An embedded small base station AIAgent autonomous decision-making loop method and system
By employing a dual-channel parallel data acquisition and autonomous decision-making loop method, the problems of resource constraints and fault detection in the operation and maintenance of small base station equipment are solved, enabling efficient and accurate equipment health assessment and operation and maintenance decisions, and supporting offline autonomy and fault detection at the telecommunications service level.
Patent Information
- Application Number
- CN202611223088.2
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2026-08-13
- Publication Date
- 2026-11-10
- Estimated Expiration
- 2046-08-13
AI Technical Summary
Existing technologies for small cell equipment operation and maintenance suffer from high costs, low efficiency, lack of convergent perception capabilities, inability to detect faults at the telecommunications service level, and the inability of existing AI Agent solutions to run on resource-constrained embedded devices, resulting in ineffective or delayed operation and maintenance decisions.
The system employs a dual-channel parallel acquisition mechanism to obtain system resource indicators and telecommunications service indicators. It avoids resource contention by using staggered triggering, generates observation snapshots by combining anomaly detection and trend prediction, calculates a comprehensive health score, realizes autonomous decision-making loops, and supports offline autonomous operation under network interruption through circuit breaker protection, asynchronous execution, and policy adaptive optimization.
It enables efficient and accurate device health assessment and operation and maintenance decisions on resource-constrained devices, ensures real-time monitoring and optimization of wireless service capabilities, supports offline autonomy under network interruption, and avoids decision oscillation and waste of operation and maintenance resources.
Smart Images

Figure CN122742015B_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of mobile communication network operation and maintenance technology, specifically to an embedded small base station AI Agent autonomous decision-making loop method and system. Background Technology
[0002] With the rapid deployment of 5G communication networks and the increasing demand for deep coverage, small base stations, as an important supplement to macro base stations, are widely used in indoor scenarios and remote areas. In the small base station deployment architecture, EU devices manage multiple RU devices, which are deployed at remote, unattended sites. The number of sites is large and their distribution is scattered, making manual on-site maintenance costly and inefficient. Currently, introducing AI agents into the field of equipment operation and maintenance has become a technological trend. However, existing AI agents or reinforcement learning solutions based on large language models typically require GB-level memory and GPU computing power, while small base station EU / RU devices are based on ARM / RISC-V architecture, with a maximum of only 128MB of memory and a single-core or dual-core processor. Resources are extremely limited, and existing solutions cannot operate under such stringent constraints. In addition, existing operation and maintenance systems lack convergence perception capabilities when equipment status fluctuates, which can easily cause the decision-making system to switch frequently between multiple states due to short-term status fluctuations, resulting in repeated execution of unnecessary operation and maintenance operations. At the same time, small base station RU devices are deployed at remote sites without fixed network connections, and network outages occur frequently. Existing solutions that rely on cloud AI engines become completely ineffective when the network is interrupted, and the devices cannot obtain timely intelligent operation and maintenance decision support.
[0003] Most existing equipment health assessment solutions rely solely on system resource indicators such as CPU utilization, memory utilization, and temperature to determine equipment status. They fail to detect faults at the telecommunications service level, such as abnormal power amplifier VSWR (antenna feeder mismatch), decreased optical module optical power (fiber aging), and deteriorating RSRP / SINR (cell coverage issues). When system resource indicators are normal but telecommunications service indicators are abnormal, traditional health scores are still marked as good, failing to reflect the actual decline in the equipment's wireless service capabilities, leading to ineffective or delayed operational decisions. Summary of the Invention
[0004] The purpose of this invention is to provide an embedded small base station AI Agent autonomous decision-making loop method and system to solve the problems mentioned in the background art, such as high cost, low efficiency, lack of convergence perception capability, and inability to perceive faults at the telecommunications service level.
[0005] To achieve the above objectives, the present invention adopts the following technical solution: According to one aspect of the present invention, an embedded small base station AI agent autonomous decision-making loop method is provided, the method comprising: The system resource indicators and telecommunications service indicators of the embedded small base station equipment are obtained through a dual-channel parallel acquisition mechanism. The dual channels adopt a staggered triggering method to avoid resource competition. Each acquired data is assigned a quality label. After anomaly detection and trend prediction, an observation snapshot is generated and stored in the memory system. Historical context is extracted from the memory system, sub-health scores are calculated for system resource indicators and telecommunications service indicators respectively, and weighted and fused into a comprehensive health score, which is mapped to the device health level. For decision rules that are successfully matched and whose confidence assessment is not lower than a preset threshold, an action plan containing actions to be executed is generated. When the reasoning is abnormal, circuit breaker protection is triggered and the system is downgraded to the backup decision path. Execute each action in the action plan asynchronously according to priority and monitor the execution status and timeout. Perform predefined rollback operations on actions that fail to execute or time out and record the execution results. The weights of the triggered rules are updated based on the execution results to achieve adaptive optimization of the strategy. The evolution trajectory of the health level over multiple consecutive cycles is analyzed to identify convergence, oscillation, or divergence state patterns and the trigger sensitivity of subsequent rules is adjusted according to the identified state patterns. After completing this round of feedback, the system returns to the data collection step to enter the next decision-making cycle.
[0006] Based on the aforementioned scheme, in the dual-channel parallel acquisition mechanism, the first channel acquires system resource indicators at a fixed period, and the second channel acquires telecommunications service indicators through an adapter plug-in at a configurable period; a preset phase offset is maintained between the trigger times of the two channels; when the configuration period of the second channel conflicts with the fixed period of the first channel as an integer multiple, the offset is automatically increased for the configuration period of the second channel to avoid periodic overlapping triggering.
[0007] Based on the aforementioned scheme, the anomaly detection is based on calculating the Z-Score of each indicator using a sliding window. When the absolute value of the Z-Score exceeds a preset threshold, it is marked as an anomaly. The trend prediction is based on historical data within the sliding window, which is fitted with a slope through linear regression. The direction of the slope is used to determine whether the indicator is trending upward, downward, or stable.
[0008] Based on the aforementioned scheme, the system resource indicators include CPU utilization, memory utilization, storage utilization, and device temperature; the telecommunications service indicators include multiple items from the following: power amplifier VSWR, optical module transmit / receive power, CPRI link status, clock synchronization status, RSRP, and SINR; the system resource sub-health score is calculated by weighted average of the abnormal contribution of each indicator, and the telecommunications service sub-health score is calculated by the maximum value of the abnormal contribution of each indicator.
[0009] Based on the aforementioned scheme, the circuit breaker protection includes three states: closed, circuit breaker, and half-open. Under normal conditions, it is in the closed state. When the number of consecutive inference failures reaches a preset number, it switches to the circuit breaker state and is downgraded to the backup decision path. After a preset cooling time, the circuit breaker state switches to the half-open state. In the half-open state, one inference attempt is allowed. If successful, it returns to the closed state; if it fails, it switches back to the circuit breaker state.
[0010] Based on the aforementioned scheme, each action in the action plan has a priority level, and actions within the same priority are executed in order of action identifier; the predefined rollback operation is triggered when an action fails or times out, and the rollback operation itself does not trigger a secondary rollback.
