A multi-node cooperative cloud-free edge autonomous perception and decision system and method

Abstract

This invention discloses a multi-node collaborative, cloud-decentralized edge autonomous perception and decision-making system and method, comprising: multiple edge perception and decision-making nodes forming a decentralized autonomous cluster through an industrial mesh self-organizing network; each node includes a data acquisition unit, a computing and perception unit, a status monitoring unit, a communication unit, a collaborative decision-making unit, a storage unit, and a microcontroller unit. The status monitoring unit collects resource status and calculates a health score; the computing and perception unit extracts lightweight anomaly labels and feature vectors; the collaborative decision-making unit performs master node election based on the health score, fuses the received external anomaly labels and feature vectors with the local node's feature vector and health score, matches an anomaly propagation rule matrix, and outputs an action when the joint triggering condition is met. This invention achieves autonomous organization, collaborative perception, and distributed decision-making of the node cluster during cloud interruptions, featuring low overhead, high fault tolerance, and low communication burden.

Description

Technical Field

[0001] This invention belongs to the fields of IoT edge computing, distributed control systems and industrial intelligent monitoring technology, and in particular relates to a cloud-free edge autonomous sensing and decision-making system and method with multi-node collaboration. Background Technology

[0002] Traditional edge sensing systems primarily rely on a centralized "cloud-edge-device" architecture. Edge nodes are mainly responsible for data collection and simple preprocessing, while all complex anomaly analysis, feature extraction, global decision-making, and long-term storage are uploaded to the cloud server. This architecture offers advantages such as strong centralized management and ease of global optimization when the cloud communication link is stable.

[0003] However, existing edge sensing and decision-making systems generally adopt a centralized design dominated by the cloud, with edge nodes having relatively fixed functions, serving only as data acquisition and forwarding units. In extreme scenarios such as deep mine tunnels, offshore platforms, and remote areas where communication is prone to interruption or instability, once the cloud link fails, experiences high latency, or suffers a prolonged interruption, the entire system loses its real-time sensing and decision-making capabilities, leading to disruptions in safe production continuity, slow response times, and even serious safety accidents. The high dependence of existing systems on cloud communication makes them extremely susceptible to paralysis under network partitioning, physical damage, electromagnetic interference, or extreme weather conditions, failing to guarantee high continuity and low latency requirements.

[0004] In existing technologies, while edge nodes can perform local threshold judgments or simple rule processing, they lack effective multi-node dynamic collaboration mechanisms and decentralized consensus capabilities. Each node often exists in an "island" state, unable to share abnormal features in real time, perceive the global state, or collaboratively judge complex events (such as combined anomalies of regional vibration, temperature, and displacement), easily leading to misjudgments, missed judgments, or decision conflicts. Simultaneously, critical decisions still need to wait for cloud-based delivery, further amplifying the risk of single-point failures and response time. Regarding inter-node collaboration, traditional solutions lack lightweight, fault-tolerant distributed communication and consensus mechanisms, making them unsuitable for bandwidth-constrained and resource-limited industrial environments.

[0005] At the decision-making and execution level, existing edge systems mostly rely on simple local rules or fixed thresholds, lacking dynamic reconfiguration and distributed consensus capabilities. When faced with joint anomalies across multiple nodes (such as regional rockbursts or cascading equipment failures), a single node struggles to obtain a global view, significantly reducing the accuracy and timeliness of decision-making. Furthermore, while some solutions introduce local edge processing, they still require a fixed central node or cloud coordination, failing to achieve a truly "de-cloudified" autonomous closed loop. Especially when the master node fails, the network partitions, or some nodes fail, the overall resilience and continuity of decision-making are difficult to guarantee.

[0006] While existing technologies have achieved basic data acquisition, uploading, and simple local judgment, they generally suffer from the following prominent drawbacks: strong reliance on cloud communication leads to a high risk of system paralysis in extreme environments; lack of dynamic collaboration and consensus mechanisms between nodes results in silos and decision-making delays; heavy communication and computing burdens (requiring frequent transmission of massive amounts of raw data); incomplete decision-making loops and long response times; and insufficient overall system fault tolerance and self-healing capabilities. The root cause of these problems lies in the fact that traditional designs focus more on "centralized intelligence in the cloud" rather than "distributed autonomy at the edge," making it difficult to meet the actual needs of mines, marine environments, and remote industrial sites for unstable communication, high continuity, low latency, and zero cloud dependency. Summary of the Invention

[0007] To address the aforementioned technical problems, this invention proposes a multi-node collaborative cloud-free edge autonomous perception and decision-making system and method to solve the problems existing in the prior art.

