A distributed resource management terminal and method based on multi-modal perception

Abstract

This invention relates to the field of distributed resource management and intelligent sensing technology, and discloses a distributed resource management terminal and method based on multimodal sensing. First, it corrects the observation time of multi-source sensors based on the BeiDou spatiotemporal reference, inversely infers the spatiotemporal nature of real events, and fuses the output uncertainty. Second, it constructs a four-layer causal graph of environment, event, observation, and system, and identifies root causes through counterfactual intervention intensity. Then, it dynamically adjusts sampling, inference, communication, and alarm strategies according to uncertainty and root cause type. Subsequently, it performs hierarchical dynamic programming by combining risk-reward, battery SOC, and photovoltaic prediction to achieve resource optimization. Finally, it completes multi-terminal task migration based on a virtual energy pool and the benefit-normalized cost inequality. The five-level modules form a closed-loop collaboration through four feedback paths, which can improve fusion reliability and critical event assurance rate, reduce false alarm rate, and extend battery life, making it suitable for scenarios such as forest fire prevention, water environment monitoring, and border security.

Description

Technical Field

[0001] This invention relates to the field of distributed resource management and intelligent decision-making technology, and in particular to a five-level closed-loop adaptive cooperative control method and terminal for multimodal sensing terminals in the field, specifically a distributed resource management terminal and method based on multimodal sensing. Background Technology

[0002] Distributed sensing terminals in the field are widely used in scenarios such as forest fire prevention, watershed water environment monitoring, border security, mine safety, oil and gas pipeline inspection, and ecological protection. These terminals are typically deployed in areas with limited power supply, communication, and maintenance conditions, and need to operate continuously for extended periods using photovoltaic, battery, or low-power communication methods. They collect environmental status and abnormal event information through various sensors, including gas, water quality, infrared thermal imaging, visible light, sound, vibration, and meteorological sensors.

[0003] Existing field sensing systems typically employ technologies such as multi-sensor data fusion, anomaly detection, edge computing, and low-power scheduling to improve event detection capabilities and system endurance. However, the following shortcomings still exist in practical applications.

[0004] First, different types of sensors have different observation mechanisms. For example, sound signals propagate at the speed of sound, thermal anomalies involve thermal diffusion, water pollutants spread with water flow and dispersion, and gas leaks are affected by wind speed, wind direction, and turbulent diffusion. If only sensor sampling times are used for simple alignment, it is easy to misjudge the temporal manifestations of the same event in different sensors as multiple independent anomalies, or to incorrectly fuse different events into a single event. Therefore, existing multimodal fusion methods lack a unified back-inference mechanism oriented towards the spatiotemporal occurrence of real events.

[0005] Secondly, existing anomaly detection methods are mostly based on fixed thresholds, statistical models, or black-box deep learning models. While they can determine whether observed values ​​are abnormal, they often struggle to explain the causes of the anomalies. For example, an increase in infrared temperature could originate from a real fire, strong sunlight reflection, or sensor drift; an abnormal gas concentration could result from a pollution leak, a change in local wind direction, or sensor aging. Due to the lack of a causal root cause identification mechanism, existing systems are prone to false alarms, missed alarms, or erroneous alerts.

[0006] Secondly, field terminals have limited computing, communication, and energy resources. Existing task scheduling strategies typically adjust sampling frequency and communication modes based on fixed periods or simple SOC thresholds, failing to incorporate fusion uncertainty, root cause type, event risk, and energy status into scheduling decisions. This results in wasted energy under low-risk conditions or insufficient resource allocation under high-risk conditions.

[0007] Furthermore, field terminals largely rely on photovoltaic power, and energy availability is significantly affected by weather, season, shading, and battery status. Existing energy management methods mostly focus on the current SOC of individual terminals, lacking forward-looking optimization that incorporates future sunlight forecasts, risk predictions, and mission benefits, making it difficult to strike a balance between ensuring critical missions and extending endurance.

[0008] Finally, in multi-terminal networking scenarios, there are often overlapping coverage and complementary resources among neighboring terminals. When a terminal is unable to continue performing critical tasks due to low battery, communication abnormalities, or excessive computing load, existing technologies lack a task migration mechanism based on coverage capability, energy margin, migration benefits, and migration costs, making it difficult to form a cross-terminal collaborative closed loop. Summary of the Invention

[0009] The present invention mainly solves the following technical problems:

[0010] (1) The problem of spatiotemporal inconsistency of multimodal sensing data

[0011] Sensors in field terminals, such as those for gas, water quality, sound, thermal imaging, and vibration, have different response delays, transmission delays, and physical propagation delays. Existing methods typically fuse data only based on sampling time, making it difficult to accurately deduce the actual time and location of events, resulting in low reliability of multimodal fusion.

[0012] (2) Problems where anomaly detection cannot explain the root cause.

[0013] Existing anomaly detection methods mostly rely on threshold models or black-box neural networks, which are difficult to distinguish between real events, environmental interference, sensor drift, communication anomalies, or equipment failures, and are prone to false alarms, missed alarms, or incorrect scheduling.

[0014] (3) The problem of disconnect between task scheduling and root cause diagnosis

[0015] Existing distributed sensing terminals typically adjust sampling frequency, communication method, and alarm level according to fixed rules, and cannot dynamically adjust resource allocation based on fusion uncertainty and root cause type.

[0016] (4) The problem of lack of risk perception and forward-looking optimization in energy management

[0017] Field terminals often rely on photovoltaics, batteries, and low-power operation strategies. Existing energy management is usually based on local control only on the current SOC, making it difficult to simultaneously consider future lighting, mission risks, critical event protection, and battery protection.

[0018] (5) Insufficient multi-terminal collaboration capability

[0019] When a single terminal has insufficient power, limited communication, or reduced coverage, the existing system lacks a task migration mechanism based on benefit and cost assessment, making it difficult to form a closed loop of cross-terminal collaborative perception and continuous monitoring.

[0020] To achieve the above objectives, the present invention adopts the following technical solution:

[0021] In a first aspect, the present invention provides a closed-loop adaptive cooperative control method based on a multimodal sensing distributed resource management terminal, comprising a five-level closed-loop control process, and forming a closed loop through four explicit feedback paths:

[0022] The first stage involves performing sensor observation time correction, real event spatiotemporal inference, and dynamic causal propagation graph construction on multimodal sensor data based on the BeiDou spatiotemporal reference. After fusion by a graph neural network, the fusion feature F(t) and fusion uncertainty U(t) are output. The U(t) is adjusted by the reverse driving sampling strategy through feedback path ①.

[0023] The second level is based on a four-layer causal diagram consisting of environmental background, event root cause, observation performance, and system state. Counterfactual intervention baselines are set according to variable type, the intervention intensity (ICS) of each candidate root cause is quantified, and the root cause type is determined based on the ICS ranking results. The root cause type drives the task priority adjustment in reverse through feedback path ②.

