Adaptive signal filtering method and system based on multi-band fusion

Abstract

The present invention provides an adaptive signal filtering method and system based on multi-band fusion. First, a multi-band signal set containing original signal components of different frequency bands in a target environment is obtained, and the multi-band signal set is subjected to frequency band decomposition to obtain multiple sub-band signals. Then, an adaptive filtering operation is performed on each sub-band signal, and the filtering coefficient is dynamically adjusted according to the noise characteristics of the sub-band signal to generate a filtered sub-band signal. The filtered sub-band signal is then subjected to multi-band fusion to obtain a target filtered signal. Finally, according to a quality evaluation result of the target filtered signal, the segmentation strategy of the frequency band decomposition or the dynamic adjustment rule of the adaptive filtering is adjusted to output an optimized target filtered signal. This method can effectively cope with complex multi-band signals, adaptively adjust the filtering process, and improve the quality and stability of the filtered signal.

Description

Technical Field

[0001] The present invention relates to the field of signal filtering technology, and in particular to an adaptive signal filtering method and system based on multi-band fusion. Background Art

[0002] In today's signal processing field, signal filtering is a crucial technology that is widely used in many fields such as communications, radar, audio processing, etc. Its purpose is to remove noise interference in the signal to obtain high-quality useful signals.

[0003] Traditional signal filtering methods mostly employ a single, fixed filtering strategy. This single filtering strategy presents significant limitations when processing multi-band signals containing raw signal components from different frequency bands. Because signals in different frequency bands have distinct characteristics, such as frequency range and noise distribution, a single filtering parameter cannot simultaneously meet the filtering requirements of signals in all frequency bands. This can lead to over-filtering of signals in some frequency bands and under-filtering of signals in others, compromising the quality of the final filtered signal.

[0004] Furthermore, existing filtering technologies often lack the ability to adapt to real-time signal changes. In practical applications, the signal characteristics of the target environment often change dynamically, and the characteristics of noise also change over time and in different environments. Traditional filtering methods are unable to dynamically adjust the filter coefficients based on the real-time noise characteristics of the signal, making it difficult to maintain stable filtering effects and provide reliable filtering results in complex and changing signal environments. Summary of the Invention

[0005] In view of the above-mentioned problems, in combination with the first aspect of the present invention, an embodiment of the present invention provides an adaptive signal filtering method based on multi-band fusion, the method comprising:

[0006] Acquire a multi-band signal set of a target environment, wherein the multi-band signal set includes original signal components of different frequency bands;

[0007] Performing frequency band decomposition processing on the multi-band signal set to obtain a plurality of sub-band signals;

[0008] Performing an adaptive filtering operation on each of the sub-band signals to generate a corresponding filtered sub-band signal; wherein the adaptive filtering operation includes dynamically adjusting a filter coefficient according to a noise characteristic of the sub-band signal;

[0009] Performing multi-band fusion processing on the filtered sub-band signals to generate a target filtered signal;

[0010] According to the signal quality evaluation result of the target filtered signal, the segmentation strategy of the frequency band decomposition processing or the dynamic adjustment rule of the adaptive filtering operation is adjusted, and an optimized target filtered signal is output.

[0011] In a possible implementation of the first aspect, performing frequency band decomposition processing on the multi-band signal set to obtain a plurality of sub-band signals includes:

[0012] Determine a target decomposition level according to an initial frequency band division rule, and dynamically iteratively adjust the number of frequency bands and bandwidth range of the target decomposition level based on the spectral characteristics of the multi-band signal set during signal processing, wherein if the frequency band energy balance characteristics after dynamic iterative adjustment do not meet a preset threshold, further optimize the frequency band division rule based on the frequency band energy proportion;

[0013] performing overlapping segmentation processing on each original signal component in the multi-band signal set based on the target decomposition level to generate an initial segmented signal;

[0014] Performing a frequency domain transform operation on the initial segmented signal to obtain a frequency domain energy distribution feature;

[0015] According to the matching degree between the frequency domain energy distribution feature and the frequency band division rule, the initial segmented signal is mapped to the corresponding target frequency band to generate the sub-band signal.

[0016] In a possible implementation of the first aspect, performing overlapping segmentation processing on each original signal component in the multi-band signal set based on the target decomposition level to generate an initial segmented signal includes:

[0017] Determining a sampling rate of the original signal component, and calculating the number of sample points corresponding to a segmented window according to a minimum bandwidth range in the target decomposition level and the sampling rate, such that the duration of the segmented window is proportional to the inverse of the minimum bandwidth;

[0018] Determining the number of overlapping sample points between adjacent segment windows based on the number of frequency bands at the target decomposition level, wherein the number of overlapping sample points is in negative proportion to the number of frequency bands;

[0019] A sliding window sequence with equal intervals is generated according to the number of sample points and the number of overlapping sample points in the segmented window.

[0020] In a possible implementation of the first aspect, performing an adaptive filtering operation on each of the sub-band signals to generate a corresponding filtered sub-band signal includes:

[0021] Extracting time domain waveform features and frequency domain energy features of the sub-band signal;

[0022] Determining a signal stationarity index based on the time domain waveform characteristics, and determining a noise energy proportion based on the energy proportion of the noise-dominant frequency band in the frequency domain energy characteristics, wherein the noise-dominant frequency band is identified by a preset noise reference interval or a dynamic noise estimation method based on a minimum statistics method, specifically including extracting background noise energy distribution in a signal silence segment and iteratively updating a noise energy threshold;

[0023] Constructing a dynamic weight factor based on the signal stationarity index and the noise energy ratio, and adjusting a preset filter parameter set according to the dynamic weight factor;

[0024] Recursively filtering the sub-band signal by calling the set of filtering parameters, and detecting the position of the filter poles in real time during the filtering process to constrain the stability of the filter. If the poles are detected to be outside the unit circle, resetting the filtering parameters to preset safe values to generate an initial filtered signal.

[0025] Performing short-time Fourier transform processing on the initial filtered signal to generate a time-frequency energy distribution matrix;

[0026] According to the frequency band division rules of the current target decomposition level, the characteristic frequency bands related to noise in the time-frequency energy distribution matrix are extracted;

[0027] The ratio of the average energy in the characteristic frequency band to the average energy in the signal dominant frequency band is calculated. If the ratio exceeds a preset noise threshold, it is determined that residual noise energy exists. The filter parameter set is iteratively updated until the residual noise energy is lower than the preset threshold, and the final filtered sub-band signal is output.

[0028] In a possible implementation of the first aspect, determining a signal stationarity index according to the time domain waveform feature includes:

[0029] Calculating the time domain variance sequence of the sub-band signal and the difference mean between adjacent sampling points;

[0030] Determining a first stationary factor according to the fluctuation amplitude of the time domain variance sequence;

[0031] determining a second stability factor according to the absolute value of the difference mean;

[0032] performing standardization conversion processing on the first stationary factor and the second stationary factor respectively to map them to a uniform proportional interval, and dynamically allocating weight coefficients of the first stationary factor and the second stationary factor after the standardization conversion according to the frequency band position of the sub-band signal, and then performing weighted summation to generate the signal stationarity index;

[0033] The weight coefficients of the first stationary factor and the second stationary factor are dynamically allocated according to the frequency band position of the sub-band signal.

[0034] In a possible implementation of the first aspect, performing multi-band fusion processing on the filtered sub-band signal to generate a target filtered signal includes:

[0035] Performing signal reconstruction processing on each of the filtered sub-band signals to generate a time domain signal component of the same length as the original signal component;

[0036] Dynamically determining a frequency band weight coefficient for each time domain signal component according to the target decomposition level and the signal quality score;

[0037] Performing time domain alignment and weighted superposition processing on the time domain signal components based on the frequency band weight coefficients to generate an initial fusion signal;

[0038] Extracting phase difference features of adjacent frequency band signal components in the initial fusion signal, and dynamically calculating the phase correction amount of each frequency band based on the phase difference features;

[0039] Based on the phase correction amount of each frequency band, phase-align the signal components of the initial fusion signal through time domain delay compensation or interpolation processing, and use cyclic shift or edge symmetric interpolation method to keep the signal components equal in length to the original signal to generate a phase-aligned signal to address the change in signal length caused by time domain delay compensation;

[0040] Performing time domain smoothing on the phase alignment signal to eliminate instantaneous mutation noise introduced by phase correction, thereby obtaining a corrected initial fusion signal;

[0041] The target filtered signal is generated according to the corrected initial fusion signal.

[0042] In a possible implementation of the first aspect, adjusting the segmentation strategy of the frequency band decomposition processing or the dynamic adjustment rule of the adaptive filtering operation according to the signal quality evaluation result of the target filtered signal includes:

[0043] Extracting the time domain signal-to-noise ratio feature, frequency band energy balance feature and phase continuity feature of the target filtered signal;

[0044] Calculating a first optimization index based on the time domain signal-to-noise ratio feature, calculating a second optimization index based on the frequency band energy balance feature, and calculating a third optimization index based on the phase continuity feature;

[0045] Performing range normalization processing on the first optimization index, the second optimization index, and the third optimization index, respectively, calculating a normalized weighted sum based on a preset weight coefficient, and generating a comprehensive optimization score;

[0046] If the comprehensive optimization score is lower than the preset optimization threshold, at least one of the following adjustment operations is performed:

[0047] (1) adjusting the number of frequency bands or bandwidth range in the frequency band division rule according to the energy balance characteristics of the frequency bands;

[0048] (2) adjusting a dynamic weight factor calculation rule in the adaptive filtering operation according to the time domain signal-to-noise ratio characteristics;

[0049] (3) Adjusting the phase correction parameters in the multi-band fusion process according to the phase continuity characteristics.

