Detection and mitigation of glitches in LISA data: A machine learning approach

The proposed Laser Interferometer Space Antenna (LISA) mission is tasked with the detection and characterization of gravitational waves from various sources in the Universe. This endeavor is challenged by transient displacement and acceleration noise artifacts, commonly called glitches. Uncalibrated glitches impact the interferometric measurements and decrease the signal quality of LISA’s time-delay interferometry (TDI) data used for astrophysical data analysis. The paper introduces a novel calibration pipeline that employs a neural network ensemble to detect, characterize, and mitigate transient glitches of diverse morphologies. A convolutional neural network is designed for anomaly detection, accurately identifying and temporally pinpointing anomalies within the TDI time series. Then, a hybrid neural network is developed to differentiate between gravitational wave bursts and glitches, while a long short-term memory (LSTM) network architecture is deployed for glitch estimation. The LSTM network acts as a TDI inverter by processing noisy TDI data to obtain the underlying glitch dynamics. Finally, the inferred noise transient is subtracted from the interferometric measurements, enhancing data integrity and reducing biases in the parameter estimation of astronomical targets. We propose a low-latency solution featuring generalized LSTM networks primed for rapid response data processing and alert service in high-demand scenarios like predicting binary black hole mergers. The research highlights the critical role of machine learning in advancing methodologies for data calibration and astrophysical analysis in LISA.

Figure 1

Normalized illustration of $g_{\mathrm{FRED}}(t)$ for three different sets of parameters. The model will be modified by integrating stochastic fluctuations to accommodate more complex glitch shapes.

Figure 2

Normalized illustration of $g_{{\mathrm{FREDS}}^2}(t)$ for the three parameter sets of Fig. 1 and $w=5000$ at 4 Hz. The same color denotes glitches with identical FRED model parameters highlighting that, while their broader behavior aligns, short-term variations can differ significantly due to the stochastic extension.

Figure 3

Configuration of the LISA constellation, including the index convention for this paper analog to [30]. MOSA, OB, and TM acronyms stand for moving optical subassembly, optical bench, and test mass. A MOSA mainly accommodates an optical bench and a test mass.

Figure 4

CNN architecture and attributes employed for anomaly detection and anomaly injection time estimation. The architecture and subsequent network architectures are visualized using the open-source tool Netron [47].

Figure 5

Example outputs of the dual-channel TTI conversion process, including ground truth labels that highlight the anomaly injection time. The subfigures illustrate the diverse ways in which an anomaly may manifest. Note that in (a), the anomaly is scarcely discernible to the human observer. This prompts the exploration of CNN capabilities in detecting such subtle artifacts in LISA’s TDI data.

Figure 6

Estimated anomaly injection times for the four scenarios presented in Fig. 5, contrasted with the corresponding ground truth data.

Figure 7

Estimated anomaly injection times compared to ground truth data in scenarios with larger estimation errors.

Figure 8

WDNN architecture and attributes employed for anomaly characterization.

Figure 9

Confusion matrix demonstrating the classification accuracy of the hybrid network for the classes outlined in Table 3. The matrix differentiates between the true classes, represented by rows, and the predicted classes, indicated by columns, offering a detailed assessment of the model’s ability to classify anomalous TDI data segments.

Figure 10

Confusion matrix for a standard deviation error in the injection times reduced from 3.995 sec of Fig. 9 to 3.0 sec. This reduction results in a decrease of misclassifications within the off diagonals.

Figure 11

Semilogarithmic ROC curves representing the model’s proficiency for binary classification using the one-vs-rest approach. Translations of class labels are provided in Table 3. Group I corresponds to the off-diagonal elements in Fig. 9.

Figure 12

LSTM architecture and attributes employed for glitch time series prediction. The parameter $c$ represents the length of the TDI inputs, which depends on the FRED model parameters acquired pretraining via differential evolution optimization.

Figure 13

LSTM neural network prediction for an optical displacement glitch injected in the interspacecraft interferometer of optical bench 12, alongside the true glitch and the LSTM inputs.

Figure 14

Semilogarithmic time-frequency spectrograms for the scenario depicted in Fig. 13, showcasing the glitch-free, glitch-uncalibrated, and glitch-calibrated data using the glitch-affected TDI Michelson $X(t)$ channel as an illustrative example. Brighter colors indicate increased noise levels. The vertical red line denotes the moment of glitch injection.

Figure 15

Performance comparison of the LSTM neural network in predicting the glitch of Fig. 13 under conditions of amplified stochastic fluctuations. The TDI inputs for the LSTM predictions are not depicted. The spectrograms for (a)–(d) can be found in the Appendix.

Figure 16

Performance evaluation of the LSTM neural network in the quick response mode considerable for low-latency pipeline applications with reduced emphasis on accuracy. The spectrograms for (e)–(g) can be found in the Appendix.

Figure 17

Outline of the glitch detection and mitigation process as detailed in this paper, with the steps involving machine learning highlighted in dark blue.

Figure 18

Examples of estimated anomaly injection times for glitch characteristics comparable to Fig. 6.

Figure 19

Spectrograms of $X(t)$ for case (a) of Fig. 15.

Figure 20

Spectrograms of $X(t)$ for case (b) of Fig. 15.

Figure 21

Spectrograms of $X(t)$ for case (e) of Fig. 16.

Figure 22

Spectrograms of $X(t)$ for case (c) of Fig. 15.

Figure 23

Spectrograms of $X(t)$ for case (f) of Fig. 16.

Figure 24

Spectrograms of $X(t)$ for case (d) of Fig. 15.

Figure 25

Spectrograms of $X(t)$ for case (g) of Fig. 16.