A method for automatic verification and abnormal positioning of a communication engineering signal debugging process

Abstract

The application relates to the field of communication system automation debugging, and proposes a communication engineering signal debugging process automatic verification and abnormal positioning method. The method first constructs a communication system digital twin model, combines historical debugging data, generates an abnormal feature-cause mapping decision tree through standardized processing and association rule mining, and simultaneously establishes a strategy library of multiple sets of debugging parameter adjustment schemes. The response of a multi-dimensional test signal real-time acquisition system is utilized, time domain, frequency domain and modulation domain features are extracted, a pre-trained model and a decision tree are combined to infer and output candidate root causes, and the confidence is dynamically calibrated to quantify uncertainty. Then, the optimal parameter adjustment scheme is adaptively selected or generated, automatic correction and verification are performed, and the cycle is iterated until the debugging is completed. The whole process data is structured and recorded to incrementally optimize the knowledge base decision logic. The method realizes rapid positioning of communication system abnormalities, accurate optimization of parameters and self-evolution of knowledge, and improves the automation and intelligence level of debugging.

Description

Technical Field

[0001] This invention relates to the field of automated debugging and anomaly localization technology for communication systems, and in particular to an automated verification and anomaly localization method for the signal debugging process in communication engineering. Background Technology

[0002] With the increasing complexity and intelligence of communication systems, automated debugging and intelligent anomaly localization have become core requirements in the signal debugging process of communication engineering. Currently, the industry mainstream adopts automated testing, fault classification based on human experience, and a certain degree of data-driven anomaly diagnosis. The typical process includes injecting pre-set test signals, acquiring responses, comparing results, and using empirical rules or simple decision rules to achieve anomaly localization. For example, some systems use a hybrid approach of expert databases and decision trees to assist technicians in efficiently and quickly locating common anomalies. Simultaneously, with the introduction of machine learning, neural networks, and other methods, various pre-trained anomaly classification models are gradually being applied to automated diagnostic scenarios for communication hardware, protocols, and signal processing links.

[0003] In practical applications, most existing technologies focus on anomaly detection and preliminary root cause inference, with diagnostic results typically presented as static quantitative indicators such as probability scores and category confidence levels. These scores are mostly derived from static probability distributions based on historical samples during offline training, or generated based on fixed rules and empirical parameters, and are not dynamically updated with the real-time environment of the actual communication system, the current verification path, and historical feedback. When the anomaly detection accuracy fails to meet requirements, the system needs repeated manual debugging, each time consuming significant manpower and resources, making it difficult to achieve automatic closed-loop control and adaptive tuning.

[0004] This leads to the following main technical limitations: First, the confidence score for anomaly root cause identification is a static output, lacking a dynamic calibration mechanism for the inference path, the current verification context, and past feedback information; second, existing systems struggle to quantify and provide feedback on uncertainties in the inference process, and cannot make corresponding adjustments for dynamic factors such as model errors and signal drift in real-time environments, easily resulting in distorted confidence scores that can mislead the debugging process; third, there is a lack of an overall technical solution that integrates knowledge bases, closed-loop feedback, and environmental adaptability, resulting in a disconnect between debugging strategies and anomaly localization, making it difficult to guarantee stability and high efficiency in engineering applications. Summary of the Invention

[0005] This application provides an automated verification and anomaly localization method for signal debugging in communication engineering, aiming to solve the problem mentioned in the background art of lacking a dynamic calibration mechanism for inference paths, current verification context and past feedback information.

[0006] This application provides an automated verification and anomaly localization method for signal debugging in communication engineering, specifically including:

[0007] S1: Obtain the digital twin model of the communication system to be debugged and historical debugging case data, construct an abnormal feature-root cause mapping decision tree based on the abnormal features and root cause mapping relationship in the historical debugging case data, and generate a debugging strategy library based on the expert experience library.

[0008] S2: Inject a preset multi-dimensional test signal sequence into the actual communication system and collect the output response signal in real time. Use the digital twin model to generate the corresponding theoretical expected response. Process the output response signal and the theoretical expected response to obtain multi-dimensional difference features.

[0009] S3: Input the multidimensional differential features into the abnormal root cause classification model, combine the abnormal feature-root cause mapping decision tree to perform layer-by-layer matching reasoning, and output at least one candidate abnormal root cause and its initial probability confidence.

[0010] S4: Perform dynamic calibration processing on the candidate anomaly root causes and their initial probability confidence scores to generate dynamically calibrated candidate anomaly root causes and dynamic confidence scores.

[0011] S5: Based on the dynamically calibrated candidate anomaly root causes and dynamic confidence scores, match the optimal debugging parameter adjustment scheme from the debugging strategy library, and automatically execute the optimal debugging parameter adjustment scheme in the actual communication system to complete the parameter correction.

[0012] S6: After the parameter correction is completed, a new round of automated verification process is automatically triggered. If the verification results show that the multidimensional difference features meet the preset pass threshold, a debugging completion report is generated. If not, steps S3-S6 are recursively executed until the preset iteration limit is reached.

[0013] The automated verification and anomaly localization method for signal debugging in communication engineering provided in this application has the following beneficial effects:

[0014] By integrating historical cases and expert experience through a digital twin model, an anomaly feature-root cause mapping decision tree and debugging strategy library are formed, enabling efficient reuse of prior knowledge and improving the accuracy and interpretability of anomaly localization. In the automated verification phase, multi-dimensional test signals are injected, and the differences between actual and expected responses in the time, frequency, and modulation domains are compared to achieve a comprehensive quantitative assessment of system behavior deviations. This overcomes the shortcomings of single-dimensional detection, which is prone to missed detections and misjudgments, significantly improving fault detection sensitivity and coverage. During anomaly localization, multi-dimensional differential features are input into the root cause classification model, combined with structured decision tree reasoning layer by layer. This not only improves the level of intelligence but also provides transparent support through a visualized decision path, solving the "black box" trust problem. The system adaptively matches or dynamically generates optimal debugging parameter schemes based on candidate root causes and their confidence levels and automatically executes closed-loop control, breaking through the bottleneck of manual trial and error and significantly shortening the debugging cycle. Each adjustment automatically triggers a new round of verification, forming a recursive optimization closed loop of "detection—localization—adjustment—re-verification," ensuring continuous convergence until performance indicators are met, and achieving dynamic calibration of the communication system. Attached Figure Description

[0015] Figure 1 This is the main flowchart of an automated verification and anomaly localization method for signal debugging in communication engineering.

[0016] Figure 2 This is a sub-flowchart of an automated verification and anomaly localization method for signal debugging in communication engineering.

[0017] Figure 3 This is another sub-flowchart of an automated verification and anomaly localization method for signal debugging in communication engineering. Detailed Implementation

[0018] Embodiments of the present invention are described in detail below, examples of which are shown in the accompanying drawings, wherein the same or similar reference numerals denote the same or similar elements or elements having the same or similar functions throughout. The embodiments described below with reference to the accompanying drawings are exemplary and are only used to explain the present invention, and should not be construed as limiting the present invention.

[0019] The following disclosure provides many different embodiments or examples for implementing different structures of the invention. To simplify the disclosure, specific examples of components and arrangements are described below. Of course, these are merely examples and are not intended to limit the invention. Furthermore, reference numerals and / or letters may be repeated in different examples; such repetition is for simplification and clarity and does not in itself indicate a relationship between the various embodiments and / or arrangements discussed.

[0020] like Figure 1As shown, this application provides an automated verification and anomaly localization method for the signal debugging process in communication engineering, specifically including:

[0021] S1: Obtain the digital twin model of the communication system to be debugged and historical debugging case data, construct an abnormal feature-root cause mapping decision tree based on the abnormal features and root cause mapping relationship in the historical debugging case data, and generate a debugging strategy library based on the expert experience library.

[0022] S2: Inject a preset multi-dimensional test signal sequence into the actual communication system and collect the output response signal in real time. Use the digital twin model to generate the corresponding theoretical expected response. Process the output response signal and the theoretical expected response to obtain multi-dimensional difference features.

[0023] S3: Input the multidimensional differential features into the abnormal root cause classification model, combine the abnormal feature-root cause mapping decision tree to perform layer-by-layer matching reasoning, and output at least one candidate abnormal root cause and its initial probability confidence.

[0024] S4: Perform dynamic calibration processing on the candidate anomaly root causes and their initial probability confidence scores to generate dynamically calibrated candidate anomaly root causes and dynamic confidence scores.

[0025] S5: Based on the dynamically calibrated candidate anomaly root causes and dynamic confidence scores, match the optimal debugging parameter adjustment scheme from the debugging strategy library, and automatically execute the optimal debugging parameter adjustment scheme in the actual communication system to complete the parameter correction.

[0026] S6: After the parameter correction is completed, a new round of automated verification process is automatically triggered. If the verification results show that the multidimensional difference features meet the preset pass threshold, a debugging completion report is generated. If not, steps S3-S6 are recursively executed until the preset iteration limit is reached.

[0027] Step S1: Obtain the digital twin model of the communication system to be debugged and historical debugging case data; construct an anomaly feature-root cause mapping decision tree based on the anomaly features and root cause mapping relationships in the historical debugging case data; and generate a debugging strategy library based on the expert experience base. Specifically, this includes:

[0028] S1.1: Obtain the digital twin model of the communication system to be debugged and historical debugging case data. Perform standardization and cleaning on the time-domain waveform distortion features, frequency-domain spectral line shift features and modulation domain constellation diagram divergence features in the historical debugging case data to generate a structured historical debugging case dataset containing multi-dimensional abnormal feature labels.

[0029] The system calls the data interface to obtain the digital twin model structure parameters of the communication system to be debugged, and reads the original log file of the historical debugging case in the storage medium. The original log file contains time-domain waveform sampling points, frequency-domain spectral line amplitude values ​​and modulation domain constellation diagram coordinate data.

[0030] Detrending processing is performed on the time-domain waveform sampling points, baseline drift interference is eliminated using a moving average filter, and waveform distortion feature vectors are extracted. The feature vectors include peak factor, impulse factor, and waveform factor.