[0011] Based on the aforementioned scheme, the weight update adopts an incremental update method with truncation, including: assigning different symbols and sizes of reward values according to whether the action is successfully executed, fails, or times out; multiplying the reward values by a preset learning rate and adding them to the current weight; and trunculating the update result to a preset numerical range.
[0012] Based on the aforementioned scheme, the identification method for the convergence, oscillation, or divergence state modes is as follows: when the health level is not worse than the downgrade level for a consecutive preset number of cycles and no new anomalies are generated, it is determined to be a convergence state; when the number of times the health level switches between two levels within a consecutive preset number of cycles reaches a preset value, it is determined to be an oscillation state; when the health level shows a continuous and progressively deteriorating trend within consecutive cycles, it is determined to be a divergence state; in the oscillation state, the rule trigger sensitivity is reduced, and in the divergence state, the rule trigger sensitivity and alarm level are increased.
[0013] Based on the aforementioned scheme, the observation snapshot includes the indicator value, quality identifier, anomaly detection result, trend prediction result, and timestamp of the current period; the memory system includes short-term memory for storing the most recent preset number of observation snapshots and long-term memory for persistently storing decision history and action execution records.
[0014] According to another aspect of the present invention, an embedded small base station AI agent autonomous decision-making loop system is provided, the system comprising: The acquisition module is used to acquire system resource indicators and telecommunications service indicators of embedded small base station equipment through a dual-channel parallel acquisition mechanism. The dual channels adopt a staggered triggering method to avoid resource competition. Each acquired data is assigned a quality label, and after anomaly detection and trend prediction, an observation snapshot is generated and stored in the memory system. The reasoning and decision-making module is used to extract historical context from the memory system, calculate sub-health scores for system resource indicators and telecommunications service indicators respectively, and weight and fuse them into a comprehensive health score, which is mapped to the device health level. For decision rules that are successfully matched and whose confidence assessment is not lower than a preset threshold, an action plan containing actions to be executed is generated. When the reasoning is abnormal, circuit breaker protection is triggered and the system is downgraded to the backup decision path. The action execution module is used to asynchronously execute each action in the action plan according to priority and monitor the execution status and timeout. It performs predefined rollback operations on actions that fail to execute or time out and records the execution results. The feedback optimization module is used to update the weights of the triggered rules based on the execution results to achieve adaptive optimization of the strategy, analyze the evolution trajectory of the health level over multiple consecutive cycles, identify convergence, oscillation or divergence state patterns and adjust the trigger sensitivity of subsequent rules according to the identified state patterns, and trigger the acquisition module to enter the next decision cycle after completing this round of feedback.
[0015] As can be seen from the above technical solution, compared with the prior art, the present invention has at least the following advantages and positive effects: The present invention adopts an event-driven architecture and thread pool, with extremely low system memory usage, fully adapting to the resource-constrained operating environment of embedded devices, overcoming the problem of excessive resource consumption in existing AI Agent solutions; through a dual-channel acquisition coordination mechanism and staggered scheduling, file read conflicts are effectively avoided, system overhead is extremely low, and it does not affect the main business operation of the device; the multi-dimensional health score is expanded from the traditional system resource dimension to a score system that integrates system resources and telecommunications services. The system resource indicators adopt a weighted average fusion, and the telecommunications service indicators adopt the most severe anomaly as the main factor, which can accurately perceive telecommunications service-level faults such as antenna mismatch, fiber optic aging, and coverage degradation, and truly reflect the wireless service capability of the device; a complete closed-loop decision-making cycle of perception, reasoning, execution, and feedback is realized, and the decision-making strategy is continuously self-optimized through a reward and punishment mechanism based on the execution result, with fast cycle response, meeting the real-time decision-making needs of embedded devices; a convergence analysis mechanism is introduced, through continuous stable judgment and state switching frequency detection... The system features a detection and progressive degradation identification mechanism that adaptively adjusts decision sensitivity under different state modes to effectively avoid decision oscillations. It incorporates a three-level circuit breaker protection system, automatically degrading to a backup decision path when inference anomalies occur, and recovering after cooling, ensuring uninterrupted core operation and maintenance capabilities during AI anomalies. It supports offline autonomous operation under network interruptions, batch synchronizing offline data after network recovery to ensure the complete traceability of all decision records during offline periods. It collects physical layer indicators of telecommunications services through a diverse adapter plug-in system and extends dedicated decision rules to achieve precise operation and maintenance tailored to the characteristics of telecommunications services. The rule engine is equipped with cooling and conflict resolution mechanisms and sets decision confidence thresholds to ensure the controllability and rationality of decisions. A systematic error code system supports end-to-end error tracing, and a hierarchical data retention strategy is adopted to prevent outdated data from interfering with current decisions.
[0016] It should be understood that the above general description and the following detailed description are exemplary and explanatory only, and are not intended to limit the invention. Attached Figure Description
[0017] The accompanying drawings, which are incorporated in and constitute a part of this specification, illustrate embodiments consistent with the invention and, together with the description, serve to explain the principles of the invention. It is obvious that the drawings described below are merely some embodiments of the invention, and those skilled in the art can obtain other drawings based on these drawings without any inventive effort. In the drawings: Figure 1 This is a schematic diagram of an embedded small base station AI Agent autonomous decision-making loop method according to the present invention; Figure 2 This is a flowchart illustrating the autonomous decision-making loop method for an embedded small base station AI Agent according to the present invention. Detailed Implementation
[0018] To more clearly illustrate the purpose, technical solutions, and advantages of the present invention, the technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. The exemplary embodiments can be implemented in many forms and should not be construed as limited to the examples set forth herein. On the contrary, these embodiments are provided so that the present invention will be more comprehensive and complete, and fully convey the concept of the exemplary embodiments to those skilled in the art.
[0019] Furthermore, the described features, structures, or characteristics can be combined in any suitable manner in one or more embodiments. Numerous specific details are provided in the following description to give a full understanding of embodiments of the invention. However, those skilled in the art will recognize that the technical solutions of the invention can be practiced without one or more of the specific details, or other methods, components, apparatuses, steps, etc., can be employed. In other instances, well-known methods, apparatuses, implementations, or operations are not shown or described in detail to avoid obscuring various aspects of the invention.
[0020] The block diagrams shown in the accompanying drawings are merely functional entities and do not necessarily correspond to physically independent entities. That is, these functional entities can be implemented in software, in one or more hardware modules or integrated circuits, or in different network and / or processor devices and / or microcontroller devices.
[0021] The flowcharts shown in the accompanying drawings are merely illustrative and do not necessarily include all content and operations / steps, nor do they necessarily have to be performed in the described order. For example, some operations / steps can be broken down, while others can be combined or partially combined; therefore, the actual execution order may change depending on the specific circumstances.
[0022] The present invention will now be described in detail with reference to specific embodiments.
