Inter national J our nal of Recongurable and Embedded Systems (IJRES) V ol. 15, No. 3, No v ember 2026, pp. 645 ∼ 655 ISSN: 2089-4864, DOI: 10.11591/ijres.v15.i3.pp645-655 ❒ 645 Deployment-awar e acoustic anomaly detection on an embedded micr ocontr oller T ruong Minh Duc, Nguy en Ngoc Minh, T ruong Cao Dung F aculty of Electronics Engineering 1, Posts and T elecommunications Institute of T echnology (PTIT), Ha Noi City , V ietnam Article Inf o Article history: Recei v ed Jun 12, 2026 Re vised Aug 31, 2026 Accepted Sep 11, 2026 K eyw ords: Acoustic anomaly detection Embedded consistenc y e v aluation Internet of things edge monitoring Recongurable systems T in y machine learning ABSTRA CT Industrial acoustic anomaly detection is often optimized in of f-de vice bench- marks, while the consistenc y of e xported thresholded decisions on embedded microcontrollers remains less studied. This paper presents a bounded embedded consistenc y-e v aluation case study for an internet of things (IoT)-class acoustic monitoring pipeline on the public MIMII f an 0 dB subset. The pipeline uses a 40 × 63 × 1 log-m el representation, compact con v olutional autoenc oder (CAE), INT8 TFLite deplo yment, and latent-space Mahalanobis scoring. The ra w M aha- lanobis r ule denes the on-de vice decision; a 2-out-of-5 temporal rule is retained only for of ine ordere d-subset analysis. F or f an/id 06/0 dB, repeated strat ied e v aluation y ielded a v erage area under the curv e (A UC) 0.96 and F1 0.82. Sup- porting all-ID e v aluation o v er id 00, id 02, id 04, and id 06 sho wed the ra w Mahalanobis path had the strongest aggre g at e metrics among tested scoring paths. Stored-sample replay on ESP32-S3 used a x ed 20-sample id 06 batch with generated C audio headers through the canonical DSP pat h. Both reference-input and digital signal processing (DSP )-path checks produced 20/20 ra w-decision agreement with the of f-de vice reference. The results support bounded reference- de vice consistenc y for the e v aluated batch and pro vide a reference-de vice basis for future embedded, IoT -edge, and block-le v el recongurable-system studies, b ut not cross-ID rob ustness, streaming, latenc y , or ener gy beha vior . This is an open access article under the CC BY -SA license . Corresponding A uthor: T ruong Minh Duc F aculty of Electronics Engineering 1, Posts and T elecommunications Institute of T echnology (PTIT) Ha Noi City , V ietnam Email: ductm@ptit.edu.vn 1. INTR ODUCTION Industrial machine health monitoring is essential for a v oiding costly do wntime, and predicti v e mainte- nance increasingly relies on acoustic analysis to detect abnormal operating states without interrupting production. In smart f actories and connected machinery , such detectors are often e xpected to mo v e from an of f-de vice de v elopment en vironment to embedded or internet of things (IoT) edge nodes close to the monitored equipment. In this conte xt, unsupervised anomalous sound detection (ASD) has attracted substantial attention because it learns normal sound patterns and ags de viations without requiring labeled f ault data, which are often scarce and unpredictable in practice. Progress in industrial ASD w as substantial ly shaped by the release of the MIMII dataset [ 1 ], which pro vides machine recordings under both normal and f aulty conditions in realistic en vironments, and by the subsequent benchmark formalization in DCASE 2020 T ask 2 [ 2 ], where systems are trained only on normal data and must detect pre viously unseen anomalies at test time. The a v ailability of public benchmarks has enable d a wide range of ASD approaches. Recent studies J ournal homepage: http://ijr es.iaescor e .com Evaluation Warning : The document was created with Spire.PDF for Python.
646 ❒ ISSN: 2089-4864 ha v e reported g ains through stronger temporal modeling [ 3 ], self-supervised representation learning [ 4 ], and compact one-class objecti v es [ 5 ], [ 6 ]. Embedded inference research has lik e wise sho wn that quantized neural netw orks and compact T in y machine learning (T in yML) backbones can e x ecute within tight microcontroller b udgets [ 7 ], [ 8 ]. Ho we v er , much of the e xisting liter ature still focuses on benchmark accurac y , representation learning, or hardw are feas ibility rather than on whether e xported scoring artif acts preserv e thresholded ra w decisions a fter quantized t ransfer to microc on t roller unit (M CU)-class de vices. Of f-de vice e v aluation alone cannot v erify whether the e xported digital signal processing (DSP) path, quantized inference runtime, and scoring artif acts preserv e the same alarm/no-alarm decisions under MCU memory and implementation constraints. This transfer chec k is important for industrial anomaly detectors because small score perturbations near the operating threshold can change maintenance alarms, operator triage, or do wnstream logging beha vior e v en when aggre g ate benchmark metrics remain strong. Ag ainst this background, the present paper studies a narro wer embedded consistenc y-e v a luation question. W ithin a x ed DCASE-style MIMII f an 0 dB setting, it asks whether an of f-de vice trained anomaly detector can be e v aluated on ESP32-S3, used here as a concrete MCU-class T in yML tar get, through stored- sample replay