An artificial intelligence-based black broadcast semantic automatic identification system and method

Abstract

This invention discloses an AI-based automatic semantic recognition system and method for illegal radio broadcasts, relating to the field of intelligent signal processing technology. The invention captures broadcast signals using dedicated equipment, analyzes spectral features using FFT technology, generates anomaly scores, and marks suspicious segments. It extracts speech segments using the VAD algorithm, transcribes them into text using ASR technology, and extracts acoustic and background sound features. The spectral, acoustic, and background sound features are fused with text information to construct a comprehensive feature vector. A BERT model is introduced to analyze emotional tone and mood patterns, and an LSTM model is used to parse logical relationships. Combining the results of both, an LLM model is used to determine potential intent and generate a semantic analysis report. By integrating the semantic analysis report and the spectral anomaly score, the degree of suspicion and potential harm of the illegal broadcast is quantitatively assessed, and corresponding response strategies are automatically selected and executed to achieve intelligent identification and handling of illegal broadcasts.

Description

Technical Field

[0001] This invention relates to the field of intelligent signal processing technology, specifically to an automatic semantic recognition system and method for illegal radio broadcasts based on artificial intelligence. Background Technology

[0002] With the popularization of radio technology, illegally set up black radio stations have become a social scourge that seriously affects social stability and the safety of people's property due to their wide range of dissemination, strong concealment, and harmful content.

[0003] Current technologies lack the ability to understand the deep semantics of speech content. Traditional methods typically employ keyword matching technology, identifying illegal content through pre-set sensitive word libraries. However, these methods easily overlook contextual information, failing to recognize synonyms, suggestive language, or analyze emotional tone and intonation patterns, resulting in low detection rates for highly concealed illegal content. Current technologies cannot analyze the logical relationships within continuous content, treating each sentence in isolation and lacking the ability to analyze logical chains between sentences and paragraphs, making it difficult to accurately assess the overall deceptiveness of the content. Furthermore, current technologies struggle to accurately determine the underlying intent of broadcasts, relying on simple thresholds or manual verification, leading to high false positive and false negative rates. They cannot differentiate between different types of harm, such as false advertising and illegal fundraising, and the generated analysis reports lack depth, failing to support efficient and accurate law enforcement decisions. Summary of the Invention

[0004] The purpose of this invention is to provide an automatic semantic recognition system and method for illegal radio broadcasts based on artificial intelligence, so as to solve the problems raised in the prior art.

[0005] To achieve the above objectives, the present invention provides the following technical solution:

[0006] In a first aspect, the present invention provides an automatic semantic recognition method for black broadcasts based on artificial intelligence, comprising:

[0007] Using dedicated receiving equipment, radio spectrum is continuously monitored to capture broadcast signals; FFT digital signal processing technology is used to analyze the spectral characteristics of the captured signals in real time, generate spectral anomaly scores, and mark suspicious signal segments;

[0008] For suspicious signal segments marked by spectrum analysis, the speech segments are identified and extracted using a speech activity detection algorithm based on short-time energy and zero-crossing rate. ASR automatic speech recognition technology is then applied to transcribe the speech content, obtain the text information of the speech content, and extract acoustic features and background sound features.

[0009] The extracted spectral features, acoustic features, background noise features, and text information output by ASR are fused to construct a comprehensive feature vector;

[0010] We introduce the BERT contextual semantic understanding model and combine it with comprehensive feature vectors to analyze the emotional color and tone patterns in speech;

[0011] Based on the analysis results of emotional color and tone patterns, the logical relationships between sentences and paragraphs in continuously broadcast content are understood through the LSTM time series analysis model.

[0012] By combining the obtained emotional color, tone pattern and logical relationship, and based on the LLM model, the potential intention of the broadcast content is determined and a semantic analysis report is generated.

[0013] Based on the semantic analysis report and spectrum anomaly score, an overall status assessment of the current broadcast signal is formed and quantified to reflect the degree of suspicion of the broadcast signal as a black broadcast and the level of potential harm.

[0014] Based on the degree of suspicion of illegal radio broadcasts and the level of potential harm, select a response strategy and execute corresponding actions.

[0015] In conjunction with the first aspect, in the first implementation of the first aspect of this application, the step of continuously monitoring the radio spectrum and capturing broadcast signals using a dedicated receiving device, and then using FFT digital signal processing technology to analyze the spectral characteristics of the captured signals in real time, generating a spectral anomaly score, and marking suspicious signal segments, includes:

[0016] A broadband radio frequency receiver is used to scan and receive the target frequency band. After the received radio frequency signal is preprocessed, it is sampled into a digital baseband signal stream using a high-speed analog-to-digital converter. The sliding window technique is used for frame processing. After processing each frame signal with the Hanning window function, a fast Fourier transform is performed to convert the time domain signal into a frequency domain signal, and the power spectral density of the signal is calculated in real time.

[0017] For each frame of spectral data after Fast Fourier Transform and power spectral density calculation, multi-dimensional spectral features are extracted; the sum of the power values ​​of all frequency points in the current frame spectrum is calculated to obtain the total signal power; the frequency point with the highest power in the current frame spectrum is found, and its corresponding power value is the main frequency power; a power threshold is set, and the frequency range occupied by all frequency points with power exceeding this threshold is calculated to obtain the power bandwidth; the ratio of the variance to the mean of the power value of the current frame spectrum is calculated to obtain the spectral flatness; a weighted average value is calculated, with the weight of the power value of each frequency point, to characterize the centroid position of the spectral energy, thus obtaining the spectral centroid.

[0018] Collect spectral feature data known to be normal background noise or legitimate broadcast signals, and train a Gaussian mixture model. Input the multidimensional feature vector extracted from the current frame into the trained Gaussian mixture model, calculate the probability p that the feature vector belongs to the normal distribution, and convert the probability p into anomaly score S using the following formula:

[0019] ;

[0020] The calculated anomaly score S of the current frame is compared with a preset dynamic threshold. When the anomaly score is higher than the threshold, the corresponding frame signal is determined to be abnormal and marked as a suspicious signal segment.

[0021] In conjunction with the first aspect, in the second implementation of the first aspect of this application, the suspicious signal segments marked by spectrum analysis are identified and extracted using a speech activity detection algorithm based on short-time energy and zero-crossing rate. ASR (Automatic Speech Recognition) technology is then applied to transcribe the speech content, obtaining the text information of the speech content and extracting acoustic features and background noise features, including:

[0022] Calculate the short-time energy of each frame of signal, which is the sum of the squares of the amplitudes of all sampling points within the corresponding frame; calculate the zero-crossing rate of each frame of signal, which is the number of times the signal waveform crosses the zero level within the corresponding frame.

[0023] In the initial silent segment of the signal segment, the short-time energy and zero-crossing rate of the background noise are calculated and updated in real time using a smooth update method, and a dynamic threshold is set. The short-time energy and zero-crossing rate of each frame are compared with the dynamic threshold. When the short-time energy is higher than the energy threshold and the zero-crossing rate is within the normal speech range, the corresponding frame is determined to be a speech frame. Continuous speech frames are merged to extract speech segments containing complete speech content.

[0024] The extracted speech segments are input into a pre-trained ASR model. The ASR model is based on the Transformer architecture and converts the audio signal into a text sequence to obtain the text information of the speech content. For each speech segment, acoustic features are extracted, including speech rate features, pitch features, volume features, and prosodic features. Background sound analysis is performed on the non-speech parts of the speech segments to extract background sound features, including music type features, rhythm features, and loop cue sound features.

