An IP video phone quality control method based on artificial intelligence perception

Abstract

The application discloses an IP video telephone quality control method based on artificial intelligence perception, and comprises the following steps: collecting call session identification and call operation data, and performing pretreatment; inputting a standardized feature sequence into an improved CTRNN network to perform multi-time scale state modeling and call quality semantic constraint recursive evolution, and outputting a quality perception state vector; constructing a quality control strategy input package based on the quality perception state vector; performing end-side capability constraint screening and encoder configuration constraint checking on control elements; distributing a video telephone quality control strategy set to a video telephone session execution module, and writing into a configurable parameter; updating network parameters based on feedback samples, and outputting an updated quality control execution result. The application adopts an improved continuous time recurrent neural network to realize dynamic perception and adaptive control of IP video telephone call quality, and has the advantages of accurate perception, stable control and continuous optimization.

Description

Technical Field

[0001] This invention relates to the field of video call quality control, and more particularly to an IP video telephony quality control method based on artificial intelligence perception. Background Technology

[0002] In existing IP video telephony systems, call quality control typically relies on static threshold rules or adaptive adjustment mechanisms based on a single network metric. These mechanisms adjust the encoding bitrate, resolution, or retransmission strategy by monitoring changes in latency, packet loss rate, or bandwidth. Some solutions introduce simple machine learning models to predict network conditions, but these are mostly trained offline and lack joint modeling of the multi-dimensional states of the network, media, and services during a call. Consequently, they fail to reflect the dynamic evolution of call quality across different time scales.

[0003] The aforementioned existing technologies generally suffer from insufficient semantic characterization of call quality and a disconnect between control strategy generation and execution feedback. This makes it difficult to continuously update the model based on actual execution results, leading to delayed or unstable policy adjustments. Furthermore, existing solutions often fail to fully integrate edge capabilities and encoder configuration constraints for unified constraint verification, which can easily result in policies becoming unexecutable or deviating from expected performance. Consequently, they struggle to achieve stable and sustainable call quality control in complex network environments.

[0004] Therefore, how to provide an IP video telephony quality control method based on artificial intelligence perception is a problem that urgently needs to be solved by those skilled in the art. Summary of the Invention

[0005] One objective of this invention is to propose an IP video telephony quality control method based on artificial intelligence perception. This invention uses an improved continuous-time recurrent neural network to achieve dynamic perception and adaptive control of IP video telephony call quality, which has the advantages of accurate perception, stable control and continuous optimization.

[0006] An IP video telephony quality control method based on artificial intelligence perception according to an embodiment of the present invention includes the following steps: After an IP video telephony call is established, call session identifiers and call operation data are collected, and the call operation data is preprocessed to obtain a standardized feature sequence. The standardized feature sequence is input into the improved CTRNN network for multi-timescale state modeling and recursive evolution of call quality semantic constraints, and the output is a quality-aware state vector representing the real-time quality state of the call. Based on the quality-aware state vector associated with the call session identifier, and by integrating the end-side capability parameters and encoder configuration constraints, a quality control strategy input package is constructed. Based on the quality control strategy input package, end-side capability constraint screening and encoder configuration constraint verification are performed on the control elements to generate a set of video telephony quality control strategies. The video call quality control strategy set is distributed to the video call session execution module, and configurable parameters are written into the module to generate the quality control execution result; Feedback samples are constructed based on the quality control execution results to update the improved CTRNN network, and the updated quality control execution results are output.

[0007] Optionally, the call operation data includes network status information, media transmission data, and service operation data, and the preprocessing includes missing value imputation, outlier removal, timestamp alignment, sliding time window segmentation, and numerical normalization.

[0008] Optionally, the output of the quality-sensing state vector specifically includes: Receive standardized feature sequences and organize them into time-series input sequences corresponding to sliding time windows according to timestamp order; The temporal input sequence is input into the feature mapping layer of the improved CTRNN network, and the normalized feature values ​​at each time step are dimensionally aligned and numerically mapped to obtain a temporal input activation sequence that is consistent with the input dimension of the state unit. The improvements of the improved CTRNN network compared to the original network include: introducing a set of time constant parameters at the state modeling level and forming a state unit grouping structure based on the time constant interval, and introducing a recursive connection relationship configuration mechanism with call quality semantic constraints in the recursive connection structure. Construct a set of state units and configure a set of time constant parameters for the set of state units. Divide the set of state units into multiple state unit groups according to the time constant interval. The state unit set is initialized, an initial mass state vector is generated based on the state initialization module, and the initial mass state vector is written into the state buffer as the recursive input at the first moment. For the current time step of the temporal input activation sequence, input activation vectors are used to calculate the input driving force and the intermediate results of the recursive input. Based on the recursive connection relationship defined by the semantics of call quality, connection constraint processing is performed on the recursive connection weights between state units, and the intermediate results of the recursive input are recursively calculated based on the constrained recursive connection weights to obtain the recursive driving quantity. The input driving quantity and the recursive driving quantity are merged to obtain the state update driving quantity, and the state update driving quantity is input into the state nonlinear transformation module for nonlinear mapping to obtain the state candidate change quantity. Based on the continuous-time evolution update module corresponding to the state update structure, the mass state vector is updated to obtain the mass state vector at the current moment. The current quality state vector is input into the network quality readout module, media quality readout module, and service quality readout module respectively for output mapping, resulting in network quality sub-vectors, media quality sub-vectors, and service quality sub-vectors. These are then concatenated and encapsulated to output a quality perception state vector representing the real-time quality state of the call.

