Network signal interference detection system based on spectral analysis

Abstract

The invention discloses a network signal interference detection system based on spectrum analysis. The system comprises a network signal acquisition module, a spectrum preprocessing module, a spectrum calculation module, a frequency separation spectrum kurtosis analysis module, a Bayesian online change point detection module, a parameter linkage module, an interference discrimination module and a result output module. The frequency separation spectrum kurtosis analysis module outputs kurtosis statistics and noise bottom estimation according to time slices based on the power spectrum; the Bayesian online change point detection module updates and outputs the change point position, the confidence coefficient and the adaptive time window parameter online; the parameter linkage module updates the frequency separation spectrum kurtosis analysis module according to the data; and the interference judgment module generates an interference event list, and the result output module outputs alarm and troubleshooting positioning data. According to the invention, automatic detection and discrimination of same-frequency superposition, adjacent-frequency superposition, burst intermittence and noise bottom uplift interference are realized, the accuracy is improved, and the troubleshooting cost is reduced.

Description

A network signal interference detection system based on spectrum analysis Technical Field

[0001] This invention relates to the field of communication engineering technology, and in particular to a network signal interference detection system based on spectrum analysis. Background Technology

[0002] In the operation and maintenance of mobile communication networks, dedicated wireless networks, and broadband access networks, spectrum monitoring and interference detection are crucial for ensuring link quality and improving network availability. Existing network-side and test-side equipment often samples received signals and calculates the power spectrum, then uses power changes over time and frequency to identify anomalies and output alarms. Interference sources include co-channel carrier superposition, adjacent-channel leakage, intermodulation products, external radiation intrusion, and sudden energy injection introduced by terminal and equipment anomalies. Interference patterns exhibit frequency domain superposition and time domain intermittency, leading to decreased throughput, latency jitter, and increased call drop rates. Existing technologies use noise floor estimation and fixed threshold comparison to detect energy anomalies, and combine sliding window statistics to construct indicators such as duration and bandwidth usage for interference discrimination; or introduce time-frequency domain block processing and multi-scale smoothing to suppress random noise, and use change point detection or anomaly segment extraction to trigger alarms; other solutions utilize higher-order statistics to characterize non-Gaussian properties, through spectral slackness... The above methods enhance narrowband interference and impulse interference using degree or related statistics, and then combine rule bases or template matching to output interference types. In engineering implementation, the above schemes are often linked with network management systems to form a closed-loop output of alarm and troubleshooting data. When co-channel and adjacent-channel interference are superimposed and accompanied by a multi-carrier structure, single noise floor thresholds and energy indicators are prone to false alarms and are difficult to reliably distinguish superposition relationships. When interference is sudden and intermittent with irregular duration, fixed window statistics and fixed thresholds are not compatible with short-term abrupt changes and long-term drifts and require repeated manual parameter tuning. When the noise floor rises and drifts slowly with the environment and load, the lag and drift of the noise floor estimation will lead to inconsistent alarm triggering and reduce the reliability of positioning. At the same time, the lack of an adaptive update mechanism for frequency band division and time slice division based on observation results makes it difficult to optimize feature extraction and change point determination in conjunction with each other and increases troubleshooting costs. Therefore, how to provide a network signal interference detection system based on spectrum analysis is a problem that urgently needs to be solved by those skilled in the art. Summary of the Invention

[0003] One objective of this invention is to propose a network signal interference detection system based on spectrum analysis. This invention enables automatic detection and discrimination of interference such as co-frequency superposition, adjacent frequency superposition, sudden intermittent interference, and noise floor rise, thereby improving accuracy and reducing troubleshooting costs.

[0004] According to an embodiment of the present invention, a network signal interference detection system based on spectrum analysis includes: a network signal acquisition module, which receives the signal of the network communication link under test and outputs a digitized time-domain sampling sequence; a spectrum preprocessing module, which performs window function segmentation, DC component removal, and amplitude normalization on the digitized time-domain sampling sequence and outputs a spectrum calculation input sequence; a spectrum calculation module, which performs discrete Fourier transform on the spectrum calculation input sequence to generate a power spectrum sequence; and a frequency separation spectrum kurtosis analysis module, which performs frequency separation processing on the power spectrum sequence to obtain multiple separated frequency band power spectrum sets, and forms a kurtosis statistics sequence and a noise floor estimation sequence according to time slices. The system comprises the following modules: a Bayesian online change point detection module, which performs online posterior updates on the kurtosis statistics sequence and the noise floor estimation sequence, and outputs the change point location sequence, change point confidence sequence, and adaptive time window parameters; a parameter linkage module, which receives the change point location sequence, change point confidence sequence, and adaptive time window parameters, updates the frequency separation spectrum kurtosis analysis module, and recalculates the kurtosis statistics sequence and the noise floor estimation sequence; an interference discrimination module, which receives the kurtosis statistics sequence, the noise floor estimation sequence, the change point location sequence, and the change point confidence sequence, and generates a list of interference events; and a result output module, which receives the list of interference events and outputs alarm information and troubleshooting location data.

[0005] Optionally, the network signal acquisition module specifically comprises: accessing the network communication link signal under test through an RF coupling interface and allocating it to a first sampling channel and a second sampling channel; performing variable bandwidth analog pre-selection filtering in the first sampling channel and determining the center frequency and passband bandwidth according to a preset frequency band table, and outputting a first filtered RF signal; generating a reference noise RF signal through an isolation terminal in the second sampling channel; performing parallel dual-gain amplification and automatic gain control on the first filtered RF signal to obtain a first high-gain intermediate frequency signal and a first low-gain intermediate frequency signal; performing in-phase quadrature demodulation on the first high-gain intermediate frequency signal and the first low-gain intermediate frequency signal to obtain a first high-gain baseband sampling pair and a first low-gain baseband sampling pair; performing in-phase quadrature demodulation on the reference noise RF signal to obtain a reference baseband sampling pair; performing unified clock sampling on the first high-gain baseband sampling pair, the first low-gain baseband sampling pair, and the reference baseband sampling pair to obtain a synchronous sampling sequence; performing amplitude and phase imbalance calibration and DC bias calibration on the synchronous sampling sequence and performing cross-gain fusion to generate a digital time-domain sampling sequence.

[0006] Optionally, the spectrum preprocessing module specifically comprises: receiving a digitized time-domain sampling sequence and generating a frame index sequence and a segmented time-domain frame sequence according to a preset frame length and a preset frame shift; performing endpoint continuity detection on the segmented time-domain frame sequence and calculating the amplitude difference between the first and last samples and the amplitude difference between the first and last first-order differences of each segmented time-domain frame, and calculating a windowing coefficient sequence; applying the windowing coefficient sequence to the segmented time-domain frame sequence to generate a windowed time-domain frame sequence, calculating the corresponding complex mean, and using the complex mean to perform DC component cancellation on the windowed time-domain frame sequence to generate a de-DC time-domain frame sequence; and calculating the amplitude square sequence on the de-DC time-domain frame sequence. The algorithm generates a normalized amplitude sequence using robust scaling estimates. Based on the normalized amplitude sequence, it performs amplitude normalization on the de-DC time-domain frame sequence to generate a normalized time-domain frame sequence, and then performs sample compression to generate a compressed time-domain frame sequence. It extracts the overlapping region from adjacent compressed time-domain frame sequences and calculates the cross-correlation function. Based on the maximum correlation hysteresis of the cross-correlation function and parabolic interpolation, it obtains the resampling ratio sequence, and performs fractional delay filtering resampling on the compressed time-domain frame sequence to generate the spectrum calculation input frame sequence. Finally, it concatenates the spectrum calculation input frame sequences according to the frame index sequence to generate the spectrum calculation input sequence.

[0007] Optionally, the execution of the Discrete Fourier Transform specifically involves: receiving the spectrum calculation input sequence and writing it into a circular buffer; forming a time-domain block sequence according to the number of transform points; selecting a twitch factor lookup table based on the sampling rate identifier; performing bit-reversal rearrangement on the time-domain block sequence and distributing the rearrangement result to a parallel butterfly operation array; performing butterfly operations in a multi-stage pipeline within the parallel butterfly operation array and calling the twitch factor lookup table at each stage to complete complex multiplication and addition, thereby obtaining a frequency-domain complex spectrum sequence; performing block floating-point scaling on the frequency-domain complex spectrum sequence to obtain a calibrated frequency-domain complex spectrum sequence; performing complex conjugate product on the calibrated frequency-domain complex spectrum sequence to obtain power spectrum unit values ​​and arranging them according to frequency index to form a single-frame power spectrum; and concatenating the single-frame power spectrum in the order of the time-domain block sequence to output a power spectrum sequence.

[0008] Optionally, the frequency separation spectral kurtosis analysis module specifically comprises: receiving a power spectrum sequence and constructing a frame power spectrum sequence in chronological order; calculating a frequency change indicator sequence along the frequency dimension of the frame power spectrum sequence; performing continuous segment clustering on the frequency index based on the frequency change indicator sequence and generating a frequency band boundary sequence; constructing multiple non-overlapping separation frequency band definition sets using the frequency band boundary sequence; performing frequency band extraction on the frame power spectrum sequence according to the separation frequency band definition sets and forming multiple separation frequency band power spectrum sets, and dividing them into frequency band time slice power block sequences according to time slices; calculating the mean sequence, variance sequence, and fourth-order central moment sequence for the frequency band time slice power block sequences, and generating a kurtosis statistics sequence based on the mean sequence, variance sequence, and fourth-order central moment sequence; performing low percentile tracking filtering on the frequency band time slice power block sequences to generate a noise floor estimation sequence; aligning the kurtosis statistics sequence and the noise floor estimation sequence according to the separation frequency band definition sets and time slice indexes and outputting them to the Bayesian online change point detection module.

[0009] Optionally, the Bayesian online change point detection module specifically comprises: receiving a kurtosis statistics sequence and a noise floor estimation sequence, and generating a frequency band time slice observation sequence by aligning the separated frequency band definition set with the time slice index; establishing a run length state set for each frequency band identifier and configuring a state parameter cache and a sufficient statistics cache; generating a burst flag sequence based on the kurtosis statistics sequence, generating a drift flag sequence based on the noise floor estimation sequence, and selecting a hazard rate sequence based on the burst flag sequence and the drift flag sequence; driving the run length state set to perform online posterior updates on the frequency band time slice observation sequence and synchronously updating the state parameter cache based on the hazard rate sequence, and outputting a run length posterior distribution sequence; generating a candidate change point location set based on the run length posterior distribution sequence; performing cross-frequency band consistency fusion on the candidate change point location set and outputting a change point location sequence and a change point confidence sequence; extracting the maximum posterior run length based on the run length posterior distribution sequence and combining it with the change point confidence sequence to generate adaptive time window parameters.

