Inter national J our nal of Electrical and Computer Engineering (IJECE) V ol. 16, No. 5, October 2026, pp. 2483 ∼ 2493 ISSN: 2088-8708, DOI: 10.11591/ijece.v16i5.pp2483-2493 ❒ 2483 Label-fr ee acoustic monitoring of honeybee swarming: An unsuper vised online lear ning appr oach Abdelmadjid Guessoum Graba 1 , Djoher Dalila Graba 2 1 Communication Netw orks, Architectures and Multimedia Laboratory , Djillali Liabes Uni v ersity , Sidi-Bel-Abbes, Algeria 2 Ev olutionary Engineering and Distrib uted Information Systems Laboratory , Djillali Liabes Uni v ersity , Sidi-Bel-Abbes, Algeria Article Inf o Article history: Recei v ed Apr 23, 2026 Re vised May 22, 2026 Accepted Jul 22, 2026 K eyw ords: Acoustic monitoring Anomaly detection Bayesian changepoint Circadian modelling Sw arm detection ABSTRA CT Colon y losses caused by hone ybee sw arming remain a major operational chal- lenge because departure occurs within minutes, although acoustic changes be gin tens of minutes earlier and could allo w timely interv ention if detected reliably . Current detection systems miss this windo w because the circadian rh ythm of in- di vidual hi v e acoustics is not modelled, making it impossible to separate genuine pre-sw arming drift from normal day-to-night spectral v ariation. Recursi v e least squares is use d to estimate a colon y-specic circadian baseline, Mahalanobis distance scoring is applied to quantify de viations, and Bayesian online change- point detection (BOCPD) accumulate s Bayesian e vidence of a re gime shift, with a threshold deri v ed from w armup data guaranteeing a controlled f alse alarm rate without labelled recordings and re g ardless of bee race, season, or micro- phone placement. A controlled simulation w as used for e v aluation, yielding 3 . 5 × higher precisi on than the best label-free baseline, sw arm anticipation e x- ceeding 25 minutes before departure, and a f alse alarm rate held at the nominal 5% tar get. The lo w per -frame cost and memory footprint mak e this algorithm deplo yable on resource-constrained embedded de vices, enabling continuous au- tonomous hi v e monitoring without e xpert supervision. This is an open access article under the CC BY -SA license . Corresponding A uthor: Abdelmadjid Guessoum Graba Communication Netw orks, Architectures and Multimedia Laboratory , Djillali Liabes Uni v ersity Sidi-Bel-Abbes, Algeria Email: abdelmadjid.graba@uni v-sba.dz 1. INTR ODUCTION Hone ybee colonies ( Apis mellifer a ) are responsible for pollinating approximately one third of global food crops. The annual economi c v alue of this service has been estimated at o v er 150 billion euros [1]. Bee- k eeping operations w orldwide are e xposed to a recurring threat kno wn as sw arming. During a sw arm e v ent, the queen departs with more than half the w ork er population. The remaining colon y loses its foraging capacity and, in most cases, its hone y yield for the season [2]. Sw arming is dif cult to manage because of its timing. Internal colon y preparation de v elops o v er se v eral days before departure occurs within minutes. W eekly inspections are not suf cient to intercept this sequence. At the scale of a professional apiary , the mismatch between visit fre- quenc y and e v ent speed results in direct economic loss. Acoustic monitoring systems address this problem. A sensor installed on each hi v e re gisters spectral changes that precede sw arming well before departure, pro viding time for beek eeper interv ention. In-hi v e acoustic recordings sho w that colon y sound has a measurable daily structure. Spectral ener gy and frequenc y content v ary with the l ight c ycle. This circadian pattern remains stable under normal hi v e condi- tions [3]. In the days before sw arming, this pattern changes progressi v ely . The dominant frequenc y decreases, J ournal homepage: http://ijece .iaescor e .com Evaluation Warning : The document was created with Spire.PDF for Python.
