User experience optimization method and system for audio and video transmission system

Abstract

The present invention provides a user experience optimization method and system for audio and video transmission systems. First, a data set of transmission quality during audio and video transmission at a target terminal is acquired. The data set contains multiple transmission session sequences, each of which is characterized by network fluctuations, encoding and decoding parameters, and user feedback. Transmission feature extraction is then performed on the data set to obtain audio and video synchronization, dynamic bit rate, and decoding delay features. A pre-trained reinforcement learning optimization model is then invoked to dynamically aggregate these features to generate a transmission quality assessment result. Finally, a dynamic transmission optimization strategy is determined based on the transmission quality assessment result and fed back to the audio and video transmission system, triggering real-time transmission parameter adjustment operations to optimize the user experience of the audio and video transmission system.

Description

Technical Field

[0001] The present invention relates to the technical field of audio and video systems, and in particular to a method and system for optimizing user experience of an audio and video transmission system. Background Art

[0002] In the field of audio and video transmission technology, with the popularization of applications such as streaming services, remote conferencing, and real-time communications, users' requirements for transmission quality are becoming increasingly stringent. Existing audio and video transmission systems struggle to maintain stable audio and video synchronization and real-time performance under complex network conditions. Furthermore, existing quality assessment systems often rely on a single technical indicator, failing to fully reflect the complex relationship between user subjective experience and multi-dimensional technical parameters. This results in a disconnect between optimization strategies and actual needs. Furthermore, existing solutions often rely on static threshold judgments or preset rule adjustments, lacking the real-time perception and adaptability to environmental changes and user feedback, making it difficult to address sudden quality degradation in dynamic scenarios.

[0003] In addition, the user feedback mechanisms of related solutions are mostly limited to post-ratings or simple complaint records, and are not deeply associated with transmission technology parameters, resulting in unclear problem location and low optimization efficiency. Summary of the Invention

[0004] In view of the above-mentioned problems, in combination with the first aspect of the present invention, an embodiment of the present invention provides a method for optimizing user experience in an audio and video transmission system, the method comprising:

[0005] Acquire a transmission quality data set generated by a target terminal during audio and video transmission, the transmission quality data set comprising multiple transmission session sequences, each transmission session sequence comprising network fluctuation characteristics, encoding and decoding parameter characteristics, and user feedback characteristics during at least one transmission period;

[0006] Performing transmission feature extraction processing on the transmission quality data set to obtain audio and video synchronization features, dynamic bit rate features, and decoding delay features of each transmission session sequence;

[0007] Invoking a pre-trained reinforcement learning optimization model to dynamically aggregate the audio and video synchronization features, the dynamic bit rate features, and the decoding delay features to generate a transmission quality evaluation result of the transmission session sequence;

[0008] A dynamic transmission optimization strategy is determined based on the transmission quality evaluation result, and the dynamic transmission optimization strategy is fed back to the audio and video transmission system to trigger a real-time transmission parameter adjustment operation.

[0009] On the other hand, an embodiment of the present invention also provides a user experience optimization system for an audio and video transmission system, including a processor and a machine-readable storage medium, wherein the machine-readable storage medium is connected to the processor, the machine-readable storage medium is used to store programs, instructions or codes, and the processor is used to execute the programs, instructions or codes in the machine-readable storage medium to implement the above method.

[0010] Based on the above aspects, the present invention systematically integrates multi-dimensional heterogeneous data such as network fluctuations, encoding and decoding parameters and user feedback, so that the transmission quality analysis can fully reflect the dynamic coupling relationship between audio and video synchronization, bit rate adaptability and decoding real-time performance during audio and video transmission. The transmission feature extraction link not only reveals the potential causes of transmission quality degradation by analyzing core features such as audio and video synchronization deviation, dynamic bit rate fluctuation and decoding delay distribution, but also establishes a quantitative mapping model between user perception and underlying technical parameters, significantly improving the accuracy of problem diagnosis. The dynamic aggregation processing mechanism based on the pre-trained reinforcement learning model effectively overcomes the static adaptation defects of the traditional rule-driven method. By learning the complex nonlinear relationship between transmission characteristics and user experience online, dynamic calibration and forward-looking prediction of transmission quality evaluation are achieved. The dynamic transmission optimization strategy finally generated not only significantly reduces the probability of typical quality problems such as audio and video desynchronization, freeze and delay by adjusting key parameters such as bit rate control, buffer management and decoding priority in real time, but also achieves the Pareto optimality of bandwidth utilization and user experience quality in a complex network environment through the adaptive parameter adjustment mechanism. Therefore, through the deep integration of data intelligence and real-time control, the robustness and user satisfaction of the audio and video transmission system in dynamic scenarios are improved. BRIEF DESCRIPTION OF THE DRAWINGS

[0011] Figure 1 It is a schematic diagram of the execution flow of the user experience optimization method for the audio and video transmission system provided by an embodiment of the present invention.

[0012] Figure 2 Schematic diagram of exemplary hardware and software components of a user experience optimization system for an audio and video transmission system provided by an embodiment of the present invention. DETAILED DESCRIPTION

[0013] The present invention will be described in detail below with reference to the accompanying drawings. Figure 1 This is a flow chart of a method for optimizing user experience of an audio and video transmission system provided by an embodiment of the present invention. The method for optimizing user experience of an audio and video transmission system is introduced in detail below.

[0014] Step S110: Acquire a transmission quality data set generated by the target terminal during the audio and video transmission process, wherein the transmission quality data set includes multiple transmission session sequences, each transmission session sequence consisting of network fluctuation characteristics, encoding and decoding parameter characteristics, and user feedback characteristics of at least one transmission period.

[0015] In this example, we use a user terminal on a video playback platform. This platform has a large number of users and experiences a massive amount of audio and video transmission activity every day. Assuming that the target terminal in this example is User A's device, we begin collecting transmission quality data while User A is watching a popular TV series. The transmission session sequences are divided into specific time intervals, for example, every five minutes.

[0016] For each transmission session sequence, relevant data can be collected to identify network fluctuation characteristics. For example, within a transmission session sequence, network fluctuation characteristics include packet loss rate, jitter rate, and bandwidth change rate. Assume that the packet loss rate sequence is [0.01, 0.02, 0.03, 0.02, 0.01] during a period, representing the packet loss rate at different times during the period; the jitter rate sequence is [5, 8, 6, 7, 5], measured in milliseconds, reflecting the degree of network jitter; and the bandwidth change rate sequence is [0.1, 0.2, -0.1, 0.1, 0.05], reflecting the bandwidth variation during the period.

[0017] Codec parameter characteristics may include the initial bitrate setting value, dynamic bitrate adjustment threshold, and buffer capacity parameters. Assume the initial bitrate setting value is 2000kbps, the dynamic bitrate adjustment threshold is ±200kbps, and the buffer capacity parameter is 500KB. This means that the system initially sets the bitrate to 2000kbps and can adjust the bitrate between 1800kbps and 2200kbps when the network fluctuates. At the same time, the terminal device has a 500KB buffer to process data.

[0018] User feedback features are collected through user actions and comments during viewing. For example, if user A provides feedback during viewing, the user feedback text is obtained. This user feedback text contains information such as the user's evaluation of the audio-visual synchronization and their perception of latency. This is used to extract the user's subjective rating features for audio-visual synchronization and their sensitivity to latency.

[0019] Step S120: performing transmission feature extraction processing on the transmission quality data set to obtain audio and video synchronization features, dynamic bit rate features, and decoding delay features of each transmission session sequence.

[0020] In this embodiment, step S120 may include:

[0021] Step S121: performing time series segmentation processing on the network fluctuation characteristics in the transmission session sequence to obtain a plurality of network fluctuation sub-segments.

[0022] In this scenario, considering the aforementioned network fluctuation characteristics, each transmission session sequence can be further segmented into time series. For example, a 5-minute transmission session sequence can be divided into five network fluctuation sub-segments by minute. Taking the first network fluctuation sub-segment as an example, it contains the packet loss rate sequence [0.01, 0.02], the jitter rate sequence [5, 8], and the bandwidth change rate sequence [0.1, 0.2] within that minute, facilitating a more detailed analysis of the characteristics of network fluctuations within different time periods.

[0023] Step S122: Call the pre-trained encoder model to encode each network fluctuation sub-segment to generate a fluctuation intensity vector for each network fluctuation sub-segment.

[0024] In this embodiment, step S122 may include:

[0025] Step S1221: Obtain the packet loss rate sequence, jitter rate sequence, and bandwidth change rate sequence contained in each network fluctuation sub-segment.

[0026] For example, taking the first network fluctuation sub-segment as an example, the packet loss rate sequence obtained therein is [0.01, 0.02], the jitter rate sequence is [5, 8], and the bandwidth change rate sequence is [0.1, 0.2].

[0027] Step S1222: normalize the packet loss rate sequence to obtain a packet loss rate distribution feature.

[0028] In this step, the packet loss rate sequence is normalized to map the data into a specific range for better comparison and analysis. Assuming the normalization method used is to divide each value in the packet loss rate sequence by the maximum value in the sequence, then for the packet loss rate sequence [0.01, 0.02], the maximum value is 0.02. After normalization, the resulting packet loss rate distribution characteristic is [0.5, 1], which represents the relative distribution of packet loss rates within this network fluctuation subsegment.

[0029] Step S1223: performing sliding window statistical processing on the jitter rate sequence to generate jitter peak-to-valley difference features and jitter duration features.

[0030] In this embodiment, sliding window statistical processing is used for the jitter rate sequence [5, 8]. Assuming the sliding window size is 2, that is, two consecutive jitter rate values are considered at a time, in this jitter rate sequence, the first window is [5, 8], and the calculated peak-to-valley difference is 8-5=3. This is the jitter peak-to-valley difference characteristic within this sliding window. The jitter duration characteristic is determined based on the number of window moves and the window size. For example, since the window size is 2 and it moves once, the jitter duration characteristic is 2 time units (here assuming that each time point represents 1 time unit).

[0031] Step S1224: performing gradient calculation on the bandwidth change rate sequence to obtain bandwidth change trend characteristics.

[0032] In this example, the gradient is calculated for the bandwidth change rate sequence [0.1, 0.2]. The gradient reflects the changing trend of the data. This gradient is calculated by subtracting the previous value from the next value, i.e., 0.2 - 0.1 = 0.1. Therefore, a bandwidth change trend characteristic of 0.1 indicates an upward trend in bandwidth within this network fluctuation subsegment.

[0033] Step S1225: input the packet loss rate distribution characteristics, jitter peak-to-valley difference characteristics, jitter duration characteristics and bandwidth change trend characteristics into the encoder model, and generate the fluctuation intensity vector through multi-layer nonlinear transformation.