[0010] Optionally, the updated frequency separation spectrum kurtosis analysis module specifically comprises: receiving a change point location sequence, a change point confidence sequence, and adaptive time window parameters, and generating a feedback event sequence by aligning the frequency band identifier with the time slice index; performing confidence segmentation filtering on the feedback event sequence and generating a locked event set; performing boundary relocation on the separation frequency band definition set based on the locked event set and generating an updated separation frequency band definition set; performing window length mapping on the time slice division rules based on the adaptive time window parameters and generating an updated time slice division rule; performing stability constraints on the updated separation frequency band definition set and the updated time slice division rule and outputting a constrained updated parameter set; writing the constrained updated parameter set into the frequency separation spectrum kurtosis analysis module and triggering the frequency separation spectrum kurtosis analysis module to re-execute frequency band extraction and time slice division according to the constrained updated parameter set, generating an updated kurtosis statistics sequence and an updated noise floor estimation sequence.

[0011] Optionally, the interference discrimination module specifically comprises: receiving a kurtosis statistics sequence, a noise floor estimation sequence, a change point location sequence, and a change point confidence sequence, and aligning them according to the frequency band identifier and time slice index to generate a discrimination record sequence; extracting kurtosis change indicators, noise floor change indicators, change point trigger indicators, and confidence weights from the discrimination record sequence and combining them to form an interference evidence vector sequence; performing time slice segmentation on the interference evidence vector sequence based on the change point location sequence and generating a candidate time period set; performing cross-frequency band association on each frequency band identifier within the candidate time period set to generate a candidate frequency band range set; establishing an interference discrimination template set for the candidate time period set and the candidate frequency band range set; reading the corresponding interference evidence vector sequence for each candidate time period and candidate frequency band range and performing template trigger condition matching with the interference discrimination template set to generate an interference segment list containing interference type identifiers and interference segment confidence; and performing merging of adjacent time periods of the same type and adjacent frequency band ranges on the interference segment list to generate an interference event list.

[0012] Optionally, the result output module specifically includes: receiving a list of interference events and parsing it to obtain an event record sequence; reading an alarm mapping table based on the event record sequence and generating an alarm record sequence; generating an alarm fingerprint for the alarm record sequence and matching it with an alarm receipt table to generate a sending flag sequence, and binding the sending flag sequence with the alarm record sequence to form a pending alarm sequence; generating a troubleshooting and location record sequence based on the event record sequence; reading data segments from the kurtosis statistics sequence cache, noise floor estimation sequence cache, and variable point location sequence cache based on the evidence index identifier and encapsulating them to generate an evidence packet sequence; associating the pending alarm sequence, troubleshooting and location record sequence, and evidence packet sequence according to the event identifier and encapsulating them to generate an output message sequence, and writing the output message sequence into an output queue; reading the output message sequence from the output queue and sending it to the network management interface and the local log interface, generating a sending receipt sequence, and writing the sending receipt sequence into an alarm receipt table.

[0013] The beneficial effects of the present invention are: (1) The present invention constructs a frame power spectrum sequence from the power spectrum sequence through the frequency separation spectrum kurtosis analysis module and calculates the frequency change indication sequence along the frequency dimension, thereby generating a separation frequency band definition set and outputting the kurtosis statistics sequence and noise floor estimation sequence according to the time slice, realizing the structured extraction of the superposition features of the same frequency and adjacent frequency, effectively improving the separability and consistency of the interference feature characterization, and enhancing the detection and discrimination ability in complex frequency domain superposition scenarios.

[0014] (2) This invention receives the kurtosis statistics sequence and the noise floor estimation sequence through the Bayesian online change point detection module and generates the frequency band time slice observation sequence. It constructs the run length state set and performs online posterior update based on the hazard rate sequence. It outputs the change point location sequence, change point confidence sequence and adaptive time window parameters, realizing online segmented identification of sudden intermittent and noise floor rise. It significantly improves the timeliness and stability of change point location and shows better adaptability in network real-time monitoring and maintenance troubleshooting scenarios.

[0015] (3) In terms of closed-loop linkage and engineering output, the present invention receives the variable point position sequence, variable point confidence sequence, and adaptive time window parameters through the parameter linkage module, updates the frequency separation spectrum kurtosis analysis module, and recalculates the kurtosis statistics sequence and noise floor estimation sequence. At the same time, the interference discrimination module generates an interference event list based on the interference evidence vector sequence and template triggering conditions, and the result output module outputs alarm information and troubleshooting location data. This effectively solves the problems of frequency band and time slice being difficult to update adaptively and alarm and location data being difficult to keep in a closed loop in the existing technology. It breaks through the bottleneck of false alarms and missed alarms and repeated parameter tuning caused by fixed thresholds and fixed windows, and realizes the improvement of automatic detection, discrimination and location linkage capabilities for network operation and maintenance. Attached Figure Description

[0016] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used together with the embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings: Figure 1 is a structural diagram of a network signal interference detection system based on spectrum analysis proposed in this invention; Figure 2 is a processing flowchart of a network signal interference detection system based on spectrum analysis proposed in this invention. Detailed Implementation

[0017] The present invention will now be described in further detail with reference to the accompanying drawings. These drawings are simplified schematic diagrams, illustrating only the basic structure of the invention, and therefore only show the components relevant to the invention.

[0018] Referring to Figures 1 and 2, a network signal interference detection system based on spectrum analysis includes: a network signal acquisition module, which receives the signal of the network communication link under test and outputs a digitized time-domain sampling sequence; a spectrum preprocessing module, which performs window function segmentation, DC component removal, and amplitude normalization on the digitized time-domain sampling sequence and outputs a spectrum calculation input sequence; a spectrum calculation module, which performs discrete Fourier transform on the spectrum calculation input sequence to generate a power spectrum sequence; and a frequency separation spectrum kurtosis analysis module, which performs frequency separation processing on the power spectrum sequence to obtain multiple separated frequency band power spectrum sets, and forms a kurtosis statistics sequence and a noise floor estimation sequence according to time slices. The Bayesian online change point detection module performs online posterior updates on the kurtosis statistics sequence and the noise floor estimation sequence, outputting the change point location sequence, change point confidence sequence, and adaptive time window parameters; the parameter linkage module receives the change point location sequence, change point confidence sequence, and adaptive time window parameters, updates the frequency separation spectrum kurtosis analysis module, and recalculates the kurtosis statistics sequence and the noise floor estimation sequence; the interference discrimination module receives the kurtosis statistics sequence, the noise floor estimation sequence, the change point location sequence, and the change point confidence sequence, and generates a list of interference events; the result output module receives the list of interference events and outputs alarm information and troubleshooting location data.

[0019] In this embodiment, the network signal acquisition module specifically comprises: receiving the network communication link signal under test through an RF coupling interface and allocating it to the first sampling channel and the second sampling channel; performing variable bandwidth analog pre-selection filtering in the first sampling channel and determining the center frequency and passband bandwidth according to a preset frequency band table, and outputting a first filtered RF signal; the variable bandwidth analog pre-selection filtering specifically comprises: setting a tunable RF bandpass filter at the input of the first sampling channel and receiving the network communication link signal under test; the filter passband center frequency is continuously adjustable by the control word output by the local oscillator phase-locked loop synthesizer driving the varactor diode tuning network; the filter passband bandwidth is adjusted by switching the filter coupling coefficient through a switched capacitor array and adjusting the transconductance amplification accordingly. The unit bias current is switchable in multiple ranges; the filter adopts a third- to fifth-order cascaded resonant structure and a temperature-compensated resistor network is configured in each resonant branch to stabilize the passband amplitude-frequency response; an adjustable attenuator is inserted in the filter passband and the attenuation is adjusted by the output control voltage of the automatic gain control loop to suppress strong signal compression; outside the filter passband, dual notch branches are used to achieve fixed-depth suppression on the upper and lower sides of the passband to reduce adjacent-channel energy leakage; a reference noise RF signal is generated in the second sampling channel through an isolation terminal; the generation of the reference noise RF signal specifically involves: connecting a 50-ohm isolation terminal at the input of the second sampling channel and placing the isolation terminal in an electromagnetic shielding cavity to isolate external RF coupling; the isolation terminal is made of gold The device integrates a membrane resistor and a coaxial connector, employing a ring-shaped grounding spring between the connector housing and cavity to achieve low-resistance grounding. Thermal noise at both ends of the isolation terminal is amplified by a broadband low-noise amplifier and its amplitude range is limited by a bandpass limiting network to prevent overload of the subsequent demodulator. The output of the broadband low-noise amplifier is connected to an adjustable attenuator, and the attenuation is adjusted by a constant voltage control loop to stabilize the amplitude of the reference noise RF signal. The output of the adjustable attenuator is passed through an RF bandpass filter identical to the first sampling channel to achieve bandpass limitation in the same frequency band and output the reference noise RF signal. Parallel dual-gain amplification and automatic gain control are performed on the first filtered RF signal to obtain a first high-gain intermediate frequency signal and a first low-gain intermediate frequency signal. The first high-gain intermediate frequency signal and the first low-gain intermediate frequency signal are then compared with the first high-gain intermediate frequency signal and the first low-gain intermediate frequency signal. A low-gain intermediate frequency (IF) signal undergoes in-phase quadrature demodulation to obtain a first high-gain baseband sampling pair and a first low-gain baseband sampling pair. The dual-gain amplification and automatic gain control are specifically as follows: the first filtered RF signal is distributed to the high-gain amplification branch and the low-gain amplification branch via a power divider while maintaining the input phase of the two branches in sync; the high-gain amplification branch is connected in series with a low-noise amplifier, a variable-gain amplifier, and an IF converter to output a first high-gain IF signal; the low-gain amplification branch is connected in series with an attenuator, a variable-gain amplifier, and an IF converter to output a first low-gain IF signal; logarithmic detectors are respectively set at the IF output terminals of the high-gain amplification branch and the low-gain amplification branch to generate a high-gain power detection voltage and a low-gain power detection voltage.The high-gain power detection voltage and the target amplitude reference voltage are input into the error amplifier to generate a high-gain gain control voltage. The low-gain power detection voltage and the target amplitude reference voltage are input into the error amplifier to generate a low-gain gain control voltage. The high-gain gain control voltage is applied to the gain control terminal of the variable gain amplifier in the high-gain amplification branch, limiting the intermediate frequency output amplitude to fall within the linear operating region. The low-gain gain control voltage is applied to the gain control terminal of the variable gain amplifier in the low-gain amplification branch, limiting the intermediate frequency output amplitude to fall within the linear operating region. The high-gain amplification branch and the low-gain amplification branch... The circuits share the same local oscillator source and employ a phase noise suppression loop to lock the local oscillator frequency to reduce phase disturbances introduced during gain control. The in-phase quadrature demodulation specifically involves: inputting the first high-gain intermediate frequency signal into the in-phase input of the first mixer and inputting the local oscillator signal into the local oscillator input of the first mixer to output the first in-phase baseband signal; inputting the first high-gain intermediate frequency signal into the quadrature input of the second mixer and inputting the local oscillator signal into the local oscillator input of the second mixer after passing through a 90-degree phase-shifting network to output the first quadrature baseband signal; and passing the first in-phase baseband signal through a low-pass filter to suppress the double intermediate frequency component and outputting the first in-phase baseband analog signal. The process involves: 1) Suppressing the double intermediate frequency component of the first quadrature baseband signal using a low-pass filter and outputting the first quadrature baseband analog signal; 2) Inputting the first low-gain intermediate frequency signal into the in-phase input of the third mixer and inputting the local oscillator signal into the local oscillator input of the third mixer to output the second in-phase baseband signal; 3) Inputting the first low-gain intermediate frequency signal into the quadrature input of the fourth mixer and inputting the local oscillator signal into the local oscillator input of the fourth mixer after passing through a 90-degree phase-shifting network to output the second quadrature baseband signal; 4) Filtering the second in-phase baseband signal through a low-pass filter to output the second in-phase baseband analog signal and filtering the second quadrature baseband signal through a low-pass filter to output the second quadrature baseband signal. Analog signals; the first in-phase baseband analog signal and the first quadrature baseband analog signal are combined to form a first high-gain baseband sampling pair, and the second in-phase baseband analog signal and the second quadrature baseband analog signal are combined to form a first low-gain baseband sampling pair; in-phase and quadrature demodulation is performed on the reference noise RF signal to obtain a reference baseband sampling pair; unified clock sampling is performed on the first high-gain baseband sampling pair, the first low-gain baseband sampling pair, and the reference baseband sampling pair to obtain a synchronous sampling sequence; amplitude and phase imbalance calibration and DC bias calibration are performed on the synchronous sampling sequence, and cross-gain fusion is performed to generate a digital time-domain sampling sequence.

