AI-based IMS call multi-service priority adaptive triggering feedback system and method
By using an AI-based IMS call multi-service priority adaptive triggering feedback system, the problems of static priority, signaling redundancy and AI bypass separation in IMS call multi-service scheduling are solved. Dynamic priority scheduling, signaling merging and AI closed-loop processing are realized, improving system performance and intelligence level.
Patent Information
- Application Number
- CN202611265955.9
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2026-08-20
- Publication Date
- 2026-11-17
AI Technical Summary
Existing IMS call multi-service scheduling solutions suffer from problems such as static priority configuration that cannot be dynamically adjusted, dispersed multi-service control signaling leading to high link load, large call latency, and poor real-time coordination due to the independence of AI voice processing from the call link.
An AI-based IMS call multi-service priority adaptive triggering feedback system is adopted. It interacts with the AI-ASR/NLP/TTS module through the IMS call access module, and combines the scheduling processor, message merging unit and event-oriented distribution unit to realize dynamic priority scheduling, unified signaling processing and event-oriented distribution, and embeds AI voice closed-loop processing.
It achieves improved performance of dynamic priority scheduling, reduced signaling link load and call latency, improved real-time performance and reliability of AI voice processing, supports hot access to services, fault degradation and fault tracing, and reduces the commercial cost of the system.
Smart Images

Figure CN123123316A_ABST
Abstract
Description
Technical Field
[0001] This invention belongs to the field of IP Multimedia Subsystem (IMS) communication technology, specifically relating to an AI-based IMS call multi-service priority adaptive triggering feedback system and method. Background Technology
[0002] The IP Multimedia Subsystem (IMS) is the core network architecture for current mobile communications and enterprise fixed-line multimedia services. It relies on the SIP session protocol to enable voice and multimedia call connections and widely supports multiple parallel services such as VoLTE, enterprise PBX, and intelligent customer service. Existing IMS call multi-service scheduling solutions have three inherent flaws and lack AI real-time voice interaction closed-loop capabilities, thus limiting system concurrency efficiency and intelligence levels: 1. Business priorities are statically and fixedly configured, lacking dynamic adaptive adjustment capabilities. Existing technology relies solely on manual preset of fixed sorting, unable to dynamically adjust based on real-time system load and user membership levels; new business access requires system shutdown and restart to update configurations, resulting in long business launch cycles; high-priority calls cannot preempt computing power and media resources from low-priority services, and concurrent multi-service operations are prone to logical conflicts.
[0003] 2. Multiple services control signaling is distributed and sent out in a dispersed manner, resulting in high link load and long call latency. Services such as playback, recording, and AI interaction generate SIP control messages independently and push them to the IMS core network one by one. The number of signaling interactions increases linearly with the number of services. Multiple messages are out of order and have a high probability of being lost, significantly increasing call setup latency. Signaling congestion is likely to occur in large-scale concurrent scenarios.
[0004] 3. The IMS call link and the AI voice processing system are deployed independently and in a bypass manner, making real-time collaboration impossible. Traditional solutions treat ASR / NLP / TTS as bypass external platforms, requiring call voice to be forwarded to the AI node, resulting in high voice transmission latency. Without unified scheduling logic, the timing of AI interaction, recording, and playback is disordered, making real-time semantic recognition and multi-turn dialogue feedback during the call impossible. Furthermore, the lack of AI fault degradation and computing power limiting mechanisms means that AI service outages directly lead to the interruption of the entire call.
[0005] In summary, existing IMS multi-service scheduling solutions suffer from multiple technical defects, such as static priority, signaling redundancy, and AI bypass fragmentation. There is an urgent need for an integrated system that combines dynamic priority adaptive scheduling, signaling merging, event-oriented distribution, and end-to-end embedded AI voice closed loop. Summary of the Invention
[0006] To address the problems existing in the prior art, the present invention aims to provide a technical solution for an AI-based IMS call multi-service priority adaptive triggering feedback system and method, applicable to VoLTE high-definition voice, government and enterprise IP calls, operator intelligent customer service, and IMS video multimedia call scenarios.