[0024] The third level involves adjusting the sampling frequency, inference depth, communication channel, and alarm level according to preset rules based on U(t) and the root cause type, and outputting a task scheduling list.

[0025] The fourth level uses the objective function composed of risk-weighted terms and battery protection penalty terms to perform hourly-minute hierarchical dynamic programming based on the SOC state equation normalized by battery capacity, outputs a resource allocation scheme, and feeds it back to the first and second levels through feedback path ③.

[0026] At the fifth level, based on the virtual energy pool construction and the benefit-normalized cost migration decision inequality, the task migration instruction is output. After the migration is completed, a cross-terminal fusion closed loop is formed through feedback path ④.

[0027] Secondly, the present invention also provides a distributed resource management terminal based on multimodal perception, including an environmental protection and installation support system, a three-level energy system, a multimodal perception integration module, a Beidou spatiotemporal reference module, an edge computing module, a multimodal communication module, and a regional broadcast alarm module;

[0028] The three-level energy system includes photovoltaic modules, energy storage batteries and supercapacitors, which are used to supply power to the distributed resource management terminal and provide the edge computing module with state of charge (SOC), power generation and energy storage status data.

[0029] The BeiDou time and space reference module provides a unified time synchronization, positioning and timekeeping reference for all sensors;

[0030] The edge computing module includes a processor and a memory, the memory storing a computer program, and the processor executing the computer program to implement the method described above.

[0031] The multi-mode communication module is used to perform event reporting, communication with neighboring terminals, cross-terminal task migration, and communication channel switching.

[0032] The regional broadcast alarm module is used to send regional alarm information to nearby terminals, management platforms or on-site personnel when the root cause type is a real event, a compound cause or an uncertain cause and the alarm conditions are met.

[0033] The terminal is also configured to communicate with a GIS visualization module to enable causal chain display, multimodal trajectory synchronous playback, and regional situation display.

[0034] The environmental protection and installation support system includes at least one of the following: waterproof and dustproof shell, wind-resistant mounting bracket, lightning protection grounding unit, temperature control and heat dissipation unit, anti-condensation structure, and field fixing structure;

[0035] The multimodal sensing integrated module includes at least three of the following: gas sensor array, water quality sensor array, infrared thermal imaging sensor, visible light camera, microphone array, vibration sensor, and meteorological sensor.

[0036] The photovoltaic modules in the three-level energy system have a power of 100W to 1000W, the energy storage battery capacity is 1kWh to 20kWh, and the supercapacitor capacity is 10F to 500F.

[0037] The BeiDou time and space reference module includes a BeiDou-3 multi-frequency receiver, a 1PPS timing interface, an RTK high-precision positioning antenna, and a crystal oscillator timing module.

[0038] In open environments, the timing accuracy of the BeiDou time and space reference module is better than 20ns;

[0039] Under RTK fixed solution conditions, the positioning accuracy of the BeiDou spatiotemporal reference module is better than 0.1m;

[0040] When the BeiDou signal is briefly lost, the crystal oscillator timing module is used to maintain the continuity of the local clock and to perform time correction after the BeiDou signal is restored.

[0041] The edge computing module includes at least one of an ARM architecture processor, an x86 architecture processor, a GPU acceleration unit, and an FPGA coprocessor, and is used to perform graph neural network fusion, causal reasoning, dynamic programming, and task migration decision-making.

[0042] The multi-mode communication module includes at least two of the following: 4G / 5G cellular communication unit, LoRa long-distance communication unit, Beidou short message communication unit, and WiFi local area network communication unit, and supports dynamic switching of communication channels based on task priority, link quality, communication power consumption, and network status.

[0043] The regional broadcast alarm module includes at least one of the following: an audible and visual alarm unit, a LoRa broadcast unit, a BeiDou short message alarm unit, and a cellular network alarm unit.

[0044] The environmental protection and installation support system is used to ensure the long-term stable operation of the terminal in the field environment. It includes at least one of the following: a waterproof and dustproof shell, a wind-resistant mounting bracket, a lightning protection grounding unit, a temperature control and heat dissipation unit, an anti-condensation structure, and a field fixing structure. The waterproof and dustproof shell protects the sensor interface, power interface, edge computing module, and communication module. The wind-resistant mounting bracket secures the terminal to a pole, shore base, mountain support, or border pole. The lightning protection grounding unit reduces damage to the terminal from lightning strikes and surges. The temperature control and heat dissipation unit maintains the operating temperature of the edge computing module and energy storage battery in high or low temperature environments.

[0045] The three-tiered energy system supplies power to the various power modules of the terminal and provides energy status data to the edge computing module. The photovoltaic modules are used to capture solar energy; the energy storage battery provides continuous power supply; and the supercapacitor handles short-term high-power loads such as camera startup, instantaneous wireless communication transmission, and edge computing acceleration, reducing the instantaneous impact on the energy storage battery. The three-tiered energy system outputs at least one of the following energy status data to the edge computing module: SOC, battery voltage, battery temperature, photovoltaic power generation, and load power.

[0046] The edge computing module includes a processor and a memory. The memory stores computer programs, sensor calibration parameters, propagation delay model parameters, causal graph parameters, task energy consumption tables, and communication strategy tables. When the processor executes the computer program, it realizes multimodal fusion based on the BeiDou spatiotemporal reference, four-layer causal graph root cause diagnosis, task scheduling, hierarchical dynamic programming, and cross-terminal task migration.

[0047] The multi-mode communication module is used to select the communication channel based on task priority, link quality, communication power consumption, and network status. When the root cause type is a real event and the task priority is P1, cellular networks or BeiDou short messages are used for high-priority reporting. When in a low-risk inspection state, LoRa or low-power communication methods are used to transmit summary data. When neighboring terminals are performing task migration or collaborative verification, LoRa, WiFi, or cellular networks are used to complete the transmission of task descriptions, model parameters, and historical data.

[0048] The regional broadcast alarm module is used to issue regional alarms when the root cause type is a real event, a compound cause, or an uncertain cause, and the alarm conditions are met. The regional broadcast alarm module can send alarm information to nearby terminals, management platforms, mobile inspection terminals, or on-site personnel via audible and visual alarms, LoRa broadcasts, BeiDou short messages, cellular networks, or local area networks.

[0049] The GIS visualization module can be deployed on a cloud server, edge gateway, command platform, or mobile terminal. The distributed resource management terminal communicates with the GIS visualization module through the multi-mode communication module to display terminal location, event source reverse location, propagation path, causal chain, task migration path, multi-modal trajectory synchronous playback, and regional situation map.

[0050] The beneficial effects of this invention are as follows:

[0051] (1) Improve the accuracy and reliability of multimodal fusion

[0052] By using BeiDou for unified timing and positioning, corrections can be made for sensor response delays, transmission delays, and physical propagation delays such as sound, heat, water flow, and gas. This allows observations from different modes to be back-referenced to the spatiotemporal basis of real events, thereby improving the consistency of the fusion results.