of a x ed test batch while preserving ra w-decision agreement when both the INT8 encoder and the e xported latent-space Mahalanobis artif acts are frozen. All model tting and threshold calibration follo w the one-class benchmark protocol: only normal MIMII f an recordings are used before e v aluation, while abnormal clips are held out for testing. Among the e v aluated f an IDs, id 06 w as carried forw ard as the deplo yment anchor because it combined the strongest repeated-split ra w metrics, the lar gest mean Mahalanobis mar gin, and the most stable anchor beha vior in the bounded per -ID comparison. The weak er ID results are k ept as supporting limitations rather than omitted from the e vidence chain. The de vice-side c heck then replays a frozen id 06 batch deri v ed from the public MIMII benchmark W A V les rather than introducing a ne w li v e-data collection. The claim boundary is limited to ra w-decision agreement between the of f -de vice reference and de vice e x ecution under that protocol. Score and latent discrepancies are treated as bounded diagnostics, the retained 2-out-of-5 rule as an of ine ordered-subset polic y rather than as a de vice-v alidated endpoint, and the de vice study as x ed-batch reproducibility plus static memory-footprint reporting rather than li v e microphone acquisition, continuous streaming, timing, or ener gy characterization. The paper mak es three contrib utions. First, it frames a bounded replay question as an embedded and IoT -edge consistenc y-e v aluation case study for MIMII f an 0 dB on ESP32-S3, with a block-decomposed DSP , INT8 encoder , Mahalanobis scoring, and thresholding path rele v ant to later recongurable-s ystem studies. Second, it freezes an INT8 encoder together with e xported latent-space Mahalanobis artif acts ( µ , Σ − 1 , τ ) for a x ed-batch reference-de vice comparison centered on thresholded ra w-decision agreement. Third, it k eeps the e vidence hierarch y e xplicit by separating the primary id 06 anchor study , supporting all-ID descripti v e conte xt, and the direct x ed-batch de vice-side consistenc y e vidence. Thi s contrib ution dif fers from prior T in yML ASD studies by testing the transferred decision path itself, not by proposing a ne w ASD backbone or a complete real-time embedded monitoring system. 2. RELA TED W ORK Prior w ork rele v ant to this study can be grouped by proximity to the present pipeli n e . The closest group is one-class, reconstruction-based, and lightweight decision-oriented ASD, although se v eral methods dif fer from the present pipeline in supervis ion, metadata usage, or score construction. Deep SVDD has been appl ied to industrial machine audio as a compact one-class baseline [ 6 ], [ 9 ] while con v olutional v ariational autoencoders and compact self-encoder pipelines remain natural reconstruction baselines [ 10 ], [ 11 ]. Acti vity-a w are detection and attenti v e neural-process scoring e xplore alternati v e decision structures be yond plain reconstruction [ 12 ], [ 13 ], h yperbolic latent modeling of fers a dif ferent geometric vie w of anomaly separation [ 14 ], and metadata- assisted rst-shot ASD and prototype-based learning e xtend this design space when anomaly estimation or partial attrib ute information is a v ailable [ 5 ], [ 15 ]. Broader of f-de vice ASD studies pro vide a second comparison group: temporal modeling with recurrent backbones and spectral-temporal fusion remains competiti v e for industrial audio [ 3 ], [ 16 ], while recent systems increasingly rely on self-supervised or contrasti v e representation learning [ 4 ], [ 17 ]. Joint generati v e-contrasti v e learning, hierarchical metadata constraints, and noise-a w are contrasti v e object i v es further rene latent representations under more demanding benchmark settings [ 18 ]–[ 20 ]. A third comparison group concerns embedded feasibility and MCU deplo yment. Quantized neural- netw ork libraries and embedded inference runtimes establish the softw are foundation for MCU-side e x ecution [ 7 ], Int J Recongurable & Embedded Syst, V ol. 15, No. 3, No v ember 2026: 645–655 Evaluation Warning : The document was created with Spire.PDF for Python.
Int J Recongurable & Embedded Syst ISSN: 2089-4864 ❒ 647 [ 21 ], and ful ly quantized T in yML backbones and related lightweight edge reports sho w that useful models can t within strict memory and compute b udgets on MCU-class hardw are [ 8 ], [ 22 ]. These embedded-inference studies moti v ate the paper’ s narro w question: whether a frozen con v olutional autoencoder (CAE)-plus-Mahalanobis detector preserv es ra w decisions in x ed-batch ESP32-S3 replay . The present study therefore complements prior T in yML and embedded-inference w ork by auditing the transferred decision path and e xported scoring artif acts, while treating hardw are-le v el realization as future w ork rather than as e vidence pro vided here. 3. RESEARCH METHOD 3.1. F eatur e r epr esentation and con v olutional autoencoder ar chitectur e This subsection denes the e xportable feature-to-latent path of the proposed pipeline. Each 2 s audio se gment is sampled at 16 kHz and con v erted into a 40 × 63 × 1 log-mel tensor [ 23 ] using n f ft = 1024, hop length = 512, and n mels = 40. The reproducible feature path uses MIMII channel 0, zero-pads or