[0007] The AI-based IMS call multi-service priority adaptive triggering feedback system is characterized by comprising: an IMS call access module, a multi-service triggering module, a service function module, a media module, and an AI-ASR / NLP / TTS module; The IMS call access module is standardized and interfaces with the IMS core network S-CSCF. It establishes an IMS call session based on the SIP RFC3261 protocol, sends and receives SIP unified control messages, synchronizes the voice stream bidirectionally through the media module, and interacts with the AI-ASR / NLP / TTS module. The multi-service triggering module integrates a scheduling processor, a priority policy storage unit, a message merging unit, and an event-oriented distribution unit. The scheduling processor is configured with an adaptive priority scheduling algorithm, dynamically sorting and generating a priority triggering sequence based on service weights, user subscription service information, and service status. The priority calculation formula is Pi = Wi * Si * Ei, where Pi is the comprehensive priority score of the i-th service, Wi is the preset basic weight score of the i-th service (range 0-1000), with a 20-point interval between each service score, allowing new services to be inserted among existing services, Si is the current user's subscription identifier for the i-th service (subscribed = 1, unsubscribed = 0), and Ei is the real-time activation status identifier for the i-th service (activated = 1, offline / fault = 0), with a default activation interval of 500ms. The system automatically refreshes the priority sequence, allowing new services to be dynamically included in the priority sequence without requiring a system restart. The priority policy storage unit stores the basic weights of services and user subscription information. The message merging unit receives distributed control events from various service function modules, removes redundant parameters and duplicate service identifier fields, saves the information related to the control events and services, and encapsulates them into a unified control message for distribution to the IMS call access module. The event-directed distribution unit receives media completion events, AI interaction completion events, and hang-up events reported by the IMS call access module, and only pushes the events to the service function modules with the corresponding service IDs, achieving data isolation and non-interference among multiple service events. The business function module executes playback, recording, and AI voice interaction business logic based on the calling and called number rules. The media module collects two-way call voice streams and pushes them to the AI-ASR / NLP / TTS module, receives synthesized speech and completes real-time playback, and performs recording and media stream forwarding operations. The AI-ASR / NLP / TTS module is deployed on GPU computing nodes and includes an AI noise reduction ASR unit, an NLP semantic interaction unit, and an adaptive TTS synthesis unit. The AI noise reduction uses a three-level noise reduction algorithm to eliminate line current noise, environmental noise, and overlapping voices from multiple people, and outputs structured call text in real time. The NLP semantic interaction unit loads multi-scenario dedicated models and completes keyword recognition, multi-turn dialogue context caching, and response text generation based on the call text. The adaptive TTS synthesis unit supports multilingual and multi-dialect speech synthesis, dynamically adjusts timbre, speech rate, and volume according to the call scenario, and converts the response text into a speech stream to be sent back to the media module.
[0008] The system is characterized in that when the scheduling processor triggers a service module according to priority, if the service module does not respond, it automatically skips to trigger the next service module. If the service module does not respond 3 times, the triggering is determined to be a failure and the service module is automatically downgraded. After receiving the unified control message, the IMS call access module replies with an ACK confirmation signaling. If no ACK is received within the timeout period, it will retransmit up to 3 times.
[0009] The system is characterized in that the event-directed distribution unit is connected to a time-series database; when the business function module is offline, events are cached in the middleware, and undelivered events are re-pushed in batches after the module is restored online; all call events are persistently stored with timestamps, business IDs, and result codes; the system automatically pushes maintenance alarms after recognizing abnormal events such as playback failure, AI recognition timeout, and recording interruption.
[0010] The system is characterized by the AI-ASR / NLP / TTS module being configured with a fault circuit interruption mechanism; when the GPU node goes offline or the ASR recognition timeout rate exceeds a threshold, the real-time NLP semantic interaction function is automatically shut down, while basic recording and fixed playback capabilities are retained, without affecting normal call functions; it supports hot model loading and A / B grayscale switching, allowing interactive model updates to be completed without interrupting online calls; it has a built-in session context cache, which caches the dialogue text throughout a single call to achieve multi-turn continuous semantic understanding, and automatically clears the cache and releases computing power after the call is hung up.
[0011] The system is characterized by: each module being deployed in a cluster; multiple service triggering modules sharing a distributed Redis cluster to store call session states; the system can resume ongoing call sessions after an instant restart, achieving session keep-alive; and load balancing scheduling is performed on multiple GPU nodes, allocating computing resources according to call priority.
[0012] The system is characterized by having a built-in call record generation unit and a security encryption unit; the call record generation unit generates a complete call record at the call hang-up node; the security encryption unit adopts the national cryptographic standard SM4, desensitizes the calling and called numbers when storing logs, encrypts and stores the original voice messages, and sets an automatic deletion period to meet communication privacy compliance requirements.
[0013] The system is characterized in that: the system is compatible with multiple IMS core network elements such as S-CSCF and I-CSCF; the media module is compatible with voice media streams and video IMS audio streams; and the AI-ASR / NLP / TTS module supports noise reduction and recognition of audio during video calls.