[0050] In a possible implementation of the first aspect, adjusting the number of frequency bands or bandwidth range in the frequency band division rule according to the frequency band energy balance feature includes:

[0051] According to the frequency band energy balance characteristics of the target filtered signal, the energy proportion sequence and standard deviation of each sub-band are obtained;

[0052] Determine whether the standard deviation exceeds a preset equalization threshold. If so, perform the following frequency band adjustment operations:

[0053] (a) Determine the upper limit of the number of frequency bands allowed for adjustment and the minimum bandwidth constraint value based on the number of frequency bands and the bandwidth range of each frequency band in the current frequency band division rules;

[0054] (b) selecting a first target frequency band corresponding to the highest energy ratio and a second target frequency band corresponding to the lowest energy ratio from the energy ratio sequence;

[0055] (c) splitting the first target frequency band: calculating a maximum number of splittable sub-frequency bands based on the minimum bandwidth constraint value, and dividing the frequency band into a plurality of sub-frequency bands according to an energy proportion gradient, ensuring that the bandwidth of each sub-frequency band is not less than the minimum bandwidth constraint value;

[0056] (d) performing frequency band merging on the second target frequency band: identifying energy proportions of adjacent frequency bands, and if the energy proportions of the adjacent frequency bands are lower than a preset merging threshold, merging the second target frequency band and the adjacent frequency band into a new frequency band, where the bandwidth of the new frequency band is the sum of the energy proportions of the adjacent frequency bands;

[0057] (e) updating the number of frequency bands and the bandwidth of each frequency band in the frequency band division rules, and recalculating the standard deviation of the adjusted energy proportion sequence of each sub-band;

[0058] Iterate steps (a) to (e) until the standard deviation is lower than the equalization threshold or reaches the upper limit of the number of frequency bands, and output the adjusted frequency band division rule.

[0059] In a possible implementation of the first aspect, adjusting a dynamic weight factor calculation rule in the adaptive filtering operation according to the time-domain signal-to-noise ratio feature includes:

[0060] According to the time domain signal-to-noise ratio characteristics of the target filtered signal, the ratio of the average energy of the signal-dominant time domain window to the average energy of the noise time domain window is obtained;

[0061] mapping the ratio to a filter strength level, wherein the filter strength level is negatively correlated to the ratio;

[0062] Adjusting the proportional relationship between the noise energy weight coefficient and the signal stationarity weight coefficient in the dynamic weight factor calculation formula based on the filtering strength level specifically includes:

[0063] (f) When the filter strength level is high, the proportion of the noise energy weight coefficient is increased so that its proportion in the dynamic weight factor is increased to a preset upper limit;

[0064] (g) When the filter strength level is low, increase the proportion of the signal stability weight coefficient so that its proportion in the dynamic weight factor is raised to a preset lower limit;

[0065] Substitute the adjusted weight coefficient ratio into the dynamic weight factor calculation formula to generate a new dynamic weight factor;

[0066] The filter parameter set of the adaptive filtering operation is updated according to the new dynamic weight factor and applied to the next round of recursive filtering processing of the sub-band signal.

[0067] For example, in a possible implementation of the first aspect, adjusting the phase correction parameter in the multi-band fusion process according to the phase continuity feature includes:

[0068] According to the phase continuity characteristics of the target filtered signal, the mean and variance of the phase difference of the adjacent frequency band signal components in the overlapping time domain window are obtained;

[0069] determining an overall phase offset based on the phase difference mean, and determining a local phase fluctuation intensity based on the variance;

[0070] Perform the following phase correction parameter adjustments:

[0071] (h) if the phase difference mean exceeds a first correction threshold, calculating a time domain delay compensation amount for each frequency band based on the overall phase offset, generating a continuous time domain signal using a polynomial interpolation method for non-integer sample point shift requirements, and achieving phase alignment through resampling;

[0072] (i) if the variance exceeds a second correction threshold, inserting interpolation points in the overlapping time domain windows, and generating a smooth transition curve based on the interpolated phase sequence to eliminate local phase mutations;

[0073] Update the time domain delay or interpolation point density in the phase correction parameters, and recalculate the phase difference mean and variance;

[0074] The parameters are iteratively adjusted until the phase difference mean and variance are both lower than corresponding thresholds, and an optimized phase correction parameter set is output.

[0075] On the other hand, an embodiment of the present invention also provides an adaptive signal filtering system based on multi-band fusion, including a processor and a machine-readable storage medium, wherein the machine-readable storage medium is connected to the processor, the machine-readable storage medium is used to store programs, instructions or codes, and the processor is used to execute the programs, instructions or codes in the machine-readable storage medium to implement the above method.

[0076] Based on the above aspects, the embodiments of the present invention can effectively deal with complex multi-band signals containing original signal components of different frequency bands, and use frequency band decomposition to subdivide the multi-band signal into multiple sub-band signals, so that subsequent filtering operations can be performed according to the characteristics of each sub-band, avoiding the limitations of a single filtering strategy for processing signals of different frequency bands. The adaptive filtering operation dynamically adjusts the filter coefficient according to the noise characteristics of the sub-band signal, which can more accurately remove the noise in each sub-band, thereby improving the pertinence and effectiveness of the filtering. The multi-band fusion processing integrates the filtered sub-band signals into the target filtered signal, ensuring the integrity and coherence of the signal. The segmentation strategy of the frequency band decomposition processing or the dynamic adjustment rules of the adaptive filtering operation are dynamically adjusted according to the signal quality evaluation results of the target filtered signal, so that the entire filtering process has adaptive and self-optimization capabilities, and can optimize the filtering effect in real time according to the actual signal conditions, significantly improving the quality and stability of the filtered signal. BRIEF DESCRIPTION OF THE DRAWINGS

[0077] Figure 1 It is a schematic diagram of the execution flow of the adaptive signal filtering method based on multi-band fusion provided by an embodiment of the present invention.

[0078] Figure 2 3 is a schematic diagram of exemplary hardware and software components of an adaptive signal filtering system based on multi-band fusion provided by an embodiment of the present invention. DETAILED DESCRIPTION

[0079] The present invention will be described in detail below with reference to the accompanying drawings. Figure 1 FIG1 is a flow chart of an adaptive signal filtering method based on multi-band fusion provided by an embodiment of the present invention. The adaptive signal filtering method based on multi-band fusion is introduced in detail below.

[0080] Step S110: Acquire a multi-band signal set of a target environment, where the multi-band signal set includes original signal components of different frequency bands.

[0081] In the field of signal processing, the target environment can be a communications network. For these environments, broadband antennas are often used as signal acquisition devices. Broadband antennas have a wide frequency response range, capable of receiving wireless signals across different frequency bands. Their operating principle is based on electromagnetic induction. When an external electromagnetic wave signal acts on the antenna, an induced current is generated within the antenna. This induced current is detected and processed, and converted into an electrical signal for output. Signal acquisition devices must adhere to certain sampling rules when acquiring signals. Sampling is the process of converting a continuous analog signal into a discrete digital signal. The core of these sampling rules is the selection of the sampling frequency. According to the Nyquist sampling theorem, the sampling frequency must be at least twice the highest frequency of the signal to ensure complete acquisition of signal information. Assuming the highest frequency of the signal in the target environment is fmax, the sampling frequency fs should satisfy fs ≥ 2fmax. In practical applications, to improve sampling accuracy and reliability, a higher sampling frequency than the theoretical value is often selected.

[0082] During the sampling process, the sampling interval and duration must also be considered. The sampling interval Δt is equal to the inverse of the sampling frequency, that is, Δt = 1 / fs. The sampling duration T is determined based on actual needs and signal characteristics. If the signal is continuously changing, the acquisition time must be long enough to capture the complete signal characteristics. If the signal is a short burst, the sampling duration can be relatively short.

[0083] The collected multi-band signal set is composed of multiple original signal components in different frequency bands. These original signal components may come from different signal sources, and the signal generated by each signal source has a unique frequency range and characteristics.

[0084] Step S120: performing frequency band decomposition processing on the multi-band signal set to obtain a plurality of sub-band signals.

[0085] Step S121: Determine the target decomposition level according to the initial frequency band division rule, and dynamically iteratively adjust the number of frequency bands and bandwidth range of the target decomposition level based on the spectral characteristics of the multi-band signal set during the signal processing process. If the frequency band energy balance characteristics after dynamic iterative adjustment do not meet the preset threshold, the frequency band division rule is further optimized according to the frequency band energy ratio.

[0086] The initial frequency band division rule forms the basis for frequency band decomposition processing and is developed based on a preliminary understanding of the target environment's signals and previous experience. For example, in a communication network environment, a rough frequency band division scheme can be pre-defined based on the frequency band ranges of different communication protocols. Assume that the initial frequency band division rule divides the entire frequency range into n bands, each with a specific bandwidth.

[0087] The target decomposition level is determined based on the initial frequency band division rules. It specifies the level of detail for decomposing the multi-band signal set. A higher target decomposition level means a more detailed signal decomposition, a greater number of frequency bands, and a narrower bandwidth for each band. After determining the target decomposition level, spectral analysis is performed on the multi-band signal set to determine its spectral characteristics.

[0088] Spectral analysis converts time-domain signals into frequency-domain signals. Spectral analysis provides a clear understanding of the signal's distribution across different frequencies. Common spectral analysis methods include Fourier transform and fast Fourier transform. For example, the fast Fourier transform (FFT) is an efficient Fourier transform algorithm that quickly converts time-domain signals into frequency-domain signals. Assuming the collected multi-band signal set is x(t), the FFT algorithm can be used to obtain its frequency-domain representation, X(f). Spectral characteristics include information such as the signal's frequency components, frequency range, frequency density, and the energy distribution of each frequency component.

[0089] Based on spectral characteristics, the number of frequency bands and bandwidth ranges at the target decomposition level are dynamically and iteratively adjusted. Specifically, if spectrum analysis results show that certain frequency bands have concentrated signal energy and a wide bandwidth, further subdivision of these bands may be necessary to better process the signal. If certain frequency bands have weak signal energy and a narrow bandwidth, they may be merged with adjacent bands to simplify processing. After each adjustment, the energy balance characteristics of the frequency bands must be calculated.

[0090] The frequency band energy balance feature measures the uniformity of energy distribution across frequency bands. One method for calculating the frequency band energy balance feature is to first calculate the energy of each frequency band. Assuming the signal of frequency band i is xi(t), its energy Ei can be obtained by integrating the square of the signal in that frequency band over a period of time, that is, Ei = ∫[xi(t)]²dt. Next, the average value Eavg and standard deviation σ of the energy of all frequency bands are calculated. The frequency band energy balance feature can be expressed as the ratio of the standard deviation to the average value, σ / Eavg. A smaller ratio indicates a relatively uniform energy distribution; a larger ratio indicates an uneven energy distribution.

[0091] The preset threshold is a standard value set based on actual needs and experience. If the frequency band energy balance characteristics after dynamic iterative adjustment do not meet the preset threshold, indicating that the energy distribution is severely uneven, the frequency band division rules need to be further optimized based on the frequency band energy contribution. The frequency band energy contribution refers to the proportion of each frequency band's energy in the total energy. Assuming the total energy is Etotal and the energy of the i-th frequency band is Ei, the energy contribution of the frequency band Pi = Ei / Etotal.