[0025] In conjunction with the first aspect, in the third implementation of the first aspect of this application, the step of fusing the extracted spectral features, acoustic features, background sound features, and text information output by ASR to construct a comprehensive feature vector includes:

[0026] The text information output by ASR is preprocessed, and each word is converted into a low-dimensional dense vector using a pre-trained BERT context-dependent language model. The vectors of all words in a sentence or text segment are weighted and summed through an attention mechanism to generate a text semantic vector.

[0027] The spectral features, acoustic features, and background sound features are normalized separately. The normalized spectral feature vector, acoustic feature vector, background sound feature vector, and vectorized text semantic vector are then concatenated to form a long vector. An attention weighting mechanism is applied to the long vector formed by the concatenation of the first and last parts, and the weighted sum is calculated according to the importance of different modal features in the black radio identification task to generate a comprehensive feature vector.

[0028] In conjunction with the first aspect, in the fourth implementation of the first aspect of this application, the introduction of the BERT contextual semantic understanding model, combined with comprehensive feature vectors, to analyze the emotional color and tone patterns in speech includes:

[0029] A BERT model pre-trained on a general large-scale text corpus was selected as the foundation to construct a training dataset for black radio broadcast recognition with annotation information. The constructed training dataset includes speech samples and their corresponding text information, and is labeled with sentiment tags and tone tags. Features were extracted from the constructed training dataset to generate corresponding comprehensive feature vectors, which were input into the initialized BERT model. Through the backpropagation algorithm, the parameters of the BERT model and downstream classification tasks were jointly optimized to obtain a BERT contextual semantic understanding model with dedicated semantic understanding capabilities. Two parallel downstream task classification heads were integrated at the top of the model, including a sentiment color classification head and a tone pattern classification head. Each classification head contains a fully connected layer, and the number of output neurons is equal to the number of preset sentiment categories and tone categories.

[0030] The comprehensive feature vector of the speech to be recognized is input into the BERT contextual semantic understanding model. BERT generates a context vector representing the overall semantics of the speech segment to be recognized through a multi-layer Transformer encoder. This context vector is simultaneously fed into the sentiment color classification head and the tone pattern classification head. The fully connected layer of each classification head performs a linear transformation on the input vector by performing matrix multiplication of the input vector with a learnable weight matrix and adding the result to a learnable bias vector to generate the original score vector. The original score vectors output by the two classification heads are applied with the Softmax activation function to transform them into a probability distribution in which all elements are positive and the sum is 1. From the probability distribution generated by Softmax, the category with the highest probability value is selected as the final sentiment label and tone label of the speech segment to be recognized, and the probability value is used as the confidence score of the corresponding label.

[0031] In conjunction with the first aspect, in the fifth implementation of the first aspect of this application, the analysis results based on emotional coloring and tone patterns, using an LSTM time series analysis model, are used to understand the logical relationships between sentences and paragraphs of continuously broadcast content, including:

[0032] A labeled training dataset of logical relationships in illegal broadcasts is constructed, including complete broadcast content confirmed as illegal and legitimate broadcasts. Comprehensive feature vectors are obtained and concatenated to form a multidimensional temporal feature sequence. Logical relationship labels are added at the level of adjacent sentences and paragraphs. The multidimensional temporal feature sequence and logical relationship labels are used as input to an LSTM model. The cross-entropy loss between the model's predicted logical relationships and the true labels is calculated using the backpropagation algorithm. The parameters of the LSTM model are jointly optimized using the gradient of the cross-entropy loss to obtain an LSTM model that can identify logical relationships in continuously broadcast content.

[0033] The continuously broadcast content to be identified is formed into a multi-dimensional temporal feature sequence and input into a trained LSTM model. The model outputs a hidden state vector at each time step, which encodes the context information up to the current time step. At the sentence level, the model performs linear transformation and Softmax classification on the hidden state vectors of two adjacent time steps, and outputs the probability of belonging to each type of logical relationship between the time steps. At the paragraph level, the model performs the same processing on the hidden state vector sequences representing different paragraphs. Based on the output probability distribution, the logical relationship between sentences and paragraphs is identified.

[0034] In conjunction with the first aspect, in the sixth implementation of the first aspect of this application, the combined emotional tone, mood pattern, and logical relationship are used to determine the underlying intent of the broadcast content based on an LLM model, and a semantic analysis report is generated, including:

[0035] The dataset consists of text information, sentiment tags, tone tags, logical relationship tags, corresponding potential intent of broadcast content, and semantic analysis reports from labeled black broadcast and normal broadcast samples. The LLM model is trained using a fine-tuning method. After training, for broadcast content to be identified, the text information, sentiment tags, tone tags, and logical relationship tags are input into the trained LLM model, which then determines the potential intent of the broadcast content and generates a semantic analysis report.

[0036] In conjunction with the first aspect, in the seventh implementation of the first aspect of this application, the step of forming an overall state assessment of the current broadcast signal based on the semantic analysis report and spectrum anomaly score, and quantifying it to reflect the degree of suspicion of the broadcast signal as a black broadcast and its potential harm level, includes:

[0037] The semantic analysis report generated by LLM is structured and parsed to extract the number of keyword triggers, sentiment tags, tone tags, number of logical contradictions, and potential intent. By comparing these with the corresponding preset thresholds, keyword scores, sentiment color scores, tone pattern scores, logical relationship scores, and potential intent scores are obtained. The semantic suspicion score is calculated by weighted summation to indicate the degree of suspicion of the broadcast signal as a black broadcast.

[0038] The spectrum anomaly score is normalized and weighted and summed with the semantic suspicion score to obtain the comprehensive suspicion score of the broadcast signal. This comprehensive suspicion score is then compared with a preset suspicion score threshold range to obtain the potential hazard level.

[0039] In conjunction with the first aspect, in the eighth implementation of the first aspect of this application, the step of selecting a response strategy and performing corresponding operations based on the degree of suspicion of black broadcasting and the level of potential harm includes:

[0040] A multi-level response strategy corresponding to different levels of suspicion and harm is predefined to form a strategy library. The calculated level of suspicion and potential harm of the black radio broadcast is used as input and matched with the strategy conditions in the strategy library. The highest priority response strategy is automatically selected and executed.

[0041] When the signal is considered low-susceptibility and low-hazard, an alarm notification is triggered, a prompt message is pushed to the responsible personnel, a continuous tracking mechanism is initiated, and the sampling and analysis frequency of the signal is increased. When the signal is considered medium-susceptibility and medium-hazard, a structured alarm work order containing a semantic analysis report and audio is automatically generated based on the alarm notification and pushed to the superior regulatory department through a preset interface for manual judgment and decision-making. When the signal is considered high-susceptibility and high-hazard, while executing the alarm-level operation, a directional interference command is issued to the radio management equipment or interference suppression equipment in the signal transmission area according to preset geofencing rules or authorized instructions to temporarily suppress the power or shift the frequency of the target signal, rendering it ineffective.

[0042] Secondly, the present invention provides an artificial intelligence-based automatic semantic recognition system for black broadcasts, comprising:

[0043] The signal acquisition and spectrum analysis module includes an RF receiving unit, a framing and FFT processing unit, a feature extraction unit, and an anomaly detection unit. The RF receiving unit scans the target frequency band, captures the RF signal, and converts it into a digital baseband signal stream. The framing and FFT processing unit performs framing and windowing on the digital baseband signal stream, executes a Fast Fourier Transform (FFT) to convert the time-domain signal into a frequency-domain signal, and calculates the power spectral density. The feature extraction unit extracts multi-dimensional spectral features. The anomaly detection unit calculates a spectral anomaly score based on a Gaussian mixture model and uses dynamic thresholds to determine suspicious signal segments.