2484 ❒ ISSN: 2088-8708 broadband ener gy increases, and the frequenc y distrib ution narro ws. These changes occur simultaneously across multiple descri ptors. This multi v ariate progression is referred to as pre-sw arming drift. At sw arm depar - ture, the acoustic signal changes abruptly as thousands of bees e xit the hi v e within seconds. This progression creates tw o detection problems that must be addressed together . Normal day-to-night spectral transitions re- semble pre-sw arming drift and cause f alse alarms in det ectors that ignore circadian v ariation. Detection must also occur during the drift phase, not at departure, to preserv e a usable interv ention windo w . Both conditions must be satised simultaneously for acoustic monitoring to be operationally ef fecti v e. The empirical foundations of acoustic hi v e monitoring were established by Ferrari et al. [2]. Their study co v ered 270 hours of continuous recordings and documented a reproducible spectral sequence before sw arming. A progressi v e centroid descent w as observ ed, follo wed by a sharp jump at departure. Bromenshenk et al. [4] de v eloped embedded systems based on frequenc y analysis. These x ed-threshold methods share a fundamental structural limitation: an y threshold calibrated on specic e xperimental colonies f ails to generalise across races, en vironments, and seasons, and none accounts for the circadian rh ythm that shapes normal hi v e acoustics. Labelled sw arming recordings are required by all deep learning approaches to beehi v e audio cl assi- cation. Accurac y le v els of up to 99% ha v e been reported under controlled conditions [5]–[8]. Noise rob ust- ness [9], T in yML compatibilit y [10], feature selection [11], and do wnsampled spectrograms [12] ha v e each been e xplored as e xtensions. The labelling requirement, ho we v er , constitutes a structural obstacle. Sw arming occurs at most once per colon y per year . The publicly a v ailable datasets — OSBH [13], we4bee [14], Ur - B AN [15], and MSPB [16] — each contain only a fe w dozen annot ated e v ents, none of which includes times- tamped episodes usable for online detection benchmarking. Se v ere performance de gradation with changes in microphone placement has also been reported [17]. In short, supervised approaches f ace a data-a v ailability ceiling that is unlik ely to be o v ercome without multi-year , multi-colon y eld campaigns. Hi v e monitoring without labelled data has been approached as an unsupervised anomaly detection problem by se v eral authors. An autoencoder architecture w as used by Libal and Biernacki [18] to distinguish hone ybee types from audio without annotations. IoT -based acoustic classication w as studied by Zg ank [19]. V ibrational spectra were used by Ramse y et al. [20], who reported sw arming prediction up to 30 days ahead with accurac y e xceeding 90%. The CUSUM statistic [21] has been applied for label-free drift detection, b ut f alse alarms are systematically generated at da wn and dusk due to its uni v ariate formulation and its complete disre g ard for circadian structure. A probabilistic alternati v e is of fered by Bayesian online changepoint detection (BOCPD) [22], [23], in which e vidence is accumulated across successi v e observ ations and isolated perturba- tions are naturally suppressed. Standardised e v aluation prot o c ols ha v e been identied as missing from this literature, and label dependenc y has been agged as a persistent issue in recent comparati v e studi es [24], [25]. In summary , no e xisting unsupervised method jointly addresses circadian modelling and sequential Bayesian changepoint detection with a statistically guaranteed f alse alarm rate. Three concurrent g aps are identied in the literature. First, no e xisting approach models the circa- dian rh ythm e xplicitly before anomaly detection. This generates systematic f alse alarms at da wn and dusk. Second, label-free methods operate on uni v ariate descriptors without statistical guarantees on the f alse alarm rate. Third, supervised methods cannot be applied without annotated sw arming recordings. No prior w ork has applied BOCPD to acoustic hi v e monitoring. No e xisting system combines adapti v e circadian modelling with sequential probabilistic changepoint detection in a single online frame w ork. The algorithm proposed in this paper addresses all three g aps jointly . 2. METHOD The proposed s w arming detection pipeline is sho wn in Figure 1. F our processing blocks are connected in sequence. Each block addresses one specic property of sw arming acoustics: spectral descriptors capture the multi v ariate signal, circadian modelling remo v es day-to-night v ariation, Mahalanobis scoring quanties joint de viation, and BOCPD accumulates Bayesian e vidence of a re gime change. A fully online, label-free algorithm is described in algorithm 1. Notation is dened in T able 1. The system operates in tw o sequential phases. During the w armup phase, model parameters con v er ge and threshold γ is calibrated from the empirical distrib ution of P C P ( t ) under normal conditions. During the monitoring phase, γ is applied to each incoming fram e for real-time detection. At each frame t , the algorithm e xtracts a v ector of four spectral descriptors x ( t ) ∈ R 4 (line 3). These descriptors capture the dif ferent dimensions of the collecti v e acoustic beha viour of the colon y . During Int J Elec & Comp Eng, V ol. 16, No. 5, October 2026: 2483-2493 Evaluation Warning : The document was created with Spire.PDF for Python.
Int J Elec & Comp Eng ISSN: 2088-8708 ❒ 2485 pre-sw arming drift, the spectral centroid C ( t ) and dominant frequenc y P F ( t ) tend to decrease, the bandwidth B W ( t ) narro ws, while the ener gy E ( t ) progressi v ely increases. The coherence of these simultaneous v ariations constitutes a rob ust signature of the phenomenon, e xploited by the Mahalanobis distance at line 11. This minimal set is suf cient: each descriptor tar gets a distinct documented dimension of pre-sw arming drift, and d =4 bounds the in v ersion cost at O (64) operations per frame. Figure 1. Proposed sw arming detection pipeline T able 1. Summary of notation Symbol Dim. Description φ ( t ) R 5 Circadian harmonic basis θ RLS R 5 × 4 Circadian model parameters P R 5 × 5 RLS co v ariance matrix K R 5 RLS g ain v ector Σ ( t ) R 4 × 4 Adapti v e residual co v ariance D ( t ) scalar Mahalanobis anomaly score P C P ( t ) [0 , 1] Posterior changepoint probability λ, λ R LS scalars EMA and RLS for getting f actors H scalar BOCPD hazard rate T o pre v ent natural day/night v ariations from being interpreted as anomalies, the algorithm e xplici tly models the circadian beha viour of the colon y via the harmonic basis dened at line 4: φ ( t ) = 1 , cos 2 π t T , sin 2 π t T , cos 4 π t T , sin 4 π t T ⊤ (1) where T is the circadian period e xpressed in frames. The RLS update (lines 5–9) ts θ RLS online, producing the circadian prediction µ circ ( t ) = φ ( t ) ⊤ θ RLS at each step. At line 6, the RLS update is sus pended whene v er S ( t − 1)=1 : admitting anomalous frames into the parameter update w ould bias θ RLS to w ard the disturbed re gime, de grading the residual and eroding future detection sensiti vity . Once this model is learned, the residual (line 10) r ( t ) = x ( t ) − φ ( t ) ⊤ θ RLS contains only the abnormal v ariations. Its normal v ariability is modelled by an adapti v e co v ariance (line 8), updated with the pre vious residual r ( t − 1) , enabling the computation of the multi v ariate anomaly score (line 11): D ( t ) = r ( t ) ⊤ ( Σ ( t ) + ε I ) − 1 r ( t ) (2) This measure accounts for correlations between descriptors and ef fecti v ely captures the joint pre-sw arming drift. Label-fr ee acoustic monitoring of hone ybee swarming ... (Abdelmadjid Guessoum Gr aba) Evaluation Warning : The document was created with Spire.PDF for Python.
