Live broadcast dynamic intelligent monitoring and early warning system for live broadcast smart machine

Abstract

The invention discloses a live broadcast dynamic intelligent monitoring and early warning system for a live broadcast smart machine, and particularly relates to the technical field of live broadcast monitoring. Comprising a live broadcast dynamic monitoring database, a live broadcast dynamic monitoring synchronous acquisition module, a live broadcast dynamic monitoring calculation module, a live broadcast dynamic monitoring evidence chain analysis module and a live broadcast dynamic monitoring intelligent early warning and feedback module. The live broadcast dynamic monitoring synchronous acquisition module acquires a live broadcast dynamic monitoring data set in a live broadcast process through a big data technology and performs synchronous processing to obtain synchronized live broadcast dynamic monitoring data set data and transmits the synchronized live broadcast dynamic monitoring data set data to the live broadcast dynamic monitoring calculation module; through the live broadcast dynamic monitoring evidence chain analysis module, the long-term behavior mode of the anchor is quantified, misjudgment caused by accidental behaviors is effectively avoided, the early warning accuracy is improved, more reliable monitoring and early warning information is provided for a live broadcast platform, and healthy development of a live broadcast environment is guaranteed.

Description

Technical Field

[0001] The present invention relates to the field of live broadcast monitoring technology, and in particular to a live broadcast dynamic intelligent monitoring and early warning system for a live broadcast smart machine. Background Art

[0002] The live broadcast smart machine is an integrated intelligent device designed specifically for live broadcast scenarios. With the rapid development of the live broadcast industry, live broadcast content covers multiple fields such as entertainment, education, e-commerce, and games. In order to ensure content compliance, create a healthy live broadcast environment, and protect user experience, live broadcast platforms have put forward extremely high requirements for the monitoring and early warning of live broadcast content. However, the existing live broadcast monitoring and early warning system has many shortcomings.

[0003] Traditional live broadcast monitoring and early warning systems use manual spot checks, analysis and monitoring based on single modal data, and monitoring based on simple rule matching. Manual spot checks arrange a special manual monitoring team to spot check live broadcast content at certain time intervals or according to specific rules. Analysis and monitoring based on single modal data, such as only analyzing audio data or only analyzing video data. Monitoring based on simple rule matching achieves the purpose of live broadcast monitoring and early warning by pre-setting a series of clear rules.

[0004] Although traditional live broadcast monitoring methods can serve as early warning, they still have many shortcomings; for example, manual spot checks are inefficient, highly subjective and have lags; analysis and monitoring based on single-modal data are difficult to automatically adjust the calculation model according to different scenarios and their own performance, and lack quantitative analysis of the host's long-term behavior patterns, which is prone to occasional misjudgments, thereby reducing the efficiency and accuracy of early warnings; monitoring based on simple rule matching may lead to deviations in semantic analysis based on audio and video due to the lack of synchronization in acquisition time, and thus cannot accurately determine whether the live broadcast content has a risk of violation; therefore, the existing live broadcast monitoring and early warning system can no longer meet the needs of the development of the live broadcast industry, and an innovative live broadcast dynamic intelligent monitoring and early warning system is urgently needed to solve the above problems. Summary of the invention

[0005] In order to overcome the above-mentioned defects of the prior art, an embodiment of the present invention provides a live broadcast dynamic intelligent monitoring and early warning system for a live broadcast smart machine to solve the problems of poor data synchronization, weak adaptive ability of computing models and lack of long-term behavioral quantitative evaluation in evidence chain analysis proposed in the above-mentioned background technology.

