A method for monitoring the quality of video conference communication based on intelligent sensors

Abstract

This invention discloses a video conferencing communication quality monitoring method based on intelligent sensors, belonging to the field of intelligent sensor technology. The method includes: collecting multi-source communication status data and preprocessing it to generate a communication quality data matrix; performing propagation mutation separation and link structure reconstruction to generate a stable propagation set of communication links; employing a local mean decomposition algorithm to obtain a local mean decomposition feature set; performing existence stability evolution to form a communication quality existence structure system; constructing an improved InceptionTime model to obtain a communication quality representation vector; performing irreversible decision generation to form communication quality levels and anomaly types; performing early warning event generation and grading, and outputting communication quality early warning information. This invention, by combining the local mean decomposition algorithm and the improved InceptionTime model, achieves real-time monitoring of video conferencing communication quality, anomaly type identification, fault link location, and tiered early warning.

Description

Technical Field

[0001] This invention relates to the field of intelligent sensor technology, and in particular to a method for monitoring the quality of video conferencing communication based on intelligent sensors. Background Technology

[0002] With the popularization of remote work, smart meetings, paperless meetings, and multimedia collaboration, video conferencing systems have gradually evolved from single audio and video transmission platforms into comprehensive communication systems integrating meeting terminals, audio equipment, video equipment, network interfaces, streaming media servers, meeting servers, and display output devices. In actual meetings, audio acquisition, video acquisition, network transmission, streaming media access, streaming media preview, projection display, screen sharing, and server processing all jointly determine the quality of meeting communication. An abnormality in any of these stages can affect the continuity of the meeting video, the clarity of the sound, and the stability of multi-terminal collaboration. When network transmission experiences delays, packet loss, jitter, or when there are abnormal terminal operating states, increased server load, streaming media link congestion, or abnormal display output response, it can easily cause problems such as video stuttering, video interruption, audio-visual desynchronization, projection failure, or abnormal screen sharing, affecting the effectiveness of speaking, presentations, voting, sharing, and remote collaboration during the meeting.

[0003] Existing methods for monitoring video conferencing communication quality typically focus on detecting single network parameters, single terminal status, or single audio / video indicators. They judge meeting quality solely based on network bandwidth, packet loss rate, terminal online status, or audio / video playback status, lacking unified collection, synchronous processing, and link-level correlation analysis of multi-source communication status data. This makes it difficult to reflect the status transmission relationships between the acquisition, transmission, processing, and presentation ends in a timely manner. Existing methods rely heavily on fixed rules, manual inspection, or single model outputs for anomaly identification, making it difficult to comprehensively judge the propagation of mutations in the communication link, inconsistencies in link status, and direct and indirect link relationships. It is also difficult to map anomaly types to specific link nodes, resulting in delayed detection of communication quality anomalies, inaccurate fault location, incomplete early warning information, and difficulty in distinguishing between network transmission anomalies, audio / video quality anomalies, terminal operation anomalies, and service link anomalies.

[0004] Therefore, how to provide a video conferencing communication quality monitoring method based on intelligent sensors 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 a video conferencing communication quality monitoring method based on intelligent sensors. This invention comprehensively utilizes multi-source communication state perception, cross-source state synchronization, structural propagation feature generation, local mean decomposition algorithms, and an improved InceptionTime model data analysis method. It details the entire process of collecting multi-source communication state data from video conferencing terminals, audio devices, video devices, network interfaces, streaming media servers, conferencing servers, and display output devices. This data undergoes preprocessing, link propagation analysis, local mean decomposition, communication quality structure generation, communication state characterization extraction, irreversible decision generation, and tiered early warning output. Structurally, it innovatively introduces structural contradiction-driven calculation, tensor relation closure calculus, structural information compression-release dual-state structure, path self-avoidance topology rearrangement, and unit phase break reconstruction processing methods, achieving real-time identification and early warning of video conferencing network transmission anomalies, audio / video quality anomalies, terminal operation anomalies, and service link anomalies. Compared with existing technologies, this invention has advantages such as strong multi-source perception capability, accurate link anomaly location, stable communication quality assessment, timely early warning output, and ease of intelligent operation and maintenance for video conferencing.

[0006] A video conferencing communication quality monitoring method based on smart sensors according to an embodiment of the present invention includes: Multi-source communication status data is collected by intelligent sensors, and the multi-source communication status data is preprocessed to generate a communication quality data matrix. Perform cross-source state synchronization and structural propagation feature generation on the communication quality data matrix, perform propagation mutation separation and link structure reconstruction, and generate a stable propagation set of communication links; The local mean decomposition algorithm is used to calculate the components of the stable propagation feature set of the communication link. The contradiction evolution is processed based on the structural contradiction-driven calculation, and the tensor relation closure calculation is introduced to perform closure completion processing to obtain the local mean decomposition feature set. Based on the local mean decomposition feature set, existential stability evolution is performed, and mutual exclusion existence relation generation and existence consistency generation processes are executed to form a communication quality existence structure system. An improved InceptionTime model is constructed to extract communication state representations from the multi-channel communication quality feature system. Based on the structural information compression-release dual-state structure, compressed state generation and release state restoration are performed. Path self-avoidance topology rearrangement is used to rearrange the topology order. Unit phase break reconstruction is introduced to determine the continuity of the phase and reorganize the broken units, thus obtaining the communication quality representation vector. Based on the communication quality characterization vector, irreversible decision generation is performed, and decision stability convergence calculation and conflict resolution calculation are carried out to form communication quality level and anomaly type; The system performs multi-level trigger event generation and processing for communication quality levels and anomaly types, conducts linked backtracking calculations, generates and classifies early warning events, and outputs communication quality early warning information.

[0007] Optionally, the smart sensor includes a conference terminal, audio equipment, video equipment, network interface, streaming media server, conference server, and display output device.

[0008] Optionally, the multi-source communication status data includes terminal operation data, audio acquisition data, video acquisition data, network transmission data, streaming media link data, server operation data, and display output data.

[0009] Optionally, generating the communication quality data matrix includes: A data acquisition task table is established based on the conference session identifier. The smart sensors are connected to the same data acquisition bus. Each data acquisition task is configured with sensor address, link node identifier, data field code, sampling period, sampling start time and data quality flag to form the original sampling record. Preprocessing is performed on the original sampling records to delete records that are uploaded repeatedly under the same meeting number, terminal number, streaming media number, data field code, and collection time. The records are aligned according to the unified sampling time, which is equal to the sampling start time plus the product of the sampling sequence number and the sampling period. Missing sample values ​​are filled in by the arithmetic mean of the previous and next valid sample values. The filled sample values ​​are then normalized. Using data field encoding as row index, uniform sampling time as column index, and normalized value as matrix element, a communication quality data matrix is ​​generated according to the same conference number, terminal number, and streaming media number. Data quality flags are written into the auxiliary identifier area of ​​the communication quality data matrix.

[0010] Optionally, the generation of a stable propagation set for communication links includes: Based on the positional relationship of each data field in the communication quality data matrix at the acquisition end, transmission end, processing end and presentation end in the video conferencing link, a state association unit composed of the start field, arrival field and unified sampling time is established. Cross-source state synchronization processing is performed on the state association unit, and the continuous change relationship, structural consistency relationship and link coupling relationship between the starting field and the arriving field are calculated in sequence to generate the state evolution tensor. The coupling propagation path between various communication quality indicators is determined based on the state evolution tensor. The path offset in the coupling propagation path is recursively corrected to generate a structurally stable propagation sequence. A propagation mutation separation process is performed on the structurally stable propagation sequence, and a set of stable propagation features of the communication link is generated based on the separated stable propagation results and mutation results.

