{"id":"2fd212e9-e675-439d-8997-c4cb5214f835","arxiv_id":"2504.21317","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":2,"one_line_summary":"A multi-level redundancy mitigation framework is validated on DED audio-visual defect detection, cutting latency by 91%, error rate by 47%, and storage by 99.4%.","lead":"This paper defines redundancy in machine-learning-based monitoring of 3D printing and proposes a framework to cut it at sample, feature, and model levels. In a directed energy deposition defect-detection test it reports 91% lower latency, 47% fewer errors, and 99.4% less storage.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Microphone-removal claim rests on a confounded comparison: Eq. (3) contrasts two systems that differ in resolution, architecture, and training, so R=1.0082 does not isolate the audio modality's contribution.","rationale":"The reader identified the same weakest assumption: the 0.8% balanced-accuracy gap and the comparability of the two systems. My stress-test sharpens this: the gap is not merely a noise-margin problem; the comparison is potentially confounded by input resolution, architecture, and training objective changes, so Eq. (3) may not measure cross-sensor redundancy at all. Even so, the concern is addressable by a controlled ablation, and the latency and storage improvements are credible direct measurements. The manuscript's other limitations (test-set-based pruning ratio selection, unreleased code/data, unclear error-rate baseline) are secondary and do not change the verdict. A conditional acceptance requiring the controlled ablation and error bars is appropriate; the framework itself is plausible and the reported efficiency gains are likely robust to the ablation.","tokens_in":17406,"tokens_out":5936,"duration_ms":55139,"concrete_test":"Run a controlled ablation on the same aligned, downscaled 80×80 dataset: train (a) the visual-only CMKT model exactly as in the paper, and (b) an otherwise identical CMKT model with an additional audio encoder branch consuming the aligned 80×80 spectrograms, using the same hyperparameter search and at least 10 random seeds. Compute test balanced accuracy means and 95% confidence intervals for both, and recompute R_cs(X_A, X_V|Y) in Eq. (3) from the paired accuracies. If the audio-inclusive model is not significantly worse than the visual-only model (paired difference not less than zero), then R is not >1 within noise and the microphone-removal claim is unsupported.","verdict_should_be":"CONDITIONAL","load_bearing_attack":"The central claim that audio is fully redundant (R=1.0082) is computed in Eq. (3) from two balanced accuracies reported in prior work [46]: 97.3% for multimodal fusion and 98.1% for visual-only CMKT. This computation is valid only if the two systems differ solely by the presence of the audio modality. The manuscript does not establish this. Per §4.2 and Figure 5, the original multimodal fusion system in this paper uses 320×320 inputs and overparameterized separate encoders, while the CMKT system uses downscaled 80×80 inputs, a shared encoder, and a contrastive semantic alignment loss. The text also states that downscaling itself improved balanced accuracy by mitigating overfitting before CMKT is applied, so the 0.8% gap cannot be attributed to audio removal alone. Even if [46] controlled for these factors, no confidence intervals are provided, and the margin is within typical seed-to-seed variation. Since sensor removal is the basis for the claimed 50% sensor-cost reduction and is the framework's key cross-sensor redundancy result, this confound is load-bearing. The latency and storage reductions are direct measurements and less affected, but the sensor-removal and accuracy-improvement claims need a controlled ablation.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a formal definition of redundancy in ML-based AM process monitoring, categorizes redundancy into sample-, feature-, and model-level subtypes, and introduces a multi-level redundancy mitigation (MLRM) framework that combines data registration, downscaling, cross-modality knowledge transfer, and model pruning. The framework is validated on a DED defect-detection case study using a public audio-visual dataset. The authors report a 91% latency reduction, a 47% error-rate decrease, and a 99.4% storage reduction, and conclude that the audio microphone is fully redundant (R=1.0082) and can be removed from the operation phase.","tokens_in":17607,"tokens_out":7587,"duration_ms":69683,"significance":"The paper addresses a real gap: most prior work addresses redundancy types in isolation, and there is no unified framework. The taxonomy and the step-by-step pipeline are useful for practitioners. The latency and storage measurements are direct and internally consistent. The entropy-based downscaling analysis is a nice touch. However, the central cross-sensor redundancy result is not established by the current comparison, and the error-rate reduction claim lacks a stated baseline. If the authors supply a controlled ablation with confidence intervals, the framework's claims could be substantiated; as it stands, the microphone-removal conclusion is not supported.","major_comments":[{"comment":"The cross-sensor redundancy calculation in Eq. (3) uses balanced accuracies of 97.3% (multimodal fusion) and 98.1% (visual-only CMKT) taken from prior work [46]. However, the two systems differ in input resolution (320×320 vs 80×80), architecture (separate encoders vs. shared