[0006] To achieve the above-mentioned purpose, the present invention provides the following technical solution: a live broadcast dynamic intelligent monitoring and early warning system for a live broadcast smart machine, comprising: Live broadcast dynamic monitoring database: through big data technology, live broadcast dynamic monitoring data is stored and updated in real time to build a live broadcast dynamic monitoring database. The live broadcast dynamic monitoring data includes live broadcast scene knowledge graph data, anchor historical violation records and real-time live broadcast data; Live streaming dynamic monitoring synchronous acquisition module: Through big data technology, live streaming dynamic monitoring data sets are collected and synchronously processed during the live streaming process, and the synchronized live streaming dynamic monitoring data sets are obtained and transmitted to the live streaming dynamic monitoring calculation module; Live streaming dynamic monitoring and calculation module: receives the synchronized live streaming dynamic data set obtained from the synchronization acquisition module, obtains the corresponding monitoring model from the live streaming dynamic monitoring database according to the live streaming tag, performs live streaming content monitoring and calculation, and transmits the monitoring and calculation results to the live streaming dynamic monitoring evidence chain analysis module; Live streaming dynamic monitoring evidence chain analysis module: Through big data technology, combined with the anchor's historical violation records, the monitoring calculation results obtained by the live streaming dynamic monitoring calculation module are analyzed for violation evidence risk, and the risk analysis results are transmitted to the live streaming dynamic monitoring intelligent early warning and feedback module; Live streaming dynamic monitoring intelligent early warning and feedback module: Through big data analysis technology, the risk analysis results of the live streaming dynamic monitoring evidence chain analysis module are compared with the threshold, different early warning levels are set according to the comparison results, and early warning feedback notifications are sent to the live streaming platform operators according to different early warning levels.

[0007] Technical effects and advantages of the present invention: 1. The present invention solves the problem of lip-sync asynchronization by using a live broadcast dynamic monitoring synchronization acquisition module, a high-precision clock chip and a PTP protocol, effectively improves the accuracy of semantic analysis of live broadcast content, reduces misjudgment caused by audio and video asynchronization, and more accurately identifies whether there is a risk of violation in the live broadcast, providing accurate data support for subsequent monitoring and early warning analysis; 2. The present invention uses a live broadcast dynamic monitoring calculation module to identify the live broadcast scene type based on the live broadcast dynamic monitoring database, and uses an image classification model, an audio classification model, and a text classification model to perform corresponding live broadcast dynamic monitoring calculations according to the target live broadcast scene type, thereby improving the monitoring efficiency, timeliness, and accuracy of monitoring; 3. The present invention uses a live broadcast dynamic monitoring evidence chain analysis module to quantify the long-term behavior pattern of the anchor, effectively avoid misjudgment caused by occasional behavior, enable the system to more scientifically evaluate the anchor's risks, improve the accuracy of early warning, and then provide more reliable monitoring and early warning information for the live broadcast platform to ensure the healthy development of the live broadcast environment. BRIEF DESCRIPTION OF THE DRAWINGS

[0008] Figure 1 It is a schematic diagram of the overall process of the present invention.

[0009] Figure 2 This is a flow chart of the live broadcast dynamic monitoring calculation module. DETAILED DESCRIPTION

[0010] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0011] See also Figure 1 As shown, the present invention provides a live broadcast dynamic intelligent monitoring and early warning system for a live broadcast smart machine, including a live broadcast dynamic monitoring database, a live broadcast dynamic monitoring synchronous acquisition module, a live broadcast dynamic monitoring calculation module, a live broadcast dynamic monitoring evidence chain analysis module and a live broadcast dynamic monitoring intelligent early warning and feedback module.

[0012] The live broadcast dynamic monitoring database is connected to the remaining modules, the live broadcast dynamic monitoring synchronous collection module is connected to the live broadcast dynamic monitoring calculation module, and the live broadcast dynamic monitoring evidence chain analysis module is respectively connected to the live broadcast dynamic monitoring calculation module and the live broadcast dynamic monitoring intelligent early warning and feedback module.

[0013] Live broadcast dynamic monitoring database: through big data technology, live broadcast dynamic monitoring data is stored and updated in real time to build a live broadcast dynamic monitoring database. The live broadcast dynamic monitoring data includes live broadcast scene knowledge graph data, anchor historical violation records and real-time live broadcast data; What needs to be specifically explained in this embodiment is that the live broadcast scene knowledge graph data includes live broadcast scene types (games / e-commerce / chat), live broadcast scene type monitoring models (such as classification models based on video images, text models based on semantic analysis, and semantic analysis text models based on audio data), live broadcast scene type violation rules, and live broadcast equipment parameters required by the live broadcast smart machine corresponding to the scene type (such as picture resolution, equipment performance, etc.); the anchor's historical violation records include the anchor's unique ID number, historical violation content, historical violation risk value, live broadcast data, etc.; real-time live broadcast data includes the unique identifier of the live broadcast session, the operating status data of the live broadcast smart machine (such as temperature, memory, resolution, etc.), data collection records, live broadcast dynamic monitoring data (such as live broadcast dynamic analysis confidence, audio sensitive words and timestamps), etc.; at the same time, the data is regularly backed up and restored through encryption technology to ensure the security and reliability of the data.

