Intelligent call capability monitoring method and system based on cloud mobile phone, and storage medium

Abstract

The invention discloses an intelligent call capability monitoring method and system based on a cloud mobile phone and a storage medium, and the method comprises the steps: collecting and preprocessing audio, video and other heterogeneous data through a QUIC protocol, dynamically evaluating the weight through a Transform architecture model to generate an abnormal confidence coefficient, triggering multi-modal data backtracking during abnormality, and carrying out the correlation analysis of a high-frequency sampling bottom hardware index. And constructing a three-level time sequence evidence chain and a root cause knowledge graph to position a fault root, and performing joint training and reinforcement learning parameter adjustment in combination with user feedback to realize online iteration of the model. The system comprises a heterogeneous data fusion acquisition unit, a dynamic weight evaluation unit and the like, and the steps of the method are correspondingly executed. According to the invention, through multiple dimensions, full-process accurate monitoring of the call data, rapid abnormal positioning and efficient solution are realized, a closed-loop optimization system is constructed, continuous evolution of a monitoring model is promoted, the cloud mobile phone call quality and the user experience are remarkably improved, and a solid technical guarantee is provided for the cloud mobile phone call service.

Description

Technical Field

[0001] The present invention relates to the field of intelligent monitoring technology, and in particular to a method, system and storage medium for intelligent monitoring of call capabilities based on a cloud phone. Background Art

[0002] "Cloud phones" are virtual phones that run in the cloud through cloud computing and virtualization technology. They offer similar functions and application environments to physical phones, enabling remote control and multiple accounts, flexible resource allocation, and reduced hardware costs. With the rise in remote work, the average daily call volume on cloud phones has exceeded 1 billion. Call quality has become a key factor impacting user experience, with users' tolerance for lag and latency dropping to below 50ms. "Call quality" primarily encompasses metrics such as voice quality, connection stability, signal strength, and call latency during a call, and is a core element in measuring call service quality.

[0003] Current call capability monitoring often focuses on basic indicators such as voice quality and signal strength, while ignoring some factors that have a significant impact on the call experience, such as background noise interference, user-end device compatibility, and the accuracy of semantic understanding during the call. In cloud phone calls, due to the diverse types of terminal devices used by users and the complex network environment, a single monitoring indicator cannot fully reflect the true quality of the call, which will result in some potential problems not being discovered and resolved in a timely manner. In addition, the call monitoring methods in the existing technology mostly rely on single-dimensional indicators, such as packet loss rate and latency, which are difficult to fully and accurately reflect the complex problems during the call process. There are also obvious deficiencies in multimodal data fusion collection, accurate location of abnormal root causes, deep mining of hardware underlying features, and dynamic weight adaptive adjustment. It cannot meet the high-quality and efficient monitoring needs of cloud phone call services in diverse scenarios.

[0004] With the popularization of cloud phone applications, users' requirements for call quality have increased, and the limitations of existing technologies have restricted development. Summary of the Invention

[0005] The present invention aims to address the deficiencies in the existing technology of cloud phone call monitoring, and provides a method, system, and storage medium for intelligent call capability monitoring based on cloud phones. It utilizes the characteristics of cloud phones in combination with intelligent technology to build an efficient and accurate call capability monitoring system, thereby ensuring the quality of cloud phone call services and providing users with a stable and high-quality call experience.