truncates each clip to the x ed 2 s duration, applies librosa mel-spectrogram def aults for parameters not listed abo v e, and con v erts po wer to decibels with ref = np.max before the x ed TFLite af ne input quantization. Let x t ∈ R 40 × 63 × 1 denote the resulting feature tensor at sequence inde x t . A c o m p a c t C A E m a p s t h i s t e n s o r t o a l o w - d i m e n s i o n a l l a t e n t r e p r e s e n t a t i o n u s e d b y t h e E S P 3 2 - S 3 s c o r i n g p a t h s t u d i e d h e r e . T h e e n c o d e r a p p l i e s a 3 × 3 c o n v o l u t i o n w i t h 8 c h a n n e l s a n d R e L U a c t i v a t i o n , a 2 × 2 m a x - p o o l i n g l a y e r , a s e c o n d 3 × 3 c o n v o l u t i o n w i t h 1 6 c h a n n e l s a n d r e c t i f i e d l i n e a r u n i t ( R e L U ) a c t i v a t i o n , a n d a s e c o n d 2 × 2 m a x - p o o l i n g l a y e r . A t h i r d 3 × 3 c o n v o l u t i o n w i t h 1 6 c h a n n e l s f o r m s a l a t e n t t e n s o r , w h i c h i s r e d u c e d b y g l o b a l a v e r a g e p o o l i n g t o t h e 1 6 - D l a t e n t v e c t o r z t = f θ ( x t ) ∈ R 1 6 . T h e r e s u l t i n g C A E h a s 7 , 1 2 1 t r a i n a b l e p a r a m e t e r s . T h i s c o m p a c t C A E d e s i g n f o l l o w s t h e g e n e r a l d i r e c t i o n o f l i g h t w e i g h t l a t e n t - e n c o d e r A S D s y s t e m s t h a t l e a r n c o m p a c t l a t e n t r e p r e s e n t a t i o n s u s e f u l f o r d o w n s t r e a m a n o m a l y s c o r i n g u n d e r l i m i t e d m o d e l c a p a c i t y [ 1 0 ] , [ 1 1 ] . During training, the latent tensor is passed through a symmetric decoder composed of tw o transposed- con v olution blocks with upsampling, follo wed by a cropping layer and a linear single-channel reconstruction head. The autoencoder is optimized of f-de vice with Adam and mean squared reconstruction loss so that the encoder learns a reconstruction-oriented latent space. F or embedded deplo yment, only the e ncoder branch is quantized and e xported as an INT8 TFLite model, while the decoder remains a training-time component. The de vice-side consistenc y study uses a tw o-stage ESP32-S3 protocol. Stage A dri v es the MCU with of f -de vice reference tensors to isolate the e xported inference and scoring artif acts. Stage B replays the same x ed batch t hrough the canonical MCU DSP path. Figure 1 summarizes the v alidated path from DSP/log-mel features and INT8 encoder inference to latent Mahalanobis scoring and thresholded ra w decisions. De vice-side v ali dation stops at ra w thresholding; temporal v oting and the terminal decision are retained only as of ine ordered-subset polic y blocks. This block-le v el separati on k eeps front-end DSP , quantized encoder inference, compact scoring artif acts, and threshold logic inspectable as implementation units. Thus, the study e v aluates x ed-batch ra w-decision consistenc y , not streaming beha vior , latenc y , or ener gy . 3.2. Latent-space scoring and temporal decisions T h i s s u b s e c t i o n d e f i n e s t h e s c o r e a r t i f a c t s a n d d e c i s i o n r u l e s t h a t o p e r a t e o n t h e 1 6 - D l a t e n t v e c t o r . T h r e e s c o r e f a m i l i e s a r e c o n s i d e r e d i n t h e s t u d y : r e c o n s t r u c t i o n m e a n s q u a r e d e r r o r ( M S E ) , l a t e n t L 2 - t o - m e a n , a n d M a h a l a n o b i s s c o r i n g i n t h e l a t e n t s p a c e . O n l y M a h a l a n o b i s s c o r i n g f o l l o w e d b y t h r e s h o l d i n g d e f i n e s t h e d e p l o y e d r a w d e t e c t o r ; r e c o n s t r u c t i o n M S E a n d l a t e n t L 2 - t o - m e a n a r e o f f - d e v i c e c o m p a r i s o n s c o r e s u s e d i n t h e a b l a t i o n a n a l y s i s . T h e f o c u s o n a c o m p a c t l a t e n t - s p a c e d e t e c t o r i s c o m p l e m e n t a r y t o r e c e n t s e l f - s u p e r v i s e d a n d c o n t r a s t i v e A S D s t r a t e g i e s [ 4 ] , [ 1 8 ] . P r o t o t y p e - b a s e d a d a p t a t i o n h a s a l s o s h o w n p r o m i s e w h e n a d d i t i o n a l a t t r i b u t e i n f o r m a t i o n i s a v a i l a b l e [ 5 ] . Mahalanobis is retained as the deplo yed ra w score because i t has a compact e xported-artif act form through µ , Σ − 1 , and τ . F or each e xport, the INT8 encoder output is dequantized back to z t ∈ R 16 ; µ is the train- normal latent mean in that space, Σ − 1 is obtained from a re gularized latent co v ariance using a pseudoin v erse when needed, and τ is set by a percentile rule on train-normal Mahalanobis scores. When a single frozen e xport b undle is used for ra w reference-v ersus-de vice agreement checks, these artif acts are held x ed, whereas repeated-split analyses re generate their o wn score artif acts. The deplo yed latent-space score is written as (1): s t = ( z t − µ ) ⊤ Σ − 1 ( z t − µ ) , (1) where µ and Σ − 1 denote the e xported train-normal latent mean and re gularized in v er se co v ariance in the dequantized latent space. Since s t is used as a ranking and thresholding quantity rather than as a normalized Deployment-awar e acoustic anomaly detection on an embedded micr ocontr oller (T ruong Minh Duc) Evaluation Warning : The document was created with Spire.PDF for Python.