[0044] The speech content extraction and feature engineering module includes a speech activity detection unit, a speech conversion unit, an acoustic feature extraction unit, and a background noise analysis unit. The speech activity detection unit detects speech frames using short-time energy and zero-crossing rate dynamic thresholds, and merges continuous speech segments. The speech conversion unit, based on an ASR model with a Transformer architecture, converts speech into text and outputs the text information. The acoustic feature extraction unit extracts acoustic features. The background noise analysis unit analyzes background features of non-speech components.

[0045] The multimodal feature fusion module includes a text vectorization unit, a feature normalization unit, and an attention fusion unit. The text vectorization unit uses the BERT model to convert the text output by ASR into a semantic vector. The feature normalization unit standardizes the spectral, acoustic, and background sound features. The attention fusion unit generates a comprehensive feature vector by weighted summation and concatenation of the multimodal features.

[0046] The semantic understanding module includes a sentiment analysis unit, a tone pattern analysis unit, and a logical relationship analysis unit. The sentiment analysis unit outputs sentiment labels and confidence scores based on the BERT sentiment classification head; the tone pattern analysis unit outputs tone labels and confidence scores based on the BERT tone classification head; and the logical relationship analysis unit identifies temporal logical relationships between sentences and paragraphs using an LSTM model.

[0047] The intent analysis and report generation module includes an LLM intent reasoning unit and a semantic report generation unit. The LLM intent reasoning unit takes text, sentiment, tone, and logical relationship tags as input and generates potential intents from the fine-tuned LLM. The semantic report generation unit outputs a structured semantic analysis report.

[0048] The comprehensive assessment and decision-making module includes a quantitative assessment unit, a strategy matching unit, and an execution control unit. The quantitative assessment unit calculates the spectral anomaly score and semantic suspicion score with weights, and outputs the comprehensive suspicion score and hazard level. The strategy matching unit matches multi-level response strategies based on the suspicion level and hazard level. The execution control unit triggers operations in the strategy library.

[0049] Compared with the prior art, the beneficial effects of the present invention are:

[0050] 1. This invention introduces the BERT contextual semantic understanding model and combines it with comprehensive feature vectors to analyze the emotional color and tone patterns in speech.

[0051] 2. Based on the analysis results of emotional color and tone patterns, this invention uses an LSTM time series analysis model to understand the logical relationships between sentences and paragraphs in continuously broadcast content.

[0052] 3. This invention combines the obtained emotional color, tone pattern and logical relationship, and based on the LLM model, judges the potential intent of the broadcast content and generates a semantic analysis report. Attached Figure Description

[0053] Figure 1 This is a schematic diagram illustrating the steps of an AI-based automatic semantic recognition method for black radio broadcasts according to the present invention.

[0054] Figure 2This is a system structure diagram of an AI-based automatic semantic recognition system for black radio broadcasts according to the present invention.

[0055] Figure 3 This is a flowchart of an AI-based method for automatic semantic recognition of illegal radio broadcasts according to the present invention. Detailed Implementation

[0056] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0057] Example: Figures 1-3 As shown, the present invention provides a technical solution.

[0058] like Figure 1 A schematic diagram illustrating the steps of an AI-based method for automatic semantic recognition of illegal radio broadcasts is provided. This invention offers an AI-based method for automatic semantic recognition of illegal radio broadcasts, comprising:

[0059] Step S100: Continuously monitor the radio spectrum and capture broadcast signals using a dedicated receiving device; analyze the spectral characteristics of the captured signals in real time using FFT digital signal processing technology, generate spectral anomaly scores, and mark suspicious signal segments;

[0060] Specifically, a broadband radio frequency receiver is used to scan and receive the target frequency band. After the received radio frequency signal is preprocessed, it is sampled into a digital baseband signal stream using a high-speed analog-to-digital converter. The sliding window technique is used for frame processing. After processing each frame of signal with the Hanning window function, a fast Fourier transform is performed to convert the time domain signal into a frequency domain signal, and the power spectral density of the signal is calculated in real time.

[0061] For each frame of spectral data after Fast Fourier Transform and power spectral density calculation, multi-dimensional spectral features are extracted; the sum of the power values ​​of all frequency points in the current frame spectrum is calculated to obtain the total signal power; the frequency point with the highest power in the current frame spectrum is found, and its corresponding power value is the main frequency power; a power threshold is set, and the frequency range occupied by all frequency points with power exceeding this threshold is calculated to obtain the power bandwidth; the ratio of the variance to the mean of the power value of the current frame spectrum is calculated to obtain the spectral flatness; a weighted average value is calculated, with the weight of the power value of each frequency point, to characterize the centroid position of the spectral energy, thus obtaining the spectral centroid.

[0062] Collect spectral feature data known to be normal background noise or legitimate broadcast signals, and train a Gaussian mixture model. Input the multidimensional feature vector extracted from the current frame into the trained Gaussian mixture model, calculate the probability p that the feature vector belongs to the normal distribution, and convert the probability p into anomaly score S using the following formula:

[0063] ;

[0064] The calculated anomaly score S of the current frame is compared with a preset dynamic threshold. When the anomaly score is higher than the threshold, the corresponding frame signal is determined to be abnormal and marked as a suspicious signal segment.

[0065] In one specific embodiment, a device deployed in a certain area of ​​the city is responsible for monitoring the 88-108MHz FM broadcast band. A broadband RF receiver with a bandwidth of 20MHz continuously scans and receives the target frequency band at a sampling rate of 10MSPS. The received RF signal undergoes preprocessing such as low-noise amplification and filtering before being sampled into a digital baseband signal stream by a high-speed analog-to-digital converter. A sliding window technique is used to divide the continuous signal stream into frames of 1024 sampling points each, with 512 sampling points between frames. The overlapping of sample points ensures the continuity of the signal. For each frame of signal, the Hanning window function is applied to reduce spectral leakage. A fast Fourier transform is performed to convert it from the time domain to the frequency domain, and the power spectral density of 1024 frequency points in the frame is calculated in real time. Four-dimensional features are extracted for each frame. The total power of a certain frame is measured to be -42.8dBm, the main frequency is located at 101.1MHz, and its power is -28.3dBm. A power threshold of -70dBm is set, the power bandwidth is 350kHz, and the spectral flatness is 12.1.

[0066] 1000 hours of signal frames known to be normal background noise and legitimate broadcasts were collected in advance. After extracting the aforementioned four-dimensional features, a Gaussian mixture model (GMM) containing three Gaussian components was trained. The feature vector of a frame extracted in real time [-42.8, -28.3, 350, 12.1] was input into the trained GMM. The probability p that the vector belongs to the normal distribution was calculated to be 0.65. The probability p was converted into an anomaly score S using a formula, and S was calculated to be 0.35. The preset dynamic threshold was 0.2. Since the anomaly score of 0.35 for the current frame was higher than the threshold, the signal of this frame was determined to be abnormal. The signal segments corresponding to the 512 sampling points starting from the start time of this frame were immediately marked as suspicious signal segments on the time axis.

[0067] Step S200: For the suspicious signal segments marked by spectrum analysis, the speech segments are identified and extracted using a speech activity detection algorithm based on short-time energy and zero-crossing rate. The ASR automatic speech recognition technology is then applied to transcribe the speech content, obtain the text information of the speech content, and extract acoustic features and background sound features.

