A Communication Data Evaluation Method and System Based on Language Analysis

Abstract

This application discloses a communication data evaluation method and system based on language analysis, belonging to the field of language analysis. This application extracts five complementary features, namely, clarity, signal-to-noise ratio, spectral distortion, temporal distortion, and coherence, and uses Bayesian networks for probabilistic inference, thereby avoiding the risk of misjudgment caused by traditional hard threshold judgment and significantly improving the robustness and fault tolerance of the system. The causal logic structure of Bayesian networks enables diagnostic results to be directly traced to specific observation indicators, providing clear interpretability and traceability, which facilitates maintenance personnel to quickly locate the root cause of the fault, thereby continuously optimizing the diagnostic accuracy to cope with dynamic challenges such as link aging and environmental changes.

Description

TECHNICAL FIELD

[0001] The present application belongs to the field of language analysis, and specifically relates to a communication data evaluation method and system based on language analysis. BACKGROUND

[0002] In existing communication scenarios such as voice communication and service data transmission, the communication link is easily affected by various factors such as channel interference, bandwidth fluctuation, signal attenuation, frequency offset, link jitter, and environmental noise, and communication quality degradation problems such as voice distortion, transmission lag, data distortion, and communication clarity reduction are prone to occur. At present, the traditional communication quality evaluation method mainly uses single signal-to-noise ratio and link delay to make simple judgment, which has the defects of single evaluation dimension, low precision, and poor pertinence. It is difficult to accurately identify the poor communication state by combining the clarity, coherence, time domain distortion, and spectral characteristics of the voice language.

[0003] The existing communication monitoring technology can only realize passive alarm of communication quality abnormality, lacks fault tracing ability based on language spectral characteristics, cannot accurately locate the core inducement leading to communication quality decline, and is difficult to distinguish different fault types such as interference, bandwidth, noise, and frequency offset, resulting in strong blindness and poor adaptability of subsequent communication optimization adjustment. The traditional communication system does not have the ability of adaptive repair of communication faults, cannot execute targeted correction methods such as frequency band switching, bandwidth allocation, gain adjustment, and spectrum calibration according to different fault reasons, and has weak self-healing optimization ability of communication quality. The existing communication transmission mechanism adopts a unified transmission strategy, cannot pre-judge and grade the service priority, semantic importance, time efficiency demand, and confidentiality level of future to-be-transmitted messages according to historical communication data and real-time link state, and cannot match different transmission schemes for messages of different importance levels. In the link quality fluctuation scenario, high-importance service message transmission delay, verification loss, and transmission distortion are prone to occur, and low-priority messages occupy a large amount of channel resources, resulting in low overall communication resource utilization, poor stability and reliability of core business communication, and difficulty in meeting the modern communication transmission demand of high precision, high reliability, and intelligence. SUMMARY

[0004] In view of the deficiencies of the prior art, the present application provides a communication data evaluation method and system based on language analysis To achieve the above-mentioned purpose, the present application provides the following technical scheme: The communication data evaluation method and system based on language analysis include the following specific steps: S1, real-time collection of voice language communication data transmitted by a communication link, preprocessing of original communication language data, and obtaining of standardized language analysis signals; S2, based on the speech language feature dimension, the standardized language analysis signal is subjected to multi-dimensional language analysis, language intelligibility features, speech signal-to-noise ratio features, language time domain distortion features, frequency spectrum features and speech coherence features are extracted, and a communication language quality feature dataset is constructed; S3, a preset communication data quality grading evaluation threshold is set, the extracted multi-dimensional language feature data is matched and compared with the preset threshold, and the overall quality grade of the current communication data is quantitatively determined; S4, if it is determined that the communication data quality is in a poor abnormal state, the real-time spectrum data and the time-frequency feature map corresponding to the current language are called, the fault causes affecting the communication quality are intelligently predicted and analyzed through spectrum distortion features, frequency attenuation interval, harmonic interference features and spectrum aliasing features; S5, the various influencing causes obtained through the tracing are subjected to weight ordering, the main influencing factors causing the communication quality degradation are screened out, and the preset fault corresponding reminding strategy and the self-adaptive optimization disposal strategy are matched.

[0005] Preferably, the real-time collection of the voice language communication data transmitted by the communication link, the pre-processing of the original communication language data to obtain the standardized language analysis signal comprises the following specific steps: S11, through the integration of multiple types of physical interfaces, the voice load data in the communication link is subjected to non-intrusive bypass collection and low-delay mirror grabbing, the voice signal is subjected to analog-digital conversion, and the data packet is subjected to real-time decoding and restoration according to the standard encoding and decoding format; S12, the various noises mixed in the collected original voice signal are subjected to noise reduction filtering processing, and the sliding window frame processing is performed on the continuous time domain signal to generate the standardized language analysis signal.

