【正文】 ations, the discrete versions, such as the discrete wavelet transform (DWT) and wavelet packet analysis (WPA), are usually used for reducing the number of calculations to be done, thereby saving puterrunning time. Ambiguity functions (AF) and Wigner–Ville distributions (WVD) have better resolution than STFT, but suffer from crossterm interference and produce results with coarser granularity than wavelet techniques do [7]. AF and WVD with excessive transformation durations are obviously unacceptable in the development of realtime monitoring systems. Almost all vehicle noises in actual working cases, such as those generated in acceleration and deceleration processes, are nonstationary. Therefore, based on the above findings, the STFT, DWT, WPA, and CWT techniques, which are suitable for feature extraction of nonstationary signals, may be considered in this study. However, their ability to characterize sounds for special purposes should be further tested in SQE system developments.The sound classification techniques in mon use have been investigated and pared in [11]. The results show that: selforganizing maps and learning vector quantization techniques are plementary to each other, while longterm statistics cannot be applied in bination with nonstationary feature extraction techniques. The Gaussian mixture model may be adopted in the unsupervised classification of musical signals. Neural networks (NNs) and fuzzy identification are widely used in the biomedical field for modeling, data analysis, and diagnostic classification [12–14]. The ability to reproduce arbitrary nonlinear functions of input, as well as ，the highly parallel and regular structure of the NNs, makes them suitable for plex pattern recognition and classification tasks [15]. Therefore, a NN algorithm is adopted to project the input signal features to the output SQE patterns of vehicle noises in this paper.Based on the above discussion, it can be concluded that sound perception, as a neural response of humans, is an ambiguous concept: Individual differences in perception among people are as much influenced by personal factors as by noise factors. Thus, it is sometimes impossible to find an exact physical model to describe the annoyance response for all people. Therefore, for a correct assessment of sound quality, intelligent methods such as fuzzy logic and NNs should naturally be considered. Based on vehicle interior noises, a new SQE technique of wavelet preprocessing neural network (WTNN) is developed in this work. The WTNN model may be directly used to predict not only nonstationary, but also stationary and even transient soundquality matrices (SQMs) in vehicle engineering. In view of the application, it may also be used in sound recognition, failure diagnostics of equipments and clinical disease diagnostics in medical treatment, etc.2. Establishment of vehicle noise databaseIn this paper, sample vehicle interior noises were prepared using the binaural recording technique. The following data acquisition parameters were used: signal length, 10 s, sampling rate, 22 050 Hz. Inevitably, distortion of the measured sounds by certain additive noises occurred, which came from both ambient background noise and the hardware of the measurement system。 