[0014] The AI-based IMS call multi-service priority adaptive triggering feedback method is characterized by the following steps when applied to any of the aforementioned systems: (1) The IMS call access module receives the SIP call message issued by the S-CSCF, establishes an IMS call session based on the SIP protocol, extracts the calling and called numbers, call forwarding parameters, and call scenario identifiers to generate a call event and pushes it to the multi-service triggering module. (2) The multi-service triggering module queries the user's subscribed service list and uses an adaptive priority scheduling algorithm to calculate and generate an ordered priority triggering sequence; the system automatically refreshes the priority sequence every 500ms, and new services are included in the sorting in real time; (3) The multi-service triggering module receives the playback, recording, and AI interaction distributed control events returned by each service function module, removes redundant parameters and duplicate service identifiers, and encapsulates them into a single unified control message and sends it to the IMS call access module. (4) The IMS call access module completes SIP call setup based on unified control messages, and triggers the linkage media module and AI-ASR / NLP / TTS module to perform media forwarding and AI voice interaction in parallel according to the control messages; (5) The media module forwards the two-way call voice stream to the AI noise reduction ASR unit in real time, completes noise reduction recognition and outputs the call text; the NLP semantic interaction unit reads the cached multi-turn dialogue context to complete semantic parsing and generate response text; the adaptive TTS synthesis unit adjusts the timbre and speech rate according to the scenario to generate speech, which is played to both parties in real time through the media module; when the AI module fails, the real-time semantic interaction is turned off, while the basic services of recording and fixed playback are retained. (6) The IMS call access module completes the playback, off-hook and hang-up events, and reports the events with the service ID to the multi-service triggering module; the event-directed distribution unit of the multi-service triggering module pushes the events to the corresponding service function modules separately according to the service identifier; when the service module is offline, the events are cached in the middleware and resent in batches after recovery; all call events are stored persistently in full, and maintenance alarms are pushed synchronously when abnormal events are detected; (7) After the call is hung up, the IMS call access module reports the hang-up event, the multi-service triggering module sends the service termination instruction to all service function modules, and simultaneously clears the AI session context cache, releases the media stream, GPU computing power, and session storage resources; the system continuously identifies dead sessions through heartbeat detection and forcibly reclaims the resources they occupy.
[0015] The beneficial effects of this invention are: 1) Significantly improved performance of dynamic priority scheduling. Traditional fixed priority schemes require system shutdown and restart when adding new services, while this invention supports hot access to services without interrupting online calls; the time for a single priority sorting is ≤10ms; dynamic priority scheduling (three-dimensional quantization formula + 500ms automatic refresh + hot access to services) reduces the failure rate of multi-service concurrent conflicts by 85%; when the system is overloaded, it automatically degrades to low-priority services, reducing call stuttering rate by 90% in scenarios with tens of thousands of concurrent connections.
[0016] 2) Significantly reduced signaling link load and call latency. Traditional solutions send N independent SIP messages for N services, while this invention merges them into a single SIP message, reducing the total signaling interaction by 60%; call setup latency is reduced by 40% compared to traditional solutions, and there is no signaling congestion in large-scale concurrent scenarios.
[0017] 3) Targeted event distribution optimizes computing power consumption and fault tracing capabilities. Events are only pushed to the corresponding business modules; middleware caching ensures no event loss for offline services; the time-series database stores all events for at least 7 days, supporting one-click fault tracing by call ID and business ID; real-time alarm push for abnormal audio playback and AI recognition improves operation and maintenance troubleshooting efficiency by 90%.
[0018] 4) Embedded AI voice closed loop, with real-time performance and reliability far exceeding bypass solutions. The AI module is deployed on the main call link, with end-to-end voice recognition latency ≤150ms, a 70% reduction in latency compared to external bypass AI platforms; it supports multi-turn dialogue context caching to achieve continuous semantic interaction; it is configured with a circuit breaker degradation mechanism, so calls are not interrupted when AI fails, only advanced semantic functions are disabled; the model is hot-loaded without downtime for updates, and it supports AI recognition in both voice and video audio scenarios.
[0019] 5) Comprehensive carrier-grade commercial support capabilities and low cost of adaptation to existing networks. Active-active clusters and distributed session storage ensure session persistence, preventing call interruptions during system upgrades; number anonymization and automatic periodic deletion of voice messages meet privacy compliance requirements in the telecommunications industry; compatible with multiple equivalent solutions including S-CSCF / I-CSCF, container / virtual machine deployment, and local GPU / cloud AI interfaces, requiring no large-scale modifications to the existing IMS core network, resulting in low deployment costs and low barriers to commercial deployment.