[0006] Preferably, the construction of the communication language quality feature dataset comprises the following specific steps: S21, the standardized language analysis signal is subjected to octave band pass filter set, the time domain envelope signal in each frequency band is separated, the envelope signal of each frequency band is subjected to second Fourier transform, the modulation depth on the modulation frequency is extracted, the envelope signal of each frequency band is subjected to second Fourier transform, the modulation depth on the modulation frequency is extracted, the modulation retention rate of each frequency band is weighted and integrated with the standard human ear hearing weight function to obtain the voice transmission index, and the intelligibility index is calculated according to the ANSIS3.5 standard, and the comprehensive intelligibility score is calculated by weighting the normalized voice transmission index and intelligibility index; S22, the minimum value recursive average algorithm is used to continuously track the minimum power spectrum value of each frequency point in the frequency domain as the real-time noise basement of the frequency point, the decision guided method is adopted, the priori signal-to-noise ratio of the current frame is calculated by combining the posteriori signal-to-noise ratio of the previous frame and the instantaneous power spectrum of the current frame, and the calculation formula of the priori signal-to-noise ratio is: wherein, is the prior SNR of the mth frame at the kth frequency point, is a smoothing factor, is the clean speech power spectrum of the kth frequency point in the previous frame, is the noise power spectrum of the kth frequency point in the previous frame, is the posterior SNR of the mth frame at the kth frequency point, is a half-wave rectification correction term, the entire effective speech is dynamically divided into high-energy segments, low-energy segments and pure noise segments according to their power, the segment average prior SNR, the minimum prior SNR and the prior SNR fluctuation variance of the three segments are counted respectively, and the speech SNR feature vector is output, including the global mean, the median, the low quantile and the high quantile; S23, language time domain distortion feature extraction; the specific content includes extracting the pitch period and peak amplitude of each frame by autocorrelation method, calculating the absolute average of the pitch period change rate between adjacent effective voiced frames, and the absolute average of the adjacent peak amplitude difference, waveform symmetry degree and peak clipping coefficient: calculating the area ratio of positive and negative half cycles of the time domain signal to obtain the asymmetric distortion degree, and simultaneously counting the proportion of samples whose peak amplitude exceeds the linear dynamic range threshold to generate the peak clipping distortion index; jump edge steepness loss: the maximum change rate between adjacent samples of the signal is calculated by using the differential operator, and compared with the slope template of the standard reference signal to extract the edge blunting coefficient.

[0007] S24, frequency spectrum feature extraction; the specific features include: power spectral density and spectral centroid: perform N-point fast Fourier transform on each windowed signal to calculate the power spectrum. Extract the spectral centroid, whose offset directly quantifies the frequency drift degree; frequency band energy ratio: divide the effective frequency band into low frequency band, middle frequency band and high frequency band; calculate the percentage of the energy of each frequency band in the total energy; when the high frequency band energy ratio drops sharply, mark it as a high frequency attenuation feature; when the middle and low frequency energy ratio abnormally rises, mark it as a low frequency hum interference feature; spectral flatness and spectral entropy: calculate the ratio of geometric mean power to arithmetic mean power. Spectral flatness tends to 1 indicates that the signal is noise-like, and tends to 0 indicates that the signal is harmonic-like. This index forms cross-validation with the prior SNR of S2.3; harmonic energy cluster distribution: extract the energy peak value and its half peak width at the fundamental frequency and its multiples. If the harmonic peak presents broadening or splitting, it is quantitatively determined as spectral aliasing distortion degree, which provides quantitative input for harmonic interference tracing in S4.

[0008] S25, voice coherence feature extraction; specific features include: long-term autocorrelation attenuation coefficient: calculate the normalized autocorrelation function envelope attenuation slope of the normalized language analysis signal in the 100ms to 1000ms lag interval. Normal speech is characterized by slow attenuation of autocorrelation due to periodicity of voiced periods; if there is a burst of packet loss in the link, the autocorrelation envelope will drop sharply at the corresponding lag, and the drop amplitude is extracted as the time sequence break index; speech rate and pause variation coefficient: use the silence interval index output by the speech activation detection in S1 to calculate the effective speech rate and pause duration distribution; compare the real-time statistical value with the standard speech rate template by dynamic time warping, and output the rhythm distortion degree, which is used to determine whether the speech speed is out of control due to time delay jitter; resonance peak trajectory continuity: extract the time-varying trajectory of the first three order resonant peaks through LPC analysis. Calculate the change rate of resonant peak frequency between adjacent frames, if the change rate exceeds the physiological limit of normal pronunciation, it is determined as abnormal jump, marked as channel switching or codec parameter mutation caused by incoherence; S26, generate original feature vectors for each type of index in steps S21-S25, calculate the mean and standard deviation of each dimension feature in the entire communication session time period independently, and perform standardization processing to obtain a standardized feature vector.

[0009] Preferably, the quantization determines the overall quality level of the current communication data, including the following specific steps: weighting each feature in the standardized feature vector to calculate a comprehensive score, corresponding to the clarity standardized value, the signal-to-noise ratio standardized value, the spectral distortion standardized value, the time domain distortion standardized value, and the coherence standardized value, respectively, based on the pre-set communication data quality classification threshold, the overall quality level of the current communication data is divided into high-quality communication state, general fluctuation state, and poor-quality abnormal state; that is, communication data greater than or equal to the high-quality threshold is set as data in the high-quality communication state, communication data less than or equal to the poor-quality threshold is set as data in the poor-quality abnormal state, and communication data between the high-quality threshold and the poor-quality threshold is set as data in the general fluctuation state. The weights and thresholds are obtained through historical data fitting.

[0010] Preferably, if the communication data quality is determined to be a poor-quality abnormal state, the real-time spectral data and time-frequency feature map corresponding to the current language are retrieved, and the fault causes affecting the communication quality are intelligently predicted and analyzed through spectral distortion features, frequency attenuation interval, harmonic interference features, and spectral aliasing features, including the following specific steps: S41, the communication data determined to be a poor abnormal state is used to call real-time spectrum data and time-frequency feature map corresponding to the current language, a time-frequency-energy three-dimensional graph matrix is generated by using short-time Fourier transform, the absolute mean value of the third derivative of the power spectrum envelope is calculated, which is denoted as the spectrum envelope roughness, and the depth of the local minimum point in the spectrum envelope relative to the adjacent peak value and the 3dB notch bandwidth are extracted by detecting the local minimum point in the spectrum envelope; S42, the effective frequency band is divided into equal-width subbands, the average power of each subband is calculated, a frequency-power linear regression curve is fitted by using the least square method, the attenuation slope is obtained, and the frequency point corresponding to the place where the power spectrum density first drops to-20dB of the average power of the low frequency band is taken as the cutoff frequency; S43, the fundamental frequency is extracted by using the cepstrum method, whether there is a peak value at the frequency doubling place is checked, the ratio of the harmonic peak value to the average value of adjacent frequency points is calculated, which is denoted as the harmonic structure integrity index; S44, for the frequency band higher than the Nyquist frequency, whether there is a mirror image of the peak value in the original effective frequency band is checked, if the mirror image component is detected, it is determined that there is spectral aliasing caused by undersampling or failure of the anti-aliasing filter; all the extracted quantitative indicators are input into the pre-trained Bayesian network, the posterior probability of the potential fault cause is calculated respectively, and the fault probability is output.