[0009] Optionally, the construction of the quality control strategy input package specifically includes: The quality-aware state vector is associated and bound with the call session identifier to obtain a quality state record with a session identifier; Obtain the set of end-side capability parameters and encapsulate them in a standardized manner. The set of end-side capability parameters includes end-side processing capability parameters, end-side coding capability parameters, and end-side network access capability parameters. Obtain the encoder configuration constraint set and organize it in a structured manner. The encoder configuration constraint set includes bit rate configuration constraints, resolution configuration constraints, frame rate configuration constraints, and keyframe interval configuration constraints. The quality-aware state vector is parsed to obtain the state field set corresponding to the network quality sub-vector, media quality sub-vector, and service quality sub-vector. Field alignment and field merging are then performed on the state field set to obtain the quality state field package. The quality status field package, the quality status record with session identifier, the set of edge capability parameters, and the set of encoder configuration constraints are associated and encapsulated to obtain the quality control strategy input package. The association and encapsulation includes field alignment, field association binding, field order arrangement, field identifier mapping, and input package structured encapsulation processing.

[0010] Optionally, the generation of the video call quality control strategy set specifically includes: Obtain the quality control strategy input package, and read the quality perception state vector field area, end-side capability parameter field area, and encoder configuration constraint field area from the quality control strategy input package; Based on the quality-aware state vector field area, the state fields of network quality sub-vector, media quality sub-vector and service quality sub-vector are extracted, and a set of control elements is generated according to the mapping relationship between the state fields and the control dimensions. Based on the end-side capability parameter field area, the control element set is subjected to capability constraint filtering processing, and the control elements that do not meet the end-side capability parameters are masked to obtain the control element set after capability constraints. Based on the encoder configuration constraint field area, the set of control elements after capability constraints is subjected to configuration constraint verification processing, and the control elements that do not meet the encoder configuration constraint set are subjected to constraint adjustment processing to obtain the set of control elements after constraint verification. Based on the set of control elements after constraint verification, the execution strategy mapping is performed to generate encoding control strategy, transmission control strategy and media transmission control strategy respectively. The encoding control strategy, transmission control strategy, and media transmission control strategy are combined and encapsulated to generate a set of video telephony quality control strategies.

[0011] Optionally, the generation of the quality control execution result specifically includes: The video call quality control strategy set is associated and bound with the call session identifier, and then encapsulated according to the field identifier and field order to generate an instruction package; The instruction packet is sent to the video call session execution module, and the video call session execution module performs parameter writing processing corresponding to the encoding control strategy on the video encoding parameter set; In the video call session execution module, the parameter writing process corresponding to the transmission control strategy is performed on the transmission parameter set; In the video call session execution module, the parameters corresponding to the media transmission control strategy are written to the media transmission behavior parameter set. The policy effectiveness confirmation process is performed in the video call session execution module. The current parameter value after the parameter is written is read from the video call session execution module and the field is encapsulated to generate the quality control execution result. The policy effectiveness confirmation process includes policy parameter writing status confirmation process, policy parameter consistency verification process, policy execution module response status confirmation process, and policy effectiveness timestamp recording process.

[0012] Optionally, the output of the updated quality control execution result specifically includes: While the parameters corresponding to the execution results of the quality control are in effect, continuously collect the adjusted network status information, media transmission data and service operation data, and organize them into the adjusted original feature sequence according to the timestamp order; Preprocessing is performed on the adjusted original feature sequence to obtain the adjusted standardized feature sequence; The quality control execution results and the adjusted standardized feature sequences are associated and encapsulated according to the call session identifier and timestamp alignment relationship to obtain feedback samples; The feedback samples are input into the update module of the improved CTRNN network to perform parameter update processing, resulting in the updated CTRNN network; The adjusted and standardized feature sequence is input into the updated CTRNN network to output the updated quality-aware state vector, and the updated quality control execution result is output based on the updated quality-aware state vector.