[0006] The present invention solves the technical problem by adopting a method for intelligently monitoring call capabilities based on a cloud phone, comprising the following steps: S1. Heterogeneous Data Fusion Collection and Preprocessing: The QUIC protocol is used to collect audio, video, and environmental sensor data streams during a call in real time and transmit them to edge nodes. At these edge nodes, the real-time packet loss rate of the audio data packets and the real-time frame rate of the video stream are calculated as basic quality indicators. Data fusion technology is used to integrate the audio, video, environmental sensor data streams, and basic quality indicators into a cross-modal feature set. After completing heterogeneous data collection, the feature vectors are dynamically weighted. The key to this step is to adapt to the changing priority of indicators in different call scenarios. S2. Dynamic Weight Intelligent Assessment and Anomaly Confidence Generation: The cross-modal feature set is input into a dynamic weight assessment model based on the Transformer architecture. This dynamic weight assessment model uses a self-attention mechanism to learn the weight priorities of various quality indicators and hardware indicators in different scenarios. Combined with a preset typical scenario database and transfer learning strategy, it performs an adaptive quantitative assessment of the current call scenario and outputs a real-time anomaly confidence score and dynamic weight change trajectory data. S3. Intelligently trigger multimodal data backtracking: A circular buffer is deployed at the edge node, cyclically storing raw audio PCM data, uncompressed video frame sequences, and raw environmental sensor sampling values ​​in a sliding time window manner. The real-time monitoring dynamic weight assessment model calculates and outputs an anomaly confidence score in real time, which reflects the probability of an anomaly in the current call quality. When the real-time anomaly confidence score exceeds a preset threshold, the backtracking mechanism is triggered. The core of this step is to ensure that the backtracking data contains the complete context before and after the anomaly. At this time, the raw audio PCM data, uncompressed video frame sequences, and raw environmental sensor sampling values ​​for a preset time period before and after the anomaly triggering moment are immediately extracted from the circular buffer and timestamps accurate to the millisecond level are added to generate a multimodal backtracking data packet. S4. High-frequency penetrating sampling of underlying features and correlation triggering: Hardware probe sampling frequency is increased, and key underlying hardware metrics such as CPU clock cycle utilization (≥1 kHz frequency) and memory access latency are collected in real time through direct access to hardware registers. This collected real-time underlying hardware metric data is input into a correlation analysis engine, which performs real-time matching analysis based on a predefined hardware feature-call quality correlation rule base. When fluctuations in underlying hardware metric data that meet abnormal conditions in the correlation rule base are detected, a hardware anomaly trigger signal is generated. S5. Construction of an abnormal root cause evidence chain and graph reasoning: In response to the real-time abnormality confidence score exceeding a preset threshold or receiving the hardware abnormality trigger signal, the root cause analysis process is initiated; a three-level temporal evidence chain is constructed, and the three-level temporal evidence chain includes: the time backtracking layer evidence is the multimodal backtracking data packet; the hardware indicator layer evidence is the underlying hardware indicator data sequence collected from step S4 and aligned with the timestamp of the multimodal backtracking data packet; the decision weight layer evidence is the weight change segment corresponding to the timestamp of the multimodal backtracking data packet in the dynamic weight change trajectory data; the three-level temporal evidence chain is input into a root cause reasoning model based on a graph neural network; the root cause reasoning model is extended to construct a root cause knowledge graph containing multimodal data nodes, hardware indicator nodes, weight nodes and their spatiotemporal correlation relationships; through graph reasoning analysis, the root cause of the call quality abnormality is located, and the top three root cause entries with the highest confidence and their corresponding confidence scores are output; and based on the association of the root cause entries with a preset abnormality handling knowledge base, corresponding handling suggestions are generated.

[0007] Preferably, step S2 also includes a hardware anomaly compensation mechanism: when step S4 detects abnormal fluctuations in hardware features, the weight priority of the corresponding hardware indicators is automatically increased according to the type of anomaly, and a compensation coefficient is generated based on the device compatibility knowledge graph to dynamically adjust the attention weight distribution of the Transformer encoder network.

[0008] Preferably, it also includes a user feedback joint training step and a reinforcement learning parameter adjustment step. The user feedback joint training step is to collect users' subjective rating data on call quality through a lightweight SDK, and construct a joint training set with the cross-modal feature set of step S1; the reinforcement learning parameter adjustment step is to use user ratings as reward signals, and dynamically optimize the parameters of the Transformer encoder network in step S2 through the Q-learning algorithm to realize online iteration of the evaluation model.