[0011] Preferably, the weight order of the various influence causes obtained by tracing back is sorted, the main influencing factors causing the communication quality degradation are screened out, and the preset fault corresponding reminding strategy and adaptive optimization disposal strategy are matched, which include the following specific steps: According to the current link type, service period and user service level agreement, the fault probability is dynamically loaded and weighted, the fault probability with the highest value is taken as the main fault cause, the preset fault corresponding reminding strategy and adaptive optimization disposal strategy are matched, the corresponding communication quality abnormal reminding information is generated and pushed to the communication management and control end, and the corresponding adaptive processing means is executed according to the main fault cause, including frequency band switching and interference shielding for channel interference, dynamic bandwidth resource allocation for bandwidth fluctuation, transmission power adaptive gain adjustment for signal attenuation, frequency spectrum calibration for frequency offset, link queue optimization for link delay jitter, and deep noise reduction parameter iterative adjustment for noise intrusion.

[0012] In the second aspect, the communication data evaluation system based on language analysis is realized based on the above-mentioned communication data evaluation method based on language analysis, which specifically includes: A communication preprocessing module is used to collect voice language communication data transmitted by a communication link in real time, pre-process the original communication language data, and obtain standardized language analysis signals. The feature analysis module performs multi-dimensional language analysis on standardized language analysis signals based on speech language feature dimensions, extracting language profundity features, speech signal-to-noise ratio features, language temporal distortion features, frequency spectrum features, and speech coherence features to construct a communication language quality feature dataset. The quality analysis module is used to preset the threshold for evaluating the quality of communication data. It matches and compares the extracted multi-dimensional language feature data with the preset threshold to quantitatively determine the overall quality level of the current communication data. If the source analysis module determines that the communication data quality is in a poor or abnormal state, it retrieves the real-time spectrum data and time-frequency feature map corresponding to the current language, and performs source analysis through spectrum distortion features, frequency attenuation range, harmonic interference features, and spectrum aliasing features to intelligently predict and analyze the causes of faults affecting communication quality. The strategy matching module is used to sort the various influencing factors obtained from the source tracing by weight, filter out the main influencing factors that lead to the degradation of communication quality, and match the preset fault-corresponding reminder strategy and adaptive optimization handling strategy.

[0013] Thirdly, an electronic device includes: a processor and a memory, wherein the memory stores a computer program that can be called by the processor; The processor executes the above-described language analysis-based communication data evaluation method by calling a computer program stored in the memory.

[0014] Fourthly, a computer-readable storage medium is characterized by storing instructions that, when executed on a computer, cause the computer to perform the above-described language analysis-based communication data evaluation method.

[0015] Compared with the prior art, the beneficial effects of this application are: By extracting five complementary features—clarity, signal-to-noise ratio, spectral distortion, temporal distortion, and coherence—and using Bayesian networks for probabilistic inference, the system avoids the risk of misjudgment caused by traditional hard threshold judgments, significantly improving its robustness and fault tolerance. At the same time, the causal logic structure of Bayesian networks allows diagnostic results to be directly traced back to specific observation indicators, providing clear interpretability and traceability, making it easier for maintenance personnel to quickly locate the root cause of the fault. Attached Figure Description

[0016] Figure 1 This is a schematic diagram of the overall process of the communication data evaluation method based on language analysis in this application; Figure 2 This is a schematic diagram illustrating the multi-dimensional language features of this application; Figure 3 This is a schematic diagram of the overall framework of the communication data evaluation system based on language analysis in this application. Detailed Implementation

[0017] The technical solutions in the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments.

[0018] Example 1

[0019] Please see Figure 1 and Figure 2 One embodiment provided in this application is as follows: Figure 1 As shown, the communication data evaluation method based on language analysis includes the following specific steps: S1. Real-time acquisition of voice and language communication data transmitted through the communication link, preprocessing of the raw communication language data to obtain standardized language analysis signals; In this embodiment, the specific implementation steps of step S1 are as follows: S11, by integrating multiple types of physical interfaces (including but not limited to E1 / T1 digital trunk interface, analog PSTN subscriber line interface, VoIPSIP / RTP protocol stack packet capture interface and software-defined radio SDR front end), the voice payload data in the communication link is non-intrusively bypassed and low-latency mirrored capture is performed. The acquisition module performs analog-to-digital conversion on the voice signal with a sampling rate of 8kHz (narrowband) and 16kHz (wideband) and 16-bit linear quantization precision, and decodes and restores the data packets in real time according to the standard codec format. S12. Perform noise reduction filtering on various noises mixed in the acquired raw speech signal (noise reduction is performed by adaptive spectral subtraction and wavelet packet transform threshold decorrelation filtering). Perform sliding window framing processing on the continuous time domain signal, set the frame length and frame shift, use Hamming window as the frame truncation window, multiply each frame of time domain data with the Hamming window coefficient to obtain the windowed short-time analysis frame, and generate a standardized language analysis signal.