[0014] Live streaming dynamic monitoring synchronous acquisition module: Through big data technology, live streaming dynamic monitoring data sets are collected and synchronously processed during the live streaming process, and the synchronized live streaming dynamic monitoring data sets are obtained and transmitted to the live streaming dynamic monitoring calculation module, including the following steps: S1.1: Live streaming dynamic monitoring data set collection: Through big data technology, live streaming dynamic monitoring data set is collected during the live streaming process, including real-time video data collection through video acquisition devices (such as cameras), synchronous audio data collection through audio acquisition devices (such as microphones), and collection of the host's interactive text (such as answers, chat information, etc.) through the host interaction event collection interface (such as user input interface). The video data includes video key frames and corresponding timestamps, video images; audio data includes audio energy peaks and corresponding timestamps, audio clips; the host's live streaming interactive data includes the timestamp and interactive text that triggers the interactive event; S1.2: Synchronous processing of live dynamic monitoring data set: First, the clock chip is integrated into the synchronous acquisition module to obtain the reference time of the live dynamic monitoring data, and the time accuracy index t ac Compare with the threshold value. If it is greater than or equal to the threshold value, it means that the reference time accuracy of the clock chip is good. Otherwise, calibrate the clock chip to obtain the reference time accuracy index t provided by the clock chip. ac , t ac =Δt / t rea , Δt represents the error between the reference time provided by the clock chip and the standard time, t rea Indicates the actual time that has passed; then an interrupt signal is sent when the video acquisition device outputs the video key frame, and the clock chip timestamp t is recorded s , detect the audio energy peak (such as the FFT peak of each 20ms frame) through audio detection technology (such as audio ADC), and record the timestamp t y When the host triggers an interaction event through the host interaction event collection interface, the interaction event timestamp t is recorded. j ; Secondly, through the constraints of the PTP protocol t se ≤t th , the timestamps of the video key frames, the timestamps of the audio energy peaks and the timestamps of the anchor interaction events are time-aligned, the t th represents the PTP protocol synchronization error threshold, constraint condition t se Indicates the PTP protocol synchronization error. , max() means taking the maximum value; finally, the synchronized live dynamic data set is obtained; What needs to be specifically explained in this embodiment is that the video key frame refers to a completely independently compressed frame in video encoding, which can be decoded without relying on the previous and next frames and contains complete image information; the audio signal is originally a time domain signal, that is, a waveform with time as the independent variable; FFT is a mathematical tool used to convert time domain signals into frequency domain signals, revealing the amplitude and phase information of different frequency components in the signal; since the audio signal is continuous, in order to analyze its frequency characteristics, it is usually divided into multiple short time frames (for example, every 20ms), and each frame is processed by FFT separately. In the spectrum, the frequency component with the largest amplitude is called the peak frequency, and its corresponding amplitude value is called the FFT peak; t th This is based on the actual business requirements for time synchronization accuracy. For example, in industrial automation, the synchronization error between PLC (programmable logic controller) and sensors needs to be less than 1ms. In 5G communication, the time synchronization between base stations needs to be less than ±1.5μs (IEEE 1588v2 standard). However, in actual deployment, the error may accumulate to 2ms due to network delays or device performance degradation.

[0015] What needs to be specifically explained in this embodiment is that in a live broadcast, due to the asynchrony of video and audio acquisition, lip sounds may be inconsistent, resulting in deviations in the semantic analysis based on audio and video, and thus it is impossible to accurately determine whether the live broadcast content has a risk of violation. At the same time, when analyzing the association between the host interaction event and the live audio and video content, the asynchrony of data will make it difficult to determine the specific live broadcast link corresponding to the host interaction behavior, reducing the accuracy of the live broadcast content analysis, and may cause misjudgment or omission of early warning.