encoder), and training objective (standard fusion vs. contrastive semantic alignment). Section 5 explicitly states that downscaling itself improved balanced accuracy by mitigating overfitting. Therefore the 0.8 percentage-point gap cannot be attributed to removing the audio modality, and R=1.0082 does not establish cross-sensor redundancy. This is load-bearing because it grounds the microphone-removal and the 50% sensor-cost reduction claims. Please provide a controlled ablation in which the same architecture, resolution, and training procedure are used with and without the audio input, reporting mean and standard deviation of balanced accuracy over multiple random seeds.","section":"§4.2, Eq. (3); §5, Fig. 5"},{"comment":"The manuscript claims a 47% decrease in error rate relative to the original system, but the balanced accuracy of the original system is never stated. Using the numbers in Eq. (3), the relative error-rate reduction from 97.3% to 98.1% is (2.7%−1.9%)/2.7% ≈ 29.6%, not 47%. The 47% figure apparently also includes the downscaling gain, but the intermediate balanced accuracy after downscaling and before CMKT is not reported. Please report the balanced accuracy after each MLRM step so the error-rate reduction can be verified.","section":"§5, 'decreased the error rate by 47%'; Abstract"}],"minor_comments":[{"comment":"The formula for cross-sensor redundancy uses the subscript 'cv' (R_{cv}) where it should be 'cs' (R_{cs}) in the displayed equation for R(F_A, F_V, Y).","section":"§3.1.2"},{"comment":"The storage reduction numbers are hard to reconcile: the text gives 4633.1 MB initial total (dataset + model), then 226.06 MB for the downscaled dataset, then 28.49 MB after audio removal. Clarify whether 4633.1 MB is the original dataset alone or includes the original model, and how the ~18-fold reduction follows from 16× downscaling.","section":"§5"},{"comment":"The redundancy calculation for downscaling (R_cv=0.833) uses the minimum entropy rather than the mean; justify this choice, since the mean and maximum entropy are barely changed.","section":"§4.2, Table 2"},{"comment":"The statement 'sensor costs were reduced from $10,000 to $5000' is presented as fact; label this as an illustrative estimate or provide a source.","section":"§5"},{"comment":"There are several presentation issues: 'groups-relative' should be 'group-relative'; the abstract contains a duplicated 'However'; and Table 2 has a formatting artifact in the 'NA *' entry.","section":"Throughout"},{"comment":"The paper does not include a data/code availability statement; since the case study relies on an open dataset and prior work [46], it would be helpful to state whether the trained models and metrics are available.","section":"General"}],"recommendation":"major_revision","confidential_remarks":"The paper relies heavily on the authors' own prior work [46] for the key accuracy numbers that drive the sensor-removal conclusion, and the confound between resolution, architecture, and modality makes the comparison difficult to interpret. The framework and the direct latency/storage measurements are solid contributions, but the central claim about cross-sensor redundancy needs a controlled experiment before publication. The paper's scope fits the journal as a systems/framework contribution, but the case study should not be oversold until the ablation is provided."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Quick take: this paper is a solid engineering integration, not a breakthrough. It defines redundancy at three levels, builds a framework (MLRM) that combines data registration, downscaling, CMKT, and pruning, and demonstrates large efficiency gains in a DED defect-detection case study. The latency reduction from 14.2 ms to 1.33 ms per sample and the storage cut from 4633 MB to 28.58 MB are straightforward measurements and they are credible. The entropy-based downscaling analysis is a nice practical touch.\n\nThe soft spots are mostly around attribution. The central claim that the microphone is fully redundant rests on Eq. (3), which plugs two numbers from the authors' earlier CMKT paper into their redundancy formula: 97.3% balanced accuracy for multimodal fusion versus 98.1% for visual-only CMKT, giving R=1.0082. Those two systems differ in input resolution (320x320 vs 80x80), architecture (separate encoders vs shared encoder), and training (standard fusion vs contrastive semantic alignment). The paper itself notes that downscaling improved balanced accuracy before CMKT was applied, so the 0.8% gap cannot be attributed to removing the audio channel. There are no error bars or multiple seeds, and the margin is smaller than typical run-to-run variation. The 47% error-rate reduction is an end-to-end comparison against the original system, not a per-step effect, and the original baseline accuracy is not stated. Also, the pruning ratios were chosen using test-set balanced accuracy; that should be validation-based. No code or training configs are released, though the dataset is open.\n\nNone of this kills the framework. The taxonomy and pipeline are useful, and the efficiency measurements stand on their own. But the sensor-removal decision, which is the most practically interesting outcome, needs a controlled ablation where the only difference is presence/absence of audio under matched resolution and architecture. The authors should also report error bars, state the baseline, and move pruning selection to validation.