[0023] Example 1
[0024] like Figure 1 , 2 As shown, this embodiment provides an embedded small base station AI Agent autonomous decision-making loop method. Figure 2 As shown, this method is driven by the engine state machine, and the state transition sequence is: IDLE→OBSERVE→REASON→ACT→FEEDBACK→IDLE; solid arrows indicate the state transition path under normal flow, and dashed arrows indicate the forced jump path triggered by timeout protection at each stage. The system operates as follows: IDLE (idle state) is a waiting node between decision-making cycles. In this state, the system maintains timer listening but does not perform any data collection or calculation operations. After the main loop timer is triggered, the state transitions to OBSERVE. OBSERVE is where step S1, data collection and perception operations, is performed, including data collection, anomaly detection, trend prediction, and observation snapshot generation. After completion, the system transitions to REASON. REASON is where step S2, reasoning and decision-making operations, is performed, including health scoring, rule matching, and action plan generation. After completion, the system transitions to ACT. ACT is where step S3, action execution operations, is performed, including asynchronous concurrent execution of actions, state monitoring, and rollback compensation. After completion, the system transitions to FEEDBACK. FEEDBACK is where step S4, feedback and optimization operations, is performed, including rule weight updates and convergence analysis. After completion, the system returns to the IDLE state, forming a complete decision-making cycle.
[0025] like Figure 1 As shown, the specific steps of this method are as follows: S1: The system resource indicators and telecommunications service indicators of the embedded small base station equipment are obtained through a dual-channel parallel acquisition mechanism. The dual channels adopt a staggered triggering method to avoid resource competition. Each acquired data is assigned a quality label. After anomaly detection and trend prediction, an observation snapshot is generated and stored in the memory system.
[0026] In this embodiment, the Observe phase is the starting point of the autonomous decision-making loop, responsible for collecting equipment operation data, performing quality control, executing anomaly detection and trend prediction, and generating standardized observation snapshots.
[0027] Two independent data acquisition channels are set up, including an AI real-time channel and a general acquisition channel, to achieve parallel acquisition of system resource indicators and telecommunications service indicators, while avoiding data source competition conflicts.
[0028] The AI real-time channel runs the agent_resource.c data collection module, triggering data collection at a fixed interval (preferably 5 seconds). It obtains system resource metrics, including CPU utilization, memory utilization, Flash storage utilization, and device core chip temperature, by reading the embedded Linux system's / proc file system (including / proc / stat, / proc / meminfo, etc.). The data collected by this channel is directly used by the Z-Score anomaly detection and ReAct decision engine.
[0029] The general data acquisition channel runs the libdaqdev.so dynamic library, and the acquisition period is configurable, preferably ranging from 5 seconds to 300 seconds (default 60 seconds). This channel acquires telecommunications service-specific metrics by loading various adapter plugins, including: the PMBUS adapter (via I... 2 (C bus reads PMBus register), used to collect voltage, current, power consumption, temperature, and VSWR of the power amplifier module; SFP_DDM adapter (via I 2 The C adapter (reads the SFP management interface) is used to collect the Tx / Rx optical power, bias current, and supply voltage of the optical module; the MODBUS adapter (executes the Modbus RTU protocol via RS485 bus) is used to collect the forward power, reverse power, gain, and temperature of the power amplifier; the HW_REG adapter (accesses hardware registers via MMIO) is used to collect the transceiver status, CPRI link status, fan speed, and clock synchronization status; the SW_COUNTER adapter (reads physical layer measurement data via shared memory SHM) is used to collect terminal-reported measurements such as RSRP, RSRQ, SINR, and CQI. These adapters together constitute the data collection source for telecommunications service indicators.
[0030] To avoid resource contention caused by two channels simultaneously accessing the / proc file system, a phase offset scheduling strategy is adopted. Based on the fixed period T=5 seconds of the AI real-time channel, its trigger time is assumed to satisfy t mod T=0; the trigger time of the general acquisition channel is offset by half a period, i.e., t mod T=2.5 seconds, ensuring that the two channels always operate at off-peak times. The system has a dynamic interval automatic adjustment function. If the configured interval of the general acquisition channel conflicts with the AI channel period as an integer multiple (i.e., the configured interval is equal to 5 seconds or an integer multiple thereof), the interval is automatically increased by a 2-second offset (e.g., adjusting 5 seconds to 7 seconds), and an INFO-level log is recorded to notify the operations and maintenance personnel.
[0031] For the parsing results of / proc / meminfo, a shared memory cache is used for optimization. After the AI real-time channel parses the data for the first time, it writes the memory usage result to shared memory and sets the time-to-live (TTL) to 1 second. The general acquisition channel reads from this shared cache first, and only performs actual file reading if the cache expires or does not exist. The read and write operations of shared memory are protected by a double-checked locking (DCL) mechanism to ensure data consistency in a multi-threaded environment.
[0032] Each collected data point is accompanied by a data quality code to identify the data's reliability. The quality codes are divided into four levels, with increasing priority: Level 0 (GOOD): High-quality data, participating in subsequent analysis with full weight (1.0); Level 1 (UNCERTAIN): Suspicious data; its confidence weight is reduced to 0.5 when included in the analysis. Level 2 (BAD): Invalid data, not included in anomaly detection and health scoring; Level 3 (MANUAL): Data is manually entered and used in decision-making with the highest priority.
[0033] The quality code is stored along with the indicator value in the observation snapshot and used as a factor to adjust the decision confidence level during the Reason phase.
[0034] For each valid metric, the system maintains a sliding window of length N=60, storing the 60 most recent historical sampled values. For the current sampled value x, the mean μ and sample standard deviation σ within the window are calculated using the following formula: ; Further, calculate the Z-Score: When the absolute value of Z-Score exceeds a preset threshold, it is marked as abnormal. In this embodiment, when |Z|>3.0, the indicator is determined to be in an abnormal state (anomaly_flag=true), otherwise it is normal.
[0035] Trend prediction is based on historical data within a sliding window, using linear regression to fit the slope. The direction of the slope determines whether the indicator is trending upward, downward, or stable. This involves analyzing the data points within the sliding window. A univariate linear regression model is performed on (i=1,...,N), and the model is as follows: Slope The calculation is performed using the least squares method, and the formula is: ; in , These are the time sequence number and the average of the indicator values, respectively. Set the trend judgment threshold. :like The trend is upward; if If the trend is positive, it indicates a decline; otherwise, it indicates stability. Anomaly markers and trend results are written as metadata to the observation snapshot.
[0036] All metric values, corresponding quality codes, anomaly markers, and trend results within the current collection period, along with the current timestamp and the Agent's own resource usage (memory MB, CPU percentage), are combined into a standardized Observation snapshot data structure. This snapshot reserves a comprehensive health score (health_score) field for later calculation in the Reason phase. This snapshot is written to short-term memory via a circular buffer with a capacity of the most recent 100 records, using a FIFO (First-In, First-Out) eviction strategy. The Observe phase has a 2-second timeout protection; after the timeout, it automatically forces entry into the next phase, ensuring that the decision loop does not stall due to collection blockage.
[0037] The memory system comprises two layers: short-term memory and long-term memory. Short-term memory is implemented using a circular buffer, storing a preset number (preferably 100) of recent observation snapshots, which are discarded according to a first-in, first-out (FIFO) strategy. This is used for real-time establishment of baselines for trend analysis and anomaly detection. Long-term memory is implemented using an SQLite relational database, persistently storing decision history, action execution results, rule weight evolution records, and convergence state analysis results. This is used for cross-period historical queries, experience accumulation, and context construction for decision optimization.
