IPTV edge node real-time quality monitoring and self-repairing system and control method

Abstract

The invention discloses an IPTV edge node real-time quality monitoring and self-repairing system and a control method, and mainly relates to the technical field of streaming media autonomous intelligent operation and maintenance. Comprising a data acquisition module used for acquiring network state data and playing state data in real time; the anomaly detection module is used for carrying a threshold detection and time sequence analysis algorithm and an One-Class SVM (Support Vector Machine) to carry out anomaly identification; the trend prediction module is used for carrying an ARIMA model and predicting a network state; the self-adaptive decision-making module is used for carrying out dynamic optimization on the code rate by carrying a QoE (Quality of Experience) model; and the actuator module is used for dynamically switching protocols and adjusting the size of the buffer area. The method has the beneficial effects that the self-adaptive dynamic optimization of the system is realized while the fault response capability of the IPTV system is improved.

Description

Technical Field

[0001] The present invention relates to the technical field of autonomous intelligent operation and maintenance of streaming media, and specifically to a system and control method for IPTV real-time playback quality monitoring and self-repair based on edge nodes. Background Art

[0002] In the current IPTV system, the service quality guarantee mainly adopts a centralized cloud monitoring system. However, this centralized cloud monitoring system often does not make full use of the computing and storage capabilities of edge devices (such as set-top boxes); moreover, the design idea of the existing centralized cloud monitoring system often relies on centralized processing, lacks autonomous intelligence on the terminal side, and does not integrate active protection technologies such as machine learning and trend prediction into the playback quality guarantee process.

[0003] This leads to the following problems in the service quality guarantee of the existing IPTV system: 1. High detection latency: The terminal needs to upload data to the cloud for analysis, usually with a latency of several seconds to dozens of seconds; 2. Weak prediction and prevention capabilities: Traditional solutions can only "detect - alarm - manual processing" and cannot achieve automatic predictive intervention; 3. Passive alarm: The traditional system relies on user complaints or alarms after batch analysis in the cloud, lacks real-time decision-making ability on the terminal side, resulting in a fault response latency of up to minutes and cannot achieve active self-healing.

[0004] Therefore, there is an urgent need for a system and control method for IPTV real-time playback quality monitoring and self-repair based on edge nodes to solve the above problems. Summary of the Invention

[0005] The purpose of the present invention is to provide an IPTV edge node real-time quality monitoring and self-repair system and control method, which improves the fault response ability of the IPTV system while realizing the adaptive dynamic optimization of the system.

[0006] To achieve the above object, the present invention is realized through the following technical solutions: On the one hand, there is provided an IPTV edge node real-time quality monitoring and self-repair system, including: A data acquisition module for: real-time collecting network status data and playback status data; An anomaly detection module for: carrying a threshold detection and time series analysis algorithm and One-Class SVM for anomaly identification; A trend prediction module for: carrying an ARIMA model to predict the network status; An adaptive decision-making module for: carrying a QoE model to dynamically optimize the bit rate; The actuator module is used for: dynamically switching protocols and adjusting the buffer size.

[0007] On the other hand, a control method applied to the real-time quality monitoring and self-healing system of the IPTV edge node as described above is provided, which is characterized by including the following steps: Step S1: The data acquisition module sets the acquisition frequency, acquires the network status data and the playback status data, and preprocesses the acquired network status data and playback status data; Step S2: The anomaly monitoring module sets static thresholds and constructs One-Class SVM according to the preprocessed network status data and playback status data, performs anomaly identification, classifies and prioritizes the identified anomaly situations, and then generates a unique event ID; Step S3: The trend prediction module performs trend prediction using the sliding window mean prediction and the LSTM long-term trend prediction according to the unique event ID; Step S4: The adaptive decision-making module sets the optimization objective of the QoE model and sets the bitrate switching rule according to the prediction result of the trend; Step S5: The actuator module switches the protocol according to the preprocessed network status data and the dynamic data of the playback status, and adjusts the buffer size.