648 ❒ ISSN: 2089-4864 lik elihood, we refer to it throughout the paper as a Mahalanobis score. A ra w anomaly decision is t h e n dened as (2): ˆ y ra w t = ( 1 , s t > τ , 0 , otherwise , (2) where τ denotes the operating threshold. Higher v alues of s t indicate lar ger de viation from the normal latent distrib ution. T emporal v oting is e v aluated as a second design layer on top of the ra w detector in of ine ordered- subset analysis; it is not the endpoint used for the de vice ra w-decision consistenc y check. Here, t inde x es successi v e 2 s se gments only after lename restoration within an ordered e v aluation subset, so temporal v oting is characterized in that setting rather than as a generic streaming benchmark. F or a temporal rule with required v ote count k and windo w length w , the aggre g ated decision at sequence inde x t is written as (3): ˆ y ( k ,w ) t = ( 1 , P t i =max(1 , t − w +1) ˆ y ra w i ≥ k , 0 , otherwise . (3) Figure 1. Proposed T in yML pipeline for acoustic anomaly detection: feature e xtraction, INT8 encoder inference, Mahalanobis scoring, thresholding, and temporal decision blocks The paper retains the 2-out-of-3 and 2-out-of-5 temporal rules for of ine ordered-subset analysis. A stricter 3-out-of-5 v ariant w as check ed only as a sensiti vity case and w as not used to reselect the detector , threshold, deplo yment anchor , or de vice-side claim. The truncated-prex denition mak es the rst w − 1 positions e xplicit without zero padding. Among the retained rules, 2-out-of-5 is treated as a pre-x ed of ine smoothing polic y , not as e vidence be yond the ra w Mahalanobis detector . The e xperiments are or g anized around three e vidence layers. The repeated f an/id 06/0 dB study pro vides t he primary single-ID anchor e vidence. The repeated all-ID study pro vides supporting descripti v e e vidence about cross-ID v ariation and remaining limitations. The x ed 20-sample test batch pro vides direct bounded refer ence-de vice consistenc y e vidence for the deplo yed ESP32-S3 path. A separate score-f amily ablation pro vides single-run design-selection e vidence and is not treated as a fourth claim-bearing layer . 3.3. Dataset, training, and export setup The study follo ws the one-class MIMII/DCASE setting on the f an 0 dB subset, with id 00, id 02, id 04, and id 06 used in the repeated all-ID e v aluation and id 06 used as the primary single-ID anchor for deplo yment-consistenc y analysis [ 1 ], [ 2 ]. W ithin this s cope, the study e v aluates 4075 normal and 1475 abnormal clips across the four f an IDs. F ollo wing the benchmark protocol, the CAE is trained from normal recordings only . In this one-class construction, abnormal clips are e v aluation-only and v alidation is formed solely from train-normal data. Model Int J Recongurable & Embedded Syst, V ol. 15, No. 3, No v ember 2026: 645–655 Evaluation Warning : The document was created with Spire.PDF for Python.
Int J Recongurable & Embedded Syst ISSN: 2089-4864 ❒ 649 tting uses Adam [ 24 ], mean squared reconstruction loss, a maximum of 50 epochs, batch size 32, and early stopping on v alidation loss with patience 8 and best-weight restora tion. W ithin each claim-bearing split, 20% of the train-normal partition is carv ed out as a shuf ed v alidation subset for early s topping only . After training, post-training quantization uses representati v e training data to e xport an INT8 TFLite encoder whose input and output types are both INT8 [ 25 ]. The e xported INT8 encoder , together with the x ed Mahalanobis artif acts µ , Σ − 1 , and τ , is synchronized with the MCU rmw are for the ra w-decision consistenc y check. 3.4. Repeated e v aluation pr otocol The main reported metrics are area under the curv e (A UC), pA UC@0.1 , and F1, with precis ion and recall used to compute the thresholded F1 summaries. Here pA UC@0.1 denotes the standardized partial recei v er operating characteristic A UC (R OC A UC) o v er the f alse-positi v e-rate range [0 , 0 . 1] . The mean mar gin reported for anchor selection is the held-out class-separation mar gin, computed as t he mean abnormal Mahalanobis score minus the mean normal Mahalanobis score and then a v eraged across repeated splits. The primary anchor study e v aluates f an/id 06/0 dB wi th repeated stratied shuf e split [ 26 ], n splits=5, and test size=0.25. W ithin the e v aluated f an IDs, id 06 serv es as the deplo yment anchor because the repeated of ine study identies it as the strongest ra w single-ID case with the lar gest mean Mahalanobis mar gin and the most stable anchor beha vior . F or each split, the encoder is trained on train-normal data and then held x ed for e v aluation. The Mahalanobis path is calibrated in the dequantized latent space of the e xported INT8 TFLite encoder: µ is the train-normal latent mean, Σ − 1 is the pseudoin v erse of the train-normal co v ariance after 1 × 10 − 4 I re gularization, and τ is the 95th percentile of the cor responding train-normal Mahalanobis scores. This label-free τ calibration uses train-norm al data o nl y; no abnormal clips, e v aluation labels, or x ed-batch samples enter threshold setting. The sensiti vity check in this study is deliberately limited to operating parameters that af f ect the e xported decision path or the of ine temporal polic y . The train-normal percentile threshold, the 1 × 10 − 4 I co v ariance re gularization, and the retained 2 - ou t -of-3 and 2-out-of-5 temporal rules were k ept e xplicit across the repeated e v aluations. A stricter 3-out -of-5 temporal v ariant w as additi onally check ed of ine as a sensiti vity v ariant b ut w as not promoted to the main operating-polic y summary because the paper’ s deplo yment claim is tied to ra w Mahalanobis thresholding rather than to temporal v oting. This compact sensiti vity scope preserv es reproducibility without turning the study into a broad h yperparameter -search benchmark. The retained Mahalanobis + 2-out -of-5 of ine operating polic y w as selected from repeated held-out comparisons and frozen before an y Stage A/B de vice tests. After each held-out subset is restored to deterministic lename order , temporal v oting is applied as ordered-subset smoothing rather than as contiguous streaming e vidence. The same repeated protocol is then applied independently to id 00, id 02, id 04, and id 06 in the all-ID study , yielding 20 e v aluated split ro ws in the aggre g ate summary . Repeated summaries are reported a s empirical mean and standard de viation, with R =5 for the id 06 anchor study and R =20 for the pooled all-ID study . 3.5. Embedded consistency-e v aluation setup The batch reference-de vice study directly e v aluates bounded consistenc y for the deplo yed detector on a x ed 20-sample id 06 test batch containing 10 normal and 10 abnormal clips. T o a v oid post hoc drift during de vice testing, this balanced id 06 batch w as frozen before Stage A/B e x ecution and then reused unchanged in both stages; no batch items were used in τ calibration or in the earlier of ine choice of the retained 2-out-of-5 operating polic y . Stage A and St age B both use one frozen id 06 deplo yment b undle consisting of the quantized encoder , synchronized µ , Σ − 1 , and τ , and t he same x ed 20-sample test batch. F or reproducibility , the retained b undle contains the INT8 TFLite enc oder , TFLite input/output scale and zero-point parameters, the Mahalanobis artif acts, the operating threshold, the x ed-batch lename list, generated C audio/header les, and Stage A/B comparison logs. In Stage A, the de vice input tensor is forced from the of f -de vice reference so that the controlled input q → latent → score → ra w decision path can be tested directly . In Stage B, the same x ed batch is re-e v aluated through the full canonical DSP path audio → DSP → input q → latent → score → ra w decision using pro v enance-clean embedded C audio headers generated from the public MIMII benchmark W A V les. This design separates an artif act-transfer consistenc y check from x ed-batch DSP-path retesting on the de vice. Stage B is a stored-sample batch replay; it does not include li v e mic ropho ne acquisition, I2S b uf fering, continuous streaming, latenc y measurement, or ener gy measurement. The de vice e vidence in this paper therefore co v ers ra w-decision reproduction under Stage A/B for that x ed b undle, not full on-de vice reproduction of the temporal polic y . Deployment-awar e acoustic anomaly detection on an embedded micr ocontr oller (T ruong Minh Duc) Evaluation Warning : The document was created with Spire.PDF for Python.