[0020] 6) The five modules are deeply linked and coupled, resulting in synergistic technical gains that cannot be achieved by deploying a single module independently. In existing technologies, priority scheduling, signaling optimization, AI interaction, and event distribution belong to independent systems. This invention integrates four innovative modules into the same IMS call processing link to form synergistic gains: dynamic priority filtering reduces the unnecessary AI computing power occupation, single unified signaling synchronously schedules media and AI to avoid timing errors, and targeted event isolation prevents interference between multiple service logics. Attached Figure Description
[0021] Figure 1. System architecture diagram of the present invention; Figure 2. Flowchart of the business triggering process of this invention. Detailed Implementation
[0022] The present invention will be further described below with reference to the accompanying drawings: This invention addresses the shortcomings of existing technologies, such as static prioritization, distributed multi-message delivery, global event broadcasting, high latency due to AI bypass, and lack of commercial support capabilities. It provides an AI-based IMS call multi-service priority adaptive triggering feedback system and method to achieve the following objectives: 1) Construct an adaptive priority scheduling mechanism to support dynamic hot access of services, automatic degradation due to overload, and preemption of high-priority service resources without restarting the system to update service configurations; 2) Merging multiple service-distributed control messages into a single unified SIP signaling message significantly reduces the number of signaling interactions and call setup latency; 3) Implement targeted event distribution based on business ID, combined with middleware caching, persistent storage, and anomaly alarms to achieve business isolation and fault traceability; 4) Embed AI-ASR / NLP / TTS into the main IMS call processing link to open up the entire closed loop of voice stream, configure fault circuit interruption, computing power rate limiting, and multi-turn dialogue context caching to ensure the real-time performance of AI interaction and the reliability of calls; 5) It is equipped with carrier-grade commercial capabilities such as active-active cluster, session keep-alive, privacy masking, and SIP attack prevention, and is compatible with S-CSCF / I-CSCF and voice / video IMS multimedia services, thus broadening the scope of scenario adaptation; 6) By deeply coupling hardware modules with scheduling algorithms, the limitations of independent operation of simple scheduling algorithms are overcome, while providing multiple equivalent alternative implementation methods.
[0023] The AI-based IMS call multi-service priority adaptive triggering feedback system proposed in this invention consists of five core hardware / logic modules: IMS call access module, multi-service triggering module, service function module, media module, and AI-ASR / NLP / TTS module. The functions and interaction logic of each module are as follows: (1) IMS call access module: As the gateway for the system to connect with the IMS core network, it connects with S-CSCF and I-CSCF network elements based on the SIP RFC3261 protocol; receives incoming SIP messages from the main network, parses the calling and called parties, call forwarding, and service subscription information to generate call events; receives unified control messages issued by the multi-service triggering module to complete call connection; synchronizes bidirectional media voice streams to the media module, and forwards uplink events such as media events and hang-up to the multi-service triggering module; (2) Multi-service triggering module: Built-in scheduling processor, priority policy storage unit, message merging unit, and event-oriented distribution unit; The priority scheduling processor calculates the priority in real time based on the comprehensive service weight, service status, and user subscription services, and refreshes automatically every 500ms, and adds new services dynamically; The message merging unit collects the scattered control events of each service module, removes redundant fields, and encapsulates and carries unified service control messages to the IMS call access module; The event-oriented distribution unit receives various service completion events reported by the IMS call access module and pushes them to the corresponding service module point-to-point according to the service ID; The middleware caches offline events, the time-series database stores the event logs in full, and identifies abnormal events to push operation and maintenance alarms; (3) Business function module: Multiple groups are deployed in parallel, each group corresponds to a type of business logic; receive trigger instructions issued by the multi-service trigger module to execute the corresponding business logic, and at the same time feed back control events to the multi-service trigger module; receive targeted push business execution completion events, and complete business status updates and log recording; (4) Media module: responsible for the acquisition, forwarding, recording and playback of two-way voice / video audio streams; real-time push of the acquired voice stream to the AI module, and playback of TTS synthesized voice to both parties in the call; supports multi-bitrate voice adaptation for VoLTE high-definition calls, has media stream heartbeat detection capability, and automatically reclaims resources from dead media channels; (5) AI-ASR / NLP / TTS module (GPU cluster deployment): AI noise reduction ASR unit: Three-level noise reduction algorithm to eliminate line current noise, environmental noise, and overlapping voices of multiple people, and recognize speech and output text in real time; NLP semantic interaction unit: Hot switching of dedicated models for multiple scenarios, caching the context of multi-turn dialogue in a single call, and continuously generating response text; Adaptive TTS unit: Multilingual and multi-dialect voice library, automatically adjusting speech rate, volume, and sentence smoothness according to the call scenario; Matching fault circuit interruption mechanism: When AI computing power is overloaded or nodes are offline, real-time semantic interaction is shut down, basic recording and playback are retained to ensure that the call is not interrupted; GPU node load balancing scheduling, high-priority calls are allocated computing power first.
[0024] The system is equipped with commercial auxiliary units: a security encryption unit and an operation and maintenance monitoring unit; the security unit realizes number de-identification and encrypted voice storage; the monitoring unit collects quantitative indicators such as priority scheduling time, message merging success rate, AI recognition latency, and concurrent call volume in real time.
