Ný rannsóknarferli fyrir stoppuð villur í hendingarskerjum
Nýlega hafa dreifiraforknýtingar fengið stærri áherslu, þar með komið hefur verið aukin notkun af hleypum í dreifilínum. En mekanískar villur sem valda óheppileikum eru auknar, sem leggja mikið ábyrgð á viðkomu og viðhald lína.
Sviktandi mekanísk prestation er aðalorsök skakka í hleypum. Margir fræðimenn rannsaka stóra umfang switchgear viðferð, með aðferðum eins og spennubúnaðarpróf, vibrerunar merki greining, hleypa ferli próf, ultrahljóð prufa og geislavarmamæling. Motor - current - based switch status detection virkar fyrir circuit breakers og disconnectors en er minni notað fyrir load - switch drive - mechanism faults.
Rannsóknir á field - running load switches sýna að orkurögn motor current signals endurspegla switch status. Mekanískar vandamál (t.d., spring jamming, rust, gear jamming) í drive mechanism breyta current signal parameters (amplitude, duration, local peaks). Með fókusi á algengu energy - storage motor rust - jamming í kuststrefum, rannsóknin skoðar fault feature extraction and identification. Skref: 1) Greina motor current characteristics, splitta waveforms í 4 stigi, og meta hverja stig. 2) Hagnýta data - acquisition device fyrir current waveforms undir mismunandi skilyrðum. 3) Búa til recording - start algorithm, feature extraction, og fault - identification methods. 4) Staðfesta með tilraunum.
1 Orkurögn Motor Current Characteristics Analysis
Load switches nota venjulega DC motors til að drifa smelluspennur fyrir orku geymslu. Í lokum motor operation, rotor - output torque og speed standa í námi samröðun við stator - circuit current. Shunt - excited DC motor electromagnetic torque og voltage equations eru eftirfarandi:
Í Equation (1), T táknar electromagnetic torque; n táknar rotational speed; Ia táknar armature current; Ra táknar armature circuit resistance, sem er fast; Ea táknar winding induced electromotive force; U táknar terminal voltage; ΔU táknar contact voltage drop, sem er fast; ϕ táknar magnetic flux; Ce táknar electromotive force constant; og CT táknar torque coefficient. Eftir Equation (1), getum við leidd út:
Eftir Equation (2), þegar load current er litill, er demagnetizing áhrif armature reaction ekki mikil, svo magnetic flux er tekið sem fast, og electromagnetic torque er hlutfall af load current. Sem load current aukast, torque stígur en rotational speed tendast til að lækkva. En demagnetizing áhrif af hærra load current lækkar magnetic flux, sem myndi auka speed. Þessar mótsædandi áhrif gerast venjulega að lýsa svipuðu lækkun í shunt - excited motor speed. Mynd 1 sýnir típísku current waveform af DC energy - storage motor í verk, skipt í 4 stigi.Mynd 1 sýnir típísku current waveform af DC energy - storage motor í verk, skipt í 4 stigi.
Stigi 1 (t0)–(t1): Motor Start - up Stigi
Á tímapunkti t0, fær load switch closing signal frá dreifiterminal eining, sem skyndar control motor til að byrja með hleypa. Motor current stigur upp í start - up peak við (tst), svo hratt lækkar til að fara í stable operation.
Stigi 2 (t1)–(t2): Motor Stable Operation Stigi
Motor drives transmission gear til idle. Í þessu stigi keyrir motor stöðugt undir ljóna hleypa, með motor current amplitude við (Ia).
Stigi 3 (t2)–(t4): Spring Energy - Storage Stigi
Sem compression spring geymir orku, motor output torque stigrar gráðuverulega, nálgast hámarks gildi við (t3); á þeim tíma, motor current nálgast stigs hámarks gildi (Im). Síðan, motor output torque lækkar gráðuverulega.
Stigi 4 (t4)–(t5): Motor Current Interruption Stigi
Við (t4), ná compression spring limit switch, sem sker af straum til motor. Motor current fallar hratt til að ná 0 við (t5), og motor stoppar að keyra.