[0011] Optionally, obtaining the local mean decomposition feature set includes: The stable propagation feature set of the communication link is formed into a link feature sequence to be decomposed according to the link node, data field and sampling time. The local mean decomposition algorithm is used to extract local extrema, generate local mean, generate local envelope, remove mean and normalize frequency modulation in sequence to obtain the product function components and residual trend term. Based on structural contradiction-driven calculation, the link state is analyzed for the product function components and residual trend terms. The component energy, component change direction and residual trend direction of each link node at different component levels are extracted. The state inconsistency relationship between link nodes is calculated based on the link function relationship between the acquisition end, transmission end, processing end and presentation end in the video conference, and structural contradiction features are generated. Tensor relation closure calculus is introduced to complete the structural contradiction features. A third-order relation tensor is constructed with the starting link node, the arriving link node, and the component level. The structural contradiction features are written into the corresponding positions of the third-order relation tensor to form direct link relations. Indirect link relations are generated based on the link transmission relations between the starting link node, intermediate link node, and arriving link node under the same component level, and the closure completion result is obtained. The product function components, residual trend terms, structural contradiction features, and closure completion results are combined according to link nodes, data fields, component levels, and sampling times to obtain a set of local mean decomposition features.

[0012] Optionally, the structured system for forming communication quality includes: Read the component records corresponding to the same meeting number, the same link node, the same data field and the same sampling time from the local mean decomposition feature set, mark the product function component as the fluctuation existence unit, mark the residual trend term as the trend existence unit, mark the structural contradiction feature as the conflict existence unit, and mark the closure completion result as the closure existence unit. For fluctuating and trend-existing units, an existence stability evolution process is performed. The existence stability value is equal to the ratio of the absolute value of the unit value at the current sampling time minus the unit value at the previous sampling time to the sum of the absolute values ​​of the unit values ​​at the current sampling time. The existence stability value at the first sampling time is the absolute value of the unit value at the current sampling time. A stable existence state is generated based on the existence stability value. Mutually exclusive existence relationship generation processing is performed on conflicting and closed existence units. Conflicting and closed existence units that exist simultaneously within the same link node are combined into mutually exclusive existence pairs. Existence consistency generation processing is performed based on the stable existence state and the mutually exclusive existence pairs. Existence units with the same link node identifier in the acquisition end, transmission end, processing end, and presentation end are arranged according to the sampling time to generate a communication quality existence structure system. A multi-channel communication quality feature system is generated based on the communication quality existence structure system.

[0013] Optionally, obtaining the communication quality characterization vector includes: An improved InceptionTime model is constructed by setting a structural information compression-release dual-state structure between the input module and the parallel feature extraction module of the original InceptionTime model, setting a path self-avoidance topology rearrangement structure between the computation paths of the parallel feature extraction module, and setting a unit phase break reconstruction structure between the residual connection module and the global convergence module. The multi-channel communication quality feature system input structure information is compressed and released into a dual-state structure. The compressed state generation process is performed on the multi-channel features under the same link node. The release state restoration process is performed based on the difference between the compressed state features and the features of each channel to obtain the dual-state structure features. The dual-state structure features are input into the path self-avoidance topology rearrangement structure. The path occupancy record of the calculated paths in the parallel feature extraction module is recorded. The order of the calculated paths is adjusted according to the repetition relationship between the candidate paths and the occupied paths to obtain the self-avoidance path features. The self-avoidance path feature is input into the phase break reconstruction structure of the unit. The continuity of the formation phase, disturbance phase and solidified phase of the computing unit is judged. The computing unit that has phase break is reconstructed to obtain the communication quality characterization vector. An improved InceptionTime model is trained using historical video conferencing communication status samples labeled with communication quality level and anomaly type. The training samples are then input into the improved InceptionTime model to obtain prediction results. The classification loss is calculated based on the prediction results and corresponding labels. The model parameters are then updated according to the classification loss to obtain the trained improved InceptionTime model.

[0014] Optionally, the formation of communication quality levels and anomaly types includes: Read the communication quality characterization vector, expand the candidate results of the communication quality characterization vector according to the communication quality level category and the anomaly type category, and generate a set of multi-path decision candidate units; Perform decision stability convergence calculation on the multi-path decision candidate unit set, and generate the convergence stability result corresponding to each candidate unit based on the characterization change relationship of each candidate unit in continuous sampling time. Conflict resolution calculations are performed on the multi-path decision candidate unit set. Based on the correspondence between the level candidate results and the anomaly type candidate results, candidate units that are inconsistent with the current communication state are screened out to form decision locking units. The decision locking unit generates communication quality levels and anomaly types. Communication quality levels include normal, minor anomaly, moderate anomaly, and severe anomaly. Anomaly types include network transmission anomaly, audio and video quality anomaly, terminal operation anomaly, and service link anomaly.

[0015] Optionally, the output communication quality warning information includes: Read the communication quality level, anomaly type, decision locking unit, and stable propagation feature set of the communication link, and generate early warning input records according to the meeting number, terminal number, streaming media number, and sampling time; Multi-level triggering events are generated based on communication quality level, anomaly type, decision locking unit, and communication link stable propagation characteristic set. These multi-level triggering events are then associated and arranged in the order of level triggering, type triggering, decision triggering, and link triggering. Taking the link node corresponding to the anomaly type as the starting point for backtracking, the linkage backtracking process is performed on the stable propagation feature set of the communication link. Combining the structural contradiction features and closure completion results in the local mean decomposition feature set, a solidified structure of the anomaly propagation path is generated. Based on the fixed structure of the abnormal propagation path, the system generates and grades early warning events, and outputs communication quality early warning information including meeting number, terminal number, streaming media number, abnormal type, abnormal link location, communication quality level, early warning level, and sampling time.

[0016] The beneficial effects of this invention are: This invention collects multi-source communication status data during video conferencing using intelligent sensors. By combining cross-source status synchronization, structural propagation feature generation, local mean decomposition, and a structural system for communication quality, it enhances the automatic identification capability of video conferencing communication quality anomalies. Compared to existing monitoring methods that rely solely on network parameters, terminal status, or manual inspection, this invention simultaneously covers key aspects such as terminal operation, audio acquisition, video acquisition, network transmission, streaming media links, server operation, and display output, fully reflecting the link relationships between the acquisition end, transmission end, processing end, and presentation end. Through structural contradiction-driven calculation and tensor relationship closure calculus, this invention can comprehensively analyze link status inconsistencies, abnormal propagation relationships, and indirect fault associations, improving the timeliness and accuracy of detecting and locating problems such as video stuttering, audio interruption, audio-visual asynchrony, and abnormal streaming media forwarding.

[0017] This invention improves the InceptionTime model to extract communication state representations from a multi-channel communication quality feature system. Utilizing a structural information compression-release dual-state structure, path self-avoidance topology rearrangement, and unit phase break reconstruction, it enhances the model's ability to identify weak anomalies, sudden anomalies, and continuous link anomalies in complex conference scenarios. Through irreversible decision generation, decision stability convergence calculation, and conflict resolution calculation, this invention can generate stable communication quality level and anomaly type judgment results. Furthermore, by combining multi-layered triggering events, linked backtracking calculation, and hierarchical output of early warning events, it generates communication quality early warning information including the location of abnormal links, early warning level, and sampling time. This invention overcomes the problems of insufficient utilization of multi-source data, unclear link fault location, delayed early warning, and weak anomaly type differentiation in existing technologies, making it suitable for real-time monitoring, intelligent operation and maintenance, and rapid fault handling in video conferencing. Attached Figure Description

[0018] 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 a video conferencing communication quality monitoring method based on intelligent sensors proposed in this invention; Figure 2 This is a structural block diagram of the local mean decomposition algorithm for a video conferencing communication quality monitoring method based on intelligent sensors proposed in this invention. Figure 3 This is a functional diagram of the improved InceptionTime model of a video conferencing communication quality monitoring method based on intelligent sensors proposed in this invention. Detailed Implementation

[0019] 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.