2486 ❒ ISSN: 2088-8708 T o inte grate the temporal dim ension of this drift, the s core D ( t ) is then analysed by the BOCPD module (line 12) [22], which estimates the posterior probability of a re gime change at instant t : P C P ( t ) = P ( r t = 0 | D 1: t ) (3) This mechanism distinguishes a persistent pre-sw arming drift from a punctual uctuation due to en vironmental noise. During the warmup phase, v alues of P C P ( t ) are accumulated in b uf fer W (line 14); once the distrib ution stabilises o v er at least 3 T frames (lines 16–17), the adapti v e threshold is learned: γ = p ercen tile( W , 100(1 − α )) (4) This threshold guarantees a f alse alarm rate controlled at le v el α without an y e xternal calibration. The nal decision is tak en at line 19 according to: S ( t ) = 1 [ warmup ∧ P C P ( t ) ≥ γ ] (5) An alarm is triggered only when the acoustic re gime has shifted persistently enough for BOCPD to assign a changepoint probability abo v e γ . Per -frame computational cost is O ( R max + d 3 + p 2 ) ≈ O (200) elementary operations with d =4 , p =5 , R max =100 . Memory stays belo w 500 oats. This theoretical analysis indicates compatibility with lo w-cost microcontrollers. Empirical benchmarking on hardw are is proposed as a priority ne xt step. Fi v e parameters go v ern the algorithm. λ R LS = 0 . 995 w as assigned to pre v ent the circadian base- line from absorbing pre-sw arming drift. λ = 0 . 05 gi v es the co v ariance estimator a half-life of 14 frames. H = 1 / 500 places a prior of one re gime change per circadian c ycle on the BOCPD model. α = 0 . 05 x es the tolerated f alse alarm rate at an operationally acceptable le v el. n min = 3 T delays surv eillance until one full day-night c ycle has been observ ed. At each ne w installation, γ is estimated from the colon y’ s o wn w armup data. No manual reconguration is needed. Sensiti vity of results to these v alues is e xamined in Experiment 4 (Section 3.4.). Algorithm 1. Adapti v e Bayesian pre-sw arming detection Input: A ( t ) , λ, λ RLS , H , α, ϵ Output: P C P ( t ) , S ( t ) 1: Init.: Σ ← I 4 , θ RLS ← 0 , P ← 10 5 I 5 , W ← ∅ , warmup ← F alse 2: f or each frame t do 3: x ( t ) ← [ C , B W , E , P F ] ⊤ 4: φ ( t ) ← [1 , cos 2 π t T , sin 2 π t T , . . . ] ⊤ 5: K ← P φ ( t ) λ RLS + φ ( t ) ⊤ P φ ( t ) − 1 6: if S ( t − 1) = 0 then 7: θ RLS ← θ RLS + K x ( t ) − φ ( t ) ⊤ θ RLS 8: Σ ← (1 − λ ) Σ + λ r ( t − 1) r ( t − 1) ⊤ 9: end if 10: r ( t ) ← x ( t ) − φ ( t ) ⊤ θ RLS 11: D ( t ) ← r ( t ) ⊤ Σ + ϵ I − 1 r ( t ) 12: Apply BOCPD on D ( t ) → P C P ( t ) 13: if ¬ warmup then 14: W ← W ∪ { P C P ( t ) } 15: end if 16: if | W | ≥ 3 T and stable ( W ) and ¬ warmup then 17: γ ← p ct( W , 100(1 − α )) ; warmup ← T rue 18: end if 19: S ( t ) ← 1 warmup ∧ P C P ( t ) ≥ γ 20: end f or 3. RESUL TS AND DISCUSSION Since no lar ge annotated corpus of hi v e recordings currently includes s w arming e v ents, we v alidate the algorithm on a controlled simulation — a standard recourse when the phenomenon of interest is both rare and dif cult to label in the eld. T o reect the actual acoustic dynamics of sw arming, the synthetic signal is con- structed in three successi v e phases: a normal baseline dri v en by the learned circadian model, a pre-sw arming period characterised by a progressi v e linear pre-sw arming drift ( δ pre = [ − 50 , − 12 , +0 . 025 , − 40] Hz/RMS), Int J Elec & Comp Eng, V ol. 16, No. 5, October 2026: 2483-2493 Evaluation Warning : The document was created with Spire.PDF for Python.