[0053] (2) Reduce false alarm rate and improve root cause interpretability

[0054] By quantifying the contribution of candidate root causes through a four-layer causal graph and counterfactual intervention intensity (ICS), it is possible to distinguish between real events, abnormal system states or sensor failures, environmental interference, compound causes, and uncertain causes, thus providing an interpretable basis for anomaly handling.

[0055] (3) Implement adaptive task scheduling

[0056] Based on the fusion uncertainty U(t) and root cause type, the sampling frequency, inference depth, communication channel and alarm level are dynamically adjusted to improve the perception intensity under critical events and reduce energy consumption under low-risk or low-uncertainty conditions.

[0057] (4) Extend terminal battery life and ensure critical missions

[0058] By using hierarchical dynamic programming that couples risk, energy, and illumination predictions, task benefits are optimized while meeting SOC safety constraints, avoiding battery over-discharge due to short-term high loads and improving critical event reliability.

[0059] (5) Enhance the robustness of distributed systems

[0060] By using virtual energy pools and task migration decision inequalities, critical tasks of low-power terminals are migrated to nearby terminals with coverage and energy margins, forming a cross-terminal fusion closed loop and reducing the risks caused by single-point energy depletion or coverage blind spots. Attached Figure Description

[0061] Figure 1 This is a schematic diagram of the five-level closed-loop adaptive cooperative control architecture of the present invention. Detailed Implementation

[0062] like Figure 1 As shown, a five-level closed-loop adaptive cooperative control method for field multimodal sensing terminals includes the following five levels:

[0063] The first stage involves multimodal fusion and uncertainty quantification based on the BeiDou spatiotemporal reference, outputting the fusion feature F(t) and fusion uncertainty U(t).

[0064] The second level is based on the counterfactual root cause tracing and intervention intensity quantification of the four-layer causal diagram, outputting the root cause type and intervention intensity ICS;

[0065] The third level is a task scheduling system driven by both uncertainty and root cause type, which outputs a quadruple of sampling rate, inference depth, communication channel, and alarm level.

[0066] The fourth level is a hierarchical dynamic programming approach based on the three-element coupling of risk, energy, and illumination prediction, which outputs a resource allocation scheme.

[0067] Level 5: Based on the virtual energy pool construction and the benefit-normalized cost migration decision inequality, output task migration instructions;

[0068] The five levels form a closed loop through four feedback paths:

[0069] Feedback path ①: U(t) → first-level sampling frequency / modal start / stop / recalibration trigger;

[0070] Feedback Pathway ②: Root Cause Type / ICS → Level 3 Task Priority;

[0071] Feedback Path ③: Level 4 resource allocation scheme → Level 1 modality activation set / Level 2 causal reasoning depth;

[0072] Feedback Path ④: Fifth-level task migration result → Cross-terminal fusion (takeover terminal sensor node access to the original terminal dynamic causal propagation graph G(t)).

[0073] (I) Multimodal fusion and uncertainty quantification based on BeiDou spatiotemporal reference

[0074] The main function of the first level is to perform time correction and event spatiotemporal inference on multimodal sensor data through the unified spatiotemporal reference of BeiDou, construct a dynamic causal propagation graph, output the fusion feature F(t) and fusion uncertainty U(t) after fusion through graph neural network, and adjust the sampling strategy based on the feedback of U(t).

[0075] Step S1.1: Synchronize with BeiDou time and space reference. All sensor sampling times are unified to the UTC time reference through the BeiDou time synchronization module, and positioning information is unified to the WGS-84 coordinate system through the BeiDou RTK module. The timing accuracy is better than 20ns, and the positioning accuracy is better than 0.1m (in open environments or under RTK fixed solution conditions).

[0076] Step S1.2, Spatiotemporal Reverse Deduction of Real Events. For modes with physical propagation delays (sound, heat, gas, water flow), the event source time is deduced by following these steps:

[0077] First, calculate the equivalent time of sensor observation:

[0078] T*obs(m)=Tsample(m)−τresponse(m)−τtransmission(m)

[0079] Where Tsample(m) is the sampling timestamp of sensor m, τresponse(m) is the sensor response delay, τtransmission(m) is the data transmission delay, and T*obs(m) is the equivalent time of sensor observation.

[0080] When the location of the event source or the candidate propagation path is known or can be estimated, calculate the time of occurrence of the event source:

[0081] Tsource(m)=T*obs(m)−τpropagation(m)

[0082] Wherein, τpropagation(m) is the physical propagation delay, which includes at least one of the following: sound propagation delay, thermal diffusion delay, water flow propagation delay, and gas propagation delay.

[0083] Step S1.3, the gas propagation delay is calculated using a three-segment model. The effective wind direction component is introduced:

[0084] veff = vwind·cos(θ−θwind)

[0085] Where vwind is the wind speed, θ is the direction angle from the event source to the sensor, θwind is the wind direction angle, and veff is the effective component of the wind direction.

[0086] When veff > vmin, the advection-dominant model is adopted:

[0087] τgas=d / veff

[0088] Where d is the distance from the event source to the sensor, vmin is the minimum effective wind speed threshold (default is 0.5m / s), and τgas is the gas propagation delay.

[0089] When |veff|≤vmin, the turbulent diffusion-dominated model is adopted:

[0090] τgas=d² / (2Deff)

[0091] Where Deff is the equivalent turbulent diffusion coefficient.

[0092] When veff < −vmin, the headwind attenuation model is used:

[0093] τgas=[d² / (2Deff)]×[1+k|cos(θ−θwind)|]

[0094] Where k is the headwind attenuation coefficient, which is set to 2.0 by default.

[0095] The three-segment model is used to avoid mathematical singularities caused by veff=0 in the calculation of gas propagation delay.

[0096] Step S1.4, the thermal diffusion delay adopts a one-dimensional thermal conduction model:

[0097] τheat = d² / (4αheat)

[0098] Where αheat is the thermal diffusivity and τheat is the thermal diffusivity delay.

[0099] Step S1.5, Calculation of water flow propagation delay. Introducing the effective component of the water flow direction:

[0100] vwater,eff=vwater·cos(θ−θwater)

[0101] Where vwater is the water flow velocity, θwater is the water flow direction angle, and vwater,eff is the effective component of the water flow direction.

[0102] When vwater,eff > vmin, the horizontal flow propagation model is used:

[0103] τwater(d,θ)=d / vwater,eff

[0104] Where τwater represents the water propagation delay.

[0105] When 0 ≤ vwater, eff ≤ vmin, the water dispersion-dominated model is adopted:

[0106] τwater(d)=d² / (2Dwater,eff)

[0107] Where Dwater,eff is the equivalent water body dispersion coefficient.

[0108] When vwater,eff < 0, it is determined to be in the countercurrent direction, and the weight of the candidate propagation edge is reduced. Alternatively, in cases with backflow, tides, or pump station backflow, a dispersion model can be used for calculation.

