Innovations Autoencoder and its Application in One-class Anomalous Sequence Detection
Authors: Xinyi Wang, Lang Tong
JMLR 2022 | Venue PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | We then demonstrate, using field-collected and synthetic datasets, the effectiveness of the proposed approach on detecting system anomalies in a microgrid (Pignati et al., 2015). 5. Performance Evaluation We present two sets of evaluations based on a combination of field collected datasets from actual systems and synthetic datasets designed to test specific properties. |
| Researcher Affiliation | Academia | Xinyi Wang EMAIL Department of Electrical and Computer Engineering Cornell University Ithaca, NY 14850, USA Lang Tong EMAIL Department of Electrical and Computer Engineering Cornell University Ithaca, NY 14850, USA |
| Pseudocode | Yes | A pseudo code that implements the IAE learning is shown in the Appendix. Appendix B. Pseudocode Algorithm 1 Training the Innovations Autoencoder |
| Open Source Code | No | IAE was implemented by adapting the Wasserstein GAN with a few modifications 6. https://keras.io/examples/generative/wgan_gp/ The text does not provide an explicit statement of code release for the methodology described in this paper, nor a direct link to a repository containing their specific implementation. |
| Open Datasets | Yes | The BESS dataset contained direct bus voltage measurements sampled at 50 k Hz at a medium-voltage (20k V) substation collected from the EPFL campus smart grid as described by Sossan et al. (2016). The second field-collected dataset (UTK) contained direct samples of voltage waveform at 6 k Hz collected at the University of Tennessee. Besides the two field datasets (BESS and UTK), we also designed several synthetic datasets to evaluate specific properties of IAE and IAE-based anomaly detections. These datasets are described in Sec. 5.2 and Sec. 5.3. Table 1: Test Synthetic Datasets. νt i.i.d U[0, 1]. 1() is the indicator function. Table 3: Data Detection Test Cases. νt i.i.d N(0, 1), ν t i.i.d U[ 1.5, 1.5] |
| Dataset Splits | No | anomaly-free training samples are given and 100,000 samples were used for training for all cases. To construct the anomaly samples, we added a comparably small Gaussian Mixture noise on the anomaly-free measurements. The paper describes using anomaly-free samples for training and anomaly samples for testing, and mentions total training sample size, but it does not specify explicit train/validation/test split percentages or methodology for partitioning the overall datasets into these sets. |
| Hardware Specification | No | The paper does not contain any specific details regarding the hardware (e.g., GPU/CPU models, memory) used for running the experiments or training the models. |
| Software Dependencies | No | IAE was implemented by adapting the Wasserstein GAN with a few modifications 6. https://keras.io/examples/generative/wgan_gp/ The paper mentions Keras and Wasserstein GAN but does not provide specific version numbers for these or any other software dependencies. |
| Experiment Setup | Yes | For all cases in this paper, similar neural network structures were used: the encoder and decoder both contained three hidden layers with 100, 50, 25 neurons respectively with hyperbolic tangent activation. The discriminator contained three hidden layers with 100, 50, and 25 neurons, of which the first two used hyperbolic tangent activation and the last one the linear activation. The tunning parameter used for each case is presented in the Appendix. Appendix C. Neural Network Parameter All the neural networks (encoder, decoder and discriminator) in the paper had three hidden layers, with the 100, 50, 25 neurons respectively. The input dimension for the generator was chosen such that n = 3m. In the paper, m = 20 was used for synthetic case, and m = 100 for real data cases. The encoder and decoder both used hyperbolic tangent activation. The first two layers of the discriminator adopted hyperbolic tangent activation, and the last one linear activation. The tuning parameter was chosen to be the same for all synthetic cases, with µ = 0.1, λ = 5, α = 0.0002, β1 = 0.9, β2 = 0.999. For the two real data cases, the hyper-parameters were set to be µ = 0.01, λ = 3, α = 0.001, β1 = 0.9, β2 = 0.999. |