[0031] Logarithmic transformation is performed on the amplitude values ​​of frequency domain spectral lines to calculate the centroid shift and sidelobe leakage index of the spectral lines, generating a frequency domain spectral line shift characteristic matrix to quantify the unexpected distribution of spectral energy.

[0032] Cluster analysis was performed on the coordinate data of the modulation domain constellation diagram to calculate the error vector magnitude (EVM) and phase deviation standard deviation of each constellation point relative to the ideal position, and a divergence feature set of the modulation domain constellation diagram was constructed.

[0033] The Z-Score standardization method is used to perform dimensionless processing on the above multidimensional features, eliminating the numerical differences between different physical dimensions and ensuring that the feature values ​​are within the same order of magnitude range.

[0034] The standardized time-domain, frequency-domain, and modulation-domain features are associated and mapped with the corresponding known fault root cause labels to form structured data records.

[0035] Abnormal records with missing values ​​exceeding a preset threshold are removed by data cleaning methods, and the integrity of the remaining data is verified to generate a structured historical debugging case dataset containing multi-dimensional abnormal feature labels.

[0036] By using the standardized cleaning process described above, the heterogeneous raw data from the previous step is transformed into a structured feature dataset in a unified format, achieving the expected technical effect of data consistency required for subsequent association rule mining.

[0037] For example, for historical debugging data of a 5G base station RF module, 1000 sets of time-domain waveform data with a sampling rate of 100MHz were read. A moving average filter with a window length of 50 points was used to remove the DC component, resulting in a mean peak factor of 3.2 and a pulse factor of 1.5. After performing an FFT transform on the frequency-domain data, the spectral centroid shift was calculated to be 0.5kHz, and the sidelobe leakage index was -40dB. Clustering of the QAM64 modulation constellation data yielded an average EVM of 2.5% and a phase deviation standard deviation of 1.2 degrees. Z-score normalization was used, setting the mean to 0 and the standard deviation to 1, mapping all feature values ​​to the [-3,3] interval. Fifty sets of records with data loss were removed, resulting in a structured historical debugging case dataset containing 950 valid records. Each record contains three-dimensional feature vectors in the time, frequency, and modulation domains, along with corresponding hardware fault or software configuration error labels, providing high-quality input for subsequent decision tree construction.

[0038] S1.2: Based on the structured historical debugging case dataset, the co-occurrence probability between multidimensional abnormal feature labels and the root causes of hardware failures and software configuration errors is analyzed using association rule mining methods, so as to construct an abnormal feature root cause mapping relationship table containing node confidence weights.

[0039] Receive the structured historical debugging case dataset output by S1.1. This dataset contains multi-dimensional abnormal feature labels such as time-domain waveform distortion, frequency-domain spectral line shift, and modulation domain constellation diagram divergence, as well as their corresponding hardware faults or software configuration error root cause identifiers.

[0040] The multidimensional abnormal feature labels in the structured historical debugging case dataset are discretized and encoded to divide the continuous signal difference features into several interval levels. Feature-root cause association records in transactional database format are constructed to ensure that each record contains a unique case ID, a set of abnormal feature items, and a definite root cause item.

[0041] The Apriori association rule mining method is used to traverse the transactional database, calculate the support between each multidimensional abnormal feature combination and a specific root cause, and filter out frequent itemsets that meet the minimum support threshold to eliminate random associations caused by low-frequency noise interference.

[0042] Based on the frequent itemsets, the confidence index of the combination of anomalous features pointing to a specific root cause is calculated. By statistically analyzing the proportion of cases where a certain type of anomalous feature and a certain type of root cause appear simultaneously to the total number of cases where that type of anomalous feature appears, the inference strength of the feature to the root cause is quantified.

[0043] The lift index is introduced to verify the validity of the calculated confidence level, evaluate the effect of the occurrence of anomalous features on the probability of root cause occurrence, eliminate invalid rules with a lift of less than 1, and ensure that the constructed mapping relationship is statistically significant rather than random coincidence.

[0044] The validated high-confidence association rules are transformed into node weight parameters, and an abnormal feature root cause mapping relationship table containing the antecedent feature set, consequent root cause identifier and corresponding confidence weight is constructed as the basis for the decision tree split attribute selection.

[0045] Through the above-mentioned association rule mining and statistical verification processing methods, the structured data in the previous step is transformed into a feature-root cause mapping relationship table with probability weight support, realizing the technical transformation from raw data to quantifiable reasoning basis, and providing a data foundation for the subsequent construction of a high-accuracy decision tree.

[0046] For example, the minimum support threshold is set to 0.05, the minimum confidence threshold to 0.8, and the minimum lift threshold to 1.2. In a communication base station debugging case library, it was found that when the feature combination of "time domain rising edge overshoot greater than 10%" and "frequency domain out-of-band spurious emissions higher than -40dBc" occurs simultaneously, the support pointing to the root cause "abnormal power amplifier bias voltage" is 0.12, the confidence is 0.92, and the lift is 3.5. This rule is retained and written into the mapping table, assigned an initial weight of 0.92. However, the combination of "constellation diagram phase rotation" and "filter cutoff frequency drift," although with a confidence of 0.85, has a lift of only 1.1 and is considered a weak correlation rule, thus being discarded. The final generated mapping table contains 50 high-confidence feature-root cause association rules, significantly improving the accuracy of subsequent decision tree splitting.

[0047] S1.3: Based on the abnormal feature root cause mapping relationship table, a recursive splitting method is used to take the high-confidence multidimensional abnormal feature labels as splitting attributes, and generate an abnormal feature root cause mapping decision tree with a clear reasoning path layer by layer, so as to form a logical judgment topology structure for subsequent root cause layer by layer matching.

[0048] Receive the root cause mapping table of abnormal features containing node confidence weights, and extract the information gain value of each multidimensional abnormal feature label in the table as the basis for splitting attribute selection.

[0049] Perform traversal calculations on the candidate split attribute set and evaluate the ability of each feature label to improve the purity of root cause classification based on the information entropy formula.

[0050] The feature label with the largest information gain is selected as the splitting attribute of the current decision tree node, and the dataset is divided into several child nodes.

[0051] The above splitting process is repeated recursively for each child node until the leaf node termination condition is met, including the number of node samples being less than a preset threshold or the information gain being less than the minimum splitting threshold.

[0052] The generated leaf nodes are labeled with the corresponding abnormal root cause type, and the cumulative confidence weights on the path are inherited to form a complete logical judgment topology.

[0053] By using a recursive splitting process, the mapping table from the previous step is transformed into an anomaly feature root cause mapping decision tree with a clear reasoning path, thereby achieving the expected technical effect of logical structuring and interpretability in root cause localization.

[0054] For example, the minimum information gain threshold for splitting is set to 0.05, and the minimum number of samples per node is 10. For 1000 historical case data points containing time-domain jitter and frequency-domain shift features, the information gain for time-domain jitter is calculated to be 0.42, and the information gain for frequency-domain shift is 0.38. Time-domain jitter is selected as the root node splitting attribute, dividing the data into two subsets: high jitter and low jitter. The information gain for frequency-domain shift is further calculated for the high-jitter subset, resulting in 0.25, which is then selected as the next-level splitting attribute. A decision tree of depth 3 is recursively generated, with leaf nodes corresponding to root causes such as amplifier nonlinear distortion and filter parameter drift, and path confidence weights of 0.92 and 0.85, respectively. This decision tree significantly improves the logical clarity of root cause localization and reduces invalid search paths.

[0055] S1.4: Call the preset expert experience library, extract the parameter adjustment rules for the root causes corresponding to the leaf nodes in the abnormal feature root cause mapping decision tree, and perform vectorized encoding processing on the parameter adjustment rules to generate a debugging strategy library containing multiple sets of debugging parameter adjustment schemes.

[0056] The expert experience database stores systematically collected, organized, and shared expert knowledge, cases, solutions, and practical experience in the field of communications.

[0057] Traverse the leaf nodes of the anomaly feature-root cause mapping decision tree, extract the hardware fault type or software configuration error identifier associated with each leaf node, and use it as the index key value of the parameter adjustment rule.

[0058] Retrieve the original debugging strategy text that matches the index key value from the preset expert experience base. The strategy text contains the name of the parameter to be adjusted, the direction of adjustment, and the step size suggestion.

[0059] Natural language parsing is performed on the retrieved original debugging strategy text to identify key parameter fields, including RF gain, filter cutoff frequency, phase offset compensation amount, and equalizer tap coefficient.

[0060] The identified key parameter fields are mapped to the register address space of the communication system's underlying control interface, and a correspondence table between parameter names and physical storage locations is established.

[0061] For continuous adjustment parameters, a linear normalization method is used to transform the qualitative descriptions in expert experience into quantitative numerical ranges, and to set the minimum adjustment step size and the maximum allowable deviation threshold.

[0062] For discrete adjustment parameters, an enumeration vector is constructed to represent the optional configuration states, such as modulation mode switching or clock source selection, and each state is assigned a unique binary code.

[0063] A parameter coupling coefficient matrix is ​​introduced to quantify the degree of mutual influence between different debugging parameters. For highly coupled parameter groups, a joint adjustment vector is defined instead of independent scalars.

[0064] By using a word embedding model to vectorize the semantic information in the parameter adjustment rules, a high-dimensional semantic vector representing the adjustment logic is generated to capture the implicit empirical constraints.

[0065] The quantitative numerical range, binary encoding, joint adjustment vector, and high-dimensional semantic vector are spliced ​​and fused to form a structured multi-dimensional debugging parameter adjustment scheme vector.

[0066] Attach metadata tags to each debugging parameter adjustment scheme vector, including applicable scenarios, risk levels, and expected convergence speed, to facilitate rapid filtering during subsequent strategy matching.

[0067] Through the above vectorized encoding process, unstructured expert experience is transformed into a computer-executable structured debugging strategy library, realizing automated connection from root cause localization to parameter correction, and significantly improving the accuracy and execution efficiency of debugging strategy generation.