[0008] To achieve the above objectives, the present invention provides a multi-node collaborative cloud-free edge autonomous perception and decision-making system, comprising: Multiple edge-aware decision-making nodes are interconnected through an industrial mesh self-organizing network to form a decentralized autonomous cluster; Each of the aforementioned edge-aware decision nodes includes: The data acquisition unit is used to connect multiple sensors and acquire raw data; A computation and perception unit is used to extract lightweight anomaly labels and feature vectors based on the raw data; The status monitoring unit is used to collect the resource status and communication quality indicators of this node and calculate the health score. A communication unit is used to broadcast lightweight anomaly labels and feature vectors between nodes; The collaborative decision-making unit is used to perform dynamic master node election based on the health scores of each node, and to fuse the external anomaly labels and feature vectors received through the communication unit with the local feature vectors currently output by the calculation and perception unit of this node and the health scores of the status monitoring unit. The fusion result is matched with a pre-stored anomaly propagation rule matrix, and the local execution action is output when the joint anomaly triggering conditions are met. Storage unit, used to store configuration information, exception event summaries, and decision logs; The microcontroller unit is used to coordinate the operation of the above-mentioned units.

[0009] Optionally, the indicators collected by the status monitoring unit include the remaining computing power utilization rate of the central processing unit, the proportion of available memory, the communication latency and packet loss rate of the industrial mesh self-organizing network, and the electrical connectivity quality of multiple sensor channels; the health score is obtained by normalizing and weighting the indicators collected by the status monitoring unit through a weighted evaluation function.

[0010] Optionally, dynamic master node election is performed based on the health scores of each node and the improved lightweight Raft consensus protocol, including: When a node changes from a follower to a candidate, it substitutes its own health score into the election timeout countdown function and determines the election timeout countdown duration based on the health score. After the countdown ends, the node initiates an election broadcast to the cluster. When multiple nodes initiate elections simultaneously, follower nodes receive candidate packets and extract the health scores carried within them, casting their votes for the candidate node with the highest health score. The elected master node broadcasts a heartbeat packet and a coarse clock synchronization signal to the industrial Mesh self-organizing network.

[0011] Optionally, the anomaly propagation rule matrix includes node identifiers, types of anomaly labels and feature vectors, logical relationships between anomaly labels of different nodes, time window constraints for anomaly occurrence at each node, and mapping relationships between decision actions and the logical relationships and time window constraints.

[0012] Optionally, when the collaborative decision-making unit determines that the joint anomaly triggering conditions are met, it outputs an execution action to the microcontroller unit, and the microcontroller unit executes a preset physical action through a local input / output interface; wherein, the joint anomaly triggering conditions include: receiving lightweight anomaly tags broadcast by multiple other nodes within a preset time window, and the local feature vector output by the calculation and perception unit of this node satisfies the threshold and logical relationship defined in the anomaly propagation rule matrix.

[0013] Optionally, the storage unit uses a non-volatile storage medium to store locally generated abnormal event summaries and decision logs during network cloud link interruptions, and uploads the abnormal event summaries and decision logs to the cloud in an asynchronous incremental manner after the network is restored.

[0014] Optionally, the computing and sensing unit performs a fast Fourier transform or a time-domain statistical algorithm on the raw data to output lightweight anomaly labels and feature vectors; the communication unit supports industrial mesh self-organizing networks for multi-hop broadcasting and point-to-point communication between nodes.

[0015] Optionally, the system also includes configuration software for zero-code configuration of the exception propagation rule matrix and joint decision-making rules, and for sending the configuration parameters to the storage units of each node, with the microcontroller unit controlling the computing and sensing units and the collaborative decision-making units to hot-load the new rules.

[0016] This invention also provides a multi-node collaborative cloud-free edge autonomous perception and decision-making method, applied to the above-mentioned system, comprising: After each edge sensing decision node is powered on, it calculates its own health score through the status monitoring unit, and performs dynamic master node election based on the health score according to the improved lightweight Raft consensus protocol. The elected master node broadcasts heartbeat and clock synchronization signals to the industrial Mesh self-organizing network through the communication unit. Follower nodes listen to the heartbeat and collect benchmark data periodically. After any node determines a local anomaly in the computing and sensing unit, it generates a lightweight anomaly label and feature vector, and broadcasts the lightweight anomaly label and feature vector generated by the node to the cluster through the communication unit. The collaborative decision-making unit of the receiving node will fuse the external anomaly labels and feature vectors received by the communication unit with the local feature vectors currently output by the calculation and perception unit of this node and the health score of the status monitoring unit, and match them with the anomaly propagation rule matrix in the storage unit. If the matching result meets the joint anomaly triggering condition, the collaborative decision-making unit outputs the execution action to the microcontroller unit. The microcontroller unit executes the preset physical action through the local input / output interface and stores a summary of the entire decision-making process in the storage unit. After the network is restored, each node will asynchronously and incrementally upload the summary of abnormal events and decision logs in the storage unit to the cloud.