[0025] The complete business process of this invention is divided into six stages: call access, priority sorting, message merging and delivery, parallel execution of media and AI, event-oriented feedback, and release of idle resources. It includes sub-steps for handling exceptions such as AI fault degradation, message retransmission, offline event resending, and session dead reclamation, so as to achieve closed-loop control of the entire process.
[0026] The adaptive priority scheduling algorithm of this invention queries all subscribed service information based on the calling and called numbers in the call event information, generates a service trigger sequence from high to low based on the priority formula Pi=Wi*Si*Ei score, and dynamically refreshes all service information every 500ms, including service addition, service deletion, service availability status, and service priority score.
[0027] The message merging unit of this invention is configured with message ACK response and automatic degradation mechanism; after the IMS call access module receives the unified control message, it replies with ACK confirmation signaling. If the multi-service triggering module does not receive the message within a timeout, it will retransmit the message at 200ms intervals, for a total of 3 times.
[0028] The event-oriented distribution unit of this invention interfaces with middleware and a time-series database; when a business function module is offline, events are cached in the middleware, and when the module is back online, undelivered events are re-pushed in batches; all call events are stored persistently with timestamp, business ID, and result code, with a storage period of no less than 7 days; the system automatically pushes maintenance alarms after recognizing abnormal events such as playback failure, AI recognition timeout, and recording interruption.
[0029] This invention's AI-ASR / NLP / TTS module is configured with a fault circuit interruption mechanism; when the GPU node goes offline or the ASR recognition timeout rate exceeds a threshold, the real-time NLP semantic interaction function is automatically shut down, while basic recording and fixed playback capabilities are retained to ensure uninterrupted calls; it supports hot model loading and A / B grayscale switching, allowing online calls to update the interactive model without interruption; it has a built-in session context cache, which caches the dialogue text throughout a single call to achieve multi-turn continuous semantic understanding, and automatically clears the cache to release computing power after the call is hung up.
[0030] Each module of this invention adopts a containerized cluster active-active deployment. The IMS call access module and the multi-service triggering module share a distributed Redis cluster to store the call session state. After the system restarts instantly, the ongoing call session can be restored, realizing session keep-alive. Multi-GPU node load balancing scheduling allocates computing resources according to call priority.
[0031] The system of this invention has a built-in security encryption unit; the security encryption unit desensitizes the calling and called numbers when storing logs, encrypts and stores the original voice messages and sets an automatic deletion period to meet communication privacy compliance requirements.
[0032] The media module of this invention is compatible with voice media streams and video IMS audio streams; the AI-ASR / NLP / TTS module supports noise reduction and recognition of audio during video calls; the system is compatible with multiple IMS core network elements such as S-CSCF and I-CSCF.
[0033] This invention also provides an AI-based IMS call multi-service priority adaptive triggering feedback method, comprising the following steps: (1) The IMS call access module receives the SIP call message issued by the S-CSCF, establishes an IMS call session based on the SIP protocol, extracts the calling and called numbers, call forwarding parameters, and call scenario identifiers to generate a call event and pushes it to the multi-service triggering module. (2) The multi-service triggering module queries the user's subscribed service list and uses an adaptive priority scheduling algorithm to calculate and generate an ordered priority triggering sequence; the system automatically refreshes the priority sequence every 500ms, and new services are included in the sorting in real time; when the system resources are overloaded, the score of low-priority services is automatically reduced, and queuing and computing power degradation are performed; (3) The multi-service triggering module receives the playback, recording, and AI interaction distributed control events returned by each service function module, removes redundant parameters and duplicate service identifiers, and encapsulates them into a single unified control message and sends it to the IMS call access module; if no ACK response is received, it retransmits. (4) The IMS call access module completes SIP call setup based on unified control messages, and performs media acquisition and AI voice interaction in parallel with the media module and AI-ASR / NLP / TTS module according to the control message requirements; (5) The media module collects the two-way call voice stream in real time and pushes it to the AI noise reduction ASR unit to complete noise reduction recognition and output the call text; the NLP semantic interaction unit reads the cached multi-turn dialogue context to complete semantic parsing and generate response text; the adaptive TTS synthesis unit adjusts the timbre and speech rate according to the scenario to generate speech, which is played to both parties in real time through the media module; when the AI module fails, the real-time semantic interaction is turned off, while the basic services of recording and fixed playback are retained. (6) The IMS call access module monitors events such as playback completion, recording start, AI interaction completion, and hang-up, and reports the events with the service ID to the multi-service triggering module; the event-directed distribution unit of the multi-service triggering module pushes the events to the corresponding service function modules individually according to the service identifier; when the service module is offline, the events are cached in the middleware and resent in batches after recovery; abnormal events are pushed synchronously to operation and maintenance alarms, and all events are persistently stored; (7) After the call is hung up, the IMS call access module reports the hang-up event, and the multi-service triggering module sends a service termination instruction to all service function modules, and simultaneously clears the AI session context cache, releases the media stream, GPU computing power, and session storage resources; when the session is dead, the occupied resources are forcibly reclaimed through heartbeat detection.