[0020] refer to Figure 1 , Figure 2 and Figure 3 A method for monitoring the quality of video conferencing communication based on intelligent sensors, comprising: Multi-source communication status data is collected by intelligent sensors, and the multi-source communication status data is preprocessed to generate a communication quality data matrix. Perform cross-source state synchronization and structural propagation feature generation on the communication quality data matrix, perform propagation mutation separation and link structure reconstruction, and generate a stable propagation set of communication links; The local mean decomposition algorithm is used to calculate the components of the stable propagation feature set of the communication link. The contradiction evolution is processed based on the structural contradiction-driven calculation, and the tensor relation closure calculation is introduced to perform closure completion processing to obtain the local mean decomposition feature set. Based on the local mean decomposition feature set, existential stability evolution is performed, and mutual exclusion existence relation generation and existence consistency generation processes are executed to form a communication quality existence structure system. An improved InceptionTime model is constructed to extract communication state representations from the multi-channel communication quality feature system. Based on the structural information compression-release dual-state structure, compressed state generation and release state restoration are performed. Path self-avoidance topology rearrangement is used to rearrange the topology order. Unit phase break reconstruction is introduced to determine the continuity of the phase and reorganize the broken units, thus obtaining the communication quality representation vector. Based on the communication quality characterization vector, irreversible decision generation is performed, and decision stability convergence calculation and conflict resolution calculation are carried out to form communication quality level and anomaly type; The system performs multi-level trigger event generation and processing for communication quality levels and anomaly types, conducts linked backtracking calculations, generates and classifies early warning events, and outputs communication quality early warning information.

[0021] In this embodiment, the intelligent sensor includes a conference terminal, audio equipment, video equipment, network interface, streaming media server, conference server, and display output device.

[0022] In this embodiment, the multi-source communication status data includes terminal operation data, audio acquisition data, video acquisition data, network transmission data, streaming media link data, server operation data, and display output data.

[0023] In this embodiment, generating the communication quality data matrix includes: A data acquisition task table is established based on the conference session identifier. The smart sensors are connected to the same data acquisition bus. Each data acquisition task is configured with sensor address, link node identifier, data field code, sampling period, sampling start time and data quality flag to form the original sampling record. Preprocessing is performed on the original sampling records to delete records that are uploaded repeatedly under the same meeting number, terminal number, streaming media number, data field code, and collection time. The records are aligned according to the unified sampling time, which is equal to the sampling start time plus the product of the sampling sequence number and the sampling period. Missing sample values ​​are filled in by the arithmetic mean of the previous and next valid sample values. The filled sample values ​​are then normalized. Using data field encoding as row index, uniform sampling time as column index, and normalized value as matrix element, a communication quality data matrix is ​​generated according to the same conference number, terminal number, and streaming media number. Data quality flags are written into the auxiliary identifier area of ​​the communication quality data matrix.

[0024] In this embodiment, generating a stable propagation set for the communication link includes: Based on the positional relationship of each data field in the communication quality data matrix at the acquisition end, transmission end, processing end and presentation end in the video conferencing link, a state association unit composed of the start field, arrival field and unified sampling time is established. Cross-source state synchronization processing is performed on the state association units, sequentially calculating the continuous change relationship between the starting field and the arriving field, the structural consistency relationship, and the link coupling relationship to generate a state evolution tensor. Specifically, the cross-source state synchronization processing for the state association units involves: Read the normalized values ​​of the starting field and the arriving field at the same unified sampling time, and calculate the continuous change relationship between them according to the sampling time order. The continuous change of the starting field is the normalized value of the starting field at the current sampling time minus the normalized value of the starting field at the previous sampling time. The continuous change of the arriving field is the normalized value of the arriving field at the current sampling time minus the normalized value of the arriving field at the previous sampling time. The continuous change of the first sampling time is zero. Calculate the structural consistency relationship. The structural consistency value is the absolute value of the difference between the normalized values ​​of the starting field and the arriving field, which represents the degree of similarity between the states of the two fields at the same sampling time. Calculate the link coupling relationship. The link coupling value is the product of the continuous change of the starting field, the continuous change of the arriving field, and the structural consistency value, which represents the strength of the synchronous change of the starting field and the arriving field at the current link position. Using the starting field as the first dimension, the arriving field as the second dimension, and the unified sampling time as the third dimension, the link coupling value is written into the corresponding tensor position to generate the state evolution tensor. The coupling propagation path between various communication quality indicators is determined based on the state evolution tensor. Path offsets in the coupling propagation path are recursively corrected to generate a structurally stable propagation sequence. Specifically, generating the structurally stable propagation sequence involves: Read the tensor elements at the same unified sampling time in the state evolution tensor. Each tensor element corresponds to a link coupling value pointing from the start field to the destination field. Connect tensor elements that share the same start field, share the same destination field, or are connected one after the other in the video conferencing link to form candidate coupling propagation paths. Calculate the path offset for each candidate coupling propagation path. The path offset is the difference between the current link coupling value and the arithmetic mean of the adjacent link coupling values ​​in the propagation path. When there are no adjacent link coupling values, the path offset is taken as the current link coupling value. Recursively correct according to the sampling time order. The recursive correction value of the current sampling time is the result of adding the recursive correction value of the previous sampling time to the current path offset. The recursive correction value of the first sampling time is taken as the current path offset. Take the arithmetic mean of the current link coupling value and the recursive correction value to obtain the structural stability propagation value at the current sampling time. Arrange the structural stability propagation values ​​in order of link node, data field, and sampling time to generate a structural stability propagation sequence. A propagation mutation separation process is performed on the structurally stable propagation sequence. Based on the separated stable propagation results and mutation results, a stable propagation feature set for the communication link is generated. Specifically, the generated stable propagation feature set for the communication link is as follows: The structural stability propagation sequence is read according to the link node, data field and sampling time. For the structural stability propagation value of each link, the propagation reference value is calculated. The propagation reference value is the arithmetic mean of the structural stability propagation value before and after the current sampling time. If the current sampling time is the first or last sampling time, the propagation reference value is the arithmetic mean of the two most recent valid structural stability propagation values. The propagation mutation amount is calculated. The propagation mutation amount is equal to the current structural stability propagation value minus the propagation reference value. It is used to characterize the degree of mutation of the current sampling time relative to the adjacent propagation state. The stable propagation result is calculated, which is equal to the current structural stable propagation value minus the propagation mutation amount. It is used to retain the stable changes in the link propagation process. The link node identifier, data field encoding, sampling time, structural stable propagation value, propagation reference value, propagation mutation amount and stable propagation result are associated and stored. They are collected according to the conference number, terminal number and streaming media number to generate a set of stable propagation features of the communication link.