[0034] In this embodiment, the previously obtained packet loss rate distribution feature [0.5, 1], jitter peak-to-valley difference feature 3, jitter duration feature 2, and bandwidth change trend feature 0.1 can be input into the pre-trained encoder model, and the encoder model processes the above features through multiple layers of nonlinear transformations. For example, the first layer of the encoder model may perform a weighted summation of the packet loss rate distribution feature, assuming the weights are 0.3 and 0.7 respectively, and calculate 0.5×0.3+1×0.7=0.85. The second layer may perform some nonlinear combination of this result with the jitter peak-to-valley difference feature, such as using a nonlinear function f(x, y)=x*y+1 to calculate f(0.85, 3)=0.85×3+1=3.55. Subsequent layers continue to perform similar processing on the data, ultimately generating a fluctuation intensity vector. Assume that after multiple layers of processing, the resulting fluctuation intensity vector is [2.5, 3.0, 1.5]. This fluctuation intensity vector comprehensively reflects the fluctuation intensity of the network fluctuation sub-segment.

[0035] Step S123: Determine the rate adaptation deviation value for each transmission period based on the correlation calculation between the fluctuation intensity vector and the encoding and decoding parameter characteristics.

[0036] In this embodiment, step S123 may include:

[0037] Step S1231: extracting the initial bit rate setting value, dynamic bit rate adjustment threshold and buffer capacity parameter from the encoding and decoding parameter characteristics.

[0038] For example, looking back at the codec parameter characteristics mentioned above, the initial bit rate setting value is 2000kbps, the dynamic bit rate adjustment threshold is ±200kbps, and the buffer capacity parameter is 500KB.

[0039] Step S1232: Calculate the cosine similarity between the fluctuation intensity vector and the dynamic bit rate adjustment threshold to obtain the influence coefficient of network fluctuation on bit rate.

[0040] For example, assuming the fluctuation intensity vector is [2.5, 3.0, 1.5], the vector corresponding to the dynamic bitrate adjustment threshold can be assumed to be [1, 1, 1]. (This is a simplified assumption for calculating cosine similarity; in practice, more complex representations may be used.) Cosine similarity is calculated by dividing the dot product of the two vectors by the product of their moduli. Specifically, first calculate the dot product: 2.5×1+3.0×1+1.5×1=7. Then calculate the modulus of the fluctuation intensity vector: √(2.5²+3.0²+1.5²)=√(6.25+9+2.25)=√17.5≈4.18. The modulus of the dynamic bitrate adjustment threshold vector is √(1²+1²+1²)=√3≈1.73. Therefore, the cosine similarity is 7 / (4.18×1.73)≈0.96, which is the coefficient of network fluctuation on bitrate, indicating a high correlation between network fluctuation and bitrate adjustment.

[0041] Step S1233: Determine the delay tolerance of the bit rate adjustment according to the buffer capacity parameter, and generate a bit rate adaptation weight in combination with the impact coefficient.

[0042] For example, let's assume the buffer capacity parameter is 500KB. Assume, based on experience or pre-defined rules, that a 500KB buffer corresponds to a rate adaptation delay tolerance of 100kbps / min (the maximum allowable rate adjustment per minute). Combined with the previously determined influence coefficient of 0.96, we generate a rate adaptation weight. Assume the weight is calculated by multiplying the influence coefficient by the delay tolerance, i.e., 0.96 × 100 = 96. This rate adaptation weight will be used in the subsequent calculation of the theoretical rate adaptation value.

[0043] Step S1234: Determine a theoretical rate adaptation value based on the product of the initial rate setting value and the rate adaptation weight.

[0044] For example, the initial bitrate setting is 2000kbps and the bitrate adaptation weight is 96. The theoretical bitrate adaptation value is calculated by multiplying the initial bitrate setting by the bitrate adaptation weight, that is, 2000 × 96 = 192000kbps² (the units here are only for illustrative purposes and their actual meaning needs to be understood in context).

[0045] Step S1235: Calculate the rate adaptation deviation value according to the difference between the theoretical rate adaptation value and the actual rate change value.

[0046] For example, if the actual bitrate change during the period is 1800 kbps, the bitrate adaptation deviation is the theoretical bitrate adaptation value minus the actual bitrate change: 192000 - 1800 = 190200 kbps² (the units here are for illustrative purposes only; their actual meaning should be understood in context). This bitrate adaptation deviation reflects the difference between the theoretical and actual bitrates.

[0047] Step S124: performing semantic analysis on the user feedback features to extract the user's subjective rating features for audio-visual synchronization and the user's sensitivity features to delay.

[0048] In this embodiment, step S124 may include:

[0049] Step S1241: Acquire a keyword set contained in the user feedback text, where the keyword set includes first-category keywords related to audio-visual synchronization and second-category keywords related to delay.

[0050] In this example, a set of keywords was extracted from user A's feedback text. For example, the first category of keywords related to audio and video synchronization includes "audio and video out of sync" and "synchronization problem"; the second category of keywords related to latency includes "significant latency" and "stuttering." These keywords will serve as key information for subsequent analysis.

[0051] Step S1242: performing sentiment polarity analysis on the first category of keywords to determine the user's satisfaction score for the audio-visual synchronization, and mapping the satisfaction score into a numerical subjective scoring feature.

[0052] For example, let's analyze the sentiment polarity of the first-category keywords "audio-video asynchrony" and "synchronization problem." Assume the sentiment analysis algorithm identifies "audio-video asynchrony" as negative, and "synchronization problem" as negative. Based on the preset scoring rules, the satisfaction score corresponding to negative sentiment is 3 out of 10. This satisfaction score is then mapped to a numerical subjective score feature. Assume the mapping rule divides the score by 10, resulting in a subjective score feature of 0.3.

[0053] Step S1243: performing frequency statistics processing on the second-category keywords to determine the occurrence frequency and contextual association strength of the delay-related keywords.

[0054] For example, the frequency of occurrence of the second-category keywords "obvious delay" and "stuttering" can be counted. Assume that "obvious delay" appears twice and "stuttering" appears three times, with a total frequency of five times. The strength of contextual association is determined by analyzing factors such as the position of the keyword in the text and its collocation with other words. For example, "obvious delay" appears in a position describing a critical moment in the video playback process, and the contextual association strength is relatively high, set at 0.8; "stuttering" appears in a relatively less critical description position, and the contextual association strength is 0.6.

[0055] Step S1244: Generate the sensitivity feature based on the weighted sum of the occurrence frequency and the context association strength.

[0056] For example, assuming the weight of occurrence frequency is 0.6 and the weight of contextual relevance strength is 0.4, the sensitivity feature is calculated by multiplying the occurrence frequency by its weight plus the contextual relevance strength by its weight, that is, 5×0.6+(0.8×2+0.6×3)×0.4=3+(1.6+1.8)×0.4=3+3.4×0.4=3+1.36=4.36. This sensitivity feature reflects the user's sensitivity to delay.

[0057] Step S125: performing feature fusion on the bit rate adaptation deviation value, the subjective scoring feature, and the sensitivity feature to generate the audio-visual synchronization feature, the dynamic bit rate feature, and the decoding delay feature.

[0058] In this step, we can fuse the previously obtained rate adaptation bias value of 190200 kbps² (the units here are for illustrative purposes only; their practical significance should be understood in context), the subjective scoring feature of 0.3, and the sensitivity feature of 4.36. Assume that the fusion method combines these features according to certain weights. For example, the rate adaptation bias value has a weight of 0.5, the subjective scoring feature has a weight of 0.2, and the sensitivity feature has a weight of 0.3. First, normalize the rate adaptation bias value to a range of 0-1, obtaining a value of 0.9 (the specific normalization method will depend on the actual situation). The weighted concatenation is then performed to obtain the audio-visual synchronization feature of [0.9×0.5, 0.3×0.2, 4.36×0.3]=[0.45, 0.06, 1.308]. The dynamic bitrate feature can be obtained by further processing the bitrate adaptation deviation value. For example, the trend and amplitude of the dynamic bitrate adjustment are determined based on the size and direction of the bitrate adaptation deviation value. Assuming the dynamic bitrate feature is [0.6, 0.4], it indicates that there is a 60% probability of increasing the bitrate and a 40% probability of reducing the bitrate. The decoding delay feature is generated by combining the sensitivity feature and other related factors. Assuming the decoding delay feature is [0.7, 0.3], it indicates that there is a 70% probability of optimizing the decoding delay and a 30% probability of maintaining the current state.

[0059] Step S130: calling a pre-trained reinforcement learning optimization model to dynamically aggregate the audio and video synchronization features, dynamic bit rate features, and decoding delay features to generate a transmission quality evaluation result of the transmission session sequence.

[0060] In this embodiment, step S130 may include:

[0061] Step S131: inputting the audio-video synchronization feature into the first strategy sub-network to generate a first candidate strategy set for audio-video synchronization optimization.

[0062] For example, the previously generated audio-video synchronization features [0.45, 0.06, 1.308] can be input into the first strategy subnetwork. The first strategy subnetwork is specifically designed for optimizing audio-video synchronization. For example, the first strategy subnetwork comprises multiple neuron layers. After inputting the audio-video synchronization features, the first layer of neurons performs a weighted summation of the features. Assuming the weights are 0.2, 0.3, and 0.5, the calculated value is 0.45 × 0.2 + 0.06 × 0.3 + 1.308 × 0.5 = 0.09 + 0.018 + 0.654 = 0.762. The second layer of neurons transforms this result using a nonlinear function, such as the ReLU function, to obtain max(0, 0.762) = 0.762. Subsequent layers continue to process the data, ultimately generating a series of strategies for optimizing audio-video synchronization, forming the first candidate strategy set. Assume that the first candidate strategy set includes Strategy P1: Adjust the video frame rate to optimize audio-video synchronization; Strategy P2: Increase the audio delay to match the video.

[0063] Step S132: Input the dynamic bit rate feature into the second strategy sub-network to generate a second candidate strategy set for bit rate adaptation optimization.

[0064] For example, the dynamic bitrate feature [0.6, 0.4] can be input into the second strategy subnetwork. The second strategy subnetwork is optimized for bitrate adaptation. Similarly, it has a multi-layer structure. After the feature is input, the first layer of neurons performs weighted processing on the feature. Assuming the weights are 0.4 and 0.6 respectively, the calculation result is 0.6×0.4+0.4×0.6=0.24+0.24=0.48. After nonlinear transformation and processing in subsequent layers, a strategy for bitrate adaptation optimization is generated. For example, the second candidate strategy set includes strategy Q1: increase the bitrate to 2200kbps; strategy Q2: reduce the bitrate to 1800kbps.

[0065] Step S133: inputting the decoding delay feature into the third strategy sub-network to generate a third candidate strategy set for delay optimization.

[0066] For example, the decoding delay feature [0.7, 0.3] can be input into the third strategy subnetwork. The third strategy subnetwork focuses on latency optimization. After inputting the decoding delay feature, the first layer of the third strategy subnetwork calculates the feature. Assuming the weights are 0.5 and 0.5, the result is 0.7 × 0.5 + 0.3 × 0.5 = 0.35 + 0.15 = 0.5. After multiple layers of processing, a third set of candidate strategies is generated. For example, the third set of candidate strategies includes strategy R1: increasing the number of decoding threads; strategy R2: optimizing the decoding algorithm to reduce latency.

[0067] Step S134: performing policy conflict detection processing on the first candidate policy set, the second candidate policy set, and the third candidate policy set, eliminating mutually contradictory candidate policies, and obtaining remaining candidate policies.