[0068] Specifically, the short-time energy of each frame of signal is calculated, which is the sum of the squares of the amplitudes of all sampling points within the corresponding frame; the zero-crossing rate of each frame of signal is calculated, which is the number of times the signal waveform crosses the zero level within the corresponding frame.

[0069] In the initial silent segment of the signal segment, the short-time energy and zero-crossing rate of the background noise are calculated and updated in real time using a smooth update method, and a dynamic threshold is set. The short-time energy and zero-crossing rate of each frame are compared with the dynamic threshold. When the short-time energy is higher than the energy threshold and the zero-crossing rate is within the normal speech range, the corresponding frame is determined to be a speech frame. Continuous speech frames are merged to extract speech segments containing complete speech content.

[0070] The extracted speech segments are input into a pre-trained ASR model. The ASR model, based on the Transformer architecture, converts the audio signal into a text sequence to obtain the textual information of the speech content. For each speech segment, acoustic features are extracted. The total number of effective words in the speech segment is calculated using the ASR output text and its corresponding timestamp information, and divided by the total duration of the speech segment to obtain the speech rate feature. For each speech frame, the fundamental frequency value of that frame is calculated using the autocorrelation function method. The fundamental frequency values ​​of all frames in the entire speech segment are concatenated to form a fundamental frequency curve. The statistical features of this curve are calculated, including the average fundamental frequency. Tonal features are obtained by analyzing the fundamental frequency, fundamental frequency standard deviation, and fundamental frequency range. For each speech frame, its short-time energy is calculated, represented as the sum of squares of the amplitudes of the sampling points within the frame. After normalizing the short-time energy values ​​of all frames, its statistical features are calculated, including average volume, volume standard deviation, and volume dynamic range, to obtain volume features. Stress detection is performed by analyzing significant spikes in local fundamental frequencies and energy to identify stressed words. The overall contour of the fundamental frequency curve is analyzed to determine whether it is a gradual descent in declarative sentences, an ascent in interrogative sentences, or a strong fluctuation in exclamatory sentences. Combined with the results of speech activity detection, the duration and frequency of pauses between sentences or phrases are analyzed to obtain prosodic features.

[0071] Background sound analysis is performed on the non-speech parts of speech segments. Based on the VAD results, the non-speech parts are extracted to form independent background sound segments. The Mel-frequency cepstral coefficients of the background sound segments are extracted and input into a pre-trained ResNet-based audio classification model to classify the background sound into a preset category, thus obtaining music type feature extraction. Spectral analysis is performed on the background sound segments. In the spectrogram, the music energy exhibits periodic peaks, corresponding to the beat of the music. The peaks are located using the autocorrelation method, and the BPM value is calculated. The BPM value reflects the tempo of the background music, thus obtaining rhythm features. Black radio often uses looped prompts to enhance deception or create a sense of professionalism. STFT is performed on the background sound segments to obtain time-spectrum diagrams. The periodic patterns are analyzed using a clustering algorithm. Once a looped prompt is identified, its loop period and timbre features are extracted to determine whether it is a standardized sound identifier with a specific meaning, thus obtaining looped prompt features.

[0072] In one specific embodiment, a suspicious signal segment identified by spectrum analysis is processed. Using a VAD algorithm based on short-time energy and zero-crossing rate, the average short-time energy of the initial silent segment is calculated to be -50.1 dBm, and the average zero-crossing rate is 15. The background noise parameters are updated in real time using a smoothing update method. The dynamic threshold is set to an energy threshold of -38.0 dBm and a zero-crossing rate range of [20, 60]. A speech segment lasting approximately 6.2 seconds is successfully extracted. This speech segment is input into a pre-trained ASR model, and the transcribed text is: "Congratulations! Your mobile phone number has been selected as a lucky audience member. Simply deposit 9980 yuan into the designated account to activate the grand prize of a sea-view apartment worth 100,000 yuan. Don't miss this opportunity, please remit the money immediately!"

[0073] Acoustic features of the speech segment were extracted, and the total number of effective words was counted as 25. Dividing this by the duration of 6.2 seconds, the speech rate was found to be as high as approximately 4 words per second, far exceeding the 2.5-3.5 words per second of normal conversation, indicating a clear sense of urgency. The fundamental frequency curve was calculated using the autocorrelation method, yielding an average fundamental frequency of 120Hz and a standard deviation of 35Hz, suggesting the speaker was deliberately imitating excitement or a high-pitched tone. The volume characteristics were calculated, with an average volume of -15dB, but a very wide dynamic range of 25dB, especially noticeable with keywords such as "immediate remittance." A sudden increase in volume was detected; three deliberate pauses exceeding 0.5 seconds were also detected in the speech, used to create suspense and emphasis; the non-speech parts before and after the speech segment were analyzed, their MFCC features were extracted and input into the ResNet classification model, and the background sound was determined to be "exciting background music". STFT analysis showed that its BPM value was 140, with an extremely fast rhythm, further enhancing the sense of urgency; based on the above features, the speech segment was determined to have the typical characteristics of black radio broadcasts: "high speech rate, high volume fluctuation, and strong urging", awaiting further semantic logic analysis.

[0074] Step S300: The extracted spectral features, acoustic features, background sound features and text information output by ASR are fused to construct a comprehensive feature vector; the BERT contextual semantic understanding model is introduced, and combined with the comprehensive feature vector, the emotional color and tone pattern in the speech are analyzed;

[0075] Specifically, the text information output by ASR is preprocessed, and each word is converted into a low-dimensional dense vector using a pre-trained BERT context-dependent language model. The vectors of all words in a sentence or text segment are weighted and summed through an attention mechanism to generate a text semantic vector.

[0076] The spectral features, acoustic features, and background noise features are normalized separately. The normalized spectral feature vector, acoustic feature vector, background noise feature vector, and vectorized text semantic vector are then concatenated to form a long vector. An attention weighting mechanism is applied to the long vector formed by concatenation, and the weighted sum is calculated according to the importance of different modal features in the black radio identification task to generate a comprehensive feature vector.

[0077] A BERT model pre-trained on a general large-scale text corpus was selected as the foundation to construct a training dataset for black radio broadcast recognition with annotation information. The constructed training dataset includes speech samples and their corresponding text information, and is labeled with sentiment tags and tone tags. Features were extracted from the constructed training dataset to generate corresponding comprehensive feature vectors, which were input into the initialized BERT model. Through the backpropagation algorithm, the parameters of the BERT model and downstream classification tasks were jointly optimized to obtain a BERT contextual semantic understanding model with dedicated semantic understanding capabilities. Two parallel downstream task classification heads were integrated at the top of the model, including a sentiment color classification head and a tone pattern classification head. Each classification head contains a fully connected layer, and the number of output neurons is equal to the number of preset sentiment categories and tone categories.

[0078] The comprehensive feature vector of the speech to be recognized is input into the BERT contextual semantic understanding model. BERT generates a context vector representing the overall semantics of the speech segment to be recognized through a multi-layer Transformer encoder. This context vector is simultaneously fed into the sentiment color classification head and the tone pattern classification head. The fully connected layer of each classification head performs a linear transformation on the input vector by performing matrix multiplication of the input vector with a learnable weight matrix and adding the result to a learnable bias vector to generate the original score vector. The original score vectors output by the two classification heads are applied with the Softmax activation function to transform them into a probability distribution in which all elements are positive and the sum is 1. From the probability distribution generated by Softmax, the category with the highest probability value is selected as the final sentiment label and tone label of the speech segment to be recognized, and the probability value is used as the confidence score of the corresponding label.