650 ❒ ISSN: 2089-4864 The embedded deplo yment tar get is ESP32-S3 (esp32-s3-de vki tc-1) through PlatformIO and Arduino. These embedded e xperim ents e v aluate x ed-batch reproducibi lity and static memory-footprint re porting rather than deplo yment rob ustness or production performance. Accordingly , this paper reports static b uild footprint and x ed-batch reference-de vice reproducibility , while lea ving rigorous latenc y and ener gy characterization to future w ork. 4. RESUL TS AND DISCUSSION 4.1. Str ongest single-ID anchor The primary depl o yment anchor is the f an/id 06/0 dB case. T able 1 reports the per -ID repeated-split screening summary used to choose this anchor from the e v aluated f an IDs. T able 1. Per -ID repeated-split ra w Mahalanobis summary for deplo yment-anchor selection ID A UC pA UC@0.1 F1 Mean mar gin Selection id 00 0.626 ± 0.051 0.544 ± 0.025 0.243 ± 0.067 0.726 Support id 02 0.721 ± 0.028 0.600 ± 0.026 0.394 ± 0.062 1.721 Support id 04 0.639 ± 0.036 0.530 ± 0.008 0.190 ± 0.034 0.448 Support id 06 0.965 ± 0.017 0.865 ± 0.054 0.844 ± 0.051 3.419 Anchor W ithin this bounded f an-only pool, id 06 combines the strongest ra w metrics with the lar gest mean Mahalanobis mar gin. The dedicated id 06 stability run used for Figure 2 is reported separately because it characterizes the selected anchor in more detail: the ra w Mahalanobis detector achie v ed 0.958 ± 0.015 A UC, 0.858 ± 0.034 pA UC@0.1 , and 0.815 ± 0.045 F1, with a split-wise threshold of 5.582 ± 0.154. T ogether , the per -ID screening summary and the dedicated anchor run pro vide the main of ine e vidence that id 06 is a suitable anchor for the subsequent x ed-batch de vice study . The weak er i d 00 and id 04 ro ws remain visible in the all-ID e vidence as limitations, so the selected anchor is positioned as the strongest bounded case ra ther than as proof of cross-ID rob ustness. Figure 2. Repeated-split stability summary for the selected f an/id 06/0 dB anchor The repeated all-ID study pro vides descripti v e e vidence about aggre g at e cross-ID v ariation and re- maining limitations be yond the primary anchor case. F or compactness, the repeated all-ID block of T able 2 reports the ra w Mahalanobis path together with the tw o temporal rules carried forw ard into the operating-polic y comparison, namely 2-out-of-3 and 2-out-of-5. Int J Recongurable & Embedded Syst, V ol. 15, No. 3, No v ember 2026: 645–655 Evaluation Warning : The document was created with Spire.PDF for Python.
Int J Recongurable & Embedded Syst ISSN: 2089-4864 ❒ 651 T able 2. Core of ine results supporting the proposed detector across the repeated all-ID study and the single-run score-f amily ablation Method A UC pA UC@0.1 F1 Repeated all-ID a g gr e gate (fan 0 dB) Raw Mahalanobis 0.738 ± 0.140 0.635 ± 0.139 0.418 ± 0.263 2-out-of-3 0 . 660 ± 0 . 124 0 . 573 ± 0 . 062 0 . 261 ± 0 . 238 2-out-of-5 0 . 673 ± 0 . 119 0 . 568 ± 0 . 049 0 . 337 ± 0 . 209 Scor e-family ablation (a g gr e gate single run) Mahalanobis 0.733 0.624 0.427 Latent L2 0.641 0.541 0.213 Reconstruction MSE 0.656 0.589 0.368 Note. The score-f amily ablation block is a single-run design-selection check, not repeated-split benchmark e vidence. Under the repeated-split protocol, the ra w Mahalanobis path produced the highest aggre g ate metrics among these reported detector v ariants. The 2-out-of-5 rule is retained as a pre-x ed of ine ordered-subset polic y x ed before de vice testing, not as a claim that temporal v oting impro v es on the ra w path. Because T able 2 is aggre g ate, the repeated all-ID study is interpret ed as descripti v e e vidence of cross-ID v ariability and present limits rather than as e vidence of rob ust deplo yment across IDs. 4.2. Anchor temporal stability The pre-x ed 2-out-of-5 temporal operating polic y is e xamined alongside the deplo yment study because it pro vides ordered-subset smoothing and a reproducible of ine rule, e v en though it is not the num erically strongest detector in the of ine tables. Under the same ordered-sequence protocol, the 2-out-of-5 rule yielded 0.828 ± 0.032 A UC, 0.651 ± 0.016 pA UC@0.1 , and 0.629 ± 0.043 F1. These v alues sho w that temporal beha vior remains sensiti v e to t he e v aluation protocol and operating objecti v e e v en on the primary single-ID anchor . Figure 2 summarizes repeated-split beha vior for this anchor . 4.3. Scor e-family ablation The score-f amily block of T able 2 supports selecting Mahalanobis within this pipeline because it gi v es the highest aggre g ate single-run metrics among the tested score f amilies. Because this ablation is report ed in a dif ferent e v aluation conte xt from the repeated-split results, it is used here as design-selection e vidence rather than as a broader benchmark claim. 