专利汇可以提供Digital adaptive finite impulse response filter with large number of coefficients专利检索,专利查询,专利分析的服务。并且A digital adaptive finite impulse response (AFIR) filter having a large number of coefficients is composed of two or more separate filter units. The first filter unit computes response samples h o - h n+1 in response to input signal samples to provide a partial estimated response during each sampling period. After the signal samples arefully processed by the first unit, they are transferred to a second filter unit which produces response samples h n+2 - h p±1 in response thereto to provide a second partial response. The sum of the two partial responses is computed to provide the total estimated system response. The two independent filter units thus act simultaneously to provide twice as many coefficients as prior art AFIR filters in the same amount of time.,下面是Digital adaptive finite impulse response filter with large number of coefficients专利的具体信息内容。
Digital adaptive transversal finite impulse response (AFIR) filters have recently become an important tool in signal processing. The theory of AFIR filters is described in "Identification Algorithm for Adaptive Filters", Comsat Technical Review, O.A. Horna, Volume 8, No. 2, Fall, 1978, pp 331-351.
Figure 1 is an illustration of one such filter. An input signal X is applied on the one hand to tap delay line 10 having a number of equi-distant taps, each of the taps equipped with an adjustable coefficient hj. The input signal X is also applied to the echo feedback path through an identified system. The identified system includes a "flat delay" tfd, and an impulse response H(t). The signal X, as well as the system output d, are applied to the delay line 10 and the error detector 30, respectively, under the selective control of sampler 35.
The output hj of each of the adjustable amplitude devices, is applied to linear combiner 15 which sums the products xijhj, where xij are samples of the input signal X present at the ith sampling interval at the j-th tap, to produce the output xiTh. The identified system output d is subtracted from the system output from combiner 15 in error detector 30 to produce error signal em, which in turn is applied to error storage device 25. The errors em are delivered from the error storage device 25 to correction processor 20. Correction processor 20 computes corrections Δhj which change the coefficients hj so as to drive the error value e m to zero. Thus, the output of the AFIR filter is as close a replica of the output of the identified system H(t) + tfd as possible.
The prior art device of Figure 1 will work properly as long as the number of coefficients is relatively small (on the order of 10 or less) or if the samples of the input reference signal X are uncorrelated over time. Wideband white noise and Poisson waves are examples of a signal having this property. However, voice, music and video signals have autocorrelation coefficients of adjacent samples 6xx > 0.8 when sampled at the Nyquist rate. Thus, for "long filters" where the number of taps is greater than 30, decorrelation of the input signals entering the correction processor must be performed in order to permit the AFIR filter to converge to a stable state, i.e., a stable coefficient vector h.
A very effective prior art decorrelation technique is illustrated in Figure 2. Both the error vector em stored in the error storage register 25 and the signal vector xi stored in the AFIR filter 5 are multiplied or "dithered" by generators 45 and 55 which produce random sequences of ones and zeros Fe and φx, respectively, before entering the correction processor.
Although the number of taps in the filter may be increased and still provide the generation of a stable vector h for correlated signals, this decorrelation method substantially reduces the speed of the adaptation process for producing the coefficients hj. Rather than multiplying the error and signal vectors by purely random sequences, a better solution exists in providing a pseudorandom dither which selects for processing only those samples xij and ei greater than a predetermined value. Such a system is described in U.S. Patent 4,064,379 to O.A. Horna.
Direct analog implementation of the AFIR filters of Figures 1 or 2 is difficult. A wideband delay line with several milliseconds delay is difficult to implement electronically, and a direct electronic permanent or semi-permanent analog memory having the requisite precision does not yet exist. Therefore, typical AFIR filters, such as the one described in U.S. Patent 4,064,379 typically use digital techniques.
Figure 3 is an illustration of a digital embodiment of a prior art AFIR filter. Input signal X is applied to the inputs of A to D converter 55 and the identified system under the, selective control of sampler 35. The output of A to D converter 55 is selectively applied to X register 60 and multiplier 70 via switch S under the control of sampler 35 and delay device 35'. The delay device 35' provides an amount of delay equal to the analog-to-digital conversion time of converter 55. At all other times, switch S is in the "a" position to allow for the recirculation of the values xj in X register 60.