[0109] τwater(d)=[d² / (2Dwater,eff)]×γreverse

[0110] Wherein, γreverse is the backflow propagation penalty coefficient, which is set to 3.0 by default.

[0111] Step S1.6, Calculation of sound propagation delay:

[0112] τsound=d / vsound

[0113] Where vsound is the speed of sound, which is taken as 343 m / s under standard atmospheric conditions, and τsound is the sound propagation delay.

[0114] Step S1.7, Construction of the dynamic causal propagation graph. With sensor nodes as vertices and physical propagation paths as directed edges, and edge weights τpropagation(m), construct the dynamic causal propagation graph G(t).

[0115] Step S1.8, sensor reliability calculation. The sensor reliability Ri(t) comprehensively considers historical calibration residuals, current SOC, and communication signal strength RSSI, and is limited to the range of 0 to 1 by the clip function:

[0116] Ri(t)=clip[w1·(1−Calibration_errori / Calibration_errorthreshold)+w2·SOCi(t)+w3·(1−|RSSIi−RSSIoptimal| / RSSIrange),0,1]

[0117] Where Calibration_errori is the historical calibration residual of sensor i, Calibration_errorthreshold is the calibration residual threshold, SOCi(t) is the state of charge of the terminal where sensor i is located, RSSIi is the communication signal strength of sensor i, RSSIoptimal is the optimal signal strength, RSSIrange is the signal strength range, w1, w2, and w3 are weighting coefficients, with the default values ​​being w1=0.4, w2=0.3, and w3=0.3. clip(·,0,1) indicates that the result is truncated to the interval between 0 and 1.

[0118] Step S1.9, Adaptive Attention Fusion. An attention mechanism is used to fuse multimodal features, with the attention weights αi adaptively learning the contributions of each sensor:

[0119] αi=exp(score(xi)) / Σjexp(score(xj))

[0120] Where xi is the feature vector of sensor i, xj is the feature vector of sensor j, score(·) is the attention scoring function, and αi is the adaptive attention weight of sensor i.

[0121] The fusion features are:

[0122] F(t) = Σiαi·fi(t)

[0123] Where fi(t) is the feature vector of sensor i at time t, and F(t) is the fused feature.

[0124] Step S1.10, Spatiotemporal consistency score calculation. For the same event, the event source time Tsource(m) derived from different sensors should satisfy spatiotemporal consistency. The spatiotemporal consistency score is calculated as follows:

[0125] Sconsistency=clip[1−Var(Tsource(m)) / Var(Tsource(m))threshold,0,1]

[0126] Where Var(Tsource(m)) is the variance of the event source time estimated by each sensor, and Var(Tsource(m))threshold is the variance threshold, which is set to 1.0s by default. 2 Sconsistency is the spatiotemporal consistency score, and clip(·,0,1) means truncate the result to the interval between 0 and 1.

[0127] Step S1.11, quantify the fusion uncertainty. The fusion uncertainty U(t) comprehensively considers the attention weight variance, spatiotemporal consistency score, and sensor reliability variance:

[0128] U(t)=clip(β1·Var^(αi)+β2·(1−Sconsistency)+β3·Var^(Ri),0,1)

[0129] Where Var^(αi) is the normalized attention weight variance, calculated as Var^(αi) = Var(αi) / Var(αi)max, where Var(αi) is the attention weight variance and Var(αi)max is the maximum value of the attention weight variance; Var^(Ri) is the normalized reliability variance, calculated as Var^(Ri) = Var(Ri) / Var(Ri)max, where Var(Ri) is the sensor reliability variance and Var(Ri)max is the maximum value of the sensor reliability variance; β1, β2, and β3 are weight coefficients, with default values ​​of β1 = 0.3, β2 = 0.4, and β3 = 0.3; clip(·,0,1) truncates the result to the interval between 0 and 1.

[0130] Step S1.12, adaptively adjust the acquisition strategy based on U(t). Adjust the acquisition strategy in stages according to the value range of U(t):

[0131] When U(t) < 0.3, it is determined to be low uncertainty, the sampling frequency is reduced to 50% of the reference frequency, and redundant modes are turned off;

[0132] When 0.3 ≤ U(t) < 0.7, it is determined to be of medium uncertainty, and the reference sampling frequency and modal configuration are maintained.

[0133] When U(t)≥0.7, it is determined to be high uncertainty, the sampling frequency is increased to 200% of the reference frequency, the redundant mode is activated, and the sensor self-calibration process is triggered.

[0134] (II) Counterfactual Root Cause Tracing and Intervention Intensity Quantification Based on a Four-Level Causal Diagram

[0135] Step S2.1, Construction of a four-layer cause-effect graph. Construct a cause-effect graph containing the following four layers:

[0136] Environmental background layer Venv: includes environmental variables such as temperature T, humidity H, wind speed vwind, wind direction θwind, water flow velocity vwater, water flow direction θwater, solar radiation intensity, and water level;

[0137] The root cause layer, Vevent, includes real physical events or external event nodes, such as Fire, Intrusion, Leakage, and Landslide.

[0138] System state layer Vsys includes Sensor_drift, Equipment_fault, State of Charge (SOC), Communication Link State (RSSI), and Computational Load (CPU_load).

[0139] The observation layer Vobs includes observations from various sensors, such as infrared temperature (T_IR), smoke concentration (Smoke), sound intensity (Sound), vibration acceleration (Vibration), pH value, and turbidity.

[0140] Step S2.2, Causal Direction Constraints. The four-layer cause-effect graph strictly follows the following causal directions:

[0141] Venv→Vevent: Environmental context influences the probability of an event occurring;

[0142] Venv→Vobs: Environmental background directly affects observation performance;

[0143] Vevent→Vobs: Real events cause sensor observation anomalies;

[0144] Vsys→Vobs: Abnormal system status or sensor failure leads to abnormal observation;

[0145] Vsys→Vevent: System state affects event handling capabilities or indirectly alters the risk of event evolution.

[0146] The NOTEARS algorithm with hierarchical constraints is used during the learning process to reset the edge weights to zero for edges that violate the above hierarchical direction.

[0147] Step S2.3, Construction of the candidate root cause set. When an anomaly occurs in a certain variable Oi in the observed manifestation layer, trace back its parent node set according to the causal graph to construct a candidate root cause set C={C1,C2,…,Cn}, where Ci may come from Venv, Vevent, or Vsys.

[0148] Step S2.4: Set counterfactual intervention baselines according to variable type. The counterfactual intervention baselines are determined separately according to variable type:

[0149] (a) For binary event variables (Fire, Intrusion, Equipment_fault, etc.): Cibaseline=0 (the event has not occurred);

[0150] (b) For continuous environmental background variables (T, H, vwind, vwater, etc.), automatically select according to the following priorities and automatically degrade according to data availability:

[0151] Priority 1: Median of the same spatial grid over the past 30 days (±1 hour window);

[0152] Priority 2: Historical average of the same type of region in the same season;

[0153] Priority 3: Safety thresholds specified by national or industry standards (e.g., PM2.5 = 35 μg / m³) 3 (Ambient temperature T=20°C)

[0154] Priority 4: Expert prior settings;

[0155] (c) For system state variables (SOC, RSSI, Sensor_health, etc.): Cibaseline=Cinormal (e.g., SOC=80%, RSSI=−70dBm, Sensor_health=1.0).