[0025] In this embodiment, obtaining the local mean decomposition feature set includes: The stable propagation feature set of the communication link is divided into a link feature sequence to be decomposed according to the link node, data field, and sampling time. The local mean decomposition algorithm is then used to sequentially extract local extrema, generate local means, generate local envelopes, remove the mean, and perform normalized frequency modulation on the link feature sequence to be decomposed, resulting in the product function components and residual trend term. Specifically, the product function components and residual trend term are obtained as follows: Arrange the stable propagation results of the same link node and the same data field in the stable propagation feature set of the communication link in chronological order at each sampling time to form a link feature sequence to be decomposed. Extract the local maximum and local minimum points in the link feature sequence to be decomposed. Calculate the local mean and local envelope points for two adjacent local extreme points. The local mean point is equal to half the sum of the values ​​of two adjacent local extreme points, and the local envelope point is equal to half the absolute value of the difference between two adjacent local extreme points. The local mean function is obtained by smoothing all local mean points, and the local envelope function is obtained by smoothing all local envelope points. The local mean function is subtracted from the link feature sequence to be decomposed to obtain the mean-removed sequence. The mean-removed sequence is divided by the local envelope function to obtain the normalized frequency modulation sequence. When the envelope of the normalized frequency modulation sequence satisfies the unit envelope condition, the current local envelope function is multiplied by the normalized frequency modulation sequence to generate a product function component. The product function component is subtracted from the link feature sequence to be decomposed to obtain the residual sequence. The residual sequence is used as the new object to be decomposed and the process is repeated until the residual sequence no longer has local extreme value fluctuations, resulting in multiple product function components and the last retained residual trend term. Based on structural contradiction-driven computation, link state analysis is performed on the product function components and residual trend terms. The component energy, component change direction, and residual trend direction of each link node at different component levels are extracted. Based on the link function relationships between the acquisition end, transmission end, processing end, and presentation end in a video conference, the state inconsistency relationships between link nodes are calculated, generating structural contradiction features. Specifically, the generation of structural contradiction features includes: Read the product function components according to the link node and component level, calculate the component energy of each link node at each component level. The component energy is equal to the sum of squares of all sampled values ​​in the current component level divided by the number of sampled values. Calculate the component change direction and residual trend direction. The component change direction is determined by the sign of the difference between the product function component value at the current sampling time and the product function component value at the previous sampling time. A positive difference indicates an upward direction, a negative difference indicates a downward direction, and a zero difference indicates a hold direction. The residual trend direction is determined by the sign of the difference between the residual trend term value at the current sampling time and the residual trend term value at the previous sampling time. Based on the link functional relationships between the acquisition end, transmission end, processing end, and presentation end in a video conference, two link nodes with a forward-backward transmission relationship or a cooperative change relationship are grouped into a contradiction calculation object. The component energy difference, component change direction difference, and residual trend direction difference between the two link nodes are calculated respectively. The component energy difference is equal to the absolute value of the component energy difference between the two link nodes. The component change direction difference is equal to the number of times the component change direction of the two link nodes is inconsistent divided by the number of samplings. The residual trend direction difference is equal to the number of times the residual trend direction of the two link nodes is inconsistent divided by the number of samplings. The component energy difference, component change direction difference, and residual trend direction difference are arithmetically averaged to obtain the structural contradiction value between the link nodes. The link node identifier, component level, sampling time, and structural contradiction value are associated and stored to generate structural contradiction features. Tensor relation closure calculus is introduced to complete the closure of structural contradiction features. A third-order relation tensor is constructed with the starting link node, the arriving link node, and the component level. The structural contradiction features are written into the corresponding positions of the third-order relation tensor to form direct link relationships. Indirect link relationships are generated based on the link transmission relationships between the starting link node, intermediate link node, and arriving link node under the same component level, resulting in the closure completion result. Specifically, the closure completion result is as follows: A third-order relation tensor structure is constructed using the starting link node, the arriving link node, and the component level. Structural contradiction features are written into the tensor elements according to the corresponding link node pairs and component level positions to obtain the direct link relation tensor. Within the same component level, for any position where there are no direct contradiction features between the starting link node and the arriving link node, intermediate link nodes are introduced to extend the path, constructing link combination relationships from the starting link node to the intermediate link node and from the intermediate link node to the arriving link node. The indirect relation value is calculated for the link combination relationship. The indirect relation value is equal to the product of the structural contradiction value from the starting link node to the intermediate link node and the structural contradiction value from the intermediate link node to the arriving link node, and then normalized. The maximum value of the indirect relation value corresponding to all possible intermediate link nodes is taken as the indirect link relation value of the node pair. The direct link relationship value and the indirect link relationship value are compared element by element, and the maximum value of the two is taken as the closed relationship value. All closed relationship values ​​are rewritten in the third-order relation tensor to form the closed completion result. The product function components, residual trend terms, structural contradiction features, and closure completion results are combined according to link nodes, data fields, component levels, and sampling times to obtain a set of local mean decomposition features.

[0026] In this embodiment, the structure system for forming communication quality includes: Read the component records corresponding to the same meeting number, the same link node, the same data field and the same sampling time from the local mean decomposition feature set, mark the product function component as the fluctuation existence unit, mark the residual trend term as the trend existence unit, mark the structural contradiction feature as the conflict existence unit, and mark the closure completion result as the closure existence unit. For both fluctuating and trend-existing units, an existence stability evolution process is performed. The existence stability value is equal to the ratio of the absolute value of the unit value at the current sampling time minus the unit value at the previous sampling time to the sum of the absolute values ​​of the unit values ​​at the current sampling time. The existence stability value at the first sampling time is taken as the absolute value of the unit value at the current sampling time. A stable existence state is generated based on the existence stability value. The existence stability evolution process is specifically as follows: Fluctuation and trend existence units under the same link node, data field, and component level are read from the local mean decomposition feature set and arranged into existence unit sequence according to the sampling time order. Starting from the second sampling time, the existence change is calculated point by point. The existence change is equal to the absolute value of the unit value at the current sampling time minus the unit value at the previous sampling time, which represents the change amplitude of the existence unit between adjacent sampling times. The existence stable value is calculated, which is equal to the existence change divided by the absolute value of the unit value at the current sampling time and the existence change. The first sampling time has no previous sampling value, and the existence stable value is the absolute value of the unit value at the current sampling time. According to the sampling time order, each existence stable value is associated with the corresponding link node, data field, component level, and existence unit type to form an existence stable evolution sequence. Based on the stable value change relationship of consecutive sampling times in the existence stable evolution sequence, the fluctuation existence unit and trend existence unit are converted into the corresponding stable existence state. For conflicting and closed-loop existence units, a mutual exclusion existence relationship generation process is performed. Conflicting and closed-loop existence units existing simultaneously within the same link node are grouped into mutually exclusive existence pairs. Based on the stable existence state and the mutually exclusive existence pairs, an existence consistency generation process is performed. Existence units with the same link node identifier in the acquisition end, transmission end, processing end, and presentation end are arranged according to the sampling time to generate a communication quality existence structure system. A multi-channel communication quality feature system is then generated based on the communication quality existence structure system. Specifically, the mutual exclusion existence relationship generation process is performed as follows: Read conflicting and closed existence units under the same meeting number, the same link node, the same data field, and the same sampling time. Pair them to form mutually exclusive existence pairs. Conflicting existence units are used to represent the degree of inconsistency between link nodes, and closed existence units are used to represent the closure strength of the relationship between link nodes. Calculate the mutually exclusive existence value. The mutually exclusive existence value is equal to the absolute value of the difference between the conflicting existence unit value and the closed existence unit value. It represents the degree of incoordination between the conflicting state and the closed relationship within the same link node. Associate the mutually exclusive existence value with the corresponding stable existence state to form the existence consistency judgment result. Its existence consistency value is equal to 1 minus the product of the mutually exclusive existence value and the stable existence value. The stable existence value is the arithmetic mean of the corresponding stable existence values ​​of the fluctuation existence unit and the trend existence unit under the same link node. Following the link sequence of acquisition end, transmission end, processing end, and presentation end, the existence units, mutually exclusive existence values, and existence consistency values ​​with the same link node identifier are arranged according to the sampling time to generate a communication quality existence structure system. The fluctuation existence state, trend existence state, mutually exclusive existence relationship, and existence consistency result in the communication quality existence structure system are respectively regarded as different channels to form a multi-channel communication quality feature system.