[0013] The beneficial effects of this invention are: This invention introduces an improved continuous-time recurrent neural network to uniformly model the network state, media transmission state, and service operation state during IP video telephony calls, achieving dynamic perception and continuous evolution of call quality across multiple time scales. Compared to existing technologies that rely solely on a single indicator or static rules, this invention recursively models the influence relationships between different states at the overall call quality semantic level. This allows the generated quality-perceived state vector to more realistically reflect the real-time call quality change trend, providing a more stable and accurate decision-making basis for the subsequent generation of quality control strategies, thereby effectively improving the targeting and reliability of call quality control.

[0014] At the quality control strategy generation and execution level, this invention achieves executability screening and constraint consistency control of control elements by uniformly encapsulating and jointly verifying the quality-aware state vector, call session identifier, end-side capability parameters, and encoder configuration constraints. The resulting video telephony quality control strategy set, after being sent to the video telephony session execution module, can be directly written with the corresponding configurable parameters, and a clear quality control execution result is formed through a strategy effectiveness confirmation mechanism. This avoids the problems of unexecutable strategies, opaque execution status, or untraceable execution effects in existing technologies, improving the determinism of the quality control process and the stability of system operation.

[0015] Furthermore, this invention constructs a feedback learning mechanism based on the execution results of quality control. By continuously collecting call operation data after the strategy takes effect and constructing feedback samples, the parameters of the improved continuous-time recurrent neural network are updated, enabling the model to continuously and adaptively adjust according to the actual control effect. This closed-loop update mechanism allows the quality control strategy to be continuously optimized as the network environment and call scenarios change, effectively reducing the risk of control lag and strategy oscillation, and improving the long-term stability and service quality assurance capability of the IP video telephony system under complex dynamic network conditions. Attached Figure Description

[0016] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings: Figure 1 This is a flowchart of an IP video telephony quality control method based on artificial intelligence perception proposed in this invention; Figure 2 This is a flowchart illustrating the quality perception state vector generation process of an IP video telephony quality control method based on artificial intelligence perception proposed in this invention. Figure 3 This is a flowchart illustrating the quality control strategy execution of an IP video telephony quality control method based on artificial intelligence perception proposed in this invention. Detailed Implementation

[0017] The present invention will now be described in further detail with reference to the accompanying drawings. These drawings are simplified schematic diagrams, illustrating only the basic structure of the invention, and therefore only show the components relevant to the invention.

[0018] refer to Figures 1-3 A quality control method for IP video telephony based on artificial intelligence perception includes the following steps: After an IP video telephony call is established, call session identifiers and call operation data are collected, and the call operation data is preprocessed to obtain a standardized feature sequence. The standardized feature sequence is input into the improved CTRNN network for multi-timescale state modeling and recursive evolution of call quality semantic constraints, and the output is a quality-aware state vector representing the real-time quality state of the call. Based on the quality-aware state vector associated with the call session identifier, and by integrating the end-side capability parameters and encoder configuration constraints, a quality control strategy input package is constructed. Based on the quality control strategy input package, end-side capability constraint screening and encoder configuration constraint verification are performed on the control elements to generate a set of video telephony quality control strategies. The video call quality control strategy set is distributed to the video call session execution module, and configurable parameters are written into the module to generate the quality control execution result; Feedback samples are constructed based on the quality control execution results to update the improved CTRNN network, and the updated quality control execution results are output.

[0019] In this embodiment, the call operation data includes network status information, media transmission data, and service operation data. The network status information includes one-way latency, round-trip latency, jitter, packet loss rate, available bandwidth, and reordering rate. The media transmission data includes receive frame interval, decoding frame rate, keyframe interval, decoding frame loss rate, and playback buffer duration. The service operation data includes session duration, audio and video synchronization deviation, terminal-side CPU utilization, and terminal-side encoding queue length. The preprocessing includes missing value imputation, outlier removal, timestamp alignment, sliding time window segmentation, and numerical normalization.