[0008] Preferably, in step S1, specifically: The data acquisition module acquires the network status data and the playback status data in real time at a frequency of 10 times per second. The network status data includes: packet loss rate , bandwidth , round-trip delay , and the playback status data includes: number of freezes , buffering duration , startup delay ; The acquired original network status data and playback status data are standardized using Z-Score to eliminate the dimension difference: ; Among them, is the historical data mean value, is the standard deviation, and the update durations of and are set to adapt to network fluctuations; The DBSCAN clustering algorithm is adopted to mark and eliminate the abnormal points. If the data point does not satisfy , then this data point is marked as an outlier.

[0009] Preferably, step S2 includes: Step S21: Anomaly detection is performed through static thresholds. If or lasting for two seconds, then trigger network anomaly determination; when or , then trigger playback anomaly determination; Step S22: Construct a One-Class SVM dynamic learning model, train the model using historical normal data, and the kernel function is RBF; Step S23: Establish an inference model and calculate the anomaly score in real time: ; wherein, the weight coefficient , the threshold is dynamically adjusted; Step S24: Classify and prioritize the abnormal situations to generate a unique event ID.

[0010] Preferably, in the step S24, classifying and prioritizing the abnormal situations includes: Network anomaly: The trigger condition is a bandwidth drop or a sudden increase in the packet loss rate, and the priority is high; Player anomaly: The trigger condition is that the number of freezes exceeds the standard, and the priority is medium; User operation anomaly: The trigger condition is that the channel switching frequency exceeds the threshold, and the priority is low.

[0011] Preferably, in the step S3, the moving window mean prediction is specifically: The input data is the bandwidth sequence of the last 5 seconds , and the prediction process is: ; The LSTM long-term trend prediction is specifically: The input layer of the model is the historical data of 5 time steps, the hidden layer of the model is 32 LSTM units, the activation function is ReLU, the Dropout rate is 0.2, and the output layer is the predicted value of the bandwidth in the next 3 seconds ; The mean square error of the loss function is used to train the model; the model optimizer is Adam, and the learning rate , the training period , the data set takes the historical data of 7 days, and 80% of the data is used as the training set, and 20% of the data is used as the validation set for training.

[0012] Preferably, the step S4 includes: Step S41: Set the QoE optimization goal to minimize the weighted sum of the freezing rate and the picture quality loss, and the calculation process is: ; wherein, StallRate(r) represents the playback freezing rate at the bit rate r, and the calculation formula is: ; QualityLoss(r) represents the picture quality loss caused by reducing the bitrate r, which is quantified by the peak signal-to-noise ratio decrease value, and the formula is: ; 0.7 and 0.3 are weight coefficients; Step S42: Set the bitrate switching rule: If , switch to the 2Mbps bitrate; if the predicted bandwidth continues to decline, the bitrate is reduced to the next gradient in advance.

[0013] Preferably, the said step S5 includes: Step S51: When the packet loss rate or the bandwidth fluctuation , switch from HTTP-FLV to the HLS protocol, and the switching delay , and the user perceives stuttering ; where the bandwidth fluctuation is the amplitude of the bandwidth change within a unit time, the switching delay is the time difference from the trigger to the effect of the protocol switching operation, including but not limited to the time consumption of protocol stack reloading and buffer resetting.

[0014] The expansion process of the buffer dynamic adjustment in step S52 is: ; Among them, is the predicted delay increment, specifically: the difference between the predicted future network delay and the current delay, which is used to dynamically adjust the buffer size, and the formula is: , when , restore the default buffer size.

[0015] Compared with the prior art, the beneficial effects of the present invention are: 1. Millisecond-level fault response ability: Since all monitoring, analysis, decision-making and execution are completed locally on the set-top box without uploading to the cloud, the end-to-end delay is less than 100ms; it is realized through the closed-loop control of the MAPE-K model (Monitor-Analyze-Plan-Execute-Knowledge).

[0016] 2. Active prediction and prevention: Introduce the moving window average method and the LSTM prediction model, which can predict the bandwidth decline trend before the fault occurs and adjust the playback strategy in advance; 3. Adaptive dynamic optimization: Dynamically adjust the system strategy according to the user experience feedback (such as the decrease in the playback interruption rate and the reduction in the buffering times), and the self-learning mechanism continuously iterates to repair the effect; 4. Bandwidth Saving and Privacy Protection: Only report summary logs without transmitting original playback data, ensuring the privacy of users' viewing behaviors and meeting data compliance requirements. 5. Low-cost Deployment: It can run based on existing low-end set-top boxes (such as RK3228), without hardware modification, and only achieved through software upgrade. Description of the Drawings

[0017] Figure 1 is a schematic diagram of the system structure of the present invention; Figure 2 is a flowchart of the control method of the present invention. Detailed Embodiments

[0018] The present invention will be further described below in conjunction with specific embodiments. It should be understood that these embodiments are only used to illustrate the present invention and not to limit the scope of the present invention. In addition, it should be understood that after reading the content taught by the present invention, those skilled in the art can make various changes or modifications to the present invention, and these equivalent forms also fall within the scope defined by this application.

