An adaptive electromagnetic wave signal anomaly detection and classification system

Abstract

The application discloses an adaptive electromagnetic wave signal anomaly detection and classification system, which can synchronously collect electromagnetic wave emission signals and power data in a monitoring period, extract a spectrum distortion index and a power fluctuation index, and fuse and generate a risk coefficient, realizes anomaly discrimination and accidental / unstable / persistent classification, and adaptively updates a reference template, a threshold value and a weight under the constraint of a credible normal sample, so that the adaptive electromagnetic wave signal anomaly detection and classification system realizes long-term stable and traceable electromagnetic wave signal anomaly detection and classification. The system is characterized in that it comprises a signal collection and synchronization module, a preprocessing module, a spectrum feature extraction module, a power feature extraction module, a risk fusion evaluation module, an anomaly discrimination module, an anomaly classification module, an adaptive updating module and a result output and tracing module, the signal collection and synchronization module is used for acquiring original observation quantities for anomaly discrimination and completing multi-source time alignment in a monitoring period.

Description

Technical Field

[0001] This invention relates to an adaptive electromagnetic wave signal anomaly detection and classification system, which is an adaptive anomaly detection and classification method and system for electromagnetic wave transmitted signals. It belongs to the field of intelligent signal processing and anomaly detection technology. In particular, it relates to an adaptive electromagnetic wave signal anomaly detection and classification system that can simultaneously collect electromagnetic wave transmitted signals and power data during a monitoring period, extract spectral distortion index and power fluctuation index and fuse them to generate risk coefficient, realize anomaly discrimination and random / unstable / persistent classification, and adaptively update the benchmark template, threshold and weight under the constraint of credible normal samples, so as to achieve long-term stable and traceable electromagnetic wave signal anomaly detection and classification. Background Technology

[0002] Currently, in the application of artificial intelligence technology for intelligent operation and maintenance and health management of electromagnetic wave signals, mainstream solutions typically use the time-domain or frequency-domain data of the transmitted signal as input, utilizing feature engineering combined with classification models, or employing unsupervised anomaly detection models (such as reconstruction error, clustering distance, probability density deviation, etc.) to achieve anomaly identification. Some solutions also use the detection results for alarm classification and attempt to adapt to changes in equipment status and electromagnetic environment through continuous learning or online updates to improve the model's generalization ability and real-time performance. Existing mainstream artificial intelligence solutions still have two prominent technical shortcomings. First, the detection criteria often lean towards a single representation or a single model output. Common practices use spectral shift, out-of-band energy, reconstruction error, or a single confidence score as the main basis, making it difficult to uniformly quantify and jointly decide on spectral morphology anomalies and power stability anomalies within the same criterion system. When multiple sources of anomalies, such as external interference, modulation distortion, power amplifier drift, and power supply fluctuations, overlap or alternate, problems such as missed detections, false alarms, or unclear alarm interpretations easily occur, making it difficult for operation and maintenance personnel to quickly locate the source of the anomaly and take corresponding measures based on the detection results. Secondly, there are insufficient adaptability and traceability. Many solutions rely on offline training and manual recalibration. When faced with changes in operating conditions, equipment aging, or environmental changes, the model and thresholds are prone to drift, leading to an increase in false alarm rates or the masking of anomalies over long-term operation. Even when online learning or self-updating mechanisms are introduced, there is often a lack of strict constraints on credible normal samples, posing a risk of introducing abnormal samples into the update process, resulting in benchmark drift and criterion degradation. At the same time, existing solutions generally lack parameter version management and result evidence chains, making it difficult to clearly define the source of the benchmark template, threshold, and weights corresponding to the detection conclusion at a certain moment. This makes the detection results difficult to verify and audit, resulting in high maintenance and iteration costs.

[0003] Publication No. CN112924749B discloses an unsupervised adversarial learning method for detecting electromagnetic spectrum anomalies. This method involves: preprocessing acquired power spectrum data to obtain a power spectral density estimate; constructing a deep learning-based electromagnetic spectrum anomaly detection model using the power spectral density estimate; determining the reconstruction error Lr and the discriminator loss Ld for any power spectrum data using the electromagnetic spectrum anomaly detection model; determining the anomaly result Aresult based on the reconstruction error Lr and the discriminator loss Ld; and determining whether the electromagnetic spectrum data is abnormal based on the anomaly result Aresult. The above detection method primarily relies on the reconstruction error Lr and discriminator loss Ld output by the deep model for anomaly determination, and sets the hyperparameter ε as the judgment criterion using a validation set according to the false alarm probability requirement. Therefore, when equipment operating conditions or electromagnetic environment distribution drift, the error distribution and threshold matching relationship easily change, causing instability in the judgment criterion. Furthermore, its anomaly conclusions are more often reflected in an increase in error, making it difficult to directly correspond to explainable causes such as spectral morphology deviation or power stability fluctuations, thus increasing engineering positioning and maintenance costs. Summary of the Invention

[0004] To improve the above situation, the present invention provides an adaptive electromagnetic wave signal anomaly detection and classification system. This system can synchronously collect electromagnetic wave emission signal and power data during the monitoring period, extract the spectral distortion index and power fluctuation index and fuse them to generate a risk coefficient, realize anomaly discrimination and accidental / unstable / persistent classification, and adaptively update the benchmark template, threshold and weight under the constraint of credible normal samples, so as to achieve long-term stable and traceable electromagnetic wave signal anomaly detection and classification.

[0005] The adaptive electromagnetic wave signal anomaly detection and classification system of the present invention is implemented as follows: The adaptive electromagnetic wave signal anomaly detection and classification system of the present invention includes a signal acquisition and synchronization module, a preprocessing module, a spectrum feature extraction module, a power feature extraction module, a risk fusion assessment module, an anomaly discrimination module, an anomaly classification module, an adaptive update module, and a result output and traceability module.

[0006] The system is characterized in that: the signal acquisition and synchronization module establishes signal interaction with the preprocessing module and the result output and traceability module; the preprocessing module establishes signal interaction with the spectrum feature extraction module and the power feature extraction module; the spectrum feature extraction module establishes signal interaction with the risk fusion assessment module and the adaptive update module; the power feature extraction module establishes signal interaction with the risk fusion assessment module; the risk fusion assessment module establishes signal interaction with the anomaly detection module and the adaptive update module; the anomaly detection module establishes signal interaction with the anomaly classification module, the adaptive update module, and the result output and traceability module; the anomaly classification module establishes signal interaction with the result output and traceability module; and the adaptive update module establishes signal interaction with the result output and traceability module.

[0007] The signal acquisition and synchronization module is used to acquire raw observations for anomaly detection within the monitoring period and complete multi-source time alignment, providing consistent input for subsequent feature construction.