[0020] In this embodiment, the spectrum preprocessing module specifically includes: receiving a digitized time-domain sampling sequence and generating a frame index sequence and a segmented time-domain frame sequence according to a preset frame length and a preset frame shift; performing endpoint continuity detection on the segmented time-domain frame sequence and calculating the amplitude difference between the first and last sample values ​​and the first and last first-order difference amplitude difference for each segmented time-domain frame, and calculating a windowing coefficient sequence; the amplitude difference between the first and last sample values ​​specifically involves: extracting the first and last sample points for each segmented time-domain frame and calculating the amplitude difference between the first and last sample values; the first and last first-order difference amplitude difference specifically involves: calculating the first-order difference value of the first point and the first-order difference value of the last point for each segmented time-domain frame and calculating the first and last first-order difference amplitude difference; the calculation of the windowing coefficient sequence specifically involves: obtaining the frame length for each segmented time-domain frame and establishing a sequence from... The sample number sequence is from 0 to frame length minus 1; a continuity weight is constructed based on the amplitude difference between the first and last sample values ​​and the amplitude difference between the first and last first-order difference, and the continuity weight is limited to the range of 0 to 1; the length of the window function transition zone is determined according to the continuity weight, and the length of the window function transition zone is limited to not less than 2 and not more than half the frame length; a cosine roll-off weight is generated within the window function transition zone, and a unit weight is generated outside the window function transition zone; the cosine roll-off weight and the unit weight are concatenated according to the sample number sequence to form the window function coefficient vector of each segmented time-domain frame; the window function coefficient vectors of each segmented time-domain frame are arranged according to the frame index sequence to form a windowing coefficient sequence; the windowing coefficient sequence is applied to the segmented time-domain frame sequence to generate a windowed time-domain frame sequence and the corresponding complex mean is calculated, and the complex mean is used to adjust the windowing time-domain frame sequence. The DC component cancellation of the windowed time-domain frame sequence generates a de-DC time-domain frame sequence. The calculation of the corresponding complex mean specifically involves: representing each windowed time-domain frame in the windowed time-domain frame sequence as a complex sampling point sequence composed of an in-phase component sequence and a quadrature component sequence; accumulating the complex sampling point sequence of each windowed time-domain frame to obtain a complex accumulated value and recording the number of valid sampling points; dividing the complex accumulated value by the number of valid sampling points to obtain the complex mean of the windowed time-domain frame. The DC component cancellation specifically involves: reading the complex mean of each windowed time-domain frame and obtaining the in-phase mean and quadrature mean; subtracting the in-phase mean from the in-phase component sequence of the windowed time-domain frame point by point to generate a de-DC in-phase component sequence; subtracting the quadrature mean from the quadrature component sequence of the windowed time-domain frame point by point to generate a de-DC quadrature component sequence; and then... The DC in-phase component sequence and the deDC quadrature component sequence are reassembled into a deDC complex sampling point sequence according to the sampling point number; the deDC complex mean is calculated for the deDC complex sampling point sequence and used for consistency verification; when the consistency verification satisfies the 0-mean condition, the deDC time-domain frame sequence is output; the amplitude square sequence and robust scaling estimate are calculated for the deDC time-domain frame sequence, and the amplitude normalized sequence is generated using the robust scaling estimate; the robust scaling estimate is specifically: the median of the amplitude square sequence of each deDC time-domain frame in the deDC time-domain frame sequence is calculated to obtain the amplitude square median; the absolute deviation from the amplitude square median is calculated for each sampling point in the amplitude square sequence to obtain the absolute deviation sequence; the median of the absolute deviation sequence is calculated to obtain the median absolute deviation value;The robust scale estimate is obtained by scaling the median absolute deviation value using a Gaussian noise uniformity scaling factor. The generation of the amplitude normalization sequence specifically involves: reading the robust scale estimate corresponding to each de-DC time-domain frame and determining if the robust scale estimate is 0; when the robust scale estimate is 0, replacing it with the root mean square square of the amplitude square sequence to form an effective scale value; performing a square root operation on the effective scale value to obtain the amplitude scale value; taking the reciprocal of the amplitude scale value to obtain the amplitude normalization coefficient; arranging the amplitude normalization coefficients according to the frame index sequence to form an amplitude normalization sequence; performing amplitude normalization on the de-DC time-domain frame sequence based on the amplitude normalization sequence to generate a normalized time-domain frame sequence, and performing sample compression to generate a compressed time-domain frame sequence. The sample compression specifically involves: calculating the amplitude squares of the in-phase and quadrature components point-by-point in each normalized time-domain frame within the normalized time-domain frame sequence and forming an amplitude square sequence; sorting the amplitude square sequences in ascending order and obtaining the mapping relationship between the sorted amplitude square sequences and the sample point indices; determining the upper truncation position based on the number of samples in the sorted amplitude square sequence, where the upper truncation position corresponds to the indices obtained by multiplying the number of samples by 0.999 and rounding up; reading the amplitude square value at the upper truncation position and taking its square root to obtain the compression amplitude threshold; calculating the amplitude for each sampling point within the normalized time-domain frame and comparing it with the compression amplitude threshold; calculating a scaling factor when the amplitude is greater than the compression amplitude threshold, where the scaling factor is the ratio of the compression amplitude threshold to the amplitude, and then separating the in-phase and quadrature components respectively. Multiplying by a scaling factor yields the compressed sampling points; keeping the sampling points unchanged when the amplitude is not greater than the compression amplitude threshold yields the uncompressed sampling points; reconstructing the compressed and uncompressed sampling points according to the sample point sequence number to form a compressed time-domain frame sequence; extracting the overlapping region from adjacent compressed time-domain frame sequences and calculating the cross-correlation function, obtaining the resampling scaling sequence based on the maximum correlation hysteresis of the cross-correlation function and parabolic interpolation, and performing fractional delay filtering resampling on the compressed time-domain frame sequence to generate the spectrum calculation input frame sequence; the calculation of the cross-correlation function specifically involves: extracting the in-phase component overlap region sequence and the orthogonal component overlap region sequence from two adjacent compressed time-domain frame sequences respectively, and forming the previous frame complex overlap region sequence and the next frame complex overlap region sequence; generating a common overlap region sequence by performing conjugate operation on the previous frame complex overlap region sequence. The preceding frame's multi-overlapping region sequence is used; an integer hysteresis search range is set and a hysteresis index sequence from negative hysteresis to positive hysteresis is generated; for each hysteresis index, the following frame's multi-overlapping region sequence is cyclically shifted according to the hysteresis index and multiplied point-by-point with the conjugate preceding frame's multi-overlapping region sequence to obtain a product sequence; the product sequence is accumulated according to the sample point index to obtain the cross-correlation value of the corresponding hysteresis index and form a cross-correlation function; the fractional delay filter resampling specifically involves: reading the resampling ratio sequence and converting the resampling ratio into a fractional delay amount, where the fractional delay amount is the fractional sampling offset corresponding to the resampling ratio; generating a finite-length fractional delay filter coefficient vector based on the fractional delay amount, where the filter coefficient vector is obtained by sampling at the fractional delay amount using a windowed sine interpolation kernel and then normalized to unity gain.The in-phase component sequence of the compressed time-domain frame sequence is convolved with the filter coefficient vector according to the sample index to obtain the in-phase resampled sequence; the orthogonal component sequence of the compressed time-domain frame sequence is convolved with the filter coefficient vector according to the sample index to obtain the orthogonal resampled sequence; the in-phase resampled sequence and the orthogonal resampled sequence are recombined according to the sample index to form a resampled complex sampling point sequence and form a resampled time-domain frame sequence; the overlapping sample points of the resampled time-domain frame sequence are replaced according to the corresponding positions of the frame index sequence to achieve inter-frame time alignment and output the spectrum calculation input frame sequence; the spectrum calculation input frame sequence is concatenated according to the frame index sequence to generate the spectrum calculation input sequence.