therefore, the signal needed to be denoised. Some techniques for white noise suppression in mon use, such as the least square, spectral subtraction, matching pursuit methods, and the wavelet threshold method have been used successfully in various applications [16–18]. The wavelet threshold method in particular has proved very powerful in the denoising of a nonstationary signal. Therefore, a DWTbased shrinkage denoising technique was applied. This technique may be performed in three steps: (a) deposition of the signal, (b) determination of threshold and nonlinear shrinking of the coefficients, and (c) reconstruction of the signal. Mathematically, the soft threshold signal is sign(x) (|x|t) if |x|t, and otherwise is 0, where t denotes the threshold. The candidate threshold rules and other options in the denoising functions in nonlinear shrinking were also carefully investigated. Finally, the selected parameters were: Daubechies wavelet “b3,” 7 levels, soft universal threshold equal to the root square of 2 log (length(f)), assuming the model is basic and with unscaled noise. As an example, a denoised interior signal is shown in Figs. 1 and 2 is the corresponding spectrum. It can be seen that the harmony and white noise ponents of the sample interior noise are wellcontrolled. The wavelet shrinkage denoising technique is effective and sufficient for denoising vehicle noises. After signal denoising, a database of vehicle interior noises was established for evaluation by using filtering technique. The parameters of the filters designed by Matlab toolbox are shown in Table 1. Eighteen noise signals were generated by filtering the interior noises from the right ‘‘ear’’ of the dummy head, and 18 from the left “ear”Fig. 2. Comparison of the interior noise spectra before and after the waveletdenoising model基于小波變換前神經(jīng)網(wǎng)絡(luò)模型的車內(nèi)非平穩(wěn)噪聲質(zhì)量預(yù)測(cè)摘 要本論文介紹了小波變換前神經(jīng)網(wǎng)絡(luò)模型，是一種新的噪聲質(zhì)量預(yù)測(cè)方法?；谲噧?nèi)噪聲小波變換的前神經(jīng)網(wǎng)絡(luò)噪聲評(píng)測(cè)模型混合了小波變換和神經(jīng)網(wǎng)絡(luò)分類的技術(shù)。作為一個(gè)神經(jīng)網(wǎng)絡(luò)模型，一個(gè)提取噪聲特征的基于小波變換的模型被建立，結(jié)果證明，經(jīng)過訓(xùn)練的神經(jīng)網(wǎng)絡(luò)模型更精確而且有效的對(duì)于非平穩(wěn)車輛噪聲信號(hào)的噪聲質(zhì)量預(yù)測(cè)。由于它非常突出的時(shí)間頻率特性，基于小波變換的神經(jīng)網(wǎng)絡(luò)模型可應(yīng)用于處理平穩(wěn)和非平穩(wěn)隨機(jī)信號(hào)，甚至短暫的信號(hào)。在傳統(tǒng)的心理模型方面，小波變換前神經(jīng)網(wǎng)絡(luò)技術(shù)不進(jìn)能夠預(yù)測(cè)、分類、比較車內(nèi)噪聲質(zhì)量（響度、尖銳度）而且可以用于其他相關(guān)的工程領(lǐng)域1. 引言降低我們生活中日益增長(zhǎng)的噪聲水平，可以提高聲音質(zhì)量，因此提高生活質(zhì)量。在過去的幾十年中，汽車噪聲占城市環(huán)境噪聲的40%，因此汽車噪聲控制已經(jīng)成為一個(gè)非常熱門的研究領(lǐng)域。汽車聲音的主要問題不是造成聽力損傷，而是聲音舒適性。為了在汽車設(shè)計(jì)中提高聲音舒適性，研究人員首先應(yīng)了解怎樣去評(píng)價(jià)噪聲。最近已進(jìn)行了大量相關(guān)的關(guān)于車輛噪聲質(zhì)量評(píng)價(jià)的相關(guān)工作。此外，一個(gè)有關(guān)人聽覺的物理過程是，人感受到的聲音與汽車發(fā)出的聲音不完全一樣；因此，許多心理聲學(xué)指標(biāo)，如響度、尖銳度、音調(diào)、粗超度、波動(dòng)量、愉悅度等等，這些都可以解釋聲音刺激和人的感覺之間的定量關(guān)系，也可用于評(píng)價(jià)汽車噪聲?？梢钥闯鲈肼曎|(zhì)量評(píng)測(cè)方面的算法復(fù)雜度和時(shí)間消耗并不能包括所有的人的情感因素和指定所有聲音的反應(yīng)。舉例來說，汽車噪聲質(zhì)量研究往往把重點(diǎn)放在對(duì)一個(gè)或兩個(gè)類型噪聲的反應(yīng)上，原因是每一種類型的噪音有每一種類型的特征。因此，開發(fā)一種新的強(qiáng)大的工具以便更精確的評(píng)價(jià)噪聲品質(zhì)是有必要、實(shí)用的。在聲音質(zhì)量評(píng)測(cè)領(lǐng)域，信號(hào)處理技術(shù)必須根據(jù)聲音的特征細(xì)心的加以選擇。文獻(xiàn)[7，8]經(jīng)常提到平穩(wěn)頻率技術(shù)和非平穩(wěn)頻率技術(shù)。前者是用來算出那些非獨(dú)一無二的特征，因此不適用于非平穩(wěn)隨機(jī)信號(hào)。時(shí)頻技術(shù)，可以用來提取聲音的瞬態(tài)特性。短時(shí)傅立葉變換在幾個(gè)不同的窗口用一個(gè)標(biāo)準(zhǔn)的傅立葉變換。小波分析技術(shù)用一個(gè)母小波函數(shù)和一個(gè)離散或連續(xù)的尺度克服