[0038] S2: Extract historical context from the memory system, calculate sub-health scores for system resource indicators and telecommunications service indicators respectively, and weight and fuse them into a comprehensive health score, which is mapped to the device health level. For decision rules that are successfully matched and whose confidence assessment is not lower than a preset threshold, generate an action plan containing actions to be executed. When the reasoning is abnormal, trigger circuit breaker protection and downgrade to the backup decision path.
[0039] The Reason phase is the core of the autonomous decision-making cycle. It involves in-depth analysis of the observation snapshots generated in the Observe phase to complete the quantitative assessment of health status, matching of decision rules, and generation of action plans.
[0040] Before proceeding with health scoring and decision reasoning, the context builder (memory_context_build()) is first invoked to extract historical information relevant to the current decision from the memory system. This builder supports multi-dimensional query conditions such as time range, event type, and device status. It extracts recent continuous observation sequences from short-term memory (a circular buffer containing the 100 most recent observation snapshots) and retrieves historical decision records, action execution effects, and abnormal event evolution trajectories under the same or similar device statuses from long-term memory (SQLite persistent storage). The constructed context data is injected into the reasoning process in a structured form, providing historical evidence for subsequent baseline references for health scoring and decision rule matching, ensuring the temporal coherence of the reasoning results.
[0041] The health score of the current device is calculated by combining the indicator data provided in the Observe phase, the anomaly detection results, and the historical references extracted by the context builder.
[0042] (1) Sub-score of system resource dimension ( ) The system resource dimension covers four general metrics: CPU utilization (weight 0.35), memory utilization (weight 0.30), storage utilization (weight 0.20), and device core temperature (weight 0.15). For each metric, the anomaly contribution is first calculated based on its Z-score. That is, when the Z-Score reaches or exceeds the anomaly threshold of 3.0, the contribution saturation is 1.0. The system resource sub-score is calculated using a weighted average method: ; The weights satisfy the following: This weighted average method reflects the complementary and superimposed characteristics of the abnormality levels among various indicators of system resources; the abnormality of any one indicator can be diluted by the partial normality of other indicators.
[0043] (2) Sub-scores for telecommunications services dimension ( ) The telecommunications service dimension categorizes specific indicators by equipment type, including physical layer indicators such as power amplifier VSWR, optical module Tx / Rx optical power, forward / reverse power, CPRI link status, clock synchronization status, RSRP, and SINR. For each collectable telecommunications service indicator, anomaly contribution is calculated based on its Z-score. The sub-score for telecommunications services is calculated using the maximum value method: ; Where m is the total number of telecommunications service indicators that the current device can actually collect; a serious abnormality in a single telecommunications service indicator is sufficient to cause a substantial loss of the device's wireless service capability, and the various telecommunications service indicators are not complementary.
[0044] (3) Overall health score (H) The two sub-scores mentioned above are weighted and fused using the following formula: ; The fusion coefficient is fixed at 1. , And satisfy The overall score H has a range of [0,1], with a larger value indicating a less healthy device.
[0045] (4) Health level mapping Based on the comprehensive health score H, the equipment health status is divided into four discrete levels, with the mapping rule as follows: When H < 0.3, the health level is GOOD. When 0.3≤H<0.5, the health level is DEGRADED. When 0.5 ≤ H < 0.7, the health grade is POOR (poor). When H ≥ 0.7, the health level is CRITICAL.
[0046] Before rule matching, a comprehensive confidence assessment of the quality of the current observation data is performed. The assessment is based on: the distribution of data quality codes for each indicator in the observation snapshot; a higher proportion of GOOD data results in a higher base confidence score, while UNCERTAIN data is weighted at 0.5. Simultaneously, data freshness is checked; if the difference between the timestamp of any indicator and the current system time exceeds a preset validity period (preferably 30 seconds), this outdated data is given a lower weight in the confidence calculation. The final output confidence score is a value between 0 and 1, serving as the admission condition for rule triggering. The Reason phase has a 3-second timeout protection; after the timeout, automatic degradation processing occurs.
[0047] The system loads the decision rule configuration file (decision_rules.json), which defines the complete rule set in JSON Schema format. Each decision rule contains the following elements: a unique rule identifier (rule_id), a trigger condition, an action, a priority (divided into five levels from P1 to P5, with P1 being the highest), a cooldown duration (cooldown_s), and a confidence threshold (threshold). Further, rule matching is performed. Based on the current health level, the set of abnormal indicators (indicators marked as true in anomaly_flags), and the trend direction, the rule set is traversed for condition evaluation. For rules that meet the conditions, their cooldown status is further checked. The difference between the current time and the rule's most recent trigger time is calculated. If the difference is greater than the configured cooldown duration, the rule's cooldown restriction is lifted. Otherwise, it will not be triggered during the cooldown period. The current inference confidence is compared with the confidence threshold of the rule. The rule is confirmed as a triggerable rule only if the confidence is not lower than the threshold (default threshold 0.7). When multiple rules meet the triggering conditions at the same time, according to the conflict resolution strategy, they are first sorted by priority from high to low (P1 is the highest), and rules of the same priority are arranged in lexicographical order of rule identifiers.
[0048] The rule system comprises at least six fundamental core rules, covering typical operational scenarios such as severe CPU anomalies, severe memory anomalies, storage anomalies, temperature anomalies, trend prediction anomalies, and fallback diagnostics for moderate anomalies. Furthermore, six dedicated rules are added for telecommunications service metrics, including: power amplifier VSWR anomalies triggering PA protection power reduction and alarms; optical module Tx optical power decline trends triggering fiber optic link checks and early warnings; optical module Rx optical power anomalies triggering link diagnostics; RSRP and SINR signal quality degradation triggering output power adjustments and alarms; CPRI link status anomalies triggering link switching and EU notifications; and clock synchronization loss triggering EU notifications and diagnostics. The telecommunications service rules and system resource rules share the same cooling mechanism, conflict resolution strategy, and confidence assessment system.
[0049] Once the matching is complete, the triggerable rules are converted into a structured action plan, which contains at least one action item to be executed, and is output to the Act stage along with the reasoning results.
[0050] To prevent the AI inference module from exhausting system resources due to consecutive exceptions, the Reason phase is built into a three-state circuit breaker. The circuit breaker has three states: CLOSED state: AI inference works normally. The consecutive failure counter is incremented each time inference fails and the counter is reset to zero when it succeeds. When the number of consecutive failures reaches a preset threshold (preferably 3 times), the fuse state switches to OPEN.
[0051] OPEN (Circuit Breaker) State: Skips the AI inference process and directly calls the L1 Lua rule engine (a deterministic rule engine based on Lua 5.4.6 scripts) to generate an action plan, ensuring uninterrupted basic operation and maintenance capabilities. After the circuit breaker enters the OPEN state, a cooldown timer is started, with a preferred cooldown period of 300 seconds. After the cooldown period expires, it automatically switches to the HALF_OPEN state.
[0052] HALF_OPEN state: Allows one AI inference probe request. If the probe inference is successful, the fuse returns to the CLOSED state and the consecutive failure counter is reset to zero; if the probe inference fails, the fuse switches back to the OPEN state and the cooldown timer is reset.