[0016] See also Figure 2 As shown, the live broadcast dynamic monitoring calculation module: receives the synchronized live broadcast dynamic data set obtained from the synchronization acquisition module, obtains the corresponding monitoring model from the live broadcast dynamic monitoring database according to the live broadcast tag to perform live broadcast content monitoring calculation, and transmits the monitoring calculation result to the live broadcast dynamic monitoring evidence chain analysis module, the live broadcast dynamic monitoring calculation module includes a live broadcast scene type recognition unit and a live broadcast scene type monitoring calculation unit, the live broadcast scene type monitoring calculation unit includes an image violation confidence monitoring calculation unit, an audio violation confidence monitoring calculation unit, an interactive text violation confidence monitoring calculation unit and a live broadcast scene type monitoring result generation unit, including the following steps: S2.1: Live scene type identification unit: At the start of the live broadcast, the target live scene type manually set by the anchor is identified, and the corresponding live scene type monitoring model is retrieved from the live broadcast dynamic monitoring database to monitor and calculate the synchronized live broadcast dynamic data set; What needs to be specifically explained in this embodiment is that the target live broadcast scene types include game live broadcast type, e-commerce live broadcast type, chat live broadcast type, etc.; for example, the corresponding game live broadcast type rules include props, gestures, etc., and the corresponding e-commerce live broadcast type rules include false prices, exaggerated language, etc.

[0017] S2.2: Live scene type monitoring calculation unit: S2.2.1: Image violation confidence monitoring calculation unit: The pre-processed images collected by the video acquisition device during the live broadcast are transmitted in real time to the ResNet-based classification model integrated in the live broadcast smart machine for image classification, and the confidence data set CD of the image classification results under the target live broadcast scene type is obtained. , cd i represents the confidence of the i-th classification result, n represents the number of classification results, such as normal, violation rule 1, violation rule 2, violation rule 3, violation rule 4 and other classification results, and the corresponding confidences are 0.1, 0.02, 0.05, 0.8, 0.03 and other confidences respectively. The preprocessing includes normalizing the image size and pixel value according to the model input requirements, matching the classification results under the target live scene type with the corresponding image violation rules of the live dynamic monitoring database, and calculating the image violation confidence IVCL of the successful match, IVCL=∑cd i , for example, if violation rule 3 is matched successfully, IVCL=0.8; It should be specifically explained in this embodiment that the classification model based on ResNet is a trained model that can accurately classify multi-source data types such as images and audio. Compared with other classification models, ResNet has a lower error rate. The flexibility of its architecture enables it to adapt to classification tasks of different scales and complexities; the ResNet classification model has achieved remarkable results in the field of computer vision; preprocessing, such as adjusting the image size to 255×255, normalizing pixel values ​​(such as dividing by 255 or using ImageNet mean / standard deviation).

[0018] S2.2.2: Audio violation confidence monitoring calculation unit: First, the audio clips collected by the audio acquisition device at the preset time are transmitted to the VGGish model in real time to obtain the feature vector VGGish_features of the audio clips; then the feature vectors are transmitted to the SVM-based classification model integrated in the live broadcast smart machine for audio classification, and the confidence data set ACD of the audio classification results under the target live broadcast scene type is obtained. , ac jrepresents the confidence of the jth classification result, k represents the number of classification results, such as normal, screaming and other classification results, the corresponding confidences are 0.1, 0.02 and other confidences, respectively. The classification results under the target live scene type are matched with the corresponding audio violation rules of the live dynamic monitoring database, and the audio violation confidence CLAV of the successful match is calculated. CLAV=∑ac j , for example, if screaming matches successfully, then CLAV=0.02, and the SVM-based classification model is: , C audio The result is -1 to 1, which is normalized to 0-1 to obtain confidence data. The normalized model is: CL=(S+1) / 2; What needs to be specifically explained in this embodiment is that the SVM-based classification model is a trained model, and finding a separation hyperplane in the feature space that can maximize the interval between two types of data is a supervised learning algorithm, which realizes classification by constructing an optimal hyperplane and is suitable for high-dimensional, nonlinear audio data; VGGish is an audio feature extraction model based on a convolutional neural network (CNN), which is usually used to extract features from audio signals and is a trained model.