[0021] In this embodiment, performing the Discrete Fourier Transform specifically involves: receiving the spectrum calculation input sequence and writing it into a circular buffer, forming a time-domain block sequence according to the number of transform points; forming the time-domain block sequence specifically involves: writing the spectrum calculation input sequence into the circular buffer in the form of a complex sampling point sequence and recording the write pointer position; setting the number of transform points and configuring the number of transform points as the data retrieval window length of the circular buffer; reading consecutive complex sampling points from the circular buffer to form a time-domain block with the number of transform points as the length during the continuous increment of the write pointer; assigning a block index to each time-domain block and establishing a mapping relationship between the block index and the starting write pointer position of the time-domain block; concatenating the tail data and the head data in a circular manner when the write pointer crosses the tail of the circular buffer to obtain a complete time-domain block; and sorting the data according to the block index order. The generated time-domain block set is output to form a time-domain block sequence; a rotation factor lookup table is selected based on the sampling rate identifier; the rotation factor lookup table specifically stores the complex rotation factor coefficient pairs required for the butterfly operation of the Discrete Fourier Transform, and the complex rotation factor coefficient pairs include in-phase coefficients and orthogonal coefficients; the rotation factor lookup table establishes a table space according to the number of transform points and establishes a hierarchical index according to the butterfly operation hierarchy, each hierarchical index corresponds to a group length and a set of rotation angle sequences; the rotation angle sequence is generated by incrementing the rotation angle steps within one rotation angle range according to the equal interval angles corresponding to the group length, and forms a rotation sequence number, with a one-to-one correspondence between the rotation sequence number and the rotation angle; for each rotation angle, the cosine value and the negative and positive sine values ​​are calculated and written into the in-phase coefficient storage field and the positive sine value, respectively. The intersection coefficient storage field forms complex twitch factor coefficient pairs; the complex twitch factor coefficient pairs are stored using fixed-point quantization and saturated rounding rules. The fixed-point bit width and decimal places are mapped by the sampling rate identifier and written into the twitch factor lookup table metadata area; the twitch factor lookup table address is constructed by concatenating the transform point index field, the hierarchical index field, and the twitch number field and supports parallel read ports. The parallel butterfly operation array generates addresses based on the butterfly index at each pipeline stage and synchronously reads complex twitch factor coefficient pairs for complex multiplication and addition; bit reversal is performed on the time-domain block sequence and the reversal result is allocated to the parallel butterfly operation array; the parallel butterfly operation array is specifically composed of multiple butterfly processing units and connected to the input reversal buffer. The system is connected to the rotation factor read port and inter-stage register links. Each butterfly processing unit includes a complex adder, a complex subtractor, a complex multiplier, and a coefficient multiply-add register, and supports parallel operation of in-phase and quadrature components. The input rearrangement buffer provides a dual-port readout of the bit-reversed rearrangement result and provides paired complex input samples to the butterfly processing unit. The rotation factor read port outputs pairs of complex rotation factor coefficients in parallel from the rotation factor lookup table and sends them to the complex multiplier of the butterfly processing unit. The butterfly processing unit performs summation and difference calculations on the paired complex input samples and performs complex multiplication with the complex rotation factor coefficient pairs to obtain the rotation difference term. The butterfly processing unit outputs the summation result and the rotation difference term and writes them to the inter-stage register links to realize multi-stage pipeline cascading.The parallel butterfly arithmetic array adopts a fixed interconnection topology and generates input read addresses and output write addresses through butterfly indexes at each stage to complete data cross-permutation. Within the parallel butterfly arithmetic array, butterfly operations are executed in a multi-stage pipeline, and a twiddle factor lookup table is called at each stage to complete complex multiplication and addition, resulting in a frequency domain complex spectrum sequence. Block floating-point scaling is performed on the frequency domain complex spectrum sequence to obtain a calibrated frequency domain complex spectrum sequence. Specifically, the block floating-point scaling involves: reading the in-phase and quadrature components of the frequency domain complex spectrum sequence frequency by frequency point and calculating the corresponding absolute values; performing a maximum value search on the absolute values ​​within the frequency range corresponding to the power spectrum of a single frame to obtain the block peak value; and then... The block peak value is calculated to determine the position of the most significant bit in binary and obtain the scaling exponent. A uniform shift is generated based on the scaling exponent, and a uniform arithmetic shift is performed on the in-phase and quadrature components of the frequency domain complex spectrum sequence to obtain the calibrated in-phase and quadrature components. The result after the uniform arithmetic shift is subjected to saturation clipping and rounding to meet the target fixed-point bit width. The scaling exponent is written into the block exponent flag and associated with the calibrated frequency domain complex spectrum sequence. The calibrated frequency domain complex spectrum sequence is subjected to complex conjugate product to obtain power spectrum unit values, which are then arranged by frequency index to form a single-frame power spectrum. The single-frame power spectra are concatenated according to the time-domain block sequence order to output the power spectrum sequence.

[0022] In this embodiment, the frequency separation spectral kurtosis analysis module specifically comprises: receiving a power spectrum sequence and constructing a frame power spectrum sequence in chronological order; the construction of the frame power spectrum sequence specifically comprises: receiving the power spectrum sequence and identifying the frame boundary marker and frame index of each single frame power spectrum; arranging the power spectrum unit values ​​corresponding to the same frame index in ascending order of frequency index to obtain a single frame power spectrum vector; performing a length consistency check on the single frame power spectrum vector and filling the missing frequency index with zero power unit values ​​to maintain the continuity of the frequency index; sequentially writing the single frame power spectrum vectors corresponding to consecutive frame indices into the frame buffer in chronological order to form a two-dimensional frame power spectrum matrix; expanding the two-dimensional frame power spectrum matrix in frame index order and outputting it to form a frame power spectrum sequence; and performing frequency-wise analysis on the frame power spectrum sequence. The frequency mutation indicator sequence is calculated using the rate dimension. Specifically, this calculation involves: calculating the power ratio sequence of adjacent frequency points for each single-frame power spectrum vector in the frame power spectrum sequence according to the frequency index, and taking the logarithm of the power ratio sequence to obtain a logarithmic power ratio sequence; calculating the local mean and local absolute deviation of the logarithmic power ratio sequence using a fixed-length sliding window to obtain a frequency-direction robust baseline sequence; subtracting the logarithmic power ratio sequence from the frequency-direction robust baseline sequence to obtain a frequency-direction mutation amplitude sequence; performing sign consistency determination on the frequency-direction mutation amplitude sequence according to the frequency index to obtain a mutation polarity sequence, and combining the mutation polarity sequence with the mutation amplitude sequence to obtain a mutation score sequence; performing threshold comparison on the mutation score sequence and outputting a binary frequency mutation indicator. The binary frequency mutation indicator sequence is used to mark positions where the mutation score is greater than a threshold as 1 and other positions as 0. The binary frequency mutation indicator sequences corresponding to the power spectrum vectors of each single frame are arranged in frame index order and output to form a frequency mutation indicator sequence. Based on the frequency mutation indicator sequence, continuous segment clustering is performed on the frequency index to generate a frequency band boundary sequence. Multiple non-overlapping separate frequency band definition sets are constructed using the frequency band boundary sequence. Specifically, constructing multiple non-overlapping separate frequency band definition sets involves: reading the frequency band boundary sequence and sorting it by frequency index from smallest to largest to obtain an ordered boundary list; performing adjacent boundary deduplication on the ordered boundary list and deleting duplicate boundaries with the same frequency index to form a unique boundary list; and then... The minimum and maximum frequency indices are added as frequency band endpoint boundaries to form a complete endpoint boundary list; the complete endpoint boundary list is used to generate a frequency band start frequency index and a frequency band end frequency index by pairing adjacent boundaries to obtain a frequency band interval list; the frequency band interval list is subjected to interval validity checks and invalid intervals with start frequency indices greater than end frequency indices are removed, as well as invalid intervals with interval lengths of 0 are removed; endpoint connection checks are performed on adjacent frequency band intervals, and when adjacent intervals overlap, the overlapping part is divided according to the midpoint frequency index to obtain a non-overlapping frequency band interval list; the non-overlapping frequency band interval list is written into the separate frequency band definition set, and a frequency band identifier and frequency index range are assigned to each frequency band interval to form multiple non-overlapping separate frequency band definition sets;According to the defined set of separated frequency bands, frequency band extraction is performed on the frame power spectrum sequence to form multiple separated frequency band power spectrum sets, and the frequency band time slice power block sequence is obtained by dividing it according to time slices; the formation of multiple separated frequency band power spectrum sets specifically involves: reading the defined set of separated frequency bands and obtaining the frequency band start frequency index and frequency band end frequency index for each separated frequency band; extracting the power spectrum unit values ​​within the range of the frequency band start frequency index to the frequency band end frequency index for each single frame power spectrum vector in the frame power spectrum sequence to form the corresponding separated frequency band frame power sub-vector; writing the separated frequency band frame power sub-vectors of each separated frequency band into the frequency band buffer according to the frame index to form a separated frequency band frame sequence; arranging the separated frequency band frame sequences of each separated frequency band in frame index order to form a sequence of separated frequency bands. Off-band power spectrum set; calculate the mean sequence, variance sequence, and fourth-order central moment sequence for the power block sequence of the frequency band time slice, and generate a kurtosis statistics sequence based on the mean sequence, variance sequence, and fourth-order central moment sequence; the fourth-order central moment sequence is specifically generated as follows: read the power block sequence of the frequency band time slice and determine the number of samples and the range of sample number for each power block; read the mean sequence of the corresponding power block and write the mean to the centering processing register; subtract the mean from each power sample in the power block to obtain the centering deviation value; perform a square operation on the centering deviation value to obtain the quadratic deviation value and perform a square operation on the quadratic deviation value to obtain the quartic deviation value; accumulate the quartic deviation value in the power block according to the sample number to obtain the quartic deviation accumulation. The fourth-order central moment value is obtained by dividing the cumulative value of the four deviations by the number of samples; the fourth-order central moment values ​​corresponding to each time slice are arranged according to the time slice index to form a fourth-order central moment sequence; the generation of kurtosis statistics sequence specifically involves: reading the mean sequence, variance sequence, and fourth-order central moment sequence corresponding to the same frequency band identifier and the same time slice index and performing index consistency verification; determining the variance sequence to have a 0 value and marking the corresponding kurtosis statistics as 0 and outputting an anomaly flag when the variance is 0; when the variance is not 0, squaring the variance to obtain the second-order variance and using it as the normalized denominator; using the fourth-order central moment as the normalized numerator and performing a division operation with the second-order variance to obtain the normalized fourth-order ratio; performing bias compensation on the normalized fourth-order ratio and generating kurtosis statistics; and then... Kujic statistics are arranged by time slice index to form a kurtosis statistics sequence; low percentile tracking filtering is performed on the frequency band time slice power block sequence to generate a noise floor estimation sequence; the low percentile tracking filtering specifically involves: reading power samples for each frequency band time slice power block and sorting them in ascending order to obtain a sorted power sequence; determining the low percentile position based on the number of samples in the sorted power sequence and reading the low percentile power value as a candidate noise floor value; calculating the difference between the candidate noise floor value and the noise floor estimation value of the previous time slice and generating an ascending update coefficient and a descending update coefficient; when the candidate noise floor value is greater than the noise floor estimation value of the previous time slice, the ascending update coefficient is used to exponentially smooth the noise floor estimation value of the previous time slice to obtain the noise floor estimation value of the current time slice;When the candidate noise floor value is not greater than the noise floor estimate of the previous time slice, a descent update coefficient is used to exponentially smooth the noise floor estimate of the previous time slice to obtain the noise floor estimate of the current time slice. The noise floor estimates of the current time slice are arranged by time slice index to form a noise floor estimate sequence. The ascending update coefficient and the descent update coefficient are used to control the update rate of the noise floor estimate sequence under different directions of change. The ascending update coefficient is applied when the candidate noise floor value is greater than the noise floor estimate of the previous time slice. The ascending update coefficient ranges from 0 to 1. The larger the value, the more the noise floor estimate follows the candidate noise floor value upwards. The faster the speed, the better; the descent update coefficient is applied when the candidate noise floor value is not greater than the noise floor estimate value of the previous time slice. The descent update coefficient ranges from 0 to 1, and the larger the value, the faster the noise floor estimate value follows the candidate noise floor value downwards; the ascending update coefficient and the descending update coefficient are used to suppress the noise floor estimate rise caused by short-term sudden interference and accelerate the noise floor estimate fall back after the interference disappears, so that the noise floor estimate sequence is more in line with the slow change characteristics of the noise floor; the kurtosis statistics sequence and the noise floor estimate sequence are aligned with the separation frequency band definition set and the time slice index and output to the Bayesian online change point detection module.