[0053] The circuit breaker maintains a continuous failure counter in the CLOSED state. This counter is reset to zero each time AI inference succeeds and incremented by 1 each time inference fails. When the counter value reaches a preset failure threshold (preferably 3 times), the circuit breaker switches to the OPEN state. This three-state switching allows for smooth degradation in case of AI inference anomalies, ensuring the continuous availability of the decision-making chain. The overall timeout for the Reason phase is set to 3 seconds. If the circuit breaker directly calls the Lua rule engine while in the OPEN state, this timeout still applies.
[0054] S3: Execute each action in the action plan asynchronously according to priority and monitor the execution status and timeout. Perform predefined rollback operations on actions that fail or time out and record the execution results.
[0055] The Act (Execution) phase transforms the action plan generated in the Reason phase into specific device operations and executes them. This phase adopts an asynchronous concurrent execution architecture, combined with priority scheduling, timeout monitoring, and rollback mechanisms to ensure the efficiency and reliability of action execution.
[0056] Upon receiving the action plan output from the Reason phase, the first step is to perform validity verification and structured parsing. The action plan includes at least: a unique action identifier (action_id), an action type identifier (type, such as degrade, alert, adjust_power, etc.), an execution priority (priority, ranging from P1 to P5, with P1 being the highest), the target device or module (target), action parameters (params, carried in key-value pairs, such as target adjustment value, execution duration, etc.), a timeout (timeout_ms, in milliseconds), a list of preconditions (preconditions, listing other action IDs that must be completed first), and a rollback action description (rollback_action, used for compensation operations in case of execution failure). The action entries in the action plan are grouped by priority, with P1 priority actions placed in the highest priority queue and P5 actions placed in the lowest priority queue. Within the same priority level, actions are arranged lexicographically by action ID. After parsing, the action queue is submitted to the action execution thread pool for scheduling and execution.
[0057] A pre-defined fixed-size thread pool (preferably 4 threads) is used to asynchronously execute each action item in the action plan, avoiding serial execution that could block the main decision loop. The thread pool employs a work-stealing or fixed-queue allocation pattern, retrieving an action from the highest-priority non-empty queue and assigning an idle thread to execute it. For actions with preconditions (not null in the preconditions field), the thread pool checks before execution whether all dependent preconditions have been successfully completed. If a precondition has not been executed or has failed, the action is temporarily deferred and re-queued until the preconditions are met or the dependency times out.
[0058] Each action has a defined lifecycle state during execution, including PENDING, RUNNING, SUCCESS, FAILED, and TIMEOUT. After retrieving an action from the queue, the worker thread in the thread pool sets the action's state to RUNNING and calls the execution function bound to the action type. Simultaneously, an independent timeout monitoring timer is started for each action. The timer duration is specified by the `timeout_ms` field in the action plan (the default value can be preset according to the action type, such as 3000ms for alarm sending and 5000ms for power adjustment). The timeout monitoring timer is implemented using the embedded Linux `timerfd` mechanism, enabling precise timing without consuming additional CPU polling resources. If the action returns an execution result (success or failure) within `timeout_ms`, the corresponding timeout timer is canceled, the actual execution time (`execution_time_ms`) is recorded, and the state is updated. If the timer triggers before the execution result, the worker thread immediately stops the execution of the action (by setting a thread cancellation flag or calling the action's built-in termination interface), marks the action's state as TIMEOUT, and records the corresponding error code.
[0059] Action execution results are uniformly encapsulated into a standardized ActionResult data structure, including action ID (action_id), execution status (status, values SUCCESS / FAILED / TIMEOUT), actual execution time in milliseconds (execution_time_ms), error code (error_code, non-zero values indicate specific error types), standard output information (output, recording key diagnostic information during execution), and side effect descriptions (side_effects, recording observable changes to the system state after the action's execution, such as a comparison of values before and after configuration changes). ActionResult is collected immediately after the action is completed.
[0060] For actions carrying the `rollback_action` field, if the execution result status is FAILED or TIMEOUT, a rollback compensation process is automatically triggered. The rollback action is constructed based on the predefined `rollback_action` description in the action plan, including the type of rollback operation, target parameters, and rollback execution conditions. The rollback action is submitted to the thread pool for execution, with a default timeout of half the original action's `timeout_ms`. The rollback action itself does not trigger secondary rollbacks to prevent recursive compensation from causing system instability. The rollback execution result is also recorded as part of the `ActionResult`, marked with the "ROLLBACK_EXECUTED" flag in the `side_effects` field for use in subsequent Feedback phases and convergence analysis.
[0061] The Act phase has a 10-second timeout protection. Upon entering the Act phase, the system starts a phase-level timer. If the thread pool fails to complete the execution of all triggered actions in the current action plan and necessary rollback compensation within 10 seconds, the phase-level timer triggers, forcibly terminating all still-executing action threads, recording the timeout event in long-term memory, and marking the final state of incomplete actions as TIMEOUT. After the timeout, regardless of the actual execution status of the actions, the system automatically proceeds to the next phase (Feedback phase), ensuring that the decision loop is not blocked for an extended period by individual time-consuming actions.
[0062] S4: Update the weights of the triggered rules based on the execution results to achieve adaptive optimization of the strategy, analyze the evolution trajectory of the health level over multiple consecutive cycles, identify convergence, oscillation or divergence state patterns and adjust the trigger sensitivity of subsequent rules according to the identified state patterns; after completing this round of feedback, return to the data collection step to enter the next decision loop.
[0063] The Feedback phase is a closed loop in the autonomous decision-making cycle. It collects the results of action execution, updates the weights of decision rules to achieve continuous strategy optimization, and performs convergence analysis to prevent decision oscillations.
[0064] After the Act phase, ActionResults for all executed actions are collected from the thread pool and the execution context of each action. Each ActionResult includes: action ID (action_id), execution status (status, values are SUCCESS, FAILED, or TIMEOUT), actual execution time (execution_time_ms), error code (error_code), output information, and side effect descriptions (side_effects). The collected ActionResults are summarized and statistically analyzed to calculate the success rate, average execution time, and failure distribution of actions in the current ActionPlan, generating a structured feedback data packet. This feedback data packet is persistently stored in long-term memory (SQLite database) for subsequent historical queries and experience accumulation; and key fields (especially the execution status of each action) are extracted as input parameters for rule weight adjustment and convergence analysis.
[0065] Based on the results of action execution, the weights of each triggered rule are incrementally updated to achieve adaptive optimization of the strategy.
[0066] For rule k, let its current weight be... The reward value r is determined based on the execution result of the action generated by this rule: If the action execution status is SUCCESS, then the reward value r = 1 (positive reward). If the execution status is FAILED, the reward value r = -0.5 (negative reward); If the execution status is TIMEOUT, the reward value r = -1 (the strongest negative reward, because timeout behavior causes more severe damage to system responsiveness).
[0067] The weight update uses a truncated linear increment rule, as shown in the following formula: ; in This embodiment uses a preset learning rate to control the step size of each update. ; The truncation function restricts the update results to the range [0,1]. This rule ensures that the weight update process is bounded and stable, and will not cause weight overflow due to continuous positive rewards or weight underflow due to continuous negative rewards.