[0017] Optionally, in the improved lightweight Raft consensus protocol, the triggering conditions for a node to change from a follower to a candidate include system initialization, failure of the current master node or heartbeat timeout, or detection of multi-master node conflict during network topology reconstruction.

[0018] Compared with the prior art, the present invention has the following advantages and technical effects: Completely achieve “de-cloudification” autonomy: In extreme scenarios where the cloud is completely interrupted or unreachable for a long time, the edge node cluster can still autonomously organize, dynamically elect master nodes, collaboratively perceive environmental anomalies, lightweightly propagate characteristics, and execute distributed decisions, greatly improving the continuity and safety production reliability of industrial sites (such as deep mine tunnels, offshore platforms, and remote environmental monitoring), and avoiding the problem of system paralysis caused by cloud link failures in traditional solutions. Multi-node dynamic collaboration and high fault tolerance: An improved lightweight Raft consensus protocol (based on a health-weighted dynamic election mechanism) and an anomaly propagation rule matrix are introduced to achieve low-overhead, high-fault-tolerant state synchronization and joint event judgment among nodes. This supports rapid reconstruction after master node failure and autonomous operation in network partitioning scenarios, significantly enhancing the overall resilience and self-healing ability of the system and overcoming the "island" state and decision conflict risks of existing edge solutions. Extremely low communication and computing burden: Only anomaly labels and key feature vectors (rather than raw massive data) are propagated, reducing the data volume by more than 90% compared to traditional cloud models, adapting to industrial field environments with limited bandwidth and resources; local sensing, cleaning and other processing are completed at the edge, reducing the burden of secondary analysis in the cloud and reducing transmission latency and traffic costs; Truly distributed closed-loop decision-making: Enables local priority autonomous decision execution (such as joint triggering of power cut-off or alarm due to vibration exceeding limits and temperature anomalies), without waiting for confirmation from the cloud or fixed central node, reducing response time to milliseconds, supporting real-time closed-loop control of complex industrial events (such as regional rock bursts and equipment cascading failures), improving decision-making accuracy and safety; Deployment and maintenance friendly: The node hardware is universal and no fixed firmware needs to be preset. The accompanying software enables visual and zero-code configuration of exception rules and decision logic, and supports remote push updates and asynchronous cloud summary synchronization. The roles of installers and users are separated, reducing the professional threshold and the complexity of on-site debugging. It is suitable for rapid deployment and long-term stable operation in large-scale and ever-changing scenarios. Attached Figure Description

[0019] The accompanying drawings, which form part of this application, are used to provide a further understanding of this application. The illustrative embodiments and descriptions of this application are used to explain this application and do not constitute an undue limitation of this application. In the drawings: Figure 1 This is a schematic diagram of the system structure according to an embodiment of the present invention; Figure 2 This is a schematic diagram of the interior of an edge node according to an embodiment of the present invention; Figure 3 This is a flowchart of the dynamic election process for the master node in an embodiment of the present invention. Detailed Implementation

[0020] It should be noted that, unless otherwise specified, the embodiments and features described in this application can be combined with each other. This application will now be described in detail with reference to the accompanying drawings and embodiments.

[0021] It should be noted that the steps shown in the flowchart in the accompanying drawings can be executed in a computer system such as a set of computer-executable instructions, and although a logical order is shown in the flowchart, in some cases the steps shown or described may be executed in a different order than that shown here.

[0022] Example 1 like Figure 1 As shown, this embodiment provides a multi-node collaborative cloud-decentralized edge autonomous perception and decision-making system. The system uses multiple edge perception and decision-making nodes as the core to construct a decentralized autonomous network.

[0023] Each edge-aware decision-making node consists of an IoT smart gateway. The gateway hardware adopts a general container design, with a microcontroller unit (MCU) at its core, responsible for cross-module task distribution, interrupt handling, and algorithm execution for collaborative decision-making units. Figure 2 As shown, the following functional units are integrated: (1) Data acquisition unit: Physically connects multiple sensors. The MCU sends acquisition instructions to the data acquisition unit through serial communication. The data acquisition unit acquires the original sensor data and returns the original sensor data to the MCU for subsequent data processing. This unit supports multiple protocol and multi-channel sensor interfaces and can connect to sensors such as vibration, temperature, displacement, strain, and current. It supports interfaces such as RS485 and Ethernet.