Int J Elec & Comp Eng ISSN: 2088-8708 ❒ 2487 and a sw arming e v ent modelled as a sudden jump ( δ sw arm = [+220 , +55 , +0 . 28 , +175] Hz/RMS). Indepen- dent Gaussian noise with v ariance diag ( σ 2 ) , σ = [15 , 6 , 0 . 008 , 12] , is superimposed at each frame; v alues for C , B W , and P F are e xpressed in Hz, and E in normalised RMS units. Both reference methods recei v e the same circadian model θ R LS learned online by our algorithm, not the true model θ true . This f air comparison — a realistic deplo yment condition where no method benets from a perfect acoustic model — isolates the specic contrib uti on of the BOCPD frame w ork. Nominal parameters are λ =0 . 05 , λ R LS =0 . 995 , H =1 / 500 , α =0 . 05 , n min =3 T . F alse positi v es are counted per continuous alarm episode, and all results are a v eraged o v er 200 independent episodes. 3.1. Experiment 1: con v er gence of the cir cadian model The rst e xperiment e v aluates the con v er gence speed and quality of the harmonic circadian model learned by RLS. This model is the direct response to the circadian problem identied in Section 1: it must rapidly learn the day-night rh ythm of the colon y so that the residual r ( t ) contains only genuine anomali es. T able 2 presents the e v olution of the RMSE between the prediction µ cir c ( t ) and the true beha viour µ cir c,tr ue ( t ) o v er se v en days. The criterion n min = 3 T ensures that the system has observ ed the colon y at all hours of the day before triggering surv eillance. Figure 2 sho ws the complete con v er gence curv e on a log scale; the RMSE starts at 9.44 Hz (no prior kno wledge), f alls to 0.39 Hz after one full circadian c ycle (Day 1), rises transiently at Day 2 during 12h harmonic renement, and stabilises at Day 3 ( n min =3 T , red v ert ical line), which marks the end of the w armup phase and the onset of surv eillance. T able 2. RMSE of the circadian RLS model at k e y instants. Elapsed time RMSE (Hz) Interpretation Start ( t = 0 ) 9.44 θ R LS = 0 , no prior kno wledge — maximum error 6h equi v alent 1.22 First signicant harmonic corrections Day 1 (1 full c ycle) 0.39 Functional model — error < 0 . 15% of dynamic range Day 2 1.62 Second-order harmonic adjustment — normal RLS beha viour Day 3 ( n min = 3 T ) ⋆ 1.00 Con v er gence established — stationarity criterion met ✓ Day 7 1.97 Residual uctuations due to measurement noise σ C = 15 Hz Figure 2. RLS model RMSE o v er time (Hz, log scale). Red line: w armup end at n min = 3 T Se v eral observ ations follo w . First, con v er gence is remarkably f ast: at the end of the rst circ adian c ycle (Day 1), the RMSE f alls to 0.39 Hz, less than 0.15% of the spectral centroid dynamic range, without an y prior data. Second, the slight increase at Day 2 reects a characteristic RLS beha viour: once the fundamental 24h component is well estimated, the system renes the 12h harmonics, introducing a transient oscillation before deniti v e stabilisation. Third, residual uctuations around 2 Hz be yond Day 3 represent the theoretical lo wer bound of the error for a second-order harmonic model subject to Gaussian noise σ C = 15 Hz. These results empirically justify the choice of n min = 3 T . 3.2. Experiment 2: system calibration The statistical f alse alarm guara ntee w as assessed o v er 200 purely normal episodes. The empiri cal distrib ution of D ( t ) does not follo w the theoretical χ 2 (4) la w: E [ D ]=2 . 81 w as measured instead of the e x- pected 4.00. This bias originates from the EMA w arm-start initialisation of Σ . The BOCPD inference is not Label-fr ee acoustic monitoring of hone ybee swarming ... (Abdelmadjid Guessoum Gr aba) Evaluation Warning : The document was created with Spire.PDF for Python.