[0023] In this embodiment, the Bayesian online change point detection module specifically comprises: receiving a kurtosis statistics sequence and a noise floor estimation sequence, and generating a frequency band time slice observation sequence by aligning the separated frequency band definition set with the time slice index; the generation of the frequency band time slice observation sequence specifically comprises: receiving the kurtosis statistics sequence and the noise floor estimation sequence, and reading the frequency band identifier list of the separated frequency band definition set; parsing the kurtosis statistics sequence according to the frequency band identifier and the time slice index to obtain kurtosis observation entries, and parsing the noise floor estimation sequence according to the frequency band identifier and the time slice index to obtain noise floor observation entries; performing frequency band identifier consistency verification and time slice index consistency verification on the kurtosis observation entries and the noise floor observation entries; and verifying the same frequency band identifier and the same time slice index that pass the verification. The slice index entries are merged to generate observation records, which include a frequency band identifier field, a time slice index field, a kurtosis statistic value field, and a noise floor estimate field. The observation records are then arranged in ascending order by frequency band identifier and ascending order by time slice index to form a frequency band time slice observation sequence. A run-length state set is established for each frequency band identifier, and state parameter caches and sufficient statistics caches are configured. Specifically, establishing the run-length state set involves: allocating a run-length state table for each frequency band identifier and creating state entries from 0 to the maximum run-length in the run-length state table; configuring a state probability field for each state entry and initializing the probability of a run-length of 0 to 1 and the probabilities of other states to 0; configuring... The system caches state parameters and writes the kurtosis statistics state parameters and noise floor estimation state parameters into the corresponding fields of the state parameter cache; it configures a sufficient statistics cache for each state entry and writes the kurtosis statistics sufficient statistics and noise floor estimation sufficient statistics into the corresponding fields of the sufficient statistics cache; it establishes an index mapping between the run length state table and the frequency band identifier and writes the index mapping into the frequency band index table; it generates a burst flag sequence based on the kurtosis statistics sequence, generates a drift flag sequence based on the noise floor estimation sequence, and selects a hazard rate sequence based on the burst flag sequence and the drift flag sequence; the burst flag sequence specifically involves: reading the kurtosis statistics sequence and obtaining the kurtosis statistics value sequence according to the frequency band identifier and time slice index; and processing the kurtosis statistics value sequence... A short-window sequence is constructed, and the change amplitude sequence of kurtosis statistics is calculated within the short window; a robust baseline amplitude is calculated from the change amplitude sequence, and an amplitude deviation sequence is obtained; a threshold comparison is performed on the amplitude deviation sequence, and a burst flag of 1 is generated at positions where the amplitude deviation is greater than the threshold, and a burst flag of 0 is generated at other positions; the burst flags are arranged according to the frequency band identifier and time slice index to form a burst flag sequence; the drift flag sequence is specifically: the noise floor estimation sequence is read, and a noise floor estimation value sequence is obtained according to the frequency band identifier and time slice index; a long-window sequence is constructed from the noise floor estimation value sequence, and the cumulative change amount sequence of the noise floor estimation value is calculated within the long window; the robust baseline drift amount is calculated from the cumulative change amount sequence, and a drift deviation sequence is obtained;A threshold comparison is performed on the drift deviation sequence, and a drift flag of 1 is generated at positions where the drift deviation is greater than the threshold, while a drift flag of 0 is generated at other positions. The drift flags are arranged according to the frequency band identifier and time slice index to form a drift flag sequence. The step of selecting a hazard rate sequence based on the burst flag sequence and drift flag sequence is as follows: the burst flag value and drift flag value are read from the frequency band time slice observation sequence according to the frequency band identifier and time slice index; when the burst flag value is 1 and the drift flag value is 0, the burst hazard rate mode is selected and the burst hazard rate sequence is output; when the burst flag value is 0 and the drift flag value is 1, the drift hazard rate mode is selected and the drift hazard rate sequence is output; when the burst flag value is 1 and the drift flag value is 1, the composite hazard rate mode is selected and the burst hazard rate sequence is output. The column and drift hazard rate sequences are fused by mean according to time slice index to obtain a composite hazard rate sequence; when the burst flag value is 0 and the drift flag value is 0, a stable hazard rate mode is selected and a stable hazard rate sequence is output; the hazard rate values ​​corresponding to the selected mode in the burst hazard rate sequence, drift hazard rate sequence, composite hazard rate sequence, and stable hazard rate sequence are written into a hazard rate sequence consistent with the frequency band identifier and time slice index; based on the hazard rate sequence, the run length state set is driven to perform online posterior update of the frequency band time slice observation sequence and synchronously update the state parameter cache, outputting the run length posterior distribution sequence; for each frequency band identifier, the hazard rate value corresponding to the current time slice index is read and the state probability field of the previous time slice index is read from the run length state set; based on... The hazard rate value is used to perform a state transition on the state probability field of the previous time slice index. The state transition includes a continuation transition to a state with a run length incremented by one and a reset transition to a state with a run length of 0, generating a post-transition state probability field. The kurtosis statistic and noise floor estimate fields are read from the frequency band time slice observation sequence, and the observation consistency weight is calculated for each run length state entry. The observation consistency weight is multiplied by the post-transition state probability field item by item according to the run length state entry to obtain the unnormalized posterior probability field. The unnormalized posterior probability field is normalized to obtain the run length posterior probability field and written back to the run length state set. Based on the run length posterior probability field, the corresponding state parameter cache and sufficient statistics cache are updated, and the next time slice is established. The index's state parameter cache version; the run-length posterior probability field is arranged in run-length order to form a run-length posterior distribution sequence; a candidate change point location set is generated based on the run-length posterior distribution sequence; the generation of the candidate change point location set specifically involves: reading the run-length posterior distribution sequence and locating the corresponding posterior probability field according to the frequency band identifier and time slice index; extracting the posterior probability with a run-length of 0 from the posterior probability field as the change point posterior probability; performing peak detection on the change point posterior probability and determining the peak time slice index; applying a minimum interval constraint to the peak time slice index and removing peak time slice indices with an interval smaller than the minimum interval; and forming candidate change point entries with the retained peak time slice indices and corresponding frequency band identifiers and writing them into the candidate change point location set.Perform cross-band consistency fusion on the candidate variable point location set and output the variable point location sequence and variable point confidence sequence; the cross-band consistency fusion specifically includes: reading the candidate variable point location set and grouping the candidate variable point entries according to the time slice index to obtain time slice candidate groups; counting the number of participating frequency band identifiers and calculating the frequency band coverage for each time slice candidate group; reading the variable point posterior probability of the corresponding frequency band identifier for each time slice candidate group and calculating the mean of the within-group posterior probability; weighting the frequency band coverage and the mean of the within-group posterior probability to obtain the fusion confidence; marking the time slice index with a fusion confidence greater than a threshold as the fusion variable point time slice and writing the fusion confidence into the variable point confidence sequence; associating the fusion variable point time slice with the set of participating frequency band identifiers and writing it into the variable point location sequence; according to the play The maximum a posteriori (MAP) run length is extracted from the run length posterior distribution sequence and combined with the change-point confidence sequence to generate adaptive time window parameters. The MAP run length posterior distribution sequence is read, and the MAP run length is extracted according to the frequency band identifier and time slice index. The MAP run length is converted into a time window length to obtain a basic time window length sequence. The change-point confidence sequence is read, and a correspondence is established between the time slice index and the basic time window length sequence. Segmentation mapping is performed on the change-point confidence sequence to obtain a confidence scaling factor sequence, with the confidence scaling factor taking values ​​between 0 and 1. The basic time window length sequence is multiplied by the confidence scaling factor sequence to obtain the adaptive time window length sequence. Minimum and maximum window length constraints are applied to the adaptive time window length sequence, and the adaptive time window parameters are output.