[0024] (2) Computation and Sensing Unit: The MCU sends the cached raw data stream into this unit, performs Fast Fourier Transform (FFT) or time-domain statistics through the preset anomaly detection model algorithm, and outputs lightweight anomaly tags and feature vectors to reduce the subsequent processing burden of the MCU. (3) Status Monitoring Unit: This unit is the foundation of system autonomy. It periodically collects data on the node's remaining CPU computing power utilization, available memory ratio, Mesh network communication latency / packet loss rate, and the electrical connectivity quality (such as noise level and disconnection status) of the connected multi-sensor channels. Each indicator is fused and calculated in real time using a preset weighted evaluation function to output a normalized quantitative indicator, namely the health score (H-Score, range, for example, 0-100 points), which is then reported to the MCU. This evaluation result dynamically reflects the node's comprehensive ability to undertake global collaborative scheduling tasks.

[0025] The status monitoring unit calculates the health score (H-Score) using a dynamic weighted evaluation function, with the specific calculation formula as follows: ; ; ; ; ; in, U cpu ( t )for T The remaining computing power utilization of the central processing unit at any time is monitored, and a preset value is used to avoid interference from instantaneous peak values. K 1 consecutive sampling period (step size Δ) t The remaining computing power is calculated using the moving average load rate within the specified range. Lcpu This represents the instantaneous load rate of the MCU within a single sampling period. M mem ( t () represents the percentage of available memory; Mem total The total physical memory capacity allocated to collaborative decision-making and computational perception tasks. Mem used ( t )for t The amount of memory currently in use; Q net ( t () represents the communication quality index; D ( t The current time interval represents the round-trip communication delay of heartbeat packets between the current node and its neighboring nodes. D max This is the maximum tolerable delay allowed by the system (used to normalize the delay). L ( t () represents the packet loss rate at the current moment; l 1 and l 2 is the preset exponential decay coefficient, used to adjust the penalty for delay and packet loss on the communication quality score; S link For the electrical connectivity quality of multiple sensor channels; C total This represents the total number of sensor channels physically connected to this node. d c ( t ) is the first c The connection status indication function of the first channel, when the status monitoring unit determines the connection status of the first channel... c When the variance of the reference data for each channel is within the normal electrical noise range (i.e., no open circuit or strong electromagnetic interference short circuit has occurred), d c ( t )=1, otherwise d c ( t )=0; oh 1. oh 2. oh 3 and oh 4 represents the weighting coefficient for each item. This coefficient is dynamically adjusted based on the node's current operating status and the network environment. The dynamic adjustment mechanism specifically includes: 1) Resource bottleneck penalty mechanism: When the status monitoring unit detects a certain resource indicator (such as...) U cpu or M memWhen the remaining computing power falls below a preset critical threshold (e.g., less than 10% of the total computing power), the system automatically triggers a penalty function, significantly increasing the weight of this indicator in the evaluation function. Because the indicator itself has an extremely low value, this amplified weight will cause the node's overall health score (H-Score) to plummet exponentially, forcing the node to be quickly eliminated in subsequent dynamic master node elections or having its trust weight reduced in joint decision-making. This effectively prevents "sick" nodes from dragging down the entire cluster. 2) Network fluctuation adaptive mechanism: when Q net ( t When drastic fluctuations occur, the status monitoring unit dynamically adjusts the communication quality weights by calculating the variance of recent communication delays. oh 3. This makes the system's master node election and collaborative decision-making more inclined to select nodes that are in the core of the network and have the most stable communication links during periods of network instability, thereby ensuring the overall communication resilience of the decentralized cluster.

[0026] (4) Storage unit: Non-volatile storage (SD card) is used. The MCU uses this unit to store the gateway's configuration information, sensor historical data, gateway abnormal event logs and decision summaries, etc. (5) Communication Unit: The MCU realizes communication between different gateways through this unit. The unit supports industrial Mesh self-organizing network or Ethernet / Wi-Fi hybrid networking, thereby realizing reliable multi-hop broadcast and point-to-point communication between nodes; (6) Collaborative Decision-Making Unit: This unit is the autonomous logic center of the system. By integrating lightweight feature information between multiple nodes and matching the anomaly propagation rule matrix, it realizes distributed event analysis and real-time linkage decision control without cloud participation according to the preset improved lightweight Raft consensus protocol.