[0034] Complete triggering process for this invention: 1) The IMS call access module receives SIP inbound call messages, generates call events, and pushes them to the multi-service triggering module; 2) The multi-service triggering module calculates priorities and generates a service triggering sequence; 3) Each business module returns distributed control events; 4) The message merging unit encapsulates a single unified control message for distribution; 5) The IMS call access module completes call setup, and the media and AI modules execute in parallel. 6) AI real-time noise reduction and recognition, multi-round semantic parsing, and TTS synthesized speech feedback to both parties in the call; 7) Media outlets and AI-generated events should report with the relevant business ID. 8) The event-oriented distribution unit pushes information to the corresponding business modules point-to-point based on the business ID; 9) Report idle events, issue unified termination commands, and release all resources across the entire chain.
[0035] Example 1: Complete Example of a Standard Carrier VoLTE Voice Call The system adopts a containerized cluster active-active deployment, with dual-machine hot standby for the IMS call access module, a 3-node cluster for load balancing of the multi-service triggering module, and 4 GPU computing nodes deployed for AI-ASR / NLP / TTS. A distributed Redis cluster stores session state, Kafka serves as an event buffer middleware, and InfluxDB time-series database stores event logs.
[0036] 1. Priority parameter configuration This includes a default global configuration table, a service configuration table, and a user subscription information table. The global configuration table includes the default triggered service, timeout, playback parameters, recording parameters, and AI parameters. The service configuration table includes the service name, connection address, priority score, timeout parameters, and availability status. The user subscription information table includes user number information, a list of subscribed services, and a list of enabled services.
[0037] 2. Example of a unified control message structure The message body JSON includes the business ID, business type, playback parameters, recording duration, and AI interaction switch; if no ACK confirmation is received, retransmission will be performed every 200ms; if three consecutive failures occur, the transaction will end and all resources will be released.
[0038] 3. AI Interaction Execution Process The call audio stream is fed into the ASR module in real time, and the text is output after eliminating line noise; NLP caches the first 10 rounds of dialogue text for contextual understanding and generates response text; TTS uses a calm business tone and moderate speaking speed to synthesize and play speech in real time; if the GPU computing node is offline, the system disables NLP semantic response, only retains the recording and fixed welcome sound, and the call connection is unaffected; the end-to-end latency of the entire AI recognition and synthesis system is ≤150ms.
[0039] 4. Event-oriented feedback process The playback completion event carries a business ID=001 and is only pushed to the playback business module; the AI interaction end event has a business ID=002 and is only pushed to the AI interaction module; when the business module is offline, the event is cached in the middleware and resent in batches after the module is restored; all events are stored in the database for 7 days.
[0040] 5. Idle resource recycling Upon receiving the SIP BYE hang-up message, the multi-service triggering module sends an end command to all service modules, clearing the Redis session cache, AI dialogue context, and media stream channel; the system checks the media and AI channels with a heartbeat every 2 seconds, and if there is no heartbeat response after 3 times, it is considered dead and CPU and GPU resources are forcibly reclaimed.
[0041] Example 2: Lightweight and Simplified Deployment Example (Small and Medium-sized Enterprise Scenario) This embodiment fully includes all the core modules and scheduling mechanisms of claim 1, with only equivalent replacements in hardware deployment scale and storage components. The system adopts a single-container, single-machine deployment, without cluster dual-active; long-term storage of the time-series database is eliminated, and 7 days of event logs are stored on the local disk; no independent GPU cluster is configured; only the core innovative functions of basic priority sorting, message merging, and event-oriented distribution are retained; it is adapted to small enterprise IP calls, reducing hardware deployment costs, and is an equivalent simplified implementation of the present invention.
[0042] Example 3: AI Model Hot Update and Grayscale Switching Example The AI module loads new models online without interrupting online calls; 1% of calls are routed to the new version model and 99% to the old version model for A / B testing. Once the recognition accuracy meets the standard, all calls are switched to the new version model, ensuring uninterrupted business operations for AI interaction and iteration.