[0068] When the same rule generates multiple actions in a single loop (e.g., simultaneously sending an alert and performing a degradation), the lowest execution status among all actions under that rule is used as the aggregated result for weight updates. Specifically, if any action is TIMEOUT, the update is based on TIMEOUT; if there is no TIMEOUT, any action FAILED is updated based on FAILED; and if all actions SUCCESS, the update is based on SUCCESS. This low-value strategy ensures that the weights are sensitive to the worst-case performance of the rule, prioritizing the avoidance of high-risk behaviors.
[0069] To prevent repeated decisions and resource waste caused by short-term state fluctuations in the system, convergence analysis is performed to identify patterns in the historical evolution trend of health scores and dynamically adjust the sensitivity of subsequent decisions.
[0070] (1) Convergence determination Check the most recent consecutive The health level sequence within a decision cycle. If the health level for all three cycles is GOOD or DEGRADED (i.e., not worse than DEGRADED), and no new anomalies are generated during this period (i.e., the anomaly label array). If all values in each period are 0, the system is considered to have entered a convergent state. Convergence indicates that the equipment's operating state has stabilized and the system can regain normal decision-making sensitivity.
[0071] (2) Oscillation detection Check the most recent consecutive The health level sequence within a decision cycle. The number of transitions between adjacent cycles in this sequence is counted, calculated using the following formula: ; in For indicator functions. When If the number of different health levels appearing in the sequence is exactly 2, the system is determined to be in an oscillation state. After oscillation detection is triggered, the sensitivity of decision rule triggering in subsequent cycles is automatically reduced to avoid unnecessary repeated maintenance operations due to frequent switching between the two levels.
[0072] (3) Diverging detection Examine the health level changes over the most recent three consecutive periods. If observed... , , If the health level deteriorates sequentially over three consecutive cycles (DEGRADED→POOR→CRITICAL), the system is considered to be in a diverging state. Once divergence detection is triggered, the triggering priority and alarm level of subsequent decision-making rules are increased, and it is recommended that operations and maintenance personnel intervene manually in a timely manner.
[0073] Based on the above three detection results, the current convergence state is determined according to the following priority (divergence takes precedence over oscillation, and oscillation takes precedence over convergence): If the divergence detection condition is met, then the convergence state is DIVERGING. Otherwise, if the oscillation detection condition is met, then the convergence state = OSCILLATING; Otherwise, if the convergence condition is met, then the convergence state is CONVERGED. Otherwise, the convergence state is CONVERGING (indicating that it is in the process of convergence, has not reached stability but has no signs of abnormal divergence or oscillation).
[0074] The convergence state is stored as metadata in long-term memory and serves as a dynamic adjustment factor in the next Reason phase, affecting the sensitivity threshold for rule triggering. For example, in the OSCILLATING state, the confidence threshold can be temporarily increased from the default value of 0.7 to 0.8 to suppress the triggering of low-confidence decisions; when identified as the DIVERGING state, the system maintains or decreases the confidence threshold to improve the sensitivity of rule triggering and automatically raises the priority of alarm actions to level P1; when the system returns to the CONVERGED state, the confidence threshold returns to the default value of 0.7.
[0075] All data generated during the Feedback phase, including ActionResult, rule weight update records, and convergence status analysis results, is structured, serialized, and written to long-term memory. The lifecycle of this data is automatically managed according to a tiered data retention strategy: Agent heartbeat records are retained for 14 days, operational logs for 30 days, AI analysis reports for 30 days, alarm history for 90 days, and configuration changes for the last 50 versions. Historical data exceeding the retention period is no longer included in the context construction of subsequent Reason phases, ensuring that decision inputs are not interfered with by outdated data.
[0076] The Feedback phase has a 2-second timeout protection. The system starts a phase-level timer upon entering the Feedback phase. If the entire process of ActionResult collection, rule weight adjustment, and convergence analysis is not completed within 2 seconds, the phase-level timer is triggered, immediately terminating the remaining operations of the current Feedback phase. Completed data persistence is still maintained, while unfinished analysis work (such as convergence calculations) is skipped in this cycle, automatically advancing to the next IDLE state and starting a new Observe cycle. This timeout protection mechanism ensures that the feedback loop will not occupy the decision loop time for an extended period due to abnormal situations (such as SQLite write blocking), maintaining the overall real-time responsiveness of the system.
[0077] After the Feedback phase ends, all state data of the current loop (including updated rule weights, convergence status, and the latest Observation snapshot) is stored in the memory system. Then, the engine state machine automatically switches from the FEEDBACK state back to the IDLE state and starts the next Observe phase according to the preset main loop cycle (baseline 30 seconds, which can be dynamically adjusted to 10 to 60 seconds according to device load). This cycle repeats to achieve continuous autonomous operation and maintenance decision-making for embedded small base station devices in unattended scenarios. When the system receives an external stop signal (such as SIGTERM or SIGINT), the engine state machine executes a graceful shutdown process after the Feedback phase of the current loop is completed, releasing all thread pool resources, closing database connections, saving the current state snapshot, and then exiting.
[0078] Example 2 This embodiment exemplifies an embedded small base station AI Agent autonomous decision-making loop system. This embodiment uses an RU device (ARM Cortex-A7 dual-core processor, 128MB memory, running embedded Linux) as the deployment platform for illustration. The system runs in the form of a daemon process and includes four core parts: a data acquisition module, an inference decision-making module, an action execution module, and a feedback optimization module. The modules interact with each other through shared memory and message queues.
[0079] The system's main loop is driven by a timer, with a default base period of 30 seconds, which can be dynamically adjusted to 10 to 60 seconds based on device load. After startup, the system operates in the IDLE→OBSERVE→REASON→ACT→FEEDBACK→IDLE state machine sequence. Modules communicate using four standardized data structures: Observation snapshots, ActionPlans, ActionResults, and feedback data packets, ensuring clear interfaces and low coupling between modules. The system employs an epoll event-driven architecture, combining a timerfd timer and an eventfd event notification mechanism to achieve efficient event scheduling and avoid CPU waste caused by polling. The overall system memory usage is kept below 10MB, meeting the resource constraints of embedded devices.
[0080] The acquisition module is responsible for performing the functions of the Observe phase. This module has built-in AI real-time channel submodule and general acquisition channel submodule. The AI real-time channel reads the / proc file system at a fixed period (preferably 5 seconds) to collect four-dimensional system resource indicators: CPU usage, memory usage, storage usage, and device temperature. The general acquisition channel collects telecommunications service indicators such as power amplifier VSWR, optical module transmit and receive power, CPRI link status, clock synchronization status, RSRP, and SINR at a configurable period (default 60 seconds) by loading adapter plugins such as PMBUS, SFP_DDM, MODBUS, HW_REG, and SW_COUNTER.
[0081] The two channels employ phase-off scheduling: the AI real-time channel triggers at t mod 5 = 0, and the general acquisition channel triggers at t mod 5 = 2.5 (i.e., offset by half a cycle), avoiding I / O contention caused by simultaneous access to the / proc file system. When the configuration cycle of the general channel conflicts with the cycle of the AI channel due to an integer multiple, the system automatically adds a 2-second offset to avoid this. For the parsing results of / proc / meminfo, the system uses a shared memory cache (TTL = 1 second, DCL protected) to achieve multiplexing between the two channels, reducing the overhead of repeated parsing.