[0068] In this embodiment, step S134 may include:

[0069] Step S1341: extracting the transmission parameter adjustment instruction set included in each candidate policy in the first candidate policy set, the second candidate policy set, and the third candidate policy set.

[0070] For example, for policy Q1 in the second candidate policy set, the transmission parameter adjustment instruction set is {increase the bitrate to 2200kbps}; the instruction set for policy Q2 is {reduce the bitrate to 1800kbps}. For policy R1 in the third candidate policy set, the transmission parameter adjustment instruction set is {increase the number of decoding threads}; the instruction set for policy R2 is {optimize the decoding algorithm to reduce latency}.

[0071] Step S1342: Detect whether there is a conflicting combination of a bit rate adjustment instruction and a buffer expansion instruction in the transmission parameter adjustment instruction set.

[0072] For example, the candidate policies listed above don't explicitly include instructions for buffer expansion, so this conflicting combination doesn't currently exist. However, if similar situations arise in other scenarios or across more policy sets, such as policies requiring both increasing the bitrate and increasing the buffer capacity, further analysis is necessary. Imagine a scenario where a policy set includes the instructions "increase the bitrate to 2500kbps" and "increase the buffer capacity from 500KB to 800KB." This would be a conflicting combination. Increasing the bitrate may increase data transmission volume, while simultaneously increasing the buffer capacity may lead to resource allocation conflicts, requiring more complex resource management and scheduling.

[0073] Step S1343: If so, calculate the failure probability of the conflicting combination in the historical transmission data. If the failure probability exceeds a preset threshold, eliminate the corresponding candidate strategy.

[0074] Suppose historical transmission data shows that when executing both the bitrate increase and buffer expansion commands simultaneously, 60 out of 100 transmissions experience transmission instability, lags, or even interruptions. This gives a 60% failure probability. If the preset threshold is 50%, since 60% exceeds the threshold, candidate policies containing this conflicting combination need to be eliminated. This is because a high failure probability means that this policy combination will not be effective in practice and may severely impact the user experience.

[0075] Step S1344: Detect whether there is an instruction combination of increasing the number of decoding threads and reducing the bit rate at the same time. If so, and the current terminal hardware performance parameters do not meet the parallel processing requirements, eliminate the corresponding candidate strategy.

[0076] For example, the current candidate policy set includes a potentially conflicting combination: policy Q2 (reducing the bitrate to 1800 kbps) and policy R1 (increasing the number of decoding threads). Assume that the current terminal hardware performance parameters are: 4 CPU cores and 8 GB / s memory bandwidth. Based on experience, increasing the number of decoding threads while reducing the bitrate requires at least 6 CPU cores and a memory bandwidth of at least 10 GB / s to meet parallel processing requirements. Because the current terminal hardware performance does not meet these requirements, the presence of this instruction combination could lead to insufficient system resources, making it impossible to effectively process audio and video data, thus affecting transmission quality and user experience. Therefore, the candidate policy containing this instruction combination is eliminated. After the aforementioned policy conflict detection process, the remaining candidate policy set is obtained.

[0077] Step S135: Perform weighted fusion based on the policy priority scores of the remaining candidate policies to generate a comprehensive transmission optimization policy, and generate the transmission quality evaluation result based on the effect prediction value of the comprehensive transmission optimization policy in the simulated transmission environment.

[0078] In this embodiment, step S135 may include:

[0079] Step S1351: Obtain the execution success rate, execution efficiency score, and user satisfaction score of each remaining candidate strategy in the historical transmission data.

[0080] Assume that after conflict detection, the remaining candidate strategies are strategy P1 (adjusting the video frame rate to optimize audio and video synchronization), strategy Q1 (increasing the bitrate to 2200 kbps), and strategy R2 (optimizing the decoding algorithm to reduce latency). Analysis of historical transmission data shows that strategy P1 has a 70% execution success rate, meaning that out of the past 100 transmissions using this strategy, 70 have successfully optimized audio and video synchronization. Its execution efficiency score is 6 out of 10, which is a comprehensive assessment of factors such as the time and resource consumption required for strategy execution. Its user satisfaction score is 7, reflecting the user experience after applying this strategy. Strategy Q1 has a 60% execution success rate, a 5% execution efficiency score, and a 6% user satisfaction score. Strategy R2 has a 75% execution success rate, a 7% execution efficiency score, and a 8% user satisfaction score.

[0081] Step S1352: Determine the strategy effectiveness coefficient according to the product of the execution success rate and the execution efficiency score.

[0082] For example, for strategy P1, the strategy effectiveness coefficient = 70% × 6 = 4.2. The strategy effectiveness coefficient for strategy Q1 = 60% × 5 = 3. The strategy effectiveness coefficient for strategy R2 = 75% × 7 = 5.25. The above strategy effectiveness coefficients reflect the overall performance of each strategy in terms of execution success and execution efficiency.

[0083] Step S1353: Determine the policy priority weight according to the ratio of the user satisfaction score to a preset satisfaction threshold.

[0084] For example, assume the preset satisfaction threshold is 5 points. For strategy P1, the strategy priority weight is 7 ÷ 5 = 1.4. The strategy priority weight for strategy Q1 is 6 ÷ 5 = 1.2. The strategy priority weight for strategy R2 is 8 ÷ 5 = 1.6. These strategy priority weights reflect the importance of user satisfaction with each strategy in the comprehensive evaluation.

[0085] Step S1354: performing weighted summation on the strategy effectiveness coefficient and the strategy priority weight to obtain a comprehensive priority score for each candidate strategy.

[0086] For example, assuming the weight of the policy effectiveness coefficient is 0.6 and the weight of the policy priority is 0.4, for policy P1, the overall priority score = 4.2 × 0.6 + 1.4 × 0.4 = 2.52 + 0.56 = 3.08. The overall priority score of policy Q1 = 3 × 0.6 + 1.2 × 0.4 = 1.8 + 0.48 = 2.28. The overall priority score of policy R2 = 5.25 × 0.6 + 1.6 × 0.4 = 3.15 + 0.64 = 3.79.

[0087] Step S1355: Select the top N candidate strategies with the highest comprehensive priority scores, remove duplicates and merge the transmission parameter adjustment instructions in the top N candidate strategies to generate the comprehensive transmission optimization strategy.

[0088] For example, assuming N=2, the two strategies with the highest overall priority scores are strategy R2 and strategy P1. The transmission parameter adjustment instruction for strategy R2 is {Optimize decoding algorithm to reduce latency}, while the transmission parameter adjustment instruction for strategy P1 is {Adjust video frame rate to optimize audio and video synchronization}. After deduplicating and merging these instructions, the resulting overall transmission optimization strategy is: simultaneously optimizing the decoding algorithm to reduce latency and adjusting the video frame rate to optimize audio and video synchronization.

[0089] Step S1356: Generate the transmission quality evaluation result according to the predicted value of the effect of the comprehensive transmission optimization strategy in the simulated transmission environment.

[0090] In this embodiment, the comprehensive transmission optimization strategy is tested in a simulated transmission environment. It is assumed that the settings of the simulated transmission environment are similar to the actual transmission environment, including parameters such as network bandwidth, packet loss rate, and terminal hardware performance. After multiple simulation tests, it was found that after applying the comprehensive transmission optimization strategy, the audio and video synchronization problem was significantly improved, the video fluency was improved, and the user feedback satisfaction was also improved. Based on the above simulation test results, it is predicted that good results can also be achieved in actual transmission. Based on the above effect prediction values, a transmission quality evaluation result is generated. For example, if the simulation test finds that the audio and video synchronization error is reduced from the original average of 50 milliseconds to 20 milliseconds, and the number of video freezes is reduced from 10 times per hour to 3 times, the transmission quality evaluation result can be rated as "good" based on the above specific data and preset evaluation criteria. The transmission quality evaluation result reflects the expected transmission quality of the current transmission session sequence after the application of the comprehensive transmission optimization strategy.

[0091] Step S140: determining a dynamic transmission optimization strategy based on the transmission quality evaluation result, and feeding back the dynamic transmission optimization strategy to the audio and video transmission system to trigger a real-time transmission parameter adjustment operation.

[0092] In this embodiment, step S140 may include:

[0093] Step S141: Divide the transmission quality evaluation result into multiple quality levels, wherein each quality level corresponds to a different optimization strategy template.

[0094] Assume that the transmission quality assessment results are divided into four levels: "excellent", "good", "average" and "poor". The optimization strategy template corresponding to the "excellent" level focuses on maintaining the current transmission status and making slight optimization adjustments to maintain high-quality transmission. For example, the bit rate may be fine-tuned to adapt to slight fluctuations in the network while maintaining efficient scheduling of decoding threads. The optimization strategy template corresponding to the "good" level aims to further improve the transmission quality. It may optimize the video encoding parameters or add some additional buffering mechanisms. The optimization strategy template corresponding to the "average" level requires a larger adjustment of the transmission parameters, such as adjusting the bit rate range, optimizing the decoding algorithm, etc. The optimization strategy template corresponding to the "poor" level may involve more radical measures such as reselecting the transmission path and changing the encoding and decoding scheme.

[0095] Step S142: Detecting the target quality level to which the transmission quality evaluation result belongs, and matching an optimization policy template corresponding to the target quality level from a predefined policy library.

[0096] For example, if the previously generated transmission quality assessment result is "Good," the corresponding optimization policy template can be searched from a predefined policy library. This library stores various optimization policy templates for different quality levels, derived through extensive experimentation and data analysis. Once the optimization policy template corresponding to "Good" is found, it may include some basic adjustment instructions, such as appropriately increasing the bitrate to enhance video quality and optimizing buffer management strategies to reduce data loss.

[0097] Step S143: Dynamically adjust the parameter thresholds in the optimization policy template according to the network fluctuation characteristics of the current transmission session sequence to generate a dynamic transmission optimization policy adapted to the current network status.

[0098] In this embodiment, step S143 may include:

[0099] Step S1431: extracting the bandwidth change rate gradient sequence, packet loss rate distribution trend, and jitter peak-valley interval duration from the network fluctuation characteristics of the current transmission session sequence.

[0100] For example, looking back at the network fluctuation characteristics obtained earlier, the bandwidth change rate gradient sequence is assumed to be [0.05, 0.1, -0.05, 0.03], which represents the bandwidth change rate gradient at different time periods in the transmission session sequence; the packet loss rate distribution trend shows a gradual increase, from 0.01 at the beginning to 0.03 later; the average duration of the jitter peak-to-valley interval is 8 milliseconds.

[0101] Step S1432: Calculate a bandwidth stability index based on the bandwidth change rate gradient sequence, and determine a packet loss risk level according to the degree of overlap between the packet loss rate distribution trend and a preset packet loss tolerance curve.