[0092] Frequency bands with high energy contributions can be further subdivided to improve signal processing accuracy. Frequency bands with low energy contributions can be merged with adjacent bands to improve energy distribution uniformity. Optimizing the frequency band division rules requires repeated dynamic iterative adjustments and calculations of frequency band energy balance characteristics until the characteristics meet a preset threshold.

[0093] Step S122: performing overlapping segmentation processing on each original signal component in the multi-band signal set based on the target decomposition level to generate an initial segmented signal.

[0094] Step S1221: Determine the sampling rate of the original signal component, and calculate the number of sample points corresponding to the segmented window according to the minimum bandwidth range in the target decomposition level and the sampling rate, so that the duration of the segmented window is proportional to the inverse of the minimum bandwidth.

[0095] The sampling rate of the original signal components is a critical parameter determined during the signal acquisition phase. It affects the precision of signal discretization and must satisfy the Nyquist sampling theorem to ensure complete acquisition of signal information. When performing overlapping segmentation, the number of sample points corresponding to the segmentation window must be calculated based on the minimum bandwidth and sampling rate at the target decomposition level.

[0096] The minimum bandwidth reflects the minimum granularity of signal segmentation at the target decomposition level. According to fundamental principles of signal processing, the duration of the segmentation window is proportional to the inverse of the minimum bandwidth. Specifically, the segmentation window duration T and the minimum bandwidth range Δfmin satisfy the relationship T = 1 / Δfmin. This is because in the frequency domain, a narrower bandwidth corresponds to a longer time window. Accurately analyzing signals in frequency bands with smaller bandwidths requires a correspondingly longer segmentation window.

[0097] When calculating the number of sample points, the time duration T is converted into the number of sample points N using the sampling rate fs. Since the sampling rate fs represents the number of samples collected per second, the number of sample points N = fs × T = fs / Δfmin. This calculation ensures that the duration of the segmented window is proportional to the inverse of the minimum bandwidth, enabling the segmented window to effectively cover signals with different bandwidths.

[0098] Step S1222: determining the number of overlapping sample points between adjacent segment windows based on the number of frequency bands at the target decomposition level, wherein the number of overlapping sample points is in negative proportion to the number of frequency bands.

[0099] After determining the number of sample points corresponding to the segmentation windows, we need to determine the number of overlapping sample points between adjacent segmentation windows. The overlap between adjacent segmentation windows is intended to avoid information loss or discontinuities during signal segmentation. The number of overlapping sample points is related to the number of frequency bands at the target decomposition level, and they are inversely proportional.

[0100] A large number of frequency bands means a more detailed signal segmentation, with each band having a relatively narrow bandwidth, and potentially more complex signal variations. In this case, excessive overlap between adjacent segmented windows increases computational complexity and processing complexity, potentially introducing unnecessary redundant information. Therefore, when the number of frequency bands is large, appropriately reducing the number of overlapping sample points can improve processing efficiency.

[0101] Conversely, when there are fewer frequency bands, the bandwidth of each band is relatively wide, and the signal changes more gradually. To ensure accurate capture of signal characteristics and avoid information loss during segmentation, it is necessary to increase the number of overlapping sample points between adjacent segmentation windows. This allows for greater information overlap between adjacent segmentation windows, thereby better connecting and transitioning between different segmented signals.

[0102] When determining the specific number of overlapping sample points, to avoid negative values and range out-of-bounds situations, assume that the number of frequency bands at the target decomposition level is n, the number of sample points in the segmentation window is N, and the number of overlapping sample points is M. First, a proportional coefficient function α(n) related to the number of frequency bands is set, which satisfies α(n) = max(0, min(1, ab×n)), where a and b are constants and a>0, b>0. Then, the number of overlapping sample points M = α(n) × N is calculated. In this way, the number of overlapping sample points can be dynamically adjusted according to the number of frequency bands at the target decomposition level, while ensuring that the number of overlapping sample points is within a reasonable range to achieve the best segmentation processing effect.

[0103] Step S1223: Generate a sliding window sequence with equal intervals according to the number of sample points and the number of overlapping sample points in the segmented window.

[0104] After determining the number of sample points N in a segmented window and the number of overlapping sample points M between adjacent segmented windows, a sliding window sequence with equal intervals can be generated. A sliding window sequence is a tool used to segment a signal. It slides across the signal sequence at set intervals, capturing a segment of the signal as a segmented window.

[0105] The specific process for generating a sliding window sequence is as follows: First, the starting point of the signal sequence is used as the starting position of the first sliding window. A segment of the signal is intercepted according to the number of sample points N in the segmented window to form the first segmented window. Then, the starting position of the next sliding window is determined based on the number of overlapping sample points M. The starting position of the next sliding window is shifted backward by the number of sample points NM relative to the starting position of the current sliding window. This process is repeated until the sliding window covers the entire signal sequence, thus generating a series of equally spaced sliding windows.

[0106] The signal segments intercepted by each sliding window are the initial segmented signals. These initial segmented signals will serve as the basis for subsequent processing. By further analyzing and processing each initial segmented signal, the signal features can be better extracted and noise can be removed.

[0107] Step S123: performing a frequency domain transformation operation on the initial segmented signal to obtain a frequency domain energy distribution feature.

[0108] The purpose of performing a frequency domain transform on the initial segmented signal is to convert the time domain signal into a frequency domain signal, thereby obtaining the energy distribution characteristics of the signal in the frequency domain. Common methods for frequency domain transform operations include Fourier transform and fast Fourier transform (FFT).

[0109] Take the Fast Fourier Transform (FFT) as an example. It's an efficient Fourier transform algorithm that quickly converts time-domain signals into frequency-domain signals. Assume the initial segmented signal is x[n], where n = 0, 1, ..., N-1, where N is the number of sample points in the segmented window. Using the FFT algorithm, we can obtain its frequency-domain representation, X[k], where k = 0, 1, ..., N-1.

[0110] The frequency domain energy distribution characteristic reflects the energy distribution of a signal at different frequencies. For a frequency domain signal X[k], the energy of each frequency component can be calculated by calculating the square of the amplitude of that frequency component. That is, the energy of the kth frequency component, Ek = |X[k]|². Arranging the energies of all frequency components in order of frequency yields the frequency domain energy distribution characteristic.

[0111] Frequency domain energy distribution features can intuitively demonstrate the concentration of signal energy at different frequencies. For example, if the energy is high within a certain frequency range, it indicates that the signal components within that frequency range are strong; if the energy is low within a certain frequency range, it indicates that the signal components within that frequency range are weak. By analyzing frequency domain energy distribution features, we can better understand the frequency characteristics of the signal.

[0112] Step S124: mapping the initial segmented signal to a corresponding target frequency band according to the matching degree between the frequency domain energy distribution feature and the frequency band division rule to generate the sub-band signal.

[0113] After obtaining the frequency domain energy distribution characteristics and determining the frequency band division rules, it is necessary to map the initial segmented signal to the corresponding target frequency band according to the matching degree between them, thereby generating a sub-band signal.

[0114] The frequency band division rule specifies the frequency ranges of different frequency bands. For example, suppose the frequency band division rule divides the entire frequency range into m frequency bands, and the frequency range of the jth frequency band is [fj1, fj2], where j = 1, 2, ..., m.

[0115] For each frequency component in the frequency domain energy distribution feature, determine the frequency band it belongs to. If the frequency fk of the kth frequency component satisfies fj1≤fk≤fj2, then the time domain signal component corresponding to the frequency component is mapped to the jth target frequency band.

[0116] The specific mapping process is as follows: First, based on the relationship between the frequency-domain signal X[k] and the time-domain signal x[n], the frequency-domain signal is converted back to the time-domain signal. This can be achieved using an inverse fast Fourier transform (IFFT). Then, the time-domain signal components belonging to the same target frequency band are extracted to form the subband signals of that target frequency band.

[0117] For example, for the jth target frequency band, the time-domain signal components corresponding to all frequency components satisfying fj1≤fk≤fj2 are extracted and, after IFFT transformation, the sub-band signal yj[n] for that frequency band is obtained. In this way, the initial segmented signal is mapped to the corresponding target frequency band, generating multiple sub-band signals. These sub-band signals will serve as input for subsequent adaptive filtering operations.

[0118] Step S130: performing an adaptive filtering operation on each of the sub-band signals to generate a corresponding filtered sub-band signal; wherein the adaptive filtering operation includes dynamically adjusting a filtering coefficient according to a noise characteristic of the sub-band signal.

[0119] Step S131: extracting the time domain waveform characteristics and frequency domain energy characteristics of the sub-band signal.

[0120] Sub-band signals contain rich information. By extracting their time domain waveform characteristics and frequency domain energy characteristics, we can better understand the characteristics of sub-band signals and provide a basis for adaptive filtering operations.

[0121] Time-domain waveform features reflect the temporal variations of subband signals. Common time-domain waveform features include amplitude, period, rise time, and fall time. There are various methods for extracting time-domain waveform features. For example, the amplitude characteristics of a signal can be described by calculating statistics such as the maximum, minimum, and average values, and the period can be determined by detecting the signal's zero crossings.

[0122] Frequency domain energy characteristics reflect the energy distribution of subband signals in the frequency domain. As previously mentioned, frequency domain signals can be obtained by performing a frequency domain transform (such as FFT) on the time domain signal. The energy of each frequency component is then calculated to obtain the frequency domain energy characteristics. Frequency domain energy characteristics can help understand the strength of different frequency components in the signal and determine the presence and frequency range of noise.

[0123] Step S132: Determine a signal stationarity index based on the time domain waveform characteristics, and determine the noise energy proportion based on the energy proportion of the noise-dominated frequency band in the frequency domain energy characteristics, wherein the noise-dominated frequency band is identified by a preset noise reference interval or a dynamic noise estimation method based on the minimum statistics method, specifically including extracting the background noise energy distribution in the signal silence segment and iteratively updating the noise energy threshold.

[0124] Signal stationarity is used to measure the temporal stability of a sub-band signal. One method for determining signal stationarity based on time-domain waveform characteristics is to comprehensively consider the signal's time-domain variance sequence and the differential mean between adjacent sampling points.