[0156] (d) For discrete state variables, the baseline is the state with the highest frequency of occurrence in the historical normal state.

[0157] (e) For continuous observation class variables or derived indicators included in the candidate root cause set, the same baseline selection priority is used to determine the counterfactual baseline.

[0158] Step S2.5, Quantify the intervention intensity. For each candidate root cause Ci, calculate its counterfactual intervention intensity ICS:

[0159] ISi=P(E=1|do(Ci=Cianomaly))−P(E=1|do(Ci=Cibaseline))

[0160] Where E represents the abnormal event to be diagnosed, Cianomaly represents the abnormal state or current abnormal value of the candidate root cause, Cibaseline represents the counterfactual baseline determined by the variable type, do(·) represents the causal intervention operation, and P(·) represents the probability.

[0161] Candidate root causes with ICSi < 0 are considered inhibitory factors and do not participate in the positive root cause ranking; the positive root cause ranking uses:

[0162] ICS+i=max(ICSi,0)

[0163] Wherein, ICS+i represents the positive intervention intensity of candidate root cause Ci.

[0164] Step S2.6, Root Cause Type Determination Rules. Sort ICS+i in descending order as follows:

[0165] ICS+1≥ICS+2≥…≥ICS+n

[0166] When the number of positive candidate root causes is less than 2, let ICS+2=0. When the following conditions are met:

[0167] When (ICS+1+ε) / (ICS+2+ε)>1.5 and ICS+1>0.3, it is determined to be a single root cause type corresponding to the level where ICS+1 is located; where ε is a very small positive number to prevent division by zero, and the default value is 10^−6.

[0168] when:

[0169] (ICS+1+ε) / (ICS+2+ε)≤1.5 and ICS+1 and ICS+2 belong to different levels, and:

[0170] When ICS+1>θICS and ICS+2>θICS, it is determined to be a compound cause, and all conditions satisfying this condition are output:

[0171] Candidate root causes whose ICS+i > θICS are selected, where θICS is the output threshold for candidate root causes, which is 0.2 by default. All other cases are considered uncertain and cross-terminal verification is triggered.

[0172] The single root cause type corresponding to the level where ICS+1 is located is determined according to the following rules:

[0173] When the candidate root cause corresponding to ICS+1 belongs to the event root cause layer Vevent, the root cause type is determined to be a real event.

[0174] When the candidate root cause corresponding to ICS+1 belongs to the system state layer Vsys, the root cause type is determined to be system state abnormality or sensor failure.

[0175] When the candidate root cause corresponding to ICS+1 belongs to the environmental background layer Venv, the root cause type is determined to be environmental interference.

[0176] When the candidate root cause corresponding to ICS+1 is simultaneously affected by high intervention intensity variables at at least two levels of the event root cause layer, the environmental background layer, or the system state layer, the root cause type is determined to be a composite cause.

[0177] When the intervention intensity of each candidate root cause fails to reach the determination threshold, or when the ranking of positive intervention intensity is unstable, the root cause type is determined to be uncertain.

[0178] Step S2.7, Root Cause Type and Task Priority Mapping. Determine the task priority P and handling rules based on the root cause type:

[0179] Table 1. Mapping Table of Root Cause Types, Task Priorities, Handling Rules, and Alarm Levels

[0180] Root cause type Priority Disposal rules Alarm Level Real events P1 (highest) Initiate deep inference (L=5) + broadcast alarm + BeiDou short message reporting Red Alert Abnormal system status or sensor malfunction P3 (Middle) Initiating self-calibration, redundant sensor takeover, communication link reselection, or device maintenance prompts. Suppress alarms Environmental interference P4 (Low) Adjust threshold or wait for recovery Suppress alarms Reasons for the combination P2 (Second highest) Parallel verification of multiple hypotheses + cross-terminal collaboration Yellow pending confirmation uncertain P2 Cross-device verification Yellow pending confirmation

[0181] The following are examples of counterfactual root cause analysis in fire hazard detection:

[0182] Deployment scenario: A forest fire prevention terminal detected abnormal infrared temperature (T_IR=85°C) and abnormal smoke concentration (Smoke=120ppm) at 15:23:40.

[0183] Candidate root cause set: C={Fire,Direct_sunlight,Sensor_drift}.

[0184] Counterfactual intervention baseline:

[0185] Firebaseline=0 (No fire)

[0186] Direct_sunlightbaseline = Historical median for the same period (15:00-16:00, sunny, Direct_sunlight=0.75)

[0187] Sensor_driftbaseline=0 (Sensor is working properly)

[0188] Intervention intensity calculation:

[0189] ICSFire=P(E=1|do(Fire=1))−P(E=1|do(Fire=0))=0.92−0.05=0.87

[0190] ICSDirect_sunlight=P(E=1|do(Direct_sunlight=1.0))−P(E=1|do(Direct_sunlight=0.75))=0.18−0.05=0.13

[0191] ICSSensor_drift=P(E=1|do(Sensor_drift=1))−P(E=1|do(Sensor_drift=0))=0.22−0.05=0.17

[0192] Root cause determination:

[0193] (ICS+Fire+ε) / (ICS+Sensor_drift+ε)=0.87 / 0.17≈5.1>1.5

[0194] And ICS + Fire = 0.87 > 0.3

[0195] If the event is determined to be real, P1 priority is triggered, and deep inference (L=5), broadcast alarm, and BeiDou short message reporting are initiated.

[0196] In contrast, existing black-box models output abnormalities but lack root cause types, leading to false positives or false negatives. This invention uses counterfactual intervention intensity quantization to clearly distinguish between Fire's ICS (0.87) and Direct_sunlight's ICS (0.13), thus avoiding false positives.

[0197] (III) Task scheduling driven by both uncertainty and root cause type

[0198] Step S3.1, Uncertainty-driven acquisition strategy adjustment. The acquisition strategy is adjusted in stages according to the range of values ​​of U(t) (as described in step S1.12).

[0199] Step S3.2, root cause type-driven task priority adjustment (described in step S2.7).

[0200] Step S3.3, task list generation. Output a task scheduling list of {sampling rate, inference depth, communication channel, alarm level} quadruple, and pass it to the fourth level.

[0201] (iv) Hierarchical dynamic programming of the risk-energy-lighting prediction ternary coupling

[0202] Step S4.1, define the state variables. This includes: the current state of charge SOC(t); the predicted solar power for the next Δ hours Psolar(t+Δ) (based on historical solar radiation curves and short-term cloud optical thickness prediction); the grid risk level Risk(g,t) (from the weighted aggregation of the second-level ICS); and the candidate task set and its energy consumption Ek and reward rk vector Task_set(t).