[0008] The signal acquisition and synchronization module includes a time-domain signal acquisition submodule, a power acquisition submodule, an operating context acquisition submodule, and a period identification and synchronization submodule.

[0009] The time-domain signal acquisition submodule is used to acquire the time-domain sampling sequence x[n] of the transmitted signal within the monitoring period T, and record the sampling rate. Number of sampling points N, channel number and timestamp,

[0010] Preferably, the monitoring period T can be set to 0.5–5s, and N can be set to 2048–65536 points to meet the spectral resolution requirements of devices in different frequency bands.

[0011] The power acquisition submodule is used to acquire the transmit power sequence p[m] within the same monitoring period and record the sampling period and timestamp.

[0012] Preferably, power sampling can be provided by a power meter, power amplifier sampling, standing wave ratio (SWR) detection and conversion, or by the device's built-in sensing channel.

[0013] The runtime context acquisition submodule is used to collect runtime information related to stability determination.

[0014] Preferably, the operating information includes at least one or a combination of the following: operating frequency band identifier, operating mode identifier, power level identifier, temperature, power supply status, or standing wave ratio (VSWR), for subsequent operating condition identification and adaptive triggering.

[0015] The cycle identifier and synchronization submodule aligns x[n], p[m] with the running information in time, and generates a cycle number and data packet identifier for each monitoring cycle, which are used to correspond one-to-one with the detection results and for traceability and review.

[0016] The signal acquisition and synchronization module also includes an integrated status acquisition box. This integrated status acquisition box includes an acquisition control board, a mounting base, a temperature sensor, an acquisition box housing, a communication connection port, a power terminal, and a power signal input terminal. The acquisition control board is fixedly mounted on the top surface of the mounting base. The temperature sensor is fixedly mounted on the mounting base and forms thermal contact with it. The temperature sensor is electrically connected to the acquisition control board. A snap-fit ​​groove is circumferentially formed along the edge of the upper surface of the mounting base. The bottom end of the acquisition box housing is open, and its bottom end has a snap-fit ​​flange that mates with the snap-fit ​​groove. The snap-fit ​​flange is embedded into the snap-fit ​​groove to achieve data acquisition. The acquisition box housing and the mounting base are positioned and snapped together. The mounting base and the acquisition box housing are also detachably fixed together by fixing screws, so that the mounting base and the acquisition box housing enclose a receiving cavity. The acquisition control board and the temperature sensor are disposed within the receiving cavity. One side of the acquisition box housing is provided with a communication connection port, a power terminal, and a power signal input terminal, which are electrically connected to the acquisition control board. Preferably, the acquisition control board integrates signal conditioning and sampling circuits, and further integrates a clock unit. The acquisition box housing is a metal shielded housing or a flame-retardant insulating housing.

[0017] The preprocessing module is used to normalize the collected data and extract effective segments, reducing the disturbance of noise, spikes, and missing data to feature calculation.

[0018] The preprocessing module includes a time-domain normalization submodule, a power sequence cleaning submodule, and an effective transmit segment extraction submodule.

[0019] The time-domain normalization submodule is used to perform DC removal, windowing, and amplitude normalization on x[n], and can set bandpass / notch filtering according to the device frequency band to suppress non-operating frequency band interference.

[0020] The power sequence cleaning submodule is used to suppress outliers, remove missing segments, or impute p[m], and output a valid power sequence that can be used for statistical analysis.

[0021] The effective transmit segment extraction submodule is used to identify effective transmit segments in pulse / gated operation mode. Only x[n] and p[m] within the effective transmit segment are included in the subsequent SDI and PFI calculations, thereby avoiding the dilution of abnormal features in the unloaded segment.

[0022] The spectral feature extraction module is used to convert the time-domain signal to the frequency domain and construct the spectral distortion index (SDI) to reflect the degree of deviation of the spectral shape from the reference template.

[0023] The spectral feature extraction module includes a frequency domain transformation submodule, a reference spectrum template management submodule, and a spectral distortion index calculation submodule.

[0024] The frequency domain transformation submodule is used to perform a discrete Fourier transform on the preprocessed x[n] to obtain the spectrum X[k], and output the amplitude or energy distribution within a set frequency band Ω.

[0025] Preferably, the frequency band Ω is configured in association with the operating frequency band identifier to ensure that the deviation metric is focused on the effective transmission frequency band of the device.

[0026] The reference spectrum template management submodule is used to maintain the reference spectrum template under normal operating conditions. And establish a correspondence between the template and the operating frequency band identifier, operating mode identifier, and power level identifier.

[0027] Preferably, the reference spectrum template management submodule performs an initialization process after the system is first deployed or the equipment is replaced: during the installation and commissioning phase, M stable monitoring periods are collected (M can be set to 50–200), and when these periods all meet the low-risk condition, their spectra are aggregated to generate the initial spectrum. If the conditions are not met, the data collection will be extended until the initialization constraints are met, in order to avoid writing abnormal states into the template.

[0028] The spectral distortion index calculation submodule is used to calculate... and The deviation is used to generate SDI, which employs an energy-normalized error metric: Where ε is a constant to prevent the denominator from being zero.

[0029] Preferably, the spectral distortion index calculation submodule can assign weights to different subbands within Ω, making the system more sensitive to key frequency points, sideband structures, or modulation characteristics. However, the weights remain fixed under the same operating conditions to ensure comparability.

[0030] The power feature extraction module is used to construct a power fluctuation index (PFI) from the power sequence to reflect the degree of deviation in transmit power stability.

[0031] The power feature extraction module includes a power statistics calculation submodule, a deviation set construction submodule, and a power fluctuation index calculation submodule.

[0032] The power statistics calculation submodule is used to calculate the mean of the power sequence within the monitoring period. with standard deviation The deviation set construction submodule is used to base on Construct a set of deviations and extract the maximum deviation. , minimum deviation and peak-valley difference ,

[0033] The power fluctuation index calculation submodule is used to normalize and combine the above statistics and generate the PFI: Where w1, w2, w3 are weights and satisfy ε is the zero-prevention constant.

[0034] Preferably, w1, w2, and w3 are fixed as equipment configuration parameters during the factory or commissioning phase and remain unchanged before the adaptive switching is triggered, so as to ensure the statistical consistency of PFI.

[0035] The risk fusion assessment module is used to fuse SDI and PFI and output a unified anomaly risk coefficient R to form a single discriminant and reduce the uncertainty of multi-indicator decision-making.

[0036] The risk fusion assessment module includes a fusion weight management submodule and a risk coefficient calculation submodule.

[0037] The fusion weight management submodule is used to provide α and β, satisfying α and β ≥ 0 and α + β = 1.

[0038] The risk coefficient calculation submodule is used to calculate the risk coefficient using the following formula: ,

[0039] Preferably, the initial values ​​of α and β are determined during the debugging phase (e.g., α can be set to 0.4–0.7, and β can be set to 0.3–0.6), and remain unchanged until the triggering condition of the adaptive update module is met, in order to avoid false alarm drift caused by frequent weight fluctuations.

