Please enable JavaScript.

Coggle requires JavaScript to display documents.

Understanding the HPFP Exchanges and data - Coggle Diagram

- - - - :fire: Q: Can we classify why each pump was exchanged (e.g. pump failure due to fuel dilution)?
        
        :fire: Q: Can we see signals in the data if we visualize the data similar to an onsite expert (e.g. how an Asset Health Supervisor would see the data and classify the failure, and maybe failure mode)?
        
        :question: Q: Can we enrich the visualizations to highlight the signals in the data?
        
        :question: Q: Can the enriched visualizations be presented into the current onsite expert toolset?
        
        :question: Q: Can the onsite citizen content developer create the enriched visualizations?
        
        :question: Q: Could the current onsite expert toolset platform be improved to implement these types of enriched visualizations?
        
        :question: Q: Do we need to first create the enriched visualizations with a tool not available to onsite experts?
        
        :question: Q: Can we build rules to find the signals in the data that the onsite expert would see with their data visualizations?
        
        :fire: Q: Can we classify why the pump was exchanged based on work order notes, and other tradesperson/engineer observations?
        
        :question: Q: Can we classify why the pump was exchanged based on pump tear downs?
      - :check: Q: Can we get additional exchange data (e.g. the months after the current set ends)?
        A: Additional records may be available if we can get them from sites.
        
        :check:: Q; Who can get new additional HPFP exchange data?
        A: Will Kearns provided the new records
        
        :check: Q: Can we ingest the new records in BigQuery and compare to the last list of targets?
        A: Yes, this data was uploaded in the failure data set. The previous recorded matched and we received an additional set of records.
        
        :fire: Q: The new target list also contains some date times for when the failure occurred. Can we compare this expert selected date to the automatic date detection?
        
        :fire: Q: Why do we see some differences in the expert selected date times compared to the automatically detected dates?
      - :black_flag: Q: Can we classify how each pump was scheduled for exchange (e.g. was is planned or unplanned)?
    - - :black_flag: Q: Can we use more than live data (for now there is mismatch between live and old tag lists)?
    - - :black_flag: Q: Who can get the data science team access to the oil analysis data?
    - - :black_flag: Q: Can we identify when each truck had a QSK60 engine installed?
- - - - :check: Q: Are the experts able to show degradation in HPFP health over time?
        A: Jed was able to provide some examples of failure modes for:
        
        Fuel dilution
        
        Plunger not returning (stuck)
        
        :check: Q: Which data was flagged as important?
        A: This data was mentioned by Jed:
        
        Engine oil pressure
        
        Crank case pressure
        
        Engine oil filter differential
        
        Horsepower
        
        Exhaust temperatures
        
        Engine speed
        
        Fuel delivery system
        
        Fuel rail pressure
        
        Temperature (ambiant?)
        
        Truck location
        
        FPPA current (not available yet)
        
        Fuel volume request
        
        Fuel volume actual
      - :fire: Q: Do we know which data points (e.g. sensors) are related to each type of failure mode?
        
        :question: Which data is related to fuel dilution?
        
        :question: Which data is related to the fuel filter being plugged?
        
        :question: Which data is related to the plunger not returning?
        
        :question: Which data is related to fuel pump actuator?
        
        :question: Which data is related to electrical issues?
        
        :question: Which data is related to the manual dump valve open (stuck?)?
    - - :fire: What are the system inputs?
        A: partial list
        
        Low pressure fuel
        
        Mechanical rotational energy
        
        Electronic/Physical control
        
        :black_flag: Q: Is there a pressure requirement on the low pressure fuel input and is it measured/observable?
        
        :black_flag: Q: What is the method for controlling and observing the fuel quality leading up to consumption by the fuel pump?
        
        :black_flag: Q: How is the mechanical energy input regulated (controlled) and can we observe it (e.g. is this is engine RPM)?
        
        :question: Could unbalanced weight on shaft cause vibration that is detrimental to the fuel pump operation or health?
        
        :black_flag: Q; What control signals does the high pressure fuel pump receive and how does it respond to these inputs?
        
        :question: Could noisy demand (irregular control) contribute to detrimental fuel pump operation and health?
      - :fire: What are the system outputs?
        A: partial list
        
        High pressure fuel
        
        :black_flag: Q: What are the expectations for high pressure fuel output?
        
        :black_flag: Q: How is performance measured, do we record this data (e.g. is this fuel rail pressure only)?
        
        :question: Is timing of the high pressure fuel response important (e.g. if the pump lags in creating pressure, does that matter)?
        
        :question: Q: Can we measure the response lag (e.g. time between control input and response)?
        
        :black_flag: Q: What are the consequences to poor performance of the high pressure fuel output?
      - :fire: What are the system internal operations?
        
        :black_flag: Q: Which on-pump sensors exist and how are they used, also can we get this data?
  - - - :check: Q: Can we identify anomalies in the data using unsupervised methods?
        A: Yes, Andrii has identified anomalies using unsupervised methods (Isolation Forest - mainly, Local Outlier Factor)
      - :check: Q: Can we classify using clusters and anomalies as targets with supervised learning methods?
        A: Yes, Andrii trained supervised algorithms to target anomalies (Random Forest mainly)
        
        :check: Q: Can we determine feature importance based on supervised methods?
        A: Yes, Andrii was able to filter out a majority of features from 50K to 7
        However, it seems like daily aggregation data does not contain any reliable signal, so most important features was just missings in columns 0:405504-J1939 and GEnopropel. When data is missing in this alarm or sensor (which is quite rare), in nearly 75% model thinks that this is anomaly behavior. It can be useful, but, for sure, it's obviously not the main signal in the data
        
        :check: Q: How did the supervised learning algorithm perform when predicting anomaly classes?
        A: Pretty good. auc = 0.97; confusion matrix: [5411, 35] for normal data [67, 414] for anomalies.
      - :check: Q: Can we identify clusters using unsupervised methods to group the data?
        A: Yes, Andrii used several methods to cluster the data (k-means clustering with following anomaly detection based on distances)
      - :check: Q: Can we identify failures in the data using supervised methods in the current data setup (daily aggregation, prediction for 7 days ahead)
        A: No, in the current data set up there is absolutely no signal to predict failures, auc is near 0.5, there is no correct failure prediction in confusion matrix
        
        :black_flag: Q: Can we improve our targets by understanding the different failure modes and timeframes for each mode?
        
        :black_flag: Q: Do we have better idea about aggregation period for sensor data? From daily analysis it seems like daily aggregates does not contain any meaningful signal
      - :fire: Q: Can we use time series anomaly detection methods to figure out possibly new ways of labelling the data or even find better time windows to aggregate or look ahead
        A: Andrii to make some research
- - - - :black_flag: Q: Do Teck sites monitor fuel quality, and can we get historical records (e.g. When is our fuel out of ASTM 2D/1D spec, or have hard particles greater than 4 microns like ISO 4406 Code 3 and above)?
        
        :question: Q: Do we have fuel with large hard particles that can be related to fuel system malfunctions?
        
        :question: Q: Do sites have additional fuel filtering before dispensing to trucks?
        
        :question: Q: When do the fuel filters get replaced, and can we get the records of replacement?
        
        :question: Q: Do we have standard fuel filters, and are we able to consistently source this inventory?
        
        :question: Q: What is the process for fuel filter maintenance?
  - - - :question: Q: Which models/types of pumps have which failure modes at Teck, what is the frequency of failure, and how impactful are each of these failures to our business?