[0009] Preferably, the reinforcement learning parameter adjustment step also includes a three-dimensional dynamic adjustment mechanism: designing a time sliding window to constrain the training data range, dynamically adjusting the learning rate of the gradient descent algorithm based on the hardware load status in the window, and calibrating the resource allocation weight in combination with the current scenario type to achieve a "time-resource-scenario" coupled weight update strategy.

[0010] Preferably, in the step S4 of high-frequency sampling and association analysis of underlying features, when it is detected that the L2 cache hit rate drops by ≥15%, the association analysis engine triggers the following composite analysis: using a CNN model to perform time-frequency domain spectrum analysis on the audio PCM data in the backtracked data packet in step S3; and performing semantic interruption feature extraction on the same audio data based on natural language processing technology.

[0011] Preferably, it also includes a device compatibility penetration monitoring step and a graph application step. The device compatibility penetration monitoring step collects the CPU hardware indicator sequence, uses principal component analysis to reduce the dimension, and inputs it into the Apache Spark distributed framework to build a compatibility knowledge graph of device model-driver version-resource occupancy mode; the graph application step is to call the graph to match the historical fault case library to generate hardware optimization suggestions when the root cause output by S5 contains a device compatibility problem.

[0012] Preferably, in the three-level temporal evidence chain S5, the time backtracking layer includes multimodal backtracking data packets and packet loss event timestamps in QUIC transmission logs; the hardware indicator layer includes high-frequency sampled register-level CPU clock cycle raw data and memory access delay sequences; the decision weight layer includes the heat map change trajectory of the attention weight distribution in the dynamic weight model and the scene switching mark points.

[0013] On the other hand, the present invention adopts a cloud phone-based call capability intelligent monitoring system for implementing the method, comprising the following units: a heterogeneous data fusion acquisition unit configured to execute S1 and output a cross-modal feature set; a dynamic weight intelligent evaluation unit configured to execute S2 and output a real-time anomaly confidence score and dynamic weight change trajectory data; a multimodal backtracking cache unit configured to execute S3 and be controlled by the anomaly confidence score of the evaluation unit; a hardware penetration analysis unit configured to execute S4 and output a hardware anomaly trigger signal; a root cause graph inference unit configured to execute S5 and input includes a backtracking data packet, a hardware indicator sequence, and weight trajectory data; a compatibility graph construction unit configured to execute the device compatibility penetration monitoring; wherein, the data interaction logic of each unit is strictly consistent with steps S1-S5.

[0014] Preferably, the hardware penetration analysis unit includes a dedicated register access interface that samples the CPU clock cycle at a frequency of ≥1KHz; the root cause graph inference unit has a built-in graph neural network acceleration chip that supports real-time spatiotemporal alignment calculation of the three-level evidence chain.

[0015] In another aspect, the present invention further adopts a computer-readable storage medium storing a computer program, which implements the steps of the method when executed by a processor.

[0016] The beneficial effects of the present invention are as follows: Heterogeneous data fusion and dynamic weight assessment are used at the data collection and analysis level to break the limitations of single data monitoring and accurately quantify call quality. Multi-dimensional anomaly monitoring and backtracking mechanisms are used to take a two-pronged approach from multimodal data and underlying hardware indicators to greatly improve the ability to capture and analyze anomalies. Root cause reasoning combined with knowledge graph applications can quickly locate the root cause of anomalies and provide optimization suggestions; closed-loop optimization and dynamic parameter adjustment ensure the continuous evolution of the monitoring model. This solution realizes accurate monitoring of the entire call data process, rapid location of anomalies and efficient resolution, significantly improving call quality and user experience, and providing strong technical support for cloud phone call services. BRIEF DESCRIPTION OF THE DRAWINGS

[0017] Figure 1 It is a flow chart of the monitoring method of the present invention; Figure 2 It is the software main flow chart of the monitoring method of the present invention; Figure 3 It is a system layer architecture diagram of the present invention; Figure 4 It is a diagram of the collaborative relationship between modules in the detection system of the present invention. DETAILED DESCRIPTION

[0018] The present invention will be further described below with reference to the accompanying drawings and examples.