[0125] First, calculate the time-domain variance sequence of the subband signal. The time-domain variance reflects the degree of signal fluctuation over a period of time. Let the subband signal be x[n], where n = 0, 1, …, N-1. Divide the signal into multiple time periods, each containing L sample points. For the i-th time period, its variance σi² can be calculated using the following formula:

[0126] σi²=(1 / L)×∑[x[n]-μi]², where n is from iL to (i+1)L-1, and μi is the average value of the signal during that time period.

[0127] Then, calculate the mean difference between adjacent sampling points. The mean difference reflects the degree of change of the signal between adjacent sampling points. Assuming the difference sequence is d[n]=x[n+1]-x[n], n=0, 1, ..., N-2, the mean difference md can be calculated by the following formula:

[0128] md=(1 / (N-1))×∑d[n].

[0129] The first stationary factor is determined based on the fluctuation amplitude of the time-domain variance sequence. The fluctuation amplitude can be expressed by calculating the standard deviation of the time-domain variance sequence. Assuming the standard deviation of the time-domain variance sequence is σv, the first stationary factor s1 can be expressed as s1 = 1 / (1 + σv). A larger value of the first stationary factor indicates smaller signal fluctuations and better stationarity.

[0130] The second stationarity factor is determined based on the absolute value of the mean difference. A larger absolute value of the mean difference indicates greater signal variation between adjacent sampling points and poorer stationarity. Therefore, the second stationarity factor s2 can be expressed as s2 = 1 / (1 + |md|).

[0131] The first and second stationary factors are each subjected to a normalization transformation to be mapped to a uniform scale interval. Normalization can be performed using a linear transformation, for example, by mapping the first and second stationary factors to the interval [0, 1]. Let the normalized first stationary factor be s1' and the normalized second stationary factor be s2'.

[0132] The weight coefficients of the first and second stationary factors after normalization are dynamically assigned based on the frequency band position of the subband signal. Signals in different frequency bands may have different characteristics, so the weight coefficients need to be dynamically adjusted based on the frequency band position. Assuming the subband signal is in the low frequency band, which is generally relatively stable, the weight of the first stationary factor may be relatively large. If the signal is in the high frequency band, where fluctuations may be more dramatic, the weight of the second stationary factor may be larger. Let the weight coefficient of the first stationary factor be w1, and the weight coefficient of the second stationary factor be w2, where w1 + w2 = 1. A weight assignment function can be established based on the frequency band position. For example, the weight coefficients can be determined based on the center frequency f0 of the frequency band. Assume that the frequency range of the low frequency band is f_low to f_mid, and the frequency range of the high frequency band is f_mid to f_high. When f0 is in the low frequency band, w1 can be set to a relatively large value, such as w1 = 0.7 and w2 = 0.3; when f0 is in the high frequency band, w1 can be set to a relatively small value, such as w1 = 0.3 and w2 = 0.7. Then, the standardized first stationary factor and the second stationary factor are weighted and summed according to the weight coefficient to generate the signal stationary index S, that is, S=w1×s1'+w2×s2'.

[0133] Next, the noise energy percentage is determined. There are two ways to identify noise-dominated frequency bands. One is through a preset noise reference interval. This preset noise reference interval is determined based on past experience or prior knowledge of the target environment's noise characteristics. Frequency components within this reference interval are considered noise-dominated frequency bands. Another method is a dynamic noise estimation method based on the minimum statistics method. This method first extracts the background noise energy distribution during periods of signal silence. Signal silence refers to periods of time when the signal contains only background noise and no valid signal components. During these periods, the signal is transformed in the frequency domain to obtain the frequency domain energy distribution of the background noise. The noise energy threshold is then iteratively updated. The initial noise energy threshold can be determined based on the statistical characteristics of the background noise energy distribution, for example, taking the average background noise energy as the initial threshold. During subsequent signal processing, the signal's frequency domain energy distribution is continuously monitored. When the energy of a frequency component falls below the current noise energy threshold, it is considered noise. As the signal changes, the noise energy threshold is continuously updated to adapt to varying noise environments.

[0134] After determining the noise-dominant frequency band, calculate the energy percentage of the noise-dominant frequency band. First, calculate the total energy E_noise of all frequency components within the noise-dominant frequency band. Then, calculate the total energy E_total of the entire sub-band signal. The noise energy percentage P_noise = E_noise / E_total.

[0135] Step S133: constructing a dynamic weight factor based on the signal stationarity index and the noise energy ratio, and adjusting a preset filter parameter set according to the dynamic weight factor.

[0136] The dynamic weight factor comprehensively considers signal stability and noise, allowing for more flexible adjustment of filtering parameters. Let the dynamic weight factor be W, which is a function of the signal stability index S and the noise energy proportion P_noise. A simple functional relationship can be constructed, such as W = a × S + b × (1-P_noise), where a and b are weight coefficients, and a + b = 1. The values of a and b can be determined based on the specific application scenario and the emphasis on signal stability and noise suppression. If signal stability is more important, the value of a can be relatively large; if noise suppression is more important, the value of b can be relatively large.

[0137] The preset filter parameter set is a set of filter parameters, such as filter coefficients, that are pre-set before an adaptive filtering operation is performed. The preset filter parameter set is adjusted based on the dynamic weight factor W. Assuming the preset filter parameter set is F_set, the adjusted filter parameter set F_set' can be calculated as follows: for each parameter f_i in the filter parameter set, f_i'=f_i×W, where f_i' is the adjusted parameter. In this way, the filter parameters are dynamically adjusted according to the actual conditions of the signal, allowing the filter to better adapt to different signal characteristics and noise environments.

[0138] Step S134: Call the filtering parameter set to perform recursive filtering on the sub-band signal, and detect the pole position of the filter in real time during the filtering process to constrain its stability. If it is detected that the pole is beyond the unit circle range, reset the filtering parameters to the preset safety value to generate an initial filtered signal.

[0139] Recursive filtering is a commonly used filtering method that uses the current and past outputs of the filter along with the current input to calculate the current output. Let the subband signal be x[n] and the filter parameter set be F_set'. The recursive filter output y[n] can be calculated using the following recursive relationship: y[n] = ∑(a_i × x[ni]) + ∑(b_j × y[nj]), where a_i and b_j are the coefficients in the filter parameter set.

[0140] During the filtering process, monitoring the filter's pole locations is necessary to ensure filter stability. For high-order filters, where the pole locations are difficult to determine analytically in real time, approximate methods can be used to determine filter stability. During filter design, the filter's structure and parameters can be analyzed beforehand to determine the approximate range of the filter's poles. During recursive filtering, monitoring the filter output amplitude and changing trends can indirectly determine potential filter instability.

[0141] For example, an amplitude threshold A is set. If the amplitude of the filter output exceeds this threshold or shows a rapid increase, the filter is considered unstable. If the filter is considered unstable, the filter parameters are reset to preset safety values. The preset safety values are a set of pre-set filter parameters that ensure filter stability. After resetting the filter parameters to the preset safety values, recursive filtering is performed again until the filter output stabilizes, generating the initial filtered signal y_initial[n].

[0142] Step S135: performing short-time Fourier transform processing on the initial filtered signal to generate a time-frequency energy distribution matrix.

[0143] The short-time Fourier transform (STFT) converts a time-domain signal into a time-frequency domain signal, displaying the signal's energy distribution across time and frequency. To perform the STFT on the initial filtered signal y_initial[n], first select an appropriate window function, such as a Hanning window or Hamming window. This window function applies a windowing effect to the signal in the time domain, making it approximately stationary locally.

[0144] Let the window function be w[n], with a length of L. Multiply the initial filtered signal y_initial[n] by the window function w[n] to obtain the windowed signal y_windowed[n]. Then perform a Fourier transform on the windowed signal to obtain the frequency domain representation at each time point. Let time point be t, and perform a Fourier transform on y_windowed[n+t] to obtain Y(t, f), where f is the frequency.

[0145] Combining the frequency domain representations at different time points yields the time-frequency energy distribution matrix M. Each element of the matrix, M(t, f), represents the signal energy at time t and frequency f. The time-frequency energy distribution matrix allows us to intuitively observe how the signal energy varies at different times and frequencies.

[0146] Step S136: extracting characteristic frequency bands related to noise in the time-frequency energy distribution matrix according to the frequency band division rule of the current target decomposition level.

[0147] The frequency band division rules for the current target decomposition level define the frequency ranges of the different bands. Based on these rules, characteristic frequency bands associated with noise are extracted from the time-frequency energy distribution matrix M. The noise-dominant frequency bands have been previously identified using a preset noise reference interval or a dynamic noise estimation method based on the minimum statistics method. In the time-frequency energy distribution matrix, the frequency range corresponding to the noise-dominant frequency band is found.

[0148] Assuming the dominant noise frequency range is f_noise1 to f_noise2, extract the elements with frequencies between f_noise1 and f_noise2 from the time-frequency energy distribution matrix M to form the noise-related characteristic frequency band matrix M_noise. The characteristic frequency band matrix M_noise shows the energy distribution of the noise at different times and frequencies.

[0149] Step S137: Calculate the ratio of the average energy in the characteristic frequency band to the average energy in the signal dominant frequency band. If the ratio exceeds the preset noise threshold, it is determined that there is residual noise energy. It iteratively update the filter parameter set until the residual noise energy is lower than the preset threshold, and output the final filtered sub-band signal.

[0150] First, calculate the average energy E_noise_avg within the characteristic frequency band. Sum all elements in the characteristic frequency band matrix M_noise and divide by the number of elements in the matrix to obtain the average energy E_noise_avg within the characteristic frequency band = ∑M_noise(t,f) / N_noise, where N_noise is the number of elements in the characteristic frequency band matrix M_noise.

[0151] The dominant signal band is the frequency band with the highest signal energy and the primary information. Based on the frequency band division rules, determine the frequency range of the dominant signal band. Extract the elements corresponding to the dominant signal band from the time-frequency energy distribution matrix M to form the dominant signal band matrix M_signal. Calculate the average energy within the dominant signal band using E_signal_avg = ∑M_signal(t, f) / N_signal, where N_signal is the number of elements in the dominant signal band matrix M_signal.

[0152] Calculate the ratio of the average energy in the characteristic frequency band to the average energy in the signal's dominant frequency band (R = E_noise_avg / E_signal_avg). The preset noise threshold is a pre-set value used to determine whether residual noise energy exists. If the ratio R exceeds the preset noise threshold, residual noise energy exists.