4.4. Fixed-batch r efer ence-de vice consistency The x ed-batch study directly compares the of f-de vice reference and ESP32-S3 beha vior on a x ed id 06 test batch. Ra w agreement is reported here as the fraction of batch items for which the reference and de vice ra w decisions match; it is an implementation-consistenc y measure, not ground-truth accurac y . Score consistenc y and latent consistenc y are summarized by batch-a v eraged absolute score and latent errors. T able 3 consolidates the x ed 20-sample id 06 batch study with the static b uild footprint. T able 3. Fix ed-batch ESP32-S3 consistenc y and static b uild footprint Stage Ra w agree. Mean score err . Mean latent err . F ixed 20-sample id 06 consistency batc h Stage A ref. input 20/20 0.030 0.001 Stage B DSP replay 20/20 0.127 0.005 Static b uild footprint ESP32-S3 static b uild 28.7% RAM/17.2% Flash In Stage A, the ESP32-S3 w as dri v en by the of f -de vice reference tensor and achie v ed 20/20 ra w- decision agreement with mean score and latent errors of 0.030 and 0.001. In Stage B, the same x ed batch passed through the full DSP path and ag ain achie v ed 20/20 agreement, wi th mean score and latent errors of 0.127 and 0.005 . Relati v e to the anchor threshold scale reported abo v e (5.582 ± 0.154), the St age B score shift remains small on the same score axis, without implying unchanged ranking or calibration. The ESP32-S3 static b uild occupied 28.7% RAM and 17.2% Flash (94,136/327,680 RAM bytes; 575,317/3,342,336 Flash bytes). The reported memory gures are static b uild-footprint v alues; runtime heap/stack high-w ater marks, latenc y , and ener gy were not measured. These results therefore support only x ed-batch i d 06 ra w-decision agreement Deployment-awar e acoustic anomaly detection on an embedded micr ocontr oller (T ruong Minh Duc) Evaluation Warning : The document was created with Spire.PDF for Python.
652 ❒ ISSN: 2089-4864 with bounded mean score and latent drift across the e v aluated batch. Figure 3 separates the Stage A/Stage B summary from the ordered per -sample Stage B mar gins. Figure 3. Reference-de vice consistenc y summary for the x ed id 06 batch on ESP32-S3 4.5. Discussion The Stage A/Stage B contrast denes the safe deplo yment interpretation in this paper . Stage A sho ws near -numerical parity along the input q → latent → score → ra w decision path because the de vice is dri v en from the of f-de vice reference tensor . Stage B sho ws that reintroducing the canonical DSP front end changes intermediate scores and latents, yet the x ed id 06 batch still preserv es 20/20 ra w decisions. Across the tested detector v ariants in T able 2, the ra w Mahalanobis path gi v es the strongest aggre g ate metrics, whereas the pre-x ed 2-out-of-5 rule is retained only as an of ine ordered-subset polic y . The score-f amily ablation supports design selection within this pipeline, not a broader benchmark result, and the deplo yment result remains limited to x ed-batch id 06 ra w-decision agreement with bounded latent and score drift rather than cross-ID rob ustness, score-ranking in v ariance, calibration in v ariance, or eld-performance guarantees. 4.6. Embedded, r econgurable, and inter net of things implications F or industrial IoT monitoring, ra w-decisi on consistenc y is practically important because an edge node typically forw ards an alarm state, e v ent log, or maintenance trigger rather than a full benchmark curv e. A detector that preserv es aggre g ate A UC b ut changes thresholded decisions near τ may still produce dif ferent alarm/no-alarm beha vior after deplo yment. The Stage A/B protocol therefore acts as a x ed-batch deterministic ra w-decision consistenc y check before broader eld trials: it v eries whether the e xported INT8 encoder , synchronized Mahalanobis artif acts, and threshold logic preserv e the x ed reference decisions on the tar get embedded de vice. The same decomposition is also rele v ant to future hardw are-oriented and netw ork ed deplo yment studies. The pipeline separates feature e xtraction, quantized encoder infe rence, compact Mahalanobis scoring, and a comparator -style threshold decision. This block-le v el structure may help frame future FPGA, ASIC, or multi-node IoT studies, b ut the present paper does not e v aluate hardw are synthesis, distrib uted depl o yment, latenc y , ener gy , or security beha vior; instead, it reports bounded reference-de vice consistenc y e vidence for the e v aluated ESP32-S3 batch. 4.7. Limitations and thr eats to v alidity The e vidence should be read as a bounded case study . The x ed-batch de vice result uses one id 06 test batch and stored-sample replay through generated headers rather than li v e acquisition. It does not establish li v e streaming beha vior , runtime perfor mance, sec ure or distrib uted deplo yment, cross-ID rob ustness, calibration or Int J Recongurable & Embedded Syst, V ol. 15, No. 3, No v ember 2026: 645–655 Evaluation Warning : The document was created with Spire.PDF for Python.