[0020] In this embodiment, the output of the quality-sensing state vector specifically includes: The system receives a standardized feature sequence and organizes it into a time-series input sequence corresponding to a sliding time window according to the timestamp order. Each moment in the time-series input sequence includes normalized feature values ​​of network status information, media transmission data, and service operation data. The temporal input sequence is input into the feature mapping layer of the improved CTRNN network, and the normalized feature values ​​at each time step are dimensionally aligned and numerically mapped to obtain a temporal input activation sequence that is consistent with the input dimension of the state unit. The improvements of the improved CTRNN network compared to the original network include: introducing a set of time constant parameters at the state modeling level and forming a state unit grouping structure based on the time constant interval, and introducing a recursive connection relationship configuration mechanism with call quality semantic constraints in the recursive connection structure. A set of state units is constructed, and a set of time constant parameters is configured for the set of state units. The set of state units is divided into multiple state unit groups according to the time constant interval. The time constant interval is obtained by segmenting the time constant value range. The time constant value range is determined according to the statistics of historical call operation data and divided into multiple non-overlapping value intervals according to the segmentation rules. Each value interval corresponds to a set of state units. The state unit set is initialized, an initial mass state vector is generated based on the state initialization module, and the initial mass state vector is written into the state buffer as the recursive input at the first moment. The generation of the initial quality state vector includes: the state initialization module receiving the quality state initialization input, which is composed of a preset initialization parameter set and the standardized feature value corresponding to the call start time; based on the quality state initialization input, generating initial state values ​​corresponding to each state unit in the structural order of the state unit set, and performing numerical mapping and range constraint processing on the initial state values ​​to obtain the initial quality state vector. For the current time step input activation vector of the temporal input activation sequence, the input driving quantity and the recursive input intermediate result are calculated respectively. The input driving quantity is calculated by weighting the feature values ​​of each dimension of the current input with the input connection weight parameters used to describe the connection relationship between the input and the state unit, and the weighted calculation results are summarized. The recursive input intermediate result is obtained by reading the quality state vector corresponding to the previous time step from the state buffer and expanding the quality state vector according to the arrangement order of the state unit set. Based on the recursive connection relationship defined by the call quality semantics, connection constraint processing is performed on the recursive connection weights between state units, and the intermediate results of the recursive input are recursively calculated based on the constrained recursive connection weights to obtain the recursive driving quantity. The call quality semantics include the feedback relationship from network quality state to media quality state, the feedback relationship from network quality state to service quality state, and the feedback relationship from media quality state to service quality state. The connection constraint processing performs filtering, retention, and masking operations on the recursive connection weights between state unit groups according to the call quality semantics. The recursive driving force is obtained by: reading the quality state vector corresponding to the previous time step from the state cache and expanding the quality state vector according to the arrangement order of the state unit set; selecting the recursive connection weights corresponding to each state unit in the quality state vector based on the recursive connection weights between the state units retained after connection constraint processing; performing item-by-item weighted calculation on the state value of each state unit in the quality state vector and its corresponding recursive connection weight, and summarizing the weighted calculation results according to the state unit group to obtain the recursive contribution value corresponding to each state unit group; and combining the recursive contribution values ​​corresponding to each state unit group to obtain the recursive driving force used for the state update calculation at the current time step. The input driving quantity and the recursive driving quantity are merged to obtain the state update driving quantity, and the state update driving quantity is input into the state nonlinear transformation module for nonlinear mapping to obtain the state candidate change quantity. Based on the continuous-time evolution update module corresponding to the state update structure, the quality state vector is updated to obtain the quality state vector at the current moment. The continuous-time evolution update module recursively updates the quality state vector at the current recursive moment by weighting the quality state vector at the previous recursive moment and the change in the state candidate. The weight of the weighted combination is determined by the time constant parameter corresponding to the state unit group and the time interval between adjacent sampling moments. The current quality state vector is input into the network quality readout module, media quality readout module, and service quality readout module respectively for output mapping, resulting in network quality sub-vectors, media quality sub-vectors, and service quality sub-vectors. These are then concatenated and encapsulated to output a quality perception state vector representing the real-time quality state of the call.

[0021] In this embodiment, the construction of the quality control strategy input package specifically includes: The quality-aware state vector is associated and bound with the call session identifier to obtain a quality state record with a session identifier; A set of edge-side capability parameters is obtained and standardized and encapsulated. The set of edge-side capability parameters includes edge-side processing capability parameters, edge-side encoding capability parameters, and edge-side network access capability parameters. The edge-side processing capability parameters include edge-side processor computing power parameters, edge-side processor load parameters, edge-side available memory capacity parameters, and edge-side parallel processing capability parameters. The edge-side encoding capability parameters include edge-side supported encoding format parameters, edge-side supported encoding resolution range parameters, edge-side supported encoding frame rate range parameters, edge-side supported bitrate adjustment range parameters, and edge-side keyframe generation capability parameters. The edge-side network access capability parameters include edge-side currently available bandwidth parameters, edge-side network latency parameters, edge-side packet loss rate parameters, edge-side network jitter parameters, and edge-side supported network access type parameters. The standardized encapsulation includes parameter format unification, value range normalization, missing item filling, outlier filtering, and field structured mapping processing. The encoder configuration constraint set is obtained and structured. The encoder configuration constraint set includes bitrate configuration constraints, resolution configuration constraints, frame rate configuration constraints, and keyframe interval configuration constraints. The bitrate configuration constraints include target bitrate upper limit parameters, target bitrate lower limit parameters, and bitrate adjustment step size parameters. The resolution configuration constraints include optional resolution set parameters, resolution switching restriction parameters, and resolution-bitrate correlation constraint parameters. The frame rate configuration constraints include target frame rate upper limit parameters, target frame rate lower limit parameters, and frame rate adjustment granularity parameters. The keyframe interval configuration constraints include keyframe interval upper limit parameters, keyframe interval lower limit parameters, and keyframe trigger condition restriction parameters. The structured organization includes the extraction, classification, hierarchical mapping, consistency verification, and encapsulation processing of the configuration constraint fields. The quality-aware state vector is parsed to obtain the state field set corresponding to the network quality sub-vector, media quality sub-vector, and service quality sub-vector. Field alignment and field merging are then performed on the state field set to obtain the quality state field package. The quality status field package, the quality status record with session identifier, the end-side capability parameter set, and the encoder configuration constraint set are associated and encapsulated to obtain the quality control policy input package. The association and encapsulation includes field alignment, field association binding, field order arrangement, field identifier mapping, and input package structured encapsulation processing. The quality control policy input package includes a quality-aware status vector field area, a call session identifier field area, an end-side capability parameter field area, and an encoder configuration constraint field area.