[0019] S2.2.3: Interaction text violation confidence monitoring calculation unit: First, the anchor's interaction text collected by the anchor interaction event collection interface is transmitted to the BERT-based classification model integrated in the live broadcast smart machine for text classification, and the confidence data set TCD of the text classification result under the target live broadcast scene type is obtained. , tc I represents the confidence of the I-th classification result, N represents the number of classification results, such as normal, absolute profit and other classification results, the corresponding confidences are 0.1, 0.2 and other confidences, respectively, the classification results under the target live scene type are matched with the corresponding interactive text violation rules of the live dynamic monitoring database, and the confidence CITV of the successfully matched interactive text violation is calculated, CITV=∑tc I , for example, if the absolute profit matching is successful, CITV=0.2; It should be specifically explained in this embodiment that BERT (Bidirectional Encoder Representations from Transformers) is a pre-trained language model based on the Transformer architecture. It is a trained model that can accurately segment and classify text.

[0020] S2.2.4: Live scene type monitoring result generation unit: Through big data analysis technology, based on the image violation confidence IVCL, audio violation confidence CLAV and interactive text violation confidence CITV, the comprehensive index CIVC of live dynamic monitoring violation confidence is obtained. ; Live dynamic monitoring evidence chain analysis module: Through big data technology, combined with the anchor's historical violation records, the monitoring calculation results obtained by the live dynamic monitoring calculation module are analyzed for violation evidence risk, and the risk analysis results are transmitted to the live dynamic monitoring intelligent early warning and feedback module. Through big data technology, combined with the anchor's historical violation records, the comprehensive index CIVC of the live dynamic monitoring violation confidence within the preset time window K is calculated. his , perform violation evidence risk analysis on the monitoring calculation results obtained by the live broadcast dynamic monitoring calculation module, and obtain the anchor's current intelligent monitoring violation evidence risk index Risk, , CIVC his J Represents the comprehensive index of the confidence level of the Jth violation, Δt J Indicates the time from the Jth violation to the present (e.g., the violation occurred 3 days ago, then Δt J =3); What needs to be specifically explained in this embodiment is that violations beyond the preset time window K are automatically invalidated to avoid interference from outdated data; the system can monitor the live broadcast content in real time, promptly identify and warn of illegal content, and send early warning feedback notifications to the live broadcast platform operators, which in turn helps the live broadcast platform take timely measures to curb the spread of illegal content and ensure the healthy development of the live broadcast environment.

[0021] Live streaming dynamic monitoring intelligent early warning and feedback module: Through big data analysis technology, the risk analysis results of the live streaming dynamic monitoring evidence chain analysis module are compared with the threshold, different early warning levels are set according to the comparison results, and early warning feedback notifications are sent to the live streaming platform operators according to different early warning levels, including the following steps: S1: Through big data analysis technology, the risk analysis results of the live broadcast dynamic monitoring evidence chain analysis module are first compared with the threshold to obtain the evaluation coefficient η(Risk) of the anchor's current intelligent monitoring violation evidence risk index, η(Risk)=Risk / Risk 0 ,Risk 0 Indicates the corresponding threshold; then different warning levels are set according to the comparison result η(Risk), that is, if η(Risk)≥0.8, it indicates a high risk, triggering the first warning level; if 0.4≤η(Risk)<0.8, it indicates a moderate risk, triggering the second warning level; if η(Risk)<0.4, it indicates a low risk, triggering the third warning level; S2: Send a forced interruption warning feedback notification according to the first-level warning level, send a flow limit and digital human prompt warning feedback notification according to the second-level warning level, and send a warning warning feedback notification according to the third-level warning level. The notification includes the anchor ID, warning level, notification content, violation evidence and timestamp.

[0022] What needs to be specifically explained in this embodiment is that through the intelligent early warning and feedback of live broadcast dynamic monitoring, the risk of the anchor's live broadcast behavior can be evaluated more comprehensively and accurately, and the rationality and foresight of the early warning can be improved; at the same time, the multi-level early warning mechanism can quantify the monitored violation risks and divide them into different warning levels (such as level 1, level 2, level 3, etc.). This hierarchical management helps the operators of the live broadcast platform to locate risks more accurately and take corresponding countermeasures according to different levels of risks; the early warning and feedback notification mechanism helps to maintain a good atmosphere of the live broadcast platform and protect the rights and interests of legal and compliant anchors and viewers. By timely identifying and handling illegal content, the system can curb the spread of bad information and improve user experience and satisfaction.