[0020] S2. Perform multi-dimensional language analysis on standardized language analysis signals based on speech language feature dimensions, extract language proficiency features, speech signal-to-noise ratio features, language temporal distortion features, frequency spectrum features, and speech coherence features, and construct a communication language quality feature dataset. In this embodiment, the specific implementation steps of step S2 are as follows: S21, the standardized speech analysis signal is passed through an octave bandpass filter bank to separate the time-domain envelope signal in each frequency band. A second Fourier transform is performed on the envelope signal of each frequency band to extract its modulation depth at the modulation frequency. The modulation depth retention rate directly reflects the transmission fidelity of the channel to the changes in the fundamental frequency of voiced speech and the transient amplitude fluctuations of unvoiced speech. A second Fourier transform is performed on the envelope signal of each frequency band to extract its modulation depth at the modulation frequency. The modulation depth retention rate directly reflects the transmission fidelity of the channel to the changes in the fundamental frequency of voiced speech and the transient amplitude fluctuations of unvoiced speech. The modulation retention rate of the band is weighted and integrated with the standard human hearing weighting function to obtain the speech transmission index. The intelligibility index is calculated according to the ANSIS 3.5 standard. The speech transmission index and the intelligibility index are normalized and then weighted to calculate the comprehensive intelligibility score. The ANSIS 3.5 standard defines a method for calculating a physical measurement that is highly correlated with speech intelligibility. This physical measurement is evaluated by a set of speech perception tests for speakers and listeners. This metric is called the Speech Intelligibility Index (SII). The SII is calculated based on acoustic measurements of speech and noise. It is a conventional technique in this field and will not be described in detail here. S22. Using the minimum recursive averaging algorithm, the minimum power spectral density value of each frequency point is continuously tracked in the frequency domain as the real-time noise basis for that frequency point. A decision-guided method is employed, combining the posterior signal-to-noise ratio of the previous frame with the transient power spectral density of the current frame, to calculate the prior signal-to-noise ratio of the current frame. The formula for calculating the prior signal-to-noise ratio is as follows: ,in, Let be the prior signal-to-noise ratio at the k-th frequency point in the m-th frame. As a smoothing factor, it controls the allocation ratio between the weights of historical frames and the observation weights of the current frame. This is the power spectrum of the clean speech at the k-th frequency point in the previous frame. This is the noise power spectrum at the k-th frequency point of the previous frame. The posterior signal-to-noise ratio (SNR) at the k-th frequency point in the m-th frame is obtained from the current observation data. This is a half-wave rectification correction term, partly called the maximum likelihood instantaneous estimator. Since the posterior signal-to-noise ratio may be less than 1 in actual communication (i.e., the signal power is lower than the noise power), directly taking γ-1 as a negative value would violate the physical meaning. Therefore, the lower limit is clamped to 0 by taking the maximum value function to avoid the prior signal-to-noise ratio from being negative. The entire effective speech segment is dynamically divided into high-energy segments (voiced sounds), low-energy segments (unvoiced and weak sounds), and pure noise segments (silent frame regions) according to their power. The segmental average prior signal-to-noise ratio, minimum prior signal-to-noise ratio, and prior signal-to-noise ratio fluctuation variance of the three segments are calculated respectively. The speech signal-to-noise ratio feature vector is output, which includes the global mean, median, lower quantile, and higher quantile to comprehensively capture the impact of dynamic compression or burst noise on the communication signal. S23. Extraction of temporal distortion features of speech; specifically, it includes extracting the pitch period and peak amplitude of each frame using the autocorrelation method, calculating the absolute average of the pitch period change rate between adjacent effective voiced frames, and the absolute average of the difference between adjacent peak amplitudes; waveform symmetry and clipping coefficient: calculating the area ratio of the positive and negative half-cycles of the time-domain signal to obtain the asymmetric distortion, and simultaneously counting the proportion of samples whose peak amplitude exceeds the linear dynamic range threshold (such as 90% of the full scale) to generate a clipping distortion index, which is used to identify AGC overcompensation or power amplifier saturation problems; edge steepness loss: using the differential operator to calculate the maximum change rate between adjacent samples of the signal, comparing it with the slope template of the standard reference signal, and extracting the edge passivation coefficient to reflect the phase delay or capacitive coupling attenuation of high-frequency components during transmission; S24. Frequency Spectrum Feature Extraction; Specific features include: Power Spectral Density and Spectral Centroid: Perform an N-point (N = 512 or 1024) Fast Fourier Transform on each windowed signal frame to calculate the power spectrum; extract the spectral centroid (i.e., the centroid frequency of the power spectrum), and its offset directly quantifies the degree of frequency drift; Band Energy Ratio: Divide the effective frequency band into low-frequency band (300Hz to 1kHz), mid-frequency band (1kHz to 2.5kHz), and high-frequency band (2.5kHz to 3.4kHz / 7kHz), and calculate the energy ratio of each band. The percentage of energy in the total energy is used to indicate high-frequency attenuation when the energy ratio in the high-frequency band drops sharply, and low-frequency humming interference when the energy ratio in the mid-to-low frequency band rises abnormally. Spectral flatness and spectral entropy are calculated as the ratio of geometric mean power to arithmetic mean power. A spectral flatness approaching 1 indicates signal-like noise (severe interference), while a flatness approaching 0 indicates signal-like harmonics (clear speech), cross-validating with the SNR in S2.3. Harmonic energy cluster distribution extracts the energy peaks and their half-peak widths at the fundamental frequency and its harmonics (second and third harmonics). If the harmonic peaks broaden or split, they are quantitatively determined as spectral aliasing distortion, providing quantitative input for harmonic interference source tracing in S4.

[0021] S25. Speech coherence feature extraction; specific features include: long-term autocorrelation decay coefficient: calculate the decay slope of the normalized autocorrelation function envelope of the standardized speech analysis signal in the lag interval of 100ms to 1000ms. Due to the periodicity of voiced sounds, the autocorrelation of normal speech decays slowly. If there is a sudden packet loss in the link, the autocorrelation envelope will drop sharply at the corresponding lag. The magnitude of this drop is extracted as the temporal breakage index; speech rate and pause variation coefficient: using the silent interval index output by speech activation detection in S1, calculate the effective speech rate (number of syllables / second) and pause duration distribution (short pause, medium pause, long pause). Compare the real-time statistical values ​​with the standard speech rate template through dynamic time warping and output the prosodic distortion degree to determine whether there is speech speed loss due to time delay jitter; formant trajectory continuity: extract the time-varying trajectory of the first three formants through LPC (linear predictive coding) analysis. Calculate the rate of change of the formant frequency between adjacent frames. If the rate of change exceeds the physiological limit of normal pronunciation, it is judged as an abnormal jump and marked as a discontinuity caused by channel switching or sudden change in encoding and decoding parameters. S26. Generate original feature vectors for the various indicators in steps S21-S25. Calculate the mean and standard deviation of each feature independently over the entire communication session time period, and perform standardization processing to obtain standardized feature vectors.