The output of X register 60 is applied to detector 80, pseudorandom sequence generator 85, and correction processor 20'. Detector 80 detects the square root of the sum of the squares of the contents of the X register. The output of detector 80 is applied on the one hand to pseudorandom generator means 85, and on the other hand to pseudorandom generator means 90. Pseudorandom generator means 90 receives a second input from the output of subtractor 30, and provides the pseudorandom sequence Fi to correction processor 20'. The output of pseudorandom processor 85 provides the pseudorandom sequence φj to correction processor 20. Correction processor 20' receives a further input directly from the output of subtractor 30. Finally, correction processor 20' receives its final input from the hn stage of H register 65. The present corrected coeffic:.ent value hn+1 as calculated by correction processor 20' is delivered to the hn+1 stage of H register 65, the value of which is applied to multiplier 70 on the one hand and recirculated to the first stage of H register 65, namely ho, to provide the recirculation of the h vector.
Each of the stages of shift registers 60 and 65 actually comprises a multibit storage element, 8 bits for example, such that an 8-bit value in each of the storage elements is sequentially clocked to an adjacent storage element. Random access memories could be used as an alternative to shift registers 60 and 65 if so desired.
The output of the multiplier 70 is applied to accumulator 75 which functions accumulate the most recent n+1 products. The output of accumulator 75 is applied to D to A converter 95, which in turn provides the system output and an input to the negative terminal of subtractor 30.
The details of correction processor 20', and pseudorandom generators 85 and 90 are given in the above cited article in Comsat Technical Review, as well as in "Echo Canceller with Adaptive Transversal Filter Utilizing Pseudologarithmic Coding" by O.A. Horna, Comsat Technical Review, Volume 7, No. 2, Fall 1977, pp 393-428, and from the above cited U.S. Patent 4,064,379.
The X register 60, having n+1 stages, and the H register 65, having n+2 stages are clocked at a rate (n+1)Fs, where Fs is the sampling rate of sampler 35, to thereby effect a convolution shift of the X and h vectors. However, the same convolution shift can be achieved using X and H registers having an equal number of stages by adding an additional clock pulse to the X register each sampling period. For each sampling period, the contents of registers 60 and 65 will have shifted n+1 stages. During the next clock pulse (n+2), the switch S is put in the "b" position under the control of sampler 35 and delay device 35', such that the next sample is placed in the X register. Since this occurs after n+2 shifts of the n+1 stage shift register, each sample effectively propagates through the registers one stage per each sampling period to thereby effectively provide the delay element 10, Figure 1, while at the same time being circulated, multiplied and accumulated within each sampling period 1/Fs to effect the summation performed by element 15, Figure 1. Thus, the sample Y(i) of the estimated response of the system H(t)+tfd is produced at the end of each cycle, where
The contents of the last stage xn of shift register 60 are destroyed once per sampling period when switch S is in position "b" to accept the new sample.
Since the processing in the digital AFIR filter is serial, i.e., during one sampling period the partial products xijhj are computed sequentially and sequentially added into the accumulator, the sampling rate and the number of coefficients are limited by the speed of the shift registers and the speed of the multiplier and accumulator. With present digital circuits and with a sampling rate of 10 kHz, the maximum number of coefficients must typically be less than 300 to thereby provide a whole impulse response for the identified system of no longer than 30 milliseconds. While these values may be sufficient to handle the responses and delays in long distance telephone networks, they are insufficient to handle broadcast quality speech or systems having a long tfd, such as acoustic feedback between microphone and loudspeaker in a public address system, or acoustic problems in teleconferencing, since tfd, as well at the system response, must be stored by the h coefficients in register 65.
Thus, it is an object of the present invention to overcome the above mentioned shortcomings of the prior art AFIR filters.
In accordance with the present invention, the AFIR is composed of two or more separate units. The first unit stores response samples ho - hn+1 in the same way as the AFIR filter shown in Figure 3. A partial estimated response is computed at the first unit during each sampling period. However, when a new sample is shifted into the X register, the oldest sample, rather than being destroyed, is read into a second shift register where samples xn+1 - xp are stored. The second unit also stores response samples hn+2 - hp+1. During the same period, a second partial response is computed. The sum of the two partial responses is computed and subtracted from the sample of the response of the identified system to produce the error signal. Thus, the two independent units act simultaneously to thereby provide twice as many coefficients as the prior art system in the same amount of time.
Further, more than two of the independent units can be cascaded in this manner. Thus, the estimate of the response is the sum of the partial responses.
In accordance with the further aspect of the present invention, a detector detects a pseudo RMS or average value of the coefficients stored in the H register, and selectively clears the H register whenever the detected value indicates that either the impulse response of the AFIR filter has not been stored or that the H register contains a "false" or "contaminated" response. Resetting the response to zero permits a buildup of a new response under new signal conditions, and reduces the processing noise of the filter.