[0040] The anomaly detection module is used to output the detection results for each monitoring period based on the risk coefficient and threshold set, and to generate a confidence level related to alarm linkage.

[0041] The anomaly detection module includes a threshold initialization and management submodule, a detection execution submodule, and a confidence generation submodule.

[0042] The threshold initialization and management submodule is used to maintain normal thresholds. With abnormal threshold (or single threshold Th), and establish a correspondence with frequency band / mode / power level.

[0043] Preferably, the threshold initialization and management submodule performs threshold initialization synchronously after template initialization: this involves initializing the risk coefficient set during the initialization phase. Robust statistics are performed by setting thresholds using quantiles and robust dispersion. Desirable The 0.95-0.99 quantiles, Desirable The median is increased by 2-4 times the MAD (median absolute deviation) to control the initial false alarm rate while taking into account individual device differences.

[0044] The discrimination execution submodule is used to compare R with the threshold and output the result. When a dual threshold strategy is adopted, it satisfies the following condition. Output abnormal cycle, satisfying Output a reliable normal cycle, satisfying Output the warning cycle and add it to the observation queue.

[0045] The confidence level generation submodule is used to map R to a confidence level C between 0 and 1. This confidence level is used by the upper-level system to set a tiered handling strategy, where C can be based on... The piecewise normalization implementation makes it closer to The early warning cycle has a higher priority.

[0046] Preferably, the anomaly detection module outputs a spectral contribution term simultaneously with the output result. With power contribution items This is used to explain the source of the anomaly and facilitate maintenance and troubleshooting.

[0047] The anomaly classification module is used to classify anomaly patterns after an anomaly or warning occurs, and output anomaly type labels that can be used for differentiated handling.

[0048] The anomaly classification module includes an observation window management submodule, a statistical feature calculation submodule, and a type determination submodule. The observation window management submodule is used to construct a sliding observation window W and aggregate the risk coefficient sequence. The window length can be set to 10-60 monitoring cycles, and the update frequency can be adjusted according to the alarm level.

[0049] The statistical feature calculation submodule is used to calculate statistical measures such as the proportion of anomalies within the window, the length of consecutive anomalies, and the mean and standard deviation of R, and to form the basis for classification.

[0050] The type determination submodule is used to output the anomaly type, which includes at least random anomalies, unstable anomalies, and persistent anomalies. Random anomalies are characterized by a low anomaly percentage and short consecutive anomaly length (e.g., an anomaly percentage not exceeding 0.1-0.2 and a consecutive anomaly length not exceeding 1-2 periods). Unstable anomalies are characterized by a moderate anomaly percentage and significant fluctuations (e.g., an anomaly percentage between 0.2-0.6 and a standard deviation higher than a set threshold). Persistent anomalies are characterized by a high anomaly percentage or a long consecutive anomaly length (e.g., an anomaly percentage not less than 0.6-0.8 or a consecutive anomaly length not less than 5-10 periods).

[0051] Preferably, the anomaly classification module uses the same window statistical framework for both the warning period and the anomaly period, enabling the system to output warning information indicating an unstable or persistent trend in the early stages of an anomaly, thereby allowing for early intervention.

[0052] The adaptive update module is used to update the template, threshold, and weights online without introducing abnormal contamination, forming a closed loop of detection-classification-self-calibration.

[0053] The adaptive update module includes a trusted normal sample pool management submodule, a drift detection and triggering submodule, a benchmark spectrum template update submodule, a threshold adaptive update submodule, and a fusion weight adaptive update submodule.

[0054] The trusted normal sample pool management submodule is used to write trusted normal cycles into the sample pool and manage them by frequency band / mode / power level to avoid cross-operating condition mixing that could cause template distortion.

[0055] The drift detection and triggering submodule is used to determine when template / threshold / weight updates are allowed. The triggering conditions include at least one of the following: First, the operating context identifier changes (e.g., frequency band, mode, or power level switch); second, within K consecutive periods (K can be set to 20-100), the median SDI, PFI, or R of the reliable normal samples deviates from the historical baseline by more than a set proportion δ (δ can be set to 20%-40%), which is determined to be statistical drift; third, the warning proportion continuously exceeds the target interval within a continuous window.

[0056] The reference spectrum template update submodule is used to update only the reliable normal spectrum within the sample pool. And use an exponential moving average: λ can be set to 0.01-0.10. When drift is triggered and a new operating condition is confirmed, λ can be briefly increased to 0.10-0.30 to accelerate convergence. After stabilization, it should be restored to its normal value.

[0057] The threshold adaptive update submodule is used to update based on the set of credible normal risk coefficients in the sample pool. , It employs a quantile and robust dispersion reestimation strategy, entering a fast reestimation mode when drift is triggered, shortening the statistical window and prioritizing updates. To suppress the spread of false alarms, updates will be made only after the sample pool capacity has recovered to the set size. ,

[0058] The fusion weight adaptive update submodule is used to adjust α and β in small steps when the trigger condition is met, so that the warning ratio returns to the target range and the abnormal classification result is stable. The adjustment is constrained by being slow and unidirectionally restricted, and the single adjustment range can be limited to 0.02-0.05 to prevent discrimination drift caused by weight oscillation.

[0059] Preferably, the adaptive update module receives external confirmation information (annotations from maintenance personnel or verification results from the upper-level system) as the basis for correction. When the external confirmation and the system's judgment are inconsistent for a long period of time, threshold re-estimation is triggered first rather than template update to avoid incorrect corrections being written into the baseline.

[0060] The result output and traceability module is used to output structured test results and establish a verifiable chain of evidence to meet the needs of equipment maintenance, quality traceability, and operation auditing.

[0061] The result output and traceability module includes a result encapsulation submodule, an alarm linkage submodule, and a data archiving submodule.

[0062] The result encapsulation submodule is used to output a result data packet that includes at least the period number, timestamp, SDI, PFI, R, discrimination result, anomaly type, confidence level, operating condition identifier, and parameter version number.

[0063] The alarm linkage submodule is used to trigger alarms under abnormal or high-confidence warning conditions and output handling suggestions. The handling suggestions include at least one or more of the following: recalibration suggestions, power supply check suggestions, power amplifier thermal management check suggestions, and antenna matching check suggestions. The suggestions are automatically prioritized based on the spectrum contribution item / power contribution item.

[0064] The data archiving submodule is used to archive key features, discrimination results, parameter versions, and necessary original data summaries. It supports review and comparison by periodic number and parameter version, thereby ensuring that the detection results of the same device at different stages are comparable and interpretable.