[0027] In this embodiment, obtaining the communication quality characterization vector includes: An improved InceptionTime model is constructed by setting a structural information compression-release dual-state structure between the input module and the parallel feature extraction module of the original InceptionTime model, setting a path self-avoidance topology rearrangement structure between the computation paths of the parallel feature extraction module, and setting a unit phase break reconstruction structure between the residual connection module and the global convergence module. The multi-channel communication quality feature system is input to a structure information compression-release dual-state structure. Compressed state generation is performed on the multi-channel features under the same link node. Release state restoration is then performed based on the differences between the compressed state features and the features of each channel, resulting in the dual-state structure features. Specifically, the dual-state features are as follows: The multi-channel communication quality feature system is read according to the conference number, link node, data channel, and sampling time. All channel feature values ​​of the same link node at the same sampling time are extracted. Compressed state generation processing is performed. The arithmetic mean of all channel feature values ​​under the same link node is calculated to obtain the compressed state feature value of the link node at the current sampling time, which represents the overall communication state of the link node. Release state restoration processing is performed. The difference between each channel feature value and the compressed state feature value is calculated. The difference value is equal to the absolute value of the channel feature value minus the compressed state feature value. The compressed state feature value is added to the difference value of the corresponding channel to obtain the release state feature value of the channel. The compressed state feature value of the same link node at the same sampling time and the release state feature values ​​of each channel are combined according to the channel order to form a dual-state structure feature. The dual-state structural features are input into the path self-avoidance topology rearrangement structure. The already calculated paths in the parallel feature extraction module are recorded for path occupancy. The order of calculated paths is adjusted based on the repetition relationship between candidate paths and already occupied paths to obtain the self-avoidance path features. Specifically, the self-avoidance path features are as follows: The bi-state structure features are input into the parallel feature extraction module of the improved InceptionTime model. The link node sequence, channel sequence, and sampling time sequence corresponding to each parallel computing path are recorded as candidate paths. After each round of calculation, the paths that have participated in the calculation are written into the path occupancy table. The path occupancy table includes occupied link nodes, occupied channels, and occupied sampling positions. The repetition relationship between candidate paths and occupied paths is calculated. The path repetition is equal to the sum of the number of link nodes, channels, and sampling positions that are the same as those in the candidate paths and occupied paths, divided by the sum of the number of link nodes, channels, and sampling positions in the candidate paths. The calculation order of candidate paths is adjusted according to the path repetition in ascending order. The output features of the adjusted paths are arranged according to the new topological order. The bi-state structure features, path repetition, and path occupancy status corresponding to each path are retained to generate self-avoidance path features. The self-avoidance path feature is input into the phase break reconstruction structure of the unit. Continuity discrimination is performed on the formation phase, disturbance phase, and solidified phase of the computation unit. Phase break reconstruction processing is performed on the computation unit where phase breakage occurs to obtain the communication quality characterization vector. Specifically, the communication quality characterization vector is as follows: The self-avoidance path feature input unit phase break reconstruction structure is reconstructed by reading the input feature value, internal output feature value and final output feature value of each computing unit according to the computing unit, link node and sampling time. The formed phase, disturbance phase and solidified phase are calculated. The formed phase value is equal to the difference between the current input feature value of the current computing unit and the input feature value of the previous sampling time. The disturbance phase value is equal to the difference between the internal output feature value of the current computing unit and the current input feature value. The solidified phase value is equal to the difference between the current final output feature value of the current computing unit and the final output feature value of the previous sampling time. Phase continuity is determined for adjacent computing units. The phase break value is equal to the sum of the absolute values ​​of the phase difference between adjacent computing units, the absolute values ​​of the phase difference between the disturbance phase, and the absolute values ​​of the phase difference between the solidified phase. For computing units with phase break values, phase break reconstruction processing is performed. The input features, self-avoidance path features, and output features of adjacent computing units are recombined in the order of phase formation, phase disturbance, and solidified phase to generate phase reconstruction features. The phase reconstruction features of the same link node at all sampling times are globally aggregated. The global aggregated value is equal to the arithmetic mean of the corresponding phase reconstruction features at all sampling times. The global aggregated values ​​of each link node are arranged in channel order to obtain the communication quality characterization vector. An improved InceptionTime model is trained using historical video conferencing communication status samples labeled with communication quality level and anomaly type. The training samples are input into the improved InceptionTime model to obtain prediction results. Classification loss is calculated based on the prediction results and corresponding labels. The model parameters are then updated according to the classification loss to obtain the trained improved InceptionTime model. Specifically, the trained improved InceptionTime model is obtained as follows: Based on historical video conferencing communication status samples, a corresponding multi-channel communication quality feature system is generated. Each sample is configured with a communication quality level label and an anomaly type label. The training samples are input into the improved InceptionTime model, which sequentially passes through a structural information compression-release dual-state structure, a path self-avoidance topology rearrangement structure, and a unit phase break reconstruction structure, and outputs the communication quality level prediction result and the anomaly type prediction result. The communication quality level classification loss and the anomaly type classification loss are calculated separately. The communication quality level classification loss is the cross-entropy loss between the true category corresponding to the communication quality level label and the predicted probability of the model output level. The anomaly type classification loss is the cross-entropy loss between the true category corresponding to the anomaly type label and the predicted probability of the model output anomaly type. The total classification loss is the sum of the communication quality level classification loss and the anomaly type classification loss. The trainable parameters in the improved InceptionTime model are updated in reverse according to the total classification loss. After each training round, the training samples are re-inputted to calculate the new total classification loss. When the training rounds reach the preset training round number of 100, the corresponding model structure parameters and output layer parameters are saved to obtain the trained improved InceptionTime model.