[0019] Example 1: This cloud phone-based intelligent call monitoring system aims to deeply integrate the characteristics of cloud phones with intelligent technologies to build an efficient and accurate call monitoring system. The system adopts a layered collaborative architecture, with a heterogeneous data fusion acquisition module as its data foundation. It implements data storage and hardware monitoring through multimodal data backtracking caching and high-frequency sampling of underlying features. It relies on a dynamic weighted intelligent assessment model for quality quantification and utilizes anomaly root cause knowledge graphs for fault tracing. Ultimately, through device underlying feature penetration monitoring and closed-loop assessment processing modules, an optimized closed loop is formed, enabling full-process call data monitoring, accurate anomaly location, and efficient resolution.

[0020] The heterogeneous data fusion acquisition module uses the QUIC protocol for low-latency data transmission, collecting audio, video, and environmental sensor data during calls and preprocessing basic metrics such as voice packet loss rate and video frame rate at the edge node. This module uses data fusion technology to integrate data from different modalities into a cross-modal feature set (or fused feature vector), providing an analytical foundation for subsequent modules such as dynamic weight intelligent evaluation, ensuring that the data obtained by each module is homogenous, timely, and accurate. QUIC was chosen over TCP primarily because of its UDP-based multiplexing capabilities, which can reduce transmission latency by over 30% in high-concurrency cloud phone call scenarios, as verified in our initial stress testing.

[0021] The multimodal data backtracking cache module deploys a circular buffer at edge nodes, cyclically storing raw audio PCM data, uncompressed video frame sequences, and raw environmental sensor sampling values ​​in a sliding time window. When the anomaly confidence level output by the dynamic weight model exceeds a preset threshold, the module automatically extracts multimodal cached data for a preset duration before and after the anomaly trigger, adds millisecond-accurate timestamps, and generates a multimodal backtracking data packet, providing a key basis for subsequent anomaly analysis. Data storage utilizes a tiered strategy, with normal call data managed as "1 day of hot data + 7 days of cold data." Abnormal data is stored for 30 days after being marked.

[0022] The low-level feature high-frequency sampling and correlation analysis module increases the hardware probe sampling frequency to ≥1 kHz, directly reading key hardware metrics such as CPU clock cycle utilization and memory access latency through dedicated registers. The 1 kHz high-frequency sampling of the hardware penetration analysis unit is crucial, as it captures transient CPU fluctuations that are invisible to traditional monitoring methods. The correlation analysis engine incorporates a built-in rule base for correlating hardware features with call quality. Upon detecting abnormal hardware fluctuations, such as a decrease of ≥15% in the L2 cache hit rate (L2 cache hit rate = (L2 cache hits / total data requests) × 100%), a CNN model is triggered to perform time-frequency domain spectral analysis of the audio PCM data. Using natural language processing techniques, the model extracts semantic interruption features to further identify the root cause of call quality anomalies.

[0023] The dynamic weighted intelligent assessment module is based on a six-layer encoder network based on the Transformer architecture. It inputs a cross-modal feature fusion vector and supports adaptive switching between 12 typical scenarios. The module utilizes a self-attention mechanism to learn the weight priorities of various quality and hardware indicators in different scenarios, and combines a preset database of typical scenarios with a transfer learning strategy to perform adaptive quantitative assessment. Specifically, when abnormal fluctuations in hardware features are detected, the module automatically increases the weight priority of the corresponding hardware indicator based on the type of anomaly, generates compensation coefficients based on the device compatibility knowledge graph, and dynamically adjusts the attention weight distribution of the Transformer encoder network.

[0024] The anomaly root cause knowledge graph construction module builds a root cause reasoning system with a three-level temporal evidence chain: the time backtracking layer includes timestamps of packet loss events in multimodal backtracking packets and QUIC transmission logs; the hardware indicator layer contains register-level CPU clock cycle raw data and memory access latency sequences aligned with backtracking packet timestamps; and the decision weight layer includes a heatmap of attention weight distribution changes and scene transition markers in the dynamic weight model. The root cause knowledge graph is constructed based on a graph neural network extension, outputting the top three most confident root cause entries and their corresponding solutions.