2488 ❒ ISSN: 2088-8708 af fected by this discrepanc y , since γ is re-estimat ed from w armup data at each deplo yment and absorbs the bias empirically . An ef fecti v e f alse alarm rate of 0 . 0475 ± 0 . 0141 w as obtained for α =0 . 05 , corresponding to a relati v e de viation of 5.0% from the nominal le v el. Figure 3 reports the calibrati on results. Fi gure 3(a) displays the density of per -episode f alse alarm rates, with the empirical mean of 0.0475 f alling close to the nominal α =0 . 05 . Figure 3(b) sho ws the ± 1 σ band around the mean, conrming that γ compensates the EMA bias without requiring e xplicit correction. (a) (b) Figure 3. System calibration o v er 200 purely normal episodes (a) f alse alarm rate density; black: observ ed mean; red dashed: α =0 . 05 and (b) de viation from nominal α with ± 1 σ band 3.3. Experiment 3: detection perf ormance This e xperiment constitutes the central e v aluation of the article. T able 3 presents the results o v er 200 independent episodes. All methods use θ RLS learned online; f alse positi v es are counted per continuous alarm episode; n TP denotes the number of episodes yielding at least one alarm before t c ; recall is computed per episode as the fraction of pre-sw arming frames correctly alarmed, then a v eraged o v er all 200 episodes; the delay is computed on n TP episodes only . T able 3. Detection performance o v er 200 independent episodes (mean ± std). † No alarm w as triggered ( n T P = n F P = 0 ): Precision, Recall, and F 1 are all reported as N/A since the detector is entirely silent. ‡ Anticipation delays are statistically indistinguishable (tw o-sided W ilcoxon signed-rank test, p =0 . 70 , n =152 ) Method Precision Recall F 1 -score Delay ( n T P ) Fix ed threshold χ 2 (4) N / A † N / A † N / A † — ( n T P = 0 / 200 ) CUSUM centroid 0 . 02 ± 0 . 00 1 . 00 ± 0 . 00 0 . 04 ± 0 . 00 25 . 7 ± 0 . 4 min ‡ ( n T P = 200 ) Our method (Alg . 1) 0 . 07 ± 0 . 05 0 . 76 ± 0 . 43 0 . 13 ± 0 . 08 25 . 5 ± 2 . 0 min ‡ ( n T P = 152 / 200 ) Figure 4 sho ws one complete post-w armup episode. Figure 4(a) plots D ( t ) o v er the three acoustic phases. Blue points co v er frames 0–1000. V alues uctuate near E [ D ] ≈ 2 . 81 , belo w the theoretical reference of 4.00. Orange points co v er frames 1000–1250. A progressi v e upw ard drift is observ ed during this pre- sw arming phase. Red points appear after frame 1250. A sharp jump is recorded at sw arming onset. The instant t c is mark ed by a black dashe d line. Figure 4(b) plots P C P ( t ) . The curv e rises steadily and crosses γ before t c . The rst alarm is indicated by a blue dashed line. Figure 4(c) o v erlays S ( t ) on ground truth. Orange and red backgrounds mark the pre-sw arming and sw arming zones. Blue bars indicate acti v e alarms. All 59 alarms precede t c . Figure 5 reports aggre g ated metrics o v er 200 runs. Figure 5(a) compares precision and recall across all three methods. Precision collapses to near zero for both baselines. A precision g ain of 3 . 5 × o v er CUSUM is achie v ed by the proposed method at equi v alent recall. Figure 5(b) plots the p e r -episode F 1 distrib ution. A bimodal shape is observ ed. This bimodality is a structural property of BOCPD: episodes are either detected with high condence or missed entirely , with no intermediate re gime. Int J Elec & Comp Eng, V ol. 16, No. 5, October 2026: 2483-2493 Evaluation Warning : The document was created with Spire.PDF for Python.
Int J Elec & Comp Eng ISSN: 2088-8708 ❒ 2489 (a) (b) (c) Figure 4. Representati v e post-w armup episode (a) D ( t ) by phase (blue/orange/red), (b) P C P ( t ) crossing γ before t c , and (c) decision S ( t ) ; all alarms precede t c (a) (b) Figure 5. Detection performance o v er 200 episodes (a) precision, recall, F 1 (mean ± std) and (b) F 1 distrib ution per episode (boxplot) Label-fr ee acoustic monitoring of hone ybee swarming ... (Abdelmadjid Guessoum Gr aba) Evaluation Warning : The document was created with Spire.PDF for Python.