[0102] In this embodiment, the bandwidth stability index is calculated by assuming that the calculation method is to sum the absolute values of the bandwidth change rate gradient sequence and then divide it by the sequence length. For the bandwidth change rate gradient sequence [0.05, 0.1, -0.05, 0.03], the absolute value sequence is [0.05, 0.1, 0.05, 0.03], the sum is 0.05+0.1+0.05+0.03=0.23, and the sequence length is 4, then the bandwidth stability index = 0.23÷4=0.0575. The smaller the bandwidth stability index, the more stable the bandwidth. The preset packet loss tolerance curve is set based on historical data and experience, and the packet loss rate distribution trend is compared with the preset packet loss tolerance curve. Assuming that the packet loss rate distribution trend is close to the preset packet loss tolerance curve in most time periods, but exceeds the tolerance curve in certain time periods, the comprehensive judgment of the packet loss risk level is "medium".

[0103] Step S1433: performing correlation analysis on the duration of the jitter peak-valley interval and the buffer capacity parameter in the encoding and decoding parameter characteristics to generate a cumulative impact coefficient of the jitter on the decoding delay.

[0104] For example, given an average jitter peak-to-valley interval duration of 8 milliseconds and a buffer capacity of 500KB, analysis of historical data and experimental results reveals a relationship between jitter peak-to-valley interval duration and decoding latency. Assume that analysis indicates that for every 1 millisecond increase in jitter peak-to-valley interval duration, decoding latency increases by 0.1 milliseconds at the current buffer capacity. Therefore, the cumulative impact of an 8-millisecond jitter peak-to-valley interval duration on decoding latency is 8 × 0.1 = 0.8.

[0105] Step S1434: adjusting the dynamic bitrate adjustment threshold in the optimization policy template according to the bandwidth stability index, lowering the buffer overflow protection threshold according to the packet loss risk level, and synchronously correcting the decoding thread scheduling interval parameter in combination with the cumulative impact coefficient to generate a dynamic parameter adjustment rule.

[0106] For example, based on the bandwidth stability index of 0.0575, since the index is relatively small, indicating that the bandwidth is relatively stable, the dynamic bitrate adjustment threshold range can be appropriately narrowed. Assume that the original dynamic bitrate adjustment threshold was ±200 kbps, and after adjustment it is ±150 kbps. Based on the "medium" packet loss risk level, the buffer overflow protection threshold is lowered. The original buffer overflow protection threshold was 450 KB, and after adjustment it is 400 KB. Combined with the cumulative impact coefficient of 0.8, the decoding thread scheduling interval parameters are revised. Assume that the original scheduling interval was 10 milliseconds, and now it is adjusted to 10 + 0.8 = 10.8 milliseconds. The above adjusted parameters form the dynamic parameter adjustment rule.

[0107] Step S1435: injecting the dynamic parameter adjustment rules into the rate adaptation module, buffer management module and decoding control module of the optimization strategy template to generate a dynamic transmission optimization strategy adapted to the current network fluctuation characteristics.

[0108] In this embodiment, the rate adaptation module generates rate adjustment step size instructions based on the difference between the dynamic rate adjustment threshold and the current actual bandwidth utilization. These rate adjustment step size instructions are dynamically weighted with the buffer capacity parameter to generate adaptive rate adjustment instructions that take into account both bandwidth fluctuation and buffer capacity. Assume that the current actual bandwidth utilization is 1800kbps and the dynamic rate adjustment threshold is ±150kbps. After calculating the difference, rate adjustment step size instructions are generated based on the magnitude and direction of the difference. For example, if the difference is positive and within a certain range, an instruction for increasing the rate by 50kbps is generated. This instruction is dynamically weighted with the buffer capacity parameter of 500KB. Assuming that the buffer capacity has a weight of 0.4 and the rate difference has a weight of 0.6, after calculation and processing, the adaptive rate adjustment instruction generated is: when the bandwidth is stable and there is room for improvement, gradually increase the rate to 2000kbps, but adjust appropriately based on buffer usage.

[0109] In the buffer management module, the memory allocation ratio between the receive buffer and the decode buffer is redistributed based on the lowered buffer overflow protection threshold. The trigger frequency for redundant data retransmission requests is dynamically set based on the packet loss risk level, generating buffer expansion and data recovery linkage instructions. The lowered buffer overflow protection threshold is 400KB. The memory allocation ratio between the receive buffer and the decode buffer is redistributed, assuming it is adjusted from the original 6:4 to 5:5. Based on the "medium" packet loss risk level, the trigger frequency for redundant data retransmission requests is dynamically set. When the packet loss rate reaches 0.03, a redundant data retransmission request is triggered, and the number and frequency of retransmissions are dynamically adjusted based on factors such as the number of consecutive packet losses and the interval time.

[0110] In the decoding control module, the priority queue of the audio and video decoding tasks is adjusted according to the decoding thread scheduling interval parameter, and the resource occupancy rate of the hardware decoder is dynamically allocated based on the cumulative impact coefficient to generate a decoding delay balancing control instruction. According to the adjusted decoding thread scheduling interval parameter of 10.8 milliseconds, the priority queue of the audio and video decoding tasks is adjusted. For video decoding tasks, if the data volume is large and may cause increased delay, it will be processed first. Based on the cumulative impact coefficient of 0.8, the resource occupancy rate of the hardware decoder is dynamically allocated. Assuming that the total resources of the hardware decoder are 100%, according to the cumulative impact coefficient, 58% of the resources are allocated to video decoding and 42% of the resources are allocated to audio decoding to achieve balanced control of decoding delay.

[0111] Cross-dependency detection is performed on the adaptive bit rate adjustment instructions, buffer expansion and data recovery linkage instructions, and decoding delay equalization control instructions to eliminate resource competition conflicts caused by threshold adjustment between instructions and generate a dynamic transmission optimization strategy under mutually exclusive operation constraints. For example, it is detected that the adaptive bit rate adjustment instructions may affect the usage of the buffer, and the buffer expansion and data recovery linkage instructions are also related to the buffer. Through analysis and adjustment, it is determined that when the bit rate is adjusted, the buffer expansion operation should be suspended first, and the relevant buffer operations should be performed after the bit rate adjustment stabilizes to avoid resource competition conflicts. After this processing, the final dynamic transmission optimization strategy is generated.

[0112] Step S144: parsing the bit rate adjustment instructions, buffer management instructions and decoding priority instructions in the dynamic transmission optimization strategy.

[0113] For example, the bitrate adjustment instructions in the dynamic transmission optimization strategy are: when bandwidth is stable and there is room for improvement, gradually increase the bitrate to 2000kbps, but adjust appropriately based on buffer usage. The buffer management instructions are: adjust the memory allocation ratio between the receive buffer and the decoding buffer to 5:5. When the packet loss rate reaches 0.03, trigger a redundant data retransmission request, and dynamically adjust the retransmission frequency based on the packet loss situation. The decoding priority instructions are: based on the adjusted decoding thread scheduling interval parameter of 10.8 milliseconds, adjust the priority queue of audio and video decoding tasks, allocating 58% of the hardware decoder resources to video decoding and 42% to audio decoding.

[0114] Step S145: adjusting the encoding bit rate of the current video stream according to the bit rate adjustment instruction, and synchronously updating the sensitivity parameter of the adaptive bit rate algorithm.

[0115] In this embodiment, according to the bit rate adjustment instruction, the encoding bit rate of the current video stream is adjusted. If the current bit rate is lower than 2000kbps and the bandwidth is stable, the bit rate is gradually increased according to the instruction. Assuming that the current bit rate is 1800kbps, the bit rate is first increased by 50kbps to 1850kbps. At the same time, the sensitivity parameter of the adaptive bit rate algorithm is synchronously updated. The sensitivity parameter of the adaptive bit rate algorithm determines the algorithm's response speed to network fluctuations. Based on the current network fluctuations and the bit rate adjustment requirements, the sensitivity parameter is adjusted from the original 0.5 to 0.6, so that the algorithm can more quickly perceive network changes and make bit rate adjustments.

[0116] Step S146: dynamically allocating the memory ratio of the receiving buffer and the decoding buffer according to the buffer management instruction, and setting a buffer overflow protection threshold.

[0117] In this embodiment, according to the buffer management instructions, the memory ratio of the receive buffer and the decode buffer is adjusted from the original 6:4 to 5:5. Assuming the total memory is 1000KB, 500KB is allocated to the receive buffer and 500KB to the decode buffer. At the same time, the buffer overflow protection threshold is set to 400KB. This means that when the amount of data in the receive buffer or the decode buffer reaches 400KB, the system will take appropriate measures, such as suspending data reception or speeding up decoding, to prevent buffer overflow.

[0118] Step S147: adjusting the scheduling order of the audio and video decoding threads according to the decoding priority instruction, and allocating the usage priority of the hardware acceleration resources.

[0119] In this embodiment, the priority queue for audio and video decoding tasks can be adjusted based on the decode priority instruction and the adjusted decode thread scheduling interval parameter of 10.8 milliseconds. During video playback, video decoding tasks with large data volumes and potential for increased latency are advanced in the priority queue. For example, when a video contains complex scenes, such as a large number of dynamic special effects or high-resolution details, the corresponding video decoding tasks will be prioritized by the decoding thread.

[0120] At the same time, based on the instructions, 58% of the hardware decoder resources are allocated to video decoding, and 42% to audio decoding. Hardware acceleration resources include the computing power of the GPU and the processing power of the dedicated decoding chip. Assuming that the hardware decoder is the video decoding functional module integrated with the GPU, the total resources are measured by the number of GPU computing cores and bandwidth. For example, 58% of the GPU computing cores and corresponding bandwidth can be allocated to video decoding tasks, and 42% to audio decoding tasks. In this way, during the decoding process, video decoding can obtain more hardware acceleration resources first, thereby processing video data more efficiently, reducing decoding delay, achieving balanced control of decoding delay, and improving the overall audio and video transmission quality and playback effect.

[0121] In this embodiment, the training process of the reinforcement learning optimization model includes:

[0122] Step S210: Collect the audio and video synchronization feature set, dynamic bit rate feature set and decoding delay feature set of the sample transmission session sequence as the model input sample set, and obtain the user feedback score, transmission interruption rate and bit rate smoothness index of the corresponding transmission session as the quality assessment true value label set.

[0123] In this example, the scenario is still centered around user A watching a popular TV series on a video playback platform. To improve the user's viewing experience, the platform conducts in-depth analysis of the audio and video transmission of many users, and user A's viewing behavior is one of the important samples.

[0124] For each streaming session sequence of user A watching a TV series, the required feature sets can be collected. Regarding the audio-visual synchronization feature set, the subjective rating feature sequence is obtained through real-time user feedback during viewing. The platform provides convenient and fast user feedback channels, such as a rating button on the playback interface, allowing user A to rate the current audio-visual synchronization at any time, ranging from 1 to 10. In a 30-minute streaming session, each minute is used as a feedback interval. For example, in the first minute, user A intuitively gives the audio-visual synchronization a score of 7. In the second minute, feeling that the synchronization is slightly better, he gives it a score of 8. In the third minute, feeling that it is similar to the first minute, he gives it a score of 7. And so on, the final subjective rating feature sequence is [7, 8, 7, 9, 8, 7, 8, 9, 8, 7, 8, 9, 8, 7, 8, 9, 8, 7, 8, 9, 8, 7, 8, 9, 8, 7, 8, 9, 8, 7].