[0022] S3. Set a preset threshold for evaluating the quality of communication data. Match and compare the extracted multi-dimensional language feature data with the preset threshold to quantitatively determine the overall quality level of the current communication data. In this embodiment, step S3 is specifically implemented as follows: The various features in the standardized feature vector (corresponding to the sharpness standardized value, signal-to-noise ratio standardized value, spectral distortion standardized value, temporal distortion standardized value, and coherence standardized value, respectively) are weighted to calculate a comprehensive score. Based on a pre-set communication data quality grading threshold, the overall quality level of the current communication data is divided into three categories: excellent communication state, general fluctuation state, and poor quality abnormal state. Specifically, communication data greater than or equal to the excellent threshold is set as excellent communication data, and communication data less than or equal to the poor quality threshold is set as poor quality abnormal state. Communication data between the excellent and poor quality thresholds is set as general fluctuation data. In the dynamic state, the weights and thresholds are obtained through historical data fitting. That is, the experimenter obtains the degree of discrimination of these communication data (i.e., the accuracy of sentence recognition in the language) by acquiring historical communication data. Then, the historical data is divided into two categories: those that can be recognized by 90% or more (i.e., the meaning of the language can be completely deduced from the language) are set as high-quality communication state; those that can be recognized by 50% or less (i.e., the meaning of the language cannot be deduced from the fragmented language) are set as low-quality abnormal state; and those in between are set as general fluctuation state. Then, the historical experimental data is imported into MATLAB fitting software for linear fitting, and the weight values ​​that meet the set accuracy are output.

[0023] S4. If the communication data quality is determined to be poor and abnormal, retrieve the real-time spectrum data and time-frequency feature map corresponding to the current language, and conduct source analysis through spectrum distortion features, frequency attenuation range, harmonic interference features, and spectrum aliasing features to intelligently predict and analyze the causes of faults affecting communication quality. In this embodiment, the specific implementation steps of step S4 are as follows: S41, for communication data that is determined to be in a poor quality abnormal state, retrieve the real-time spectrum data and time-frequency feature map corresponding to the current language, use short-time Fourier transform with a window length of 40ms and an overlap rate of 75% (step 10ms) to generate a three-dimensional time-frequency-energy spectrum matrix, calculate the absolute mean of the third derivative of the power spectrum envelope, and record it as the spectrum envelope roughness, quantify the degree of irregular fluctuation of the spectrum curve, and extract its depth relative to the adjacent peak and 3dB notch bandwidth by detecting local minimum points (notch) in the spectrum envelope. S42. Divide the effective frequency band into equal-width sub-bands, calculate the average power for each sub-band, fit the frequency-power linear regression curve using the least squares method, obtain the attenuation slope, and find the frequency point corresponding to the first drop of the power spectral density to -20dB of the average power in the low-frequency band as the cutoff frequency. S43. Extract the fundamental frequency using the cepstral method, check whether there is a peak at the harmonic frequency, and calculate the ratio of the harmonic peak value to the mean of the adjacent frequency points, which is denoted as the harmonic structure integrity index. S44. For frequency bands higher than the Nyquist frequency, check whether there is a mirror mapping of the peak value within the original effective frequency band. If a mirror component is detected (power > -15dB of the corresponding original peak), it is determined that there is spectral aliasing caused by undersampling or failure of the anti-aliasing filter. Input all extracted quantization indicators into the pre-trained Bayesian network, calculate the posterior probability for each potential fault cause, and output the probability of each fault. The network structure and conditional probability table have been trained with hundreds of thousands of simulated and measured fault samples before system deployment. The fault causes include channel interference, bandwidth fluctuation, signal attenuation, frequency offset, link delay jitter, and noise intrusion. The specific steps are as follows: When constructing a Bayesian network for spectrum sensing, it is first necessary to clarify the network topology, that is, define the nodes and edges. Network nodes are divided into two categories: fault cause nodes (parent nodes) and observation indicator nodes (child nodes). Fault cause nodes typically include channel interference, bandwidth fluctuations, signal attenuation, frequency offset, link delay jitter, and noise intrusion. Each node is set to a discrete state (e.g., normal and abnormal) and assigned an initial prior probability based on expert experience or historical fault statistics. Observation indicator nodes correspond to quantitative features extracted from spectrum data, such as spectral envelope roughness, notch depth and bandwidth, attenuation slope, cutoff frequency, harmonic structure integrity index, and aliasing image power ratio. These indicators are usually also... The data is discretized into several levels (e.g., high / medium / low). During the structure learning phase, hundreds of thousands of historical sample data containing fault labels and corresponding spectral features are used. The optimal directed acyclic graph structure is found within a given set of candidate parent nodes using the K2 scoring algorithm or hill-climbing search method. The core objective is to maximize the scoring function (e.g., BIC or Bayesian Dirichlet score) to determine which fault cause nodes each observed index node depends on. For example, the algorithm may find that the notch depth directly depends on channel interference and signal attenuation, while the attenuation slope depends only on signal attenuation and frequency shift.During the parameter learning phase, after the structure is determined, maximum likelihood estimation or Bayesian estimation is used to calculate the conditional probability table (CPT) for each node. For the root node (fault cause), the CPT is its prior probability distribution. For child nodes (observation indicators), the CPT records the probability of the observation indicator taking each level under various combinations of parent node states. For example, under the conditions of abnormal channel interference and normal signal attenuation, the probability of observing a high notch depth is 0.85, while the probability of a low notch depth is only 0.03. These probability values ​​are obtained entirely from the frequency statistics of the corresponding state combinations in the sample data. If the sample is sparse, Laplace smoothing can be used. Finally, after the system is built, it is tested in practice. During inference, the spectral index vector of the current signal is extracted in real time and input into the Bayesian network. Through joint probability decomposition and variable elimination algorithms (such as cluster tree propagation), the posterior probability distribution of each fault cause is quickly calculated, and the fault type with the highest posterior probability is output as the diagnostic conclusion. This process organically integrates prior knowledge, causal structure, data-driven approaches, and probabilistic reasoning. This allows the network to maintain basic performance when the sample size is insufficient, relying on expert priors, while correcting its structure with sufficient data. It possesses strong robustness and interpretability. After construction, the system inputs the currently extracted index vector during inference, uses the Bayesian formula to calculate the posterior probability of each fault cause, and outputs the fault type with the highest probability. The advantages of this construction are: First, the Bayesian network can naturally integrate prior knowledge (such as expert experience) with measured data, maintaining robustness even with a small sample size; second, its probabilistic reasoning mechanism can quantify uncertainty, avoiding the risk of misjudgment due to hard threshold judgments; third, the network structure is transparent and interpretable, facilitating reverse tracing of the causal path between fault causes and features, thereby improving the accuracy and maintainability of communication quality diagnosis.

