Sensor based IoT Predictive Maintenance — Why Digital Signal Processing is must for Machine Learning
The industrial plants consist of several types of assets. Sensor based IoT is employed for asset diagnostics and prognostics. The rotating parts of machine assets are often subjected to mechanical wear and tear. If monitoring is not done of this wear and tear, it may lead to the breakdown in the machines and unexpected shutdown in the plant. Apart from mechanical faults, machines can also develop electrical faults as well. Therefore, condition monitoring of these machines is very important for early stage fault detection to avoid unscheduled repairs, minimize downtime and hence, guarantee reliability, up-time, and sustainability of machines. Several non-invasive machine condition monitoring techniques are used using sensors. Most are based on sensing Current, Vibration, Acoustic emission from the machines[2]. To get these signals, transducers are needed. Remote monitoring needs IoT pipelines in place.
The Condition Monitoring of Equipment’s — both electrical and non-electrical — is one of the major requirements for Industrial IoT Industry 4.0. While the sensor based data driven solution looks like a Machine Learning problem, it is not possible to address the component level diagnostics with ML. E.g, the sensory data can be noisy; and depending on the SNR, traditional ML approach would train noise instead of required signal itself. Another problem is time domain data signatures are unable to differentiate, isolate or understand the underline problem with the machine without Signal transforms like Fourier, Wavelets, Time-Frequency, Hilbert etc.
In the above example, the data with fault in Pinion is shown in time-domain. The signal, when trained on ML, will have difficulty to infer anything as it cannot differentiate noise from anomalies. Thresholding approach is also not a good idea under dynamic loads and noise.
By using Fourier Transformation on the signal, we can map many signatures that point to particular anomaly in the machine. Since the patterns have distinct characteristics, this approach is unsupervised and do not require large historical data to create inference.
Let’s look into another example of Elevator System Vibration. Difference between normal operations and abnormal pattern signatures are clearly visible in spectral domain.
Time domain signature of this system does not differentiate anything between normal and abnormal patterns. The low frequency harmonics spikes are clearly visible in Spectral domain.
One more example where the spectrogram of the Acoustics of door operation is shown. The signal contains various operations of the door, and it is possible to identify the certain operations and create degradation models using Signal Processing.
What Qualifies as Digital Signal Processing
After seeing above examples, lets first identify what is not DSP. Well, Digital Signal Processing is not about Fourier Transform (FFT) or FFT is not DSP as widely misconstrued in the Data Science community. There are literally hundreds of Algorithms that fall under the umbrella of Digital Signal Processing. There is another stream which is fusion of Statistics and Signal Processing — Statistical Signal Processing. It applies ML and Signal Processing to derive inferences.
Table below shows some selected algorithms
In the example below, a Digital Filter is applied to specific cut-off frequency defined by ISO specification to the noisy signature that filters out only specific frequency components. It would otherwise be impossible to achieve using Moving Average. The Data is 3 Axial acceleration profile of an elevator @ 100Hz.
Digital signal processing and analog signal processing are streams within engineering that deals with sensor signals. DSP applications include audio and speech processing, sonar, radar and other sensor array processing, spectral density estimation, statistical signal processing, digital image processing, data compression, video coding, audio coding, image compression, signal processing for telecommunications, control systems, biomedical engineering, and seismology, among others.
DSP can involve linear or nonlinear operations. Nonlinear signal processing is closely related to nonlinear system identification and can be implemented in the time, frequency, and spatio-temporal domains. [1]
How to use DSP for AI/ML
There are two ways of applying DSP.
DSP based Data Transformation–
In this methodology, DSP is simply used as an ETL tool. The transformed data is used as another feature of the dataset. The downstream blocks of ML/AI consume the data in statistical sense, and learns the underline patterns to derive inferences. Most Data Scientist use this approach but there are many use cases where this approach is sub-optimal as domain inference and understanding of transforms are missing.
DSP based Design and Inferences -
In this method, the data is prepossessed/transformed into other domain using DSP, but the domain knowledge of DSP is used to derive the inferences that are further used to enhance the downstream AI/ML algorithms. Here, deep understanding of Signals and Systems is prerequisite to create the best ML models.
The use cases, where diagnostic of inner components of assets are needed — not just trend — then this method is mandatory. As shown in Figure 1 , the gear-pinion issue cannot be detected by raw ML or DSP as ETL + ML. We need to understand what harmonics means, what is the meaning of energy in a frequency band, what is the level of noise, attenuation, phase relationships, ability to differentiate signal from noise etc. This applies to almost all asset classes — Motors, Pumps, Compressors, Generators, Conveyors, Engines etc.
DSP and AI Workflow of IoT Sensor Applications
The Workflow of DSP based AI applications is shown above.
Data Acquisition Process —
This process is the most critical for any applications. Decision to select Sensor Types, Sensor Specification, Sampling Frequency, Analog to Digital Conversions, Sensor interface play vital role in the success of the end goal of Diagnostics and Predictive Maintenance.
I have highlighted some key components of Sensor Interface in the below diagram. During the process, I had analysed some of the vendors, so their names are visible. However, there are many players in the market with different specifications, and selection of particular component depends on many domain related specification of the Machine/Source of interest and the end goal.
Many off the shelf smart sensor providers have integrated ADC, Interface and connectivity.
Next is the Critical block of DSP — As explained earlier, this block performs the Signal Processing operations ranging from Filtering, transforms into multiple domains — Table 1
After DSP comes the Feature Engineering part where the features are extracted from the transformed data. These features are addition to original features from the dataset that provide great insights into the data that would otherwise not be possible.
Next we split into two branches. In the domain specific path, we infer the signatures that we got through multiple transformation from domain perspective and derive a diagnostic model from it. Further down the advance model training with diagnostic signatures can be used to train the AI model to derive RUL and other predictive maintenance related metrics.
The other path is statistical inference based approach that is used to give additional information to model building but is purely statistical in nature.
Conclusions
We learned what is Digital Signal Processing and what is its significance for IoT sensor based use cases.
We saw the Difference between traditional Statistical Based AI modeling and Signal Processing based approach.
We saw the Sensor Interface Specification with reference to upstream and downstream integration, ADC, Sampling Rate, Connectivity.
We discussed DSP based AI workflow.
References
- https://en.wikipedia.org/wiki/Digital_signal_processing
- A Machine Condition Monitoring Framework Using Compressed Signal Processing, Meenu Rani, Sanjay Dhok and Raghavendra Deshmukh, Sensors.
- Digital Signal Processing: Principles, Algorithms, and Applications, J.Proakis, 2007.
- Fundamentals of Statistical Signal Processing, Volume III: Practical Algorithm Development, Steven Kay.
- Smart Sensors and Systems — 2015, Youn-Long Lin, et-al, Springer.