[0125] Acquiring a sequence of rate adaptation deviation values is relatively complex, requiring comprehensive consideration of multiple factors, including network fluctuation characteristics and codec parameters. During a transmission session, network fluctuation characteristics include data such as packet loss rate, jitter rate, and bandwidth variation rate. This network data is collected in real time and combined with information such as the initial bitrate setting, dynamic bitrate adjustment threshold, and buffer capacity parameters within the codec parameters, through a series of calculations. For example, network stability is first assessed based on data such as packet loss rate, jitter rate, and bandwidth variation rate. Then, combined with the initial bitrate setting and dynamic bitrate adjustment threshold, the bitrate that should theoretically be adapted to achieve optimal transmission performance at different times is calculated. This is then compared with the actual bitrate to determine the rate adaptation deviation value. After this series of operations, the resulting rate adaptation bias sequence is [40, -20, 30, 50, -10, 20, -5, 40, 60, 30, -20, 40, 50, 30, -10, 20, 30, -5, 40, 50, 30, -10, 20, 40, 50, 30, -10, 20, 30, -5]. The subjective rating feature sequence and the rate adaptation bias sequence are mapped one-to-one according to the transmission time period to construct a two-dimensional input vector. Specifically, the two-dimensional input vector corresponding to the first minute is composed of the subjective rating of 7 points and the rate adaptation bias value of 40, that is, [7, 40]; the two-dimensional input vector corresponding to the second minute is [8, -20]; the two-dimensional input vector corresponding to the third minute is [7, 30], and so on. This completes the audio-video synchronization feature set.

[0126] The collection of dynamic bitrate feature sets is equally important. The bitrate adaptation deviation value sequence is derived from the complex calculations previously described: [40, -20, 30, 50, -10, 20, -5, 40, 60, 30, -20, 40, 50, 30, -10, 20, 30, -5, 40, 50, 30, -10, 20, 40, 50, 30, -10, 20, 30, -5]. The bandwidth change trend sequence is derived through real-time monitoring of network bandwidth. During the entire 30-minute transmission session, the system continuously tracks changes in network bandwidth, recording them on a minute-by-minute basis. For example, in the first minute, the bandwidth change rate is 0.08; in the second minute, it becomes 0.06; in the third minute, it becomes 0.09, and so on. The resulting bandwidth change trend sequence is [0.08, 0.06, 0.09, 0.1, 0.07, 0.08, 0.06, 0.09, 0.11, 0.09, 0.06, 0.08, 0.1, 0.09, 0.07, 0.08, 0.09, 0.06, 0.09, 0.1, 0.09, 0.07, 0.08, 0.09, 0.1, 0.09, 0.07, 0.08, 0.09, 0.06]. This bandwidth change trend sequence clearly reflects the minute-by-minute trend of network bandwidth change. Combining the rate adaptation offset sequence with the bandwidth change trend sequence constructs the dynamic rate fluctuation input matrix. To describe it in words, the bit rate adaptation deviation value and the corresponding bandwidth change rate at each moment are put together to form a data set with a matrix-like structure, which facilitates the subsequent analysis and processing of dynamic bit rate characteristics.

[0127] The acquisition of the decoding delay feature set relies on accurate monitoring of the delay during the decoding process. In this transmission session, the system implements a precise timestamp recording mechanism during the decoding phase, recording detailed decoding delay data at minute intervals. For example, the decoding delay in the first minute is 12 milliseconds, in the second minute it is 10 milliseconds, in the third minute it is 11 milliseconds, and so on. The resulting decoding delay feature set is [12, 10, 11, 13, 10, 12, 10, 11, 14, 11, 10, 12, 13, 11, 10, 12, 11, 10, 11, 13, 11, 10, 12, 11, 10], with units measured in milliseconds. This decoding delay feature set intuitively displays the decoding delay at every moment throughout the transmission session.

[0128] At the same time, the system also collects user feedback scores, transmission interruption rates, and bitrate smoothness metrics for this transmission session as the ground truth label set for quality assessment. The user feedback score is a comprehensive rating given by user A based on their overall viewing experience after watching the entire TV series. After completing this viewing session, user A rated the overall experience as good and gave it an 8. The transmission interruption rate is determined by calculating the ratio of the number of interruptions during the transmission process to the total transmission duration. In this 30-minute transmission session, there were two brief transmission interruptions. The transmission interruption rate is 2 divided by 30, which is approximately 0.067. The bitrate smoothness metric measures bitrate fluctuations throughout the transmission process. It is calculated by analyzing and analyzing the temporal variation of the bitrate and reflects its stability. In this transmission session, after a series of complex calculations and analyses (which consider multiple factors such as the amplitude and frequency of bitrate fluctuations), the bitrate smoothness metric is determined to have a specific value (assuming a realistic value). This metric will serve as an important reference for measuring transmission quality in subsequent model training.

[0129] Step S220: Input the audio and video synchronization feature set into the first strategy sub-network for audio and video synchronization strategy inference training to generate a first quality impact coefficient sequence, and at the same time input the dynamic bit rate feature set into the second strategy sub-network for bit rate adaptation strategy inference training to generate a second quality impact coefficient sequence, and input the decoding delay feature set into the third strategy sub-network for delay control strategy inference training to generate a third quality impact coefficient sequence.

[0130] Step S221: Input the audio-video synchronization feature set into the first strategy sub-network to perform audio-video synchronization strategy inference training to generate a first quality impact coefficient sequence.

[0131] Step S2211: extract the subjective rating feature sequence and the rate adaptation deviation value sequence from the audio-visual synchronization feature, and construct a two-dimensional input vector in the order of the transmission period.

[0132] For example, in the example of user A watching a TV series, the subjective rating feature sequence [7, 8, 7, 9, 8, 7, 8, 9, 8, 7, 8, 9, 8, 7, 8, 9, 8, 7, 8, 9, 8, 7, 8, 9, 8, 7, 8, 9, 8, 7] and the rate adaptation deviation value sequence [40, -20, 30, 50, -10, 20, -5, 40, 60, 30, -20, 40, 50, 30, -10, 20, 30, -5, 40, 50, 30, -10, 20, 40, 50, 30, -10, 20, 30, -5] are accurately extracted from the collected audio and video synchronization feature set.

[0133] Then, strictly following the order of the transmission time periods, the corresponding elements in these two sequences are combined one by one to construct a two-dimensional input vector. Specifically, in the first minute of the transmission period, the subjective score of 7 and the rate adaptation bias of 40 are combined to form the first two-dimensional input vector [7, 40]. At the second minute, the subjective score of 8 and the rate adaptation bias of -20 are combined to form the two-dimensional input vector [8, -20]. At the third minute, the two-dimensional input vector [7, 30] is formed, and so on, until the 30th minute, a complete set of two-dimensional input vectors is formed. Each two-dimensional input vector in this set represents the characteristic information related to audio and video synchronization for a specific transmission period.

[0134] Step S2212: Perform a nonlinear transformation on the two-dimensional input vector through the three-layer fully connected layer of the first strategy sub-network, and output the quality attenuation coefficient of the user's perception of the audio and video synchronization anomaly in each transmission period.

[0135] The three fully connected layers of the first strategy sub-network process the input two-dimensional input vector step by step and in detail. Taking the first two-dimensional input vector [7, 40] as an example, when it enters the first fully connected layer, it will undergo weighted summation and nonlinear transformation operations.

[0136] The first fully connected layer has specific weights and bias parameters set internally. Assume the weights are 0.3 for a subjective score of 7, 0.7 for a rate adaptation bias of 40, and 0.1 for a bias parameter. Then, after a weighted summation calculation, the result is: 0.3 × 7 + 0.7 × 40 + 0.1 = 2.1 + 28 + 0.1 = 30.2. The calculation process here multiplies each element of the two-dimensional input vector by the corresponding weight, then adds the products and adds the bias value.

[0137] Next, the result is transformed through a nonlinear function. Let's assume the nonlinear function used here is the ReLU function (Rectified Linear Unit), which is defined as follows: when the input value is greater than 0, the function output is equal to the input value; when the input value is less than or equal to 0, the function output is 0. For the calculated result of 30.2, because it is greater than 0, after the ReLU function transformation, the output value remains 30.2.

[0138] The result of this first layer processing serves as the input to the second fully connected layer. This second fully connected layer also has its own independent weights and bias parameters. Assume that the weights corresponding to the output values of the first layer are 0.5 and 0.5, respectively (the weights here are assumed for simplicity and will be adjusted based on model training in practice), and the bias parameter is 0.3. The weighted sum is calculated again: 0.5 × 30.2 + 0.5 × 30.2 + 0.3 = 15.1 + 15.1 + 0.3 = 30.5. This result is then transformed nonlinearly using the ReLU function. Since 30.5 is greater than 0, the output value after transformation is still 30.5.

[0139] Finally, the output of the second layer enters the third fully connected layer. This layer also has corresponding weights and biases. Assume the weights are 0.8 and 0.2, respectively, and the bias is 0.5. The weighted sum is: 0.8 × 30.5 + 0.2 × 30.5 + 0.5 = 24.4 + 6.1 + 0.5 = 31. After the ReLU function nonlinear transformation, the final output value is 31, which represents the user-perceived quality attenuation coefficient of the audio-video synchronization anomaly during the first transmission period.

[0140] Following the same method, the two-dimensional input vector corresponding to each transmission period is processed through the same three-layer fully connected layer. This yields the user-perceived quality attenuation coefficient for audio-video synchronization anomalies within each transmission period. For example, the second two-dimensional input vector [8, -20], after undergoing the same three-layer fully connected layer processing, also yields a corresponding quality attenuation coefficient (the specific calculation process is similar to the above, ultimately resulting in a value consistent with the model's calculation results). In this way, a corresponding quality attenuation coefficient is generated for each transmission period.

[0141] Step S2213: Align and calibrate the quality attenuation coefficient with the occurrence time of synchronization-related complaint events in the user feedback score to generate a first quality impact coefficient sequence reflecting the correlation strength between the audio-visual synchronization feature and the quality assessment result.

[0142] While user A is watching a TV series, the time at which synchronization-related complaints occur is also recorded in the user feedback score. For example, at the 10th minute, user A submits a complaint about audio and video synchronization through the feedback channel, indicating that the audio and video were significantly out of sync.

[0143] The quality attenuation coefficient for each transmission period has been obtained through the processing of the first strategy sub-network. For example, the quality attenuation coefficient for the first minute is 31 (an assumed value), the quality attenuation coefficient for the second minute is another calculated value, and the quality attenuation coefficient for the tenth minute is 40 (an assumed value).

[0144] Now, the mass attenuation coefficient is aligned with the time of occurrence of the synchronous related complaint event. Specifically, the entire mass attenuation coefficient sequence is adjusted and correlated with the time point 10 minutes after the complaint event occurs.

[0145] Based on the occurrence of a complaint, the system will re-examine the previously calculated mass attenuation coefficient. The mass attenuation coefficient before the complaint may be appropriately weighted based on the time distance from the complaint moment and other relevant factors. For example, the mass attenuation coefficient at the 9th minute, due to its proximity to the complaint moment, may be given a relatively high weight, giving it a more important position in the final first mass influence coefficient sequence. Assume that after a series of complex calibration algorithms (involving a comprehensive consideration of multiple factors such as time distance and the severity of the complaint), the mass attenuation coefficient at the 9th minute is adjusted to a new value (assuming it is 38).