[0027] Under normal circumstances, without external commands or abnormal triggers, the gateway's status monitoring unit continuously runs in the background, monitoring the node's remaining CPU utilization, available memory ratio, Mesh network communication latency / packet loss rate, and the electrical connectivity quality (such as noise level and disconnection status) of the connected multi-sensor channels in real time. It calculates a real-time health score (H-Score) and periodically reports it to the MCU via the internal bus. The MCU then passes this score to the collaborative decision-making unit, which uses an improved lightweight Raft consensus protocol to maintain the current node role (i.e., leader or follower). Simultaneously, the communication unit maintains the connectivity of the industrial Mesh self-organizing network based on the status of the collaborative decision-making unit. If the node is the leader, it broadcasts heartbeat packets and coarse clock synchronization signals to downstream nodes via the communication unit; if it is a follower, it listens for heartbeat packets. Furthermore, the MCU sends acquisition commands to the data acquisition unit via a serial port (such as RS485) according to the pre-configured polling settings in its local storage unit (i.e., pre-configured by the user based on the gateway's sensor acquisition needs). The data acquisition unit obtains data and returns it to the MCU. The MCU then sends the data to the computing and sensing unit for time-domain statistics to verify the effectiveness of the sensor's physical connection and prevent blind spots.

[0028] When the gateway's communication unit receives a command data packet from the supporting software or master node, the MCU extracts the valid information from the data packet and performs protocol parsing. If it is a configuration command, the MCU directly overwrites the parsed new sensing model parameters, anomaly judgment formulas, or anomaly propagation rule matrices into the local storage unit, and instructs the computation and sensing unit and collaborative decision-making unit to hot-load the new rules to achieve dynamic reconstruction. If it is a data acquisition command, the MCU issues an acquisition command containing specific baud rates and channel characteristics to the data acquisition unit. The data acquisition unit acquires sensor data from each channel and returns the raw data to the MCU. The MCU pushes the raw data stream into the computation and sensing unit to execute algorithms such as Fast Fourier Transform (FFT), extracts and outputs lightweight anomaly tags and feature vectors, and transmits them through the communication unit. These lightweight tags are broadcast to other nodes in the Mesh network, and the operation logs are stored in the local storage unit. If the lightweight anomaly tag is broadcast by another node, the MCU immediately wakes up the collaborative decision-making unit, which merges the received external anomaly tag with the local state (such as local temperature characteristics) currently output by the local computing and sensing unit. If the joint anomaly triggering conditions are met, the collaborative decision-making unit outputs the execution action to the MCU, and the MCU immediately executes the preset physical action (such as cutting off the device power, triggering the relay alarm, etc.) through the local I / O interface, and writes a summary of the entire process of this joint decision-making to the storage unit, and waits for the network to recover before reporting asynchronously.

[0029] For each gateway, there is accompanying software (available as a mobile app or dedicated management terminal) providing a visual interface for configuration. Users can configure system parameters, anomaly rules, and decision logic without programming. The system does not have pre-set firmware; instead, node behavior is dynamically defined through software, enabling adaptive access from various sensors, local anomaly detection, and distributed decision-making. After collecting raw sensor data, nodes immediately perform dynamic parsing, feature extraction, and cleaning, generating anomaly tags and lightweight features. Only key information (not raw data) is propagated between nodes, and collaborative judgment and local decision execution are achieved through an improved consensus mechanism. After network recovery, asynchronous uploading of event summaries to the cloud is supported for post-event analysis and optimization.

[0030] Different nodes form autonomous clusters through a mesh network, eliminating the need for a fixed central hub or cloud-based coordination, and supporting dynamic reconstruction and fault tolerance for partial node failures.

[0031] like Figure 3 As shown, this embodiment also provides a multi-node collaborative cloud-free edge autonomous perception and decision-making method, the specific implementation steps of which include: 1. Physical connection and dynamic initialization of system parameters; Various types of sensors required for each monitoring point (including but not limited to vibration, temperature, displacement, strain, current, etc.) are physically connected to the corresponding edge sensing decision nodes through the gateway's multi-protocol channel interface according to their respective communication protocols (such as RS485, Ethernet), and then the node power is turned on. The accompanying visual configuration software (mobile or management terminal) scans the node device identifier to complete node addition, configuration of system operating parameters, anomaly rules, and decision logic. For anomaly detection rules, the user inputs sample data, and the software breaks down the raw data stream into byte units and displays them visually. The user constructs anomaly judgment formulas, propagation matrices, and joint decision rules by selecting and dragging building block components (sign bits, coefficients, constants, thresholds, time windows, etc.), naming each anomaly segment, specifying units, and inputting test data to verify accuracy. The entire process relies on a zero-code, visual, block-based interactive interface, allowing users to intuitively complete the definition of sensing rules, propagation logic, and decision closed loop. Specifically, the propagation matrix includes node IDs, anomaly labels and feature vectors, logical relationships between anomaly labels between nodes, the temporal correlation of anomalies occurring at different nodes, and the mapping relationship of decision actions. After the node is powered on, it automatically uses the communication unit (industrial Mesh self-organizing network) to build a decentralized autonomous network topology.

