A convolutional neural network (CNN) is proposed to learn multiple useful feature representations for a classification from low level (raw pixels) to high level (object). Convolutional kernels are initialized by the learned filter kernels that come from sparse auto-encoders. Unlike some traditional methods, which divide the feature abstracting and classifier training into two separated processes, a discriminative feature vector and a single multi-class classifier of softmax regression are learned simultaneously during the training process. Based on the learned high-quality feature representation, the classification can be efficiently performed. A real-world case of surface defects on steel sheet, which evaluates the classification performance of the proposed method, is depicted in detail. The experimental results indicate that the proposed method is quite simple, effective and robustness for the classification of surface defects on hot-rolled steel sheet. Keywords: convolutional neural networks, classification, surface defects, steel sheet, convolutional kernels, sparse auto-encoder Konvolucijska nevronska mre`a (CNN) je predlagana za u~enje {tevilnih koristnih predstavitev pri klasifikaciji od nizkega nivoja (grobe slikovne pike) do visokega nivoja (predmet). Konvolucijska jedra so inicializirana z nau~enimi filtrirnimi jedri, ki izhajajo iz redkih samoenkoderjev. Razli~no od nekaterih klasi~nih metod, ki delijo funkcijo abstrakcije in trening klasifikacije v dva lo~ena procesa, se vektor nediskriminativne funkcije v enostavnem ve~razrednem klasifikatorju regresije softmax, u~i hkrati med procesom treninga. Na osnovi nau~ene predstavitve z visoko kvalitetno funkcijo, se lahko klasifikacija u~inkovito izvede. Primer iz resni~nega sveta povr{inske napake na jekleni plo~evini, ki ocenjujejo zmogljivost klasifikacije je prikazan v podrobnostih. Rezultati eksperimentov ka`ejo, da je predlagana metoda razmeroma preprosta, u~inkovita in robustna pri klasifikaciji povr{inskih napak na vro~e valjani jekleni plo~evini.

Lithium-ion battery refers to a complex nonlinear system. Real-time diagnosis and accurate prediction of battery state of charge(SOC) parameters are hotspots and critical issues in battery research. To reduce the dependence of state of charge prediction on battery model accuracy and speed, and achieve real-time online estimation, a SOC prediction model of lithium-ion battery system is developed based on the model of support vector machine (SVM). SVM parameter is optimized using an algorithm of particle swarm optimization, and the performance of prediction model is assessed using cross-validation. The obtained experimental data is simulated, involving the comparison with the support vector machine model, and the prediction simulation of the battery in the state of fault. The results reveal that this model with a better performance than that of the support vector machine exhibits high accuracy and generalization ability.

No abstract

No abstract

Summary Under complex working conditions in variable temperatures, the accuracy of SOC is reduced due to the low robustness of the lithium‐ion battery model online parameter identification method as well as the SOC estimation approach. Given this problem, a parameter identification method called FF‐AGLS (alternative generalized least squares with forgetting factor) is proposed. The proposed method was combined with the robust H∞‐CKF (cubature Kalman filter) based on singular value decomposition (SVD) in order to achieve an accurate estimation of lithium‐ion battery SOC. FF‐AGLS, which adopts unbiased estimation, has strong parameter tracking ability in low temperatures and low SOC regions, as well as high model parameter identification accuracy. As a result, combining the H∞ filter with SVD‐CKF can maintain the robustness of the algorithm when the model parameters are uncertain, which may solve issues related to the decrease in SOC estimation accuracy caused by temperature changes. Finally, a series of experiments were conducted on the proposed method at different temperatures, while its performance was verified with the current under different working conditions. Accordingly, the joint algorithm based on FF‐AGLS and H∞‐SVD‐CKF was able to accurately track model parameters and SOC with a strong degree of robustness.