2490 ❒ ISSN: 2088-8708 3.4. Experiment 4: h yper parameter sensiti vity As sho wn in T able 4, thirteen congurations were tested in total. The F1-score ranges from 0.117 to 0.184 across all congurations, a spread of 0.067 points. No catastrophic collapse is observ ed for an y reasonable parameter v alue. λ R LS and H sho w lo w sensiti vity . F 1 v aries by at most 0.013 points for λ R LS and 0.019 points for H . Both parameters can be x ed from ph ysical ar guments without numerical tuning. A non-monotone prole is observ ed for λ . The e xtreme v alues outperform the intermediate v alue λ = 0 . 01 ( F 1 = 0 . 131 ). The nominal v alue λ = 0 . 05 f alls in a stable re gion ( F 1 = 0 . 136 ± 0 . 082 ). The f alse alarm le v el α is the most operationally signicant parameter . The highest mean F 1 of 0.169 is obtained at α = 0 . 02 , b ut v ariance is mark edly ele v ated (std = 0.198). At α = 0 . 10 , F 1 f alls to 0 . 117 ± 0 . 048 . This parameter encodes the practitioner’ s f alse alarm b udget. It is not a free h yperparameter — it should be chosen operationally , as a signicance le v el is selected in a statistical test. The nominal conguration produces stable, representati v e performance rather than an o v ertted optimum. T able 4. Sensiti vity analysis — F 1 -score (mean ± std, 200 episodes). ⋆ denotes the nominal v alue P arameter V alue F 1 (mean ± std) λ 0.001 0 . 1 63 ± 0 . 093 0.01 0 . 1 31 ± 0 . 072 0.05 ⋆ 0 . 136 ± 0 . 082 0.10 0 . 1 84 ± 0 . 123 λ R LS 0.990 0 . 1 54 ± 0 . 078 0,995 ⋆ 0 , 141 ± 0 . 085 0.999 0 . 1 48 ± 0 . 096 H 1 / 200 0 . 134 ± 0 . 067 1 / 500 ⋆ 0 . 138 ± 0 . 084 1 / 1000 0 . 153 ± 0 . 103 α 0.02 0 . 1 69 ± 0 . 198 0.05 ⋆ 0 . 135 ± 0 . 081 0.10 0 . 1 17 ± 0 . 048 3.5. Discussion Three methods are compared in T able 3. The x ed threshold produces no detections under f air condi- tions ( n T P = n F P = 0 / 200 ). W ithout access to the true circadian model, D ( t ) ne v er reaches χ 2 (4) = 9 . 49 . Precision, Recall, and F 1 are undened for this baseline. CUSUM res on e v ery episode. A ratio of 49 f alse alarms per true detection is recorded. Continuous surv eillance is operationally impractical at this rate. The proposed algorithm achie v es F 1 = 0 . 13 ± 0 . 08 , a f actor of 3.25 o v er CUSUM. A precision g ain of 3 . 5 × is obtained at equi v alent recall. An anticipation delay of 25 . 5 ± 2 . 0 min is measured. This v alue is statistically indistinguishable from CUSUM ( 25 . 7 ± 0 . 4 min, tw o-sided W ilcoxon signed-rank test, p = 0 . 70 , n = 152 ). The impro v ement is therefore in reliability rather than speed: fe wer unjustied alerts are produced without an y loss in anticipation time. These results distinguish the proposed method from e xisting approaches. No prior calibration is required, unlik e Ferrari et al. [2]. No annotated sw arming e xamples are needed, unlik e supervised approaches [5], [7]. T o the best of t h e authors’ kno wledge, the proposed method is the rs t to e x- plicitly model the circadian rh ythm before anomaly detection, eliminating the main source of f alse alarms in nai v e approaches. Deep learning baselines such as autoencoders are architecturally incompatible with the strict online, single-pass, and embedded constraints of the proposed system; their training requirement and memory footprint preclude direct comparison in this deplo yment conte xt. Such comparison is deferred to a future of ine e v aluation. Connected beehi v e monitoring systems represent a direct application tar get for the propose d algo- rithm. Detection runs in real time on a lo w-cost microcontroller paired with a single in-hi v e microphone. Memory occupation stays belo w 500 oating-point v alues. Per -frame cost is approximately 200 elementary operations. At each ne w installation, the detection threshold is computed from the rst three days of colon y recordings. No sw arming e v ents and no manual setup are required. A mean anticipation windo w of 25 minutes is a v ailable to the beek eeper before sw arm departure. At α = 5% , one f alse alert per twenty sw arming sea- sons per hi v e is e xpected. W eekly inspection schedules are replaced by continuous autonomous surv eillance. Dependence on e xpert a v ailability is reduced. T imely interv ention becomes possible across tens of hi v es simul- taneously . Precision apiculture platforms aimed at reducing unplanned colon y losses are a natural deplo yment conte xt for this system. Int J Elec & Comp Eng, V ol. 16, No. 5, October 2026: 2483-2493 Evaluation Warning : The document was created with Spire.PDF for Python.