[0082] The data acquisition module assigns a data quality code (4 levels: GOOD / UNCERTAIN / BAD / MANUAL) to each collected data point for subsequent confidence assessment. After data acquisition, the module calculates the Z-Score for each indicator using a sliding window (N=60) for anomaly detection (|Z|>3.0 is marked as anomaly), and fits the trend direction (rising / falling / stable) using univariate linear regression. Finally, the indicator values, quality codes, anomaly markers, and trend results are encapsulated as observation snapshots, written to a short-term memory circular buffer (100 records, FIFO eviction), and persisted to long-term memory (SQLite database). The acquisition module has a 2-second timeout protection.
[0083] The reasoning and decision-making module is responsible for executing the functions of the Reason phase. This module consists of four sub-modules: a context builder, a multidimensional health scorer, a decision rule engine, and a circuit breaker.
[0084] The context builder extracts the 60 most recent observation sequences from short-term memory and retrieves historical decision records of the same or similar device states in the last 7 days from long-term memory to generate a structured context for subsequent reasoning.
[0085] The multidimensional health scorer calculates the comprehensive health score according to a predetermined formula. The system resource sub-score H_sys is a weighted average of the abnormal contribution of four indicators (CPU weight 0.35, memory weight 0.30, storage weight 0.20, and temperature weight 0.15). The telecommunications service sub-score H_tel is the maximum value of the abnormal contribution of each telecommunications service indicator (VSWR, optical power, CPRI status, RSRP, SINR, etc.). The comprehensive health score H = 0.6 × H_sys + 0.4 × H_tel, and is mapped to four health levels according to the H value: GOOD (H < 0.3), DEGRADED (0.3 ≤ H < 0.5), POOR (0.5 ≤ H < 0.7), and CRITICAL (H ≥ 0.7).
[0086] The decision rule engine loads the rule set defined in decision_rules.json. During rule matching, it checks the triggering conditions, cooldown status (no triggering during the cooldown period), and confidence threshold (default 0.7) sequentially. When multiple rules trigger simultaneously, they are executed in order of priority (P1 is the highest, and rules with the same priority are executed in lexicographical order by rule_id). The rule system includes 6 basic rules (covering CPU / memory / storage / temperature anomalies, trend warnings, and fallback diagnostics) and 6 telecommunications service-specific rules (covering scenarios such as VSWR anomalies, optical power degradation, signal quality deterioration, CPRI interruptions, and clock lockout). Successfully matched rules are converted into structured ActionPlan output by the action plan generator.
[0087] The fuse has three states: CLOSED, OPEN, and HALF_OPEN. In the CLOSED state, after three consecutive inference failures, it switches to OPEN. In the OPEN state, it degrades to the L1 Lua rule engine and initiates a 300-second cooldown. After cooling, it enters HALF_OPEN for detection and recovery. The inference decision module has a 3-second timeout protection.
[0088] The action execution module is responsible for performing the functions of the Act phase. This module includes an action queue manager, an execution thread pool, and an action monitor.
[0089] After receiving an ActionPlan, the Action Queue Manager assigns each action to its corresponding priority queue according to priority (P1 to P5), and within the same priority, actions are arranged in lexicographical order by action ID. For actions with prerequisite dependencies, the manager checks whether the prerequisite actions have been successfully completed before execution; otherwise, scheduling is postponed.
[0090] The execution thread pool has four pre-configured worker threads, which use a work-stealing mode to retrieve actions asynchronously from the highest-priority non-empty queue. The execution functions bound to each action type include, but are not limited to, send_alarm, degrade, adjust_collection_freq, storage_clean, pa_protect, fiber_inspect, adjust_power, switch_channel, and notify_eu, which correspond to maintenance operations such as alarm sending, degradation, collection cycle adjustment, storage cleanup, PA protection, fiber diagnostics, power adjustment, link switching, and notification EU, respectively.
[0091] The action monitor starts an independent timerfd for each action to monitor execution status and timeouts. The action execution result is encapsulated as an ActionResult (containing action_id, status, execution_time_ms, error_code, output, and side_effects). For actions that fail or time out, if rollback_action is defined in the ActionPlan, a rollback operation is automatically triggered; the rollback operation itself does not trigger secondary rollbacks. The entire Act phase has a 10-second timeout protection.
[0092] The Feedback Optimization module is responsible for executing the functions of the Feedback phase. This module includes a result collector, a rule weight updater, and a convergence analyzer.
[0093] The results collector aggregates all ActionResults, calculates the success rate, average execution time, and failure distribution of actions, and generates feedback data packets which are written to long-term memory.
[0094] The rule weight updater updates the weights of triggered rules based on the action execution results. The reward value is r=1 for SUCCESS, r=-0.5 for FAILED, and r=-1 for TIMEOUT. The update formula is w'=clip(w+0.1×r, 0, 1). When multiple actions are executed for the same rule, the action with the lowest execution status participates in the update. The updated weights are persisted to an SQLite database.
[0095] The convergence analyzer detects the convergence status. A condition is classified as CONVERGED if the health level is not worse than DEGRADED for three consecutive cycles and no new anomalies appear; OSCILLATING if the health level switches between the two states ≥2 times within five consecutive cycles; and DIVERGING if the health level progressively deteriorates from DEGRADED to POOR to CRITICAL for three consecutive cycles. In the OSCILLATING state, the confidence threshold is increased from 0.7 to 0.8 to suppress frequent decisions; in the DIVERGING state, the threshold is decreased and the alarm priority is increased; and in the CONVERGED state, the threshold is restored to 0.7. The feedback optimization module has a 2-second timeout protection.
[0096] After completing the feedback, the system automatically returns to the acquisition module to enter the next decision loop, until a stop signal is received and a graceful shutdown is performed.
[0097] In addition to local deployment on RU devices, this system also supports the deployment of centralized AI Agent instances on the EU device side. The EU-side system receives observation data reported by each RU device via the MQTT protocol and performs inference decisions independently on each RU device in a cascading manner. The EU side provides a REST API for operations personnel to query via the Mongoose HTTP server, and implements LLM streaming response push and Chart.js front-end visualization through WebSocket / SSE. Decision data generated during RU offline operation is batch synchronized to the EU side via three types of frames—HEARTBEAT_HISTORY, ALARM_HISTORY, and OPS_LOG_SYNC—after network recovery, ensuring data traceability during offline periods.
[0098] The key parameters of this system are preset in the configuration file and can be hot-loaded and adjusted. These mainly include: main loop cycle 30 seconds (dynamically adaptive 10-60 seconds), AI real-time channel cycle 5 seconds, general acquisition channel cycle 60 seconds (configurable from 5-300 seconds), dual-channel phase offset 2.5 seconds, sliding window size 60, Z-Score anomaly threshold 3.0, trend judgment threshold 0.05, circuit breaker failure threshold 3 times, circuit breaker cooling 300 seconds, learning rate 0.1, convergence cycle number 3, oscillation cycle number 5, default confidence threshold 0.7, short-term memory capacity 100 records, action thread pool 4, timeout for each stage (Observe 2 seconds / Reason 3 seconds / Act 10 seconds / Feedback 2 seconds), data freshness threshold 30 seconds, and L1 hit rate alarm threshold 70%.