[0025] The device's underlying feature penetration monitoring module collects 28 hardware metrics, including CPU cycle utilization and memory bandwidth. Principal component analysis is used to reduce the dimensionality of these metrics and then inputs them into the Apache Spark distributed framework to construct a compatibility knowledge graph that categorizes device model, driver version, and resource usage patterns. If the root cause of an anomaly identified in the knowledge graph includes device compatibility issues, the graph is then matched against a database of historical failure cases to generate hardware optimization recommendations.

[0026] The closed-loop evaluation processing module establishes a closed-loop evaluation system combining scenario perception, dynamic weight adjustment, and hardware status compensation. A lightweight SDK collects user subjective ratings of call quality, which are then combined with a cross-modal feature set to form a joint training set. Using user ratings as reward signals, the evaluation model parameters are dynamically optimized using a Q-learning algorithm. A time sliding window is designed to constrain the training data range. The learning rate of the gradient descent algorithm is dynamically adjusted based on the hardware load status within the window. Resource allocation weights are calibrated based on the current scenario type, implementing a weight update strategy that couples time, resources, and scenarios.

[0027] Example 2: An intelligent monitoring method based on the system described in Example 1, such as Figure 1 As shown in the figure, 110 - heterogeneous data fusion collection and preprocessing step; 120 - dynamic weight intelligent assessment and anomaly confidence generation step; 130 - multimodal data intelligent trigger backtracking step; 140 - underlying feature high-frequency penetration sampling and correlation triggering step; 150 - anomaly root cause evidence chain construction and graph reasoning step; 160 - device underlying feature penetration monitoring step; 170 - closed-loop assessment processing step; 210 - QUIC protocol encapsulation. Figure 2 The execution process of the method includes the following steps.

[0028] This intelligent monitoring method first performs heterogeneous data fusion collection and preprocessing, and collects audio streams, video streams and environmental sensor data streams during the call in real time through the QUIC protocol. After transmitting them to the edge node, it calculates the real-time packet loss rate of the audio data packet and the real-time frame rate of the video stream as basic quality indicators, and then uses data fusion technology to integrate the multimodal data streams and basic quality indicators into a cross-modal feature set.

[0029] Secondly, in the dynamic weight intelligent evaluation and anomaly confidence generation steps, considering that cloud phone users may run multiple tasks such as games and videos simultaneously, the CPU load fluctuates violently, which makes it difficult for traditional fixed weight models to adapt. Therefore, we introduced a Transformer-based dynamic weight mechanism. The self-attention mechanism adjusts the indicator priority in real time. The above cross-modal feature set is input into the dynamic weight evaluation model based on the Transformer architecture. The self-attention mechanism learns the weight priority of each indicator in different scenarios. Combined with a preset typical scenario database and transfer learning strategy, it performs an adaptive quantitative evaluation of the current call scenario and outputs a real-time anomaly confidence score and dynamic weight change trajectory data. When abnormal fluctuations in hardware features are detected, the weight priority of the corresponding hardware indicator is automatically increased, and a compensation coefficient is generated based on the device compatibility knowledge graph to dynamically adjust the attention weight distribution of the Transformer encoder network.

[0030] Multimodal data is then intelligently triggered for backtracking. A circular buffer is deployed at the edge node, cyclically storing raw audio PCM data, uncompressed video frame sequences, and raw environmental sensor samples in a sliding time window. The anomaly confidence score output by the dynamic weighted assessment model is monitored in real time. When the score exceeds a preset threshold, the raw data for a preset duration around the anomaly triggering moment is immediately extracted from the circular buffer, timestamped to millisecond accuracy, and a multimodal backtracking data packet is generated.