[0032] The anomaly detection formula specifically includes a time-domain feature detection sub-formula, a frequency-domain feature detection sub-formula, and a rate of change detection sub-formula, which output the local anomaly confidence score through weighted fusion. The specific formula is as follows: ; ; ; ; in, E local ( t ) represents the local anomaly confidence level; f time [ v ( t [)] represents the time-domain feature determination value; f freq [ P ( f [)] represents the frequency domain feature determination value; v ( t This represents the raw sensor data at the current moment. m and s These represent the historical mean and standard deviation within the sliding time window, respectively. P ( f k () represents discrete frequency points f k The power spectral density after the fast Fourier transform; P ref ( f k The preset normal reference power spectral density is denoted as . M The total number of discrete frequency points; α , β , c This is a preset, custom weight coefficient. When... E local ( t () greater than the preset local security threshold Th local At that time, the sensing unit determined that there was a local anomaly.

[0033] The joint decision-making rule is specifically composed of four dimensions: spatial correlation constraint, time window constraint, logical threshold, and node health trust weight. If within the preset joint time window... T 0, T 0+Δ T Within, joint trigger function F joint If the output is 1, then the corresponding action will be executed. The specific formula is as follows: in, N The total number of associated nodes participating in the collaboration (spatial correlation constraint, defined by the anomaly propagation rule matrix); I ( Tags i) is an indicator function, when node i In the time window Δ T The value is 1 when the specified exception label is broadcast internally, and 0 otherwise (time window constraint). C i For nodes i Abnormal confidence level of broadcast (i.e., the aforementioned) E local ( t )); W i For nodes i The dynamic trust weight is positively correlated with the health score (H-Score) output by the node's current state monitoring unit, i.e. W i = f (H-score i Nodes with higher health levels have greater weight in their anomaly labels during joint decision-making. Furthermore, the joint decision-making rule also includes a Boolean logic decision tree based on multi-source heterogeneous anomaly labels. The collaborative decision-making unit executes the quantization formula. F joint At the same time, the received abnormal tags of each node Tags i The mapping is performed as a set of Boolean variables, and logical combination operations are performed according to the logical operators (including but not limited to "AND", "OR", and "NOT") defined in the anomaly propagation rule matrix. Only when the quantization formula... F joint The collaborative decision-making unit will only output the execution action when the result of the Boolean logic decision tree is true (True) and the result of the Boolean logic decision tree is also true.

[0034] 2. Node health status monitoring and dynamic master node election; After powering on, each node initially enters a follower state, assesses its own health, and runs an improved Raft algorithm to dynamically elect a leader responsible for heartbeat broadcasting and coarse clock synchronization. The improved Raft algorithm is triggered by three conditions: system initialization (when all nodes power on), leader failure or link interruption (follower nodes do not receive a heartbeat packet from the current leader within the set heartbeat timeout window), and network topology reconstruction (e.g., when a network partition is re-merged and multiple leader conflicts are detected). If any one of these three conditions is met, the node immediately changes from follower to candidate, triggering the improved Raft election process. In the traditional Raft algorithm, the election timeout is completely random. However, in the improved Raft algorithm of this invention, nodes incorporate their own health score into a countdown function; nodes with higher H-scores receive shorter election timeout countdowns. This means that the node with the best physical condition and most abundant resources will finish its countdown first and initiate a campaign broadcast to the network. Furthermore, when multiple nodes initiate elections simultaneously due to similar timeout periods, follower nodes compare the H-scores carried in the candidate packets during the voting logic, prioritizing their votes for candidate nodes with higher health. This mechanism ensures that the system always has the most robust node currently in power as the master node, responsible for subsequent heartbeat broadcasts and coarse clock synchronization, significantly reducing the risk of system crashes due to master node hardware performance bottlenecks.

[0035] 3. Edge autonomous perception and collaborative feature propagation; Under normal monitoring conditions, nodes, configured as usual, receive low-frequency acquisition commands from the MCU to the data acquisition unit to obtain baseline data. The "computation and sensing unit" performs extremely lightweight time-domain statistics, verifying only the validity of sensor connections and preventing blind spots. If a node is a follower, it listens for heartbeats; if it is a master node, it sends heartbeat and synchronization signals. When a specific acquisition command is received or a trigger condition is met, the node sends a command to the corresponding sensor to acquire the raw data stream at high frequency. The data is sent to the local "computation and sensing unit" to perform Fast Fourier Transform (FFT) or time-domain statistical model algorithms. If the health status calculation result indicates an anomaly, the node immediately performs dynamic parsing, feature extraction, and cleaning, generating only lightweight anomaly tags and feature vectors (clearly identifying key information such as node, event type, and level, rather than massive amounts of raw data), which are then broadcast multi-hop to other nodes in the Mesh network via the communication unit.