[0099] Other embodiments of the invention will readily occur to those skilled in the art upon consideration of the specification and practice of the invention disclosed herein. This application is intended to cover any variations, uses, or adaptations of the invention that follow the general principles of the invention and include common knowledge or customary techniques in the art not disclosed herein. The specification and examples are to be considered exemplary only, and the true scope and spirit of the invention are indicated by the claims. It should be understood that the invention is not limited to the precise structures described above and shown in the drawings, and various modifications and changes can be made without departing from its scope. The scope of the invention is limited only by the appended claims.
Claims
1. A method for autonomous decision-making loop of an embedded small base station AI agent, characterized in that, The method includes: The system resource indicators and telecommunications service indicators of the embedded small base station equipment are obtained through a dual-channel parallel acquisition mechanism. The dual channels adopt a staggered triggering method to avoid resource competition. Each acquired data is assigned a quality label. After anomaly detection and trend prediction, an observation snapshot is generated and stored in the memory system. Historical context is extracted from the memory system, sub-health scores are calculated for system resource indicators and telecommunications service indicators respectively, and weighted and fused into a comprehensive health score, which is mapped to the device health level. For decision rules that are successfully matched and whose confidence assessment is not lower than a preset threshold, an action plan containing actions to be executed is generated. When the reasoning is abnormal, circuit breaker protection is triggered and the system is downgraded to the backup decision path. Execute each action in the action plan asynchronously according to priority and monitor the execution status and timeout. Perform predefined rollback operations on actions that fail to execute or time out and record the execution results. The weights of the triggered rules are updated based on the execution results to achieve adaptive optimization of the strategy. The evolution trajectory of the health level over multiple consecutive cycles is analyzed to identify convergence, oscillation, or divergence state patterns and the trigger sensitivity of subsequent rules is adjusted according to the identified state patterns. After completing this round of feedback, the system returns to the data collection step to enter the next decision-making cycle.
2. The embedded small base station AI Agent autonomous decision-making loop method according to claim 1, characterized in that, In the dual-channel parallel acquisition mechanism, the first channel acquires system resource indicators at a fixed period, and the second channel acquires telecommunications service indicators through an adapter plug-in at a configurable period; a preset phase offset is maintained between the trigger times of the two channels; when the configuration period of the second channel conflicts with the fixed period of the first channel as an integer multiple, the offset is automatically increased for the configuration period of the second channel to avoid periodic overlapping triggering.
3. The embedded small base station AI Agent autonomous decision-making loop method according to claim 1, characterized in that, The anomaly detection is based on calculating the Z-Score of each indicator using a sliding window. When the absolute value of the Z-Score exceeds a preset threshold, it is marked as an anomaly. The trend prediction is based on historical data within the sliding window, which is fitted with a slope through linear regression. The direction of the slope determines whether the indicator is trending upward, downward, or stable.
4. The embedded small base station AI Agent autonomous decision-making loop method according to claim 1, characterized in that, The system resource indicators include CPU utilization, memory utilization, storage utilization, and device temperature; the telecommunications service indicators include multiple indicators such as power amplifier VSWR, optical module transmit / receive power, CPRI link status, clock synchronization status, RSRP, and SINR; the system resource sub-health score is calculated by weighted average of the abnormal contribution of each indicator, and the telecommunications service sub-health score is calculated by the maximum value of the abnormal contribution of each indicator.
5. The embedded small base station AI Agent autonomous decision-making loop method according to claim 1, characterized in that, The circuit breaker protection includes three states: closed, circuit breaker, and half-open. Under normal conditions, it is in the closed state. When the number of consecutive inference failures reaches a preset number, it switches to the circuit breaker state and is downgraded to the backup decision path. After a preset cooling time, the circuit breaker state switches to the half-open state. In the half-open state, one inference attempt is allowed. If successful, it returns to the closed state. If it fails, it switches back to the circuit breaker state.
6. The embedded small base station AI Agent autonomous decision-making loop method according to claim 1, characterized in that, Each action in the action plan has a priority level, and actions within the same priority are executed in order of action identifier; the predefined rollback operation is triggered when an action fails or times out, and the rollback operation itself does not trigger a secondary rollback.
7. The embedded small base station AI Agent autonomous decision-making loop method according to claim 1, characterized in that, The weight update adopts an incremental update method with truncation, which includes: assigning different symbols and sizes of reward values according to whether the action is executed successfully, failed, or timed out; multiplying the reward values by a preset learning rate and adding them to the current weight; and trunculating the update result to a preset numerical range.
8. The embedded small base station AI Agent autonomous decision-making loop method according to claim 1, characterized in that, The identification method for the convergence, oscillation, or divergence state modes is as follows: when the health level is not worse than the downgrade level for a consecutive preset number of cycles and no new anomalies are generated, it is determined to be a convergence state; when the number of times the health level switches between two levels within a consecutive preset number of cycles reaches a preset value, it is determined to be an oscillation state; when the health level shows a continuous and progressively deteriorating trend within consecutive cycles, it is determined to be a divergence state; in the oscillation state, the rule trigger sensitivity is reduced, and in the divergence state, the rule trigger sensitivity and alarm level are increased.
9. The embedded small base station AI Agent autonomous decision-making loop method according to claim 1, characterized in that, The observation snapshot includes the indicator value, quality identifier, anomaly detection result, trend prediction result and timestamp of the current period; the memory system includes short-term memory for storing the most recent preset number of observation snapshots and long-term memory for persistently storing decision history and action execution records.
10. An embedded small base station AI Agent autonomous decision-making loop system, used to implement the method as described in any one of claims 1-9, characterized in that, include: The acquisition module is used to acquire system resource indicators and telecommunications service indicators of embedded small base station equipment through a dual-channel parallel acquisition mechanism. The dual channels adopt a staggered triggering method to avoid resource competition. Each acquired data is assigned a quality label, and after anomaly detection and trend prediction, an observation snapshot is generated and stored in the memory system. The reasoning and decision-making module is used to extract historical context from the memory system, calculate sub-health scores for system resource indicators and telecommunications service indicators respectively, and weight and fuse them into a comprehensive health score, which is mapped to the device health level. For decision rules that are successfully matched and whose confidence assessment is not lower than a preset threshold, an action plan containing actions to be executed is generated. When the reasoning is abnormal, circuit breaker protection is triggered and the system is downgraded to the backup decision path. The action execution module is used to asynchronously execute each action in the action plan according to priority and monitor the execution status and timeout. It performs predefined rollback operations on actions that fail to execute or time out and records the execution results. The feedback optimization module is used to update the weights of the triggered rules based on the execution results to achieve adaptive optimization of the strategy, analyze the evolution trajectory of the health level over multiple consecutive cycles, identify convergence, oscillation or divergence state patterns and adjust the trigger sensitivity of subsequent rules according to the identified state patterns, and trigger the acquisition module to enter the next decision cycle after completing this round of feedback.
Citation Information
Patent Citations
Adaptive power distribution system and method for micro cellular base station, and medium
CN112118632A
System and method for monitoring dynamic environment and equipment state of high-voltage power distribution station house
CN120489255A