Figure 4 is an illustration of the AFIR Filter in accordance with the present invention. Specifically, the filter of Figure 4 is divided into first and second modular units 100 and 200, respectively. Each of the modular units 100 and 200 is substantially identical to the unit illustrated in Figure 3, reference numerals in Figure 4 identical to those of Figure 3 identifying identical system elements. pseudorandom generator means 80, 85 and 90 are not explicitly shown in Figure 4, but may of course be included therein.'
The two independent units 100 and 200 are connected as follows. The output of shift register 60a is applied on the one hand to the "a" terminal of switch Sl, as in the Figure 3 embodiment, but is also applied to the "b" terminal of switch S2, corresponding to the input switch S of Figure 3. Thus, when the switch Sl is shifted to position "b" to shift a new sample into the shift register 60a, the oldest sample xn is not truncated, but rather is read into the second shift register 60b where the samples xn+1 - xp are stored.
Therefore, unit 100 stores samples xo - xn in X register 60a, and ho - hn+1 in H register 65a, to thereby produce the first partial response
The two partial responses Y1 and Y2 are applied to subtractor 110 where they are subtracted from the output of the identified system to produce the error signal e as before. As in the Figure 3 embodiment, the error signal e is returned to both correction processors 20'a and 20'b to thereby update the estimates.
It can now be seen that the number of coefficients ho - hp+1 has been doubled over that of the Figure 3 embodiment, while each of the registers 60a, 60b, 65a, and 65b are clocked at the same rate as that in the Figure 3 embodiment. Thus, twice as many coefficients may be provided in the Figure 4 embodiment as in the Figure 3 embodiment, while requiring the same amount of processing time.
Furthermore, more than two of the units 100 and 200 can be cascaded in a similar manner to thereby increase the total number of coefficients, while maintaining the overall processing time equal to the time required for any one unit to process samples. The estimate of the response for a system having m units is the sum of the partial responses
Rather than combining the output of the identified system with the partial responses Y1 and Y2 in subtractor 110 as shown in Figure 4, the signals may be combined as illustrated in Figure 5. Partial responses Y1 and Y2 from accumulators 75a and 75b, respectively, are digitally added in adder 120, and then converted to analog in a single D/A converter 130 for application to subtractor 135, to thereby reduce the required number of D/A converters.
Figure 6 illustrates a third and preferred technique for combining the partial responses Y1 and Y2. A switch S3 is provided for receiving the signal from the first multiplier 70a in the first unit 100 during the first convolution period (the inverse of the sampling frequency as determined in sampler 35). Before the start of the next convolution period, switch S3 is then connected to the output of the second accumulator 75b from the second unit 200, the contents of accumulator 75b being further added to accumulator 140. The contents of accumulator 140 at the end of the convolution period represents the sum of the partial responses Y1 plus Y2. The sum of the responses is applied to a single D to A converter 150 for further application to analog subtractor 135. It can be seen that the independent units can easily be cascaded using this arrangement by merely providing additional poles on switch S3, as illustrated by contact "c" receiving the partial value Yn of a further processing unit.
Figure 7 illustrates a further aspect of the present invention. In most cases, the sampled impulse response stored in the H register 65 occupies only a small portion of the H register. Typically, the system impulse response can be stored using approximately 20 to 50 h coefficients. The other coefficients in H register 65 are therefore used to provide the delay tfd. Thus, most of the coefficients hj have more or less random values _close to zero. However, since even the coefficients produced to emulate the delay tfd are typically a non-zero number, they contribute td background or processing noise when multiplied by the stored signal samples, thereby increasing the error em. This is especially troublesome when a large number of coefficients are provided as in the present invention.
A second situation occurs when a strong, highly correlated interference signal outside the transmitted bandwidth can establish under certain conditions in the H register a "false" response which sometimes cannot be corrected by the correction algorithm used in the correction processor 20'.
Both of these problems can be dealt with effectively if each of the sections 100 and 200 of the.AFIR filter is equipped with' the autocorrection device illustrated in Figure 7. The output of the H register 65 is applied to a detector 230 which computes one of (i) RMS, (ii) average or (iii) normalized sum of the absolute values of each of the contents of H register 65. The detected value is then applied to comparators 235 and 240, where the detected level is compared to references 1 and 2, respectively. Comparator 235 detects the condition where the detected level from detector 230 is greater than the level of REF 1, to thereby indicate that the contents of the H register 65 have been "contaminated" by a "false" response which cannot be corrected in correction processor 20'. In the event that such a level is detected in comparator 235, it delivers a signal to switch S4 via OR gate 215 and latch means 210. In this manner, switch S4 may be kept open for one full sample period of converter 55, whereby the entire contents of the H register 65 will be completely destroyed, thereby extinguishing the "false" response.