[0146] The mass attenuation coefficient after a complaint event is also adjusted accordingly. For example, the mass attenuation coefficient at the 11th minute may be adjusted differently than if no complaint event had occurred, due to the impact of the complaint event. After calibration, a new value (assuming it is 35) is obtained.

[0147] By comprehensively and meticulously calibrating the quality attenuation coefficients of each transmission period based on the occurrence of synchronization-related complaint events, a first quality impact coefficient sequence is ultimately generated that reflects the strength of the correlation between the audio-visual synchronization features and the quality assessment results. Each element in this first quality impact coefficient sequence is no longer simply a quality attenuation coefficient. Instead, it more accurately reflects the degree of correlation between the audio-visual synchronization features and the overall quality assessment results after comprehensively considering synchronization-related complaint events in user feedback. For example, after calibration, the first quality impact coefficient sequence may be [30, 32, 31, 33, 32, 31, 32, 33, 34, 40, 38, 36, 35, 34, 33, 32, 33, 34, 35, 36, 35, 34, 33, 32, 33, 34, 35, 36, 35, 34]. This first quality impact coefficient sequence will serve as an important basis for subsequent model training and evaluation.

[0148] Step S222: Input the dynamic bit rate feature set into the second strategy sub-network to perform bit rate adaptation strategy inference training to generate a second quality impact coefficient sequence.

[0149] Step S2221: extracting the bit rate adaptation deviation value sequence and the bandwidth change trend sequence from the dynamic bit rate feature, and constructing a dynamic bit rate fluctuation input matrix.

[0150] In the transmission session of user A watching a TV series, the rate adaptation deviation value sequence [40, -20, 30, 50, -10, 20, -5, 40, 60, 30, -20, 40, 50, 30, -10, 20, 30, -5, 40, 50, 30, -10, 20, 40, 50, 30, -10, 20, 30, -5] and the bandwidth change trend sequence [0.08, 0.06, 0.09, 0.1, 0.07, 0.08, 0.06, 0.09, 0.11, 0.09, 0.06, 0.08, 0.1, 0.09, 0.07, 0.08, 0.09, 0.06, 0.09, 0.1, 0.09, 0.07] are extracted from the dynamic rate feature set.

[0151] These two sequences are constructed into a dynamic bitrate fluctuation input matrix. Using the transmission period as the dimension, the rate adaptation deviation and bandwidth trend values corresponding to each period are combined. In the first transmission period, the rate adaptation deviation is 40 and the bandwidth trend is 0.08, forming a combined unit. In the second transmission period, the rate adaptation deviation is -20 and the bandwidth trend is 0.06, forming another combined unit. This combination is repeated for all 30 transmission periods.

[0152] Step S2222: The correlation pattern between the bit rate mutation event and the transmission interruption event is captured by the temporal convolution layer of the second strategy sub-network, and the negative impact weight of the bit rate instability on the quality assessment in each time period is output.

[0153] The second strategy subnetwork's temporal convolutional layer begins an in-depth analysis of the input matrix of dynamic bitrate fluctuations. This layer uses specially designed convolutional kernels that scan and process data along the temporal dimension to capture potential correlations between bitrate fluctuations and transmission interruptions.

[0154] During the first transmission period, the convolution kernel first acts on the first unit of the dynamic bitrate fluctuation input matrix (composed of a bitrate adaptation bias value of 40 and a bandwidth change trend value of 0.08) and several adjacent units (here we assume the first three units, but the actual number will be determined based on the size and design of the convolution kernel). The convolution kernel performs specific weighted calculations and combinations on the bitrate adaptation bias values and bandwidth change trend values in the above units. For example, a weight of 0.6 is assigned to the bitrate adaptation bias value of 40, and a weight of 0.4 is assigned to the bandwidth change trend value of 0.08. After a series of calculations (such as multiplication and addition), an intermediate result is obtained. This intermediate result undergoes a nonlinear transformation (such as the ReLU function transformation) to highlight important features in the data.

[0155] As time passes, the convolution kernel gradually moves along the time dimension, performing the same operation on the unit data of each transmission period. During this movement, it pays special attention to changes in the bit rate adaptation deviation value. If the bit rate adaptation deviation value changes significantly in a certain period of time, such as suddenly changing from a small value to a larger value (such as suddenly changing from -20 to 60), this may indicate the occurrence of a bit rate mutation event. At the same time, combined with relevant data such as the transmission interruption rate (in this scenario, it is assumed that the system will somehow incorporate the occurrence information of the transmission interruption event into the analysis process, such as marking the period of the transmission interruption event with a special identifier), it is analyzed to see whether there is a correlation between the bit rate mutation event and the transmission interruption event.

[0156] After traversing and processing the entire dynamic bitrate fluctuation input matrix, the temporal convolutional layer calculates the negative impact weight of bitrate instability on quality assessment for each time period based on the captured correlation patterns. For example, for the first transmission period, after a series of complex calculations and analyses, the negative impact weight is 0.3; for the second period, due to the specific combination of bitrate adaptation deviation and bandwidth variation trend, as well as the correlation analysis with possible transmission interruption events, the negative impact weight is 0.2. Each period is assigned a corresponding negative impact weight based on its own data characteristics and its correlation with transmission interruption events. This negative impact weight reflects the potential negative impact of bitrate instability on the overall quality assessment in each period.

[0157] Step S2223: performing sliding window averaging processing on the negative impact weights in combination with the bit rate smoothness true value label to generate a second quality impact coefficient sequence.

[0158] After determining the weights for the negative impact of bitrate instability on quality assessment during each time period, these weights need to be further processed in conjunction with the true bitrate smoothness label. The true bitrate smoothness label is a key metric obtained during the sample transmission session sequence. It reflects the actual smoothness of the bitrate during that transmission session.

[0159] Suppose we use a sliding window of size 3 to average the negative impact weights. Taking the first period as an example, the sliding window includes the negative impact weights for the first period and its two adjacent periods (i.e., the first, second, and third periods). The negative impact weight for the first period is 0.3, the second period is 0.2, and the third period is assumed to be 0.4.

[0160] First, calculate the sum of the negative impact weights within the sliding window: 0.3 + 0.2 + 0.4 = 0.9. Then, divide this sum by the sliding window size of 3 to obtain the average negative impact weight within the window: 0.9 ÷ 3 = 0.3. This average value is the sliding window average result for the first period and serves as the first element of the second quality impact coefficient sequence.

[0161] Next, the sliding window moves backward one period to include the negative impact weights of the second, third, and fourth periods. Assuming the negative impact weight of the fourth period is 0.5, the sum within this window is 0.2 + 0.4 + 0.5 = 1.1, and the average negative impact weight is 1.1 ÷ 3 ≈ 0.37, which is the second element of the second quality impact coefficient sequence.

[0162] In this way, the sliding window continues to move backward, calculating a new average each time, until all time periods have been processed. For example, when the sliding window moves to include the 28th, 29th, and 30th time periods, assuming the negative impact weight for the 28th time period is 0.25, the 29th time period is 0.3, and the 30th time period is 0.2, then the sum is 0.25 + 0.3 + 0.2 = 0.75, and the average negative impact weight is 0.75 ÷ 3 = 0.25, which is the 30th element in the second quality impact coefficient sequence.

[0163] Through this sliding window averaging process, a second quality impact coefficient sequence is generated. This second quality impact coefficient sequence comprehensively considers the negative impact weight of bitrate instability on quality assessment in each period and the overall situation of bitrate smoothness, and more accurately reflects the impact of dynamic bitrate characteristics on quality assessment. For example, the generated second quality impact coefficient sequence may be [0.3, 0.37, 0.33, 0.4, 0.35, 0.32, 0.3, 0.33, 0.38, 0.35, 0.32, 0.3, 0.33, 0.36, 0.34, 0.31, 0.3, 0.32, 0.35, 0.34, 0.32, 0.3, 0.31, 0.33, 0.35, 0.33, 0.31, 0.3, 0.32, 0.25].

[0164] Step S223: Input the decoding delay feature set into the third strategy sub-network to perform delay control strategy inference training to generate a third quality impact coefficient sequence.

[0165] Step S2231: extract the decoding delay value sequence from the decoding delay feature set, and construct a one-dimensional input vector in the order of the transmission period.

[0166] In the scenario where user A is watching a TV series, the decoding delay value sequence is directly extracted from the decoding delay feature set [12, 10, 11, 13, 10, 12, 10, 11, 14, 11, 10, 12, 13, 11, 10, 12, 11, 10, 11, 13, 11, 10, 12, 11, 13, 11, 10, 12, 11, 10]. These values are constructed into a one-dimensional input vector in the order of the transmission time periods. The decoding delay value of the first transmission period is 12, forming the first element of the one-dimensional input vector. The decoding delay value of the second transmission period is 10, forming the second element. The decoding delay value of the third transmission period is 11, forming the third element. And so on, all the way to the decoding delay value of 10 in the 30th transmission period, ultimately constructing the one-dimensional input vector [12, 10, 11, 13, 10, 12, 10, 11, 14, 11, 10, 12, 13, 11, 10, 12, 11, 10, 11, 13, 11, 10, 12, 11, 10]. This one-dimensional input vector will serve as the input data of the third policy subnetwork for subsequent analysis and processing.

[0167] Step S2232: performing feature extraction and transformation on the one-dimensional input vector through a specific neuron layer of the third strategy sub-network, and outputting the influence coefficient of the decoding delay in each time period on the quality assessment.

[0168] A specific neuron layer of the third strategy sub-network begins to process the input one-dimensional input vector. This specific neuron layer has a specially designed neuron structure and connection weights, which is intended to extract and decode key delay-related features and convert the above features into impact coefficients that are meaningful for quality assessment.

[0169] When a one-dimensional input vector enters a specific neuron layer, the first layer of neurons first performs a weighted summation on each element of the input vector. Assume that the weight corresponding to the first element (decoding delay value 12) is 0.4, the weight corresponding to the second element (decoding delay value 10) is 0.3, and the weight corresponding to the third element (decoding delay value 11) is 0.3 (the weights here are assumed for simplicity and will be adjusted in practice based on model training). This weighted summation yields an intermediate result: 0.4 × 12 + 0.3 × 10 + 0.3 × 11 = 4.8 + 3 + 3.3 = 11.1.

[0170] This intermediate result is transformed using a nonlinear function, such as the Sigmoid function (f(x)=frac{1}{1+e^{-x}}). Substituting 11.1 into the Sigmoid function yields a transformed value (assuming it's 0.95, but the actual result depends on the function), which is then passed to the next layer of neurons.

[0171] The second layer of neurons further processes the output of the first layer. This may involve another weighted summation and nonlinear transformation. Suppose the second layer assigns a weight of 0.6 to the output of the first layer. After this weighted calculation and nonlinear transformation (e.g., another nonlinear function), a new result is obtained. This process continues across multiple layers of neurons, each processing the data according to its own design to extract and decode deeper features related to latency.