[0153] When it is determined that there is residual noise energy, the filter parameter set is iteratively updated. The dynamic weight factor W can be adjusted according to the size of the ratio R, and then the filter parameter set F_set' can be adjusted. For example, if the ratio R is large, it means that the noise energy is relatively high, and it is necessary to increase the part of the dynamic weight factor related to noise suppression, that is, increase the value of b and decrease the value of a. Then, the filter parameter set is recalculated based on the adjusted dynamic weight factor, and the recursive filtering process, short-time Fourier transform process, feature frequency band extraction and ratio calculation steps are performed again until the ratio R is lower than the preset noise threshold. At this time, the final filtered sub-band signal y_final[n] is output.

[0154] Step S140: performing multi-band fusion processing on the filtered sub-band signals to generate a target filtered signal.

[0155] Step S141: performing signal reconstruction processing on each of the filtered sub-band signals to generate a time domain signal component of the same length as the original signal component.

[0156] The filtered sub-band signals are signals that have been processed by adaptive filtering, usually in the frequency domain. In order to fuse these filtered sub-band signals, they need to be converted back to the time domain and generate time domain signal components of the same length as the original signal components.

[0157] Each filtered subband signal y_final[n] is reconstructed using an inverse short-time Fourier transform (ISTFT). The inverse of the STF converts the time-frequency domain signal back to the time domain. The same window function w[n] as used for the STF is used for the ISTFT.

[0158] Let the time-frequency representation of the filtered subband signal be Y_final(t, f). This signal is converted back to the time-domain signal y_reconstructed[n] using the inverse short-time Fourier transform. During the reconstruction process, it is important to pay attention to the signal length, ensuring that the generated time-domain signal components are of the same length as the original signal components. This can be achieved by appropriately truncating or zero-padding the reconstructed signal.

[0159] Step S142: Dynamically determine the frequency band weight coefficient of each time domain signal component according to the target decomposition level and the signal quality score.

[0160] The target decomposition level specifies the number of frequency bands into which the signal is decomposed and the bandwidth of each band. Different frequency bands may have different importance in the signal, so the frequency band weight coefficient of each time-domain signal component needs to be dynamically determined based on the target decomposition level and signal quality score.

[0161] The signal quality score (SQS) is a metric used to evaluate the quality of each time-domain signal component. It can be calculated based on metrics such as the signal-to-noise ratio (SNR) and distortion. For example, for a particular time-domain signal component, its SNR (Signal-to-Noise Ratio) can be calculated. A higher SNR indicates better signal quality.

[0162] Assume that the target decomposition level divides the signal into m frequency bands, the time-domain signal component of the jth frequency band is y_reconstructed_j[n], and its signal quality score is Q_j. Based on the target decomposition level and signal quality score, the frequency band weight coefficient w_j for each time-domain signal component is determined. A functional relationship can be established, such as w_j = k × Q_j / ∑Q_i, where k is a normalization coefficient that ensures that the sum of all frequency band weight coefficients is 1, and i ranges from 1 to m. In this way, time-domain signal components with better signal quality have larger frequency band weight coefficients and thus account for a larger proportion in the subsequent fusion process.

[0163] Step S143: performing time domain alignment and weighted superposition processing on the time domain signal components based on the frequency band weight coefficients to generate an initial fusion signal.

[0164] Before performing weighted superposition processing, the time domain signal components need to be aligned in the time domain. Since signals in different frequency bands may have different time delays during processing, they need to be aligned in the time domain to ensure correct signal fusion.

[0165] The time domain alignment method can determine the signal delay by detecting the signal's characteristic points, such as zero crossing points and peak points. Then, the time domain signal components with different delays are shifted to align their characteristic points.

[0166] Assume that the time domain signal component of the jth frequency band after time domain alignment is y_aligned_j[n] and the frequency band weight coefficient is w_j. All time domain signal components are weighted and superimposed according to the frequency band weight coefficients to generate the initial fusion signal y_initial_fusion[n]=∑(w_j×y_aligned_j[n]), where j ranges from 1 to m.

[0167] Step S144: extracting phase difference features of signal components in adjacent frequency bands in the initial fusion signal, and dynamically calculating the phase correction amount of each frequency band based on the phase difference features.

[0168] Phase differences between adjacent frequency band signal components can affect signal fusion and cause signal distortion. Therefore, it is necessary to extract the phase difference characteristics of adjacent frequency band signal components in the initial fused signal and dynamically calculate the phase correction amount for each frequency band based on these characteristics.

[0169] Perform a frequency domain transform on the initial fusion signal y_initial_fusion[n] to obtain its frequency domain representation Y_initial_fusion(f). Based on the frequency band division rule, the frequency domain representation is divided into signal components in different frequency bands. For two adjacent frequency bands j and j+1, their phase information φ_j(f) and φ_j+1(f) is extracted.

[0170] Calculate the phase difference Δφ(f) = φ_j + 1(f) - φ_j(f) between signal components in adjacent frequency bands. Perform statistical analysis on the phase difference Δφ(f) to obtain its mean Δφ_avg and variance Δφ_var.

[0171] The phase correction amount for each frequency band is dynamically calculated based on the phase difference characteristics. If the phase difference mean Δφ_avg is large, it indicates that there is an overall phase offset, and the overall phase correction of the frequency band is required. If the phase difference variance Δφ_var is large, it indicates that there is a local phase fluctuation, and the local phase correction of the frequency band is required.

[0172] Let the phase correction amount for the jth frequency band be Δφ_correction_j(f). The phase correction amount can be determined based on the mean and variance of the phase difference. For example, for the overall phase offset, Δφ_correction_j(f) can be set to -Δφ_avg. For local phase fluctuations, the phase correction amount can be determined by interpolation or filtering based on the specific phase difference.

[0173] Step S145: Based on the phase correction amount of each frequency band, the signal components of the initial fusion signal are phase-aligned through time domain delay compensation or interpolation processing. In view of the change in signal length caused by time domain delay compensation, a cyclic shift or edge-symmetric interpolation method is used to keep the signal components equal in length to the original signal, thereby generating a phase-aligned signal.

[0174] The signal components of the initial fused signal are phase-aligned based on the phase correction amount Δφ_correction_j(f) for each frequency band. Phase alignment can be achieved through time-domain delay compensation or interpolation.

[0175] Time-domain delay compensation calculates the corresponding time delay based on the phase correction amount and then delays the signal components. Let the phase correction amount be Δφ_correction_j(f), and the corresponding time delay be Δt_j = Δφ_correction_j(f) / (2πf). The signal component y_aligned_j[n] in the jth frequency band is delayed to obtain the delayed signal component y_delayed_j[n].

[0176] Interpolation involves inserting new sampling points into a signal to adjust its phase. This can be done using methods such as linear interpolation and polynomial interpolation. For example, for two adjacent sampling points, y_aligned_j[n] and y_aligned_j[n+1], a new sampling point, y_interpolated[n+0.5], is inserted between them. This is calculated using the linear interpolation formula y_interpolated[n+0.5] = 0.5 × (y_aligned_j[n] + y_aligned_j[n+1]).

[0177] Time-domain delay compensation may cause changes in signal length. To maintain the same length of signal components as the original, cyclic shifting or edge-symmetric interpolation can be used. Circular shifting moves a portion of the signal to the other end, preserving the signal length. Edge-symmetric interpolation inserts symmetrical sampling points along the edges of the signal to fill gaps caused by delay.

[0178] After phase alignment processing, a phase-aligned signal y_phase_aligned[n] is generated.

[0179] Step S146: performing time domain smoothing processing on the phase-aligned signal to eliminate the instantaneous mutation noise introduced by the phase correction, and obtaining a corrected initial fusion signal.

[0180] The phase correction process may introduce transient mutation noise, which will affect the signal quality. In order to eliminate this noise, the phase alignment signal needs to be smoothed in the time domain.

[0181] Time-domain smoothing can be performed using methods such as moving average filtering and median filtering. Moving average filtering calculates the average of the adjacent sampling points for each signal sampling point, using this average as the smoothed value for that sampling point. Let the phase-aligned signal be y_phase_aligned[n] and the moving average filter window length be L. The smoothed signal y_smoothed[n] = (1 / L) × ∑y_phase_aligned[n+i], where i ranges from -(L-1) / 2 to (L-1) / 2.

[0182] Median filtering takes the median of its adjacent sampling points as the smoothing value of each sampling point. Through time domain smoothing, the instantaneous mutation noise introduced by phase correction is eliminated to obtain the corrected initial fusion signal y_corrected_fusion[n].

[0183] Step S147: generating the target filtered signal according to the corrected initial fusion signal.

[0184] The corrected initial fusion signal, y_corrected_fusion[n], has undergone signal reconstruction, weighted superposition, phase alignment, and time-domain smoothing to remove noise and interference while ensuring phase consistency. The corrected initial fusion signal is output as the target filtered signal. This target filtered signal has good quality, contains the useful information from the original signal, and reduces the effects of noise and interference.

[0185] Step S150: adjusting the segmentation strategy of the frequency band decomposition process or the dynamic adjustment rule of the adaptive filtering operation according to the signal quality evaluation result of the target filtered signal, and outputting the optimized target filtered signal.

[0186] Step S151: extracting the time domain signal-to-noise ratio feature, frequency band energy balance feature and phase continuity feature of the target filtered signal.

[0187] The time-domain signal-to-noise ratio (SNR) reflects the ratio of useful signal to noise in the target filtered signal. To calculate the SNR, the target filtered signal is first divided into signal segments and noise segments. The signal segment is the time period containing useful signal, while the noise segment is the time period containing only noise. This distinction can be made by examining signal characteristics such as amplitude and frequency.

[0188] Let's denote the signal in the signal segment as s[n] and the signal in the noise segment as n[n]. The energy of the signal segment can be calculated by summing the squares of the signals within the segment: signal energy E_s = sum of the squares of s[n] within the signal segment. The energy of the noise segment can be calculated by summing the squares of the signals within the noise segment: noise energy E_n = sum of the squares of n[n] within the noise segment. The time-domain signal-to-noise ratio (SNR) is the ratio of signal energy E_s to noise energy E_n.

