Signal Processing For Adaptive Detection And Location Of Water Pipe Leaks

Water pipe leak detection and location is such an important activity that it can have profound effects on our society, economy and environment. At present acoustic leak processing methods have been widely employed for detecting and locating the water pipe leakage. In particular, correlators are the most popular ones, which utilize the generalized cross correlation method to analyze the leak sound signal. A cross correlator needs a priori knowledge of the signal and noise spectra to design its proper filter before correlating the two measurements. Accurate spectra acquisition is crucial for the following correlating process and finally getting correct delay estimation. However, for practical water pipe leak detection, the leak sound varies with different pipe parameters and its buried soil environments, the different sound propagation paths and so on, and all kinds of interferences from the unpredictable noisy background differ all the time. It is very difficult to get accurate spectra of the signal and the noise seperately beforehand, and to design filters is not easy because there is no proper information provided for reference. Even one changing factor will make wrong estimates of the signal or noise spectra, and will lead to the failure of the whole detection.To overcome the traditional cross correlator's weakness, our research group recently applies adaptive time delay estimation, which employs adaptive filtering instead of the generalized cross correlation method, into buried water pipeline leak detection. The adaptive filter can be updated iteratively by adapting to the actual acquired signal's characteristics and get the correlation results directly from its own system parameters, which doesn't need any prior knowledge of the signal or the noise thus it can track the changing detection targets and time-varying environment. Besides, the LMS time delay estimation is simple and easy to compute, and can be implemented on line as well. However, the practical detection environment is always complicated. It is difficult to use traditional ways to determine when the adaptive algorithm converges in a dynamic complex noisy background. Besides, the leaking sound is contaminated by many different interferences coming from the inside and the outside of the pipe, thus the acquired signals are always with very low Signal-Noise-Ratio(SNR). The noises not only introduce the biased estimates for the adaptive filter, but also have an effect on the system's detectible SNR limit and its optimal detection result. Moreover, when it is mixed with nonstationary bursting noises, the location of the convergent adaptive solution varies with different signal characteristics at different time. To solve these problems confronted in the practical application of adaptive time delay estimation, several signal processing methods are proposed in this paper to dynamically determine when the adaptation process acctually converges, to elimate the noise influence with the bias-free adaptive algorithm, and to eradicate the unpredictable bursting interferences, in order to improve the adaptive leak detection and location performance for the water pipe in nonstationary, or low-SNR, or other abominable noisy environments. The detail of the work is described as follows:①It is studied the leak location principle based on adaptive time delay estimation method, and its pinpointing error and the time delay estimation error. From the practical problems confronted with the leak detection in complicated and abominable noisy environments, it is pointed out three key issues to be solved when adaptive method is adopted: dynamic discrimination for the converging state of the adaptive process, biased estimation induced by the noises in the adaptive algorithm, and the influence of the interferences for the Wiener solution. All of these are the work foundation for the whole paper.②Based on the principle of orthogonality, an on-line constraint of the adaptive detection system which can be used to dynamically determine the LMS algorithm's convergence, is derived when the error and output get approximately orthogonal due to the gradient noise practically. The dynamic discriminant obtained through estimated mean-square values of the desired, output and error signals at each iteration, can be updated along with the adaptation. While avoiding the weakness of the traditional discrimination methods, it can dynamically evaluate quality of the adaptive solution and solves the first problem in practical detection. Through ending the iteration timely and accurately it can get the optimal estimation even with complicated nonstationary interferences. As a result, the performance and efficiency of time delay estimation are improved simultaneously while using adaptive algorithms. The experiment results show that it is applicable to adaptive detection systems in nonstationary as well as stationary environments, and the estimation variances of the adaptive parameter identification and time delay estimation applications are significantly decreased because of the dynamic convergence determination.③For the measurement data of lower SNR with stationary noises, the biased Wiener solution induced by the input noise will result in deteriorated performance in LMS time delay estimator (LMSTDE). Then a modified bias-free scheme based on Treichler's r -LMS algorithm is developed for eliminating the input noise iteratively, in which the input noise variance can be simply obtained from the steady output of the traditional adaptive filter by the geometric interpretation of the best approximation projection of relative signals. It utilizes available information from the adaptive filter itself without any a priori knowledge of the interference, to enhance the peak of the optimal Wiener solution, thus it can improve the performance of the LMSTDE in lower SNR environments. Simulation and real data application are both provided to validate its improved effectiveness: it not only makes the detectible SNR level to be reduced to -20dB from -16dB, but also can resist the nonstationary interferences in the actual leak detection to some degree.④It is analyzed the distinctive features of leak sound and other interference noises. Due to the absolutely different mechanics of generation of these two types of signals, and the local time and frequency distribution property of the interferences in time-frequency plane, an application of matching pursuit with a Time-Frequency Dictionary is proposed to remove the unpredictable but determined interferences, which could come from many different sources. A proper time-frequency dictionary is built, then adaptive matching pursuit method is used to iteratively decompose the original mixture signal into a linear expansion of waveforms that are selected from the time-frequency dictionary. These waveforms are chosen in order to best match the signal structure. After the adaptive time-frequency transformation, the main determined interferences in the mixture signals are first extracted, while the leak signal is left behind and recovered at last. In the end, the signal quality is enhanced greatly after the removal of strong time-varying noises and the adaptive leak detector no longer convergences to multiple stable Wiener solutions. The detector is then much more immune to the influence of bursting interferences in practical complicated noisy environments, and more reliable and robust under nonstationary dynamic detection condition.⑤An intelligent leak detection and location instrument is developed by our group based on the adaptive time delay estimation with leak sound. By applying the above signal processing methods into this leak detection instrument, lots of experiments are employed with practical acquired leak data to validate their effectiveness respectively. It shows that through the real-time dynamic convergence discrimination, the bias-free LMSTDE, and the burst interference removal, this adaptive system's detection sensitiviry, location precision and stability in nonstationary noisy environment, and its interference resistance capability after the adaptation process enters its steady state, are all improved dramatically. The actual leak detection error in varying complicated environments can be controlled in less than 2 meters(mostly less than 1 meter) when using the dynamic discriminant, which are close to that in stationary background. And the adaptation process time can be reduced to 1/5 of the original time. The bias-free LMSTDE can not only reduce the detectible SNR limit level but also improve the reliablity and other detection performances of the adaptive system in changeable environments. And after the burst interference removal for practical measurements, the adaptive location error range when the filter enters its steady state can be reduced to 1/30 to 1/2 of the original case, the detection system's robustness is thus enhanced. When all the above signal processing methods are applied into this intelliget leak detection and location instrument, the detector's location error can be controlled in no more than 1 meter, and the system's performance is improved in the dynamic complicated noisy environments.