حذف نويز صوتي تناوبي با روش فعال و هوشمند

در اين تحقيق، روشي جديد مبتني بر تكنيك‌هاييادگيري تقويتي با تكيه بر روش تفاوت زماني صفر در مبحث يادگيري تقويتي، براي حذف نويز صوتي به صورت فعال در سيگنال‌هاي باند باريك ارايه گرديده است. براي پياده سازي اين روش مبتني بر بازخورد، اطلاعاتي از ديناميك مسير اوليه و ثانويه لازم نمي باشد و در نتيجه اين روش نسبت به تغيير اين پارامترها به صورت مقاوم عمل مي نمايد. حجم بسيار كم محاسبات رياضي و حافظه بسيار اندك محاسبات، از ديگر مزاياي روش پيشنهادي مي باشد. بدين منظور در اولين گام با انجام تعاريف مناسب براي متغيرهاي حالت، اعمال ممكن و سيگنال پاداش، مسيله به يكمسيله مناسب براي حل توسط يادگيري تقويتي تبديل شده و با روش يادگيري كيو حل مي گردد. در گام بعدي با استفاده از اطلاعات به دست آمده از مدل‌سازي در حوزه فركانس، مسيله ساده گرديده و با روش تفاوت زماني صفر مسيله حل خواهد گرديد. آخرين مرحله پيشنهاديك روش چند سطحي مي باشد كه در اين روش اخير بدون نياز به بالاتر بردن حافظه مورد نياز مي توان، دقت را تا مقدار لزوم افزايش داد. In this paper a new method based on temporal difference reinforcement learning algorithm is proposed to active noise control of periodic acoustic signal. This method does not need dynamic information and therefore it is fully robust with respect to the dynamic changes. Very low computational burden and low needed memory are some advantages of this method. In the first step, through well definition of the state variables, actions, and reward signal, a reinforcement learning problem is formed and then it is solved by Q-learning technique. In the next step, the problem is simplified with the aid of frequency domain modeling information and it is solved by temporal difference methods. Finally, a multi-level approach is presented which can increase the precision without increasing of the memory size. It is shown by simulation that this method works well and the improvement of the multi-level temporal difference method with respect to the Q-learning method is addressed.

In this paper a new method based on temporal difference reinforcement learning algorithm is proposed to active noise control of periodic acoustic signal. This method does not need dynamic information and therefore it is fully robust with respect to the dynamic changes. Very low computational burden and low needed memory are some advantages of this method. In the first step, through well definition of the state variables, actions, and reward signal, a reinforcement learning problem is formed and then it is solved by Q-learning technique. In the next step, the problem is simplified with the aid of frequency domain modeling information and it is solved by temporal difference methods. Finally, a multi-level approach is presented which can increase the precision without increasing of the memory size. It is shown by simulation that this method works well and the improvement of the multi-level temporal difference method with respect to the Q-learning method is addressed