[0024] S5. Rank the various influencing factors obtained from the source tracing by weight, filter out the main influencing factors that lead to the degradation of communication quality, and match the preset fault-corresponding reminder strategy and adaptive optimization handling strategy. In this embodiment, the specific implementation steps of step S5 are as follows: based on the current link type, service time period, and user service level protocol, the fault probability is dynamically loaded and sorted by weight, and the fault probability with the highest probability is taken as the main fault cause. The preset fault-corresponding reminder strategy and adaptive optimization handling strategy are matched to generate targeted communication quality abnormality reminder information and push it to the communication management terminal. Corresponding adaptive processing measures are executed according to the main fault cause, including performing frequency band switching and interference shielding for channel interference, performing dynamic bandwidth resource allocation for bandwidth fluctuation, performing adaptive gain adjustment of transmit power for signal attenuation, performing spectrum calibration and correction for frequency offset, performing link queue optimization for link delay jitter, and performing iterative adjustment of deep noise reduction parameters for noise intrusion.

[0025] The advantages of this embodiment compared to existing technologies are as follows: by extracting five complementary features, including sharpness, signal-to-noise ratio, spectral distortion, temporal distortion, and coherence, and using Bayesian networks for probabilistic inference, the risk of misjudgment caused by traditional hard threshold judgment is avoided, significantly improving the robustness and fault tolerance of the system; at the same time, the causal logic structure of Bayesian networks allows the diagnostic results to be directly traced to specific observation indicators, providing clear interpretability and traceability, making it easier for maintenance personnel to quickly locate the root cause of the fault.

[0026] Example 2

[0027] like Figure 3 As shown, the communication data evaluation system based on language analysis is implemented based on the above-mentioned communication data evaluation method based on language analysis. The communication preprocessing module is used to collect voice and language communication data transmitted through the communication link in real time, preprocess the raw communication language data, and obtain standardized language analysis signals. The feature analysis module performs multi-dimensional language analysis on standardized language analysis signals based on speech language feature dimensions, extracting language profundity features, speech signal-to-noise ratio features, language temporal distortion features, frequency spectrum features, and speech coherence features to construct a communication language quality feature dataset. The quality analysis module is used to preset the threshold for evaluating the quality of communication data. It matches and compares the extracted multi-dimensional language feature data with the preset threshold to quantitatively determine the overall quality level of the current communication data. If the source analysis module determines that the communication data quality is in a poor or abnormal state, it retrieves the real-time spectrum data and time-frequency feature map corresponding to the current language, and performs source analysis through spectrum distortion features, frequency attenuation range, harmonic interference features, and spectrum aliasing features to intelligently predict and analyze the causes of faults affecting communication quality. The strategy matching module is used to sort the various influencing factors obtained from the source tracing by weight, filter out the main influencing factors that lead to the degradation of communication quality, and match the preset fault-corresponding reminder strategy and adaptive optimization handling strategy.

[0028] Example 3

[0029] This embodiment provides an electronic device, including: a processor and a memory, wherein the memory stores a computer program that can be called by the processor; The processor executes the aforementioned language analysis-based communication data evaluation method by calling computer programs stored in memory.

[0030] The electronic device can vary considerably depending on its configuration or performance. It may include one or more processors and one or more memories, wherein the memory stores at least one computer program, which is loaded and executed by the processor to implement the language analysis-based communication data evaluation method provided in the above-described embodiment of the language analysis-based communication data evaluation method. The electronic device may also include other components for implementing device functions; for example, it may have wired or wireless network interfaces and input / output interfaces for data input and output. Details will not be elaborated upon in this embodiment.

[0031] Example 4

[0032] This embodiment proposes a computer-readable storage medium on which an erasable and rewritable computer program is stored. When a computer program runs on a computer device, it causes the computer device to perform the aforementioned language analysis-based communication data evaluation method.

[0033] For example, computer-readable storage media can be read-only memory, random access memory, read-only optical disc, magnetic tape, floppy disk, and optical data storage devices.