[0172] After processing through multiple layers of neural networks, a coefficient is output for each transmission period to reflect the impact of decoding delay on quality assessment. For example, after a series of complex processing, the impact coefficient for the first transmission period is 0.8; the impact coefficient for the second transmission period is 0.7. Each transmission period is assigned a coefficient value based on its decoding delay value and the processing in the neural network layer, reflecting the degree of impact of the decoding delay on quality assessment during that period.

[0173] Step S2233: adjusting the influence coefficient in combination with the transmission interruption rate true value label to generate a third quality influence coefficient sequence.

[0174] After obtaining the influence coefficient of decoding delay on quality assessment in each period, it is necessary to adjust the above influence coefficient in combination with the true value label of transmission interruption rate. The true value label of transmission interruption rate reflects the actual transmission interruption situation in the entire transmission session.

[0175] Assume the true value label for the transmission interruption rate is 0.067. The impact coefficient for each time period can be adjusted based on the correlation analysis of the impact of transmission interruption events on decoding delay. If a transmission interruption event occurs near a certain transmission time period, the impact coefficient of the decoding delay in that time period on the quality assessment may be appropriately amplified.

[0176] For example, suppose a transmission interruption occurs at the 15th minute. The impact coefficients of decoding delay on quality assessment for the 14th, 15th, and 16th periods before and after the interruption, respectively, are 0.7, 0.8, and 0.7. Due to the transmission interruption, the system analyzes the potential relationship between the interruption and decoding delay. Considering that the interruption may cause decoding instability, which in turn affects quality assessment, the impact coefficients for these periods are adjusted. The assumed adjustment rule is to multiply the impact coefficients for the period of the interruption and the periods before and after by 1.2. (This adjustment rule is for illustrative purposes only; the actual determination will be based on extensive experimentation and data analysis.)

[0177] Then, the adjusted impact coefficient of the 14th period is 0.7×1.2=0.84; the adjusted impact coefficient of the 15th period is 0.8×1.2=0.96; and the adjusted impact coefficient of the 16th period is 0.7×1.2=0.84.

[0178] For other time periods not directly affected by the transmission outage, fine-tuning may also be performed based on the transmission outage rate and overall evaluation logic. For example, for time periods farther from the transmission outage, the impact coefficient may be adjusted at a smaller rate based on the transmission outage rate. Assume that for time periods farther from the transmission outage, the adjustment rule is to multiply the impact coefficient by (1 + transmission outage rate × 0.1). For time period 5, the original impact coefficient was 0.6, but after adjustment, it becomes 0.6 × (1 + 0.067 × 0.1) = 0.6 × (1 + 0.0067) = 0.6 × 1.0067 = 0.60402.

[0179] By combining the transmission interruption rate true value label to comprehensively adjust the impact coefficient of each time period, a third quality impact coefficient sequence is ultimately generated. This third quality impact coefficient sequence comprehensively considers the impact of decoding delay itself on quality assessment and the correction of the impact of transmission interruption events on decoding delay, more accurately reflecting the role of decoding delay characteristics in the overall quality assessment. For example, the generated third quality impact coefficient sequence may be [0.60402, 0.72, 0.75, 0.8, 0.7, 0.78, 0.7, 0.75, 0.84, 0.8, 0.7, 0.78, 0.84, 0.8, 0.84, 0.96, 0.84, 0.8, 0.75, 0.8, 0.8, 0.7, 0.78, 0.8, 0.84, 0.8, 0.7, 0.78, 0.8, 0.75]. This sequence will be used for subsequent fusion with other quality impact coefficient sequences and model training.

[0180] Step S230: Time window alignment is performed on the first quality impact coefficient sequence, the second quality impact coefficient sequence and the third quality impact coefficient sequence, and the comprehensive quality score prediction value of each transmission time period is calculated through the weighted fusion network. The comprehensive quality score prediction value and the quality assessment true value label set are subjected to error back propagation, and the parameters of the first strategy sub-network, the second strategy sub-network and the weighted fusion network are iteratively updated until convergence.

[0181] Step S231: performing time window alignment processing on the first quality impact coefficient sequence, the second quality impact coefficient sequence, and the third quality impact coefficient sequence.

[0182] In this embodiment, the first, second, and third quality impact coefficient sequences are all generated based on the same transmission period. However, in actual processing, slight differences in the calculation process or data acquisition may cause slight misalignment of the sequences in the time dimension. Therefore, time window alignment is required.

[0183] For example, the first quality influence coefficient sequence is [0.3, 0.4, 0.35, ...], the second quality influence coefficient sequence is [0.25, 0.3, 0.28, ...], and the third quality influence coefficient sequence is [0.4, 0.45, 0.42, ...]. Using the first transmission period as the benchmark, the three sequences are strictly aligned in time to ensure that the coefficients corresponding to the same transmission period in each sequence accurately match. After alignment, the three sequences are completely aligned in time.

[0184] Step S232: Calculate the comprehensive quality score prediction value of each transmission period through the weighted fusion network.

[0185] Step S2321: splicing the aligned first mass influence coefficient, the second mass influence coefficient, and the third mass influence coefficient into a three-dimensional input tensor according to the transmission period.

[0186] The three sequences, after time window alignment, are concatenated according to transmission time. In the first transmission time period, the first quality impact coefficient 0.3, the second quality impact coefficient 0.25, and the third quality impact coefficient 0.4 are concatenated together to form a three-dimensional vector [0.3, 0.25, 0.4]. In the second transmission time period, the corresponding coefficients are similarly concatenated into a three-dimensional vector [0.4, 0.3, 0.45]. This concatenation is repeated for all transmission time periods, ultimately forming a three-dimensional input tensor. Each dimension of this tensor corresponds to one of the three quality impact coefficient sequences, and the time dimension corresponds to each transmission time period, comprehensively integrating quality impact information from all three aspects.

[0187] Step S2322: Capture the temporal dependencies across feature dimensions through a double-layer gated recurrent unit, and output the joint influence strength of multiple factors in each time period.

[0188] A two-layer gated recurrent unit (GRU) processes a three-dimensional input tensor. When the three-dimensional vector [0.3, 0.25, 0.4] in the first transmission period enters the first GRU layer, the GRU performs a comprehensive analysis of the three coefficients based on its internal gating mechanism and weight parameters. For example, by controlling the update and reset gates, it determines the importance of each coefficient at the current moment and captures the relationships between them. Suppose that after processing by the first GRU layer, the intermediate result vector [0.32, 0.27, 0.41] is obtained.

[0189] This intermediate result vector enters the second-layer GRU, which further processes it and deeply explores the temporal dependencies across feature dimensions. For example, it considers the changes in coefficients between the current period and the previous period, as well as the long-term and short-term dependencies between different feature coefficients. After processing by the second-layer GRU, the output is the combined impact strength of multiple factors for each period. For the first transmission period, assuming the output joint impact strength is 0.35, it comprehensively reflects the combined impact of the three factors of audio and video synchronization, bit rate adaptation, and decoding delay on the quality during this period.

[0190] Step S2323: Map the joint impact strength to a quality score interval, superimpose the baseline quality parameter of the transmission session level, and generate an end-to-end comprehensive quality score prediction value.

[0191] Assume the quality score interval is [0, 10]. Use a mapping function to map the combined impact strength output in each time period to the quality score interval. For example, if the combined impact strength output in the first transmission period is 0.35, the mapped score is (10 * 0.35 = 3.5).

[0192] At the same time, consider the baseline quality parameter at the transmission session level. Assume the baseline quality parameter is 5, a fixed parameter that comprehensively considers factors such as the quality of the video content itself and the platform's basic service level. The mapped score is superimposed on the baseline quality parameter to obtain the predicted comprehensive quality score for the first transmission period: (3.5 + 5 = 8.5). This calculation is repeated for each transmission period, ultimately generating a sequence of end-to-end comprehensive quality score predictions.

[0193] Step S233: performing error back propagation on the comprehensive quality score prediction value and the quality assessment true value label set, and iteratively updating the parameters of the first strategy sub-network, the second strategy sub-network and the weighted fusion network until convergence.

[0194] Step S2331: Calculate the mean square error between the predicted comprehensive quality score and the true value of the user feedback score as the main loss function.

[0195] Assume that the true user feedback rating is 8 points, and the predicted comprehensive quality rating for the first transmission period is 8.5 points. The mean squared error (MSE) is calculated by averaging the squares of the differences between the predicted and true values. For the first transmission period, the error is (8.5 - 8 = 0.5), which is squared to (0.5^2 = 0.25). This calculation is repeated for all transmission periods and then averaged to obtain the value of the main loss function. Assume that the average mean squared error for all 30 transmission periods is 0.3. This main loss function value reflects the overall error between the predicted comprehensive quality rating and the true user feedback rating.

[0196] Step S2332: Add the cross entropy loss between the output of the first strategy sub-network and the true value of the rate adaptation deviation as an auxiliary constraint item.

[0197] Assume that the output of the first policy subnetwork related to the rate adaptation bias is [predicted bias 1, predicted bias 2, …], and the true rate adaptation bias is [true bias 1, true bias 2, …]. Cross-entropy loss measures the difference between two probability distributions. Here, it measures the difference between the output of the first policy subnetwork and the true rate adaptation bias. By calculating the cross-entropy loss, we obtain the value of an auxiliary constraint. Assume that the calculated value of this auxiliary constraint is 0.2.

[0198] Step S2333: Use a dynamic weight distribution algorithm to balance the ratio of the main loss function and the auxiliary constraint terms, and control the parameter update amplitude.

[0199] In this embodiment, the dynamic weight allocation algorithm automatically adjusts the weights of the main loss function and auxiliary constraints based on the training process. For example, in the early stages of training, the weight of the main loss function might be set to 0.8, and the weight of the auxiliary constraints might be set to 0.2. As training progresses, the weights are dynamically adjusted based on the model's convergence and performance. Suppose that after several iterations, the weight of the main loss function is adjusted to 0.7, and the weight of the auxiliary constraints is adjusted to 0.3.

[0200] Based on the adjusted weights, calculate the total loss. Assume the total loss is (0.8*0.3+0.2*0.2=0.24+0.04=0.28). Then, perform error backpropagation based on this total loss. Use the backpropagation algorithm to calculate the gradient and update the parameters of the first and second policy sub-networks and the weighted fusion network.

[0201] In each iteration, the process of calculating loss, adjusting weights, backpropagating errors, and updating parameters is repeated until the model's loss stops decreasing, indicating convergence. At this point, the parameters of the first and second strategy subnetworks, as well as the weighted fusion network, have reached optimal levels, enabling more accurate assessment and optimization of audio and video transmission quality.

[0202] In practical application scenarios, the above user experience optimization method can be applied to various online video platforms, live streaming platforms, and video conferencing systems. For example, online video platforms host a large number of users accessing the platform daily from diverse network environments and terminal devices. This optimization method monitors the transmission quality of each user terminal in real time and enables targeted optimization based on network fluctuations, codec parameters, and user feedback.

[0203] It is worth noting that in terms of data security and privacy protection, this embodiment strictly adheres to relevant laws, regulations, and privacy policies when acquiring and processing transmission quality data. User data is encrypted to ensure that user personal information and transmitted data are not leaked. During data storage and transmission, secure storage mechanisms and encryption protocols are employed to prevent unauthorized access and tampering.