Int J Elec & Comp Eng ISSN: 2088-8708 ❒ 2491 Real-w orld deplo yment raises conditions that the simulation does not co v er . Inter -race acoustic v ari - ability , seasonal amplitude shifts, and microphone positioning artef acts are absent from the synthetic sig- nal [17]. A single sw arming trajectory is used per episode. Irre gular or interrupted pre-sw arming drifts, which are documented in real colonies, are not represented. T w o g aps are therefore identied in the current v ali- dation. The Gaussian noise model co v ers normal hi v e statistics b ut not real deplo yment v ariability . Episode di v ersity is limited to one trajectory type. Field recordings from multiple hi v es are needed to close these g aps before operational deplo yment. Three additional limitations are identied at the algorithmic le v el. The EMA bias ( E [ D ] = 2 . 81 ̸ = 4 ) is compensated empirically by γ . An NIG prior calibrated from w armup data w ould be theoretically more rigorous. The threshold γ is x ed at w armup and does not re-adapt to s lo w seasonal drifts. Periodic online recalibration is identied as a priority e xtension. The four spectral descriptors co v er the main documented dimensions of pre-sw arming drift. Richer features such as harmonicity or spectral entrop y are proposed as future e xtensions. F our ne xt steps are identied in order of priority: partial v alidation of the circadian modelling and f alse alarm calibration on real continuous hi v e audio; end-to-end sw arming detection on eld recordings with timestamped e v ents; periodic recalibration of γ ; embedded deplo yment conrmed by microcontroller benchmarks; and e xploration of h ybrid semi-supervised e xtensions once annotated sw arming recordings become a v ailable. 4. CONCLUSION A fully online, label-free algorithm for acoustic hone ybee sw arm detection w as introduced. No prior kno wledge of the monitored hi v e is required. Three components are coupled to address both: RLS-based circadian modelling, Mahalanobis distance scoring, and BOCPD. A w armup-calibrated threshold γ holds the f alse alarm rate at α without e xternal calibration. Performance w as e v aluated o v er 20 0 independent episodes under f air comparison condit ions. Ag ainst CUSUM, precision impro v ed by 3 . 5 × and F 1 by 3 . 25 × , both at an equi v alent anticipation delay of 25 . 5 ± 2 . 0 min. A f alse alarm rate of 4 . 75% w as measured at α =5% . The per - frame cost of O ( R max + d 3 + p 2 ) and memory belo w 500 oats conrm compatibility with lo w-cost embedded hardw are. All quantitati v e e v aluations were conducted on simulated data. The synthetic signal reproduces the statistical structure of pre-sw arming acoustics. Real-w orld v alidation on annotated eld recordings remains indispensable before operational deplo yment. The measured F 1 = 0 . 13 ± 0 . 08 is modest in absolute terms. It w as achie v ed without labelled e xamples, under a strict online single-pass constraint, and with a statistically guaranteed f alse alarm rate at the user -specied α — three simultaneous constraints absent from supervised comparators — and still yields 3 . 25 × higher F 1 than the best a v ailable label-free baseline. Field recordings with timestamped sw arming e v ents need to be collected for end-to-end v alidation. Periodic recalibration of γ is required to handle seasonal drift. Replacement of the EMA co v ariance estimator with an NIG prior calibrated from w armup data is proposed as a theoretically more rigorous alternati v e. Hybrid semi-supervised e xtensions become rele v ant once annotated sw arming recordings are a v ailable. FUNDING INFORMA TION Authors state no funding in v olv ed. A UTHOR CONTRIB UTIONS ST A TEMENT This journal uses the Contri b ut or 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 Abdelmadjid Guessoum Graba ✓ ✓ ✓ ✓ ✓ ✓ ✓ ✓ ✓ ✓ Djoher Dalila Graba ✓ ✓ ✓ ✓ ✓ ✓ ✓ 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 Label-fr ee acoustic monitoring of hone ybee swarming ... (Abdelmadjid Guessoum Gr aba) Evaluation Warning : The document was created with Spire.PDF for Python.
2492 ❒ ISSN: 2088-8708 CONFLICT OF INTEREST ST A TEMENT Authors state no conict of interest. D A T A A V AILABILITY The data supporting this study’ s ndings are a v ailable from the corresponding author , A GG, upon reasonable request. REFERENCES [1] N. Gallai, J.