[0036] 4. Distributed joint decision-making and autonomous closed loop; Upon receiving the broadcast, the node immediately activates the "Collaborative Decision-Making Unit," fusing the received external anomaly label with the local state (such as local feature vector and health status) currently output by its "Computation and Sensing Unit." The decision-making unit matches the fused information with the locally stored "Anomaly Propagation Rule Matrix." If the joint anomaly triggering conditions defined in the matrix are met (e.g., multiple nodes in the region experiencing specific types of vibration exceeding limits + the node's temperature rising sharply within a preset time window), then the distributed decision-making logic is triggered. The node does not need to wait for confirmation from the cloud or a fixed central node; the "Collaborative Decision-Making Unit" directly outputs the execution action to the MCU. The MCU executes preset physical actions (such as cutting off power to associated devices, triggering relay audible and visual alarms, adjusting device operating parameters, etc.) through its local I / O interface and stores a summary of the entire collaborative assessment process (including decision basis, action content, and timestamp) in the storage unit.

[0037] 5. Asynchronous digest synchronization after network recovery; The system monitors the cloud network connection status in real time through a status monitoring unit. In the autonomous mode where the cloud link is completely interrupted or extremely unstable, the node cluster maintains autonomous operation by relying on the above steps. When the network is detected to have recovered (or the periodic reporting conditions are met), the system autonomously and asynchronously uploads locally stored historical anomaly event summaries, decision logs, etc., in an incremental manner that is adaptive to the network status. The cloud (or host computer software) presents the received summaries and decision results in a visual form (such as status graphs, line graphs, heatmaps) for users to perform post-event analysis and global production optimization, thereby realizing dual-mode adaptive operation of cloud-edge collaboration and full edge autonomy.

[0038] This invention improves the continuity and reliability of industrial production by enabling an edge-aware decision-making node cluster to autonomously organize, dynamically elect a master node, collaboratively perceive environmental anomalies, lightweightly propagate features, and execute distributed decisions in extreme scenarios where the cloud is completely interrupted or unavailable for extended periods. Through an improved lightweight Raft consensus protocol and anomaly propagation rule matrix, this invention achieves low-overhead, high-fault-tolerance state synchronization and joint event judgment among nodes, supporting rapid reconstruction after master node failure and autonomous operation in network partitioning scenarios. This invention only propagates anomaly tags and key feature vectors, adapting to bandwidth- and resource-constrained industrial environments; local perception, feature extraction, and cleaning are completed at the edge, reducing the burden of secondary analysis in the cloud. The local-priority autonomous decision-making execution of this invention does not require cloud confirmation, supporting real-time closed-loop control of complex industrial events. The node hardware of this invention is universal and requires no pre-set fixed firmware; zero-code configuration and hot loading are achieved through accompanying software, supporting remote push updates and asynchronous cloud digest synchronization.

[0039] The above are merely preferred embodiments of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

Claims

1. A multi-node collaborative, cloud-free edge autonomous sensing and decision-making system, characterized in that, include: Multiple edge-aware decision-making nodes are interconnected through an industrial mesh self-organizing network to form a decentralized autonomous cluster; Each of the aforementioned edge-aware decision nodes includes: The data acquisition unit is used to connect multiple sensors and acquire raw data; A computation and perception unit is used to extract lightweight anomaly labels and feature vectors based on the raw data; The status monitoring unit is used to collect the resource status and communication quality indicators of this node and calculate the health score. A communication unit is used to broadcast lightweight anomaly labels and feature vectors between nodes; The collaborative decision-making unit is used to perform dynamic master node election based on the health scores of each node, and to fuse the external anomaly labels and feature vectors received through the communication unit with the local feature vectors currently output by the calculation and perception unit of this node and the health scores of the status monitoring unit. The fusion result is matched with a pre-stored anomaly propagation rule matrix, and the local execution action is output when the joint anomaly triggering conditions are met. Storage unit, used to store configuration information, exception event summaries, and decision logs; The microcontroller unit is used to coordinate the operation of the above-mentioned units.