[0019] In the present invention, terms such as "upper", "lower", "left", "right", "front", "rear", "vertical", "horizontal", "side", "bottom", etc. indicate the orientation or positional relationship based on the orientation or positional relationship shown in the drawings. They are only relationship terms determined to facilitate the description of the structural relationship of each component or element of the present invention, and do not specifically refer to any component or element of the present invention, and should not be construed as a limitation to the present invention.

[0020] Embodiment: As Figure 1 shown, this embodiment provides an IPTV edge node real-time quality monitoring and self-healing system, including: A data acquisition module, used for: real-time acquisition of network status data and playback status data; An anomaly detection module, used for: carrying a threshold detection and time series analysis algorithm and One-Class SVM for anomaly identification; A trend prediction module, used for: carrying an ARIMA model to predict the network status; An adaptive decision-making module, used for: carrying a QoE model to dynamically optimize the bit rate; An actuator module, used for: dynamically switching protocols and adjusting the buffer size.

[0021] In this embodiment, the test scenario is: Area A with high user density and frequent network interference was used as the test site. The device used was a mainstream set-top box on the market, equipped with the Android 4.4 system and 2GB RAM. The network environment where the device was located was: home broadband (bandwidth fluctuated between 2 and 12 Mbps) and the peak latency reached 280 ms. The test content was: live watching sports events and the bitrate gradient was: [8M, 6M, 4M, 2M].

[0022] As Figure 2 shown, this embodiment also provides a control method applied to the above IPTV edge node real-time quality monitoring and self-repair system. Combining the test scenario, it specifically includes the following steps: Step S1: The data acquisition module sets the acquisition frequency, acquires the network status data and the playback status data, and preprocesses the acquired network status data and playback status data; Specifically: The data acquisition module acquires the network status data and the playback status data in real time at a frequency of 10 times per second. Among them, the network status data includes: packet loss rate , bandwidth , round-trip delay , and the playback status data includes: number of freezes , buffering duration , startup delay ; The acquired original network status data and playback status data are standardized using Z-Score to eliminate the dimension difference: ; Among them, is the historical data mean, is the standard deviation. The update duration of and is set to 5 seconds to adapt to network fluctuations; The DBSCAN clustering algorithm is adopted (parameters: neighborhood radius , minimum number of samples ), mark and remove the abnormal points. If the data point does not satisfy , then mark this data point as an outlier.

[0023] Step S2: The anomaly monitoring module sets static thresholds and constructs One-Class SVM based on the preprocessed network status data and playback status data for anomaly identification. After classifying and prioritizing the identified anomalies, a unique event ID is generated; specifically: Anomaly detection is performed through static thresholds, or lasts for two seconds, then the network anomaly determination is triggered. When or , a playback anomaly determination is triggered; Construct a One-Class SVM dynamic learning model, train the model using historical normal data (window is the last 30 minutes), and the kernel function is RBF ( ); Model inference, real-time calculation of anomaly scores: ; Among them, the weight coefficient , and the threshold are dynamically adjusted; Classify and prioritize abnormal situations. Among them, network anomalies: the trigger condition is a decrease in bandwidth or a sudden increase in packet loss rate, and the priority is high (immediate handling required); player anomalies: the trigger condition is that the number of freezes exceeds the standard, and the priority is medium (response required within 5 seconds); user operation anomalies: the trigger condition is that the channel switching frequency exceeds the threshold (such as 10 times / minute), and the priority is low (only log records); In this embodiment, the calculation result of the anomaly score model is , exceeding the threshold of 0.75; the initial judgment of the anomaly type is "network fluctuation freeze", and the system records the anomaly event IDEX_20250428_#1; no complaint event has been triggered at the user end, but the warning system has been pushed to the foreground status bar.