[0043] See appendix Figure 2 A business triggering flowchart of a specific example of the present invention: 1) After receiving the SIP call message from the main network, the IMS call access module generates a call event including the calling number, the called number, call forwarding information, and the triggering process, and sends it to the multi-service triggering module; 2) The multi-service triggering module queries the number subscription service information based on the calling and called number information, triggers services according to priority, and triggers different service modules in sequence. Service module 1 returns a playback control event. 3) Business module 2 returns a recording control event; 4) Business module 3 returns AI interaction control events; 5) The multi-service triggering module synthesizes a unified control event (including all triggering service functions such as playback, recording, and AI) based on the control events returned by different service modules and sends it to the IMS call access module; 6) After the IMS call access module completes the SIP call setup and plays the audio, it generates a playback completion event and reports it to the multi-service triggering module. The multi-service triggering module then triggers the event only to service module 1. 7) After recording begins, a recording start event is generated and reported to the multi-service triggering module. The multi-service triggering module then triggers the event only to service module 2. 8) The IMS call access module, media module, and AI-ASR / NLP / TTS module complete the AI interaction process; 9) After the AI interaction is completed, an AI interaction completion event is generated and reported to the multi-business triggering module. The multi-business triggering module then triggers the event only to business module 3. 10) When the IMS call access module receives a network hang-up event, it reports the hang-up event to the multi-service triggering module; 11-13) The multi-service triggering module sends an end event to service module 1, service module 2, and service module 3 respectively.
[0044] This invention addresses the shortcomings of existing technologies, such as static prioritization, signaling redundancy, event obfuscation, high latency due to AI bypass, and lack of commercial support. It quantitatively reduces signaling load by 60%, call latency by 40%, and concurrent faults by 85%, significantly improving the efficiency of IMS multi-service collaboration and intelligent interaction. It can be widely applied to scenarios such as VoLTE, government and enterprise IP calls, and intelligent customer service for operators. The technical solution features deep coupling between hardware and algorithms, demonstrating outstanding creativity and licensing stability.
[0045] Note: AI (Artificial Intelligence): Artificial Intelligence IMS (IP Multimedia Subsystem): IP Multimedia Subsystem ASR (Automatic Speech Recognition): Automatic speech recognition NLP (Natural Language Processing) TTS (Text-To-Speech): Text to Speech SIP (Session Initiation Protocol): The protocol for initiating a session. CSCF (Serving-Call Session Control Function): Serving call session control function VoLTE (Voice over LTE): Long Term Evolution voice bearer (4G HD voice) GPU (Graphics Processing Unit): Graphics Processor Redis (Remote Dictionary Server): A remote dictionary server Kafka (Apache Kafka): Distributed event streaming platform / High-throughput message middleware InfluxDB: Time Series Database
Claims
1. An AI-based IMS call multi-service priority adaptive trigger feedback system, characterized in that include: IMS call access module, multi-service triggering module, service function module, media module, AI-ASR / NLP / TTS module; The IMS call access module is standardized and interfaces with the IMS core network S-CSCF. It establishes an IMS call session based on the SIP RFC3261 protocol, sends and receives SIP unified control messages, synchronizes the voice stream bidirectionally through the media module, and interacts with the AI-ASR / NLP / TTS module. The multi-service triggering module has a built-in scheduling processor, priority policy storage unit, message merging unit, and event-directed distribution unit. The scheduling processor is configured with an adaptive priority scheduling algorithm, which dynamically sorts and generates a priority trigger sequence based on service weight, user subscription service information, and service status. The priority calculation formula is Pi = Wi * Si * Ei, where Pi is the comprehensive priority score of the i-th service, Wi is the preset basic weight score of the i-th service with a value range of 0 to 1000, and the score interval for each service is 20, allowing new services to be inserted among existing services, Si is the subscription identifier of the current user for the i-th service with a value of subscribed = 1 and unsubscribed = 0, and Ei is the real-time activation status identifier of the i-th service with a value of activated = 1 and offline / fault = 0. The priority sequence is automatically refreshed every 500ms by default, and new services can be dynamically included in the priority sequence without restarting the system. The priority strategy storage unit stores the basic weights of services and user subscription information; the message merging unit receives the distributed control events issued by each service function module, removes redundant parameters and duplicate service identifier fields, saves the control events and service association information, and encapsulates them into a unified control message to be sent to the IMS call access module. The event-directed distribution unit receives media completion events, AI interaction completion events, and hang-up events reported by the IMS call access module, and only pushes the events to the business function modules with the corresponding business IDs, thereby achieving data isolation and non-interference of multiple business events. The business function module executes playback, recording, and AI voice interaction business logic based on the calling and called number rules. The media module collects two-way call voice streams and pushes them to the AI-ASR / NLP / TTS module, receives synthesized speech and completes real-time playback, and performs recording and media stream forwarding operations. The AI-ASR / NLP / TTS module is deployed on GPU computing nodes and includes an AI noise reduction ASR unit, an NLP semantic interaction unit, and an adaptive TTS synthesis unit. The AI noise reduction uses a three-level noise reduction algorithm to eliminate line current noise, environmental noise, and overlapping voices from multiple people, and outputs structured call text in real time. The NLP semantic interaction unit loads multi-scenario dedicated models and completes keyword recognition, multi-turn dialogue context caching, and response text generation based on the call text; the adaptive TTS synthesis unit supports multilingual and multi-dialect speech synthesis, dynamically adjusts timbre, speech rate, and volume according to the call scenario, and converts the response text into a speech stream to be sent back to the media module.