[0079] In one specific embodiment, the text output by ASR, "Congratulations! Your mobile number has been selected as a lucky audience member. Simply deposit 9,980 yuan into the designated account to activate the grand prize of a sea-view room worth 100,000 yuan. Don't miss this opportunity, please remit the money immediately!", is preprocessed and input into the pre-trained BERT-base model to obtain a 768-dimensional vector for each word. The vector is then weighted and summed using BERT's built-in attention mechanism to generate a 768-dimensional text semantic vector. The normalized spectral features, acoustic features, and background sound features are then normalized. These four vectors are concatenated end to end to form a 976-dimensional long vector. A learnable attention weight layer is applied to this concatenated vector, and after weighted summation, it is compressed to generate a 512-dimensional comprehensive feature vector.

[0080] This comprehensive feature vector is input into a BERT model that has been fine-tuned on a dedicated dataset for identifying illegal radio broadcasts containing 100,000 samples. The top of this model is connected to both a sentiment classification head and a tone classification head. After processing by a multi-layer Transformer encoder in BERT, a context vector representing the overall semantics is generated, which is then fed into both classification heads simultaneously. The fully connected layer of the sentiment classification head linearly transforms the vector from 512 dimensions to 4 dimensions, obtaining the original score vector [-0.8, 2.5, 1.2, -2.1]. After applying Softmax, this is converted into a probability distribution [0.0...]. [8, 0.71, 0.18, 0.03], the maximum value of 0.71, corresponding to "enticement", was selected as the sentiment label with a confidence level of 71%; the fully connected layer of the tone classification head linearly transformed the vector from 512 dimensions to 3 dimensions, obtaining the original score vector [0.9, -0.5, 2.3], which was then converted into a probability distribution [0.21, 0.09, 0.70] after applying Softmax, and the maximum value of 0.70, corresponding to "command", was selected as the tone label with a confidence level of 70%; based on comprehensive analysis, the audio segment was judged to be a "command" tone with "enticement" sentiment, and is highly suspected to be black broadcast content.

[0081] Step S400: Based on the analysis results of emotional color and tone patterns, use the LSTM time series analysis model to understand the logical relationships between sentences and paragraphs of the continuously broadcast content; combine the obtained emotional color, tone patterns and logical relationships, and use the LLM model to determine the potential intent of the broadcast content and generate a semantic analysis report.

[0082] Specifically, a labeled training dataset of logical relationships in illegal broadcasts is constructed, including complete broadcast content confirmed as illegal and legitimate broadcasts. Comprehensive feature vectors are obtained and concatenated sequentially to form a multidimensional temporal feature sequence. Logical relationship labels are labeled at the level of adjacent sentences and paragraphs. The multidimensional temporal feature sequence and logical relationship labels are used as input to an LSTM model. Through the backpropagation algorithm, the cross-entropy loss between the model's predicted logical relationships and the true labels is calculated. The parameters of the LSTM model are jointly optimized using the cross-entropy loss gradient to obtain an LSTM model that can identify logical relationships in continuously broadcast content.

[0083] The continuously broadcast content to be identified is formed into a multi-dimensional temporal feature sequence and input into a trained LSTM model. At each time step, the model outputs a hidden state vector, which encodes the context information up to the current time step. At the sentence level, the hidden state vectors corresponding to adjacent sentences are taken and concatenated to obtain a high-dimensional vector. This concatenated high-dimensional vector is then input into a linear layer for feature extraction and dimensionality reduction, with the output dimension matching the number of logical relationship categories. The output of the linear layer is then passed through a Softmax function to convert it into a probability distribution, where the sum of all values ​​is 1. The category with the highest probability value is selected as the final recognition result. At the paragraph level, the hidden state vectors corresponding to adjacent paragraphs are taken, concatenated, and input into a classifier for the same processing as at the sentence level, yielding a probability distribution of logical relationships between paragraphs and identifying the logical relationships between them.

[0084] The dataset consists of text information, sentiment tags, tone tags, logical relationship tags, corresponding potential intent of broadcast content, and semantic analysis reports from labeled black broadcast and normal broadcast samples. The LLM model is trained using a fine-tuning method. After training, for broadcast content to be identified, the text information, sentiment tags, tone tags, and logical relationship tags are input into the trained LLM model, which then determines the potential intent of the broadcast content and generates a semantic analysis report.

[0085] In one specific embodiment, a comprehensive feature vector is extracted from each of the 2000 labeled illegal broadcasts and 800 normal broadcasts in the complete broadcast content. These vectors are then concatenated in the broadcast order to form a multidimensional temporal feature sequence. Eight logical relationship labels, including "cause and effect," "contrast," "parallel," and "progressive," are added at the adjacent sentence level and paragraph level to construct a training dataset. This dataset is then input into a model containing two layers of bidirectional LSTM units. Using backpropagation and cross-entropy as the loss function, after 100 epochs of training, the model achieves a logical relationship classification accuracy of 91.5% on the validation set.

[0086] The continuous broadcast message to be identified, "Congratulations! Your mobile number has been selected as a lucky viewer. Simply deposit 9980 yuan into the designated account to activate a grand prize of a sea-view apartment worth 100,000 yuan. Don't miss this opportunity! Please transfer the money immediately! Failure to do so will affect your personal credit!" is input into a trained LSTM model. At each time step, the model outputs a hidden state vector. At the sentence level, the vectors corresponding to "Congratulations!" and "Your mobile number..." are concatenated and input into the classifier. The Softmax output probability distribution is [0.05, 0.88, 0.02, 0.05], identifying a "causal" relationship with a confidence level of 88%. At the paragraph level, the vectors corresponding to the paragraphs inducing remittance and those intimidating and demanding payment are concatenated and input into the classifier. The Softmax output probability distribution... The values ​​are [0.03, 0.91, 0.02, 0.04], identifying a "transition" relationship with a confidence level of 91%. The above text information, the sentiment label "enticement", the tone label "command", the logical relationship labels "cause and effect" and "transition", and 200 labeled "intent-report" samples constitute an instruction fine-tuning dataset, which is input into an instruction-fine-tuned LLaMA-2-7B model. Based on the input label combination, the model determines that the potential intent is "financial fraud" and generates a semantic analysis report: "The broadcast content adopts the rhetorical structure of 'enticement-command', induces users to pre-deposit funds through the lie of 'winning the prize', and uses 'affecting credit score' as a scare transition. The logical chain is complete and it belongs to a typical financial fraud black broadcast. It is recommended to take blocking measures immediately."

[0087] Step S500: Based on the semantic analysis report and spectrum anomaly score, form an overall status assessment of the current broadcast signal, quantify it, and reflect the degree of suspicion of the broadcast signal being a black broadcast and the level of potential harm; based on the degree of suspicion of black broadcast and the level of potential harm, select a response strategy and execute the corresponding operation.

[0088] Specifically, the semantic analysis report generated by LLM is structured and parsed to extract the number of keyword triggers, sentiment tags, tone tags, number of logical contradictions, and potential intent. By comparing these with the corresponding preset thresholds, keyword scores, sentiment color scores, tone pattern scores, logical relationship scores, and potential intent scores are obtained. The semantic suspicion score is calculated by weighted summation, which indicates the degree of suspicion of the broadcast signal as a black broadcast.

[0089] The spectrum anomaly score is normalized and weighted and summed with the semantic suspicion score to obtain the comprehensive suspicion score of the broadcast signal. This comprehensive suspicion score is then compared with a preset suspicion score threshold range to obtain the potential hazard level.