2 Fault Diagnosis for Energy - Storage Motor Jamming
2.1 Fault Simulation & Data Acquisition
Jamming fault test var simulaður á load switch frá electrical equipment factory (scenario í Fig. 2(a)). Eftir að draga saman switch, í motor stable operation og spring energy - storage stigi, var rocker notaður til að apply reverse locked - rotor forces til að simula gear/spring jamming. Custom current acquisition device (Fig. 2(b)) notaði ARM STM32F103 chip til að safna signals frá HSTS016L Hall current transformer (DC input: 0–30A). Þar sem opening signal manglar target current waveform, rannsóknin fokuserar á closing current signal.
2.2 Waveform Recording Start Algorithm
Frá Figure 1, effective signal waveform spans the time window t0 to t5, consisting of 4 stages with diverse current changes. Additionally, there are significant differences in signal amplitudes among different drive motors. Thus, using a simple current amplitude threshold as the start criterion for signal waveform recording is clearly inappropriate. Therefore, this study adopts the current change rate Kt within a unit time window and the mean value Imean as the start criteria to achieve effective waveform recording.Current change rate of the unit time window:
Mean current of each time window:
In Equations (3) and (4), I(i) represents the current signal; M is the number of sampling points in the unit time window; Δt is the time length of the unit time window, and Δt = 0.02s in this paper; I(1) is the first sampling point in the unit time window.
2.3 Time - Domain Feature Extraction
To identify the jamming fault of the energy - storage motor, the expressive information of the curve is extracted through some time - domain indicators. The kurtosis K can characterize the smoothness of the current signal; the root mean square Irms can characterize the average energy of the current signal; the skewness sk is a measure of the direction and degree of skewness of the statistical data distribution; the form factor sh and the peak factor C are used to characterize the extreme degree of the current peak in the waveform.
The Random Forest (RF)classification algorithm integrates multiple decision trees. Its output category is determined by the mode of individual decision - tree categories, featuring high accuracy, good tolerance for abnormal data, and low overfitting risk.
2.4 Random Forest Algorithm
RF relies on Bootstrap sampling (with - replacement sampling to form n sample sets from the original dataset) and Bagging voting. Bagging generates n training sets via Bootstrap, each training an independent weak classifier. Final decisions come from voting on weak - classifier outputs, with the majority vote as the result.
RF uses CART decision trees (binary trees splitting top - down from the root, minimizing the Gini index for splits, formula (5)). Per Liu Min et al. 100 decision trees optimize classification performance. Thus, this study uses 100 CART trees for the random forest.
3 Case Analysis
3.1 Feature Selection
The Gini index in the random forest is used to evaluate the importance of each feature. The results are shown in Figure 3, where the ordinate represents the proportional coefficient. It can be seen that four feature quantities, namely the peak factor C, skewness sk, root mean square Irms, and kurtosis K, are highly important and can effectively characterize the differences in different states of the load switch. The four feature quantities, including the form factor sh, maximum starting current Ist, motor operating time t, and Tm, are of low importance. Therefore, this study selects C, sk, Irms, and K as the feature vectors.
3.2 Random Forest Diagnosis Results
The RF algorithm classifies two load - switch states (normal/jammed) using 300 samples per state for training (total 600) and 30 samples for testing. The confusion matrix (Figure 4) shows perfect normal - state identification, 97% accuracy for jamming, and 98.33% average classification accuracy.
3.3 Comparison of Different Classification Algorithms
To test the performance of the random forest classifier, a Support Vector Machine (SVM) and an Extreme Learning Machine (ELM) are trained simultaneously for comparison. The test results are shown in Table 1.
From Table 1, among the three classifiers, the Random Forest (RF) algorithm takes a relatively long diagnosis time of 6.9 ms for test set samples. In terms of accuracy, the Support Vector Machine (SVM) achieves 95% for two operating states, lower than RF. Due to random hidden - layer weights, the Extreme Learning Machine (ELM) has accuracy fluctuating between 85% - 96.67% and poorer robustness than RF. Thus, the RF algorithm used has high accuracy and good robustness.
4 Conclusion
This paper proposes a load - switch mechanical fault detection method using energy - storage motor current time - domain features and the Random Forest (RF) algorithm. It extracts representative time - domain features from motor current waveforms and uses an RF classifier for state identification. The proposed recording - wave start criterion effectively acquires motor current signals. Leveraging the Gini index in RF, it evaluates feature importance and selects four key features (peak factor, skewness, root mean square, kurtosis) to characterize load - switch states. Experiments show the method effectively identifies motor jamming states with 98.33% accuracy.