[0031] High-frequency penetration sampling and correlation triggering of underlying features are then performed, increasing the hardware probe sampling frequency to ≥1 kHz. Key underlying hardware metrics, such as CPU clock cycle utilization and memory access latency, are collected in real time by directly accessing hardware registers. This data is then fed into the correlation analysis engine for real-time matching analysis based on a predefined hardware feature-call quality correlation rule base. When a drop in the L2 cache hit rate of ≥15% is detected, the CNN model is triggered to perform time-frequency domain spectrum analysis on the audio PCM data. Natural language processing techniques are then used to extract semantic interrupt features from the same audio data. If fluctuations in underlying hardware metrics meet the abnormal conditions specified in the correlation rule base, a hardware anomaly trigger signal is generated.

[0032] The system constructs an evidence chain for the root cause of an anomaly and performs graph reasoning. In response to a real-time anomaly confidence score exceeding a preset threshold or a hardware anomaly trigger signal, the root cause analysis process is initiated, constructing a three-level temporal evidence chain consisting of a time backtracking layer, a hardware indicator layer, and a decision weight layer. This is like building an evidence building: the time backtracking layer is the foundation, the hardware indicator layer is the support, and the decision weight layer is the roof; all three are indispensable. The three-level temporal evidence chain is input into a root cause reasoning model based on a graph neural network. This model then constructs a root cause knowledge graph consisting of multimodal data nodes, hardware indicator nodes, weight nodes, and their spatiotemporal relationships. Through graph reasoning analysis, the root cause of the call quality anomaly is located, and the top three root cause entries with the highest confidence scores are output. These entries are then linked to a pre-configured anomaly handling knowledge base to generate actionable recommendations.

[0033] The device compatibility penetration monitoring module collects CPU hardware metric sequences, reduces their dimensionality using principal component analysis, and then feeds them into the Apache Spark distributed framework to construct a compatibility knowledge graph based on device model, driver version, and resource usage patterns. If root cause analysis reveals a device compatibility issue, this graph is then matched against a database of historical failure cases to generate hardware optimization recommendations.

[0034] Closed-loop evaluation can also be performed. The user feedback joint training module collects user subjective call quality ratings via a lightweight SDK and combines them with fused feature vectors to construct a joint training set. The reinforcement learning parameter adjustment module uses user ratings as reward signals and dynamically optimizes the parameters of the Transformer encoder network in the dynamic weight evaluation model using the Q-learning algorithm. A time sliding window is designed to constrain the training data range. The learning rate of the gradient descent algorithm is dynamically adjusted based on the hardware load status within the window. Resource allocation weights are calibrated based on the current scenario type, implementing a weight update strategy that couples "time, resources, and scenario." Future plans call for further integration of user feedback data with the device compatibility map to create a three-dimensional optimization model combining "user experience, hardware configuration, and network environment." This requires incorporating more scenario-specific inputs into the reinforcement learning algorithm.

[0035] Example 3: A computer-readable storage medium based on the system described in Example 1 stores a computer program of the method described in Example 2, which implements the various steps of the method for intelligently monitoring call capabilities based on cloud phones when executed by a processor. A computer-readable storage medium stores a computer program for implementing the above method, which implements the various steps of the method for intelligently monitoring call capabilities based on cloud phones when executed by a processor. The storage medium is adapted to the storage architecture of cloud phone server clusters, edge computing nodes or terminal devices, supports distributed storage and centralized storage modes, can run in a cloud phone virtualization environment based on Linux, Windows or Android, and realizes hardware indicator collection and data fusion through the underlying interface of the virtual machine. The program modules correspond one-to-one to the system modules, and functions such as heterogeneous data collection, multimodal backtracking, hardware sampling, knowledge graph construction and dynamic weight evaluation are realized through software logic.

[0036] Through modular program design, the storage medium converts the above seven core processes of cloud phone call monitoring into executable code with cross-platform compatibility. Through reinforcement learning and hardware compensation mechanism, the monitoring model can dynamically evolve with the business scenario; the three-level evidence chain system based on knowledge graph improves the accuracy of anomaly positioning.

[0037] On the other hand, the present invention further discloses a computer device, comprising a memory and a processor, wherein the memory stores a computer program, and when the computer program is executed by the processor, the processor executes the steps of the above method.