[0090] A multi-level response strategy corresponding to different levels of suspicion and harm is predefined to form a strategy library. The calculated level of suspicion and potential harm of the black radio broadcast is used as input and matched with the strategy conditions in the strategy library. The highest priority response strategy is automatically selected and executed.

[0091] When the signal is considered low-susceptibility and low-hazard, an alarm notification is triggered, a prompt message is pushed to the responsible personnel, a continuous tracking mechanism is initiated, and the sampling and analysis frequency of the signal is increased. When the signal is considered medium-susceptibility and medium-hazard, a structured alarm work order containing a semantic analysis report and audio is automatically generated based on the alarm notification and pushed to the superior regulatory department through a preset interface for manual judgment and decision-making. When the signal is considered high-susceptibility and high-hazard, while executing the alarm-level operation, a directional interference command is issued to the radio management equipment or interference suppression equipment in the signal transmission area according to preset geofencing rules or authorized instructions to temporarily suppress the power or shift the frequency of the target signal, rendering it ineffective.

[0092] In one specific embodiment, a broadcast signal is processed, and its LLM-generated semantic analysis report is as follows: "The broadcast content contains keywords such as 'winning,' 'grand prize,' and 'remittance' a total of 8 times. The sentiment label is 'enticement,' and the tone label is 'command.' Logically, there are 2 contradictions between 'high returns' and 'pre-deposited funds,' and the potential intent is determined to be 'financial fraud.'" The report is then analyzed using preset scoring rules: keywords exceeding the threshold 5 times receive full marks (100 points); the sentiment color "enticement" corresponds to a high score (90 points); and the tone pattern "command" corresponds to a high score. The initial score is 95. Two logical inconsistencies resulted in a deduction of 80 points. The potential intent of "financial fraud" received the highest score of 100. Weighted summation of these scores (0.2, 0.2, 0.2, 0.15, 0.25) yielded a semantic suspicion score of 93.5. The normalized spectrum anomaly score was 88. Weighted summation of the semantic suspicion and spectrum anomaly scores (0.7 and 0.3 respectively) resulted in a comprehensive suspicion score of 91.45. This score falls within the pre-defined range of [80, 100], thus classifying the potential harm level as high.

[0093] Using the high suspicion level and high potential harm level of black broadcasts as input, the system matches the strategy library and automatically selects and executes the highest priority response strategy: while pushing high-level alarm work orders to the regulatory platform, it automatically issues directional interference instructions to radio management equipment in the target transmission area according to preset geofencing rules, temporarily shifting the frequency of the signal by 5MHz, making it ineffective within 30 seconds, and preventing potential financial fraud.

[0094] like Figure 2The system architecture diagram of an AI-based automatic semantic recognition system for illegal radio broadcasts is shown. This invention provides an AI-based automatic semantic recognition system for illegal radio broadcasts, comprising:

[0095] The signal acquisition and spectrum analysis module includes an RF receiving unit, a framing and FFT processing unit, a feature extraction unit, and an anomaly detection unit. The RF receiving unit scans the target frequency band, captures the RF signal, and converts it into a digital baseband signal stream. The framing and FFT processing unit performs framing and windowing on the digital baseband signal stream, executes a Fast Fourier Transform (FFT) to convert the time-domain signal into a frequency-domain signal, and calculates the power spectral density. The feature extraction unit extracts multi-dimensional spectral features. The anomaly detection unit calculates a spectral anomaly score based on a Gaussian mixture model and uses dynamic thresholds to determine suspicious signal segments.

[0096] The speech content extraction and feature engineering module includes a speech activity detection unit, a speech conversion unit, an acoustic feature extraction unit, and a background noise analysis unit. The speech activity detection unit detects speech frames using short-time energy and zero-crossing rate dynamic thresholds, and merges continuous speech segments. The speech conversion unit, based on an ASR model with a Transformer architecture, converts speech into text and outputs the text information. The acoustic feature extraction unit extracts acoustic features. The background noise analysis unit analyzes background features of non-speech components.

[0097] The multimodal feature fusion module includes a text vectorization unit, a feature normalization unit, and an attention fusion unit. The text vectorization unit uses the BERT model to convert the text output by ASR into a semantic vector. The feature normalization unit standardizes the spectral, acoustic, and background sound features. The attention fusion unit generates a comprehensive feature vector by weighted summation and concatenation of the multimodal features.

[0098] The semantic understanding module includes a sentiment analysis unit, a tone pattern analysis unit, and a logical relationship analysis unit. The sentiment analysis unit outputs sentiment labels and confidence scores based on the BERT sentiment classification head; the tone pattern analysis unit outputs tone labels and confidence scores based on the BERT tone classification head; and the logical relationship analysis unit identifies temporal logical relationships between sentences and paragraphs using an LSTM model.

[0099] The intent analysis and report generation module includes an LLM intent reasoning unit and a semantic report generation unit. The LLM intent reasoning unit takes text, sentiment, tone, and logical relationship tags as input and generates potential intents from the fine-tuned LLM. The semantic report generation unit outputs a structured semantic analysis report.

[0100] The comprehensive assessment and decision-making module includes a quantitative assessment unit, a strategy matching unit, and an execution control unit. The quantitative assessment unit calculates the spectral anomaly score and semantic suspicion score with weights, and outputs the comprehensive suspicion score and hazard level. The strategy matching unit matches multi-level response strategies based on the suspicion level and hazard level. The execution control unit triggers operations in the strategy library.

[0101] like Figure 3 A flowchart of an AI-based method for automatic semantic recognition of illegal radio broadcasts is shown. This invention provides an AI-based method for automatic semantic recognition of illegal radio broadcasts, comprising:

[0102] A broadband radio frequency receiver is used to scan and receive the target frequency band. The received radio frequency signal is preprocessed and then sampled into a digital baseband signal stream using a high-speed analog-to-digital converter. The digital baseband signal stream is then segmented into frames, and each frame is processed using a Hanning window function. A fast Fourier transform is performed to convert the time-domain signal into a frequency-domain signal. The power spectral density of the signal is calculated in real time. Multi-dimensional spectral features are extracted, including total signal power, main frequency power, power bandwidth, spectral flatness, and spectral centroid.

[0103] Suspicious signal segments are marked by comparing the spectral features of the current frame with those of normal background noise or legitimate broadcast signals. For the marked suspicious signal segments, speech segments are identified and extracted using an algorithm based on short-time energy and zero-crossing rate. Automatic speech recognition technology is applied to transcribe the extracted speech segments to obtain the text information of the speech content. The spectral features, acoustic features, background noise features, and text information output by ASR are fused to construct a comprehensive feature vector.

[0104] This paper introduces the BERT contextual semantic understanding model and combines it with comprehensive feature vectors to analyze the emotional color and tone patterns in speech. Using an LSTM time series analysis model, it understands the logical relationships between sentences and paragraphs in continuously broadcast content. Combining emotional color, tone patterns, and logical relationships, and based on an LLM model, it determines the underlying intent of the broadcast content and generates a semantic analysis report. Based on the semantic analysis report and spectral anomaly score, it forms an overall state assessment of the current broadcast signal, quantifies it, and reflects the degree of suspicion of illegal broadcasting and the potential harm level of the broadcast signal. Based on the degree of suspicion of illegal broadcasting and the potential harm level, it selects a response strategy and executes corresponding operations.