[0022] In this embodiment, the generation of the video call quality control strategy set specifically includes: Obtain the quality control strategy input package, and read the quality perception state vector field area, end-side capability parameter field area, and encoder configuration constraint field area from the quality control strategy input package; Based on the quality-aware state vector field region, the state fields of network quality sub-vector, media quality sub-vector and service quality sub-vector are extracted, and a set of control elements is generated according to the mapping relationship from the state fields to the control dimensions. The set of control elements includes bit rate control elements, resolution control elements, frame rate control elements, keyframe interval control elements, redundancy ratio control elements, retransmission control elements, congestion control elements, frame loss control elements, keyframe request control elements and transmission rhythm control elements. Based on the end-side capability parameter field area, the control element set is subjected to capability constraint filtering processing, and the control elements that do not meet the end-side capability parameters are masked to obtain the control element set after capability constraints. Based on the encoder configuration constraint field area, the set of control elements after capability constraints is subjected to configuration constraint verification processing, and the control elements that do not meet the encoder configuration constraint set are subjected to constraint adjustment processing to obtain the set of control elements after constraint verification. The constraint adjustment processing includes the trimming, replacement, rearrangement and conflict resolution processing of control element values. Based on the set of control elements after constraint verification, the execution strategy mapping generates encoding control strategy, transmission control strategy and media transmission control strategy respectively. The encoding control strategy includes target bit rate, target resolution, target frame rate and target keyframe group length. The transmission control strategy includes forward error correction redundancy ratio parameter, retransmission parameter and congestion control parameter. The media transmission control strategy includes frame dropping strategy, keyframe request strategy and transmission rhythm control parameter. The encoding control strategy, transmission control strategy, and media transmission control strategy are combined and encapsulated to generate a set of video telephony quality control strategies.

[0023] In this embodiment, the generation of the quality control execution result specifically includes: The video call quality control strategy set is associated and bound with the call session identifier, and then encapsulated according to the field identifier and field order to generate an instruction package. The instruction package includes a call session identifier field, an encoding control strategy field, a transmission control strategy field, and a media transmission control strategy field. The instruction packet is sent to the video call session execution module, and the video call session execution module performs parameter writing processing corresponding to the encoding control strategy on the video encoding parameter set. The video encoding parameter set is the encoder configurable parameter set used by the video call session execution module to receive and execute the parameter writing corresponding to the encoding control strategy. In the video call session execution module, parameter writing processing corresponding to the transmission control strategy is performed on the transmission parameter set. The transmission parameter set is a set of configurable transmission control parameters used by the video call session execution module to receive and execute parameter writing corresponding to the transmission control strategy. In the video call session execution module, the parameter writing process corresponding to the media sending control strategy is performed on the media sending behavior parameter set. The media sending behavior parameter set is a set of configurable parameters for media sending behavior used in the video call session execution module to receive and execute the parameter writing corresponding to the media sending control strategy. The policy effectiveness confirmation process is performed in the video call session execution module. The current parameter values ​​after parameter writing are read from the video call session execution module and encapsulated into fields to generate quality control execution results. The policy effectiveness confirmation process includes policy parameter writing status confirmation process, policy parameter consistency verification process, policy execution module response status confirmation process, and policy effectiveness timestamp recording process. The quality control execution results include the encoding parameter set after parameter writing takes effect, the transmission parameter set after parameter writing takes effect, and the media transmission behavior record after parameter writing takes effect.