[0034] The above embodiments can be implemented, in whole or in part, by software, hardware, firmware, or any other combination thereof. When implemented using software, the above embodiments can be implemented, in whole or in part, as a computer program product. A computer program product includes one or more computer instructions or computer programs. When the computer instructions or computer programs are loaded or executed on a computer, all or part of the processes or functions according to the embodiments of this application are generated. The computer can be a general-purpose computer, a special-purpose computer, a computer network, or other programmable device. Computer instructions can be stored in a computer-readable storage medium or transmitted from one computer-readable storage medium to another. For example, computer instructions can be transmitted from one website, computer, server, or data center to another website, computer, server, or data center via a wired network and / or wireless network. A computer-readable storage medium can be any available medium that a computer can access or a data storage device such as a server or data center that includes one or more sets of available media. The available medium can be a magnetic medium (e.g., a floppy disk), an optical medium (e.g., a DVD), or a semiconductor medium. A semiconductor medium can be a solid-state drive.

Claims

1. A communication data evaluation method based on language analysis, characterized in that, It includes the following specific steps: S1. Real-time acquisition of voice and language communication data transmitted through the communication link, preprocessing of the raw communication language data to obtain standardized language analysis signals; S2. Perform multi-dimensional language analysis on standardized language analysis signals based on speech language feature dimensions, extract language proficiency features, speech signal-to-noise ratio features, language temporal distortion features, frequency spectrum features, and speech coherence features, and construct a communication language quality feature dataset. S3. Set a preset threshold for evaluating the quality of communication data. Match and compare the extracted multi-dimensional language feature data with the preset threshold to quantitatively determine the overall quality level of the current communication data. S4. If the communication data quality is determined to be poor and abnormal, retrieve the real-time spectrum data and time-frequency feature map corresponding to the current language, and conduct source analysis through spectrum distortion features, frequency attenuation range, harmonic interference features, and spectrum aliasing features to intelligently predict and analyze the causes of faults affecting communication quality. S5. Rank the various influencing factors obtained from the source tracing by weight, filter out the main influencing factors that lead to the degradation of communication quality, and match the preset fault-corresponding reminder strategy and adaptive optimization handling strategy.

2. The communication data evaluation method based on language analysis as described in claim 1, characterized in that, The real-time acquisition of voice and language communication data transmitted through the communication link, and the preprocessing of the raw communication language data to obtain standardized language analysis signals, include the following specific steps: S11. By integrating multiple types of physical interfaces, non-intrusive bypass acquisition and low-latency mirror capture of voice payload data in the communication link are performed, analog-to-digital conversion of voice signals is performed, and data packets are decoded and restored in real time according to standard encoding and decoding formats. S12. Perform noise reduction filtering on various types of noise mixed in the acquired raw speech signal, perform sliding window framing processing on the continuous time domain signal, and generate a standardized language analysis signal.

3. The communication data evaluation method based on language analysis as described in claim 2, characterized in that, The construction of the communication language quality feature dataset includes the following specific steps: S21. The standardized speech analysis signal is filtered through an octave bandpass filter bank to separate the time-domain envelope signal in each frequency band. A second Fourier transform is performed on the envelope signal of each frequency band to extract its modulation depth at the modulation frequency. The modulation retention rate of each frequency band is weighted and integrated with the standard human hearing weight function to obtain the speech transmission index. The clarity index is calculated according to the ANSIS 3.5 standard. The speech transmission index and the clarity index are normalized and then weighted to calculate the comprehensive clarity score. S22. Using the minimum recursive averaging algorithm, the minimum power spectral density value of each frequency point is continuously tracked in the frequency domain as the real-time noise basis for that frequency point. A decision-guided method is then employed, combining the posterior signal-to-noise ratio of the previous frame with the transient power spectrum of the current frame, to calculate the prior signal-to-noise ratio of the current frame. ,in, Let be the prior signal-to-noise ratio at the k-th frequency point in the m-th frame. As a smoothing factor, This is the power spectrum of the clean speech at the k-th frequency point in the previous frame. This is the noise power spectrum at the k-th frequency point of the previous frame. Let be the posterior signal-to-noise ratio of the m-th frame and the k-th frequency point. This is a half-wave rectification correction term; S23. Extraction of temporal distortion features of speech; specifically, it includes extracting the pitch period and peak amplitude of each frame using the autocorrelation method, calculating the absolute average of the pitch period change rate between adjacent effective voiced frames, and the absolute average of the difference between adjacent peak amplitudes; waveform symmetry and clipping coefficient: calculating the area ratio of the positive and negative half-cycles of the temporal signal to obtain the asymmetric distortion, and simultaneously calculating the proportion of samples whose peak amplitude exceeds the linear dynamic range threshold to generate the clipping distortion index; edge steepness loss: using the differential operator to calculate the maximum change rate between adjacent samples of the signal, comparing it with the slope template of the standard reference signal, and extracting the edge passivation coefficient; S24. Frequency Spectrum Feature Extraction; Specific features include: Power Spectral Density and Spectral Centroid: Perform Fourier transform on each windowed signal frame to calculate the power spectrum; extract the spectral centroid, and its offset directly quantifies the degree of frequency drift; Band Energy Ratio: Divide the effective frequency band into low-frequency, mid-frequency, and high-frequency bands; calculate the percentage of energy in each band to the total energy. When the energy ratio of the high-frequency band drops sharply, it is marked as a high-frequency attenuation feature; when the energy ratio of the mid-to-low frequency band increases abnormally, it is marked as a low-frequency humming interference feature; Spectral Flatness and Spectral Entropy: Calculate the ratio of geometric mean power to arithmetic mean power; Harmonic Energy Cluster Distribution: Extract the energy peak value and half-peak width at the fundamental frequency and its harmonics. If the harmonic peaks show broadening or splitting, it is quantitatively determined as the degree of spectral aliasing distortion. S25. Speech coherence feature extraction; specific features include: long-term autocorrelation decay coefficient: calculate the decay slope of the normalized autocorrelation function envelope of the standardized speech analysis signal in the lag interval of 100ms to 1000ms; due to the periodicity of voiced sounds, the autocorrelation of normal speech decays slowly; if there is a sudden packet loss in the link, the autocorrelation envelope will drop sharply at the corresponding lag, and the drop amplitude is extracted as the temporal break index; speech rate and pause variation coefficient: using the silent interval index output by speech activation detection in S1, calculate the effective speech rate and pause duration distribution, compare the real-time statistical value with the standard speech rate template for dynamic time warping, and output the prosodic distortion degree to determine whether there is speech speed loss due to time delay jitter; formant trajectory continuity: extract the time-varying trajectory of the first three formants through linear predictive coding analysis; calculate the rate of change of formant frequency between adjacent frames, if the rate of change exceeds the physiological limit of normal pronunciation, it is judged as an abnormal jump, and marked as discontinuity caused by channel switching or sudden change of encoding and decoding parameters. S26. Generate original feature vectors for the various indicators in steps S21-S25. Calculate the mean and standard deviation of each feature independently over the entire communication session time period, and perform standardization processing to obtain standardized feature vectors.