[0068] S1.5: Index and associate the abnormal feature root cause mapping decision tree with the debugging strategy library to establish an automatic jump mechanism from the end point of the reasoning path to the specific debugging parameter adjustment scheme, so as to output a communication system debugging knowledge base instance with complete closed-loop capability.

[0069] Obtain the unique identifier of the leaf node of the abnormal feature-root cause mapping decision tree generated in S1.3, and the vector index of each parameter adjustment scheme in the debugging strategy library generated in S1.4. Perform a depth-first traversal on the leaf nodes of the decision tree to extract the complete path sequence from the root node to the current leaf node. This sequence includes the threshold range of all split attributes along the way and the logical judgment direction. Convert the extracted path sequence into a fixed-length binary hash code as the unique path fingerprint of the specific fault reasoning path. Read the applicable root cause label corresponding to each strategy in the debugging strategy library and calculate the semantic similarity between the root cause label and the root cause pointed to by the leaf node of the decision tree. When the semantic similarity exceeds the preset matching threshold, establish a bidirectional mapping pointer between the leaf node path fingerprint and the strategy library index. Construct a fast retrieval structure based on a hash table, using the path fingerprint as the key and the associated strategy index list as the storage object, to achieve strategy positioning with O(1) time complexity. Perform consistency verification on the established mapping relationship to ensure that each leaf node is associated with at least one valid debugging strategy. Leaf nodes without associated strategies are marked as awaiting expert intervention. Through the above index association and binding process, the discrete decision tree structure and strategy library generated in the previous step are transformed into a communication system debugging knowledge base instance with automatic jump capability, realizing a seamless connection from the root cause reasoning endpoint to the parameter correction action, and significantly improving the automated execution efficiency of the debugging closed loop.

[0070] Step S2: Inject a preset multi-dimensional test signal sequence into the actual communication system and collect the output response signal in real time. Use the digital twin model to generate the corresponding theoretical expected response. Process the output response signal and the theoretical expected response to obtain multi-dimensional difference features. Specifically, this includes:

[0071] S2.1: Based on a predefined communication protocol standard, obtain multi-dimensional test signal sequence configuration parameters, use an arbitrary waveform generator to perform digital-to-analog conversion processing on the configuration parameters to generate an analog excitation signal, and inject the analog excitation signal into the input port of the communication system to be debugged to obtain the real-time physical excitation state.

[0072] The S2.1 sub-step aims to generate a high-fidelity analog excitation signal based on a predefined communication protocol standard, providing an accurate physical input benchmark for subsequent multidimensional differential feature extraction.

[0073] The predefined communication protocol standard file is parsed to extract key configuration parameters such as carrier frequency, modulation method, symbol rate and pulse shaping filter coefficients, and to construct a digital descriptor for a multi-dimensional test signal sequence.

[0074] Based on digital descriptors, a baseband IQ data sequence is constructed in the memory of an arbitrary waveform generator. Then, a direct digital synthesis technique is used to convert discrete digital sampling points into analog voltage signals in the continuous time domain.

[0075] The generated analog excitation signal is subjected to linearity calibration and amplitude normalization to compensate for nonlinear distortion in the digital-to-analog conversion process and ensure that the dynamic range of the injected signal meets the receiving sensitivity requirements of the communication system under test.

[0076] The calibrated analog excitation signal is injected into the input port of the communication system under test through the radio frequency coupling interface to establish the real-time excitation state of the physical link and synchronously trigger the sampling clock of the high-speed data acquisition card to establish the time domain reference zero point.

[0077] Through the above-mentioned analog-to-digital conversion and physical injection processing, the abstract configuration parameters are transformed into physical excitation signals with definite time-frequency characteristics, realizing standardized excitation of the input end of the system to be debugged, laying the foundation for obtaining highly consistent output response signals in the future.

[0078] For example, in a 5G NR communication base station debugging scenario, the 3GPP TS 38.104 protocol standard is analyzed, with the carrier frequency set to 3.5GHz, subcarrier spacing to 30kHz, and modulation scheme to 64QAM. An arbitrary waveform generator generates baseband IQ data containing 100 resource blocks, with a sampling rate set to 245.76Msps. After digital-to-analog conversion, an analog excitation signal with a peak power of -10dBm is output, and the linearity error is controlled within 0.5dB. The signal is injected into the base station's RF input terminal through a directional coupler, while a TTL trigger signal is sent to the data acquisition card to ensure that the time alignment error between excitation and acquisition is less than 1ns, thereby guaranteeing the high-precision comparability between subsequent digital twin model simulation and actual physical response.

[0079] S2.2: Based on the real-time physical excitation state, a high-speed data acquisition card is used to synchronously sample the output port of the actual communication system to obtain the original output response signal, and the original output response signal is subjected to denoising and normalization preprocessing to obtain the output response signal.

[0080] S2.3: Based on the configuration parameters of the multi-dimensional test signal sequence, drive the digital twin model of the communication system under debugging to perform parallel simulation and deduction to generate the theoretical expected response, and perform the same time-frequency alignment processing on the theoretical expected response as the output response signal to obtain the aligned theoretical expected response signal.

[0081] The configuration parameters of the multi-dimensional test signal sequence generated by S2.1 and the output response signal obtained by S2.2 are received as input boundary conditions for the parallel simulation of the digital twin model.

[0082] The kernel of the digital twin model of the communication system to be debugged is invoked, the configuration parameters are mapped as simulation excitation sources, and a parallel simulation process synchronized with the physical system is started.

[0083] Based on the channel impulse response model and the nonlinear characteristic equation of the radio frequency front end, the propagation loss, phase rotation and amplitude distortion of the signal in the virtual link are calculated to generate the original theoretical expected response flow.

[0084] The timestamp reference and carrier frequency offset of the output response signal are obtained, and a time-frequency aligned reference frame is constructed to correct the spatiotemporal deviation of the theoretically expected response.

[0085] Interpolation resampling is performed on the original theoretical expected response to make its sampling rate consistent with the sampling rate of the output response signal, thus eliminating sample point misalignment caused by clock asynchrony.

[0086] A cross-correlation algorithm is used to calculate the time delay estimate between the theoretical expected response signal and the output response signal. Based on this time delay estimate, a cyclic shift operation is performed on the theoretical expected response to achieve precise time domain alignment.

[0087] Based on the aligned time-domain signal, the instantaneous phase error is extracted using the phase-locked loop method, and the carrier frequency deviation is compensated by digital down-conversion technology to complete the dual calibration of frequency domain phase and amplitude.

[0088] The output is a theoretically expected response signal that has been strictly aligned with time and frequency, ensuring that it is point-to-point comparable with the output response signal on the time and frequency axes.

[0089] By using the time-frequency alignment processing method described above, the original simulation data generated in the previous step is transformed into an alignment theoretical expected response signal with high spatiotemporal consistency, thereby achieving the expected technical effect of benchmark unification for subsequent multidimensional difference feature extraction.

[0090] For example, in a 5G NR base station signal debugging scenario, the configuration parameters include a center frequency of 3.5GHz, a bandwidth of 100MHz, and a subcarrier spacing of 30kHz. The digital twin model is built based on the 3GPP TS 38 series protocol and includes a power amplifier memory effect model. The original theoretical response data sampling rate is 245.76Msps, and the actual acquired signal, after resampling, is also 245.76Msps. Cross-correlation calculation yields a time delay estimate of 12 sampling points, which is then cyclically shifted right by 12 bits. The phase-locked loop bandwidth is set to 100kHz to track the remaining frequency offset and compensate for the phase rotation angle. The final output aligned theoretical expected response signal has a time-domain alignment error of less than 0.1ns with the actual signal, and the residual frequency offset is less than 1Hz, significantly improving the signal-to-noise ratio and fault location accuracy in subsequent residual calculations.

[0091] S2.4: Based on the output response signal and the alignment theory expected response signal, perform point-by-point subtraction operation using the difference method to obtain the time-domain residual sequence, and perform fast Fourier transform processing on the time-domain residual sequence to obtain the frequency-domain difference spectrum.

[0092] The system receives the output response signal and the theoretically expected response signal, constructs a time-domain difference calculation matrix, and performs strict alignment sampling point matching on the time axis for the two signals.

[0093] Perform point-by-point subtraction to calculate the difference between the actual response value and the theoretical expected value for each sampling time t, and generate the original time-domain residual sequence.

[0094] The original time-domain residual sequence is weighted using the Hanning window function to suppress spectral leakage and ensure the energy concentration of the subsequent frequency domain transformation.

[0095] Perform a Fast Fourier Transform (FFT) on the weighted time-domain residual sequence to convert the time-domain discrete signal into frequency-domain complex spectral data.

[0096] The amplitude spectrum of the frequency domain difference spectral lines is calculated, and the energy deviation values ​​of each frequency component are extracted to form a frequency domain feature vector characterizing the abnormal frequency response of the system.

[0097] By using the aforementioned differential method and spectral analysis processing, the alignment signal from the previous step is transformed into a time-domain residual sequence and a frequency-domain difference spectrum, achieving multi-dimensional feature decoupling from time-domain waveform deviation to frequency-domain energy distribution deviation, thus providing accurate spectral anomaly basis for subsequent modulation domain analysis and root cause localization.

[0098] S2.5: Based on the frequency domain difference spectral lines, the signal phase and amplitude information is demodulated using the constellation diagram mapping method to obtain the modulation domain error vector. The time domain residual sequence, the frequency domain difference spectral lines, and the modulation domain error vector are then combined for feature fusion processing to finally obtain multidimensional difference features.