[0024] In this embodiment, the output of the updated quality control execution result specifically includes: While the parameters corresponding to the execution results of the quality control are in effect, continuously collect the adjusted network status information, media transmission data and service operation data, and organize them into the adjusted original feature sequence according to the timestamp order; Preprocessing is performed on the adjusted original feature sequence to obtain the adjusted standardized feature sequence; The quality control execution results and the adjusted standardized feature sequences are associated and encapsulated according to the call session identifier and timestamp alignment relationship to obtain feedback samples; The feedback samples are input into the update module of the improved CTRNN network to perform parameter update processing, resulting in the updated CTRNN network; The parameter update process includes: using the adjusted standardized feature sequence and the corresponding quality control execution result in the feedback sample as update input, constructing a time-series sample sequence for network update according to the timestamp order; calculating the state output of the improved CTRNN network at each time step based on the time-series sample sequence, aligning the state output with the quality control execution result in the feedback sample, and generating state deviation information; performing gradient calculation processing on the input connection weight parameters, recursive connection weight parameters, and time constant parameters corresponding to the state units in the improved CTRNN network according to the state deviation information, performing iterative correction processing on the above parameters according to a preset update step size, and ending the parameter correction processing after reaching a preset number of updates, writing the updated parameters into the improved CTRNN network to obtain the updated CTRNN network; The adjusted and standardized feature sequence is input into the updated CTRNN network to output the updated quality-aware state vector, and the updated quality control execution result is output based on the updated quality-aware state vector.

[0025] Example 1: To verify the feasibility of this invention in practice, it was applied to a mobile IP video call scenario in a public network environment of a city. In this scenario, the two parties are in an indoor wireless network environment and an outdoor cellular network environment, respectively. During the call, the network status fluctuates continuously with the user's movement and changes in network load, which can easily lead to problems such as video stuttering, sudden drop in clarity, and audio-visual desynchronization. Traditional quality control methods based on fixed thresholds or static rules are difficult to respond to these changes in a timely and accurate manner, resulting in an unstable call experience and an inability to balance video quality and transmission continuity.

[0026] In this scenario, after a video call is established, the terminal continuously collects call session identifiers and corresponding call operation data, including network status information, media transmission data, and service operation data. The collected data undergoes unified preprocessing to form a standardized feature sequence. This standardized feature sequence is input into an improved CTRNN network. By introducing a multi-timescale state modeling mechanism, the network can simultaneously characterize the instantaneous fluctuations of the network and the medium- to long-term evolution trend of call quality. Furthermore, semantic constraints on call quality are introduced during the recursive evolution process, ensuring that the network state update process conforms to the actual correlation between network quality, media quality, and service quality. This results in the output of a quality-aware state vector that reflects the current call quality in real time. Subsequently, the system associates this quality-aware state vector with the call session identifier and, combined with the terminal's current processing power, encoding capabilities, and network access capabilities, while also introducing encoder configuration constraints, constructs a quality control policy input package for subsequent policy decisions.

[0027] During the strategy generation phase, the system maps control elements based on the network quality, media quality, and service quality state elements parsed from the quality-aware state vector. It then filters and verifies these elements in conjunction with terminal capability parameters and encoder configuration constraints to avoid generating control strategies that exceed terminal capabilities or violate coding constraints. Based on this, the system forms a set of video telephony quality control strategies, including encoding control strategies, transmission control strategies, and media transmission control strategies, and distributes this strategy set to the video telephony session execution module. During the call, the execution module dynamically writes the corresponding encoding parameters, transmission parameters, and media transmission behavior parameters, confirms the strategy's effectiveness, and obtains the quality control execution result. This allows video encoding, data transmission, and media transmission behavior to adaptively adjust to changes in call quality.

[0028] During the call, the system continuously collects adjusted network status information, media transmission data, and service operation data based on the effective quality control execution results. This data is then correlated and encapsulated with the quality control execution results to form feedback samples. These feedback samples are used to update the improved CTRNN network online, enabling network parameters to be continuously corrected according to changes in the actual call environment. This improves the accuracy of perceiving subsequent call quality status and the adaptability of strategy generation. Through this application method, in actual call scenarios, this invention can effectively alleviate the quality fluctuations caused by network volatility, making video image changes smoother, significantly improving audio-visual synchronization and call stability. Simultaneously, it avoids the impact of frequent and large-scale parameter adjustments on terminal and network resources, demonstrating the practical application value of this invention for video telephony quality control in complex and dynamic network environments.

[0029] Table 1. Performance Comparison of the Invention and Traditional IP Video Telephony Quality Control Methods

[0030] As can be clearly seen from Table 1, the method of the present invention is superior to the traditional method in many indicators.

[0031] The end-to-end latency metrics show that the average latency under the traditional rule-based control method is 182ms, while the method of this invention reduces this value to 165ms. Achieving a 17ms latency reduction without significantly increasing the bitrate burden demonstrates that this invention can anticipate network changes during the quality state modeling stage, allowing encoding and transmission parameter adjustments to better align with network evolution, thereby reducing the cumulative latency caused by buffering and retransmissions.