[0203] Step S4.2, hierarchical structure. The upper layer is hourly with a 24-hour field of view, determining the hourly energy budget B(h); the lower layer is minute-level with a 60-minute field of view, determining the specific task execution sequence under the constraint of B(h).

[0204] Step S4.3, ternary coupling objective function. The objective function and constraints are as follows:

[0205] maxΣt[Risk(t)×Reward(Task(t))]−λ×Penalty(SOCminviolation)

[0206] Constraints:

[0207] SOC(t+1)=SOC(t)+[η×Psolar(t)×Δt−E(Task(t))] / Cbat

[0208] SOCmin≤SOC(t)≤SOCmax,∀t

[0209] Σt∈hE(Task(t))≤B(h),∀h

[0210] Where Risk(t) is the risk level at time t, Reward(Task(t)) is the reward of Task(t), λ is the penalty coefficient, Penalty(SOCminviolation) is the penalty for defaulting on the minimum SOC value, η is the photovoltaic conversion efficiency, Psolar(t) is the solar power at time t, Δt is the time step, E(Task(t)) is the energy consumption of Task(t), Cbat is the battery capacity, SOCmin is the minimum state of charge, SOCmax is the maximum state of charge, and B(h) is the energy budget for hour h.

[0211] Penalty (SOCminviolation) is defined as follows:

[0212] Penalty(SOCminviolation)=Σtmax(SOCmin−SOC(t),0)

[0213] This penalty term is used to suppress low SOC schemes during the evaluation phase of dynamic programming candidate actions; the final output resource allocation scheme still needs to satisfy the hard constraint that SOCmin≤SOC(t)≤SOCmax.

[0214] Step S4.4, Dynamic Programming Solution. The above optimization problem is solved using a value iteration algorithm, and a resource allocation scheme is output.

[0215] The following is an example of migration decision calculation:

[0216] Deployment scenario: Environmental monitoring terminal #2 currently has a SOC of 15% and predicts insufficient light in the next 6 hours, making it unable to support gas sampling and inference tasks; neighboring terminal #4 currently has a SOC of 68% and a coverage overlap rate of 35%.

[0217] Migration decision calculation:

[0218] Task T to be migrated: Gas sampling + inference, energy consumption per run is 0.8Wh, frequency is 1 time / 10min, i.e., 6 times per hour. The energy saved per hour by the original terminal #2 is:

[0219] Esave = 0.8Wh × 6 = 4.8Wh / h

[0220] To facilitate a unified comparison of benefits, the range improvement was normalized using a reference energy-saving scale of Esave,ref=12Wh / h:

[0221] ΔSurvival(#2)=min(Esave / Esave,ref,1)=4.8 / 12=0.40

[0222] Where ΔSurvival(#2) represents the improvement in battery life for terminal #2, and Esave,ref is the reference energy-saving standard.

[0223] The risk weights are:

[0224] Riskweight(T) = 1.0

[0225] Where Riskweight(T) is the risk weight of task T.

[0226] Coverage gain is calculated jointly using coverage overlap rate and available energy margin at the takeover terminal:

[0227] Coveragegain(#4)=0.35×(SOC(#4)−SOCreserve) / (1−SOCreserve)

[0228] Wherein, SOC(#4) is the state of charge of terminal #4, SOCreserve is the reserved state of charge, which is 0.25 by default, and Coveragegain(#4) is the coverage gain of terminal #4.

[0229] therefore:

[0230] Coveragegain(#4)=0.35×(0.68−0.25) / (1−0.25)=0.35×0.43 / 0.75=0.20

[0231] If we take weights w1=0.4, w2=0.4, and w3=0.2, then:

[0232] Gain=w1×ΔSurvival(#2)+w2×Riskweight(T)+w3×Coveragegain(#4)

[0233] =0.4×0.40+0.4×1.0+0.2×0.20

[0234] =0.16+0.40+0.04

[0235] =0.60

[0236] Gain represents migration gains.

[0237] Costnorm calculation:

[0238] Ecomm+Erelocate=1.5J,Eref=10J

[0239] Latency = 0.3s, Ttol = 1.0s

[0240] Where Ecomm is the communication energy consumption, Erelocate is the migration energy consumption, Eref is the reference energy consumption, Latency is the migration delay, and Ttol is the tolerable delay.

[0241] If we take μ1 = 0.6 and μ2 = 0.4, then:

[0242] Costnorm=μ1×(Ecomm+Erelocate) / Eref+μ2×Latency / Ttol

[0243] =0.6×1.5 / 10+0.4×0.3 / 1.0

[0244] =0.09+0.12

[0245] =0.21

[0246] Where Costnorm is the normalization cost.

[0247] therefore:

[0248] Gain−Costnorm=0.60−0.21=0.39>Θmigrate=0.3

[0249] Where Θmigrate is the migration decision threshold.

[0250] Decision: Migration successful. Terminal #2 initiates a task migration request to Terminal #4.

[0251] (v) Task migration and cross-terminal collaboration based on virtual energy pool

[0252] Step S5.1, Virtual Energy Pool Construction. When Terminal A's SOC falls below a safety threshold or is predicted to be unable to complete critical tasks within the next Δ hours, it broadcasts an energy request to neighboring terminals to construct a virtual energy pool:

[0253] Epool=Σi∈Nmax(SOCi−SOCreserve,0)×Cbat,i×Coverageoverlap(A,i)

[0254] Where N is the set of neighboring terminals, SOCi is the state of charge of terminal i, SOCreserve is the reserved state of charge, Cbat,i is the battery capacity of terminal i, Coverageoverlap(A,i) is the coverage overlap rate between terminal A and terminal i, Epool is the virtual energy pool capacity, and max(SOCi−SOCreserve,0) is used to prevent terminals with a state of charge lower than the reserved state of charge from contributing negative values ​​to the virtual energy pool.

[0255] Step S5.2, Calculation of migration gains. Migration gains (Gain) comprehensively consider improvements in battery life, risk weighting, and coverage gain:

[0256] Gain=w1×ΔSurvival(A)+w2×Riskweight(T)+w3×Coveragegain(B)

[0257] in:

[0258] ΔSurvival(A)=clip(Esave / Esave,ref,0,1)

[0259] Coveragegain(B)=Coverageoverlap(A,B)×clip((SOCB−SOCreserve) / (1−SOCreserve),0,1)

[0260] Where ΔSurvival(A) represents the battery life improvement of the original terminal A, Riskweight(T) represents the risk weight of the task to be migrated T, Coveragegain(B) represents the coverage gain of the takeover terminal B, and w1, w2, and w3 are weight coefficients, with the default values ​​being w1=0.4, w2=0.4, and w3=0.2. clip(·,0,1) truncates the result to the interval between 0 and 1. Through the above normalization process, when the SOC of the takeover terminal B is not higher than SOCreserve, its coverage gain is 0, thus preventing low-battery terminals from being selected as the takeover terminal.