[0065] Beneficial effects

[0066] First, it can perform a joint assessment of the spectrum distortion index and the power fluctuation index and integrate them into a single risk coefficient, thereby achieving unified detection of anomalies from multiple sources such as spectrum distortion, out-of-band interference, power amplifier drift, and power supply fluctuation, and improving the detection accuracy and interpretability under complex operating conditions.

[0067] Second, it can adaptively update the benchmark spectrum template, discrimination threshold and fusion weight under the constraint of reliable normal samples, and archive and trace the parameter version, so that the system can maintain the stability of the judgment scale and the results can be verified under the conditions of operating condition switching, equipment aging and environmental changes, thereby reducing long-term false alarms and maintenance costs. Attached Figure Description

[0068] Figure 1 This is a three-dimensional structural diagram of the integrated status acquisition box of the present invention.

[0069] In the attached diagram: mounting base (1), acquisition control board (2), temperature sensor (3), acquisition box housing (4), communication interface (5), power input terminal (6), power monitoring input terminal (7). Detailed Implementation

[0070] Example 1:

[0071] This invention discloses an adaptive electromagnetic wave signal anomaly detection and classification system, comprising a signal acquisition and synchronization module, a preprocessing module, a spectrum feature extraction module, a power feature extraction module, a risk fusion assessment module, an anomaly discrimination module, an anomaly classification module, an adaptive update module, and a result output and traceability module.

[0072] The system is characterized in that: the signal acquisition and synchronization module establishes signal interaction with the preprocessing module and the result output and traceability module; the preprocessing module establishes signal interaction with the spectrum feature extraction module and the power feature extraction module; the spectrum feature extraction module establishes signal interaction with the risk fusion assessment module and the adaptive update module; the power feature extraction module establishes signal interaction with the risk fusion assessment module; the risk fusion assessment module establishes signal interaction with the anomaly detection module and the adaptive update module; the anomaly detection module establishes signal interaction with the anomaly classification module, the adaptive update module, and the result output and traceability module; the anomaly classification module establishes signal interaction with the result output and traceability module; and the adaptive update module establishes signal interaction with the result output and traceability module.

[0073] The signal acquisition and synchronization module is used to acquire raw observations for anomaly detection within the monitoring period and complete multi-source time alignment, providing consistent input for subsequent feature construction.

[0074] The signal acquisition and synchronization module includes a time-domain signal acquisition submodule, a power acquisition submodule, an operating context acquisition submodule, and a period identification and synchronization submodule.

[0075] The time-domain signal acquisition submodule is used to acquire the time-domain sampling sequence x[n] of the transmitted signal within the monitoring period T, and record the sampling rate. Number of sampling points N, channel number and timestamp,

[0076] Preferably, the monitoring period T can be set to 0.5–5s, and N can be set to 2048–65536 points to meet the spectral resolution requirements of devices in different frequency bands.

[0077] The power acquisition submodule is used to acquire the transmit power sequence p[m] within the same monitoring period and record the sampling period and timestamp.

[0078] Preferably, power sampling can be provided by a power meter, power amplifier sampling, standing wave ratio (SWR) detection and conversion, or by the device's built-in sensing channel.

[0079] The runtime context acquisition submodule is used to collect runtime information related to stability determination.

[0080] Preferably, the operating information includes at least one or a combination of the following: operating frequency band identifier, operating mode identifier, power level identifier, temperature, power supply status, or standing wave ratio (VSWR), for subsequent operating condition identification and adaptive triggering.

[0081] The cycle identifier and synchronization submodule aligns x[n], p[m] with the running information in time, and generates a cycle number and data packet identifier for each monitoring cycle, which are used to correspond one-to-one with the detection results and for traceability and review.

[0082] The signal acquisition and synchronization module also includes an integrated status acquisition box, which includes an acquisition control board (2), a mounting base (1), a temperature sensor (3), an acquisition box housing (4), a communication connection port (5), a power terminal (6), and a power signal input terminal (7). The acquisition control board (2) is fixedly mounted on the top surface of the mounting base (1), and the temperature sensor (3) is fixedly mounted on the mounting base (1) and forms thermal contact with the mounting base (1). The temperature sensor (3) is electrically connected to the acquisition control board (2). The upper surface edge of the mounting base (1) is provided with a snap-fit ​​groove along the circumferential direction. The bottom end of the acquisition box housing (4) is an open structure, and its bottom end is provided with a snap-fit ​​flange that cooperates with the snap-fit ​​groove. The snap-fit ​​flange is embedded in the snap-fit ​​groove to realize the acquisition box housing (4) Positioning engagement with mounting base (1), mounting base (1) and acquisition box housing (4) are also detachably fixedly connected by fixing screws, so that mounting base (1) and acquisition box housing (4) enclose to form an accommodating cavity, acquisition control board (2) and temperature sensor (3) are set in the accommodating cavity, one side of acquisition box housing (4) is provided with communication connection port (5), power terminal (6) and power signal input terminal (7), the communication connection port (5), power terminal (6) and power signal input terminal (7) are electrically connected to acquisition control board (2) respectively, preferably, acquisition control board (2) integrates signal conditioning and sampling circuit, acquisition control board (2) further integrates clock unit, acquisition box housing (4) is metal shielding housing or flame retardant insulating housing,

[0083] The preprocessing module is used to normalize the collected data and extract effective segments, reducing the disturbance of noise, spikes, and missing data to feature calculation.

[0084] The preprocessing module includes a time-domain normalization submodule, a power sequence cleaning submodule, and an effective transmit segment extraction submodule.

[0085] The time-domain normalization submodule is used to perform DC removal, windowing, and amplitude normalization on x[n], and can set bandpass / notch filtering according to the device frequency band to suppress non-operating frequency band interference.

[0086] The power sequence cleaning submodule is used to suppress outliers, remove missing segments, or impute p[m], and output a valid power sequence that can be used for statistical analysis.

[0087] The effective transmit segment extraction submodule is used to identify effective transmit segments in pulse / gated operation mode. Only x[n] and p[m] within the effective transmit segment are included in the subsequent SDI and PFI calculations, thereby avoiding the dilution of abnormal features in the unloaded segment.

[0088] The spectral feature extraction module is used to convert the time-domain signal to the frequency domain and construct the spectral distortion index (SDI) to reflect the degree of deviation of the spectral shape from the reference template.

[0089] The spectral feature extraction module includes a frequency domain transformation submodule, a reference spectrum template management submodule, and a spectral distortion index calculation submodule.

[0090] The frequency domain transformation submodule is used to perform a discrete Fourier transform on the preprocessed x[n] to obtain the spectrum X[k], and output the amplitude or energy distribution within a set frequency band Ω.

[0091] Preferably, the frequency band Ω is configured in association with the operating frequency band identifier to ensure that the deviation metric is focused on the effective transmission frequency band of the device.