[0189] The frequency band energy balance feature is used to measure the uniformity of the target filtered signal's energy distribution across frequency bands. First, the target filtered signal is decomposed into frequency bands to obtain the signal components of each frequency band. Assume that the target filtered signal is decomposed into m frequency bands, and the signal component of the jth frequency band is s_j[n]. The energy of this frequency band, E_j, is calculated by summing the squares of s_j[n] within the corresponding frequency band. The average energy of all frequency bands, E_avg, is calculated by summing the energy of all frequency bands, E_j, divided by the number of frequency bands, m. The deviation of each frequency band's energy from the average is then calculated, and the sum of the squares of the deviations is divided by the number of frequency bands, m, to obtain the variance, σ². The frequency band energy balance feature can be represented by the variance, σ². A smaller variance indicates a more balanced energy distribution.

[0190] The phase continuity feature reflects the degree of continuity of the phase variation of the target filtered signal between different frequency bands. The target filtered signal is transformed into a frequency domain representation, which is then divided into signal components of different frequency bands according to the frequency band division rule. For two adjacent frequency bands j and j+1, their phase information φ_j(f) and φ_j+1(f) are extracted. The phase difference Δφ(f) = φ_j+1(f) - φ_j(f) between the signal components of adjacent frequency bands is calculated. Statistical analysis of the phase difference Δφ(f) is performed to obtain the mean Δφ_avg and variance Δφ_var of the phase difference. The mean phase difference reflects the overall phase offset, while the variance reflects the local phase fluctuation.

[0191] Step S152: Calculate a first optimization index based on the time domain signal-to-noise ratio feature, calculate a second optimization index based on the frequency band energy balance feature, and calculate a third optimization index based on the phase continuity feature.

[0192] The first optimization metric is related to the time-domain signal-to-noise ratio (SNR) of the target filtered signal. A higher SNR indicates more useful information relative to noise, and therefore better signal quality. The time-domain SNR can be transformed to obtain the first optimization metric I_1. For example, a linear transformation can be performed: I_1 = k_1 × SNR, where k_1 is a scaling factor used to adjust the scale of the first optimization metric.

[0193] The second optimization metric is based on the energy balance characteristic of the frequency band. This energy balance characteristic is represented by the variance σ². A smaller variance indicates more balanced energy distribution and better signal quality. The second optimization metric I_2 can be obtained by transforming the variance σ². For example, using the inverse proportional transformation I_2 = k_2 / (1 + σ²), where k_2 is a scaling factor, the smaller the variance, the larger the value of the second optimization metric.

[0194] The third optimization metric is calculated based on the phase continuity feature. The phase continuity feature includes the phase difference mean Δφ_avg and variance Δφ_var. These two values can be combined to obtain the third optimization metric I_3. For example, a weighted summation method is used: I_3=k_31×(1 / (1+|Δφ_avg|))+k_32×(1 / (1+Δφ_var)), where k_31 and k_32 are weight coefficients, and k_31+k_32=1. By adjusting the weight coefficients, the emphasis on overall phase offset and local phase fluctuation can be adjusted according to actual needs.

[0195] Step S153: performing range normalization processing on the first optimization index, the second optimization index, and the third optimization index respectively, calculating the normalized weighted sum based on a preset weight coefficient, and generating a comprehensive optimization score.

[0196] Range normalization maps the values of each optimization metric to the interval [0, 1]. For the first optimization metric I_1, its maximum value is I_1_max and its minimum value is I_1_min. The normalized first optimization metric I_1_norm = (I_1 - I_1_min) / (I_1_max - I_1_min).

[0197] For the second optimization index I_2, its maximum value is set to I_2_max, and its minimum value is set to I_2_min. The normalized second optimization index I_2_norm=(I_2-I_2_min) / (I_2_max-I_2_min).

[0198] For the third optimization index I_3, its maximum value is set to I_3_max, and its minimum value is set to I_3_min. The normalized third optimization index I_3_norm=(I_3-I_3_min) / (I_3_max-I_3_min).

[0199] Preset weight coefficients are used to determine the importance of each normalized optimization metric in the overall optimization score. Let the weight coefficient for the first optimization metric be w_1, the weight coefficient for the second optimization metric be w_2, and the weight coefficient for the third optimization metric be w_3, with w_1 + w_2 + w_3 = 1. The overall optimization score (Score) = w_1 × I_1_norm + w_2 × I_2_norm + w_3 × I_3_norm.

[0200] Step S154: If the comprehensive optimization score is lower than a preset optimization threshold, at least one of the following adjustment operations is performed.

[0201] The preset optimization threshold is a pre-set value used to determine whether the quality of the target filter signal meets the requirements. If the comprehensive optimization score is lower than the preset optimization threshold, it means that the quality of the target filter signal can be improved and adjustment operations are required.

[0202] (1) Adjusting the number of frequency bands or bandwidth range in the frequency band division rule according to the energy balance characteristics of the frequency bands.

[0203] Step S1541: Obtain the energy proportion sequence and standard deviation of each sub-band according to the frequency band energy balance characteristics of the target filtered signal.

[0204] We previously calculated the energy E_j of each frequency band of the target filtered signal. The total energy E_total = the sum of all band energies E_j. The energy contribution of each sub-band, P_j, = E_j / E_total. This gives us the energy contribution sequence {P_1, P_2, …, P_m} for each sub-band.

[0205] Calculate the standard deviation σ_P of the energy contribution sequence. First, calculate the average value P_avg of the energy contribution sequence by summing all energy contributions P_j and dividing it by the number of frequency bands m. Then calculate the deviation of each energy contribution from the average value. The sum of the squared deviations is then divided by the number of frequency bands m to obtain the variance σ_P². The standard deviation σ_P is the square root of the variance σ_P².

[0206] Step S1542: Determine whether the standard deviation exceeds a preset equalization threshold; if so, perform the following frequency band adjustment operation.

[0207] The preset equalization threshold is a pre-set value used to determine whether the frequency band energy distribution is balanced. If the standard deviation σ_P exceeds the preset equalization threshold, it indicates that the frequency band energy distribution is uneven and the frequency band needs to be adjusted.

[0208] (a) Based on the number of frequency bands and the bandwidth range of each frequency band in the current frequency band division rules, determine the upper limit of the number of frequency bands allowed for adjustment and the minimum bandwidth constraint value.

[0209] The current frequency band division rules specify the number of frequency bands (m) and the bandwidth range of each band. The upper limit N_max of the number of frequency bands that can be adjusted is a pre-set value based on actual conditions and processing capabilities; the number of frequency bands cannot be increased indefinitely. The minimum bandwidth constraint B_min is also pre-set to ensure that the bandwidth of each frequency band is not too small, otherwise it will increase processing complexity and computational effort.

[0210] (b) Filter out the first target frequency band corresponding to the highest energy ratio and the second target frequency band corresponding to the lowest energy ratio from the energy ratio sequence.

[0211] In the energy proportion sequence {P_1, P_2, ..., P_m}, the maximum value P_max and the minimum value P_min are found by comparing the values of each energy proportion. The frequency band corresponding to the maximum value P_max is the first target frequency band, and the frequency band corresponding to the minimum value P_min is the second target frequency band.

[0212] (c) performing frequency band splitting on the first target frequency band: calculating a maximum number of splittable sub-frequency bands based on the minimum bandwidth constraint value, and dividing the frequency band into a plurality of sub-frequency bands according to energy distribution, ensuring that the bandwidth of each sub-frequency band is not less than the minimum bandwidth constraint value.

[0213] Assume the bandwidth of the first target frequency band is B_1. Based on the minimum bandwidth constraint value B_min, the maximum number of sub-bands that can be split is n_max = (B_1 / B_min), rounded down. Further analysis is performed on the signal within the first target frequency band, dividing the frequency range into several smaller frequency intervals and calculating the energy in each smaller frequency interval. Based on the energy distribution, regions with large energy differences are divided into different sub-bands. During the division process, ensure that the bandwidth of each sub-band is not less than the minimum bandwidth constraint value B_min.

[0214] (d) performing frequency band merging on the second target frequency band: identifying the energy proportion of its adjacent frequency bands; if the energy proportion of the adjacent frequency bands is lower than a preset merging threshold, and the bandwidth of the new frequency band after merging does not violate the minimum bandwidth constraint and the maximum number of frequency bands, then merging the second target frequency band and the adjacent frequency band into a new frequency band, and the bandwidth of the new frequency band is the sum of the two.

[0215] The preset merging threshold is a value pre-set based on historical data and experience, and is used to determine whether to merge frequency bands. For the second target frequency band, its adjacent frequency bands are identified. Let the energy proportions of the adjacent frequency bands be P_adj1 and P_adj2, respectively. If P_adj1 or P_adj2 is lower than the preset merging threshold, and the bandwidth of the new frequency band after merging is not less than the minimum bandwidth constraint value B_min, and the number of frequency bands does not exceed the upper limit N_max of the number of frequency bands allowed for adjustment, then the second target frequency band and the corresponding adjacent frequency band are merged into a new frequency band. The bandwidth of the new frequency band is the sum of the bandwidth of the second target frequency band and the bandwidth of the adjacent frequency band.

[0216] (e) updating the number of frequency bands and the bandwidth of each frequency band in the frequency band division rule, and recalculating the standard deviation of the adjusted energy proportion sequence of each sub-band.

[0217] After the frequency band splitting and merging operations, update the number of frequency bands and the bandwidth of each frequency band in the frequency band division rule. Perform frequency band decomposition of the target filtered signal again, calculating the energy E_j' of each frequency band. The total energy E_total' is equal to the sum of all frequency band energies E_j'. The energy contribution of each sub-band, P_j', is calculated as E_j' / E_total'. This yields the adjusted energy contribution sequence {P_1', P_2', …, P_m'} for each sub-band. Recalculate the standard deviation σ_P' of this adjusted energy contribution sequence using the same method as above.

[0218] Iterate steps (a) to (e) until the standard deviation is lower than the equalization threshold or reaches the upper limit of the number of frequency bands, and output the adjusted frequency band division rule.

[0219] Repeat the above frequency band adjustment operation, calculating the standard deviation of the energy proportion sequence after each adjustment, until the standard deviation σ_P' falls below the equalization threshold or the number of frequency bands reaches the upper limit of the number of frequency bands allowed to be adjusted N_max. At this time, the adjusted frequency band division rule is output and subsequently used for frequency band decomposition processing.

[0220] (2) Adjusting the dynamic weight factor calculation rule in the adaptive filtering operation according to the time domain signal-to-noise ratio characteristics.

[0221] Step S1543: According to the time-domain signal-to-noise ratio characteristics of the target filtered signal, the ratio of the average energy of the signal-dominant time-domain window to the average energy of the noise time-domain window is obtained.