[0028] In this embodiment, the formation of communication quality level and anomaly type includes: Read the communication quality representation vector, expand the vector into candidate results according to the communication quality level category and the anomaly type category, and generate a multi-path decision candidate unit set. Specifically, the generation of the multi-path decision candidate unit set is as follows: Read the communication quality representation vector and assign the vector elements to pre-set communication quality level categories and anomaly type categories. The communication quality level categories include normal, minor anomaly, moderate anomaly, and severe anomaly. The anomaly type categories include network transmission anomaly, audio and video quality anomaly, terminal operation anomaly, and service link anomaly. Calculate the candidate support value for each communication quality level category and each anomaly type category. The candidate support value is equal to the sum of the absolute values ​​of the vector elements belonging to the category divided by the sum of the absolute values ​​of all vector elements in the communication quality representation vector. Combine each communication quality level category with each anomaly type category to form a level-type candidate pair. Write the corresponding level candidate support value, type candidate support value, sampling time, link node identifier, and conference number into the candidate pair. Arrange all level-type candidate pairs according to sampling time and link node identifier to generate a multi-path decision candidate unit set. A decision stability convergence calculation is performed on the set of multi-path decision candidate units. Based on the characteristic changes of each candidate unit at consecutive sampling times, a convergence and stability result corresponding to each candidate unit is generated. Specifically, the decision stability convergence calculation is performed as follows: The multi-path decision candidate unit set is read according to the meeting number, link node identifier, and sampling time. The candidate support values ​​of the same level-type candidate pair at consecutive sampling times are arranged into a candidate support sequence. The support change between adjacent sampling times is calculated. The support change is equal to the absolute value of the candidate support value at the current sampling time minus the candidate support value at the previous sampling time. The support change at the first sampling time is zero. The convergence stability value of the candidate unit is calculated. The convergence stability value is equal to 1 divided by 1 plus the support change. The smaller the change in the candidate support value, the larger the corresponding convergence stability value. The arithmetic mean of the convergence stability values ​​of the same candidate unit at multiple consecutive sampling times is calculated to obtain the time-period stability result of the candidate unit. The level category, anomaly type category, candidate support value, support change, convergence stability value, and time-period stability result are associated and written into the corresponding candidate unit to generate the convergence stability result corresponding to each candidate unit. Conflict resolution calculations are performed on the multi-path decision candidate unit set. Based on the correspondence between the rank candidate results and the anomaly type candidate results, candidate units inconsistent with the current communication state are eliminated to form decision-locked units. The conflict resolution calculations are specifically performed as follows: Read the candidate results of the level and the candidate results of the anomaly type under the same sampling time and the same link node in the multi-path decision candidate unit set. Establish a candidate consistency table according to the correspondence between communication quality level and anomaly type. Normal level corresponds to no anomaly state. Minor anomaly, moderate anomaly and severe anomaly correspond to one or more anomaly states in network transmission anomaly, audio and video quality anomaly, terminal operation anomaly and service link anomaly, respectively. Calculate the conflict value of each level-type candidate pair. The conflict value is equal to the absolute value of the difference between the level candidate support value and the anomaly type candidate support value. When there is a mismatch in the candidate consistency table, the conflict value is incremented by one. Calculate the candidate locking value. The candidate locking value is equal to the arithmetic mean of the level candidate support value, the anomaly type candidate support value and the convergence stability value minus the conflict value. Arrange the candidate pairs of each level and type in descending order of candidate lock value, filter out candidate units with candidate lock value less than zero and candidate units that are inconsistent with the current communication state, and take the first-ranked and not-filtered level and type candidate pair as the decision lock unit, and write the corresponding communication quality level, anomaly type, link node identifier and sampling time into the decision lock result. The decision-locking unit generates communication quality levels and anomaly types. Communication quality levels include normal, minor anomaly, moderate anomaly, and severe anomaly. Anomaly types include network transmission anomalies, audio / video quality anomalies, terminal operation anomalies, and service link anomalies. Specifically, the generation of communication quality levels and anomaly types based on the decision-locking unit is as follows: Read the candidate results of the level, candidate results of the anomaly type, candidate lock value, link node identifier and sampling time recorded in the decision locking unit. Use the candidate results of the level as the video conferencing communication quality level at the current sampling time, and use the candidate results of the anomaly type as the video conferencing anomaly type at the current sampling time. When there are multiple decision locking units corresponding to the same conference number, read the candidate lock value of each decision locking unit respectively, and use the decision locking unit with the largest candidate lock value as the main decision unit. Output the final communication quality level and anomaly type based on the main decision unit. When multiple decision-making locking units have the same candidate locking value, the decision-making locking unit with the higher corresponding level is selected as the main decision-making unit in the order of severe anomaly, moderate anomaly, slight anomaly, and normal. The level candidate results in the main decision-making unit are written into the communication quality level field, and the anomaly type candidate results are written into the anomaly type field. The link node identifier, sampling time, and candidate locking value are synchronously written into the result record to obtain the communication quality level and anomaly type. The communication quality level includes normal, slight anomaly, moderate anomaly, and severe anomaly. The anomaly type includes network transmission anomaly, audio and video quality anomaly, terminal operation anomaly, and service link anomaly.

[0029] In this embodiment, the output communication quality warning information includes: Read the communication quality level, anomaly type, decision locking unit, and stable propagation feature set of the communication link, and generate early warning input records according to the meeting number, terminal number, streaming media number, and sampling time; Multi-level triggering events are generated based on communication quality level, anomaly type, decision locking unit, and communication link stable propagation characteristic set. These multi-level triggering events are then associated and arranged in the order of level triggering, type triggering, decision triggering, and link triggering. Taking the link node corresponding to the anomaly type as the starting point for backtracking, a linked backtracking process is performed on the stable propagation feature set of the communication link. Combining the structural contradiction features and closure completion results in the local mean decomposition feature set, a solidified structure of the anomaly propagation path is generated. Specifically, the generation of the solidified structure of the anomaly propagation path is as follows: The corresponding backtracking starting point link node is determined according to the type of anomaly. For network transmission anomalies, the backtracking starting point is the transmission end link node; for audio and video quality anomalies, the backtracking starting point is the acquisition end or presentation end link node; for terminal operation anomalies, the backtracking starting point is the terminal side link node; and for service link anomalies, the backtracking starting point is the streaming media server or conferencing server link node. Following the link sequence of acquisition, transmission, processing, and presentation, the stable propagation results and propagation mutation amounts of adjacent link nodes are read backward from the backtracking starting point in the stable propagation feature set of the communication link. The backtracking correlation value between adjacent link nodes is calculated. The backtracking correlation value is equal to the product of the absolute value of the propagation mutation amount of the current link node and the absolute value of the propagation mutation amount of the previous link node, and then multiplied by the reciprocal of the absolute value of the difference between the stable propagation results of the current link node and the stable propagation results of the previous link node. The backtracking correlation value is matched with the structural contradiction features and closure completion results in the local mean decomposition feature set. The path solidification value is equal to the arithmetic mean of the backtracking correlation value, structural contradiction value, and closure completion value. The order of link nodes through which the abnormal propagation passes is determined according to the path solidification value from large to small. The backtracking starting point, nodes through which the propagation passes, corresponding abnormal type, path solidification value, sampling time, and meeting number are associated and solidified to generate an abnormal propagation path solidification structure. Based on the fixed structure of the abnormal propagation path, the system performs early warning event generation and hierarchical processing, outputting communication quality early warning information including meeting number, terminal number, streaming media number, abnormal type, abnormal link location, communication quality level, early warning level, and sampling time. Specifically, the early warning event generation and hierarchical processing involves: Read the conference number, terminal number, streaming media number, anomaly type, anomaly link node, path fixed value and sampling time from the anomaly propagation path fixed structure, read the communication quality level, merge the anomaly propagation path fixed structures under the same conference number, the same terminal number, the same streaming media number and the same sampling time into one warning event, calculate the warning intensity value, the warning intensity value is equal to the arithmetic mean of the path fixed value, the communication quality level value and the number of anomaly links; Normal, minor anomaly, moderate anomaly, and severe anomaly are assigned values ​​of 0, 1, 2, and 3, respectively. The number of abnormal links is the number of abnormal link nodes contained in the fixed structure of the abnormal propagation path. The warning is graded according to the warning intensity value. When the warning intensity value is equal to 0, a no-warning state is generated. When the warning intensity value is greater than 0 and less than or equal to 1, a prompt warning is generated. When the warning intensity value is greater than 1 and less than or equal to 2, a general warning is generated. When the warning intensity value is greater than 2 and less than or equal to 3, an important warning is generated. When the warning intensity value is greater than 3, an emergency warning is generated. The meeting number, terminal number, streaming media number, anomaly type, abnormal link location, communication quality level, warning level, fixed path value, and sampling time are written into the warning event record, and the communication quality warning information is output.