[0105] It will be apparent to those skilled in the art that the present invention is not limited to the details of the exemplary embodiments described above, and that the invention can be implemented in other specific forms without departing from its spirit or essential characteristics. Therefore, the embodiments should be considered in all respects as exemplary and non-limiting, and the scope of the invention is defined by the appended claims rather than the foregoing description. Thus, all variations falling within the meaning and scope of equivalents of the claims are intended to be included within the present invention. No reference numerals in the claims should be construed as limiting the scope of the claims.

Claims

1. An automatic semantic recognition method for illegal radio broadcasts based on artificial intelligence, characterized in that, include: By using dedicated receiving equipment, radio spectrum is continuously monitored to capture broadcast signals; The FFT digital signal processing technology is used to analyze the spectral characteristics of the captured signal in real time, generate spectral anomaly scores, and mark suspicious signal segments. For suspicious signal segments marked by spectrum analysis, the speech segments are identified and extracted using a speech activity detection algorithm based on short-time energy and zero-crossing rate. ASR automatic speech recognition technology is then applied to transcribe the speech content, obtain the text information of the speech content, and extract acoustic features and background sound features. The extracted spectral features, acoustic features, background noise features, and text information output by ASR are fused to construct a comprehensive feature vector; We introduce the BERT contextual semantic understanding model and combine it with comprehensive feature vectors to analyze the emotional color and tone patterns in speech; Based on the analysis results of emotional color and tone patterns, the logical relationships between sentences and paragraphs in continuously broadcast content are understood through the LSTM time series analysis model. By combining the obtained emotional color, tone pattern and logical relationship, and based on the LLM model, the potential intention of the broadcast content is determined and a semantic analysis report is generated. Based on the semantic analysis report and spectrum anomaly score, an overall status assessment of the current broadcast signal is formed and quantified to reflect the degree of suspicion of the broadcast signal as a black broadcast and the level of potential harm. Based on the degree of suspicion of illegal radio broadcasts and the level of potential harm, select a response strategy and execute corresponding actions.

2. The method for automatic semantic recognition of illegal radio broadcasts based on artificial intelligence according to claim 1, characterized in that, The method involves continuously monitoring the radio spectrum and capturing broadcast signals using dedicated receiving equipment. The FFT digital signal processing technique is used to analyze the spectral characteristics of the captured signal in real time, generate spectral anomaly scores, and mark suspicious signal segments, including: A broadband radio frequency receiver is used to scan and receive the target frequency band. After the received radio frequency signal is preprocessed, it is sampled into a digital baseband signal stream using a high-speed analog-to-digital converter. The sliding window technique is used for frame processing. After processing each frame signal with the Hanning window function, a fast Fourier transform is performed to convert the time domain signal into a frequency domain signal, and the power spectral density of the signal is calculated in real time. For each frame of spectral data after Fast Fourier Transform and power spectral density calculation, multi-dimensional spectral features are extracted; the sum of the power values ​​of all frequency points in the current frame spectrum is calculated to obtain the total signal power; the frequency point with the highest power in the current frame spectrum is found, and its corresponding power value is the main frequency power; a power threshold is set, and the frequency range occupied by all frequency points with power exceeding this threshold is calculated to obtain the power bandwidth; the ratio of the variance to the mean of the power value of the current frame spectrum is calculated to obtain the spectral flatness; a weighted average value is calculated, with the weight of the power value of each frequency point, to characterize the centroid position of the spectral energy, thus obtaining the spectral centroid. Collect spectral feature data known to be normal background noise or legitimate broadcast signals, and train a Gaussian mixture model. Input the multidimensional feature vector extracted from the current frame into the trained Gaussian mixture model, calculate the probability p that the feature vector belongs to the normal distribution, and convert the probability p into anomaly score S using the following formula: ； The calculated anomaly score S of the current frame is compared with a preset dynamic threshold. When the anomaly score is higher than the threshold, the corresponding frame signal is determined to be abnormal and marked as a suspicious signal segment.

3. The method for automatic semantic recognition of illegal radio broadcasts based on artificial intelligence according to claim 1, characterized in that, The suspicious signal segments identified by spectrum analysis are confirmed and extracted using a speech activity detection algorithm based on short-time energy and zero-crossing rate. Automatic speech recognition (ASR) technology is then applied to transcribe the speech, obtaining the text information of the speech content and extracting acoustic and background noise features, including: Calculate the short-time energy of each frame of signal, which is the sum of the squares of the amplitudes of all sampling points within the corresponding frame; calculate the zero-crossing rate of each frame of signal, which is the number of times the signal waveform crosses the zero level within the corresponding frame. In the initial silent segment of the signal segment, the short-time energy and zero-crossing rate of the background noise are calculated and updated in real time using a smooth update method, and a dynamic threshold is set. The short-time energy and zero-crossing rate of each frame are compared with the dynamic threshold. When the short-time energy is higher than the energy threshold and the zero-crossing rate is within the normal speech range, the corresponding frame is determined to be a speech frame. Continuous speech frames are merged to extract speech segments containing complete speech content. The extracted speech segments are input into a pre-trained ASR model. The ASR model is based on the Transformer architecture and converts the audio signal into a text sequence to obtain the text information of the speech content. For each speech segment, acoustic features are extracted, including speech rate features, pitch features, volume features, and prosodic features. Background sound analysis is performed on the non-speech parts of the speech segments to extract background sound features, including music type features, rhythm features, and loop cue sound features.

4. The method for automatic semantic recognition of illegal radio broadcasts based on artificial intelligence according to claim 1, characterized in that, The process of fusing the extracted spectral features, acoustic features, background noise features, and text information output by ASR to construct a comprehensive feature vector includes: The text information output by ASR is preprocessed, and each word is converted into a low-dimensional dense vector using a pre-trained BERT context-dependent language model. The vectors of all words in a sentence or text segment are weighted and summed through an attention mechanism to generate a text semantic vector. The spectral features, acoustic features, and background sound features are normalized separately. The normalized spectral feature vector, acoustic feature vector, background sound feature vector, and vectorized text semantic vector are then concatenated to form a long vector. An attention weighting mechanism is applied to the long vector formed by the concatenation of the first and last parts, and the weighted sum is calculated according to the importance of different modal features in the black radio identification task to generate a comprehensive feature vector.

5. The method for automatic semantic recognition of illegal radio broadcasts based on artificial intelligence according to claim 1, characterized in that, The introduction of the BERT contextual semantic understanding model, combined with comprehensive feature vectors, analyzes the emotional color and tone patterns in speech, including: A BERT model pre-trained on a general large-scale text corpus was selected as the foundation to construct a training dataset for black radio broadcast recognition with annotation information. The constructed training dataset includes speech samples and their corresponding text information, and is labeled with sentiment tags and tone tags. Features were extracted from the constructed training dataset to generate corresponding comprehensive feature vectors, which were input into the initialized BERT model. Through the backpropagation algorithm, the parameters of the BERT model and downstream classification tasks were jointly optimized to obtain a BERT contextual semantic understanding model with dedicated semantic understanding capabilities. Two parallel downstream task classification heads were integrated at the top of the model, including a sentiment color classification head and a tone pattern classification head. Each classification head contains a fully connected layer, and the number of output neurons is equal to the number of preset sentiment categories and tone categories. The comprehensive feature vector of the speech to be recognized is input into the BERT contextual semantic understanding model. BERT generates a context vector representing the overall semantics of the speech segment to be recognized through a multi-layer Transformer encoder. This context vector is simultaneously fed into the sentiment color classification head and the tone pattern classification head. The fully connected layer of each classification head performs a linear transformation on the input vector by performing matrix multiplication of the input vector with a learnable weight matrix and adding the result to a learnable bias vector to generate the original score vector. The original score vectors output by the two classification heads are applied with the Softmax activation function to transform them into a probability distribution in which all elements are positive and the sum is 1. From the probability distribution generated by Softmax, the category with the highest probability value is selected as the final sentiment label and tone label of the speech segment to be recognized, and the probability value is used as the confidence score of the corresponding label.