[0024] Step S3: The trend prediction module performs trend prediction based on the unique event ID using moving window mean prediction and LSTM long-term trend prediction; specifically: Moving window mean prediction: The input data is the bandwidth sequence in the last 5 seconds , and the prediction formula is: ; LSTM long-term trend prediction: The input layer of the model is the historical data of 5 time steps, the hidden layer of the model is 32 LSTM units, the activation function is ReLU, the Dropout rate is 0.2, and the output layer is the predicted value of the bandwidth in the next 3 seconds , model training: The loss function is mean squared error (MSE), optimizer: Adam, learning rate , training cycle , the dataset takes historical data of 7 days, and is trained in a ratio of 80% for training and 20% for validation; Based on the event ID in step S2 above: Input data: The bandwidth sequence in the last 5 seconds is: [6.1, 5.3, 4.6, 3.9, 3.3] Mbps; Model parameters: p = 2, d = 1, q = 2, determined by minimizing the AIC criterion; Prediction result: The trend mean in the next 10 seconds ; The system determines that there is a trend of "continuous bandwidth decline".

[0025] Step S4: The adaptive decision-making module sets the optimization objective of the QoE model and sets the bitrate switching rule according to the prediction result of the trend; specifically: Set the QoE optimization objective to minimize the weighted sum of the stuttering rate and the picture quality loss, and the calculation formula is ; Set the bitrate switching rule: If , switch to the 2Mbps bitrate; if it is predicted that the bandwidth continues to decline, reduce the bitrate to the next gradient in advance.

[0026] Step S5: The actuator module switches the protocol according to the preprocessed network status data and the dynamic data of the playback status, and adjusts the buffer size; specifically: Protocol switching operation, when the packet loss rate or the bandwidth fluctuates , switch from HTTP-FLV to the HLS protocol, and the switching delay , ; Dynamic adjustment of the buffer: Expansion formula:

[0027] Among them, ΔT is the predicted delay increment (for example ), when the network is stable ( ), gradually restore the default buffer size.

[0028] Step S6: Adjust the priority of the kernel thread: Raise the priority of the decoding thread to the real-time level (RTpriority99) (RTpriority represents the real-time thread priority (Real-Time Priority), and the range is 0-99 (the higher the value, the higher the priority)), and limit the CPU occupancy rate of non-critical threads (such as log upload) ≤ 10%; Feedback optimization and reinforcement learning, configure the Q-Learning model training: State definition: ; Action space , Reward function design: ; ; ; Synchronize the cloud log data at 3:00 am every day, and retrain the LSTM and Q-Learning models; Update the terminal policy library and push it to the set-top box through differential update (DeltaUpdate); Local logging event ID, timestamp, exception type, repair action, execution delay, QoE metric change. Stored in JSON compression, single log size ≤ 1KB; Asynchronous reporting uses the MQTT protocol (QoS = 1) to ensure at least one delivery. The data uses the LZ4 algorithm with a compression rate ≥ 80%, and the transmission rate is limited to 10 kbps to avoid affecting the playback traffic; After analyzing the logs in the cloud, global optimization instructions are generated (such as adjusting thresholds or updating model parameters); The instructions are sent to the terminal through an HTTPS encrypted channel, and the terminal executes the update during idle periods.

[0029] The above is a specific description of the preferred embodiment of the present invention, but the present invention is not limited to the described embodiment. Those skilled in the art can make various equivalent deformations or substitutions without departing from the spirit of the present invention, and these equivalent deformations or substitutions are all included in the scope defined by the claims of this application.

Claims

1. A real-time quality monitoring and self-repair system for IPTV edge nodes, characterized in that: include: Data collection module, used to collect network status data and playback status data in real time; Anomaly detection module, used to: carry out anomaly identification with threshold detection and timing analysis algorithm and One-ClassSVM; Trend prediction module, used to: use ARIMA model to predict network status; Adaptive decision module, used to: carry out QoE model and dynamically optimize bit rate; Actuator module, used to: dynamically switch protocols and adjust buffer size.