[0261] Step S5.3, normalized cost calculation. The normalized cost (Costnorm) comprehensively considers communication energy consumption and migration delay:

[0262] Costnorm=μ1×(Ecomm+Erelocate) / Eref+μ2×Latency / Ttol

[0263] Where Ecomm is the communication energy consumption, Erelocate is the migration energy consumption, Eref is the reference energy consumption, Latency is the migration delay, Ttol is the tolerable delay, and μ1 and μ2 are weighting coefficients, with the default values ​​being μ1=0.6 and μ2=0.4.

[0264] Step S5.4, migration decision inequality. When satisfied:

[0265] When Gain−Costnorm>Θmigrate, the migration is considered successful, where Θmigrate is the migration decision threshold, which is 0.3 by default.

[0266] Step S5.5, Three-Phase Handshake Protocol. The migration process uses a three-phase handshake protocol:

[0267] Phase 1: The original terminal A sends a migration request to the candidate takeover terminal B, which includes a task description, energy consumption estimate, and latency requirements;

[0268] Phase 2: Candidate takeover terminal B assesses its own SOC, computing resources, and communication bandwidth, and returns an accept / reject response;

[0269] Phase 3: After receiving the acceptance response, the original terminal A transmits the task code, model parameters, and historical data to the takeover terminal B. After the takeover terminal B confirms the successful takeover, the original terminal A stops executing the task.

[0270] Step S5.6, Cross-terminal fusion closed loop. The takeover terminal B adds its sensor nodes to the dynamic causal propagation graph G(t) of the original terminal A, forming a cross-terminal fusion closed loop and realizing the access of virtual sensor nodes.

[0271] Examples of integrated watershed water environment monitoring are as follows:

[0272] Deployment Scenario: Six terminals of this invention are deployed along a 2km stretch of a river in a certain basin. Each terminal integrates five types of water quality sensors: pH, DO (dissolved oxygen), turbidity, conductivity, and water temperature. Event: At 14:23:00 upstream, industrial discharge occurred, causing an abnormal drop in the pH of the local water body to 5.2.

[0273] Terminal #3 is approximately 210m from the suspected discharge point, and Terminal #5 is approximately 250m downstream of Terminal #3. The real-time water flow velocity is estimated using water level gauges and a flow velocity model as follows:

[0274] vwater=0.4m / s

[0275] The direction of water flow is basically consistent with the direction of the line connecting terminals #3 and #5.

[0276] Terminal #3 detected a pH decrease at 14:31:42. According to the water flow propagation model, the propagation delay from the discharge point to terminal #3 is:

[0277] τwater,3=210 / 0.4=525s

[0278] After combining sensor response and transmission delay correction, the time of the source event was deduced to be approximately 14:22:57, which is highly consistent with the actual emission time of 14:23:00.

[0279] Terminal #5 detected an anomaly in turbidity at 14:42:18. The propagation delay between Terminal #3 and Terminal #5 is:

[0280] τwater,35=250 / 0.4=625s

[0281] The time difference between the detection time of terminal #3 (14:31:42) and the detection time of terminal #5 (14:42:18) is 636s, which is close to the propagation model estimate of 625s.

[0282] Existing one-way fusion methods do not incorporate water flow propagation delay for back-calculation, which can easily lead to misjudging the pH anomaly at terminal #3 and the turbidity anomaly at terminal #5 as two independent anomalies or as a large-scale regional pollution event. This invention uses a unified spatiotemporal reference based on BeiDou and water flow propagation delay modeling to identify the two as temporal propagation manifestations of the same upstream discharge event.

[0283] In this example, the existing method outputs a fusion uncertainty U=0.71, corresponding to a fusion confidence 1−U=0.29; the method of the present invention outputs a fusion uncertainty U=0.13, corresponding to a fusion confidence improved to 0.87, and triggers upstream emission root cause diagnosis and downstream terminal early warning sampling strategies.

[0284] The ablation experiment is as follows:

[0285] In a simulation and semi-physical playback test of one embodiment, 30 consecutive days of operation data from 12 distributed resource management terminals in a forest fire prevention scenario were selected. Under the same initial SOC, the same illumination playback conditions, the same event sequence, and the same communication environment parameters, the complete scheme and the degradation scheme with different feedback paths closed were compared and evaluated.

[0286] The configurations are compared as follows:

[0287] Complete solution: Implements a five-level closed-loop control process and four feedback paths;

[0288] Closed path ①: U(t) is not fed back to the sampling strategy, and the sampling frequency and mode start / stop are no longer adaptively adjusted according to the fusion uncertainty;

[0289] Close Pathway ②: Root cause type is not fed back to task priority, and task priority is no longer adjusted based on real events, environmental interference, abnormal system status, or compound causes;

[0290] Closed Pathway ③: Resource allocation schemes are not fed back to modality activation and inference depth, and the first-level modality activation set and the second-level causal inference depth are no longer constrained by energy planning results;

[0291] Close the path ④: The task migration result is not fed back to the cross-terminal fusion, and the sensor node of the takeover terminal is no longer connected to the dynamic causal propagation graph G(t) of the original terminal;

[0292] Degradation solution: Close all feedback paths and degrade to a one-way data collection, analysis, and alarm architecture.

[0293] The evaluation indicators are defined as follows:

[0294] False alarm rate = (Number of false alarms / Total number of alarms) × 100%

[0295] Critical incident assurance rate = (Number of critical incidents effectively discovered, confirmed, and reported / Actual number of critical incidents) × 100%

[0296] Average battery life = the average effective working time of each terminal from the start of operation until it reaches SOCmin or is unable to complete the baseline critical task under the same initial SOC and the same light playback conditions.

[0297] The experimental results are shown in the table below:

[0298] Table 2 Comparison of ablation experiment results for four feedback pathways in forest fire prevention scenarios.

[0299] Configuration False alarm rate (%) Critical event coverage rate (%) Average battery life Complete solution 3.2 98.7 187 hours and 12 minutes Close the passage ① 14.2 97.1 182 hours and 24 minutes Close the passage ② 28.1 84.0 180 hours Close the passage ③ 4.1 97.8 117 hours and 36 minutes Close the passage ④ 3.8 92.0 108 hours Degradation scheme 32.4 71.0 100 hours and 48 minutes

[0300] As shown in Table 2, under the test conditions of this embodiment, after closing feedback path ①, the false alarm rate increased from 3.2% to 14.2% because the fusion uncertainty U(t) could not drive the sampling strategy adjustment in reverse; after closing feedback path ②, the false alarm rate increased to 28.1% and the critical event guarantee rate decreased to 84.0% because the root cause type could not drive the task priority adjustment in reverse; after closing feedback path ③, the average battery life decreased from 187 hours and 12 minutes to 117 hours and 36 minutes because the resource allocation scheme could not constrain modality activation and inference depth; after closing feedback path ④, the critical event guarantee rate decreased to 92.0% and the average battery life decreased to 108 hours because the task migration results could not be integrated into cross-terminal fusion; when all feedback paths were closed, the system degenerated into a unidirectional architecture, the false alarm rate increased to 32.4%, the critical event guarantee rate decreased to 71.0%, and the average battery life decreased to 100 hours and 48 minutes.

