Low-Rank and Structured Sparse Tensor Decomposition for Anomaly Detection in Multivariate Functional Data

• High-dimensional correlations challenge functional monitoring, and FPCA can obscure localized anomalies when they affect multiple principal components.Traditional statistical process control methods are also described as inadequate for complex, high-dimensional data streams.Source: §1 · p. 2
• Tensor decompositions preserve interactions across modes such as sensors and time, making them suitable for multivariate functional monitoring.Prior work cited in the passage demonstrates tensor methods’ potential to extract comprehensive features and between-mode relationships.Source: §1 · pp. 2–3
• ES-CP uses entrywise sparsity for localized anomalies, whereas FG-Lasso adds fiberwise sparsity for deviations along sample-sensor temporal trajectories.Both represent nominal behavior with a low-rank CP tensor and use the sparse component to localize anomalous samples, sensors, and time points.Source: §1 · p. 3
• ES-CP is particularly effective for localized faults, while FG-Lasso provides the strongest overall performance when anomalies exhibit structured dependence across temporal fibers.The study evaluates both methods through two simulation studies and a multichannel forging-process case study, using alternating procedures with closed-form sparse-component updates.Source: §1 · p. 3

• Classical monitoring charts were developed for scalar or low-dimensional data, while FPCA can lack robustness to outliers that dominate its component space.The review also notes that classical methods may not capture structural dependencies across spatial or temporal dimensions.Source: §2 · p. 3
• Correlation-based sampling can reduce costs by exploiting sensor redundancy, but the reviewed approach depends on accurate correlation estimates and an existing monitoring model.It does not explicitly separate smooth signals from anomalies or inherently detect structured and temporally localized sparse anomalies.Source: §2 · p. 3
• Sparse decomposition methods use l1-norm or fused Lasso penalties to isolate anomalies, with reviewed examples spanning manufacturing tensors and photovoltaic images.The passage also describes compressed smooth sparse decomposition for reducing computation and data-transfer costs in high-resolution image streams.Source: §2 · p. 3
• Tensor decompositions preserve structure across modes, while structured regularizers can improve interpretability but may depend on predefined graphs or strong smoothness assumptions.The review also cites functional tensor SVD, tensor-on-tensor regression, graphical Lasso, and spatial Lasso approaches.Source: §2 · p. 4
• Other decomposition and structured-penalty methods target video denoising, online sampling, or gradual degradation, with assumptions that may limit their fit to abrupt or localized faults.The reviewed approaches include SS-RTD, adaptive sampling, and SGL-PCA; the latter emphasizes monotonic health progression.Source: §2 · p. 4
• Existing methods rarely unify smoothness, sparsity, and interpretability in one tensor-based monitoring framework; the paper proposes ES-CP and FG-Lasso to address this gap.ES-CP targets isolated anomalies, while FG-Lasso adds group-wise penalties for temporally or spatially extended deviations.Source: §2 · p. 4

3.1 Basic Notation and Multilinear Algebra

• The method represents nominal process structure with a rank-R CP decomposition whose factor matrices span sample, sensor, and temporal modes.Mode-wise matricizations turn factor estimation into matrix least-squares problems used in alternating optimization.Source: §3.1 · pp. 5–6

3.2 Model I: Entrywise Sparse CP Decomposition

• ES-CP models localized faults as sparse entries added to a low-rank nominal tensor, with entrywise sparsity controlling which residual departures are retained.Its temporal-factor ridge term stabilizes estimation but does not directly impose temporal smoothness; CP rank controls nominal-model complexity.Source: §3.2 · p. 6

3.3 Optimization of the ES-CP Model

• ES-CP alternates between fitting a rank-R CP model to the anomaly-adjusted tensor and soft-thresholding residual entries to update the sparse component.Residual magnitudes no greater than α1/2 are set to zero; larger values are retained after shrinkage.Source: §3.3 · pp. 6–7