4. The communication data evaluation method based on language analysis as described in claim 3, characterized in that, The quantitative determination of the overall quality level of the current communication data includes the following specific steps: The comprehensive score is calculated by weighting each feature in the standardized feature vector. Based on the pre-set communication data quality grading threshold, the overall quality level of the current communication data is divided into high-quality communication status, general fluctuation status, and poor quality abnormal status. Communication data that is greater than or equal to the high-quality threshold is set to high-quality communication status, while communication data that is less than or equal to the low-quality threshold is set to low-quality abnormal status; communication data that falls between the high-quality and low-quality thresholds is set to normal fluctuation status.

5. The communication data evaluation method based on language analysis as described in claim 4, characterized in that, If the communication data quality is determined to be poor and abnormal, the real-time spectrum data and time-frequency feature map corresponding to the current language are retrieved. Source analysis is performed using spectral distortion characteristics, frequency attenuation ranges, harmonic interference characteristics, and spectral aliasing characteristics to intelligently predict and analyze the causes of faults affecting communication quality. The specific steps include: S41. For communication data that is determined to be in a poor or abnormal state, retrieve the real-time spectrum data and time-frequency feature map corresponding to the current language, use short-time Fourier transform to generate a three-dimensional time-frequency-energy spectrum matrix, calculate the absolute mean of the third derivative of the power spectrum envelope, and denot it as the spectrum envelope roughness. By detecting local minima in the spectrum envelope, extract its depth relative to adjacent peaks and 3dB notch bandwidth. S42. Divide the effective frequency band into equal-width sub-bands, calculate the average power for each sub-band, fit the frequency-power linear regression curve using the least squares method, obtain the attenuation slope, and find the frequency point corresponding to the first drop of the power spectral density to -20dB of the average power in the low-frequency band as the cutoff frequency. S43. Extract the fundamental frequency using the cepstral method to check if there is a peak at the harmonic frequency. Calculate the ratio of the harmonic peak value to the mean of the adjacent frequency points and record it as the harmonic structure integrity index. S44. For frequency bands higher than the Nyquist frequency, check whether there is a mirror mapping of the peak in the original effective frequency band. If a mirror component is detected, it is determined that there is undersampling or spectral aliasing caused by the failure of the anti-aliasing filter. Input all the extracted quantization indicators into the pre-trained Bayesian network, calculate the posterior probability for each potential fault cause, and output the probability of each fault.

6. The communication data evaluation method based on language analysis as described in claim 5, characterized in that, The process of ranking the various influencing factors obtained from the source tracing by weight, filtering out the main influencing factors that lead to the degradation of communication quality, and matching the preset fault-corresponding reminder strategy and adaptive optimization handling strategy includes the following specific steps: Based on the current link type, service period, and user service level agreement, the probability of failure is dynamically loaded and weighted. The failure with the highest probability is identified as the primary cause of failure. Pre-set failure-related alert strategies and adaptive optimization handling strategies are matched to generate targeted communication quality anomaly alerts and push them to the communication management terminal. Corresponding adaptive handling measures are executed according to the primary cause of failure, including frequency band switching and interference shielding for channel interference, dynamic bandwidth resource allocation for bandwidth fluctuations, adaptive gain adjustment of transmit power for signal attenuation, spectrum calibration and correction for frequency offset, link queue optimization for link delay jitter, and iterative adjustment of deep noise reduction parameters for noise intrusion.

7. A communication data evaluation system based on language analysis, implemented based on the communication data evaluation method based on language analysis as described in any one of claims 1-6, characterized in that, Specifically, it includes: The communication preprocessing module is used to collect voice and language communication data transmitted through the communication link in real time, preprocess the raw communication language data, and obtain standardized language analysis signals. The feature analysis module performs multi-dimensional language analysis on standardized language analysis signals based on speech language feature dimensions, extracting language profundity features, speech signal-to-noise ratio features, language temporal distortion features, frequency spectrum features, and speech coherence features to construct a communication language quality feature dataset. The quality analysis module is used to preset the threshold for evaluating the quality of communication data. It matches and compares the extracted multi-dimensional language feature data with the preset threshold to quantitatively determine the overall quality level of the current communication data. If the source analysis module determines that the communication data quality is in a poor or abnormal state, it retrieves the real-time spectrum data and time-frequency feature map corresponding to the current language, and performs source analysis through spectrum distortion features, frequency attenuation range, harmonic interference features, and spectrum aliasing features to intelligently predict and analyze the causes of faults affecting communication quality. The strategy matching module is used to sort the various influencing factors obtained from the source tracing by weight, filter out the main influencing factors that lead to the degradation of communication quality, and match the preset fault-corresponding reminder strategy and adaptive optimization handling strategy.

8. An electronic device, comprising: A processor and a memory, wherein the memory stores a computer program that can be called by the processor; The processor is characterized in that it executes the communication data evaluation method based on language analysis as described in any one of claims 1-6 by calling a computer program stored in the memory.

9. A computer-readable storage medium, characterized in that, The system stores instructions that, when executed on a computer, cause the computer to perform the communication data evaluation method based on language analysis as described in any one of claims 1-6.

Images