the process, but where ultimate quality inspections can only be done by destructive testing, which can be an expensive process. With A2IoT’s technologies, sound analytics is used to identify the health of a welding assembly during the process. It can inform the users about failures in real time. The benefits of an AI-driven, edge-computing device for analyzing welding operations are significant in terms of reducing both rework and product failures.
Non-invasive, Non-touch Intelligence at the Edge
The company’s initial product, called Equilips 4.0, is a small device with aural sensing, enough intelligence and computational power to accurately analyze what it hears, and wireless communications for system engineers and plant operators to receive insightful information about their operations. Equilips 4.0 is completely non-invasive and non-touch; all computing is done at the edge, eliminating the need for a network and external cloud computing services. With Equilips 4.0, A2IoT offers organizations a way to apply industrial IoT in their operations at a low total cost of deployment and ownership. The company makes it easy for factories to embrace Industry 4.0. Equilips 4.0 is now being used to accurately analyze the quality and efficiency of welding operations.
Accelerating Training by 5.6X[ii]
A2IoT began development of the Equilips 4.0 device on Raspberry Pi, coded their model in Python, and trained it using the TensorFlow framework with a custom data set for the AlexNet topology. When they engaged with the Intel AI Builders technical team, their baseline training took nearly four hours[ii] to train on a custom sound data set from welding operations.
The Intel team worked with A2IoT developers to significantly improve training of the welding data set by applying several optimization capabilities, which include:
- Using the Intel® Distribution of Python and Intel® Optimizations for TensorFlow
- Adding the Intel® Math Kernel Library for Deep Neural Networks (Intel® MKL-DNN)
- Trying different combinations of OMP_NUM_THREADS, KMP_BLOCKTIME, and KMP_AFFINITY to achieve maximum performance
These basic optimizations returned a 1.4X improvement[ii] in training time, a reduction of the original four hours on the welding data set to slightly over two and a half hours.
As the next step, developers ran their training workload using Horovod-based distributed training in TensorFlow on Intel® Xeon® Scalable processors in the Intel® AI DevCloud. The Intel AI DevCloud is a cloud-hosted hardware and software platform available to developers, researchers, and startups to learn, sandbox, and get started on their AI projects. Developers can experiment with Intel hardware, apply Intel-optimized frameworks, and use Intel software tools to build and tune their applications. Through this additional software optimization effort, the total training was accelerated by nearly 5.6X[ii] with negligible impact on accuracy (see chart).
With this improvement in performance, A2IoT can accelerate training iterations and eventually speed deployment of their solutions to the field for new customers with new data sets. The optimizations can help accelerate company initiatives to embrace Industry 4.0.
An Intel Cloud to Edge Optimization Process
Additionally, A2IoT continues to optimize their edge AI solution on Intel architecture. They are currently moving the Equilips 4.0 device to an Intel® Atom™ processor-based instance, and running the Intel® OpenVINO™ toolkit to help accelerate inferencing at the edge. The company is developing an integrated, optimized, edge-to-cloud design process that is powered by Intel technologies and architecture, and delivering improved machine learning performance. With these improvements in their edge-based analytics, A2IoT can bring faster insights around real-time overall equipment effectiveness and other data points to decision-makers, system engineers, and operators, which will help improve manufacturing processes while reducing errors and costs.
Intelligent Listening for More Use Cases
In addition to welding operations, some ongoing applications where Equilips technology is currently used include:
- Monitoring and prediction of tool life, tool breakage, and other quality issues in milling machines
- Predictive maintenance of electric motors
- Real time monitoring and anomaly detection in building construction
- Real time detection of air/gas leakage
For some interesting demos of their solution, visit these links:
Sound recognition and classification demo
Car engine speed analysis demo
Learn more about the Intel AI Builders program.
Visit the solution catalog to learn more about A2IoT’s market-ready solution for your enterprise.
Intel technologies’ features and benefits depend on system configuration and may require enabled hardware, software or service activation. Performance varies depending on system configuration. No product or component can be absolutely secure.
For more complete information about performance and benchmark results, visit www.intel.com/benchmarks.
© Intel Corporation. Intel, the Intel logo, and other Intel marks are trademarks of Intel Corporation or its subsidiaries. Other names and brands may be claimed as the property of others.
[i] https://www2.deloitte.com/us/en/insights/focus/industry-4-0/digital-leaders-in-manufacturing-fourth-industrial-revolution.html
[ii] Configuration: Test by A2IOT(ISV) Test date 07/19/2019 Platform SKL # Nodes 1 # Sockets 2 CPU 8153 Cores/socket, Threads/socket 16/32 Serial No cpu0 Serial No cpu1 ucode 0x200005e HT On Turbo On BIOS version (including microcode version: cat /proc/cpuinfo | grep microcode –m1) SE5C620.86B.02.01.0008.031920191559 (0x200005e) System DDR Mem Config: slots / cap / run-speed 12 slots / 32 GB / 2666 MTs / DDR4 DIMM Total Memory/Node (DDR+DCPMM) 376GB Storage – boot INTEL SSDSC2KB48 480GB (1GB boot partition) Storage – application drives INTEL SSDSC2KB48 480GB (444GB application partition) NIC Ethernet Controller 10G X550T PCH Intel C621 Other HW (Accelerator) n/a OS Ubuntu 18.04.2 LTS Kernel 4.15.0-52-generic IBRS (0=disable, 1=enable) 0 eIBRS (0=disable, 1=enable) 0 Retpoline (0=disable, 1=enable) 0 IBPB (0=disable, 1=enable) 0 PTI (0=disable, 1=enable) 0 Mitigation variants (1,2,3,3a,4, L1TF) https://github.com/speed47/spectre-meltdown-checker Mitigated Workload & version Classification Training Compiler gcc (Ubuntu 7.3.0-27ubuntu1~18.04) 7.3.0 Libraries scikit-learn (0.20.3) numpy (0.16.4) mkl (2019.4) Pillow(6.0.0) Frameworks version TensorFlow (1.12.0) MKL DNN Version Dataset Custom Topology AlexNet Batch Size 120 Baseline (Provided by Customer) – Sec 13326.65 Optimized multiprocessing solution with Horovod – Sec 2380.58 Improvement 5.598x