[0099] The frequency domain difference spectral data output from step S2.4 is received and used as the input basis for modulation domain demodulation analysis. An inverse fast Fourier transform is performed on the frequency domain difference spectral data to recover the time-domain complex baseband signal sequence, preserving the complete amplitude and phase information of the signal. Based on the constellation diagram mapping rules defined by the communication protocol, an ideal constellation point reference set is constructed, including standard coordinate points for modulation formats such as QPSK, 16QAM, or 64QAM. Using the minimum Euclidean distance method, each sampling point in the recovered time-domain complex baseband signal is mapped to the nearest ideal constellation point, and the vector deviation between the actual sampling point and the ideal constellation point is calculated. This vector deviation is the modulation domain error vector, containing amplitude error components and phase error components. The following formula is used to calculate the... Error vector magnitude of each symbol :

[0100]

[0101] in, and These are the in-phase and quadrature components of the actual signal. and These are the in-phase and orthogonal components corresponding to the ideal constellation points. For all sign periods... Root mean square (RMS) statistics are performed to generate a modulation domain error vector feature matrix characterizing modulation quality. The time-domain residual sequence and frequency-domain difference spectrum generated in step S2.4 are obtained, and timestamp alignment and dimension matching are performed on all three. A weighted concatenation strategy is used to fuse the statistical features of the time-domain residual sequence, the peak offset of the frequency-domain difference spectrum, and the modulation domain error vector feature matrix into a unified high-dimensional feature vector. Through multi-domain feature fusion, the independent difference data from the previous step are transformed into multi-dimensional difference features, achieving a comprehensive quantitative description of the abnormal state of the communication system and providing highly discriminative input data for subsequent root cause classification models.

[0102] like Figure 2 As shown, step S3 involves inputting the multidimensional differential features into the anomaly root cause classification model, combining the anomaly feature-root cause mapping decision tree for layer-by-layer matching inference, and outputting at least one candidate anomaly root cause and its initial probability confidence. Specifically, this includes:

[0103] S3.1: Perform vector concatenation and normalization on multidimensional differential features, and construct a unified high-dimensional fault feature vector based on feature fusion algorithm to obtain standardized model input data.

[0104] S3.2: Input the standardized model input data into the pre-trained abnormal root cause classification model, and perform feature mapping calculation based on the forward propagation mechanism of the deep neural network to output an initial set of fault types containing multiple potential fault types and their corresponding original probability distributions.

[0105] The system receives the standardized high-dimensional fault feature vector generated by S3.1 and loads it as an input tensor into a pre-trained anomaly root cause classification model. This model employs a deep neural network architecture based on an attention mechanism, comprising an input layer, multiple hidden layers, and an output layer.

[0106] The input tensor is mapped to a high-dimensional latent space through a fully connected layer. The nonlinear transformation capability is introduced by using the modified linear unit activation function to extract the deep coupling features between the time-domain residual, frequency-domain spectral lines and modulation-domain error vector.

[0107] A multi-head self-attention mechanism is introduced into the hidden layer to calculate the correlation weight matrix between each dimension within the feature vector, dynamically enhancing the response strength to key fault-sensitive features and suppressing feature redundancy caused by noise interference.

[0108] The softmax activation function is used to normalize the neuron node values ​​in the output layer, transforming the original logits output of the neural network into a probability distribution. The specific calculation process follows the formula below:

[0109]

[0110] in, Indicates the first The original probability confidence of the potential fault type For the output layer The weighted sum of the values ​​of the neurons. For the output layer The weighted sum of the values ​​of the neurons. This represents the total number of preset potential fault types.

[0111] Traverse all output nodes, filter out fault types with probability values ​​greater than the preset threshold of 0.05, form an initial fault type set, and retain the corresponding original probability distribution data.

[0112] Through the aforementioned deep feature mapping and probability normalization processes, the abstract high-dimensional differential features are transformed into a concrete set of initial fault types and their original probability distributions, realizing a preliminary mapping from signal features to the fault semantic space, and providing a data foundation for subsequent decision tree logic verification.

[0113] S3.3: Use the abnormal feature-root cause mapping decision tree to perform logical verification on the initial set of fault types. Based on the node splitting rules of the decision tree, perform layer-by-layer feature matching and path search to filter out the candidate abnormal root cause list that meets the logical constraints and remove contradictory items.

[0114] It receives the initial set of fault types and their original probability distributions output by the pre-trained anomaly root cause classification model, and loads the topological structure data of the pre-built anomaly feature-root cause mapping decision tree.

[0115] We analyze the splitting attribute definition and threshold conditions of each non-leaf node in the decision tree and establish a mapping index table from the multidimensional differential feature space to the path of the decision tree node.

[0116] Iterate through each potential fault type in the initial fault type set, locate the corresponding leaf node in the decision tree, and backtrack to generate a complete reasoning path sequence from the root node to that leaf node.

[0117] Extract the multidimensional difference feature vector of the current input, and perform feature matching and verification operations for each intermediate node in the inference path sequence.

[0118] The logical consistency test method is used to determine whether the currently observed multidimensional difference feature values ​​meet the splitting threshold constraints set by each node on the path.

[0119] For any given node, if the actual feature value falls outside the abnormal range defined for that node, then the reasoning path is determined to have a logical contradiction, and the corresponding potential fault type is marked as invalid.

[0120] The path matching score is calculated based on the distance metric between the feature value and the node threshold boundary, quantifying the strength of the current feature's support for a specific inference path.

[0121] Remove all potential fault types marked as invalid, and retain fault types that pass logical verification and have a path confidence weight higher than a preset threshold as candidate root causes of anomalies.

[0122] The retained list of candidate anomaly root causes is deduplicated, and duplicate items that point to the same physical root cause but are derived through different feature paths are merged, and their path confidence weights are accumulated.

[0123] By using logical verification and path search processing, the initial set of fault types from the previous step is transformed into a list of candidate anomaly root causes that conform to the physical constraints of the system, thereby achieving the expected technical effect of eliminating false alarms and logical contradictions.

[0124] For example, assuming the initial fault type set includes "power amplifier linearity distortion" and "local oscillator phase noise", the root node of the decision tree splits based on a frequency domain spectral offset greater than 5dB. The current measured frequency domain offset is 3dB, which does not meet the root node splitting condition; therefore, the "power amplifier linearity distortion" path is judged as logically contradictory and eliminated. The "local oscillator phase noise" path, however, splits based on a modulation domain error vector amplitude greater than 0.05, with a measured value of 0.08, satisfying the condition. The calculated path confidence weight for "local oscillator phase noise" is 0.85, and the final output is a candidate anomaly root cause list containing only "local oscillator phase noise," significantly improving the logical rigor of root cause localization.

[0125] S3.4: For each item in the candidate anomaly root cause list, perform weighted fusion processing with the original probability distribution, and perform probability correction operation based on the Bayesian inference framework to generate candidate anomaly root causes and their confidence scores carrying initial probability confidence.

[0126] Receive the list of candidate anomaly root causes and their corresponding original probability distributions, and extract the split attribute confidence weights of each node in the decision tree inference path.

[0127] For each candidate root cause, its original output probability in the deep neural network is used as the likelihood evidence, and the historical statistical frequency of the decision tree path nodes is used as the prior probability.

[0128] Based on the Bayesian inference framework, a probability correction model is constructed to calculate the posterior probability in order to integrate data-driven and rule-driven dual-reset reliability information.

[0129] The calculated posterior probabilities are normalized to ensure that the sum of the confidence scores of all candidate root causes is 1, thus eliminating the influence of dimensional differences.

[0130] The candidate root causes are sorted in descending order based on the normalized posterior probabilities to generate a list of candidate abnormal root causes carrying initial probability confidence levels.

[0131] By using Bayesian probabilistic fusion processing, the original probabilities and logical constraint weights from the previous step are transformed into interpretable initial confidence scores, achieving the expected technical effect of complementary multi-source information.

[0132] For example, in a scenario of abnormal signal frequency offset at a base station, the deep learning model outputs an initial probability of 0.7 for root cause A (clock synchronization failure) and 0.3 for root cause B (filter aging). The decision tree path shows that the prior weight of root cause A is 0.6 and the prior weight of root cause B is 0.4. Substituting into the formula, the unnormalized posterior value of root cause A is 0.7 × 0.6 = 0.42, and that of root cause B is 0.3 × 0.4 = 0.12. With a normalized denominator of 0.54, the initial confidence of root cause A is finally corrected to 0.778, and that of root cause B is corrected to 0.222. This process significantly improves the ranking priority of root causes with high logical matching degree and effectively suppresses occasional misjudgments in a purely data-driven model.

[0133] like Figure 3 As shown, step S4 involves performing dynamic calibration on the candidate anomaly root causes and their initial probability confidence scores to generate dynamically calibrated candidate anomaly root causes and dynamic confidence scores. Specifically, this includes:

[0134] S4.1: Obtain the candidate anomaly root cause list and its corresponding initial probability confidence from the current anomaly root cause classification model, and extract the split attribute sequence of the decision tree node in the current inference path as context information input. Use a sequence encoding algorithm to perform vectorization mapping processing on the split attribute sequence to generate an inference path context vector representing the current inference logic topology.

[0135] The system receives the list of candidate anomaly root causes and their corresponding initial probability confidence scores from the anomaly root cause classification model, which serve as the basic data input for dynamic calibration.

[0136] Extract the split attribute sequence of the decision tree node in the current inference path. This sequence records the feature judgment logic from the root node to the leaf node, forming the context skeleton of the inference process.

[0137] The split attribute sequence is structured and parsed to identify the feature dimension identifier, threshold range and branch direction of each node, forming a discrete logical state symbol string.

[0138] The Word2Vec sequence encoding algorithm is used to perform vectorization mapping on the logical state symbol string, transforming discrete symbols into continuous high-dimensional dense vector representations.

[0139] Long Short-Term Memory (LSTM) networks are used to capture the dependencies between split attribute sequences and to calculate the contribution weight of each node in the sequence to the final root cause determination.

[0140] By aggregating weighted node vectors through an attention mechanism, a reasoning path context vector representing the current reasoning logic topology is generated.

[0141] By using sequence encoding and attention aggregation, the discrete reasoning path from the previous step is transformed into quantitative reasoning path context vector data, thereby realizing a digital representation of the reasoning logic context and providing standardized input for subsequent historical case similarity matching.