[0030] Example 1: To verify the feasibility of this invention in practice, it was applied to a multi-venue video conferencing communication quality monitoring scenario, including a main venue terminal, branch venue terminals, remote participant terminals, a streaming media server, a conference server, audio acquisition devices, video acquisition devices, network interface devices, and display output devices. To verify the method of this invention, historical video conferencing communication status samples were constructed as training data, including 4200 normal samples, 1600 network transmission anomaly samples, 1350 audio / video quality anomaly samples, 980 terminal operation anomaly samples, and 1120 service link anomaly samples. Each sample contains 60 sampling points, and each sampling point contains 7 types of data and 52 monitoring fields. A total of 1800 test samples were generated, including 720 normal samples and 1080 anomaly samples, used for comparison with traditional threshold monitoring methods.

[0031] During a simulated conference run, multi-source communication status data was collected with a sampling period of 1 second. A total of 3240 original sampling records were generated, including 286 duplicate records, 173 missing records, and 91 abnormal spike records. After preprocessing, duplicate records were deleted, and missing values ​​were filled in using the arithmetic mean of the preceding and following valid samples. After normalization, a communication quality data matrix of 52 rows and 900 columns was formed. During the stable phase, the normalized network latency ranged from 0.18 to 0.24, the packet loss rate from 0.03 to 0.06, the video frame rate from 0.91 to 0.96, the server load from 0.38 to 0.45, and the display refresh rate from 0.88 to 0.94.

[0032] Within the sampling window, simulated shared screen stuttering and intermittent remote audio occurred. Traditional methods only detected a decrease in the frame rate of the remote terminal; this invention performs cross-source state synchronization and structural propagation feature generation. After the network transmission field and the streaming media link field form state association units, the continuous change relationship increases from 0.09 to 0.42, the structural consistency relationship decreases from 0.86 to 0.51, the link coupling relationship increases from 0.08 to 0.31, and the propagation mutation amount from the transmission end to the processing end reaches 0.44. After path offset recursive correction, the structural stable propagation value increases from 0.27 to 0.63, and the propagation mutation separation yields a stable propagation result of 0.46 and a mutation result of 0.17.

[0033] Local mean decomposition was performed on the stable propagation feature set of the communication link, yielding three product function components and one residual trend term. The first product function component has an energy of 0.38, corresponding to short-term network jitter; the second product function component has an energy of 0.52, corresponding to streaming media queue backlog; and the third product function component has an energy of 0.21, corresponding to display refresh fluctuations. The residual trend term increased from 0.19 to 0.47. Structural inconsistency-driven calculations showed an inconsistency between the transmission and processing ends at 0.71, and between the processing and presentation ends at 0.66. Tensor relationship closure calculations yielded a closure completion value of 0.79, indicating that the anomaly spread from network transmission fluctuations to the service link.

[0034] The local mean decomposition feature set was transformed into a communication quality existence structure system. The stability value of the fluctuating existence unit was 0.63, the stability value of the trend existence unit was 0.58, the mutual exclusion existence value was 0.35, and the existence consistency value was 0.72. After inputting into the improved InceptionTime model, the structural information compression-release bi-state structure generated compressed state features of 0.61 and released state features of 0.76. Path self-avoidance topology rearrangement reduced the proportion of repeated paths from 38.5% to 14.2%, and the phase break reconstruction detected a phase break value of 0.69. The candidate support value for network transmission anomalies was 0.72, for service link anomalies 0.64, for audio / video quality anomalies 0.43, and for terminal operation anomalies 0.28, ultimately narrowing it down to moderate anomalies and network transmission anomalies.

[0035] Using the link node corresponding to the anomaly type as the starting point for backtracking, the anomaly propagation path is obtained as the transmission end, streaming media server, and remote presentation end. The path fixation value is 0.74, the number of anomaly links is 3, and the warning intensity value is 2.58, generating an important warning. Traditional threshold methods only issue an alarm when the video frame rate is below 20 frames per second, while this invention outputs a warning when the frame rate is still 24.8 frames per second, the audio-visual synchronization deviation is 118 milliseconds, and the streaming media queue length is 67 frames, 6 sampling windows in advance. Test results show that the anomaly identification accuracy of the traditional method is 81.6%, while that of this invention is 93.4%; the anomaly type identification accuracy of the traditional method is 76.8%, while that of this invention is 91.2%; the link location accuracy of the traditional method is 69.5%, while that of this invention is 88.7%; the average warning lag of the traditional method is 9.4 sampling windows, while that of this invention is 3.1 sampling windows; and the false alarm rate of the traditional method is 12.6%, while that of this invention is 5.3%. The simulation process shows that the present invention has clear data input, calculation process and output results in each processing stage, which can effectively solve the problems of difficulty in unified perception of multiple source states, delayed detection of communication anomalies, inaccurate location of fault links and unclear early warning classification in video conferencing systems.

[0036] 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 monitoring the quality of video conferencing communication based on intelligent sensors, characterized in that, include: Multi-source communication status data is collected by intelligent sensors, and the multi-source communication status data is preprocessed to generate a communication quality data matrix. Perform cross-source state synchronization and structural propagation feature generation on the communication quality data matrix, perform propagation mutation separation and link structure reconstruction, and generate a stable propagation set of communication links; The local mean decomposition algorithm is used to calculate the components of the stable propagation feature set of the communication link. The contradiction evolution is processed based on the structural contradiction-driven calculation, and the tensor relation closure calculation is introduced to perform closure completion processing to obtain the local mean decomposition feature set. Based on the local mean decomposition feature set, existential stability evolution is performed, and mutual exclusion existence relation generation and existence consistency generation processes are executed to form a communication quality existence structure system. An improved InceptionTime model is constructed to extract communication state representations from the multi-channel communication quality feature system. Based on the structural information compression-release dual-state structure, compressed state generation and release state restoration are performed. Path self-avoidance topology rearrangement is used to rearrange the topology order. Unit phase break reconstruction is introduced to determine the continuity of the phase and reorganize the broken units, thus obtaining the communication quality representation vector. Based on the communication quality characterization vector, irreversible decision generation is performed, and decision stability convergence calculation and conflict resolution calculation are carried out to form communication quality level and anomaly type; The system performs multi-level trigger event generation and processing for communication quality levels and anomaly types, conducts linked backtracking calculations, generates and classifies early warning events, and outputs communication quality early warning information.

2. The video conferencing communication quality monitoring method based on intelligent sensors according to claim 1, characterized in that, The intelligent sensor includes a conference terminal, audio equipment, video equipment, network interface, streaming media server, conference server, and display output device.

3. The video conferencing communication quality monitoring method based on intelligent sensors according to claim 1, characterized in that, The multi-source communication status data includes terminal operation data, audio acquisition data, video acquisition data, network transmission data, streaming media link data, server operation data, and display output data.