3.4 Model II: Fiberwise Sparse-Group Lasso CP Decomposition

• FG-Lasso combines entrywise and temporal-fiber sparsity to detect localized deviations and select anomalous sample–sensor trajectories.Nonzero fibers identify affected trajectories, while nonzero entries localize departure times; the penalty does not impose temporal smoothness or contiguity.Source: §3.4 · pp. 8–9

3.5 Optimization of the FG-Lasso Model

• FG-Lasso alternates regularized rank-R CP fitting with a two-stage fiber update that soft-thresholds entries before shrinking or eliminating each fiber.Its fixed iteration counts and random initialization mean objective monotonicity and convergence to a global minimizer are not asserted.Source: §3.5 · pp. 9–11

3.6 Sample-Level Anomaly Detection and Localization

• The sample-level anomaly score is the largest estimated anomalous deviation across a sample’s multichannel functional observation.The anomaly tensor also represents deviations by channel and time point.Source: §3.6 · p. 11
• A sample is classified as faulty if its estimated sparse component has at least one nonzero entry; the numerical studies use ε_s = 0.The same sample-level rule is applied to ES-CP, FG-Lasso, and TRPCA, while PCA-based benchmarks use labels from Isolation Forest, Local Outlier Factor, and One-Class SVM.Source: §3.6 · p. 11
• The support of the sparse component identifies anomalous channels and the corresponding time locations, enabling hierarchical detection and localization.Channel-level indicators identify affected functional channels, and nonzero entries identify time points.Source: §3.6 · p. 11

4.1 Simulation I: Synthetic Multichannel Functional Process Simulation

• Simulation I varies casewise and componentwise faults across one or all ten channels and four severity levels, using 1,000 normal samples and 5% contaminated samples.The methods are compared with TRPCA and three PCA-based detectors; their contamination parameters use known anomaly proportions, making that comparison oracle-assisted.Source: §4.1 · p. 12, §4 · p. 12
• FG-Lasso achieves or ties the highest macro F1 in 15 of 16 Simulation I scenario–severity combinations, averaging 0.799 versus 0.741 for ES-CP and 0.706 for TRPCA.Across all settings, the three PCA-based detectors average 0.554–0.587; the exception is mild all-channel casewise contamination, where PCA+LOF scores 0.559 versus 0.543 for FG-Lasso and ES-CP.Source: §4.1 · p. 13
• For all-channel faults, FG-Lasso macro F1 rises from 0.617 to 0.933 under componentwise contamination and from 0.543 to 0.956 under casewise contamination, with the latter exceeding ES-CP’s 0.792 and TRPCA’s 0.824 at highest severity.At highest severity under all-channel componentwise contamination, FG-Lasso exceeds ES-CP and TRPCA by 0.133 and 0.107; under all-channel casewise contamination, its precision and recall are 0.953 and 0.958.Source: §4.1 · p. 13, Table 1 · p. 14, Table 2 · p. 14
• For one-channel faults, FG-Lasso and ES-CP perform similarly: their componentwise macro F1 scores tie at severity levels 2 and 4, and their casewise scores both reach 0.969 at level 4.At severity levels 1 and 3 in the componentwise scenario, FG-Lasso’s advantages are 0.017 and 0.006; the added fiberwise penalty provides only a small improvement when anomalies are confined to one trajectory.Source: §4.1 · p. 13, Table 3 · p. 14, Table 4 · p. 14

4.2 Simulation II: Multivariate Functional Data with Missingness and Anomalies

• Simulation II uses 200 samples, five correlated channels, and 100 time points, with trajectory masking, additive perturbations, and rank-one mean-shift anomalies in 5% of samples.Anomaly magnitude varies over five severity levels.Source: §4.2 · p. 14
• FG-Lasso has the highest average Simulation II macro F1 at 0.711 versus 0.681 for ES-CP, and leads at four of five severities; at severity 3.00, their scores are 0.917 and 0.904.At severity 0.25, PCA+LOF and PCA+IF lead with 0.579; FG-Lasso and PCA+OCSVM follow at 0.578. At severity 2.00, FG-Lasso and ES-CP score 0.785 and 0.778, versus TRPCA at 0.663.Source: §4.2 · p. 14, Table 5 · p. 14