[0142] S4.2: Based on the inference path context vector, retrieve the verification feedback records in the historical debugging case data, filter out the historical verification sample set with feature similarity to the inference path context vector, and use the statistical distribution estimation method to fit the deviation distribution between the root cause confirmation result and the initial prediction result in the historical verification sample set to generate prior uncertainty distribution parameters characterizing the historical inference deviation characteristics.

[0143] The system receives the inference path context vector generated in the previous steps and uses it as the index key to retrieve historical debugging case data. Within the structured historical debugging case dataset, it calculates the cosine similarity between the current inference path context vector and the decision tree split attribute sequence vector in each historical record. A similarity threshold is set to filter records, and all records with similarities higher than the preset threshold are extracted to form a historical verification sample set. This sample set includes historical root cause confirmation result labels and corresponding initial prediction probability distribution data.

[0144] The initial predicted probabilities and the final confirmed root causes in the historical validation sample set are compared, and the deviation value between the two is calculated. If the predicted root cause and the confirmed root cause are consistent, the deviation value is zero; if they are inconsistent, a non-zero deviation weight is assigned according to the error type. The deviation values ​​of all samples are collected to form a deviation dataset.

[0145] A kernel density estimation method is used to fit the statistical distribution of the bias dataset. A Gaussian kernel function is selected as the smoothing kernel, and the optimal bandwidth parameter is determined through cross-validation. The probability density function of the bias data is calculated to obtain the prior uncertainty distribution characterizing the historical inference bias. This distribution includes two key parameters: mean and variance, which reflect the systematic bias tendency and the degree of random fluctuation, respectively.

[0146] The mean of the prior uncertainty distribution is extracted as the bias correction offset, and the variance is extracted as the uncertainty metric. These two parameters are encapsulated into a prior uncertainty distribution parameter object. This object is used in subsequent steps for Bayesian updates based on real-time signal noise level data.

[0147] By using the statistical distribution estimation method described above, the discrete bias data of the historical verification sample set is transformed into continuous prior uncertainty distribution parameters, thereby realizing the quantitative modeling of the historical inference bias characteristics and providing a statistical basis for dynamic confidence calibration.

[0148] For example, a similarity threshold of 0.85 is set. In the cosine similarity calculation between the current inference path context vector and historical records, 120 highly similar records are selected to form a historical verification sample set. Analysis of the prediction bias of these 120 records reveals a mean bias of 0.05 and a standard deviation of 0.12. Using Gaussian kernel density estimation with a bandwidth parameter of 0.05, a prior uncertainty distribution is fitted. This distribution has a mean of 0.05 and a variance of 0.0144. These parameters indicate a slight positive bias tendency and some volatility in the model under similar inference paths. These parameters will be used for subsequent calibration of the current initial probability confidence level to eliminate systematic bias and quantify uncertainty risk.

[0149] S4.3: Receive the prior uncertainty distribution parameters and the real-time signal noise level data of the current communication system, construct a multidimensional uncertainty influence matrix including signal drift factor and model bias factor, and use the Bayesian update mechanism to perform probability redistribution operation on the initial probability confidence using the multidimensional uncertainty influence matrix as a likelihood function to generate an intermediate calibration probability confidence that integrates real-time environmental interference features.

[0150] The system receives the prior uncertainty distribution parameters output by S4.2 and the real-time acquired communication system signal noise level data as the input basis for dynamic confidence calibration.

[0151] Spectral analysis is performed on real-time signal noise level data to extract in-band noise power spectral density and out-of-band spurious components, calculate signal-to-noise ratio degradation index, and quantify the current environmental interference intensity.

[0152] A signal drift factor is constructed based on the signal-to-noise ratio degradation index. This factor characterizes the degree of shift of the feature vector in multidimensional space caused by environmental noise, reflecting the source of uncertainty in the measurement data.

[0153] By calling the historical simulation residual statistics library of the digital twin model, the mean and variance of the deviation between the model's predicted value and the actual value under the current operating conditions are calculated, and the model deviation factor is constructed to characterize the inherent uncertainty caused by the system modeling error.

[0154] A multidimensional uncertainty influence matrix is ​​constructed by performing a tensor product operation on the signal drift factor and the model bias factor. This matrix fully describes the combined influence weights of environmental interference and model defects on the root cause judgment at the current inference moment.

[0155] By using the Bayesian update mechanism, the multidimensional uncertainty influence matrix is ​​transformed into a likelihood function, which is used to correct the initial probability confidence generated by S3.4.

[0156] By performing probability redistribution operations and incorporating real-time environmental noise and model bias information into the probability distribution through the aforementioned Bayesian inference process, an intermediate calibration probability confidence level that incorporates real-time environmental interference features is generated. This achieves dynamic correction of the static confidence level and significantly improves the robustness of anomaly localization in complex noise environments.

[0157] S4.4: Calculate the information entropy value of each candidate abnormal root cause based on the intermediate calibration probability confidence level to quantify the dispersion of the single inference result, and perform weighted fusion processing in combination with the variance component of the prior uncertainty distribution parameter to generate an uncertainty quantification index characterizing the credible boundary of the current root cause localization result.

[0158] Receive the intermediate calibration probability confidence vector output from step S4.3 and the variance component data from the prior uncertainty distribution parameters generated in step S4.2. For each item in the candidate root cause list, extract its corresponding intermediate calibration probability value and construct a set of discrete probability distributions. Use the Shannon information entropy formula to calculate the information entropy value of each candidate root cause probability distribution to quantify the dispersion and disorder of a single inference result.

[0159] Obtain the variance component of the prior uncertainty distribution parameter, which characterizes the fluctuation range of prediction bias under historically similar inference paths. Normalize the calculated information entropy value and map it to the interval of 0 to 1 to obtain the entropy weight coefficient. Perform a weighted fusion operation between the entropy weight coefficient and the variance component of the prior uncertainty distribution parameter to generate a comprehensive uncertainty quantification index.

[0160]

[0161] in, As a quantitative indicator of uncertainty, The normalized information entropy value. For the variance components of the prior uncertainty distribution parameters, and These are preset weighting coefficients used to balance the impact of current inference dispersion and historical bias fluctuations.

[0162] By using the weighted fusion processing method described above, the probability distribution characteristics and historical statistical properties of the previous step are transformed into uncertainty quantitative indicators that characterize the credibility boundary of the current root cause localization result, thereby achieving accurate quantification of the uncertainty in the reasoning process and providing a basis for subsequent confidence correction.

[0163] S4.5: The uncertainty quantification index is used to perform normalization correction on the intermediate calibration probability confidence level, eliminating the probability overstatement caused by high uncertainty, and outputting the finally determined candidate anomaly root cause after dynamic calibration and the dynamic confidence score carrying uncertainty measurement information, thus completing the online quantification of uncertainty in the reasoning process and dynamic calibration of confidence.

[0164] The intermediate calibration probability confidence set and the corresponding uncertainty quantification index generated in the preceding steps are received as input data for dynamic correction.

[0165] A nonlinear normalized mapping function based on uncertainty penalty is constructed to transform the uncertainty quantification index into a confidence decay coefficient, so as to suppress the phenomenon of inflated probability in high uncertainty scenarios.

[0166] A sigmoid activation function is used to construct a decay factor calculation model to ensure confidence stability in the low uncertainty range and achieve rapid decay in the high uncertainty range.

[0167] The dynamic decay coefficient of the a-th candidate root cause is calculated using the following formula:

[0168]

[0169] in, For the first The decay coefficient of each candidate root cause This refers to the uncertainty quantification index output by the corresponding preceding steps. To control the attenuation steepness, This is the threshold for triggering uncertainty.

[0170] The calculated attenuation coefficient is used to weight and correct the intermediate calibration probability confidence level. Element-by-element multiplication is then performed to generate the initially corrected probability value.

[0171] Softmax normalization is performed on the initially corrected set of probability values ​​to eliminate the deviation of the probability sum caused by the decay operation and ensure that the probability distribution of all candidate root causes satisfies the axiomatic constraints.

[0172] The final dynamic confidence score is calculated using the following formula:

[0173]

[0174] in, The final dynamic confidence score for the a-th candidate root cause is given. This represents the intermediate calibration probability confidence level. This represents the total number of candidate root causes.

[0175] Set a confidence level filtering threshold to remove candidates whose final dynamic confidence scores are below the threshold, and retain the list of root causes with high confidence and their scores.

[0176] By using nonlinear decay and normalization correction methods, the intermediate probability results of the previous step are transformed into dynamic confidence scores with uncertainty measurement information, realizing online quantification of uncertainty in the reasoning process and dynamic calibration of confidence, which significantly reduces the risk of misjudgment caused by environmental noise or model bias.

[0177] Step S5: Based on the dynamically calibrated candidate anomaly root causes and dynamic confidence scores, the optimal debugging parameter adjustment scheme is matched from the debugging strategy library, and the optimal debugging parameter adjustment scheme is automatically executed in the actual communication system to complete the parameter correction. Specifically, this includes:

[0178] S5.1: Obtain the dynamically calibrated candidate anomaly root causes and dynamic confidence scores as input conditions, perform priority calculation on multiple candidate anomaly root causes based on the weighted ranking method, and output the target anomaly root cause identifier with the highest confidence weight to establish the decision benchmark for subsequent strategy matching.

[0179] Receive a dynamically calibrated list of candidate anomaly root causes and their corresponding dynamic confidence scores, and construct a weighted decision vector set containing root cause identifiers, confidence scores, and uncertainty quantification indicators.

[0180] Multidimensional feature normalization is performed on the weighted decision vector set to eliminate the influence of different dimensions on priority ranking and generate a standardized root cause evaluation matrix.

[0181] The uncertainty penalty factor for each candidate root cause is calculated based on the information entropy theory to obtain the final priority score.

[0182] The calculated final priority scores are sorted in descending order to generate a candidate root cause priority queue.

[0183] Extract the root cause identifier ranked first in the priority queue as the target anomaly root cause and establish a single decision criterion for subsequent strategy matching.

