Analytics courses
Elaborate planning is required to reap full benefits of digital technologies. After digitising processes and making organisational changes, next is the decision making stage, where new technologies such as machine learning and optimisation are exploited for better operational efficiency. With technology so fast changing, keeping track of relevant algorithms is a real challenge. While there are many courses on theory, there is a lack of real implementation experience for managers. To fill this gap, we design a series of short courses to bring implementers and managers up to speed quickly using our algorithms, survey of practices and discussion of application papers
Our courses are distilled from best practices gained from our techniques and hence unique. We provide an instructor for a one-day course at client's premise, and software for hands-on practice if any. The following are possible modes
10 to 20 pax on technological insights for programme managers, systems integrator, systems architect, engineers and scientists
4 to 10 pax with hands-on for implementers. Attendees are to use their laptops, preferably win-x64
We cover explainable data analytics, machine anomalous sound detection and fleet maintenance analytics. Further courses are possible for complex applications. For instance, machine anomalous sound detection works better when tuned with data, the tuning parameters are problem-specific can be a sequel to Course 2
Quadratic model (right plot) fits better, especially at top right corner
Course 1: Explainable data analytics
Decision-tree algorithms such as random forests achieve good accuracy, but are often hard to explain. Support vector machines with quadratic expansion give high accuracy and yet easy to understand. For complex scenarios, a kernel is required and proximity to support vectors provides explainability. This course discusses accurate and explainable models which are easy to script, easy to analyse and easy to deploy
The contents are listed below, it starts off with quadratic expansion and quantile regression which are superior to traditional least square method. It includes feature selection, classification and a visualisation technique guaranteed to preserve connectivity called Topomap. A software running Ubuntu-22.04-amd64 on Windows Subsystem for Linux will be provided for hands-on practice
Improved regression via quadratic expansion
Range estimate via quantile regression
Feature selection via backward elimination
Non-linear regression via kernel functions
Classification via support vector machine
Visualisation via Topomap
Application paper discussion: Modeling of the Relationship Between Speed Limit and Characteristic Speed of Expressway Traffic Flow
Application paper discussion: TopoMap: A 0-dimensional Homology Preserving Projection of High-Dimensional Data
Preferred expertise: Basic Python scripting, basic knowledge of regression analysis
Left: Fourier transform. Right: Gabor transform
Course 2: Machine anomalous sound detection
Outlier classifier algorithms are attractive for anomalous sound detection because they require only normal training samples. A good range of models covering outlier to binary and multiple classification can be constructed with support vector machines (SVM). At the onset, a one-class SVM can be adopted for outlier detection. As abnormal data become available, binary SVM may be used. With more granularity on abnormal samples, multi-class SVM may be applicable. Using SVM to build a comprehensive detection system is discussed in this course
The contents are listed below, it starts off with extracting sound spectrogram from wav files using Gabor transform, which is a special case of the short-time Fourier transform. It then discusses various methods of one-class classification followed by one-class, multi-class SVM and practical issues such as class imbalance. Hands-on practice of sound spectrogram extraction using Python packages are included
Sound spectrogram extraction
Survey paper discussion: One-Class Classification: A Survey
One class SVM
Binary and multi classification via SVM
Addressing class imbalance
Preferred expertise: Basic Python scripting, basic knowledge of (data science) classification algorithms
Assigning systems to demands, S1 requires PM at TBlk7; S5 at TBlk6; S4 has no demand and is not required
Course 3: Fleet maintenance analytics
To sustain continuous operation, extra equipment are typically needed for preventive maintenance (PM) and as standbys in event of breakdowns. Whereas the latter is unpredictable with little scope of optimization, it is possible for PM to be spaced out evenly. PM is intrinsically linked to demand. When there is no usage, minimal PM suffices. For optimal results, PM and demand scheduling should be integrated so that
demand pattern can be exploited optimally. For instance, schedule PM when demand is low, and delay PM (within specified period) when demand is high
when demand exceeds some threshold, we can impose non-PM restriction on a corresponding number of equipment
if circumstances permits, instead of buying new equipment, we can overproduce and store inventory to hedge against high demand
This course discusses how fleet maintenance can be optimised. The contents are listed below, it starts with real experience on spare ratio in the transport industry. This is followed by benefits of optimisation - reducing spare ratio from operation viewpoint, and flattening out peak workshop load from maintenance viewpoint. Condition-based maintenance with analytics on equipment health may change PM frequency. A paper discussion on a fault diagnosis model is hence included
Survey paper discussion: Transit Cooperative Research Program 109 - System-Specific Spare Bus Ratios Update
Optimising spare ratio
Optimising workshop peak loading
Application paper discussion: Bayesian-Optimized Hybrid Kernel SVM for Rolling Bearing Fault Diagnosis
Preferred expertise: Basic knowledge of preventive maintenance practices