Process Data Insights provides services and solutions to help businesses increase profits, improve product quality, reduce downtime, increase product and process understanding, and optimize their manufacturing or business processes. We solve problems – for example, excessive variation, poor quality, defects, disturbances, equipment failure, and sub-optimal operation.
NDA = Non-Disclosure Agreement, MSA = Master Service Agreement, SOW = Statement of Work, ML = Machine Learning
P R O J E C T · PLANS
Determine causes for excessive variation and optimize processes
Hourly consulting and teaching
RATE PER HOUR
Ongoing data analytics and ML support
RATE PER WEEK / MONTH
Client data can come from many different sources and in many different formats.
Sensors are used to measure physical properties of the process or product, e.g. temperature, speed, vibration, distance, pressure, voltage, dimensions, etc. These physical measurements are converted into electrical signals for data acquisition. Data is pre-processed as needed, and saved on client PCs.
Clients will be given temporary authenticated access to a designated AWS S3 (Amazon Web Services Simple System Storage) data storage to allow them securely to upload (copy) their data files to Process Data Insight VPC (Virtual Private Cloud). They will use the AWS console to make the upload. They and PDI employees will have exclusive access to the data, and will be able to upload additional files, replace existing, or delete files. You can store virtually any kind of data in any format on AWS S3, e.g. common alphanumeric (ASCII) *.csv and *.txt files.
Data can also be selected and read from client on-premise SQL databases into the Amazon RDS (Relational Database Service) databases using SQL (Structured Query Language).
Streaming data is generated continuously by multiple data sources typically simultaneously. Data needs to be processed sequentially and incrementally over sliding time windows. A wide variety of analytics, including correlations and statistical aggregations are applied to data to derive information. Data also be input to predictive models. Amazon Kinesis software can be used to collect and process streaming data.
We work collaboratively with our clients to analyze the data and generate actionable process knowledge We use this knowledge to suggest practical process improvements.
Usually raw data is not ready for analytics. For best results, we preprocess it first. That can include removing noise and outliers, scaling, time-synchronization, encoding, and feature selection. This is often the most important and time-consuming step in the process as the analysis results are only as good as data quality is. For best modeling results, we ensure that we find an optional representation of your data.
Our advanced visualization tools allow us to review and share many different graphical views of the process data. This can quickly make hidden important information visible and easy-to-understand. Especially when there are a lot of data variables, visualization is a must.
Based on our experience we have found it is important to calculate all the necessary statistical properties of data to learn and draw conclusions. We also use Advanced methods like regression analysis, classification, PCA (Principal Component Analysis) and ICA (Independent Component Analysis) to make more sense from the data.
Using cross-correlations we can often find sources of excessive variation and other problems in time-series data. Autocorrelation is used to study signal similarity at different time lags for time series data.
Frequency Domain Analysis
Often data contains periodic variation indicating problems. e.g. an out-of-round roll in a continuous process. These can be made visible and analyzed by using frequency domain tools, e.g. spectrograms, FFT (Fast Fourier Transform) and wavelet transforms.
Process data can be used to develop models that can predict out-of-control situations or equipment failure.
Machine learning algorithms build a mathematical model based on sample data, known as “training data”, in order to make predictions or decisions without being explicitly programmed to perform the task. We develop accurate predictive models using machine learning methods like XGBoost, DeepAR, and Pytorch image recognition.
Predicting Out-of-Control Situations
Using the models, we can predict how the process will behave, and take corrective actions as needed. E.g. important process control handles can be adjusted to prevent process problems and poor-quality products.
The models can predict equipment failure so that maintenance can be scheduled without unexpected process disruptions and breakdowns. E.g. ball bearings in rolls continuous processes can gradually deteriorate.