[0184] By using weighted ranking and uncertainty penalty mechanisms, multiple candidate root causes are transformed into a unique, highly credible target root cause identifier, enabling precise targeting of debugging decisions.

[0185] S5.2: Receive the target anomaly root cause identifier and associated dynamic confidence score as input conditions, and use the strategy library index mapping mechanism to perform multi-dimensional feature retrieval and similarity matching processing in the pre-built debugging strategy library to output an initial debugging strategy set containing basic debugging parameter adjustment vectors.

[0186] The system receives the target anomaly root cause identifier and associated dynamic confidence score, parses the root cause type encoding to locate the main index key value in the debugging strategy library. Based on the main index key value, it retrieves a pre-built debugging strategy library, extracting all candidate debugging strategy entries bound to that root cause type and their metadata information. The extracted candidate debugging strategy entries undergo multi-dimensional feature vectorization, mapping the parameter adjustment direction, magnitude level, and applicable scenario labels in the strategy description to high-dimensional strategy feature vectors. It calculates the weighted similarity between the dynamic confidence score and the strategy's historical success rate, using the cosine similarity method to evaluate the spatial distance between the current root cause context vector and the feature vectors of each candidate strategy. A similarity threshold is set to filter low-matching strategies, retaining a subset of strategies with similarity higher than the preset threshold as preliminary screening results. The strategies in the preliminary screening results are sorted in descending order according to their historical average convergence speed, prioritizing debugging parameter combinations with fast response characteristics. Combining the uncertainty component of the dynamic confidence score, a risk penalty correction is applied to the sorted strategy set, reducing the weight of high-risk strategies corresponding to high-uncertainty root causes. An initial set of debugging strategies is generated, containing adjustment vectors for basic debugging parameters. Each vector element corresponds to a specific register address offset and numerical modification step size. Through multi-dimensional feature retrieval and similarity matching, the abstract root cause identifier is transformed into a concrete set of executable parameter adjustment schemes, achieving a precise mapping from diagnostic conclusions to execution actions, significantly improving the targeting and effectiveness of strategy matching.

[0187] For example, the target anomaly root cause is identified as "power amplifier linearity distortion," with a dynamic confidence score of 0.85. The parsed encoding is located to the policy library index Key_PA_LIN. Five candidate policies are retrieved: S1 (predistortion coefficient fine-tuning), S2 (bias voltage adjustment), S3 (gain backoff), S4 (impedance matching optimization), and S5 (filter bandwidth switching). The policy metadata is mapped to a 5-dimensional feature vector, such as the S1 vector being [0.9, 0.1, 0.8, 0.2, 0.1]. The cosine similarity between the current context vector and S1-S5 is calculated, which are 0.92, 0.75, 0.60, 0.45, and 0.30, respectively. A threshold of 0.70 is set, retaining S1 and S2. Based on historical convergence speeds, S1 averages 3 iterations, and S2 averages 5 iterations, ranking S1 higher. Considering the low uncertainty corresponding to a confidence score of 0.85, S1 is not significantly penalized. The output initial set of debugging strategies includes the basic adjustment vector for S1: the predistortion polynomial coefficients a3 are increased by 0.05 and a5 are decreased by 0.02. This process ensures that high-precision fine-tuning strategies are prioritized under high confidence levels, avoiding blindly trying inefficient parameters and significantly improving debugging efficiency.

[0188] S5.3: Extract the basic debugging parameter adjustment vector and the real-time state context information of the current system from the initial debugging strategy set as input conditions, and apply the adaptive parameter optimization method to perform dynamic gain compensation and boundary constraint correction on the basic debugging parameter adjustment vector to generate the optimal debugging parameter adjustment scheme with environmental adaptability.

[0189] We analyze the real-time operational status data fed back from the digital twin model of the current communication system, including channel signal-to-noise ratio, temperature drift coefficient, and load fluctuation rate, and construct a real-time state context vector that characterizes the current physical environment.

[0190] Calculate the Euclidean distance between the real-time state context vector and the historical standard operating condition vector to quantify the degree of deviation between the current environment and the ideal debugging environment, and generate an environment deviation index.

[0191] Based on the environmental deviation index, a preset nonlinear gain mapping table is queried to obtain the dynamic compensation coefficient matrix for the adjustment vector of the basic debugging parameters. This matrix includes the amplitude scaling factor and the phase offset correction amount.

[0192] Perform a Hadamard product operation on the dynamic compensation coefficient matrix and the basic debugging parameter adjustment vector to obtain an intermediate parameter adjustment vector that has undergone preliminary environmental adaptability correction.

[0193] By introducing the hardware security threshold boundary specified by the communication protocol, a multi-dimensional parameter constraint hyperplane is constructed, and the intermediate parameter adjustment vector is projected into the legal parameter space using the projection gradient descent method.

[0194] Calculate the residual norm between the projected vector and the original base vector. If the residual exceeds the preset convergence threshold, iteratively perform projection correction until the boundary constraint conditions are met, and output the final optimal debugging parameter adjustment scheme.

[0195] Through the above dynamic gain compensation and boundary constraint correction processing, the basic strategy of the previous step is transformed into an optimal debugging parameter adjustment scheme that conforms to the current physical environment and meets hardware security constraints, thereby achieving environmental adaptability and execution security of the debugging strategy.

[0196] For example, the basic debugging parameter adjustment vector is [0.5dB, -2deg], and the real-time signal-to-noise ratio is 15dB, which is lower than the standard value of 20dB. The calculated environmental deviation is 0.8, and the amplitude compensation coefficient is 1.2 and the phase compensation coefficient is 1.1 obtained from the table. The Hadamard product is performed to obtain the intermediate vector [0.6dB, -2.2deg]. The maximum gain threshold is set to 1.0dB and the minimum phase threshold is -5deg. After projection verification, the intermediate vector is within the legal space, and the optimal solution is directly output. If the intermediate vector is [1.2dB, -6deg], the projection is corrected to [1.0dB, -5deg] to ensure that the hardware is not damaged.

[0197] S5.4: Load the optimal debugging parameter adjustment scheme and the control interface protocol of the communication system to be debugged as input conditions, and convert the optimal debugging parameter adjustment scheme into a specific register configuration instruction sequence through the automated execution engine and send it to the hardware control unit to complete the real-time correction operation of the actual communication system operating parameters.

[0198] It receives the optimal debugging parameter adjustment scheme after dynamic calibration and the control interface protocol description file of the communication system to be debugged, parses the register address mapping table, data bit width definition and read / write permission constraints in the protocol, and constructs the instruction generation context environment of the hardware abstraction layer.

[0199] Based on the parameter key-value pairs in the optimal debugging parameter adjustment scheme, the query instruction generates the address mapping relationship in the context environment, converts the logical parameter name into a physical register address index, and performs data type conversion according to the data bit width definition, encoding the floating-point or integer parameter value into a hexadecimal raw data frame that meets the bus transmission requirements.

[0200] To address complex parameter adjustments involving multiple register linkages, a transaction integrity verification mechanism is introduced. Based on the timing requirements of the control interface protocol, the dependency order of write operations to each register is calculated, and an atomic operation sequence containing the starting address, operation type, data payload, and checksum is generated to ensure the atomicity and consistency of the parameter distribution process.

[0201] Through the underlying driver interface of the automated execution engine, the generated atomic operation sequence is encapsulated into a standard communication frame. By using direct memory access technology to bypass the operating system kernel scheduling, the instruction sequence is directly written into the send buffer of the hardware control unit, triggering a hardware interrupt to start the parameter loading process.

[0202] The system monitors the status register feedback of the hardware control unit in real time, captures parameter writing completion signals and abnormal error codes, and immediately triggers a retry mechanism or rolls back to the previous stable state if a verification failure or timeout response is detected, until all register configurations return a success confirmation flag.

[0203] By using the above-mentioned instruction serialization and atomic delivery processing method, the abstract parameter adjustment scheme is transformed into a hardware-executable low-level control instruction stream, realizing millisecond-level real-time correction of the communication system's operating parameters, ensuring the accuracy of debugging strategy execution and the stability of the system state.

[0204] For example, in a scenario of abnormal gain in the RF front-end of a 5G base station, the optimal debugging scheme requires adjusting the value of the control register address 0x4A02 of the variable gain amplifier VGA from 0x1F to 0x2B. The system parsing protocol reveals that this register is 8 bits wide and uses an SPI interface for communication. The automation engine encodes the logic value 0x2B into a single-byte data 0x2B, and combines this with the write command 0x02 and the high and low bit splitting of address 0x4A02 to construct an SPI command frame [0x02, 0x4A, 0x02, 0x2B]. The engine writes this frame to the SPI controller TX FIFO via DMA and monitors the status register. When an ACK signal is received and there is no CRC error, the write is considered successful. This process takes less than 10 microseconds, significantly improving the real-time performance and reliability of parameter delivery and avoiding debugging delays caused by software scheduling latency.

[0205] S5.5: Collect system feedback status signals after the parameter correction operation is executed and the execution log of the optimal debugging parameter adjustment scheme as input conditions. Execute the parameter effectiveness verification logic to perform consistency verification processing, and output a successful execution flag to confirm that the parameter correction is completed, triggering the next round of automated verification process.

[0206] The system receives system feedback status signals after parameter correction operations are executed, as well as execution logs of the optimal debugging parameter adjustment scheme, as input data sources for consistency verification.

[0207] The timestamp, target address, and expected value of register write instructions are parsed from the execution log to construct a parameter distribution baseline sequence and establish the theoretical expected state of the hardware configuration.

[0208] The system feedback status signal is processed by protocol decoding to extract the actual register read values ​​of each functional module of the current communication system, forming the actual response sequence of the hardware configuration.

[0209] A bit-by-bit comparison method is used to perform a full match between the parameter distribution reference sequence and the actual response sequence, and the Hamming distance between the two is calculated to quantify the degree of configuration deviation.

[0210] Configuration consistency is determined based on Hamming distance. If the Hamming distance is zero, the parameters are determined to be correct and a configuration consistency flag is generated. If the Hamming distance is non-zero, the difference bit is located and a configuration anomaly alarm is generated.