[0222] When calculating the time-domain SNR, signal segments and noise segments are already defined. The signal-dominant time-domain window is the signal segment, and the noise-dominant time-domain window is the noise segment. The average energy (E_s_avg) in the signal-dominant time-domain window is calculated by dividing the energy (E_s) of the signal segment by the length of the signal segment. The average energy (E_n_avg) in the noise time-domain window is calculated by dividing the energy (E_n) of the noise segment by the length of the noise segment. The ratio (R_energy) is calculated as E_s_avg / E_n_avg.

[0223] Step S1544: Map the ratio to a filtering strength level, wherein the filtering strength level is negatively correlated with the ratio.

[0224] Set multiple filter strength levels, for example, divided into three levels: low, medium, and high. Map the ratio R_energy to the corresponding filter strength level according to its size. Since the filter strength level is negatively correlated with the ratio, that is, the larger the ratio, the more useful information in the signal relative to the noise, and the filter strength can be relatively low; the smaller the ratio, the more noise there is, and a higher filter strength is required. Mapping can be performed by setting different threshold intervals. For example, when R_energy is greater than the threshold T_1, it is mapped to a low filter strength level; when R_energy is between the thresholds T_1 and T_2, it is mapped to a medium filter strength level; when R_energy is less than the threshold T_2, it is mapped to a high filter strength level.

[0225] Step S1545: adjusting the proportional relationship between the noise energy weight coefficient and the signal stationarity weight coefficient in the dynamic weight factor calculation formula based on the filtering strength level.

[0226] The dynamic weight factor calculation formula is W = a × S + b × (1-P_noise), where a is the signal stationarity weight coefficient, b is the noise energy weight coefficient, and a + b = 1.

[0227] (f) When the filtering intensity level is high, the proportion of the noise energy weight coefficient is increased so that its proportion in the dynamic weight factor is raised to a preset upper limit.

[0228] The preset upper limit is a pre-set value that limits the maximum value of the noise energy weighting factor. When the filter strength level is high, it indicates relatively high noise levels and requires greater emphasis on noise suppression. Increase the noise energy weighting factor b to the preset upper limit b_max. Accordingly, the signal stationarity weighting factor a = 1 - b_max.

[0229] (g) When the filter strength level is low, the proportion of the signal stability weight coefficient is increased so that its proportion in the dynamic weight factor is raised to the preset lower limit.

[0230] The preset lower limit is a pre-set value that limits the minimum value of the signal stationarity weight coefficient. When the filtering strength level is low, it indicates that the signal contains relatively more useful information and does not require excessive filtering. Increase the signal stationarity weight coefficient a to the preset lower limit a_min. Correspondingly, the noise energy weight coefficient b = 1 - a_min.

[0231] Step S1546: Substitute the adjusted weight coefficient ratio into the dynamic weight factor calculation formula to generate a new dynamic weight factor.

[0232] Substitute the adjusted signal stationarity weight coefficient a and noise energy weight coefficient b into the dynamic weight factor calculation formula W=a×S+b×(1-P_noise) to generate a new dynamic weight factor W'.

[0233] Step S1547: updating the filter parameter set of the adaptive filtering operation according to the new dynamic weight factor, and applying it to the next round of recursive filtering processing of the sub-band signal.

[0234] According to the new dynamic weight factor W', the filter parameter set F_set'' for the adaptive filtering operation is updated according to the above method for adjusting the filter parameter set. When the new sub-band signal is subsequently recursively filtered, the updated filter parameter set F_set'' is used.

[0235] (3) Adjusting the phase correction parameters in the multi-band fusion process according to the phase continuity characteristics.

[0236] Step S1548: According to the phase continuity characteristics of the target filtered signal, the mean and variance of the phase differences of the adjacent frequency band signal components in the overlapping time domain window are obtained.

[0237] When calculating the phase continuity feature previously, we already obtained the phase difference Δφ(f) between the adjacent frequency band signal components. For the phase difference between the adjacent frequency band signal components within the overlapping time domain window, extract the phase difference data within the window and calculate its mean Δφ_avg_overlap and variance Δφ_var_overlap.

[0238] Step S1549: determining the overall phase offset based on the phase difference mean, and determining the local phase fluctuation intensity based on the variance.

[0239] To solve the phase wrapping problem, a circular statistical method is used to calculate the mean and variance of the phase difference. The phase difference Δφ(f) between signal components in adjacent frequency bands is first converted to the main value interval (-π to π).

[0240] To determine the overall phase offset, a weighted average method is used to calculate the average phase difference. Assume the phase difference sequence is {Δφ_1, Δφ_2, …, Δφ_N}. First, convert each phase difference Δφ_i to a principal value interval, and then calculate the average phase difference Δφ_avg. The calculation process is as follows: First, convert all phase differences into complex form, that is, z_i = exp(j × Δφ_i). Then, calculate the average value of the complex number Z_avg = (∑z_i) / N. Finally, calculate the phase of Z_avg to obtain the average phase difference Δφ_avg = angle(Z_avg). This average phase difference is the overall phase offset.

[0241] To determine the strength of local phase fluctuations, the variance calculation is also based on the phase differences converted to the principal value interval. First, the deviation of each phase difference from the average phase difference is calculated. Again, these deviations are converted to the principal value interval. The squared sum of these deviations is then divided by the number of samples to obtain the variance Δφ_var. The strength of local phase fluctuations can be expressed using this variance Δφ_var. A larger variance indicates more severe local phase fluctuations.

[0242] Perform the following phase correction parameter adjustment operations.

[0243] (h) If the phase difference mean exceeds a first correction threshold, the time domain delay compensation amount for each frequency band is calculated based on the overall phase offset. For non-integer sample point shift requirements, a polynomial interpolation method is used to generate a continuous time domain signal, and phase alignment is achieved through resampling.

[0244] The first correction threshold is a pre-set value used to determine whether there is a significant overall phase offset. If the average phase difference, Δφ_avg_overlap, exceeds the first correction threshold, it indicates a significant overall phase offset. The time domain delay compensation, Δt, is calculated for each frequency band based on the overall phase offset, Δφ_avg_overlap. For non-integer sample point shifts, polynomial interpolation is used to insert new sampling points into the signal to generate a continuous time domain signal. Then, through resampling, the signal is shifted by the time domain delay compensation amount to achieve phase alignment.

[0245] (i) If the variance exceeds a second correction threshold, interpolation points are inserted in the overlapping time domain windows, and a smooth transition curve is generated based on the interpolated phase sequence to eliminate local phase mutations.

[0246] The second correction threshold is a pre-set value used to determine whether there are large local phase fluctuations. If the variance Δφ_var_overlap exceeds the second correction threshold, it indicates a significant local phase abrupt change. Interpolation points are inserted within the overlapping time domain window, and a new phase value is generated using an interpolation algorithm. Based on the interpolated phase sequence, a smooth transition curve is generated using methods such as curve fitting to eliminate local phase abrupt changes.

[0247] Then, the time domain delay amount or interpolation point density in the phase correction parameters is updated, and the phase difference mean and variance are recalculated.

[0248] According to the above phase correction parameter adjustment operation, the time domain delay or interpolation point density in the phase correction parameter is updated. The target filtered signal is processed again to calculate the phase difference mean and variance of the adjacent frequency band signal components in the overlapping time domain window.

[0249] The parameters are iteratively adjusted until the phase difference mean and variance are both lower than corresponding thresholds, and an optimized phase correction parameter set is output.

[0250] Repeat the above steps of adjusting the phase correction parameters and calculating the mean and variance of the phase difference until both the mean and variance of the phase difference fall below their corresponding thresholds (the first and second correction thresholds). At this point, the optimized phase correction parameter set is output and subsequently used for phase correction in the multi-band fusion process.

[0251] After the possible adjustments described above, the adjusted frequency band division rules are used for frequency band decomposition, the updated filter parameter set is used for adaptive filtering, and the optimized phase correction parameter set is used for multi-band fusion processing, ultimately resulting in an optimized target filtered signal. This optimized target filtered signal has improved signal-to-noise ratio (SNR) in the time domain, frequency band energy balance, and phase continuity, resulting in better signal quality.

[0252] Figure 2 A schematic diagram illustrates exemplary hardware and software components of an adaptive signal filtering system 100 based on multi-band fusion, which can implement the concepts of the present application, as provided in some embodiments of the present application. For example, the processor 120 can be used in the adaptive signal filtering system 100 based on multi-band fusion and perform the functions of the present application.

[0253] The adaptive signal filtering system 100 based on multi-band fusion can be a general-purpose server or a special-purpose server, both of which can be used to implement the adaptive signal filtering method based on multi-band fusion of the present application. Although only one server is shown in this application, for convenience, the functions described in this application can be implemented in a distributed manner on multiple similar platforms to balance the processing load.

[0254] For example, the adaptive signal filtering system 100 based on multi-band fusion can include a network port 110 connected to a network, one or more processors 120 for executing program instructions, a communication bus 130, and storage media 140 in different forms, such as a disk, ROM, or RAM, or any combination thereof. Exemplarily, the adaptive signal filtering system 100 based on multi-band fusion can also include program instructions stored in ROM, RAM, or other types of non-transitory storage media, or any combination thereof. The method of the present application can be implemented according to these program instructions. The adaptive signal filtering system 100 based on multi-band fusion also includes an I / O interface 150 between the computer and other input and output devices.

[0255] For ease of explanation, only one processor is described in the adaptive signal filtering system 100 based on multi-band fusion. However, it should be noted that the adaptive signal filtering system 100 based on multi-band fusion in the present application may also include multiple processors, so the steps performed by one processor described in the present application may also be performed jointly or individually by multiple processors. For example, if the processor of the adaptive signal filtering system 100 based on multi-band fusion performs step A and step B, it should be understood that step A and step B may also be performed jointly by two different processors or performed individually in one processor. For example, the first processor performs step A, the second processor performs step B, or the first processor and the second processor perform steps A and B together.

[0256] In addition, an embodiment of the present invention further provides a readable storage medium, in which computer-executable instructions are preset. When a processor executes the computer-executable instructions, the above-mentioned adaptive signal filtering method based on multi-band fusion is implemented.

[0257] It should be noted that in order to simplify the description of the present invention and thus help understand one or more embodiments of the invention, in the foregoing description of the embodiments of the present invention, multiple features are sometimes combined into one embodiment, figure or description thereof.