[0038] It is understandable that the system, device and storage medium provided in the embodiments of the present invention correspond to the method provided in the embodiments of the present invention, and the explanation, examples and beneficial effects of the relevant contents can refer to the corresponding parts of the above methods.

[0039] In the above embodiments, all or part of the embodiments can be implemented using software, hardware, firmware, or any combination thereof. When implemented using software, all or part of the embodiments can be implemented in the form of a computer program product. The computer program product includes one or more computer instructions. When the computer program instructions are loaded and executed on a computer, all or part of the processes or functions described in the embodiments of the present application are generated. The computer can be a general-purpose computer, a special-purpose computer, a computer network, or other programmable device. The computer instructions can be stored in a computer-readable storage medium or transmitted from one computer-readable storage medium to another. For example, the computer instructions can be transmitted from one website, computer, server, or data center to another website, computer, server, or data center via a wired (e.g., coaxial cable, optical fiber, digital subscriber line (DSL)) or wireless (e.g., infrared, wireless, microwave, etc.) method. The computer-readable storage medium can be any available medium that can be accessed by a computer or a data storage device such as a server or data center that includes one or more available media. The available medium can be a magnetic medium (e.g., a floppy disk, hard disk, tape), an optical medium (e.g., a DVD), or a semiconductor medium (e.g., a solid-state drive (SSD)).

[0040] It should be noted that, in this document, relational terms such as first and second, etc., are used only to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply the existence of any such actual relationship or order between these entities or operations. Moreover, the terms "comprises," "comprising," or any other variants thereof are intended to cover non-exclusive inclusion, so that a process, method, article, or device comprising a series of elements includes not only those elements, but also other elements not explicitly listed, or elements inherent to such process, method, article, or device. In the absence of further limitations, an element defined by the phrase "comprising a ..." does not exclude the presence of other identical elements in the process, method, article, or device comprising the element.

[0041] Each embodiment in this specification is described in a related manner. Similar parts between the various embodiments can be referred to in conjunction with each other. Each embodiment focuses on the differences between the other embodiments. In particular, the system embodiment is generally similar to the method embodiment, so the description is relatively simple. For related parts, refer to the description of the method embodiment.

[0042] The above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit the same. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the aforementioned embodiments, or make equivalent replacements for some of the technical features therein. However, these modifications or replacements do not deviate the essence of the corresponding technical solutions from the spirit and scope of the technical solutions of the various embodiments of the present invention.

Claims

1. A method for intelligent monitoring of call capabilities based on cloud phones, characterized in that: The steps include: S1. Heterogeneous data fusion acquisition and preprocessing: Collect and transmit audio, video, and environmental sensor data to edge nodes in real time through the QUIC protocol, calculate basic indicators such as audio packet loss rate and video frame rate, and integrate multimodal data and indicators to form a cross-modal feature set; S2. Dynamic Weight Intelligent Assessment and Anomaly Confidence Generation: The cross-modal feature set is input into a dynamic weight assessment model based on the Transformer architecture. The model uses a self-attention mechanism to learn the weight priorities of indicators in different scenarios. Combining a preset scenario library with a transfer learning strategy, the model outputs a real-time anomaly confidence score and weight change trajectory. S3, Multimodal Data Intelligent Triggering Backtracking: A ring buffer is deployed on the edge node to store raw data. When the anomaly confidence level exceeds a threshold, multimodal data with a preset time period before and after the anomaly is extracted and a millisecond-level timestamp is added to generate a backtracking data packet. S4. High-frequency penetrating sampling of underlying features and correlation triggering: This tool collects underlying hardware metrics such as CPU clock cycles and memory access latency at a frequency of ≥1 kHz. The correlation analysis engine matches abnormal fluctuations based on a rule base to generate hardware anomaly trigger signals. S5. Construction of abnormal root cause evidence chain and graph reasoning: In response to abnormal signals, a three-level temporal evidence chain is constructed: time traceback layer, hardware indicator layer, and decision weight layer. The input graph neural network is expanded to build a root cause knowledge graph, locate the root cause, and output the top 3 root causes and processing suggestions.