[0211] Under the premise of consistent configuration, monitor the status of the watchdog timer and the voltage and current telemetry data inside the system to confirm that the hardware unit has not reset or overload protection action after the parameters are loaded.

[0212] By combining the configuration consistency flag and hardware health status, a logical AND operation is performed. Only when both are true, a confirmation flag indicating successful execution and parameter correction is output.

[0213] Through the above consistency verification and status monitoring mechanism, the parameter distribution result of the previous step is transformed into a reliable execution success flag, realizing closed-loop verification of the parameter correction process and ensuring that the subsequent automated verification process starts based on the correct system status.

[0214] Step S6: After parameter correction, a new round of automated verification process is automatically triggered. If the verification results show that the multidimensional difference features meet the preset threshold, a debugging completion report is generated. If not, steps S3-S6 are recursively executed until the preset iteration limit is reached. Specifically, this includes:

[0215] S6.1: Based on the actual communication system state after the previous round of parameter correction, a preset multi-dimensional test signal sequence is re-injected and a new output response signal is acquired in real time. The corresponding latest theoretical expected response is generated synchronously using a digital twin model. By aligning the new output response signal with the latest theoretical expected response, the time-domain waveform data, frequency-domain spectral data, and modulation domain constellation diagram data to be verified are obtained.

[0216] The system receives a system status confirmation signal from the previous round of parameter correction as a trigger command for the new round of automated verification. Based on predefined multi-dimensional test signal sequence configuration parameters, an arbitrary waveform generator is invoked to generate a simulated excitation signal completely consistent with the initial debugging phase. This simulated excitation signal is injected into the input port of the communication system under test via a high-precision RF interface to ensure consistency of physical excitation conditions. A high-speed data acquisition card is used to synchronously sample the output port of the actual communication system to obtain the original output response signal after parameter correction. Noise filtering and clock synchronization correction are performed on the original output response signal to eliminate environmental noise interference and sampling clock jitter, generating a standardized actual output response signal. The digital twin model of the communication system under test is driven in parallel, inputting the same test signal sequence configuration parameters and the currently corrected system status parameters. Real-time simulation of the digital twin model is performed to generate the theoretical expected response corresponding to the current system state. The theoretical expected response undergoes the same preprocessing procedures as the actual signal, including time-frequency alignment and amplitude normalization, to generate a standardized theoretical expected response signal. A cross-correlation algorithm is used to calculate the time delay offset between the standardized actual output response signal and the standardized theoretical expected response signal. Phase alignment is performed on the two signals based on the aforementioned time delay offset to eliminate the time delay difference introduced by the transmission link. The aligned time-domain waveform data is directly used as the basic data for time-domain verification. Fast Fourier Transform is performed on the aligned time-domain signals to extract frequency-domain spectral data, which is used as the basic data for frequency-domain verification. The in-phase and quadrature components are recovered from the aligned time-domain signals using a quadrature demodulation method, and the resulting modulation domain constellation data is mapped and used as the basic data for modulation domain verification. Through the above parallel acquisition, simulation, and multi-dimensional alignment processing, the parameter correction results from the previous step are transformed into time-domain waveforms, frequency-domain spectral lines, and modulation domain constellations. Figure 3 Using data to be verified, a comprehensive quantitative benchmark for the effect of parameter correction was constructed.

[0217] For example, in a 5G base station power amplifier debugging scenario, an OFDM test signal with a center frequency of 3.5 GHz and a bandwidth of 100 MHz is injected. The data acquisition card obtains the actual response at a sampling rate of 200 MSps, and the digital twin model synchronously generates the theoretical response. The time delay offset is calculated to be 2.5 ns through cross-correlation, and time shift compensation is performed on the theoretical response accordingly. The final output has 2048 aligned time-domain waveform points, a frequency domain spectral resolution of 100 kHz, and a modulation domain constellation diagram containing 5000 16QAM symbols, providing accurate data input for the subsequent difference feature calculation in step S6.2.

[0218] S6.2: Based on the obtained time-domain waveform data, frequency-domain spectral data and modulation-domain constellation diagram data to be verified, perform time-domain root mean square error calculation, frequency-domain power spectral density difference calculation and modulation-domain error vector amplitude calculation respectively to quantify the deviation between the actual response and the theoretical response, and generate the current round multidimensional difference feature containing time-domain difference value, frequency-domain difference value and modulation-domain difference value.

[0219] The time-domain waveform data, frequency-domain spectral data, and modulation-domain constellation diagram data to be verified, output from step S6.1, are used as the input source for the multidimensional difference feature calculation of the current round.

[0220] A point-by-point differential operation is performed on the time-domain waveform data to obtain the residual sequence of the actual response signal and the theoretical expected response on the time axis, so as to eliminate the influence of DC bias and highlight the transient distortion characteristics.

[0221] Based on the residual sequence, the root mean square method is used to calculate the time-domain difference value and quantify the overall deviation of the time-domain waveform.

[0222] Power spectral density estimation is performed on the frequency domain spectral data. The power spectra of the actual signal and the desired signal are calculated in segments using the Welch method to suppress random noise interference and improve spectral resolution.

[0223] Calculate the absolute difference between the actual power spectrum and the expected power spectrum at each frequency point, and integrate the difference across the entire frequency band to generate a frequency domain difference value that characterizes the shift in spectral energy distribution.

[0224] The error vector magnitude (EVM) is calculated on the modulation domain constellation diagram data, the coordinates of the actual constellation points are mapped to the complex plane, and the Euclidean distance between them and the ideal constellation points is calculated.

[0225] The normalized modulation domain error vector amplitude is calculated using the following formula to quantify the modulation accuracy loss:

[0226]

[0227] in, For the error vector magnitude, For the total number of constellation points, Let E be the complex value of the actual constellation point. Let E be the complex value of the ideal constellation point.

[0228] The calculated root mean square error in the time domain, the integral value of the power spectral density difference in the frequency domain, and the amplitude of the modulation domain error vector are vector-encapsulated to form a multidimensional difference feature of the current round that includes the time domain difference value, the frequency domain difference value, and the modulation domain difference value.

[0229] By using the above-mentioned multi-domain parallel computing processing method, the original waveform and spectrum data obtained in the previous step are transformed into standardized quantization deviation indicators, realizing accurate quantization and abnormal feature extraction of the signal quality of the communication system in multiple dimensions, and providing high-precision data support for subsequent threshold comparison and judgment.

[0230] S6.3: Based on the multidimensional difference features of the current round, call the preset pass threshold set to perform domain-by-domain comparison and judgment. If the time domain difference value, frequency domain difference value and modulation domain difference value are all less than or equal to their respective preset pass thresholds, a verification pass flag will be generated; otherwise, a verification fail flag and a list of specific out-of-standard difference features will be generated.

[0231] The system receives the current round's time-domain difference, frequency-domain difference, and modulation-domain difference as input data. It then calls the corresponding time-domain root mean square error threshold, frequency-domain power spectral density deviation threshold, and modulation-domain error vector amplitude threshold from a preset threshold set. A multi-dimensional difference judgment logic matrix is ​​constructed, and the measured difference values ​​in each domain are compared element-by-element with the corresponding thresholds to generate an intermediate state vector containing Boolean judgment results. For time-domain difference values, if the value is less than or equal to the time-domain root mean square error threshold, the time-domain verification item is marked as passed; otherwise, it is marked as failed and the excess amplitude is recorded. For frequency-domain difference values, if the value is less than or equal to the frequency-domain power spectral density deviation threshold, the frequency-domain verification item is marked as passed; otherwise, it is marked as failed and the main offset frequency points are extracted. For modulation-domain difference values, if the value is less than or equal to the modulation-domain error vector amplitude threshold, the modulation-domain verification item is marked as passed; otherwise, it is marked as failed and the constellation diagram divergence region index is recorded. By combining the verification sub-item statuses of the three domains, a logical AND operation is performed to determine if all sub-items are in a passed state. A global verification pass flag is generated, indicating that the current parameter correction is effective and the system performance meets the standards. If any one or more sub-items are in a failed state, a global verification fail flag is generated, triggering an anomaly backtracking mechanism. All domains marked as failed and their specific out-of-specification values ​​are collected to construct an out-of-specification difference feature list. This list includes fields for difference type, measured value, threshold, and deviation ratio, used to guide root cause re-location in subsequent recursive iterations. Through multi-dimensional threshold domain-by-domain comparison and logical aggregation, the quantitative difference indicators calculated in the previous step are transformed into explicit verification pass or fail flags and a specific out-of-specification difference feature list. This achieves binary determination of verification results and accurate retention of fault characteristics in the debugging closed loop, providing a deterministic decision-making basis for subsequent report generation or recursive optimization.

[0232] For example, the time-domain root mean square error threshold is set to 0.05V, the frequency-domain power spectral density deviation threshold is set to 3dB, and the modulation-domain error vector amplitude threshold is set to 5%. In the current round, the calculated time-domain difference is 0.03V, the frequency-domain difference is 2.1dB, and the modulation-domain difference is 4.8%. Upon comparison, 0.03V is less than 0.05V, so the time-domain verification passes; 2.1dB is less than 3dB, so the frequency-domain verification passes; and 4.8% is less than 5%, so the modulation-domain verification passes. The logical AND operation result is true, generating a verification pass flag, and the system outputs a debugging completion report. In another scenario, the time-domain difference is 0.06V, the frequency-domain difference is 2.1dB, and the modulation-domain difference is 4.8%. Upon comparison, 0.06V is greater than 0.05V, so the time-domain verification fails, and the excess amplitude is recorded as 0.01V; the other two values ​​pass. If the logical AND operation result is false, a verification failure flag is generated, a list of differential features containing time-domain out-of-range information is constructed, and a recursive instruction is triggered to restart the anomaly localization step.