4. The video conferencing communication quality monitoring method based on intelligent sensors according to claim 1, characterized in that, The generation of the communication quality data matrix includes: A data acquisition task table is established based on the conference session identifier. The smart sensors are connected to the same data acquisition bus. Each data acquisition task is configured with sensor address, link node identifier, data field code, sampling period, sampling start time and data quality flag to form the original sampling record. Preprocessing is performed on the original sampling records to delete records that are uploaded repeatedly under the same meeting number, terminal number, streaming media number, data field code, and collection time. The records are aligned according to the unified sampling time, which is equal to the sampling start time plus the product of the sampling sequence number and the sampling period. Missing sample values ​​are filled in by the arithmetic mean of the previous and next valid sample values. The filled sample values ​​are then normalized. Using data field encoding as row index, uniform sampling time as column index, and normalized value as matrix element, a communication quality data matrix is ​​generated according to the same conference number, terminal number, and streaming media number. Data quality flags are written into the auxiliary identifier area of ​​the communication quality data matrix.

5. The video conferencing communication quality monitoring method based on intelligent sensors according to claim 1, characterized in that, The generated stable propagation set of the communication link includes: Based on the positional relationship of each data field in the communication quality data matrix at the acquisition end, transmission end, processing end and presentation end in the video conferencing link, a state association unit composed of the start field, arrival field and unified sampling time is established. Cross-source state synchronization processing is performed on the state association unit, and the continuous change relationship, structural consistency relationship and link coupling relationship between the starting field and the arriving field are calculated in sequence to generate the state evolution tensor. The coupling propagation path between various communication quality indicators is determined based on the state evolution tensor. The path offset in the coupling propagation path is recursively corrected to generate a structurally stable propagation sequence. A propagation mutation separation process is performed on the structurally stable propagation sequence, and a set of stable propagation features of the communication link is generated based on the separated stable propagation results and mutation results.

6. The video conferencing communication quality monitoring method based on intelligent sensors according to claim 1, characterized in that, The obtained local mean decomposition feature set includes: The stable propagation feature set of the communication link is formed into a link feature sequence to be decomposed according to the link node, data field and sampling time. The local mean decomposition algorithm is used to extract local extrema, generate local mean, generate local envelope, remove mean and normalize frequency modulation in sequence to obtain the product function components and residual trend term. Based on structural contradiction-driven calculation, the link state is analyzed for the product function components and residual trend terms. The component energy, component change direction and residual trend direction of each link node at different component levels are extracted. The state inconsistency relationship between link nodes is calculated based on the link function relationship between the acquisition end, transmission end, processing end and presentation end in the video conference, and structural contradiction features are generated. Tensor relation closure calculus is introduced to complete the structural contradiction features. A third-order relation tensor is constructed with the starting link node, the arriving link node, and the component level. The structural contradiction features are written into the corresponding positions of the third-order relation tensor to form direct link relations. Indirect link relations are generated according to the link transmission relations between the starting link node, intermediate link node, and arriving link node under the same component level, and the closure completion result is obtained. The product function components, residual trend terms, structural contradiction features, and closure completion results are combined according to link nodes, data fields, component levels, and sampling times to obtain a set of local mean decomposition features.

7. The video conferencing communication quality monitoring method based on intelligent sensors according to claim 1, characterized in that, The aforementioned structured system for forming communication quality includes: Read the component records corresponding to the same meeting number, the same link node, the same data field and the same sampling time from the local mean decomposition feature set, mark the product function component as the fluctuation existence unit, mark the residual trend term as the trend existence unit, mark the structural contradiction feature as the conflict existence unit, and mark the closure completion result as the closure existence unit. For fluctuating and trend-existing units, an existence stability evolution process is performed. The existence stability value is equal to the ratio of the absolute value of the unit value at the current sampling time minus the unit value at the previous sampling time to the sum of the absolute values ​​of the unit values ​​at the current sampling time. The existence stability value at the first sampling time is the absolute value of the unit value at the current sampling time. A stable existence state is generated based on the existence stability value. Mutually exclusive existence relationship generation processing is performed on conflicting and closed existence units. Conflicting and closed existence units that exist simultaneously within the same link node are combined into mutually exclusive existence pairs. Existence consistency generation processing is performed based on the stable existence state and the mutually exclusive existence pairs. Existence units with the same link node identifier in the acquisition end, transmission end, processing end, and presentation end are arranged according to the sampling time to generate a communication quality existence structure system. A multi-channel communication quality feature system is generated based on the communication quality existence structure system.

8. The video conferencing communication quality monitoring method based on intelligent sensors according to claim 1, characterized in that, The obtained communication quality characterization vector includes: An improved InceptionTime model is constructed by setting a structural information compression-release dual-state structure between the input module and the parallel feature extraction module of the original InceptionTime model, setting a path self-avoidance topology rearrangement structure between the computation paths of the parallel feature extraction module, and setting a unit phase break reconstruction structure between the residual connection module and the global convergence module. The multi-channel communication quality feature system input structure information is compressed and released into a dual-state structure. The compressed state generation process is performed on the multi-channel features under the same link node. The release state restoration process is performed based on the difference between the compressed state features and the features of each channel to obtain the dual-state structure features. The dual-state structure features are input into the path self-avoidance topology rearrangement structure. The path occupancy record of the calculated paths in the parallel feature extraction module is recorded. The order of the calculated paths is adjusted according to the repetition relationship between the candidate paths and the occupied paths to obtain the self-avoidance path features. The self-avoidance path feature is input into the phase break reconstruction structure of the unit. The continuity of the formation phase, disturbance phase and solidified phase of the computing unit is judged. The computing unit that has phase break is reconstructed to obtain the communication quality characterization vector. An improved InceptionTime model is trained using historical video conferencing communication status samples labeled with communication quality level and anomaly type. The training samples are then input into the improved InceptionTime model to obtain prediction results. The classification loss is calculated based on the prediction results and corresponding labels. The model parameters are then updated according to the classification loss to obtain the trained improved InceptionTime model.

9. The video conferencing communication quality monitoring method based on intelligent sensors according to claim 1, characterized in that, The formation of communication quality levels and anomaly types includes: Read the communication quality characterization vector, expand the candidate results of the communication quality characterization vector according to the communication quality level category and the anomaly type category, and generate a set of multi-path decision candidate units; Perform decision stability convergence calculation on the multi-path decision candidate unit set, and generate the convergence stability result corresponding to each candidate unit based on the characterization change relationship of each candidate unit in continuous sampling time. Conflict resolution calculations are performed on the multi-path decision candidate unit set. Based on the correspondence between the level candidate results and the anomaly type candidate results, candidate units that are inconsistent with the current communication state are screened out to form decision locking units. The decision locking unit generates communication quality levels and anomaly types. Communication quality levels include normal, minor anomaly, moderate anomaly, and severe anomaly. Anomaly types include network transmission anomaly, audio and video quality anomaly, terminal operation anomaly, and service link anomaly.

10. A video conferencing communication quality monitoring method based on intelligent sensors according to claim 1, characterized in that, The output communication quality warning information includes: Read the communication quality level, anomaly type, decision locking unit, and stable propagation feature set of the communication link, and generate early warning input records according to the meeting number, terminal number, streaming media number, and sampling time; Multi-level triggering events are generated based on communication quality level, anomaly type, decision locking unit, and communication link stable propagation characteristic set. These multi-level triggering events are then associated and arranged in the order of level triggering, type triggering, decision triggering, and link triggering. Taking the link node corresponding to the anomaly type as the starting point for backtracking, the linkage backtracking process is performed on the stable propagation feature set of the communication link. Combining the structural contradiction features and closure completion results in the local mean decomposition feature set, a solidified structure of the anomaly propagation path is generated. Based on the fixed structure of the abnormal propagation path, the system generates and grades early warning events, and outputs communication quality early warning information including meeting number, terminal number, streaming media number, abnormal type, abnormal link location, communication quality level, early warning level, and sampling time.

Images