2. The multi-node collaborative cloud-free edge autonomous sensing and decision-making system according to claim 1, characterized in that, The indicators collected by the status monitoring unit include the remaining computing power utilization of the central processing unit, the proportion of available memory, the communication latency and packet loss rate of the industrial Mesh self-organizing network, and the electrical connectivity quality of multiple sensor channels; the health score is obtained by normalizing and weighting the indicators collected by the status monitoring unit through a weighted evaluation function.

3. The multi-node collaborative cloud-free edge autonomous sensing and decision-making system according to claim 1, characterized in that, Dynamic master node election is performed based on the health scores of each node and the improved lightweight Raft consensus protocol, including: When a node changes from a follower to a candidate, it substitutes its own health score into the election timeout countdown function and determines the election timeout countdown duration based on the health score. After the countdown ends, the node initiates an election broadcast to the cluster. When multiple nodes initiate elections simultaneously, follower nodes receive candidate packets and extract the health scores carried within them, casting their votes for the candidate node with the highest health score. The elected master node broadcasts a heartbeat packet and a coarse clock synchronization signal to the industrial Mesh self-organizing network.

4. The multi-node collaborative cloud-free edge autonomous sensing and decision-making system according to claim 1, characterized in that, The anomaly propagation rule matrix includes node identifiers, anomaly labels and feature vector types, logical relationships between different node anomaly labels, time window constraints for anomaly occurrence at each node, and mapping relationships between decision actions and the logical relationships and time window constraints.

5. The multi-node collaborative cloud-free edge autonomous sensing and decision-making system according to claim 1, characterized in that, When the collaborative decision-making unit determines that the joint anomaly triggering conditions are met, it outputs an execution action to the microcontroller unit, which then executes a preset physical action through its local input / output interface. The joint anomaly triggering conditions include: receiving lightweight anomaly tags broadcast by multiple other nodes within a preset time window, and the local feature vector output by the computation and perception unit of this node satisfying the threshold and logical relationship defined in the anomaly propagation rule matrix.

6. The multi-node collaborative cloud-free edge autonomous sensing and decision-making system according to claim 1, characterized in that, The storage unit uses a non-volatile storage medium to store locally generated abnormal event summaries and decision logs during network cloud link interruptions, and uploads the abnormal event summaries and decision logs to the cloud asynchronously and incrementally after the network is restored.

7. The multi-node collaborative cloud-free edge autonomous sensing and decision-making system according to claim 1, characterized in that, The computing and sensing unit performs fast Fourier transform or time-domain statistical algorithm on the raw data and outputs lightweight anomaly labels and feature vectors; the communication unit supports industrial mesh self-organizing networks for multi-hop broadcasting and point-to-point communication between nodes.

8. The multi-node collaborative cloud-free edge autonomous sensing and decision-making system according to claim 1, characterized in that, The system also includes configuration software for zero-code configuration of the exception propagation rule matrix and joint decision-making rules, and sends the configuration parameters to the storage units of each node. The microcontroller unit controls the computing and sensing units and the collaborative decision-making units to hot-load the new rules.

9. A multi-node collaborative de-cloudified edge autonomous sensing and decision-making method, applied to the system of any one of claims 1 to 8, characterized in that, include: After each edge sensing decision node is powered on, it calculates its own health score through the status monitoring unit, and performs dynamic master node election based on the health score according to the improved lightweight Raft consensus protocol. The elected master node broadcasts heartbeat and clock synchronization signals to the industrial Mesh self-organizing network through the communication unit. Follower nodes listen to the heartbeat and collect benchmark data periodically. After any node determines a local anomaly in the computing and sensing unit, it generates a lightweight anomaly label and feature vector, and broadcasts the lightweight anomaly label and feature vector generated by the node to the cluster through the communication unit. The collaborative decision-making unit of the receiving node will fuse the external anomaly labels and feature vectors received by the communication unit with the local feature vectors currently output by the calculation and perception unit of this node and the health score of the status monitoring unit, and match them with the anomaly propagation rule matrix in the storage unit. If the matching result meets the joint anomaly triggering condition, the collaborative decision-making unit outputs the execution action to the microcontroller unit. The microcontroller unit executes the preset physical action through the local input / output interface and stores a summary of the entire decision-making process in the storage unit. After the network is restored, each node will asynchronously and incrementally upload the summary of abnormal events and decision logs in the storage unit to the cloud.

10. The multi-node collaborative cloud-free edge autonomous perception and decision-making method according to claim 9, characterized in that, In the improved lightweight Raft consensus protocol, the triggering conditions for a node to change from a follower to a candidate include system initialization, failure of the current master node or heartbeat timeout, or detection of multi-master node conflict during network topology reconstruction.

Images