[0024] In this embodiment, the updated frequency separation spectral kurtosis analysis module specifically includes: receiving the change point location sequence, the change point confidence sequence, and adaptive time window parameters, and generating a feedback event sequence by aligning the frequency band identifier with the time slice index; performing confidence segmentation filtering on the feedback event sequence and generating a locked event set; the confidence segmentation filtering specifically includes: reading the change point confidence sequence and establishing a correspondence between it and the feedback event sequence by aligning the frequency band identifier with the time slice index; dividing the change point confidence into high confidence intervals, medium confidence intervals, and low confidence intervals according to the numerical range and assigning an interval identifier to each interval; reading the corresponding change point confidence for each feedback event and, based on the interval... The system identifies feedback events as high-confidence, medium-confidence, and low-confidence events. High-confidence events are retained and added to the locked event set; medium-confidence events undergo neighborhood consistency verification before being added to the locked event set; and low-confidence events are removed. Based on the locked event set, the system performs boundary relocation on the separated frequency band definition set and generates an updated separated frequency band definition set. Boundary relocation includes converging the frequency band boundaries within the locked frequency band range towards the frequency change indicator position and pushing the frequency band boundaries outside the locked frequency band range towards the stable frequency position. Based on the adaptive time window parameters, the system performs window length mapping on the time slice division rules and generates an updated time slice division rule. Further steps are taken to relocate the separated frequency band definition set. The new separated frequency band definition set and the updated time slice division rules are subjected to stability constraints, and the parameter set is updated after the constraints are output. Specifically, the stability constraints are as follows: the updated separated frequency band definition set is read, and the frequency band start frequency index and frequency band end frequency index of each frequency band are extracted to form a frequency band interval list; the frequency band interval list is sorted in ascending order by the frequency band start frequency index, and the boundary spacing between adjacent frequency band intervals is calculated; adjacent frequency band intervals with boundary spacing less than the minimum boundary spacing threshold are subjected to boundary backtracking until the boundary spacing reaches the minimum boundary spacing threshold; the frequency band interval list is subjected to non-overlap verification, and when overlapping intervals are detected, the overlapping portion is segmented according to the midpoint frequency index of the overlapping interval to generate... A list of non-overlapping frequency band intervals is generated; the update time slice division rules are read and the time slice length and time slice overlap length are extracted to form a window length parameter pair; the rate of change of the window length parameter pair and the previous window length parameter pair is calculated, and when the rate of change is greater than the maximum rate of change threshold, the window length parameter pair is restricted to the change amplitude corresponding to the maximum rate of change threshold; the list of non-overlapping frequency band intervals and the rate-restricted window length parameter pair are written into the constraint-updated parameter set; the constraint-updated parameter set is written into the frequency separation spectrum kurtosis analysis module and the frequency separation spectrum kurtosis analysis module is triggered to re-execute frequency band extraction and time slice division according to the constraint-updated parameter set, generating an updated kurtosis statistics sequence and an updated noise floor estimation sequence.

[0025] In this embodiment, the interference discrimination module specifically comprises: receiving a kurtosis statistics sequence, a noise floor estimation sequence, a change point location sequence, and a change point confidence sequence, and aligning them according to the frequency band identifier and time slice index to generate a discrimination record sequence; extracting kurtosis change indicators, noise floor change indicators, change point trigger indicators, and confidence weights from the discrimination record sequence and combining them to form an interference evidence vector sequence; performing time slice segmentation on the interference evidence vector sequence based on the change point location sequence and generating a candidate time period set; and performing cross-frequency band association on each frequency band identifier within the candidate time period set, wherein the cross-frequency band association includes consistency determination of interference evidence vectors in adjacent frequency bands and... Frequency band connectivity is determined, generating a candidate frequency band set. An interference discrimination template set is established for the candidate time period set and the candidate frequency band set. This template set includes templates for co-frequency superposition, adjacent-frequency superposition, sudden intermittent interference, and noise floor rise, and also includes a template triggering condition field. For each candidate time period and candidate frequency band, the corresponding interference evidence vector sequence is read and matched with the interference discrimination template set using the template triggering condition, generating an interference segment list containing interference type identifiers and interference segment confidence levels. Specifically, the template triggering condition is determined by the kurtosis of the interference evidence vector sequence. The change indicator, noise floor change indicator, change point trigger indicator, and confidence weight constitute and are bound to the candidate time period set and the candidate frequency band range set; the template trigger condition of the same-frequency superposition discrimination template limits the frequency band range of the candidate frequency band range set to converge to a single frequency band and limits the change point trigger indicator within the candidate time period set to show a single abrupt change and the kurtosis change indicator to a positive spike; the template trigger condition of the adjacent-frequency superposition discrimination template limits the candidate frequency band range set to cover more than two adjacent frequency bands and limits the kurtosis change indicator to show a synchronous spike between adjacent frequency bands and the frequency band range connectivity determination to be true; the template trigger condition of the sudden intermittent discrimination template limits the candidate time period. The set of change point trigger indicators are multiple discrete triggers, and the kurtosis change indicator is limited to repeated occurrences near multiple trigger points, and the confidence weight is higher than the minimum confidence at each trigger point; the template trigger condition of the noise floor rise discrimination template is limited to the noise floor change indicator being continuously rising, and the duration of the candidate time period set is limited to being greater than the minimum duration, and the kurtosis change indicator being in a low fluctuation state within the time period; the template trigger condition matching is based on the time slice test results within the time slice index range of the candidate time period set and outputs the trigger flag; the interference segment list is merged with adjacent time periods of the same type and adjacent frequency band ranges to generate an interference event list.

[0026] In this embodiment, the result output module specifically performs the following steps: receiving an interference event list and parsing it to obtain an event record sequence; reading an alarm mapping table based on the event record sequence and generating an alarm record sequence, wherein the alarm record sequence includes an alarm code, alarm level, event identifier, frequency band identifier, time slice index start point, time slice index end point, interference type identifier, and interference segment confidence level; the alarm mapping table specifically comprises: an alarm mapping table with interference type identifier as the primary key and alarm code and alarm level fields as output fields; the alarm mapping table includes alarm code and alarm level mapping entries corresponding to co-frequency superimposed interference identifiers and alarm code and alarm level mapping entries corresponding to adjacent-frequency superimposed interference identifiers; the alarm mapping table includes... The alarm mapping table includes alarm codes and alarm level mapping items corresponding to sudden intermittent interference identifiers, and alarm codes and alarm level mapping items corresponding to noise floor rise interference identifiers; the alarm mapping table includes a confidence level segmentation field and sets multiple confidence level segmentation mapping items for each interference type identifier to achieve alarm level changes with the confidence level of the interference segment; the alarm mapping table includes a time duration segmentation field and sets multiple time duration segmentation mapping items for each interference type identifier to achieve alarm level changes with the interval between the start and end of the time slice index; the alarm mapping table includes a bandwidth segmentation field and sets multiple bandwidth segmentation mapping items for each interference type identifier to achieve alarm code changes with the coverage range of the bandwidth identifier; the alarm record sequence... The alarm fingerprint is generated and matched with the alarm receipt table to generate a sending flag sequence. The sending flag sequence is then bound with the alarm record sequence to form a sequence of alarms to be sent. The alarm receipt table is used to record the receipt information after the result output module sends alarms to the network management interface and the local log interface, and is used for alarm deduplication control. The alarm receipt table contains an alarm fingerprint field and uses the alarm fingerprint field as an index key to locate sent alarm records. The alarm receipt table contains alarm code field, alarm level field, event identifier field, frequency band identifier field, time slice index start field, time slice index end field, interference type identifier field, interference segment confidence field, and is consistent with the alarm record sequence fields to support consistency comparison. The alarm receipt table includes a sending status field and a receipt time field. The sending status field is used to mark three states: sent, sent failed, and duplicate suppressed. The receipt time field is used to record the corresponding time slice index generated by the receipt. The alarm receipt table includes a receipt source field and marks the receipt as coming from the network management interface or the local log interface to support multi-channel receipt management. The alarm receipt table includes a failure control field and allows the same alarm fingerprint to re-enter the alarm sequence to be sent when the failure control field meets the failure conditions. A troubleshooting location record sequence is generated based on the event record sequence. The troubleshooting location record sequence includes a frequency band identifier, a time slice index start point, a time slice index end point, an interference type identifier, an interference segment confidence level, and an evidence index identifier.Based on the evidence index identifier, data segments within the range from the start to the end of the time slice index are read from the kurtosis statistics sequence cache, the noise floor estimation sequence cache, and the variable point location sequence cache, and encapsulated to generate an evidence package sequence; the alarm sequence to be sent, the troubleshooting and location record sequence, and the evidence package sequence are associated according to the event identifier and encapsulated to generate an output message sequence, which is then written to the output queue; the output message sequence is read from the output queue and sent to the network management interface and the local log interface, generating a sending receipt sequence and writing the sending receipt sequence to the alarm receipt table.

[0027] Example 1: To verify the feasibility of this invention in practice, a network signal interference detection system based on spectrum analysis disclosed in this invention was deployed in an online monitoring scenario of a wireless backhaul link carrying multiple services. The network communication link signal under test was accessed by the network signal acquisition module through the radio frequency coupling interface and distributed to the first sampling channel and the second sampling channel. The first sampling channel performed variable bandwidth analog pre-selective filtering to suppress strong out-of-band signals and output the first filtered radio frequency signal. The second sampling channel generated a reference noise radio frequency signal through an isolation terminal to characterize the noise contribution of the receiving link. The network signal acquisition module performed parallel dual-gain amplification and automatic gain control on the first filtered radio frequency signal and completed in-phase quadrature demodulation to obtain the first high-gain baseband sampling pair and the first low-gain baseband sampling pair. At the same time, it performed in-phase quadrature demodulation on the reference noise radio frequency signal to obtain the reference baseband sampling pair. The three baseband sampling pairs obtained a synchronous sampling sequence under a unified clock and were calibrated by amplitude and phase imbalance and DC bias before cross-gain fusion to generate a digital time-domain sampling sequence. In this scenario, common methods trigger alarms using noise floor thresholds and fixed window statistics. However, these methods are prone to alarm jitter and misjudgments when faced with the coexistence of overlapping co-frequency and adjacent frequencies, as well as sudden intermittent events. Furthermore, when the noise floor rises slowly and drifts, detection lag and inconsistencies between the location data and the alarm time can occur. This invention addresses these issues by establishing a loop linkage between a frequency separation spectrum kurtosis analysis module and a Bayesian online change point detection module. The updated frequency separation spectrum kurtosis analysis module performs closed-loop updates of the frequency band and time slices.