2. A control method for the IPTV edge node real-time quality monitoring and self-repairing system as claimed in claim 1, characterized in that: The following steps are involved: Step S1: The data collection module sets the collection frequency, collects network status data and playback status data, and pre-processes the collected network status data and playback status data; Step S2: The anomaly monitoring module sets a static threshold and constructs a One-ClassSVM according to the preprocessed network status data and playback status data to identify anomalies, and generates a unique event ID after classifying and prioritizing the identified anomalies; Step S3: The trend prediction module performs trend prediction based on the unique event ID using sliding window mean prediction and LSTM long-term trend prediction; Step S4: The adaptive decision module sets the optimization target of the QoE model and sets the bit rate switching rule according to the trend prediction result; Step S5: The executor module adjusts the buffer size according to the pre-processed network status data and the playback status dynamic data switching protocol.

3. A control method according to claim 2, characterized in that: The step S1 is specifically: The data acquisition module collects network status data and playback status data in real time at a frequency of 10 times per second. The network status data includes: packet loss rate ,bandwidth , round trip delay The playback status data includes: the number of freezes , Buffering time , Startup delay ; Use Z-Score to normalize the collected original network status data and playback status data to eliminate dimensional differences: ； in, is the mean of historical data, is the standard deviation, set and The update time is used to adapt to network fluctuations; Use DBSCAN clustering algorithm to mark and remove outliers. If the data point Dissatisfied , then mark this data point as an outlier.

4. A control method according to claim 2, characterized in that: The step S2 comprises: Step S21: Perform anomaly detection using static thresholds. or If it lasts for two seconds, it will trigger a network anomaly determination. or , then the abnormal playback judgment is triggered; Step S22: construct a One-ClassSVM dynamic learning model, use historical normal data to train the model, and use RBF as the kernel function; Step S23: Establish an inference model and calculate anomaly scores in real time: ； Among them, the weight coefficient , threshold Dynamic adjustment; Step S24: Classify and prioritize the abnormal situations and generate a unique event ID.

5. A control method according to claim 4, characterized in that: In step S24, the abnormal situations are classified and prioritized, including: Network anomaly: The trigger condition is bandwidth drop or packet loss rate increase, and the priority is high; Player abnormality: The trigger condition is that the number of freezes exceeds the limit, and the priority is medium; Abnormal user operation: The trigger condition is that the channel switching frequency exceeds the threshold and the priority is low.

6. A control method according to claim 2, characterized in that: In step S3, the sliding window mean prediction is specifically as follows: The input data is the bandwidth sequence of the last 5 seconds , the prediction process is: ； The specific LSTM long-term trend prediction is: The model input layer is the historical data of 5 time steps, the model hidden layer is 32 LSTM units, the activation function is ReLU, the Dropout rate is 0.2, and the output layer is the bandwidth prediction value for the next 3 seconds. ; Use the loss function mean square error to train the model; The model optimizer is Adam, and the learning rate , training cycle ,The data set takes 7 days of historical data, 80% of the data is used as the training set and 20% of the data is used as the validation set for training.

7. A control method according to claim 2, characterized in that: The step S4 comprises: Step S41: Set the QoE optimization target to minimize the weighted sum of the freeze rate and the image quality loss. The calculation process is: ； Wherein, StallRate(r) represents the playback stall rate at bit rate r, and the calculation formula is: ;QualityLoss(r) represents the image quality loss caused by reducing the bit rate r, using the peak signal-to-noise ratio The drop value is quantified, and the formula is: ; 0.7 and 0.3 are weight coefficients; Step S42: Setting the bit rate switching rule: If , switch to 2Mbps bit rate; if the predicted bandwidth continues to decrease, reduce the bit rate to the next gradient in advance.

8. A control method according to claim 2, characterized in that: The step S5 comprises: Step S51: Packet loss rate or bandwidth fluctuations When switching from HTTP-FLV to HLS protocol, the switching delay , user-perceived lag ; Bandwidth fluctuations is the bandwidth variation per unit time, Switching delay The time difference from the triggering to the taking effect of the protocol switching operation, including but not limited to the time taken for protocol stack reloading and buffer reset; Step S52: The expansion process of the buffer zone is dynamically adjusted as follows: ； in, To predict the delay increment, specifically: predict the difference between the future network delay and the current delay, which is used to dynamically adjust the buffer size. The formula is: ,when , restores the default buffer size.

Images