[0204] Figure 2 A schematic diagram illustrates exemplary hardware and software components of a user experience optimization system 100 for an audio and video transmission system, which can implement the concepts of the present invention, according to some embodiments of the present invention. For example, a processor 120 can be used in the user experience optimization system 100 for an audio and video transmission system to perform the functions of the present invention.

[0205] The user experience optimization system 100 for audio and video transmission systems can be a general-purpose server or a special-purpose server, both of which can be used to implement the user experience optimization method for audio and video transmission systems of the present invention. Although only one server is shown in the present invention, for convenience, the functions described in the present invention can be implemented in a distributed manner on multiple similar platforms to balance the processing load.

[0206] For example, the user experience optimization system 100 for an audio and video transmission system may include a network port 110 connected to a network, one or more processors 120 for executing program instructions, a communication bus 130, and storage media 140 in different forms, such as a disk, ROM, or RAM, or any combination thereof. Exemplarily, the user experience optimization system 100 for an audio and video transmission system may also include program instructions stored in ROM, RAM, or other types of non-transitory storage media, or any combination thereof. The method of the present invention can be implemented according to the above-mentioned program instructions. The user experience optimization system 100 for an audio and video transmission system also includes an input / output (I / O) interface 150 between the computer and other input and output devices.

[0207] For ease of explanation, only one processor is described in the user experience optimization system 100 for an audio and video transmission system. However, it should be noted that the user experience optimization system 100 for an audio and video transmission system in the present invention may also include multiple processors, so the steps performed by one processor described in the present invention may also be performed jointly or individually by multiple processors. For example, if the processor of the user experience optimization system 100 for an audio and video transmission system executes step A and step B, it should be understood that step A and step B may also be performed jointly by two different processors or individually in one processor. For example, the first processor executes step A, the second processor executes step B, or the first processor and the second processor execute steps A and B together.

[0208] In addition, an embodiment of the present invention further provides a readable storage medium, in which computer-executable instructions are preset. When a processor executes the computer-executable instructions, the above-mentioned user experience optimization method for the audio and video transmission system is implemented.

[0209] It should be noted that in order to simplify the description of the present invention and thus help understand one or more embodiments of the invention, in the foregoing description of the embodiments of the present invention, multiple features are sometimes combined into one embodiment, figure or description thereof.

Claims

1. A method for optimizing user experience of an audio and video transmission system, characterized in that: The method comprises: Acquire a transmission quality data set generated by a target terminal during audio and video transmission, the transmission quality data set comprising multiple transmission session sequences, each transmission session sequence comprising network fluctuation characteristics, encoding and decoding parameter characteristics, and user feedback characteristics during at least one transmission period; Performing transmission feature extraction processing on the transmission quality data set to obtain audio and video synchronization features, dynamic bit rate features, and decoding delay features of each transmission session sequence; Invoking a pre-trained reinforcement learning optimization model to dynamically aggregate the audio and video synchronization features, the dynamic bit rate features, and the decoding delay features to generate a transmission quality evaluation result of the transmission session sequence; Determining a dynamic transmission optimization strategy based on the transmission quality evaluation result, and feeding back the dynamic transmission optimization strategy to the audio and video transmission system to trigger a real-time transmission parameter adjustment operation; The performing transmission feature extraction processing on the transmission quality data set to obtain audio and video synchronization features, dynamic bit rate features, and decoding delay features of each transmission session sequence includes: Performing time series segmentation processing on the network fluctuation characteristics in the transmission session sequence to obtain a plurality of network fluctuation sub-segments; Call the pre-trained encoder model to encode each network fluctuation sub-segment and generate the fluctuation intensity vector of each network fluctuation sub-segment; Determining a rate adaptation deviation value for each transmission period based on a correlation calculation between the fluctuation intensity vector and the encoding and decoding parameter characteristics; Performing semantic analysis on the user feedback features to extract the user's subjective rating features on audio-visual synchronization and sensitivity features to delay; The bit rate adaptation deviation value, the subjective scoring feature and the sensitivity feature are fused to generate the audio-visual synchronization feature, the dynamic bit rate feature and the decoding delay feature.

2. The user experience optimization method for an audio and video transmission system according to claim 1, characterized in that: The calling of the pre-trained encoder model to encode each network fluctuation sub-segment and generate a fluctuation intensity vector for each network fluctuation sub-segment includes: Obtain the packet loss rate sequence, jitter rate sequence, and bandwidth change rate sequence contained in each network fluctuation sub-segment; Normalizing the packet loss rate sequence to obtain a packet loss rate distribution feature; Performing sliding window statistical processing on the jitter rate sequence to generate jitter peak-to-valley difference features and jitter duration features; Performing gradient calculation on the bandwidth change rate sequence to obtain bandwidth change trend characteristics; The packet loss rate distribution characteristics, jitter peak-to-valley difference characteristics, jitter duration characteristics, and bandwidth change trend characteristics are input into the encoder model, and the fluctuation intensity vector is generated through multi-layer nonlinear transformation.

3. The user experience optimization method for an audio and video transmission system according to claim 1, characterized in that: The determining of the rate adaptation deviation value for each transmission period based on the correlation calculation between the fluctuation intensity vector and the encoding and decoding parameter feature includes: Extracting an initial bit rate setting value, a dynamic bit rate adjustment threshold, and a buffer capacity parameter from the encoding and decoding parameter characteristics; Calculating the cosine similarity between the fluctuation intensity vector and the dynamic bit rate adjustment threshold to obtain the influence coefficient of network fluctuation on bit rate; Determining a delay tolerance for rate adjustment based on the buffer capacity parameter, and generating a rate adaptation weight in combination with the impact coefficient; Determining a theoretical rate adaptation value based on a product of the initial rate setting value and the rate adaptation weight; The rate adaptation deviation value is calculated according to the difference between the theoretical rate adaptation value and the actual rate change value.

4. The user experience optimization method for an audio and video transmission system according to claim 1, characterized in that: The semantic parsing of the user feedback features to extract the user's subjective rating features for audio-visual synchronization and the user's sensitivity features to delay includes: Obtaining a keyword set contained in the user feedback text, the keyword set including a first category of keywords related to audio and video synchronization and a second category of keywords related to delay; Performing sentiment polarity analysis on the first category of keywords to determine a user's satisfaction score for audio-visual synchronization, and mapping the satisfaction score into a numerical subjective scoring feature; Performing frequency statistics on the second category of keywords to determine the frequency of occurrence and contextual relevance strength of delay-related keywords; The sensitivity feature is generated based on a weighted sum of the occurrence frequency and the context association strength.

5. The method for optimizing user experience of an audio and video transmission system according to claim 1, wherein: The reinforcement learning optimization model includes a first strategy sub-network, a second strategy sub-network, and a third strategy sub-network. The pre-trained reinforcement learning optimization model is called to dynamically aggregate the audio and video synchronization features, the dynamic bit rate features, and the decoding delay features to generate a transmission quality evaluation result of the transmission session sequence, including: Inputting the audio-video synchronization feature into a first strategy sub-network to generate a first candidate strategy set for audio-video synchronization optimization; Inputting the dynamic bit rate feature into a second strategy sub-network to generate a second candidate strategy set for bit rate adaptation optimization; Inputting the decoding delay feature into a third strategy sub-network to generate a third candidate strategy set for delay optimization; Performing policy conflict detection on the first candidate policy set, the second candidate policy set, and the third candidate policy set, eliminating conflicting candidate policies, and obtaining remaining candidate policies; A weighted fusion is performed based on the policy priority scores of the remaining candidate policies to generate a comprehensive transmission optimization policy, and the transmission quality evaluation result is generated based on the effect prediction value of the comprehensive transmission optimization policy in the simulated transmission environment.

6. The method for optimizing user experience of an audio and video transmission system according to claim 5, wherein: The performing policy conflict detection on the first candidate policy set, the second candidate policy set, and the third candidate policy set to eliminate conflicting candidate policies includes: Extracting a transmission parameter adjustment instruction set included in each candidate policy in the first candidate policy set, the second candidate policy set, and the third candidate policy set; Detecting whether there is a conflicting combination of a bit rate adjustment instruction and a buffer expansion instruction in the transmission parameter adjustment instruction set; If so, the failure probability of the conflicting combination in the historical transmission data is calculated, and if the failure probability exceeds a preset threshold, the corresponding candidate strategy is eliminated; Check whether there is an instruction combination that increases the number of decoding threads and reduces the bit rate at the same time. If so, and the current terminal hardware performance parameters do not meet the parallel processing requirements, the corresponding candidate strategy is eliminated.

7. The method for optimizing user experience of an audio and video transmission system according to claim 5, wherein: The weighted fusion of the policy priority scores based on the remaining candidate policies to generate a comprehensive transmission optimization policy includes: Obtain the execution success rate, execution efficiency score, and user satisfaction score of each remaining candidate strategy in historical transmission data; Determining a strategy effectiveness coefficient based on the product of the execution success rate and the execution efficiency score; Determining a policy priority weight based on a ratio of the user satisfaction score to a preset satisfaction threshold; The strategy effectiveness coefficient and the strategy priority weight are weighted and summed to obtain a comprehensive priority score for each candidate strategy; The top N candidate strategies with the highest comprehensive priority scores are selected, and the transmission parameter adjustment instructions in the top N candidate strategies are de-duplicated and merged to generate the comprehensive transmission optimization strategy.

8. The method for optimizing user experience of an audio and video transmission system according to claim 1, wherein: Determining a dynamic transmission optimization strategy based on the transmission quality evaluation result includes: Dividing the transmission quality evaluation result into multiple quality levels, wherein each quality level corresponds to a different optimization strategy template; Detecting a target quality level to which the transmission quality assessment result belongs, and matching an optimization policy template corresponding to the target quality level from a predefined policy library; Dynamically adjusting the parameter thresholds in the optimization strategy template according to the network fluctuation characteristics of the current transmission session sequence to generate a dynamic transmission optimization strategy adapted to the current network status; And, feeding back the dynamic transmission optimization strategy to the audio and video transmission system to trigger a real-time transmission parameter adjustment operation, including: Parsing the bit rate adjustment instructions, buffer management instructions and decoding priority instructions in the dynamic transmission optimization strategy; Adjust the encoding bit rate of the current video stream according to the bit rate adjustment instruction, and synchronously update the sensitivity parameter of the adaptive bit rate algorithm; Dynamically allocating the memory ratio of the receiving buffer and the decoding buffer according to the buffer management instruction, and setting a buffer overflow protection threshold; The scheduling order of the audio and video decoding threads is adjusted according to the decoding priority instruction, and the usage priority of the hardware acceleration resources is allocated.

9. A user experience optimization system for an audio and video transmission system, characterized in that: The user experience optimization system for the audio and video transmission system includes a processor and a memory, the memory is connected to the processor, the memory is used to store programs, instructions or codes, and the processor is used to execute the programs, instructions or codes in the memory to implement the user experience optimization method for the audio and video transmission system described in any one of claims 1 to 8.

Images