IoT | Machine Learning | AI
Solution Overview
Data Ingestion – Cranes operational data gathered from various sensors across geo locations was ingested into AWS storage.
Data Modeling – Custom ARIMA models built to forecast diverse faults across several bins
Data Insights – Predict the imminent faults of different components
Our Approach – The solution used ARIMA algorithm to predict the overall fault volumes as well as location and machinery specific faults & Generated proactive alerts for each machinery highlighting the possible event such as hoist or brake failure
4 Phased Solution – Machine Learning
DATA ANALYSIS PHASE – Initial Data Gathering : Heavy Machinery enterprise had over 10,000 Crane assets enabled with sensors which had to be analysed and as an outcome of the analysis , a prediction was to be made on probable asset failures and malfunctioning. Data Categorisation: Out of 10,000+ assets a total of 100 million records were captured over a period of 9 months duration . Out of these 100 million records about 10 million error records were obtained with over 20 fault codes and error metrics. This faulty data set was analysed across its length and breadth to forecast predictions along with survival & sustenance metrics.
DATA PROCESSING PHASE – Overall 9 month of data was zipped with varied datatypes (~10GB), and the other fault file including the parameter description and other faults was zipped with csv file (130MB) . This data has been analysed in the data processing phase along with parameter file that had about 20 parameter descriptions and almost equal number of fault codes. Data consistency and multi-dimensional analysis was carried out in this phase across asset-id, fault-code and time/months.Used R for basic file processing and got the basic statistics and summaries
DATA FORECAST PHASE – As an outcome of the Data Processing phase , in alignment with the analysis development phase of the algorithms were accomplished across the buckets of Classification, Clustering , Forecast and Survival. Time series models – ARIMA and Holt winters was been implemented on given data. Based on the Data Processing phase findings, fault forecast analysis has been confined only to emergency stops, Hoist overload and Over temperature data. Illustrative application of models has been carried-out on select set of assets.
AUTOMATION AND SURVIVAL ANALYSIS PHASE – ARIMA based fault forecast for all assets has been completed with internal fine tuning and automatic display of performing (Amber and Green) and nonperforming (Red) assets. RAG status has been determined based on the initial 6 months’ average that was used for training the model. Asset performance/faults predictions have been made for subsequent weeks/months and highlighted some of the assets under each of R, A, G categories.
