Drilling for oil and gas is one of the world's most dangerous job, workers face from all small equipment failure to the risk of offshore drilling platform events such as fire, fortunately, the application of deep learning in predicting asset maintenance helps prevent natural disasters and man-made disasters,
Due to sensor and iot devices, we on the device has more information than ever before, but we are still in the research the method of processing data, so it is very useful to prevent the catastrophic events, this is a source of deep learning, prediction model of data from multiple sources can be used for training, help oil and gas companies predict impending disaster, enabling them to take proactive approach to
By using PyTorch on Microsoft Azure deep learning framework, accenture has helped a large oil and gas companies such a predictive asset maintenance solution, the solution will plays a role in the protection of employees and the environment,
What is the forecast asset maintenance?
Predictive asset maintenance is chemical plant the core elements of digital transformation, it through a lot of cost-effective sensors, enhanced data processing, automated functional and forecast analysis of progress achieved, it involves information from real-time and historical data is converted into a simple, accessible and operational point of view, this is in order to be able to help identify and eliminate otherwise it will lead to failure of defects, for example, by simply connecting pipe seal in the early defect, we can prevent could lead to catastrophic collapse of the potential failure of the whole gas turbine,
Maintenance budget assets combination based on state monitoring technology, performance analysis, statistical process control and equipment in the factory from different sources of data intuitively clear visualization, in this way can better monitoring and equipment operation, optimization of process, better control process, and improve energy management,
It is worth noting that the process is the core of forecast analysis did not tell what will happen in the future plant operators, and completely determined, on the contrary, they are at an acceptable level of reliability to predict the future may happen, it can also provide "assumption" scenario as well as to the assessment of risks and opportunities,
Asset maintenance maturity matrix
Figure 1 - asset maintenance maturity matrix (source: accenture)
The challenge of the oil and gas
Event prediction is one of the key elements of assets maintenance, for most of the prediction problem, each model has enough sample to create the model to identify them, unfortunately, in some industries, such as oil and gas, all of them in order to avoid failure, failure mode of hot example is rare, it will mean either there will be most standard modeling method without experienced human execution, or can not work,
Accenture and PyTorch and Azure solution
Although there are only a small amount of fault sample, but there are a large number of available time series and data checking,
The method of predictive maintenance
Figure 2 - predictive maintenance methods (source: accenture)
Prepare data in the first phase, in the second stage USES PyTorch built a two stage deep learning solutions, first of all, the recurrent neural network (RNN) combined with short-term memory (LSTM) structure for a long time for training, the structure is the first stage of the second phase, the neural network architecture is used in the solution Koprinkova - Hristova et al., 2011, and inspired Aydin and Guldamlasioglu 2017, the RNN time series model to predict the important variables, such as important sealing temperature, then the forecast input classifier algorithm (random forest) to identify the variables in a safe range, if it is, the algorithm experts can check and solve the potential causes of sorting, this effectively allows experts to solve potential disaster before a potential disaster, the root causes of
Below is used for training and implementation of the solution system diagram:
System architecture
Figure 3 - system architecture
Select the architecture is to ensure that customers in the modeling, training and perform complex machine learning with maximum flexibility when working process requirements is the use of Microsoft Azure, in operation, to meet the requirements of these services is HDInsights virtual machine (DSVM) and data science, if the project have been implemented today, Azure machine learning services will be used to use HDInsights or Azure Databricks data processing training/reasoning,
Use PyTorch because has great flexibility in design calculation when performing a figure, and unlike other deep learning framework, which is bound to the static calculation, perform in figure PyTorch another important advantage is that can use the standard python control process, and the model of each sample can be different, for example, can easily create a tree RNN, PyTorch also supports the use of python debugger, so you can stop at any time the program to check for variables, gradient, etc., this flexibility is very useful in training and adjustment cycle,
PyTorch compared with other deep learning framework, optimization solution to shorten the training time by more than 20%, and the reasoning speed increased by 12%, these improvements in the team work time key environment is critical, please note that the test version is 0.3 PyTorch,
In this project description of the benefits of using PyTorch:
Training time
On the overview of Azure architecture using PyTorch will average training time shortened by 22%,
Debug/bug fixes
Dynamic calculation execution graph with the Python standard features, the combination of overall development time will be shortened by 10%,
Visual
Directly integrated into the Power of BI, get high-end users from the first day of recognition,
Using distributed training experience
Combined with flow control dynamic calculation execution graph allows us to create a simple distributed training model, and get significant improvement in overall training time,
Accenture is how to carry out the final model?
Since the first day of the project, extensibility and operation is a key design consideration, because the customer wants to expand the prototype to the entire fleet of several other assets, as a result, select all components in the system architecture as the standard, in addition, customers want to be able to use Azure Data Factory add more Data sources, Azure service of machine learning and its model management functions for manipulating the model, the figure below illustrates the use of deployment of workflow,
Deployment process
Figure 4 - the deployment of workflow
Deployment model is integrated into the continuous integration/continuous delivery (CI/CD) working process, as shown below,
CI CD workflow
Figure 5 - CI/CD workflow
Azure PyTorch: better together
Product has the function of the combination of Azure AI PyTorch was proved to be the deep learning of cultivating and using the project architecture very effective way of rapid iteration, these choices greatly shortens the training time, improve the efficiency of the data scientists,
Azure is committed to using any language, any frame and any development tool for developers to bring enterprise AI progress, customers can easily to integrate the Azure AI products into the machine learning any part of the life cycle, to mass production project, without having to lock any tools or platforms,
CodePudding user response:
Never know, to learn