6. The method for automatic semantic recognition of illegal radio broadcasts based on artificial intelligence according to claim 1, characterized in that, The analysis results based on emotional coloring and tone patterns, through an LSTM time series analysis model, are used to understand the logical relationships between sentences and paragraphs in continuously broadcast content, including: A labeled training dataset of logical relationships in illegal broadcasts is constructed, including complete broadcast content confirmed as illegal and legitimate broadcasts. Comprehensive feature vectors are obtained and concatenated to form a multidimensional temporal feature sequence. Logical relationship labels are added at the level of adjacent sentences and paragraphs. The multidimensional temporal feature sequence and logical relationship labels are used as input to an LSTM model. The cross-entropy loss between the model's predicted logical relationships and the true labels is calculated using the backpropagation algorithm. The parameters of the LSTM model are jointly optimized using the gradient of the cross-entropy loss to obtain an LSTM model that can identify logical relationships in continuously broadcast content. The continuously broadcast content to be identified is formed into a multi-dimensional temporal feature sequence and input into a trained LSTM model. The model outputs a hidden state vector at each time step, which encodes the context information up to the current time step. At the sentence level, the model performs linear transformation and Softmax classification on the hidden state vectors of two adjacent time steps, and outputs the probability of belonging to each type of logical relationship between the time steps. At the paragraph level, the model performs the same processing on the hidden state vector sequences representing different paragraphs. Based on the output probability distribution, the logical relationship between sentences and paragraphs is identified.

7. The method for automatic semantic recognition of illegal radio broadcasts based on artificial intelligence according to claim 1, characterized in that, The combined emotional tone, intonation patterns, and logical relationships, based on an LLM model, determine the underlying intent of the broadcast content and generate a semantic analysis report, including: The dataset consists of text information, sentiment tags, tone tags, logical relationship tags, corresponding potential intent of broadcast content, and semantic analysis reports from labeled black broadcast and normal broadcast samples. The LLM model is trained using a fine-tuning method. After training, for broadcast content to be identified, the text information, sentiment tags, tone tags, and logical relationship tags are input into the trained LLM model, which then determines the potential intent of the broadcast content and generates a semantic analysis report.

8. The method for automatic semantic recognition of illegal radio broadcasts based on artificial intelligence according to claim 1, characterized in that, The overall status assessment of the current broadcast signal is formed based on the semantic analysis report and spectrum anomaly score, and is quantified to reflect the degree of suspicion of the broadcast signal as a black broadcast and its potential harm level, including: The semantic analysis report generated by LLM is structured and parsed to extract the number of keyword triggers, sentiment tags, tone tags, number of logical contradictions, and potential intent. By comparing these with the corresponding preset thresholds, keyword scores, sentiment color scores, tone pattern scores, logical relationship scores, and potential intent scores are obtained. The semantic suspicion score is calculated by weighted summation to indicate the degree of suspicion of the broadcast signal as a black broadcast. The spectrum anomaly score is normalized and weighted and summed with the semantic suspicion score to obtain the comprehensive suspicion score of the broadcast signal. This comprehensive suspicion score is then compared with a preset suspicion score threshold range to obtain the potential hazard level.

9. The method for automatic semantic recognition of illegal radio broadcasts based on artificial intelligence according to claim 1, characterized in that, The process of selecting a response strategy and executing corresponding operations based on the degree of suspicion of illegal broadcasting and the level of potential harm includes: A multi-level response strategy corresponding to different levels of suspicion and harm is predefined to form a strategy library. The calculated level of suspicion and potential harm of the black radio broadcast is used as input and matched with the strategy conditions in the strategy library. The highest priority response strategy is automatically selected and executed. When the signal is considered low-susceptibility and low-hazard, an alarm notification is triggered, a prompt message is pushed to the responsible personnel, a continuous tracking mechanism is initiated, and the sampling and analysis frequency of the signal is increased. When the signal is considered medium-susceptibility and medium-hazard, a structured alarm work order containing a semantic analysis report and audio is automatically generated based on the alarm notification and pushed to the superior regulatory department through a preset interface for manual judgment and decision-making. When the signal is considered high-susceptibility and high-hazard, while executing the alarm-level operation, a directional interference command is issued to the radio management equipment or interference suppression equipment in the signal transmission area according to preset geofencing rules or authorized instructions to temporarily suppress the power or shift the frequency of the target signal, rendering it ineffective.

10. An AI-based automatic semantic recognition system for illegal radio broadcasts, using the AI-based automatic semantic recognition method for illegal radio broadcasts as described in any one of claims 1-9, characterized in that, include: The signal acquisition and spectrum analysis module includes an RF receiving unit, a framing and FFT processing unit, a feature extraction unit, and an anomaly detection unit. The RF receiving unit scans the target frequency band, captures the RF signal, and converts it into a digital baseband signal stream. The framing and FFT processing unit performs framing and windowing on the digital baseband signal stream, executes a Fast Fourier Transform (FFT) to convert the time-domain signal into a frequency-domain signal, and calculates the power spectral density. The feature extraction unit extracts multi-dimensional spectral features. The anomaly detection unit calculates a spectral anomaly score based on a Gaussian mixture model and uses dynamic thresholds to determine suspicious signal segments. The speech content extraction and feature engineering module includes a speech activity detection unit, a speech conversion unit, an acoustic feature extraction unit, and a background noise analysis unit. The speech activity detection unit detects speech frames using short-time energy and zero-crossing rate dynamic thresholds, and merges continuous speech segments. The speech conversion unit, based on an ASR model with a Transformer architecture, converts speech into text and outputs the text information. The acoustic feature extraction unit extracts acoustic features. The background noise analysis unit analyzes background features of non-speech components. The multimodal feature fusion module includes a text vectorization unit, a feature normalization unit, and an attention fusion unit. The text vectorization unit uses the BERT model to convert the text output by ASR into a semantic vector. The feature normalization unit standardizes the spectral, acoustic, and background sound features. The attention fusion unit generates a comprehensive feature vector by weighted summation and concatenation of the multimodal features. The semantic understanding module includes a sentiment analysis unit, a tone pattern analysis unit, and a logical relationship analysis unit. The sentiment analysis unit outputs sentiment labels and confidence scores based on the BERT sentiment classification head; the tone pattern analysis unit outputs tone labels and confidence scores based on the BERT tone classification head; and the logical relationship analysis unit identifies temporal logical relationships between sentences and paragraphs using an LSTM model. The intent analysis and report generation module includes an LLM intent reasoning unit and a semantic report generation unit. The LLM intent reasoning unit takes text, sentiment, tone, and logical relationship tags as input and generates potential intents from the fine-tuned LLM. The semantic report generation unit outputs a structured semantic analysis report. The comprehensive assessment and decision-making module includes a quantitative assessment unit, a strategy matching unit, and an execution control unit. The quantitative assessment unit calculates the spectral anomaly score and semantic suspicion score with weights, and outputs the comprehensive suspicion score and hazard level. The strategy matching unit matches multi-level response strategies based on the suspicion level and hazard level. The execution control unit triggers operations in the strategy library.

Images