[0032] Regarding latency jitter, the traditional method has a jitter value of 38ms, while the method of this invention reduces it to 29ms. This reduction reflects that after introducing multi-timescale state modeling, this invention can distinguish between short-term network fluctuations and medium-term quality changes, avoiding frequent and drastic parameter oscillations and making latency changes more stable.

[0033] Regarding packet loss rate, the traditional method results in 1.62%, while the method of this invention reduces it to 1.18%. This improvement does not rely on simply increasing redundancy, but rather on a fine characterization of network quality sub-states through a quality-aware state vector, enabling the transmission control strategy to adjust before the risk of packet loss increases, thereby reducing the probability of actual packet loss.

[0034] Regarding the average video bitrate, the method of this invention achieves an average bitrate of 1420kbps, slightly lower than the 1480kbps of the traditional method. With a slightly lower bitrate, the number of stutters decreased from 2.4 times / min to 1.6 times / min, indicating that this invention does not improve the user experience by simply increasing the bitrate, but rather improves bitrate utilization efficiency through more reasonable bitrate allocation and transmission rhythm control.

[0035] The sharpness fluctuation range decreased from 3.1dB to 2.4dB, indicating a smoother change in image quality. This effect stems from the introduction of quality semantic constraints during the strategy generation stage, ensuring that adjustments to resolution, bitrate, and frame rate follow a continuous principle, thus avoiding abrupt changes in sharpness.

[0036] The number of audio-visual asynchrony events decreased from 1.3 times / 10min to 0.8 times / 10min, indicating that after the service quality status was incorporated into the unified state modeling and strategy decision-making process, the audio-visual collaborative adjustment capability was significantly enhanced, effectively reducing the asynchrony problem between the decoding and playback sides.

[0037] The terminal CPU utilization rate decreased from 27.5% to 24.9%, while the parameter switching frequency decreased from 3.6 times / min to 2.2 times / min. This indicates that the present invention reduces the computation and scheduling overhead caused by terminal-side parameter reconfiguration by reducing unnecessary policy switching, thereby improving the operational stability at the system level.

[0038] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A method for quality control of IP video telephony based on artificial intelligence perception, characterized in that, Includes the following steps: After an IP video telephony call is established, call session identifiers and call operation data are collected, and the call operation data is preprocessed to obtain a standardized feature sequence. The standardized feature sequence is input into the improved CTRNN network for multi-timescale state modeling and recursive evolution of call quality semantic constraints, and the output is a quality-aware state vector representing the real-time quality state of the call. Based on the quality-aware state vector associated with the call session identifier, and by integrating the end-side capability parameters and encoder configuration constraints, a quality control strategy input package is constructed. Based on the quality control strategy input package, end-side capability constraint screening and encoder configuration constraint verification are performed on the control elements to generate a set of video telephony quality control strategies. The video call quality control strategy set is distributed to the video call session execution module, and configurable parameters are written into the module to generate the quality control execution result; Feedback samples are constructed based on the quality control execution results to update the improved CTRNN network, and the updated quality control execution results are output.

2. The method for quality control of IP video telephony based on artificial intelligence perception according to claim 1, characterized in that, The call operation data includes network status information, media transmission data, and service operation data. The preprocessing includes missing value imputation, outlier removal, timestamp alignment, sliding time window segmentation, and numerical normalization.

3. The method for quality control of IP video telephony based on artificial intelligence perception according to claim 1, characterized in that, The output of the quality-sensing state vector specifically includes: Receive standardized feature sequences and organize them into time-series input sequences corresponding to sliding time windows according to timestamp order; The temporal input sequence is input into the feature mapping layer of the improved CTRNN network, and the normalized feature values ​​at each time step are dimensionally aligned and numerically mapped to obtain a temporal input activation sequence that is consistent with the input dimension of the state unit. The improvements of the improved CTRNN network compared to the original network include: introducing a set of time constant parameters at the state modeling level and forming a state unit grouping structure based on the time constant interval, and introducing a recursive connection relationship configuration mechanism with call quality semantic constraints in the recursive connection structure. Construct a set of state units and configure a set of time constant parameters for the set of state units. Divide the set of state units into multiple state unit groups according to the time constant interval. The state unit set is initialized, an initial mass state vector is generated based on the state initialization module, and the initial mass state vector is written into the state buffer as the recursive input at the first moment. For the current time step of the temporal input activation sequence, input activation vectors are used to calculate the input driving force and the intermediate results of the recursive input. Based on the recursive connection relationship defined by the semantics of call quality, connection constraint processing is performed on the recursive connection weights between state units, and the intermediate results of the recursive input are recursively calculated based on the constrained recursive connection weights to obtain the recursive driving quantity. The input driving quantity and the recursive driving quantity are merged to obtain the state update driving quantity, and the state update driving quantity is input into the state nonlinear transformation module for nonlinear mapping to obtain the state candidate change quantity. Based on the continuous-time evolution update module corresponding to the state update structure, the mass state vector is updated to obtain the mass state vector at the current moment. The current quality state vector is input into the network quality readout module, media quality readout module, and service quality readout module respectively for output mapping, resulting in network quality sub-vectors, media quality sub-vectors, and service quality sub-vectors. These are then concatenated and encapsulated to output a quality perception state vector representing the real-time quality state of the call.