[0092] The reference spectrum template management submodule is used to maintain the reference spectrum template under normal operating conditions. And establish a correspondence between the template and the operating frequency band identifier, operating mode identifier, and power level identifier.

[0093] Preferably, the reference spectrum template management submodule performs an initialization process after the system is first deployed or the equipment is replaced: during the installation and commissioning phase, M stable monitoring periods are collected (M can be set to 50–200), and when these periods all meet the low-risk condition, their spectra are aggregated to generate the initial spectrum. If the conditions are not met, the data collection will be extended until the initialization constraints are met, in order to avoid writing abnormal states into the template.

[0094] The spectral distortion index calculation submodule is used to calculate... and The deviation is used to generate SDI, which employs an energy-normalized error metric: Where ε is a constant to prevent the denominator from being zero.

[0095] Preferably, the spectral distortion index calculation submodule can assign weights to different subbands within Ω, making the system more sensitive to key frequency points, sideband structures, or modulation characteristics. However, the weights remain fixed under the same operating conditions to ensure comparability.

[0096] The power feature extraction module is used to construct a power fluctuation index (PFI) from the power sequence to reflect the degree of deviation in transmit power stability.

[0097] The power feature extraction module includes a power statistics calculation submodule, a deviation set construction submodule, and a power fluctuation index calculation submodule.

[0098] The power statistics calculation submodule is used to calculate the mean of the power sequence within the monitoring period. with standard deviation The deviation set construction submodule is used to base on Construct a set of deviations and extract the maximum deviation. , minimum deviation and peak-valley difference ,

[0099] The power fluctuation index calculation submodule is used to normalize and combine the above statistics and generate the PFI: Where w1, w2, w3 are weights and satisfy ε is the zero-prevention constant.

[0100] Preferably, w1, w2, and w3 are fixed as equipment configuration parameters during the factory or commissioning phase and remain unchanged before the adaptive switching is triggered, so as to ensure the statistical consistency of PFI.

[0101] The risk fusion assessment module is used to fuse SDI and PFI and output a unified anomaly risk coefficient R to form a single discriminant and reduce the uncertainty of multi-indicator decision-making.

[0102] The risk fusion assessment module includes a fusion weight management submodule and a risk coefficient calculation submodule.

[0103] The fusion weight management submodule is used to provide α and β, satisfying α and β ≥ 0 and α + β = 1.

[0104] The risk coefficient calculation submodule is used to calculate the risk coefficient using the following formula: ,

[0105] Preferably, the initial values ​​of α and β are determined during the debugging phase (e.g., α can be set to 0.4–0.7, and β can be set to 0.3–0.6), and remain unchanged until the triggering condition of the adaptive update module is met, in order to avoid false alarm drift caused by frequent weight fluctuations.

[0106] The anomaly detection module is used to output the detection results for each monitoring period based on the risk coefficient and threshold set, and to generate a confidence level related to alarm linkage.

[0107] The anomaly detection module includes a threshold initialization and management submodule, a detection execution submodule, and a confidence generation submodule.

[0108] The threshold initialization and management submodule is used to maintain normal thresholds. With abnormal threshold (or single threshold Th), and establish a correspondence with frequency band / mode / power level.

[0109] Preferably, the threshold initialization and management submodule performs threshold initialization synchronously after template initialization: this involves initializing the risk coefficient set during the initialization phase. Robust statistics are performed by setting thresholds using quantiles and robust dispersion. Desirable The 0.95-0.99 quantiles, Desirable The median is increased by 2-4 times the MAD (median absolute deviation) to control the initial false alarm rate while taking into account individual device differences.

[0110] The discrimination execution submodule is used to compare R with the threshold and output the result. When a dual threshold strategy is adopted, it satisfies the following condition. Output abnormal cycle, satisfying Output a reliable normal cycle, satisfying Output the warning cycle and add it to the observation queue.

[0111] The confidence level generation submodule is used to map R to a confidence level C between 0 and 1. This confidence level is used by the upper-level system to set a tiered handling strategy, where C can be based on... The piecewise normalization implementation makes it closer to The early warning cycle has a higher priority.

[0112] Preferably, the anomaly detection module outputs a spectral contribution term simultaneously with the output result. With power contribution items This is used to explain the source of the anomaly and facilitate maintenance and troubleshooting.

[0113] The anomaly classification module is used to classify anomaly patterns after an anomaly or warning occurs, and output anomaly type labels that can be used for differentiated handling.

[0114] The anomaly classification module includes an observation window management submodule, a statistical feature calculation submodule, and a type determination submodule. The observation window management submodule is used to construct a sliding observation window W and aggregate the risk coefficient sequence. The window length can be set to 10-60 monitoring cycles, and the update frequency can be adjusted according to the alarm level.

[0115] The statistical feature calculation submodule is used to calculate statistical measures such as the proportion of anomalies within the window, the length of consecutive anomalies, and the mean and standard deviation of R, and to form the basis for classification.

[0116] The type determination submodule is used to output the anomaly type, which includes at least random anomalies, unstable anomalies, and persistent anomalies. Random anomalies are characterized by a low anomaly percentage and short consecutive anomaly length (e.g., an anomaly percentage not exceeding 0.1-0.2 and a consecutive anomaly length not exceeding 1-2 periods). Unstable anomalies are characterized by a moderate anomaly percentage and significant fluctuations (e.g., an anomaly percentage between 0.2-0.6 and a standard deviation higher than a set threshold). Persistent anomalies are characterized by a high anomaly percentage or a long consecutive anomaly length (e.g., an anomaly percentage not less than 0.6-0.8 or a consecutive anomaly length not less than 5-10 periods).

[0117] Preferably, the anomaly classification module uses the same window statistical framework for both the warning period and the anomaly period, enabling the system to output warning information indicating an unstable or persistent trend in the early stages of an anomaly, thereby allowing for early intervention.

[0118] The adaptive update module is used to update the template, threshold, and weights online without introducing abnormal contamination, forming a closed loop of detection-classification-self-calibration.

[0119] The adaptive update module includes a trusted normal sample pool management submodule, a drift detection and triggering submodule, a benchmark spectrum template update submodule, a threshold adaptive update submodule, and a fusion weight adaptive update submodule.

[0120] The trusted normal sample pool management submodule is used to write trusted normal cycles into the sample pool and manage them by frequency band / mode / power level to avoid cross-operating condition mixing that could cause template distortion.

[0121] The drift detection and triggering submodule is used to determine when template / threshold / weight updates are allowed. The triggering conditions include at least one of the following: First, the operating context identifier changes (e.g., frequency band, mode, or power level switch); second, within K consecutive periods (K can be set to 20-100), the median SDI, PFI, or R of the reliable normal samples deviates from the historical baseline by more than a set proportion δ (δ can be set to 20%-40%), which is determined to be statistical drift; third, the warning proportion continuously exceeds the target interval within a continuous window.