Int J Recongurable & Embedded Syst ISSN: 2089-4864 ❒ 653 ranking in v ariance, or on-de vice reproduction of the temporal polic y . The all-ID ro ws e xpose weak er id 00 and id 04 beha vior; those ro ws are descripti v e w arnings rather than deplo yment e vidence. 5. CONCLUSION This study e v aluated a bounded embedded consistenc y-e v aluation question for a T in yML ASD pipeline on MIMII f an 0 dB using an ESP32-S3. The main nding is restricted to the x ed 20-sample f an/id 06/0 dB test batch. F or this batch, the e xported INT8 encoder and synchronized latent-space Mahalanobis artif acts achie v ed complete agreement in ra w decisions between the of f-de vice reference and the ESP32-S3. Stage A e x ercised the controlled artif act-transfer path by feeding of f-de vice reference tensors through the de vice inference/scoring path, while Stage B retested generated C audio headers as x ed stored-sample replay through the canonical DSP path. Both stages sho wed 20/20 ra w-decision agreement; bounded score and latent drift are reported only as mean diagnostics, not as e vidence of calibration or ranking in v ariance. W ithin this e vidence boundary , we conclude reference-de vice ra w-decision consistenc y for the e v aluated x ed f an/id 06/0 dB batch. The result is rele v ant to embedded and IoT -edge deplo yment studies because it checks whether x ed binary alarm decisions survi v e transfer to MCU-class e x ecution and e xposes a block-le v el decision path for future recongurable-hardw are in v estig ati ons. These results do not establish li v e microphone acquisition, continuous streaming, embedded reproduction of the of ine ordered-subset temporal-v oting polic y , broader calibration in v ariance, score-ranking in v ariance, cros s-ID rob ustness, or complete eld characterization. Future w ork should e xtend the same protocol to additional test batches, machine IDs, acoustic conditions, continuous streams, netw ork ed deplo yments, security-a w are inte gration, and reproducible runtime characterization. A CKNO WLEDGMENTS The authors ackno wledge institutional support from A VIS Lab, Posts and T elecommunications Institute of T echnology (PTIT), V ietnam. FUNDING INFORMA TION This research recei v ed no specic e xternal grant; institutional support w as pro vided by A VIS Lab, PTIT , V ietnam. A UTHOR CONTRIB UTIONS ST A TEMENT This journal uses the Contrib utor Roles T axonomy (CRediT) to recognize indi vidual author contrib u- tions, reduce authorship disputes, and f acilitate collaboration. Name of A uthor C M So V a F o I R D O E V i Su P Fu T ruong Minh Duc ✓ ✓ ✓ ✓ Nguyen Ngoc Minh ✓ ✓ ✓ T ruong Cao Dung ✓ ✓ ✓ C : C onceptualization I : I n v estig ation V i : V i sualization M : M ethodology R : R esources Su : Su pervision So : So ftw are D : D ata Curation P : P roject Administration V a : V a lidation O : Writing - O riginal Draft Fu : Fu nding Acquisition F o : F o rmal Analysis E : Writing - Re vie w & E diting CONFLICT OF INTEREST ST A TEMENT The authors declare no conicts of interest related to this w ork. D A T A A V AILABILITY The MIMII dataset used in this study is publicly a v ailable from Zenodo. Its DOI is 10.5281/zenodo. 3384388. The DCASE one-class benchmark protocol is described in the cited DCASE sources. The original Deployment-awar e acoustic anomaly detection on an embedded micr ocontr oller (T ruong Minh Duc) Evaluation Warning : The document was created with Spire.PDF for Python.
654 ❒ ISSN: 2089-4864 MIMII audio les are not r edistrib uted by the authors. Author -generated materials that support checking the reported x ed-batch reference-de vice result, including the id 06 batch lename list, threshold and scoring- conguration notes, summary tables, and Stage A/B comparison logs, are a v ailable from the corresponding author upon reasonable request. REFERENCES [1] H. Purohit et al. , “MIMII dataset: Sound dataset for malfunctioning industria l machine in v estig ation and inspection, ” in Pr oceed- ings of the Detection and Classication of Acoustic Scenes and Events 2019 W ork shop (DCASE2019) , 2019, pp. 209–213, doi: 10.33682/m76f-d618. [2] Y . K oizumi et al. , “Description and discussion on DCASE 2020 challenge task 2: Unsupervised anomalous sound detection for machine condition monitoring, ” in Pr oceedings of the 5th W orkshop on Detection and Classication of Acoust ic Scenes and Events (DCASE 2020) , 2020, pp. 81–85. [3] T .-T .