4. The method for quality control of IP video telephony based on artificial intelligence perception according to claim 1, characterized in that, The construction of the quality control strategy input package specifically includes: The quality-aware state vector is associated and bound with the call session identifier to obtain a quality state record with a session identifier; Obtain the set of end-side capability parameters and encapsulate them in a standardized manner. The set of end-side capability parameters includes end-side processing capability parameters, end-side coding capability parameters, and end-side network access capability parameters. Obtain the encoder configuration constraint set and organize it in a structured manner. The encoder configuration constraint set includes bit rate configuration constraints, resolution configuration constraints, frame rate configuration constraints, and keyframe interval configuration constraints. The quality-aware state vector is parsed to obtain the state field set corresponding to the network quality sub-vector, media quality sub-vector, and service quality sub-vector. Field alignment and field merging are then performed on the state field set to obtain the quality state field package. The quality status field package, the quality status record with session identifier, the set of edge capability parameters, and the set of encoder configuration constraints are associated and encapsulated to obtain the quality control strategy input package. The association and encapsulation includes field alignment, field association binding, field order arrangement, field identifier mapping, and input package structured encapsulation processing.

5. The method for quality control of IP video telephony based on artificial intelligence perception according to claim 1, characterized in that, The generation of the video call quality control strategy set specifically includes: Obtain the quality control strategy input package, and read the quality perception state vector field area, end-side capability parameter field area, and encoder configuration constraint field area from the quality control strategy input package; Based on the quality-aware state vector field area, the state fields of network quality sub-vector, media quality sub-vector and service quality sub-vector are extracted, and a set of control elements is generated according to the mapping relationship between the state fields and the control dimensions. Based on the end-side capability parameter field area, the control element set is subjected to capability constraint filtering processing, and the control elements that do not meet the end-side capability parameters are masked to obtain the control element set after capability constraints. Based on the encoder configuration constraint field area, the set of control elements after capability constraints is subjected to configuration constraint verification processing, and the control elements that do not meet the encoder configuration constraint set are subjected to constraint adjustment processing to obtain the set of control elements after constraint verification. Based on the set of control elements after constraint verification, the execution strategy mapping is performed to generate encoding control strategy, transmission control strategy and media transmission control strategy respectively. The encoding control strategy, transmission control strategy, and media transmission control strategy are combined and encapsulated to generate a set of video telephony quality control strategies.

6. The method for quality control of IP video telephony based on artificial intelligence perception according to claim 1, characterized in that, The generation of the quality control execution results specifically includes: The video call quality control strategy set is associated and bound with the call session identifier, and then encapsulated according to the field identifier and field order to generate an instruction package; The instruction packet is sent to the video call session execution module, and the video call session execution module performs parameter writing processing corresponding to the encoding control strategy on the video encoding parameter set; In the video call session execution module, the parameter writing process corresponding to the transmission control strategy is performed on the transmission parameter set; In the video call session execution module, the parameters corresponding to the media transmission control strategy are written to the media transmission behavior parameter set. The policy effectiveness confirmation process is performed in the video call session execution module. The current parameter value after the parameter is written is read from the video call session execution module and the field is encapsulated to generate the quality control execution result. The policy effectiveness confirmation process includes policy parameter writing status confirmation process, policy parameter consistency verification process, policy execution module response status confirmation process, and policy effectiveness timestamp recording process.

7. The method for quality control of IP video telephony based on artificial intelligence perception according to claim 1, characterized in that, The output of the updated quality control execution result specifically includes: While the parameters corresponding to the execution results of the quality control are in effect, continuously collect the adjusted network status information, media transmission data and service operation data, and organize them into the adjusted original feature sequence according to the timestamp order; Preprocessing is performed on the adjusted original feature sequence to obtain the adjusted standardized feature sequence; The quality control execution results and the adjusted standardized feature sequences are associated and encapsulated according to the call session identifier and timestamp alignment relationship to obtain feedback samples; The feedback samples are input into the update module of the improved CTRNN network to perform parameter update processing, resulting in the updated CTRNN network; The adjusted and standardized feature sequence is input into the updated CTRNN network to output the updated quality-aware state vector, and the updated quality control execution result is output based on the updated quality-aware state vector.

Images