Claims

1. An adaptive signal filtering method based on multi-band fusion, characterized in that: The method comprises: Acquire a multi-band signal set of a target environment, wherein the multi-band signal set includes original signal components of different frequency bands; Performing frequency band decomposition processing on the multi-band signal set to obtain a plurality of sub-band signals; Performing an adaptive filtering operation on each of the sub-band signals to generate a corresponding filtered sub-band signal; wherein the adaptive filtering operation includes dynamically adjusting a filter coefficient according to a noise characteristic of the sub-band signal; Performing multi-band fusion processing on the filtered sub-band signals to generate a target filtered signal; Adjusting the segmentation strategy of the frequency band decomposition processing or the dynamic adjustment rule of the adaptive filtering operation according to the signal quality evaluation result of the target filtered signal, and outputting an optimized target filtered signal; The performing frequency band decomposition processing on the multi-band signal set to obtain a plurality of sub-band signals includes: Determine a target decomposition level according to an initial frequency band division rule, and dynamically iteratively adjust the number of frequency bands and bandwidth range of the target decomposition level based on the spectral characteristics of the multi-band signal set during signal processing, wherein if the frequency band energy balance characteristics after dynamic iterative adjustment do not meet a preset threshold, further optimize the frequency band division rule based on the frequency band energy proportion; performing overlapping segmentation processing on each original signal component in the multi-band signal set based on the target decomposition level to generate an initial segmented signal; Performing a frequency domain transform operation on the initial segmented signal to obtain a frequency domain energy distribution feature; According to the matching degree between the frequency domain energy distribution feature and the frequency band division rule, the initial segmented signal is mapped to the corresponding target frequency band to generate the sub-band signal.

2. The adaptive signal filtering method based on multi-band fusion according to claim 1, characterized in that: The performing overlapping segmentation processing on each original signal component in the multi-band signal set based on the target decomposition level to generate an initial segmented signal includes: Determining a sampling rate of the original signal component, and calculating the number of sample points corresponding to a segmented window according to a minimum bandwidth range in the target decomposition level and the sampling rate, such that the duration of the segmented window is proportional to the inverse of the minimum bandwidth; Determining the number of overlapping sample points between adjacent segment windows based on the number of frequency bands at the target decomposition level, wherein the number of overlapping sample points is in negative proportion to the number of frequency bands; A sliding window sequence with equal intervals is generated according to the number of sample points and the number of overlapping sample points in the segmented window.

3. The adaptive signal filtering method based on multi-band fusion according to claim 1, characterized in that: The performing of an adaptive filtering operation on each of the sub-band signals to generate a corresponding filtered sub-band signal comprises: Extracting time domain waveform features and frequency domain energy features of the sub-band signal; Determining a signal stationarity index based on the time domain waveform characteristics, and determining a noise energy proportion based on the energy proportion of the noise-dominant frequency band in the frequency domain energy characteristics, wherein the noise-dominant frequency band is identified by a preset noise reference interval or a dynamic noise estimation method based on a minimum statistics method, specifically including extracting background noise energy distribution in a signal silence segment and iteratively updating a noise energy threshold; Constructing a dynamic weight factor based on the signal stationarity index and the noise energy ratio, and adjusting a preset filter parameter set according to the dynamic weight factor; Recursively filtering the sub-band signal by calling the set of filtering parameters, and detecting the position of the filter poles in real time during the filtering process to constrain the stability of the filter. If the poles are detected to be outside the unit circle, resetting the filtering parameters to preset safe values to generate an initial filtered signal. Performing short-time Fourier transform processing on the initial filtered signal to generate a time-frequency energy distribution matrix; According to the frequency band division rules of the current target decomposition level, the characteristic frequency bands related to noise in the time-frequency energy distribution matrix are extracted; The ratio of the average energy in the characteristic frequency band to the average energy in the signal dominant frequency band is calculated. If the ratio exceeds a preset noise threshold, it is determined that residual noise energy exists. The filter parameter set is iteratively updated until the residual noise energy is lower than the preset threshold, and the final filtered sub-band signal is output.

4. The adaptive signal filtering method based on multi-band fusion according to claim 3, characterized in that: Determining the signal stability index according to the time domain waveform characteristics includes: Calculating the time domain variance sequence of the sub-band signal and the difference mean between adjacent sampling points; Determining a first stationary factor according to the fluctuation amplitude of the time domain variance sequence; determining a second stability factor according to the absolute value of the difference mean; performing standardization conversion processing on the first stationary factor and the second stationary factor respectively to map them to a uniform proportional interval, and dynamically allocating weight coefficients of the first stationary factor and the second stationary factor after the standardization conversion according to the frequency band position of the sub-band signal, and then performing weighted summation to generate the signal stationarity index; The weight coefficients of the first stationary factor and the second stationary factor are dynamically allocated according to the frequency band position of the sub-band signal.

5. The adaptive signal filtering method based on multi-band fusion according to claim 1, characterized in that: The performing multi-band fusion processing on the filtered sub-band signal to generate a target filtered signal includes: Performing signal reconstruction processing on each of the filtered sub-band signals to generate a time domain signal component of the same length as the original signal component; Dynamically determining a frequency band weight coefficient for each time domain signal component according to the target decomposition level and the signal quality score; Performing time domain alignment and weighted superposition processing on the time domain signal components based on the frequency band weight coefficients to generate an initial fusion signal; Extracting phase difference features of adjacent frequency band signal components in the initial fusion signal, and dynamically calculating the phase correction amount of each frequency band based on the phase difference features; Based on the phase correction amount of each frequency band, phase-align the signal components of the initial fusion signal through time domain delay compensation or interpolation processing, and use cyclic shift or edge symmetric interpolation method to keep the signal components equal in length to the original signal to generate a phase-aligned signal to address the change in signal length caused by time domain delay compensation; Performing time domain smoothing on the phase alignment signal to eliminate instantaneous mutation noise introduced by phase correction, thereby obtaining a corrected initial fusion signal; The target filtered signal is generated according to the corrected initial fusion signal.

6. The adaptive signal filtering method based on multi-band fusion according to claim 1, characterized in that: The adjusting, according to the signal quality evaluation result of the target filtered signal, the segmentation strategy of the frequency band decomposition processing or the dynamic adjustment rule of the adaptive filtering operation includes: Extracting the time domain signal-to-noise ratio feature, frequency band energy balance feature and phase continuity feature of the target filtered signal; Calculating a first optimization index based on the time domain signal-to-noise ratio feature, calculating a second optimization index based on the frequency band energy balance feature, and calculating a third optimization index based on the phase continuity feature; Performing range normalization processing on the first optimization index, the second optimization index, and the third optimization index, respectively, calculating a normalized weighted sum based on a preset weight coefficient, and generating a comprehensive optimization score; If the comprehensive optimization score is lower than the preset optimization threshold, at least one of the following adjustment operations is performed: (1) adjusting the number of frequency bands or bandwidth range in the frequency band division rule according to the energy balance characteristics of the frequency bands; (2) adjusting a dynamic weight factor calculation rule in the adaptive filtering operation according to the time domain signal-to-noise ratio characteristics; (3) Adjusting the phase correction parameters in the multi-band fusion process according to the phase continuity characteristics.

7. The adaptive signal filtering method based on multi-band fusion according to claim 6, characterized in that: The adjusting the number of frequency bands or bandwidth range in the frequency band division rule according to the frequency band energy balance characteristics includes: According to the frequency band energy balance characteristics of the target filtered signal, the energy proportion sequence and standard deviation of each sub-band are obtained; Determine whether the standard deviation exceeds a preset equalization threshold. If so, perform the following frequency band adjustment operations: (a) Determine the upper limit of the number of frequency bands allowed for adjustment and the minimum bandwidth constraint value based on the number of frequency bands and the bandwidth range of each frequency band in the current frequency band division rules; (b) selecting a first target frequency band corresponding to the highest energy ratio and a second target frequency band corresponding to the lowest energy ratio from the energy ratio sequence; (c) splitting the first target frequency band: calculating a maximum number of splittable sub-frequency bands based on the minimum bandwidth constraint value, and dividing the frequency band into a plurality of sub-frequency bands according to an energy proportion gradient, ensuring that the bandwidth of each sub-frequency band is not less than the minimum bandwidth constraint value; (d) performing frequency band merging on the second target frequency band: identifying energy proportions of adjacent frequency bands, and if the energy proportions of the adjacent frequency bands are lower than a preset merging threshold, merging the second target frequency band and the adjacent frequency band into a new frequency band, where the bandwidth of the new frequency band is the sum of the energy proportions of the adjacent frequency bands; (e) updating the number of frequency bands and the bandwidth of each frequency band in the frequency band division rules, and recalculating the standard deviation of the adjusted energy proportion sequence of each sub-band; Iterate steps (a) to (e) until the standard deviation is lower than the equalization threshold or reaches the upper limit of the number of frequency bands, and output the adjusted frequency band division rule.

8. The adaptive signal filtering method based on multi-band fusion according to claim 6, characterized in that: The step of adjusting a dynamic weight factor calculation rule in the adaptive filtering operation according to the time domain signal-to-noise ratio feature includes: According to the time domain signal-to-noise ratio characteristics of the target filtered signal, the ratio of the average energy of the signal-dominant time domain window to the average energy of the noise time domain window is obtained; mapping the ratio to a filter strength level, wherein the filter strength level is negatively correlated to the ratio; Adjusting the proportional relationship between the noise energy weight coefficient and the signal stationarity weight coefficient in the dynamic weight factor calculation formula based on the filtering strength level specifically includes: (f) When the filter strength level is high, the proportion of the noise energy weight coefficient is increased so that its proportion in the dynamic weight factor is increased to a preset upper limit; (g) When the filter strength level is low, increase the proportion of the signal stability weight coefficient so that its proportion in the dynamic weight factor is raised to a preset lower limit; Substitute the adjusted weight coefficient ratio into the dynamic weight factor calculation formula to generate a new dynamic weight factor; The filter parameter set of the adaptive filtering operation is updated according to the new dynamic weight factor and applied to the next round of recursive filtering processing of the sub-band signal.

9. An adaptive signal filtering system based on multi-band fusion, characterized in that: It includes a processor and a memory, the memory is connected to the processor, the memory is used to store programs, instructions or codes, and the processor is used to execute the programs, instructions or codes in the memory to implement the adaptive signal filtering method based on multi-band fusion as described in any one of claims 1 to 8.

Images