[0023] Secondly: In the drawings of the embodiments disclosed in the present invention, only the structures related to the embodiments disclosed in the present invention are involved, and other structures can refer to the general design. In the absence of conflict, the same embodiment and different embodiments of the present invention can be combined with each other; Finally: The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principles of the present invention should be included in the protection scope of the present invention.

Claims

1. A live broadcast dynamic intelligent monitoring and early warning system for a live broadcast smart machine, characterized in that: include: Live broadcast dynamic monitoring database: through big data technology, live broadcast dynamic monitoring data is stored and updated in real time to build a live broadcast dynamic monitoring database. The live broadcast dynamic monitoring data includes live broadcast scene knowledge graph data, anchor historical violation records and real-time live broadcast data; Live streaming dynamic monitoring synchronous acquisition module: Through big data technology, live streaming dynamic monitoring data sets are collected and synchronously processed during the live streaming process, and the synchronized live streaming dynamic monitoring data sets are obtained and transmitted to the live streaming dynamic monitoring calculation module; Live streaming dynamic monitoring and calculation module: receives the synchronized live streaming dynamic data set obtained from the synchronization acquisition module, obtains the corresponding monitoring model from the live streaming dynamic monitoring database according to the live streaming tag, performs live streaming content monitoring and calculation, and transmits the monitoring and calculation results to the live streaming dynamic monitoring evidence chain analysis module; Live streaming dynamic monitoring evidence chain analysis module: Through big data technology, combined with the anchor's historical violation records, the monitoring calculation results obtained by the live streaming dynamic monitoring calculation module are analyzed for violation evidence risk, and the risk analysis results are transmitted to the live streaming dynamic monitoring intelligent early warning and feedback module; Live streaming dynamic monitoring intelligent early warning and feedback module: Through big data analysis technology, the risk analysis results of the live streaming dynamic monitoring evidence chain analysis module are compared with the threshold, different early warning levels are set according to the comparison results, and early warning feedback notifications are sent to the live streaming platform operators according to different early warning levels.

2. According to claim 1, a live broadcast dynamic intelligent monitoring and early warning system for a live broadcast smart machine is characterized in that: The live broadcast dynamic monitoring synchronous acquisition module includes the following steps: S1.1: Live streaming dynamic monitoring data set collection: Use big data technology to collect live streaming dynamic monitoring data sets during the live streaming process, including real-time video data collection through video acquisition equipment, synchronous audio data collection through audio acquisition equipment, and collection of the host's interactive text through the host interaction event collection interface; S1.2: Synchronous processing of live dynamic monitoring data set: First, the clock chip is integrated into the synchronous acquisition module to obtain the reference time of the live dynamic monitoring data, and the time accuracy index t ac Compare with the threshold value, if it is greater than or equal to the threshold value, it means that the reference time accuracy of the clock chip is good, otherwise the clock chip is calibrated; then send an interrupt signal when the video acquisition device outputs the video key frame, and record the clock chip timestamp t s , detect the audio energy peak through audio detection technology, and record the timestamp t y When the host triggers an interaction event through the host interaction event collection interface, the interaction event timestamp t is recorded. j ; Secondly, through the constraints of the PTP protocol t se ≤t th , the timestamps of the video key frames, the timestamps of the audio energy peaks and the timestamps of the anchor interaction events are time-aligned, the t th represents the PTP protocol synchronization error threshold, constraint condition t se Indicates the PTP protocol synchronization error; finally, the synchronized live dynamic data set is obtained.

3. According to claim 1, a live broadcast dynamic intelligent monitoring and early warning system for a live broadcast smart machine is characterized in that: The live broadcast dynamic monitoring calculation module includes a live broadcast scene type identification unit and a live broadcast scene type monitoring calculation unit. The live broadcast scene type monitoring calculation unit includes an image violation confidence monitoring calculation unit, an audio violation confidence monitoring calculation unit, an interactive text violation confidence monitoring calculation unit and a live broadcast scene type monitoring result generation unit.