[0028] During online system operation, the spectrum preprocessing module receives the digitized time-domain sampling sequence and generates a frame index sequence and a segmented time-domain frame sequence. The frame length is set to 4096 complex sampling points, and the frame shift is set to 1024 complex sampling points. The spectrum preprocessing module performs endpoint continuity detection on the segmented time-domain frame sequence and calculates a windowing coefficient sequence based on the amplitude difference between the first and last sample values ​​and the amplitude difference between the first and last first-order difference. The windowing coefficient sequence is applied to obtain a windowed time-domain frame sequence, and the corresponding complex mean is calculated to perform DC component cancellation and generate a de-DC time-domain frame sequence. The spectrum preprocessing module calculates the amplitude square sequence on the de-DC time-domain frame sequence and constructs a matrix based on the median absolute deviation of the amplitude square sequence. A robust scaling estimate is constructed, and an amplitude-normalized sequence is generated using the robust scaling estimate. After amplitude normalization, a normalized time-domain frame sequence is obtained. Then, sample compression is performed, with the compression threshold set to the 0.999th percentile of the normalized amplitude sequence to suppress pulse spikes. The spectrum preprocessing module extracts the overlapping area of ​​adjacent compressed time-domain frame sequences and calculates the cross-correlation function. The hysteresis search range is set to ±8 sampling points, and parabolic interpolation is used near the maximum correlation hysteresis to obtain fractional hysteresis. Based on the fractional hysteresis, a resampling ratio sequence is generated, and fractional delay filtering resampling is implemented using a 32-tap windowed interpolation kernel. Finally, the spectrum calculation input sequence is output.

[0029] The spectrum calculation module writes the spectrum calculation input sequence into a circular buffer and forms a time-domain block sequence according to the number of transformation points (4096). Based on the sampling rate identifier, it selects a rotation factor lookup table, performs bit reversal rearrangement, and then distributes the results to a parallel butterfly operation array. Within a multi-stage pipeline, the rotation factor lookup table is called to complete complex multiplication and addition to obtain the frequency domain complex spectrum sequence. To control the fixed-point dynamic range, the spectrum calculation module performs block floating-point scaling on the frequency domain complex spectrum sequence. The scaling exponent is determined by the maximum absolute value of a single frame of the frequency domain complex spectrum sequence. The calibrated frequency domain complex spectrum sequence is subjected to complex conjugate product to obtain the power spectrum unit values, which are then arranged according to the frequency index to form a single frame of power spectrum. The single frame of power spectrum is spliced ​​together according to the time-domain block sequence to output the power spectrum sequence. The sampling rate is configured to be 30.72MHz, corresponding to a frequency resolution of approximately 7.5kHz.

[0030] The frequency separation spectral kurtosis analysis module receives the power spectrum sequence and constructs a frame power spectrum sequence. It calculates a frequency abrupt change indicator sequence along the frequency dimension of the frame power spectrum sequence. The abrupt change score uses the logarithmic power ratio of adjacent frequency points and suppresses smooth spectral fluctuations with robust baseline amplitude. The abrupt change score threshold is set to 3.2. Based on the frequency abrupt change indicator sequence, it performs continuous segment clustering on the frequency indices to generate frequency band boundary sequences and constructs multiple non-overlapping separation frequency band definition sets. The minimum boundary interval is set to 4 frequency indices. According to the separation frequency band definition sets, it performs frequency band extraction on the frame power spectrum sequence to form multiple separation frequency band power spectrum sets, and divides them into time slices. The frequency band time slice power block sequence is obtained, with a time slice length of 32 frames and an overlap of 16 frames. The mean sequence, variance sequence, and fourth-order central moment sequence are calculated for the frequency band time slice power block sequence, and a kurtosis statistics sequence is generated. The noise floor estimation sequence is obtained by low percentile value tracking filtering, with the low percentile position set to 0.1, the rising update coefficient set to 0.15, and the falling update coefficient set to 0.35, to ensure that the estimated rise is more gradual when the noise floor rises and falls more quickly after the interference disappears. The kurtosis statistics sequence and the noise floor estimation sequence are aligned with the separated frequency band definition set and time slice index and output to the Bayesian online change point detection module.

[0031] The Bayesian online change point detection module generates frequency band time-slice observation sequences and establishes a run-length state set for each frequency band identifier. The maximum run-length is configured to 200 to control the complexity of online updates. A burst marker sequence is generated based on the kurtosis statistics sequence, with the deviation of the change amplitude within a short window from the robust baseline as the trigger criterion, and a threshold of 2.8. A drift marker sequence is generated based on the noise floor estimation sequence, with the cumulative change deviation from the robust baseline within a long window as the trigger criterion, and a threshold of 1.6. A hazard rate sequence is selected based on the burst marker sequence and the drift marker sequence: a stationary hazard rate of 0.02, a burst hazard rate of 0.12, and a drift hazard rate of 0.06. The combined hazard rate is the average of the sudden hazard rate and the drift hazard rate. The run length state set is updated online posteriorly to the frequency band time slice observation sequence under the drive of the hazard rate sequence, and the run length posterior distribution sequence is output. The candidate change point location set is obtained by detecting the peak with a run length of 0 in the run length posterior distribution sequence, and the minimum peak interval is configured to be 2 time slices. The candidate change point location set is fused across frequency bands to obtain the change point location sequence and the change point confidence sequence. The confidence threshold is set to 0.7 for locking event generation. The maximum posterior run length is combined with the change point confidence sequence to generate adaptive time window parameters and output to the update frequency separation spectral kurtosis analysis module.

[0032] The frequency separation spectrum kurtosis analysis module receives the change point location sequence, change point confidence sequence, and adaptive time window parameters, and generates a feedback event sequence. It then performs confidence segmentation filtering on the feedback event sequence to generate a locked event set. Based on the locked event set, it performs boundary relocation on the separation frequency band definition set and performs window length mapping on the time slice division rules. Simultaneously, it applies stability constraints to the updated separation frequency band definition set and the updated time slice division rules to limit the minimum boundary interval and the rate of change of the window length. Within the first 20 minutes of online operation, the system completes 7 rounds of closed-loop updates. The number of separation frequency bands converges from 12 to 9, the adjustment amount of the main interference frequency band boundary decreases from 11 frequency indices to 2 frequency indices, the time slice length is adaptively adjusted from 32 frames to 48 frames and remains stable during the noise floor rise segment, the intra-time slice variance of the main interference frequency band kurtosis statistics sequence decreases by approximately 30%, and the peak positions of the change point confidence sequence become more concentrated. The interference discrimination module receives the kurtosis statistics sequence, noise floor estimation sequence, change point location sequence, and change point confidence sequence, and generates a discrimination record sequence. The discrimination record sequences are combined to form an interference evidence vector sequence, and a candidate time period set is generated based on the change point location sequence. Cross-frequency band correlation is performed within the candidate time period set to obtain a candidate frequency band range set. After the interference discrimination template set is matched under the template triggering conditions, an interference segment list is output and merged to generate an interference event list. The result output module generates an alarm record sequence from the interference event list and performs alarm deduplication by matching the alarm fingerprint with the alarm receipt table. At the same time, it generates a troubleshooting location record sequence and an evidence packet sequence, encapsulates the output message sequence, sends it to the network management interface and the local log interface, and writes a receipt.

[0033] To quantify the effect, six types of interference events were recorded during 2 hours of continuous online monitoring. The measured data were obtained by synchronizing independent spectrum records and link performance counters. The system output was obtained by parsing the interference event list and evidence packet sequence. The system triggered alarms for all six events without generating additional alarms. The event-level recall rate was 1.00, the event-level precision rate was 1.00, and the interference type consistency rate was 0.83. Compared with the control operation without enabling the frequency separation spectrum kurtosis analysis module, the control operation had two alarm jitters and one duplicate alarm. The mean absolute error of the start time decreased from 5.87ms to 2.16ms, the mean absolute error of the center frequency decreased from 25.5kHz to 9.7kHz, and the mean absolute error of the occupied bandwidth decreased from 41.8kHz to 24.2kHz. This indicates that the closed-loop linkage has a direct improvement effect on positioning stability and alarm consistency.

[0034] Table 1 Comparison of Measured and System Output Results for Interference Event Identification

[0035] As shown in Table 1, E01 and E02 maintain consistency in interference start and end times, center frequency, occupied bandwidth, and interference type identification. The start time errors are 1.5ms and 1.2ms, respectively, and the center frequency errors are 3.5kHz and 7.5kHz, respectively. This indicates that the separated frequency band definition set constructed from the frequency change indicator sequence under the single form of co-frequency and adjacent-frequency superposition can stably support the online segmentation of the kurtosis statistics sequence and the noise floor estimation sequence. E03 and E06 exhibit short-duration characteristics of sudden intermittent events. The system output start and end time errors are both controlled within the 3ms range, and the center frequency error is less than 10kHz, indicating that the run-length state set driven by the sudden hazard rate sequence can be quickly reset and form clear candidate change points. E04 exhibits continuous drift characteristics of noise floor rise, with the end time output... The output shows a slight lag but remains within the 4ms range, indicating that the combination of drift flag sequence and low percentile tracking filter can provide stable observation for change point detection. The interference type output of E05 differs from the actual measurement because the synchronous spike of the interference evidence vector in adjacent frequency bands and the connectivity of the frequency band range make the template triggering condition closer to the same-frequency superposition discrimination template. However, its time and frequency positioning error is still within the usable range and can be assisted by the kurtosis statistics and noise floor fragments provided by the evidence packet sequence for obstacle removal and positioning. This embodiment verifies that the present invention has the ability to perform online detection, discrimination and positioning output under interference modes such as same-frequency adjacent frequency superposition, sudden intermittent and noise floor rise, and verifies that the cyclic linkage of Bayesian online change point detection module and updated frequency separation spectrum kurtosis analysis module can reduce alarm jitter and improve positioning consistency.