[0122] The reference spectrum template update submodule is used to update only the reliable normal spectrum within the sample pool. And use an exponential moving average: λ can be set to 0.01-0.10. When drift is triggered and a new operating condition is confirmed, λ can be briefly increased to 0.10-0.30 to accelerate convergence. After stabilization, it should be restored to its normal value.

[0123] The threshold adaptive update submodule is used to update based on the set of credible normal risk coefficients in the sample pool. , It employs a quantile and robust dispersion reestimation strategy, entering a fast reestimation mode when drift is triggered, shortening the statistical window and prioritizing updates. To suppress the spread of false alarms, updates will be made only after the sample pool capacity has recovered to the set size. ,

[0124] The fusion weight adaptive update submodule is used to adjust α and β in small steps when the trigger condition is met, so that the warning ratio returns to the target range and the abnormal classification result is stable. The adjustment is constrained by being slow and unidirectionally restricted, and the single adjustment range can be limited to 0.02-0.05 to prevent discrimination drift caused by weight oscillation.

[0125] Preferably, the adaptive update module receives external confirmation information (annotations from maintenance personnel or verification results from the upper-level system) as the basis for correction. When the external confirmation and the system's judgment are inconsistent for a long period of time, threshold re-estimation is triggered first rather than template update to avoid incorrect corrections being written into the baseline.

[0126] The result output and traceability module is used to output structured test results and establish a verifiable chain of evidence to meet the needs of equipment maintenance, quality traceability, and operation auditing.

[0127] The result output and traceability module includes a result encapsulation submodule, an alarm linkage submodule, and a data archiving submodule.

[0128] The result encapsulation submodule is used to output a result data packet that includes at least the period number, timestamp, SDI, PFI, R, discrimination result, anomaly type, confidence level, operating condition identifier, and parameter version number.

[0129] The alarm linkage submodule is used to trigger alarms under abnormal or high-confidence warning conditions and output handling suggestions. The handling suggestions include at least one or more of the following: recalibration suggestions, power supply check suggestions, power amplifier thermal management check suggestions, and antenna matching check suggestions. The suggestions are automatically prioritized based on the spectrum contribution item / power contribution item.

[0130] The data archiving submodule is used to archive key features, discrimination results, parameter versions, and necessary raw data summaries. It supports review and comparison by periodic number and parameter version, thereby ensuring that the detection results of the same device at different stages are comparable and interpretable.

[0131] During initial system deployment, maintenance personnel configure relevant operating condition indicators (such as frequency band, working mode, power level, etc.) for the electromagnetic wave transmitting equipment, and set the monitoring period T (0.5–5s) and sampling parameters (N number of time-domain signal sampling points). Simultaneously, the integrated status acquisition box is installed near the transmitting equipment and wiring is completed, connecting the power supply terminal to the power supply line and the power signal input terminal to the equipment's power monitoring signal. A communication connection is established between the communication port and the signal acquisition and synchronization module. The signal acquisition and synchronization module acquires the time-domain data x[n] of the transmitted signal in each monitoring period and collects power data p[m] and operating status information such as temperature through the integrated status acquisition box. Subsequently, the data is time-aligned and denoised, normalized, and the effective transmission segment is extracted using the preprocessing module. During the initialization phase, the system generates an initial reference spectrum template through the reference spectrum template management submodule and generates an initial threshold based on the collected reliable normal samples. To facilitate subsequent normal monitoring and anomaly detection, during normal system operation, the spectrum feature extraction module performs Discrete Fourier Transform on the acquired signals to calculate the Spectral Distortion Index (SDI), and the power feature extraction module calculates the Power Fluctuation Index (PFI) based on the power sequence output by the integrated status acquisition box. These features are then fused by the risk fusion assessment module to generate an anomaly risk coefficient (R). The anomaly discrimination module compares the risk coefficient with a preset threshold, outputs whether it is an anomaly, and calculates the corresponding confidence level. The anomaly classification module classifies anomalies into random, unstable, or persistent anomalies based on their duration and fluctuation characteristics. If the system detects the following situations... One mechanism is adaptive update triggering: if the SDI or PFI offset from the baseline exceeds a set threshold within multiple consecutive monitoring periods, or if the operating condition indicator changes, an update is triggered. After this, the baseline spectrum template is updated based on reliable normal samples, the threshold is recalculated to adapt to the current operating state, and the fusion weights (α and β) are appropriately adjusted to ensure stable system operation. Finally, the system outputs the detection results, including SDI, PFI, risk coefficient R, anomaly type, and confidence level, and provides corresponding handling suggestions. All detection results and parameter versions, along with the period number information output by the integrated status acquisition box, are archived to ensure that the detection process and system updates are traceable and verifiable.

[0132] The goal is to achieve the ability to simultaneously collect electromagnetic wave emission signals and power data during the monitoring period, extract the spectral distortion index and power fluctuation index and fuse them to generate a risk coefficient, realize anomaly discrimination and random / unstable / persistent classification, and adaptively update the benchmark template, threshold and weight under the constraint of credible normal samples, so as to achieve the purpose of long-term stable and traceable electromagnetic wave signal anomaly detection and classification.

[0133] Other similar embodiments of the invention will readily occur to those skilled in the art upon consideration of the specification and practice of the invention disclosed herein. This application is intended to cover any variations or adaptations of the invention that follow the general principles of the invention and include common knowledge or customary techniques in the art that are not disclosed herein.

[0134] The above embodiments are preferred embodiments of the present invention. Due to space limitations, the applicant has not used other embodiments, but this is not intended to limit the scope of the present invention. Any person skilled in the art can make some modifications without departing from the scope of the present invention; that is, all equivalent modifications made in accordance with the present invention should be covered by the scope of the present invention.

Claims

1. An adaptive electromagnetic wave signal anomaly detection and classification system, comprising a signal acquisition and synchronization module, a preprocessing module, a spectrum feature extraction module, a power feature extraction module, a risk fusion assessment module, an anomaly discrimination module, an anomaly classification module, an adaptive update module, and a result output and traceability module, characterized in that: The signal acquisition and synchronization module is used to acquire raw observations for anomaly detection within the monitoring period and complete multi-source time alignment, providing consistent input for subsequent feature construction. The preprocessing module is used to normalize the acquired data and extract effective segments, reducing the disturbance of noise, spikes, and missing data to feature calculation. The spectrum feature extraction module is used to convert the time-domain signal to the frequency domain and construct the spectral distortion index (SDI) to reflect the degree of deviation of the spectrum shape from the reference template. The power feature extraction module is used to construct the power fluctuation index (PFI) from the power sequence to reflect the degree of deviation of the transmit power stability. The risk fusion assessment module is used to fuse the SDI and PFI and output a unified anomaly. The risk coefficient R is used to form a single discriminant and reduce the uncertainty of multi-indicator decision-making. The anomaly discrimination module is used to output the discrimination result for each monitoring period based on the risk coefficient and threshold set, and generate confidence levels related to alarm linkage. The anomaly classification module is used to classify the anomaly pattern after the anomaly or warning occurs, and output anomaly type labels that can be used for differentiated handling. The adaptive update module is used to update the template, threshold and weight online without introducing anomaly contamination, forming a detection-classification-self-calibration closed loop. The result output and traceability module is used to output structured detection results and establish a verifiable evidence chain to meet the needs of equipment maintenance, quality traceability and operation audit.