\n\nWho this is for: researchers working on multimodal monitoring for AM and anyone thinking about sensor-reduction in industrial ML. It deserves a serious peer review, but with a request for a real ablation before acceptance.","headline":"Useful integration of redundancy-mitigation methods, but the headline sensor-removal claim is built on a confounded comparison and needs a controlled ablation.","tokens_in":18177,"tokens_out":2274,"would_cite":true,"duration_ms":22643,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"This paper argues that redundancy in ML-based process monitoring can be defined, measured, and largely removed, cutting defect-detection latency from 14.2 ms to 1.33 ms per sample and marking the microphone as fully redundant.","keywords":["additive manufacturing","process monitoring","redundancy","machine learning","cross-modality knowledge transfer","model pruning","defect detection","directed energy deposition"],"falsifier":"Retrain the visual-only CMKT and the multimodal fusion classifier many times with different random seeds on the same dataset and compute confidence intervals for balanced accuracy; if the intervals overlap, the measured gap that produces $R=1.0082$ can be training noise, and removing the microphone is not supported by the evidence as stated.","tokens_in":17147,"feed_emoji":"🖨️","tokens_out":6909,"duration_ms":67060,"temperature":0.7,"pith_summary":"This paper tries to establish that redundancy—duplicative, correlated, or non-essential components in data and models—is a hidden driver of cost, latency, and degraded accuracy in machine-learning-based process monitoring for additive manufacturing, and that a structured pipeline can remove it. The paper defines redundancy with a quantitative measure, classifies it as sample-level, feature-level, or model-level, and combines existing mitigation techniques into a multi-level redundancy mitigation (MLRM) framework. In a directed-energy-deposition defect-detection case study, the framework reduces per-sample detection latency from 14.2 ms to 1.33 ms, lowers the error rate by 47%, and shrinks storage from 4633.1 MB to 28.58 MB. The most consequential claim is that the audio modality is fully redundant given the visual modality after cross-modality knowledge transfer, so the microphone can be removed in the operation phase.","feed_headline":"Redundancy audit cuts 3D-print defect detection latency by 91%","feed_subtitle":"A unified redundancy framework prunes data and model weights, cutting storage by 99.4% and error rate by 47%.","key_machinery":"The machinery has three load-bearing pieces. First is the redundancy measure $R(C,K)=1-[(P(K\\cup C)-P(K))/|P(K)|]$, a task-relative score that treats a component as redundant when adding it fails to improve—or worsens—a chosen performance, similarity, or information metric. Second is cross-modality knowledge transfer (CMKT), a shared-encoder architecture with contrastive semantic alignment that aligns latent representations of visual and audio inputs so one modality can predict from the other; this is what makes the microphone removable. Third is iterative $\\ell^1$-norm-based model pruning, which removes the smallest-magnitude parameters until performance drops, turning earlier qualitative observations of overparameterization into a measured redundancy value.","core_discovery":"The central discovery is that redundancy is not merely harmless waste in ML-based AM monitoring; it actively degrades performance, and a systematic audit across sample, feature, and model levels can produce simultaneous gains in speed, accuracy, and compactness. Using the redundancy measure $R(C,K)=1-[(P(K\\cup C)-P(K))/|P(K)|]$, the authors quantify when adding a component helps or hurts: $R<1$ means it contributes, $R=1$ means it is fully redundant but neutral, and $R>1$ means it is fully redundant and damaging. Applied to the audio-visual DED dataset, visual-only cross-modality knowledge transfer reaches 98.1% balanced accuracy versus 97.3% for multimodal fusion, giving $R=1.0082$ for the microphone, and pruning removes 92.9% and 99.7% of parameters from the two networks with $R=1$, i.e., zero measured accuracy loss. Together these steps cut error rate by 47% and latency by an order of magnitude, establishing redundancy mitigation as a way to build cheaper, faster, and more accurate monitoring systems.","pith_inferences":["Because the redundancy measure is relative to the chosen metric $P$ and task $Y$, the label 'fully redundant' is conditional: the same audio features could be informative for a different monitoring target, such as melt-pool size regression or uncertainty estimation.","An immediate testable extension is to run the same MLRM audit on thermal-plus-camera or structure-borne-sound monitoring data; if the microphone result generalizes, many production sensor suites carry fully removable channels.","Since the authors explicitly leave redundancy's positive uses for future work, a natural next step is designing systems that deliberately keep a small, controlled amount of redundancy in safety-critical monitoring paths while still pruning the rest.","The pruning results suggest that model capacity was not the limiting factor in this dataset; if that holds elsewhere, collecting more data or building larger architectures may matter less than cleaning and aligning the data already being collected."],"forward_implications":["A monitoring system built through the MLRM pipeline can run at 1.33 ms per sample, well below the 33.3 ms camera