[0036] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A network signal interference detection system based on spectrum analysis, characterized in that, include: The network signal acquisition module receives the network communication link signal under test and outputs a digital time-domain sampling sequence; The spectrum preprocessing module performs window function segmentation, DC component removal, and amplitude normalization on the digitized time-domain sampled sequence and outputs the spectrum calculation input sequence; the spectrum calculation module performs discrete Fourier transform on the spectrum calculation input sequence to generate a power spectrum sequence. The frequency separation spectrum kurtosis analysis module performs frequency separation processing on the power spectrum sequence to obtain multiple separated frequency band power spectrum sets, and forms a kurtosis statistics sequence and a noise floor estimation sequence according to time slices; the Bayesian online change point detection module performs online posterior update on the kurtosis statistics sequence and the noise floor estimation sequence, and outputs the change point location sequence, the change point confidence sequence, and the adaptive time window parameters; the parameter linkage module receives the change point location sequence, the change point confidence sequence, and the adaptive time window parameters, updates the frequency separation spectrum kurtosis analysis module, and recalculates the kurtosis statistics sequence and the noise floor estimation sequence; the interference discrimination module receives the kurtosis statistics sequence, the noise floor estimation sequence, the change point location sequence, and the change point confidence sequence, and generates a list of interference events; The results output module receives a list of interference events and outputs alarm information and troubleshooting location data.

2. The network signal interference detection system based on spectrum analysis according to claim 1, characterized in that, The network signal acquisition module specifically comprises: accessing the network communication link signal under test through a radio frequency coupling interface and allocating it to the first sampling channel and the second sampling channel; performing variable bandwidth analog pre-selection filtering in the first sampling channel and determining the center frequency and passband bandwidth according to a preset frequency band table, and outputting the first filtered radio frequency signal; A reference noise radio frequency signal is generated through an isolation terminal in the second sampling channel; Parallel dual-gain amplification and automatic gain control are performed on the first filtered RF signal to obtain a first high-gain intermediate frequency signal and a first low-gain intermediate frequency signal. In-phase quadrature demodulation is performed on the first high-gain intermediate frequency signal and the first low-gain intermediate frequency signal to obtain a first high-gain baseband sampling pair and a first low-gain baseband sampling pair. In-phase quadrature demodulation is performed on the reference noise RF signal to obtain a reference baseband sampling pair. A unified clock sampling is performed on the first high-gain baseband sampling pair, the first low-gain baseband sampling pair, and the reference baseband sampling pair to obtain a synchronous sampling sequence. Amplitude-phase imbalance calibration and DC bias calibration are performed on the synchronous sampling sequence, and cross-gain fusion is performed to generate a digital time-domain sampling sequence.

3. The network signal interference detection system based on spectrum analysis according to claim 1, characterized in that, The spectrum preprocessing module specifically comprises: receiving a digitized time-domain sampling sequence and generating a frame index sequence and a segmented time-domain frame sequence according to a preset frame length and a preset frame shift; performing endpoint continuity detection on the segmented time-domain frame sequence and calculating the amplitude difference between the first and last samples and the first and last first-order difference amplitude difference for each segmented time-domain frame, and calculating a windowing coefficient sequence; applying the windowing coefficient sequence to the segmented time-domain frame sequence to generate a windowed time-domain frame sequence, calculating the corresponding complex mean, and using the complex mean to perform DC component cancellation on the windowed time-domain frame sequence to generate a de-DC time-domain frame sequence; and calculating the amplitude square sequence and... Robust scaling estimates are used to generate amplitude-normalized sequences. Based on the amplitude-normalized sequences, amplitude normalization is performed on the de-DC time-domain frame sequences to generate normalized time-domain frame sequences, and sample compression is performed to generate compressed time-domain frame sequences. The overlapping area of ​​adjacent compressed time-domain frame sequences is extracted and the cross-correlation function is calculated. The resampling ratio sequence is obtained based on the maximum correlation hysteresis of the cross-correlation function and parabolic interpolation. Fractional delay filtering and resampling are performed on the compressed time-domain frame sequences to generate the spectrum calculation input frame sequence. The spectrum calculation input frame sequences are concatenated according to the frame index sequence to generate the spectrum calculation input sequence.

4. The network signal interference detection system based on spectrum analysis according to claim 1, characterized in that, The execution of the Discrete Fourier Transform (DFT) specifically involves: receiving the spectrum calculation input sequence and writing it into a circular buffer; forming a time-domain block sequence according to the number of transform points; selecting a twitch factor lookup table based on the sampling rate identifier; performing bit-reversal rearrangement on the time-domain block sequence and distributing the rearrangement result to a parallel butterfly operation array; performing butterfly operations in a multi-stage pipeline within the parallel butterfly operation array and calling the twitch factor lookup table at each stage to complete complex multiplication and addition, obtaining a frequency-domain complex spectrum sequence; performing block floating-point scaling on the frequency-domain complex spectrum sequence to obtain a calibrated frequency-domain complex spectrum sequence; performing complex conjugate multiplication on the calibrated frequency-domain complex spectrum sequence to obtain power spectrum unit values ​​and arranging them according to frequency index to form a single-frame power spectrum; and concatenating the single-frame power spectrum in the order of the time-domain block sequence to output a power spectrum sequence.

5. The network signal interference detection system based on spectrum analysis according to claim 1, characterized in that, The frequency separation spectral kurtosis analysis module specifically comprises: receiving a power spectrum sequence and constructing a frame power spectrum sequence in chronological order; calculating a frequency change indicator sequence along the frequency dimension of the frame power spectrum sequence; performing continuous segment clustering on the frequency index based on the frequency change indicator sequence and generating a frequency band boundary sequence; constructing multiple non-overlapping separation frequency band definition sets using the frequency band boundary sequence; performing frequency band extraction on the frame power spectrum sequence according to the separation frequency band definition sets and forming multiple separation frequency band power spectrum sets, and dividing the frequency band time slice power block sequence according to time slices; calculating the mean sequence, variance sequence, and fourth-order central moment sequence for the frequency band time slice power block sequence, and generating a kurtosis statistics sequence based on the mean sequence, variance sequence, and fourth-order central moment sequence; performing low percentile tracking filtering on the frequency band time slice power block sequence to generate a noise floor estimation sequence; aligning the kurtosis statistics sequence and the noise floor estimation sequence according to the separation frequency band definition set and time slice index and outputting them to the Bayesian online change point detection module.

6. The network signal interference detection system based on spectrum analysis according to claim 1, characterized in that, The Bayesian online change point detection module specifically includes: receiving a kurtosis statistics sequence and a noise floor estimation sequence, and generating a frequency band time slice observation sequence by aligning the separated frequency band definition set with the time slice index; establishing a run length state set for each frequency band identifier and configuring a state parameter cache and a sufficient statistics cache; A burst marker sequence is generated based on the kurtosis statistics sequence, and a drift marker sequence is generated based on the noise floor estimation sequence. A hazard rate sequence is selected based on the burst marker sequence and the drift marker sequence. The hazard rate sequence drives the run length state set to perform online posterior updates on the frequency band time slice observation sequence and synchronously updates the state parameter cache, outputting the run length posterior distribution sequence. A candidate change point location set is generated based on the run length posterior distribution sequence. Perform cross-band consistency fusion on the candidate variable point location set and output the variable point location sequence and variable point confidence sequence; extract the maximum a posteriori run length based on the run length posterior distribution sequence and combine it with the variable point confidence sequence to generate adaptive time window parameters.

7. The network signal interference detection system based on spectrum analysis according to claim 1, characterized in that, The updated frequency separation spectrum kurtosis analysis module specifically comprises: receiving a change point location sequence, a change point confidence sequence, and adaptive time window parameters, and generating a feedback event sequence by aligning the frequency band identifier with the time slice index; performing confidence segmentation filtering on the feedback event sequence and generating a locked event set; performing boundary relocation on the separation frequency band definition set based on the locked event set and generating an updated separation frequency band definition set; performing window length mapping on the time slice division rules based on the adaptive time window parameters and generating updated time slice division rules; performing stability constraints on the updated separation frequency band definition set and the updated time slice division rules and outputting a constrained updated parameter set; writing the constrained updated parameter set into the frequency separation spectrum kurtosis analysis module and triggering the frequency separation spectrum kurtosis analysis module to re-execute frequency band extraction and time slice division according to the constrained updated parameter set, generating an updated kurtosis statistics sequence and an updated noise floor estimation sequence.

8. The network signal interference detection system based on spectrum analysis according to claim 1, characterized in that, The interference discrimination module specifically comprises: receiving a kurtosis statistics sequence, a noise floor estimation sequence, a change point location sequence, and a change point confidence sequence, and aligning them according to the frequency band identifier and time slice index to generate a discrimination record sequence; extracting kurtosis change indicators, noise floor change indicators, change point trigger indicators, and confidence weights from the discrimination record sequence and combining them to form an interference evidence vector sequence; performing time slice segmentation on the interference evidence vector sequence based on the change point location sequence and generating a candidate time period set; performing cross-frequency band association on each frequency band identifier within the candidate time period set to generate a candidate frequency band range set; establishing an interference discrimination template set for the candidate time period set and the candidate frequency band range set; reading the corresponding interference evidence vector sequence for each candidate time period and candidate frequency band range and performing template trigger condition matching with the interference discrimination template set to generate an interference segment list containing interference type identifiers and interference segment confidence scores; The interference segment list is merged by combining adjacent time periods of the same type and adjacent frequency bands to generate an interference event list.

9. The network signal interference detection system based on spectrum analysis according to claim 1, characterized in that, The result output module specifically performs the following steps: receiving a list of interference events and parsing it to obtain an event record sequence; reading an alarm mapping table based on the event record sequence and generating an alarm record sequence; generating an alarm fingerprint from the alarm record sequence and matching it with an alarm receipt table to generate a sending flag sequence, and binding the sending flag sequence with the alarm record sequence to form a pending alarm sequence; generating a troubleshooting and location record sequence based on the event record sequence; reading data segments from the kurtosis statistics sequence cache, noise floor estimation sequence cache, and variable point location sequence cache based on the evidence index identifier and encapsulating them to generate an evidence packet sequence; associating the pending alarm sequence, troubleshooting and location record sequence, and evidence packet sequence according to the event identifier and encapsulating them to generate an output message sequence, and writing the output message sequence into an output queue; reading the output message sequence from the output queue and sending it to the network management interface and the local log interface, generating a sending receipt sequence, and writing the sending receipt sequence into an alarm receipt table.

Images