2. The adaptive electromagnetic wave signal anomaly detection and classification system according to claim 1, characterized in that... The signal acquisition and synchronization module includes a time-domain signal acquisition submodule, a power acquisition submodule, an operating context acquisition submodule, and a period identification and synchronization submodule. The time-domain signal acquisition submodule is used to acquire the time-domain sampling sequence x[n] of the transmitted signal within the monitoring period T and record the sampling rate. The number of sampling points N, channel number, and timestamp are specified. The monitoring period T can be set to 0.5–5s, and N can be set to 2048–65536 points to meet the spectral resolution requirements of devices in different frequency bands. The power acquisition submodule is used to acquire the transmit power sequence p[m] within the same monitoring period and record the sampling period and timestamp. Power sampling can be provided by a power meter, power amplifier sampling, standing wave ratio detection conversion, or the device's built-in sensor channel. The operation context acquisition submodule is used to acquire and determine operation information related to stability. The operation information includes at least one or a combination of the following: operating frequency band identifier, operating mode identifier, power level identifier, temperature, power supply status, or standing wave ratio, for subsequent operating condition identification and adaptive triggering. The period identifier and synchronization submodule aligns x[n], p[m], and operation information in time and generates a period number and data packet identifier for each monitoring period for one-to-one correspondence with the detection results and for traceability review. The signal acquisition and synchronization module also includes an integrated status acquisition box, which includes an acquisition control board, a mounting base, and a temperature sensor. The system comprises a data acquisition box housing, a communication connection port, a power terminal, and a power signal input terminal. A data acquisition control board is fixedly mounted on the top surface of a mounting base. A temperature sensor is fixedly mounted on the mounting base and forms thermal contact with it. The temperature sensor is electrically connected to the data acquisition control board. A snap-fit ​​groove is circumferentially formed along the upper surface edge of the mounting base. The bottom end of the data acquisition box housing is open and has a snap-fit ​​flange that mates with the snap-fit ​​groove. The snap-fit ​​flange is embedded in the snap-fit ​​groove to achieve positioning and snap-fit ​​between the data acquisition box housing and the mounting base. The mounting base and the data acquisition box housing are also detachably fixed together by fixing screws, forming a receiving cavity. The data acquisition control board and the temperature sensor are housed within this cavity. A communication connection port, a power terminal, and a power signal input terminal are correspondingly provided on one side of the data acquisition box housing. These terminals are electrically connected to the data acquisition control board. Preferably, the data acquisition control board integrates signal conditioning and sampling circuitry and further integrates a clock unit. The data acquisition box housing is a metal shielded housing or a flame-retardant insulating housing.

3. The adaptive electromagnetic wave signal anomaly detection and classification system according to claim 1, characterized in that... The preprocessing module includes a time-domain normalization submodule, a power sequence cleaning submodule, and an effective transmit segment extraction submodule. The time-domain normalization submodule is used to perform DC removal, windowing, and amplitude normalization on x[n], and can set bandpass / notch filtering according to the device frequency band to suppress non-operating frequency band interference. The power sequence cleaning submodule is used to suppress outliers, remove missing segments, or interpolate p[m], and output an effective power sequence that can participate in statistics. The effective transmit segment extraction submodule is used to identify the effective transmit segment in pulse / gated operation mode, and only x[n] and p[m] within the effective transmit segment participate in the subsequent SDI and PFI calculations, thereby avoiding the dilution of abnormal features in the idle segment.

4. The adaptive electromagnetic wave signal anomaly detection and classification system according to claim 1, characterized in that... The spectrum feature extraction module includes a frequency domain transformation submodule, a reference spectrum template management submodule, and a spectrum distortion index calculation submodule. The frequency domain transformation submodule performs a discrete Fourier transform on the preprocessed x[n] to obtain the spectrum X[k], and outputs the amplitude or energy distribution within a set frequency band Ω. The frequency band Ω is configured in association with the working frequency band identifier to ensure that the deviation measurement focuses on the effective transmission frequency band of the device. The reference spectrum template management submodule is used to maintain the reference spectrum template under normal operating conditions. The template is then associated with the operating frequency band identifier, operating mode identifier, and power level identifier. The reference spectrum template management submodule performs an initialization process after the system is first deployed or the equipment is replaced: during the installation and commissioning phase, M stable monitoring cycles (M can be set to 50–200) are collected. When all these cycles meet the low-risk condition, their spectra are aggregated to generate the initial spectrum. If the conditions are not met, the acquisition process is extended until the initialization constraints are met to avoid writing abnormal states into the template. The spectral distortion index calculation submodule is used to calculate... and The deviation is used to generate SDI, which employs an energy-normalized error metric: ε is a constant to prevent the denominator from being zero. The spectral distortion index calculation submodule can set weights for different subbands within Ω, making the system more sensitive to key frequency points, sideband structures, or modulation characteristics. However, the weights remain fixed under the same operating conditions to ensure comparability.

5. The adaptive electromagnetic wave signal anomaly detection and classification system according to claim 1, characterized in that... The power feature extraction module includes a power statistics calculation submodule, a deviation set construction submodule, and a power fluctuation index calculation submodule. The power statistics calculation submodule is used to calculate the mean of the power sequence within the monitoring period. with standard deviation The deviation set construction submodule is used to base on Construct a set of deviations and extract the maximum deviation. , minimum deviation and peak-valley difference The power fluctuation index calculation submodule is used to normalize and combine the above statistics and generate the PFI: Where w1, w2, w3 are weights and satisfy ε is a zero-prevention constant. w1, w2, and w3 are fixed as equipment configuration parameters during the factory or commissioning phase and remain unchanged before the adaptive switching is triggered to ensure the statistical consistency of PFI. The risk fusion assessment module includes a fusion weight management submodule and a risk coefficient calculation submodule. The fusion weight management submodule provides α and β, satisfying α and β ≥ 0 and α + β = 1. The risk coefficient calculation submodule calculates the risk coefficient using the following formula: The initial values ​​of α and β are determined during the debugging phase (for example, α can be set to 0.4–0.7 and β can be set to 0.3–0.6), and remain unchanged until the triggering condition of the adaptive update module is met, so as to avoid false alarm drift caused by frequent weight swings.