Comparator 240 detects the condition where the level detected in detector 230 is lower than a selected reference level, REF 2, to thereby indicate that the impulse response in this section of the AFIR filter has not been stored, and is not participating in the correction process. In this event, the random non-zero values in the H register 65, which contribute to background noise, are extinguished by opening switch S4 for an entire sampling period. However, the contents of the H register 65 will typically be at a value less than REF 2 during the normal adaptation period, until adaptation is achieved. Thus, integrator 225 and comparator 220 are further provided at the output of comparator 240 to ensure that the "low value" condition of H register 65 is maintained for an amount of time determined by integration time constant of the integrator 225 and the reference level of comparator 220. The output of the comparator 220 is applied to switch S4 through OR gate 215 and latch 210, such that in the event that comparator 220 detects that H register 65 has not achieved adaptation, and is merely adding background noise to the system, the contents of the H register can be set to zero.
The RMS, average, or average of absolute values of the coefficients stored in the H register 65 may be computed digitally or analogically in detector circuit 230. The choice of RMS, average, or average of absolute values, depends upon the nature of the transmitted signal X, e.g., voice, data, video, and upon the nature of the identified network, e.g., electric, electro-acoustic, and the like. For example, in good quality telephone transmission lines, the normalized sum of absolute values may be detected in detector 230. Comparator 235 would provide a signal to open switch S4 in the event that the normalized sum of the absolute values were greater than 5 percent of the maximum possible value, and comparator 240 would provide a signal to integrator network 225 in the event that the detected normalized sum of the absolute values were less than 0.1 percent of the maximum possible value. Integrator 225 and comparator 220 would pass the output from comparator 240 whenever the output of the comparator were less than the 0.1 percent value for at least several continuous seconds, e.g., 10 seconds. Integrator 225 would periodically be reset in order to avoid a buildup of comparator outputs.
As a further refinement, a double-talk detector 245, as fully discussed in U.S. Patent 4,064,379, may provide a "double-talk" signal to comparators 235 and 240. In the event that a "double-talk" condition is detected by detector 245, comparator 240 should be disabled from providing an output to integrator 225, while the output of comparator 235 may, under certain circumstances and operating conditions, be stopped. Again, the exact technique of operating the system in accordance with the present invention will inevitably be a function of the system usage and conditions.
Thus, it can be seen that the present invention provides an arbitrary increase in the number of coefficients used in an AFIR filter, while at the same time maintaining a constant processing speed. The present invention is extremely flexible and lends itself to a wide field of applications. By simply cascading the independent modules to provide as many coefficients as desired, the AFIR filter can be used in long-distance telephone networks, broadcast quality speech systems, or to reduce acoustic feedback between microphone and loudspeaker in a public address system or in teleconferencing.
标题 | 发布/更新时间 | 阅读量 |
---|---|---|
一种永磁同步电机迭代反馈整定控制及其优化方法 | 2020-05-11 | 70 |
补偿通信信号的群延迟的通信接收器 | 2020-05-11 | 866 |
用于自适应像素滤波的系统和方法 | 2020-05-13 | 615 |
在频域线性均衡器中处理数字信号的装置和方法 | 2020-05-08 | 550 |
在基于MIMO系统中用于自适应空间分集的设备及方法 | 2020-05-12 | 387 |
用于局部改善电容式触摸传感器中的信噪比的技术 | 2020-05-11 | 194 |
一种基于经验模态分解的对肌电图的预处理方法 | 2020-05-11 | 727 |
谐波电流的补偿方法、装置、设备和可读存储介质 | 2020-05-11 | 575 |
一种适用于双轴或三轴进给驱动系统的有限控制集模型预测轮廓控制方法 | 2020-05-12 | 871 |
用于单脉冲雷达的数字中频接收机及方法 | 2020-05-13 | 901 |
高效检索全球专利专利汇是专利免费检索,专利查询,专利分析-国家发明专利查询检索分析平台,是提供专利分析,专利查询,专利检索等数据服务功能的知识产权数据服务商。
我们的产品包含105个国家的1.26亿组数据,免费查、免费专利分析。
专利汇分析报告产品可以对行业情报数据进行梳理分析,涉及维度包括行业专利基本状况分析、地域分析、技术分析、发明人分析、申请人分析、专利权人分析、失效分析、核心专利分析、法律分析、研发重点分析、企业专利处境分析、技术处境分析、专利寿命分析、企业定位分析、引证分析等超过60个分析角度,系统通过AI智能系统对图表进行解读,只需1分钟,一键生成行业专利分析报告。