-M. Salles, J. Settele, and B. E. V aissi ` ere, “Economic v aluati on of the vulnerability of w orld agriculture confronted with pollinator decline, ” Ecolo gical Economics , v ol. 68, no. 3, pp. 810–821, 2009, doi: 10.1016/j.ecolecon.2008.06.014. [2] S . Ferrari, M. Silv a, M. Guarino, and D. Berckmans, “Monitoring of sw arming sounds in bee hi v es for early detection of the sw arming period, ” Computer s and Electr onics in Agricultur e , v ol. 64, no. 1, pp. 72–77, 2008, doi: 10.1016/j.compag.2008.05.010. [3] T . Cejro wski, J. Szyma ´ nski, and D. Logof ˘ atu, “Buzz-based recognition of the hone ybee colon y circadian rh ythm, ” Computer s and Electr onics in Agricultur e , v ol. 175, p. 105586, 2020, doi: 10.1016/j.compag.2020.310486. [4] J. J. Bromenshenk, C. B. Henderson, and B. K. Smith, “ Acoustic sensor s for monitoring hi v e health, ” American Bee J ournal , v ol. 149, pp. 1073–1078, 2009. [5] I. Nolasco, A. T erenzi, S. Cecchi, S. Orcioni, H. L. Bear , and E. Benetos, “ Audio-based identication of beehi v e states, ” in Pr oceedings IEEE International Confer ence Acoustics, Speec h and Signal Pr ocessing (ICASSP) , pp. 8256–8260, 2019, doi: 10.1109/ICASSP .2019.8682981. [6] V . K ulyukin, S. Mukherjee, and P . Amlathe, “T o w ard audio beehi v e monitoring: Deep learning vs. standard machine learning in classifying beehi v e audio samples, ” Applied Sciences , v ol. 8, no. 9, p. 1573, 2018, doi: 10.3390/app8091573. [7] K. Iqbal, B. Alabdullah, N. Al Muda wi, A. Alg arni, A. Jalal, and J. P ark, “Empirical analysis of hone ybees acoustics as biosensors signals for sw arm prediction in beehi v es, ” IEEE Access , v ol. 12, pp. 148405–148421, 2024, doi: 10.1109/A CCESS.2024.3471895. [8] A. Bricout, P . Lele ux, P . Acco, C. Escriba, J.-Y . F ourniols, G. Soto-Romero, and R. Floquet, “Bee T ogether: Joining bee audio datasets for hi v e e xtrapolation in AI-based monitoring, ” Sensor s , v ol. 24, no. 18, p. 6067, 2024, doi: 10.3390/s24186067. [9] I. Ardekani, S. Pour , and H. Sharifzadeh, “Real-time sw arming detection in hone ybees: Le v eraging audio signal processing and machine learning techniques, ” J ournal of the Acoustical Society of America , v ol. 154, no. 4, pp. A216–A217, 2023, doi: 10.1121/10.0023330. [10] A. De Simone, L. Barbisan, G. T urv ani, and F . Riente, “ Adv ancing beek eeping: IoT and T in yML for queen bee monitoring using au- dio signals, ” IEEE T r ansact ions on Instrumentation and Measur ement , v ol. 73, p. 2527309, 2024, doi: 10.1109/TIM.2024.3449981. [11] F . Rustam, M. Z. Sharif, W . Aljedaani, E. Lee, and I. Ashraf, “Bee detection in bee hi v es using selecti v e features from acoustic data, ” Multimedia T ools and Applications , v ol. 83, pp. 23269–23296, 2024, doi: 10.1007/s11042-023-15192-5. [12] A. Orlo wska, D. F ourer , J.-P . Ga vini, and D. Cassou-Ribehart, “Hone y bee queen presence detection from audio eld recordings using summarised spectrogram and con v olutional neural netw orks, ” in Lectur e Notes in Networks and Systems , v ol. 418 LNNS, 2022, pp. 83–92. doi: 10.1007/978-3-030-96308-8 8. [13] “Ope n Source Beehi v es Project, ” OSBH Bee Audio Dataset, 2018. https://zenodo.org/record/1321278 [14] “we 4bee Consortium, ” we4bee Open Hi v e Monitoring Dataset, 2020. https://www.we4bee.org [15] “UrB AN Project, ” Urban Bee Acoustic Monitoring Dataset, 2022. https://zenodo.org/records/11004682 [16] Y . Zhu, M. Abdollahi, S. Maucourt, N. Coallier , H. R. Guimar ˜ aes, P . Gio v enazzo, and T . H. F alk, “MSPB: A longitudinal multi- sensor dataset with phenotypic trait measurements from hone y bees, ” Scientic Data , v ol. 11, p. 860, 2024, doi: 10.1038/s41597- 024-03695-1. [17] P . Ksiazek, U. Libal, and A. Kr ´ ol-No w ak, “Spectral components of hone y bee sound signals recorded inside and outside the bee- hi v e: An e xplai nable machine learning approach to diurnal pattern recognition, ” Sensor s , v ol. 25, no. 14, p. 4424, 2025, doi: 10.3390/s25144424. [18] U. Libal and P . Biernacki, “Non-intrusi v e system for hone ybee recognition based on audio signals and maximum lik elihood classi- cation by autoencoder , ” Sensor s , v ol. 24, no. 16, p. 5389, 2024, doi: 10.3390/s24165389. [19] A . Zg ank, “Bee sw arm acti vity acoustic classication for an IoT -based f arm service, ” Sens or s , v ol. 20, no. 1, p. 21, 2019, doi: 10.3390/s20010021. [20] M.-T . Ramse y , M. Bencsik, M. I. Ne wton, M. Re yes, M. Pioz, D. Crauser , N. S. Delso, and Y . Le Conte, “The prediction of sw arming in hone ybee colonies using vibrational spectra, ” Scientic Reports , v ol. 10, p. 9798, 2020, doi: 10.1038/s41598-020-66115-5. [21] E. S. P age, “Continuous inspection schemes, ” Biometrika , v ol. 41, no. 1-2, pp. 100–115, 1954, doi: 10.1093/biomet/41.1-2.100. [22] R. P . Adams and D. J. C. MacKay , “Bayesian online changepoint detection, ” arXiv pr eprint arXiv:0710.3742 , 2007. [23] M. Altamirano, F .-X. Briol, and J. Knoblauch, “Rob ust and scalable Bayesian online changepoint detection, ” in Pr oceedings of the 40th International Confer ence on Mac hine Learning (ICML) , PMLR 202, pp. 642–663, 2023. [24] C. Uthof f, M. N. Homsi, and M. v on Ber gen, “ Acoustic and vibration monitoring of hone ybee colonies for beek eeping-rele v ant aspects of presence of queen bee and sw arming, ” Computer s and Electr onics in Agricultur e , v ol. 205, p. 107589, 2023, doi: 10.1016/j.compag.2022.107589. [25] A. Robles-Guerrero, T . Saucedo-Anaya, C. A. Guerrero-Mendez, S. G ´ omez-Jim ´ enez, and D. J. Na v arro-Sol ´ ıs, “Comparati v e study of machine learning models for bee colon y acoustic pattern classication on lo w computational resources, ” Sensor s , v ol. 23, no. 1, p. 460, 2023, doi: 10.3390/s23010460. Int J Elec & Comp Eng, V ol. 16, No. 5, October 2026: 2483-2493 Evaluation Warning : The document was created with Spire.PDF for Python.