6. The adaptive electromagnetic wave signal anomaly detection and classification system according to claim 1, characterized in that... The anomaly detection module includes a threshold initialization and management submodule, a detection execution submodule, and a confidence generation submodule. The threshold initialization and management submodule is used to maintain normal thresholds. With abnormal threshold (or a single threshold Th), and establish a correspondence with the frequency band / mode / power level. The threshold initialization and management submodule synchronously performs threshold initialization after the template initialization is completed: for the risk coefficient set in the initialization phase. Robust statistics are performed by setting thresholds using quantiles and robust dispersion. Desirable The 0.95-0.99 quantiles, Desirable The median is added to 2-4 times the MAD (Median Absolute Deviation) to control the initial false alarm rate while taking into account individual device differences. The discrimination execution submodule is used to compare R with the threshold and output the result. When a dual threshold strategy is adopted, it satisfies... Output abnormal cycle, satisfying Output a reliable normal cycle, satisfying The system outputs the early warning cycle and enters it into the observation queue. The confidence level generation submodule maps R to a confidence level C between 0 and 1. The confidence level is used by the upper-level system to set a tiered handling strategy, where C can be based on... The piecewise normalization implementation makes it closer to The early warning cycle has a higher priority, and the anomaly detection module outputs the spectral contribution term simultaneously with the output result. With power contribution items This is used to explain the source of the anomaly and facilitate maintenance and location.

7. The adaptive electromagnetic wave signal anomaly detection and classification system according to claim 1, characterized in that... The anomaly classification module includes an observation window management submodule, a statistical feature calculation submodule, and a type determination submodule. The observation window management submodule is used to construct a sliding observation window W and aggregate the risk coefficient sequence. The window length can be set to 10-60 monitoring cycles, and the update frequency is adjusted according to the alarm level. The statistical feature calculation submodule is used to calculate the percentage of anomalies within the window, the length of consecutive anomalies, the mean and standard deviation of R, and other statistical quantities, and form the basis for classification. The type determination submodule is used to output the anomaly type. The anomaly type includes at least random anomalies, unstable anomalies, and persistent anomalies. Random anomalies meet the following criteria: low percentage of anomalies and short length of consecutive anomalies (e.g., percentage of anomalies not higher than 0.1-0.2 and length of consecutive anomalies not exceeding 1-2 cycles). Unstable anomalies meet the following criteria: medium percentage of anomalies and significant fluctuations (e.g., percentage of anomalies between 0.2-0.6 and standard deviation higher than the set threshold). Persistent anomalies meet the following criteria: high percentage of anomalies or long length of consecutive anomalies (e.g., percentage of anomalies not lower than 0.6-0.8 or length of consecutive anomalies not less than 5-10 cycles). The anomaly classification module uses the same window statistical framework for the warning cycle and the anomaly cycle, so that the system can output the warning information of tending to be unstable / tending to be persistent in the early stage of anomalies, thereby intervening in advance.

8. The adaptive electromagnetic wave signal anomaly detection and classification system according to claim 1, characterized in that... The adaptive update module includes a trusted normal sample pool management submodule, a drift detection and triggering submodule, a baseline spectrum template update submodule, a threshold adaptive update submodule, and a fusion weight adaptive update submodule. The trusted normal sample pool management submodule writes trusted normal periods into the sample pool and manages them in groups by frequency band / mode / power level to avoid cross-condition mixing that could cause template distortion. The drift detection and triggering submodule determines when template / threshold / weight updates are allowed. The triggering condition includes at least one of the following: first, a change in the operating context identifier (e.g., switching between frequency band, mode, or power level); second, within K consecutive periods (K can be set to 20-100), the median SDI, PFI, or R of trusted normal samples deviates from the historical baseline by more than a set proportion δ (δ can be set to 20%-40%), which is considered statistical drift; third, the warning proportion continuously exceeds the target range within a continuous window. The baseline spectrum template update submodule uses only trusted normal spectrum from the sample pool for updates. And use an exponential moving average: λ can be set to 0.01-0.

10. When drift is triggered and a new working condition is confirmed, λ can be briefly increased to 0.10-0.30 to accelerate convergence. After stabilization, it returns to its normal value. The threshold adaptive update submodule is used to update the set of credible normal risk coefficients in the sample pool. , It employs a quantile and robust dispersion reestimation strategy, entering a fast reestimation mode when drift is triggered, shortening the statistical window and prioritizing updates. To suppress the spread of false alarms, updates will be made only after the sample pool capacity has recovered to the set size. The fusion weight adaptive update submodule is used to adjust α and β in small steps when the triggering condition is met, so that the warning ratio returns to the target range and the abnormal classification result is stable. The adjustment is constrained by being slow and unidirectionally limited, and the single adjustment range can be limited to 0.02-0.05 to prevent the judgment drift caused by weight oscillation. The adaptive update module receives external confirmation information (marked by maintenance personnel or reviewed by the upper system) as the basis for correction. When the external confirmation and the system judgment are inconsistent for a long time, the threshold re-evaluation is triggered first rather than the template update to avoid incorrect correction written to the benchmark.

9. The adaptive electromagnetic wave signal anomaly detection and classification system according to claim 1, characterized in that... The result output and traceability module includes a result encapsulation submodule, an alarm linkage submodule, and a data archiving submodule. The result encapsulation submodule is used to output a result data packet containing at least the period number, timestamp, SDI, PFI, R, discrimination result, anomaly type, confidence level, operating condition identifier, and parameter version number. The alarm linkage submodule is used to trigger an alarm under abnormal or high-confidence warning conditions and output handling suggestions. The handling suggestions include at least one or more of the following: recalibration suggestions, power supply inspection suggestions, power amplifier thermal management inspection suggestions, and antenna matching inspection suggestions. The suggestions are automatically prioritized based on the spectrum contribution item / power contribution item. The data archiving submodule is used to archive key features, discrimination results, parameter versions, and necessary original data summaries. It supports review and comparison by period number and parameter version, thereby ensuring that the detection results of the same device at different stages are comparable and interpretable.

10. The adaptive electromagnetic wave signal anomaly detection and classification system according to claim 1, characterized in that... The signal acquisition and synchronization module establishes signal interaction with the preprocessing module and the result output and traceability module. The preprocessing module establishes signal interaction with the spectrum feature extraction module and the power feature extraction module. The spectrum feature extraction module establishes signal interaction with the risk fusion assessment module and the adaptive update module. The power feature extraction module establishes signal interaction with the risk fusion assessment module. The risk fusion assessment module establishes signal interaction with the anomaly detection module and the adaptive update module. The anomaly detection module establishes signal interaction with the anomaly classification module, the adaptive update module, and the result output and traceability module. The anomaly classification module establishes signal interaction with the result output and traceability module. The adaptive update module establishes signal interaction with the result output and traceability module.

Images