sampling interval, so inference no longer limits real-time control.","Removing the audio channel halves the sensor bill in the case study, from $10,000 to $5,000, and deletes the associated transmission, storage, and maintenance overhead.","Storage falls by 99.4%, from 4633.1 MB to 28.58 MB, making edge and embedded deployment realistic.","Redundancy removal is not a trade-off against accuracy: in the case study the error rate drops by 47% while the model shrinks, so overparameterization and redundant data actively hurt performance.","The same audit can be applied to other multimodal monitoring systems because the framework is not specific to DED or to audio-visual data."],"supporting_citations":[{"why":"Supplies the CMKT architecture, the 98.1% / 97.3% balanced-accuracy numbers used in Eq. (3), and the Bayesian optimization settings.","marker":"[46]"},{"why":"Supplies the open-access audio-visual DED dataset with melt-pool images, spectrograms, and optical-microscope defect labels used in the case study.","marker":"[22]"},{"why":"Introduces audio-visual cross-modality knowledge transfer for laser AM, the mechanism the paper uses to make the audio modality redundant during operation.","marker":"[12]"},{"why":"Demonstrates inference of melt-pool visual characteristics from acoustic and thermal emission, providing the empirical correlation that justifies treating audio as a redundant channel.","marker":"[14]"},{"why":"Supplies the deep-learning data-registration method used to align multimodal monitoring data before redundancy analysis.","marker":"[45]"},{"why":"Provides the model-pruning-plus-edge-computing precedent for large inference-time reductions in AM monitoring, which the paper extends through L1-norm pruning.","marker":"[27]"}],"fun_headline_variants":["Redundancy audit cuts 3D-print monitoring latency 91%, storage 99.4%","Redundancy mitigation in 3D-print ML monitoring cuts errors 47%","Multi-level redundancy audit makes 3D-print monitoring faster, cheaper","Pruning redundant data and models improves 3D-print monitoring accuracy","Redundancy analysis framework yields 99.4% storage cut in 3D-print ML"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that the 0.8 percentage-point difference between visual-only CMKT (98.1%) and multimodal fusion (97.3%) is a real, reproducible performance gap; on that difference hinge $R=1.0082$ and the verdict that the microphone is fully redundant.","fun_headline_variants_meta":{"raw":{"variants":["Redundancy audit cuts 3D-print monitoring latency 91%, storage 99.4%","Redundancy mitigation in 3D-print ML monitoring cuts errors 47%","Multi-level redundancy audit makes 3D-print monitoring faster, cheaper","Pruning redundant data and models improves 3D-print monitoring accuracy","Redundancy analysis framework yields 99.4% storage cut in 3D-print ML"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000701,"raw_usage":{"total_tokens":3211,"prompt_tokens":1041,"completion_tokens":2170,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":657,"completion_tokens_details":{"reasoning_tokens":2058}},"tokens_in":657,"tokens_out":2170,"duration_ms":16492,"temperature":1.0,"reasoning_tokens":2058,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-16T05:06:09.982364+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Retrain the visual-only CMKT and the multimodal fusion classifier many times with different random seeds on the same dataset and compute confidence intervals for balanced accuracy; if the intervals overlap, the measured gap that produces $R=1.0082$ can be training noise, and removing the microphone is not supported by the evidence as stated.","supporting_citations":[{"cited_title":"Audio-visual cross-modality knowledge transfer for machine learning-based in-situ monitoring in laser additive manufacturing","cited_arxiv_id":null,"evidence_quote":"Supplies the CMKT architecture, the 98.1% / 97.3% balanced-accuracy numbers used in Eq. (3), and the Bayesian optimization settings."},{"cited_title":"Multisensor fusion-based digital twin for localized quality prediction in robotic laser-directed energy deposition","cited_arxiv_id":null,"evidence_quote":"Supplies the open-access audio-visual DED dataset with melt-pool images, spectrograms, and optical-microscope defect labels used in the case study."},{"cited_title":"Audio-visual cross-modality knowledge transfer for machine learning-based in-situ monitoring in laser additive manufacturing","cited_arxiv_id":"2408.05307","evidence_quote":"Introduces audio-visual cross-modality knowledge transfer for laser AM, the mechanism the paper uses to make the audio modality redundant during operation."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Demonstrates inference of melt-pool visual characteristics from acoustic and thermal emission, providing the empirical correlation that justifies treating audio as a redundant channel."},{"cited_title":"Deep learning-based data registration of melt-pool-monitoring images for laser powder bed fusion additive manufacturing","cited_arxiv_id":null,"evidence_quote":"Supplies the deep-learning data-registration method used to align multimodal monitoring data before redundancy analysis."},{"cited_title":"In-situ monitoring additive manufacturing process with AI edge computing","cited_arxiv_id":null,"evidence_quote":"Provides the model-pruning-plus-edge-computing precedent for large inference-time reductions in AM monitoring, which the paper extends through L1-norm pruning."}],"review_version":1}