[0233] S6.4: Based on the generated verification pass flag, directly call the report generation module to integrate the debugging data of the current round and the previous rounds, and output a debugging completion report containing the final parameter configuration and verification results; if based on the generated verification fail flag, read the current recursion counter value and compare it with the preset iteration count limit. If the limit is not reached, generate a recursion trigger instruction to restart the anomaly location and parameter correction steps. If the limit is reached, generate a comprehensive diagnostic report and terminate the debugging process.

[0234] Receive the verification pass flag or verification fail flag and the list of out-of-limit difference features generated in step S6.3 as the input conditions for the logical judgment of this step.

[0235] If the verification pass flag is received, the system immediately calls the report generation module to extract the complete context data of the current debugging session from the database, including the initial system state parameters, the multidimensional difference feature sequence in each iteration, the candidate root cause confidence trajectory after dynamic calibration, and the final effective parameter configuration vector.

[0236] The extracted historical debugging data is timestamped and structured to generate a visual debugging completion report that includes time-domain waveform comparison charts, frequency-domain spectral evolution curves, and modulation domain constellation diagram convergence status. This report is then stored in non-volatile memory, marking the successful completion of this automated debugging process.

[0237] If a verification failure flag is received, the system reads the current value of the recursive counter maintained in memory. This value records the cumulative number of times the exception location and parameter correction loop has been executed since the debugging started.

[0238] The current value of the recursion counter is compared with a preset upper limit threshold for the number of iterations. This threshold is preset based on the hardware response latency of the communication system and the convergence characteristics of the algorithm to prevent resource exhaustion caused by infinite loops.

[0239] If the recursion counter value is less than the preset iteration limit, the system generates a recursion trigger instruction, resets some temporary cache variables, and jumps the control flow back to step S3 to restart the abnormal root cause reasoning process based on the latest system state, thereby achieving continuous optimization through closed-loop iteration.

[0240] If the recursion counter value reaches or exceeds the preset iteration limit, the system determines that the current debugging strategy cannot make the signal index converge within the limited number of steps, and then terminates the recursion loop and generates a comprehensive diagnostic report.

[0241] The comprehensive diagnostic report records in detail the reasons for failure, characteristics of remaining residuals, and adjustment strategies tried in all historical rounds, providing a complete traceability basis for manual intervention while releasing the occupied computing resources.

[0242] Through the above branch logic processing, the verification results of the previous step are transformed into explicit process control instructions or final delivery documents, realizing automated closed-loop control of the debugging process and orderly exit of abnormal situations.

[0243] For example, the preset maximum number of iterations is set to 5. After the 3rd iteration, S6.3 outputs a verification failure flag, at which point the recursion counter value is 3. The system compares the values ​​and finds that 3 is less than 5, so it generates a recursion trigger instruction and returns to S3. If the verification still fails after the 6th iteration, the counter value is 6, which is greater than the upper limit of 5. The system then generates a comprehensive diagnostic report of up to 20 pages, including the residual spectrum diagram and confidence change table for 6 iterations, and forcibly terminates the process to avoid system deadlock.

[0244] For those skilled in the art, various other corresponding changes and modifications can be made based on the technical solutions and concepts described above, and all such changes and modifications should fall within the protection scope of the claims of this invention.

[0245] Unless otherwise defined, the technical or scientific terms used herein shall have the ordinary meaning as understood by one of ordinary skill in the art to which this application pertains. The terms “first,” “second,” “third,” and similar terms used in this patent application specification and claims do not indicate any order, quantity, or importance, but are merely used to distinguish different components. Similarly, the terms “an” or “a” and similar terms do not indicate a quantity limitation, but rather indicate the presence of at least one. The terms “comprising” or “including” and similar terms mean that the elements or objects preceding “comprising” or “including” encompass the elements or objects listed following “comprising” or “including” and their equivalents, and do not exclude other elements or objects. The “multiple” mentioned in the embodiments of this application refers to two or more. A and / or B indicate three possibilities: A; B; and A and B.

[0246] The above description is merely an exemplary embodiment of this application, but the scope of protection of this application is not limited thereto. Any person skilled in the art can easily conceive of various equivalent modifications or substitutions within the technical scope disclosed in this application, and such modifications or substitutions should all be covered within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

Claims

1. An automated verification and anomaly localization method for signal debugging in communication engineering, characterized in that, Specifically, it includes: S1: Obtain the digital twin model of the communication system to be debugged and historical debugging case data, construct an abnormal feature-root cause mapping decision tree based on the abnormal features and root cause mapping relationship in the historical debugging case data, and generate a debugging strategy library based on the expert experience library. S2: Inject a preset multi-dimensional test signal sequence into the actual communication system and collect the output response signal in real time. Use the digital twin model to generate the corresponding theoretical expected response. Process the output response signal and the theoretical expected response to obtain multi-dimensional difference features. S3: Input the multidimensional differential features into the abnormal root cause classification model, combine the abnormal feature-root cause mapping decision tree to perform layer-by-layer matching reasoning, and output at least one candidate abnormal root cause and its initial probability confidence. S4: Perform dynamic calibration processing on the candidate anomaly root causes and their initial probability confidence scores to generate dynamically calibrated candidate anomaly root causes and dynamic confidence scores. S5: Based on the dynamically calibrated candidate anomaly root causes and dynamic confidence scores, match the optimal debugging parameter adjustment scheme from the debugging strategy library, and automatically execute the optimal debugging parameter adjustment scheme in the actual communication system to complete the parameter correction. S6: After the parameter correction is completed, a new round of automated verification process is automatically triggered. If the verification results show that the multidimensional difference features meet the preset pass threshold, a debugging completion report is generated. If not, steps S3-S6 are recursively executed until the preset iteration limit is reached.

2. The automated verification and anomaly localization method for signal debugging process in communication engineering according to claim 1, characterized in that, The expert experience database stores systematically collected, organized, and shared expert knowledge, cases, solutions, and practical experience in the field of communications.

3. The automated verification and anomaly localization method for signal debugging process in communication engineering according to claim 1, characterized in that, The abnormal root cause classification model adopts a deep neural network architecture based on the attention mechanism, which includes an input layer, multiple hidden layers and an output layer.

4. The automated verification and anomaly localization method for signal debugging process in communication engineering according to claim 1, characterized in that, S2 specifically includes: Based on a predefined communication protocol standard, multi-dimensional test signal sequence configuration parameters are obtained, and the real-time physical excitation state is obtained by converting the multi-dimensional test signal sequence configuration parameters. Based on the real-time physical excitation state, a high-speed data acquisition card is used to synchronously sample the output port of the actual communication system to obtain the original output response signal, and the original output response signal is subjected to denoising and normalization preprocessing to obtain the output response signal. Based on the configuration parameters of the multi-dimensional test signal sequence, the digital twin model of the communication system under debugging is driven to perform parallel simulation and deduction to generate the theoretical expected response. The theoretical expected response is then subjected to the same time-frequency alignment processing as the output response signal to obtain the aligned theoretical expected response signal. Frequency domain difference spectral lines are obtained by performing calculations based on the output response signal and the expected response signal of the alignment theory. Feature fusion is then performed based on the frequency domain difference spectral lines to finally obtain multidimensional difference features.

5. The automated verification and anomaly localization method for signal debugging process in communication engineering according to claim 4, characterized in that, The process of converting the configuration parameters of the multi-dimensional test signal sequence to obtain the real-time physical excitation state includes: An arbitrary waveform generator is used to perform digital-to-analog conversion on the configuration parameters of the multi-dimensional test signal sequence to generate an analog excitation signal, and the analog excitation signal is injected into the input port of the communication system to be debugged to obtain the real-time physical excitation state.

6. The automated verification and anomaly localization method for signal debugging process in communication engineering according to claim 4, characterized in that, The process of obtaining the frequency domain difference spectrum based on the output response signal and the alignment theory expected response signal includes: The output response signal and the alignment theory expected response signal are subtracted point by point using the difference method to obtain the time-domain residual sequence, and the time-domain residual sequence is processed by fast Fourier transform to obtain the frequency-domain difference spectrum.

7. The automated verification and anomaly localization method for signal debugging process in communication engineering according to claim 6, characterized in that, The feature fusion based on frequency domain difference spectral lines ultimately yields multidimensional difference features, including: Based on the frequency domain difference spectral lines, the signal phase and amplitude information is demodulated using the constellation diagram mapping method to obtain the modulation domain error vector. The time domain residual sequence, the frequency domain difference spectral lines, and the modulation domain error vector are then combined for feature fusion processing to finally obtain multidimensional difference features.

8. The automated verification and anomaly localization method for signal debugging process in communication engineering according to claim 1, characterized in that, S5 specifically includes: The candidate anomaly root causes and dynamic confidence scores after dynamic calibration are obtained as input conditions. Based on the weighted ranking method, the priority of multiple candidate anomaly root causes is calculated and processed to output the target anomaly root cause identifier with the highest confidence weight. The system receives the target anomaly root cause identifier and its associated dynamic confidence score as input conditions, performs multi-dimensional feature retrieval and similarity matching processing in a pre-built debugging strategy library, and outputs an initial debugging strategy set. The initial set of debugging strategies is processed to generate an optimal debugging parameter adjustment scheme. The optimal debugging parameter adjustment scheme and the control interface protocol of the communication system to be debugged are loaded as input conditions. The optimal debugging parameter adjustment scheme is converted into a specific register configuration instruction sequence and sent to the hardware control unit to complete the real-time correction operation of the actual communication system operating parameters.

9. The automated verification and anomaly localization method for signal debugging process in communication engineering according to claim 8, characterized in that, The initial set of debugging strategies includes a basic debugging parameter adjustment vector.

10. The automated verification and anomaly localization method for signal debugging process in communication engineering according to claim 9, characterized in that, The process of processing the initial set of debugging strategies to generate the optimal debugging parameter adjustment scheme includes: The basic debugging parameter adjustment vector and the real-time state context information of the current system are extracted from the initial debugging strategy set as input conditions. An adaptive parameter optimization method is applied to perform dynamic gain compensation and boundary constraint correction on the basic debugging parameter adjustment vector to generate the optimal debugging parameter adjustment scheme.

Images