2. The system according to claim 1, characterized in that... When the scheduling processor triggers a service module according to priority, if the service module does not respond, it automatically skips to trigger the next service module. If the service module does not respond 3 times, the triggering is determined to be a failure and the service module is automatically downgraded. After receiving the unified control message, the IMS call access module replies with an ACK confirmation signal. If no ACK is received within the timeout period, it will retransmit up to 3 times.
3. The system according to claim 1, characterized in that... The event-directed distribution unit connects to the time-series database; when the business function module is offline, events are cached in the middleware, and undelivered events are re-pushed in batches after the module is restored online; all call events are stored persistently with timestamp, business ID, and result code; the system automatically pushes operation and maintenance alarms after recognizing abnormal events such as playback failure, AI recognition timeout, and recording interruption.
4. The system according to claim 1, characterized in that... The AI-ASR / NLP / TTS module is configured with a fault circuit interruption mechanism; when the GPU node goes offline or the ASR recognition timeout rate exceeds the threshold, the real-time NLP semantic interaction function is automatically shut down, while basic recording and fixed playback capabilities are retained, without affecting normal call functions; it supports hot model loading and A / B grayscale switching, and interactive model updates can be completed without interrupting online calls; it has a built-in session context cache, which caches the dialogue text throughout a single call to achieve multi-turn continuous semantic understanding, and automatically clears the cache to release computing power after the call is hung up.
5. The system according to claim 1, characterized in that: Each module is deployed in a cluster, and multiple service triggering modules share a distributed Redis cluster to store the call session state; the system can resume ongoing call sessions after a momentary restart, thus achieving session keep-alive; load balancing scheduling is performed on multiple GPU nodes, and computing resources are allocated according to call priority.
6. The system according to claim 1, characterized in that: The system has a built-in call record generation unit and a security encryption unit; the call record generation unit generates a complete call record at the call hang-up node; the security encryption unit adopts the national cryptographic standard SM4, desensitizes the calling and called numbers when storing logs, encrypts and stores the original voice messages and sets an automatic deletion period to meet communication privacy compliance requirements.
7. The system according to claim 1, characterized in that: The system is compatible with multiple IMS core network elements such as S-CSCF and I-CSCF; the media module is compatible with voice media streams and video IMS audio streams, and the AI-ASR / NLP / TTS module supports noise reduction and recognition of audio during video calls.
8. An AI-based IMS call multi-service priority adaptive triggering feedback method, characterized in that: Applied to the system of any one of claims 1 to 7, the method includes the following steps: (1) The IMS call access module receives the SIP call message issued by the S-CSCF, establishes an IMS call session based on the SIP protocol, extracts the calling and called numbers, call forwarding parameters, and call scenario identifiers to generate a call event and pushes it to the multi-service triggering module. (2) The multi-service triggering module queries the user's subscribed service list and uses an adaptive priority scheduling algorithm to calculate and generate an ordered priority triggering sequence; the system automatically refreshes the priority sequence every 500ms, and new services are included in the sorting in real time; (3) The multi-service triggering module receives the playback, recording, and AI interaction distributed control events returned by each service function module, removes redundant parameters and duplicate service identifiers, and encapsulates them into a single unified control message and sends it to the IMS call access module. (4) The IMS call access module completes SIP call setup based on unified control messages, and triggers the linkage media module and AI-ASR / NLP / TTS module to perform media forwarding and AI voice interaction in parallel according to the control messages; (5) The media module forwards the two-way call voice stream to the AI noise reduction ASR unit in real time, completes noise reduction recognition and outputs the call text; the NLP semantic interaction unit reads the cached multi-turn dialogue context to complete semantic parsing and generate response text; the adaptive TTS synthesis unit adjusts the timbre and speech rate according to the scenario to generate speech, which is played to both parties in real time through the media module; when the AI module fails, the real-time semantic interaction is turned off, while the basic services of recording and fixed playback are retained. (6) The IMS call access module completes the playback, off-hook and hang-up events, and reports the events with the service ID to the multi-service triggering module; the event-directed distribution unit of the multi-service triggering module pushes the events to the corresponding service function modules separately according to the service identifier; when the service module is offline, the events are cached in the middleware and resent in batches after recovery; all call events are stored persistently in full, and maintenance alarms are pushed synchronously when abnormal events are detected; (7) After the call is hung up, the IMS call access module reports the hang-up event, the multi-service triggering module sends the service termination instruction to all service function modules, and simultaneously clears the AI session context cache, releases the media stream, GPU computing power, and session storage resources; the system continuously identifies dead sessions through heartbeat detection and forcibly reclaims the resources they occupy.