5 Case Study

• The analysis uses 513 profiles—306 normal and 207 faulty—arranged as a samples-by-crank-angle-by-sensor tensor and evaluated without supervision.The dataset retains three fault groups after excluding conditions that were visually evident; the profiles are block-averaged to 200 crank-angle positions.Source: §5 · p. 15
• FG-Lasso has faulty-class precision of 0.95 and recall of 0.69, while ES-CP has precision of 0.83 and recall of 0.73.FG-Lasso correctly classifies 299 of 306 normal profiles, with normal-class recall 0.98 and false-alarm rate 0.02.Source: §5 · p. 15
• PCA+OCSVM is the strongest PCA-based detector, with F1 scores of 0.75 for normal profiles and 0.63 for faulty profiles.TRPCA’s corresponding F1 scores are 0.65 and 0.47; TRPCA and PCA+LOF each detect 94 of 207 faulty profiles, for faulty-class recall 0.45.Source: §5 · p. 15
• FG-Lasso’s sparse reconstructions localize faults into crisp patches in affected channels, whereas TRPCA produces a more diffuse anomaly map.FG-Lasso’s low-rank reconstruction retains smooth process-driven patterns, while TRPCA retains noticeable fault signatures near peak-force instants.Source: §5 · pp. 15–16, Figure 3 · p. 16, Figure 4 · p. 16
• 0.85 and 0.82 are the macro F1 scores for FG-Lasso and ES-CP, respectively, confirming FG-Lasso’s overall advantage.Table 6 reports class-specific precision, recall, and F1 for six unsupervised methods on the forging-process profiles.Source: §5 · p. 15, Table 6 · p. 16

6 Conclusion

• ES-CP uses entrywise sparsity for localized deviations, while FG-Lasso adds fiberwise sparsity for anomalies along temporal trajectories or across sensor channels.Both methods model nominal behavior with a low-rank component; temporal-factor ridge regularization stabilizes its estimation.Source: §6 · pp. 16–17
• The results indicate that the appropriate sparsity structure depends on the anticipated fault mechanism.ES-CP is competitive for isolated or channel-specific anomalies, while FG-Lasso offers improved precision and stronger performance for temporal-fiber or multi-channel deviations.Source: §6 · p. 17
• The framework supports anomaly detection and interpretable fault localization for industrial processes with correlated, noisy, and incomplete functional measurements.Source: §6 · p. 17
• Future work includes online updating, automatic CP-rank and regularization selection, uncertainty quantification, irregular or partially observed profiles, and heterogeneous multimodal data.Source: §6 · p. 17

6.1 Proof of Proposition 1

• The Proposition 1 proof reduces the sparse-component problem to independent scalar minimizations and obtains the solution by considering positive, negative, and zero cases.The scalar objective is strictly convex, and applying the resulting solution entrywise completes the proof.Source: §6.1 · pp. 18–19

6.2 Proof of Proposition 2

• The Proposition 2 proof separates the sparse-component optimization across temporal fibers and derives a unique fiber-level minimizer using entrywise soft-thresholding.The objective is strongly convex; the solution is applied independently to each residual fiber.Source: §6.2 · pp. 19–21

6.3 Details of Simulation I

• Simulation I samples functional data at 100 uniformly spaced time points and considers componentwise and casewise contamination affecting individual or all channels.Its covariance construction uses cross-channel dependence and eigenvalue-based truncation at L*=225.Source: §6.3 · p. 22

6.4 Details of Simulation II

• Simulation II uses N=200 samples, p=5 channels, T=100 time points, and cross-channel correlation ρc=0.75, with trajectory-level masking and anomalies injected into 5% of samples.Each selected sample-channel trajectory is masked; injected anomalies share temporal support across channels and are applied before masking.Source: §6.4 · p. 23
• Simulation II assigns sample-level anomaly labels according to whether anomaly injection was selected, not according to exact entrywise anomaly support.Source: §6.4 · p. 23