2. The method according to claim 1, characterized in that Step S2 also includes a hardware anomaly compensation mechanism: when step S4 detects abnormal fluctuations in hardware features, the weight priority of the corresponding hardware indicators is automatically increased according to the type of anomaly, and a compensation coefficient is generated based on the device compatibility knowledge graph to dynamically adjust the attention weight distribution of the Transformer encoder network.

3. The method according to claim 1 or 2, characterized in that It also includes a user feedback joint training step and a reinforcement learning parameter adjustment step. The user feedback joint training step is to collect users' subjective rating data on call quality through a lightweight SDK, and construct a joint training set with the cross-modal feature set of step S1; the reinforcement learning parameter adjustment step uses user ratings as reward signals and dynamically optimizes the parameters of the Transformer encoder network in step S2 through the Q-learning algorithm to achieve online iteration of the evaluation model.

4. The method according to claim 3, characterized in that The reinforcement learning parameter adjustment step also includes a three-dimensional dynamic adjustment mechanism: designing a time sliding window to constrain the training data range, dynamically adjusting the learning rate of the gradient descent algorithm based on the hardware load status within the window, and calibrating the resource allocation weights based on the current scenario type to implement a "time-resource-scenario" coupled weight update strategy.

5. The method according to claim 1, wherein In the step S4 of high-frequency sampling and association analysis of underlying features, when it is detected that the L2 cache hit rate drops by ≥15%, the association analysis engine triggers the following composite analysis: using a CNN model to perform time-frequency domain spectrum analysis on the audio PCM data in the backtracked data packet in step S3; and extracting semantic interruption features from the same audio data based on natural language processing technology.

6. The method according to claim 1, characterized in that It also includes a device compatibility penetration monitoring step and a graph application step. The device compatibility penetration monitoring step collects CPU hardware indicator sequences, uses principal component analysis to reduce the dimension, and then inputs them into the Apache Spark distributed framework to build a compatibility knowledge graph of device model-driver version-resource occupancy mode; the graph application step is to call the graph to match the historical fault case library to generate hardware optimization suggestions when the root cause output by S5 contains device compatibility issues.

7. The method according to claim 1, characterized in that In the three-level temporal evidence chain S5, the time backtracking layer includes multimodal backtracking data packets and packet loss event timestamps in QUIC transmission logs; the hardware indicator layer includes high-frequency sampled register-level CPU clock cycle raw data and memory access delay sequences; the decision weight layer includes the heat map change trajectory of the attention weight distribution in the dynamic weight model and the scene switching mark points.

8. A cloud phone-based call capability intelligent monitoring system, used to implement the method described in any one of claims 1 to 7, characterized in that: include: Heterogeneous data fusion acquisition unit: configured to execute S1 and output a cross-modal feature set; Dynamic weight intelligent evaluation unit: configured to execute S2 and output real-time anomaly confidence scores and dynamic weight change trajectory data; A multimodal backtracking cache unit: configured to execute S3 and controlled by the anomaly confidence score of the evaluation unit; Hardware penetration analysis unit: configured to execute S4 and output a hardware anomaly trigger signal; Root cause graph inference unit: configured to execute S5, with input including backtrace data packets, hardware indicator sequences, and weight trajectory data; A compatibility map construction unit: configured to perform the device compatibility penetration monitoring according to claim 6; Among them, the data interaction logic of each unit is strictly consistent with steps S1-S5 of claim 1.

9. The system according to claim 8, characterized in that: The hardware penetration analysis unit includes a dedicated register access interface that samples the CPU clock cycle at a frequency of ≥1KHz; the root cause graph inference unit has a built-in graph neural network acceleration chip that supports real-time spatiotemporal alignment calculation of the three-level evidence chain.

10. A computer-readable storage medium storing a computer program, characterized in that: When the program is executed by a processor, the steps of the method according to any one of claims 1 to 7 are implemented.

Images