-H. Le, A. A. Adiputra, J. Y un, and H. Kim, “ Anomaly detection in industrial machine sounds using high-frequenc y features and g ate recurrent unit netw orks, ” IEEE Access , v ol. 13, pp. 77165–77186, 2025, doi: 10.1109/A CCESS.2025.3565812. [4] K. W ilkinghof f, “Self-supervised learning for anomalous sound detection, ” i n ICASSP 2024 - 2024 IEEE International Confer ence on Acoustics, Speec h and Signal Pr ocessing (ICASSP) , 2024, pp. 276–280, doi: 10.1109/ICASSP48485.2024.10447156. [5] A. Jiang et al. , “ Adapti v e prototype learning for anomalous sound detection with partially kno wn attrib utes, ” in ICASSP 2025 - 2025 IEEE International Confer ence on Acoustics, Speec h and Signal Pr ocessing (ICASSP) , 2025, pp. 1–5, doi: 10.1109/ICASSP49660.2025.10889514. [6] S. Kilickaya et al. , “ Audio-based anomaly detection in industrial machines using deep one-class support v ector data description, ” in 2025 IEEE Symposium on Computational Intellig ence on Engineering/Cyber Physical Systems Companion (CIES Companion) , pp. 1–5, 2025, doi: 10.1109/CIESCompanion65073.2025.11010815. [7] A. Garof alo, M. Rusci, F . Conti, D. Rossi, and L. Benini, “PULP-NN: A computing library for quantized neural netw ork inference at the edge on RISC-V base d parallel ultra lo w po wer clusters, ” in 2019 26th IEEE International Confer ence on Electr onics, Cir cuits and Systems (ICECS) , 2019, pp. 33–36, doi: 10.1109/ICECS46596.2019.8965067. [8] J. Moosmann, M. Giordano, C. V ogt, and M. Magno, “T in yissimoY OLO: A quantized, lo w-memory footprint, T in yML object detection netw ork for lo w po wer microcontrollers, ” in 2023 IEEE 5th International Confer ence on Articial Intellig ence Cir cuits and Systems (AICAS) , 2023, pp. 1–5, doi: 10.1109/AICAS57966.2023.10168657. [9] L. Ruf f et al. , “Deep one-class classication, ” in Pr oceedings of the 35th International Confer ence on Mac hine Learning , 2018, pp. 4393–4402. [10] M.-H. Nguyen, D.-Q. Nguyen, D.-Q. Nguyen, C.-N. Pham, D. Bui, and H.-D. Han, “Deep con v olutional v ariational autoencoder for anomalous sound detection, ” in 2020 IEEE Eighth International Confer ence on Communications and Electr onics (ICCE) , 2021, pp. 313–318, doi: 10.1109/ICCE48956.2021.9352085. [11] W . Dong, F . Guo, and T . Cheng, “Machine anomalous sound detection based on a multi-dimensional feature e xtraction self-encoder model, ” in 2024 5th International Confer ence on Computer Engineering and Application (ICCEA) , 2024, pp. 1165–1169, doi: 10.1109/ICCEA62105.2024.10604121. [12] T . Nishida, K. Dohi , T . Endo, M. Y amamoto, and Y . Ka w aguchi, “ Anomalous sound detection based on machine acti vity detection, ” in 2022 30th Eur opean Si gnal Pr oce ssing Confer e nce (EUSIPCO) , 2022, pp. 269–273, doi: 10.23919/EUSIPCO55093.2022.9909901. [13] G. W ichern, A. Chakrabarty , Z.-Q. W ang, and J. L. Roux, “ Anomalous sound detection using at tenti v e neural processes, ” in 2021 IEEE W orkshop on Applications of Signal Pr ocessing to A udio and Acoustics (W ASP AA) , 2021, pp. 186–190, doi: 10.1109/W AS- P AA52581.2021.9632762. [14] F . G. Germain, G. W ichern, and J. Le Roux, “Hyperbolic unsupervised anomalous sound detection, ” in 2023 IEEE W orkshop on Applications of Signal Pr ocessing to A udio and Acoustics (W ASP AA) , 2023, pp. 1–5, doi: 10.1109/W ASP AA58266.2023.10248092. [15] H. Zhang et al. , “First-shot unsupervised anomalous sound detection with unkno wn anomal ies estimated by metadata-assisted audio generation, ” in ICASSP 2024 - 2024 IEEE International Confer ence on Acoustics, Speec h and Signal Pr ocessing (ICASSP) , 2024, pp. 1271–1275, doi: 10.1109/ICASSP48485.2024.10448451. [16] Y . Liu, J. Guan, Q. Zhu, and W . W ang, “ Anomalous sound detection using spectral-temporal information fusion, ” in ICASSP 2022 - 2022 IEEE International C onfer ence on Acoustics, Speec h and Signal Pr ocessing (ICASSP) , 2022, pp. 816–820, doi: 10.1109/ICASSP43922.2022.9747868. [17] H. Chen, Y . Song, L.-R. Dai, I. McLoughlin, and L. Liu, “Self-supervised representation learning for unsupervised anomalous sound detection under domain shift, ” in ICASSP 2022 - 2022 IEEE International Confer ence on Acoustics, Speec h and Signal Pr ocessing (ICASSP) , 2022, pp. 471–475, doi: 10.1109/ICASSP43922.2022.9747863. [18] X.-M. Zeng et al. , “Joint generati v e-contrasti v e representation learning for anomalous sound detection, ” in ICASSP 2023 - 2023 IEEE International Confer ence on Acoustics, Speec h and Signal Pr ocessing (ICASSP) , 2023, pp. 1–5, doi: 10.1109/ICASSP49357.2023.10095568. [19] H. Lan, Q. Zhu, J. Guan, Y . W ei, and W . W ang, “Hierarchical metadata information constrained self-supervised learning for anomalous sound detection under domain shift, ” in ICASSP 2024 - 2024 IEEE International Confer ence on Acoustics, Speec h and Signal Pr ocessing (ICASSP) , 2024, pp. 7670–7674, doi: 10.1109/ICASSP48485.2024.10446044. [20] S. Huang, Z. F ang, and L. He, “Noise supervised contrasti v e learning and feature-perturbed for anomalous sound detection, ” in ICASSP 2025 - 2025 IEEE International Confer ence on Acoustics, Speec h and Signal Pr ocessing (ICASSP) , 2025, pp. 1–5, doi: 10.1109/ICASSP49660.2025.10888995. [21] R. Da vid et al. , “T ensorFlo w Lite Micro: Embedded machine learning for T in yML systems, ” Pr oceedings of Mac hine Learning and Systems 3 (MLSys 2021) , 2021, pp. 800–811. [22] J. Moosmann, H. M ¨ uller , N. Zimmerman, G. Rutishauser , L. Benini, and M. Magno, “Fle xible and fully quantized lightweight T in yis- simoY OLO for ultra-lo w-po wer edge systems, ” IEEE Access , v ol. 12, pp. 75093–75107, 2024, doi: 10.1109/A CCESS.2024.3404878. [23] B. McFee et al. , “librosa: Audio and music signal analysis in Python, ” in Pr oceedings of the 14th Pyt hon in Science Confer ence Int J Recongurable & Embedded Syst, V ol. 15, No. 3, No v ember 2026: 645–655 Evaluation Warning : The document was created with Spire.PDF for Python.