4. The live broadcast dynamic intelligent monitoring and early warning system for a live broadcast smart machine according to claim 3, characterized in that: The live broadcast scene type identification unit: identifies the target live broadcast scene type manually set by the anchor at the beginning of the live broadcast, and retrieves the corresponding live broadcast scene type monitoring model through the live broadcast dynamic monitoring database to monitor and calculate the synchronized live broadcast dynamic data set.

5. The live broadcast dynamic intelligent monitoring and early warning system for a live broadcast smart machine according to claim 3, characterized in that: The image violation confidence monitoring calculation unit in the live scene type monitoring calculation unit: transmits the pre-processed images collected by the video acquisition device during the live broadcast process to the ResNet-based classification model integrated in the live broadcast smart machine in real time for image classification, and obtains the confidence data set CD of the image classification result under the target live scene type, , cd i represents the confidence of the i-th classification result, n represents the number of classification results, and the preprocessing includes normalizing the image size and pixel value according to the model input requirements, matching the classification results of the target live scene type with the corresponding image violation rules of the live dynamic monitoring database, and calculating the image violation confidence IVCL of the successful match; The audio violation confidence monitoring calculation unit: firstly transmits the audio segment collected by the audio collection device at the preset time to the VGGish model in real time to obtain the feature vector VGGish_features of the audio segment; then transmits the feature vector to the SVM-based classification model integrated in the live broadcast smart machine for audio classification, and obtains the confidence data set ACD of the audio classification result under the target live broadcast scene type. , ac j represents the confidence of the jth classification result, k represents the number of classification results, the classification result under the target live scene type is matched with the corresponding audio violation rules of the live dynamic monitoring database, and the audio violation confidence CLAV of the successful match is calculated. The SVM-based classification model is: , C audio The result is -1 to 1, which is normalized to 0-1 to obtain confidence data.

6. The live broadcast dynamic intelligent monitoring and early warning system for a live broadcast smart machine according to claim 3, characterized in that: The interactive text violation confidence monitoring calculation unit in the live broadcast scene type monitoring calculation unit: firstly, the interactive text of the anchor collected by the anchor interaction event collection interface is transmitted to the BERT-based classification model integrated in the live broadcast smart machine for text classification, and the confidence data set TCD of the text classification result under the target live broadcast scene type is obtained. , tc I represents the confidence of the first classification result, N represents the number of classification results, the classification result under the target live scene type is matched with the corresponding interactive text violation rules of the live dynamic monitoring database, and the confidence CITV of the successful matching interactive text violation is calculated; The live scene type monitoring result generating unit obtains the comprehensive index CIVC of the live dynamic monitoring violation confidence based on the image violation confidence IVCL, the audio violation confidence CLAV and the interactive text violation confidence CITV through big data analysis technology.

7. The live broadcast dynamic intelligent monitoring and early warning system for a live broadcast smart machine according to claim 1, characterized in that: The live broadcast dynamic monitoring evidence chain analysis module obtains the risk analysis result: through big data technology, combined with the anchor's historical violation records in the preset time window K, the live broadcast dynamic monitoring violation confidence index CIVC his , perform violation evidence risk analysis on the monitoring calculation results obtained by the live broadcast dynamic monitoring calculation module to obtain the anchor's current intelligent monitoring violation evidence risk index Risk.

8. The live broadcast dynamic intelligent monitoring and early warning system for a live broadcast smart machine according to claim 1, characterized in that: The live broadcast dynamic monitoring intelligent early warning and feedback module includes the following steps: S1: Through big data analysis technology, first compare the risk analysis results of the live broadcast dynamic monitoring evidence chain analysis module with the threshold value to obtain the evaluation coefficient η(Risk) of the current intelligent monitoring violation evidence risk index of the anchor; then set different warning levels according to the comparison result η(Risk), that is, if η(Risk)≥0.8, it indicates a high risk and triggers the first-level warning level; if 0.4≤η(Risk)<0.8, it indicates a moderate risk and triggers the second-level warning level; if η(Risk)<0.4, it indicates a low risk and triggers the third-level warning level; S2: Send a forced interruption warning feedback notification according to the first-level warning level, send a flow limit and digital human prompt warning feedback notification according to the second-level warning level, and send a warning warning feedback notification according to the third-level warning level. The notification includes the anchor ID, warning level, notification content, violation evidence and timestamp.

Images