[0301] The results above indicate that, under these test conditions, the four feedback pathways contribute to reducing false alarm rates, improving critical event reliability, and extending battery life, respectively.

[0302] In one embodiment, the distributed resource management terminal includes an environmental protection and installation support system, a three-level energy system, a multimodal sensing integration module, a BeiDou spatiotemporal reference module, an edge computing module, a multimode communication module, a regional broadcast alarm module, as well as sensor interfaces, power management interfaces, and communication interfaces.

[0303] The environmental protection and installation support system includes a waterproof and dustproof shell, a wind-resistant mounting bracket, a lightning protection grounding unit, a temperature control and heat dissipation unit, and an anti-condensation structure. The waterproof and dustproof shell houses the edge computing module, energy storage battery, power management board, and communication module; the wind-resistant mounting bracket allows the terminal to be installed on forest poles, riverbanks, border posts, mining area supports, or oil and gas pipeline inspection points; the lightning protection grounding unit directs lightning surge current into the grounding system; and the temperature control and heat dissipation unit maintains the internal temperature of the terminal through air cooling, heat-conducting plates, heat dissipation fins, or a heating film.

[0304] The three-tiered energy system comprises photovoltaic modules, energy storage batteries, and supercapacitors. The photovoltaic modules power the terminals and charge the energy storage batteries during the day; the energy storage batteries power the terminals at night, on cloudy or rainy days, or when sunlight is insufficient; and the supercapacitors support high-power tasks such as communication transmission, camera activation, infrared thermal imaging activation, and GPU inference. The three-tiered energy system outputs SOC, battery voltage, battery temperature, photovoltaic power, load power, and charge / discharge status to the edge computing module via a power management interface.

[0305] The multimodal sensing integration module is connected to the edge computing module via a sensor interface and includes at least three of the following: a gas sensor array, a water quality sensor array, an infrared thermal imaging sensor, a visible light camera, a microphone array, a vibration sensor, and a meteorological sensor. Data collected by different sensors are all provided with unified timestamps and spatial coordinates by the BeiDou spatiotemporal reference module, forming multimodal observation data that can be used for spatiotemporal back-calculation of real-world events.

[0306] The BeiDou spatiotemporal reference module includes a BeiDou-3 multi-frequency receiver, a 1PPS timing interface, an RTK high-precision positioning antenna, and a crystal oscillator timekeeping module. The BeiDou-3 multi-frequency receiver receives BeiDou satellite signals; the 1PPS timing interface provides a unified second pulse to various sensors and edge computing modules; the RTK high-precision positioning antenna acquires the terminal's spatial location; and the crystal oscillator timekeeping module maintains local clock continuity during short-term BeiDou signal loss and performs time correction after BeiDou signal recovery.

[0307] The edge computing module includes a processor and a memory. The processor can be at least one of an ARM processor, an x86 processor, a GPU acceleration unit, or an FPGA coprocessor. The memory stores computer programs, sensor calibration parameters, propagation delay model parameters, causal graph structure parameters, task energy consumption tables, communication strategy tables, and historical observation data. When the processor executes the computer program, it performs the following functions: sensor observation time correction based on the BeiDou spatiotemporal reference, spatiotemporal inversion of real events, construction of dynamic causal propagation graphs, graph neural network fusion, calculation of fusion uncertainty, root cause diagnosis of four-layer causal graphs, task scheduling, hierarchical dynamic programming, construction of virtual energy pools, and cross-terminal task migration.

[0308] The multi-mode communication module includes at least two of the following: a 4G / 5G cellular communication unit, a LoRa long-range communication unit, a BeiDou short message communication unit, and a WiFi local area network communication unit. The edge computing module controls the multi-mode communication module to switch communication channels based on task priority, communication link quality, communication power consumption, and network status. For example, when the root cause type is a real event and the communication link is available, 4G / 5G cellular communication is preferentially used to upload images, videos, or fused features; when the cellular network is unavailable, BeiDou short message service is used to send compressed alarm information; when performing collaborative verification with neighboring terminals or task migration, LoRa or WiFi is used to transmit task descriptions, model parameters, and historical data.

[0309] The regional broadcast alarm module is used to execute local or regional alarms based on the alarm level output at the third level. When the alarm level is red, the regional broadcast alarm module sends an alarm to nearby terminals and the management platform via audible and visual alarms, LoRa broadcast, BeiDou short message, or cellular network; when the alarm level is yellow and pending confirmation, the regional broadcast alarm module sends a collaborative verification request to nearby terminals; when the alarm is suppressed, the regional broadcast alarm module does not issue a serious alarm externally, but only records local logs or sends maintenance information.

[0310] The embodiments described above are merely examples of several implementations of the present invention, and while the descriptions are relatively specific and detailed, they should not be construed as limiting the scope of the present invention. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of the present invention, and these modifications and improvements all fall within the scope of protection of the present invention.

Claims

1. A closed-loop adaptive cooperative control method based on a distributed resource management terminal with multimodal sensing, characterized in that, It includes a five-level closed-loop control process, which is formed by four explicit feedback paths: The first stage involves performing sensor observation time correction, real event spatiotemporal inference, and dynamic causal propagation graph construction on multimodal sensor data based on the BeiDou spatiotemporal reference. After fusion by a graph neural network, the fusion feature F(t) and fusion uncertainty U(t) are output. The U(t) is adjusted by the reverse driving sampling strategy through feedback path ①. The second level is based on a four-layer causal diagram consisting of environmental background, event root cause, observation performance, and system state. Counterfactual intervention baselines are set according to variable type, the intervention intensity (ICS) of each candidate root cause is quantified, and the root cause type is determined based on the ICS ranking results. The root cause type drives the task priority adjustment in reverse through feedback path ②. The third level involves adjusting the sampling frequency, inference depth, communication channel, and alarm level according to preset rules based on U(t) and the root cause type, and outputting a task scheduling list. The fourth level uses the objective function composed of risk-weighted terms and battery protection penalty terms to perform hourly-minute hierarchical dynamic programming based on the SOC state equation normalized by battery capacity, outputs a resource allocation scheme, and feeds it back to the first and second levels through feedback path ③. At the fifth level, based on the virtual energy pool construction and the benefit-normalized cost migration decision inequality, the task migration instruction is output. After the migration is completed, a cross-terminal fusion closed loop is formed through feedback path ④.

2. The method according to claim 1, characterized in that, The first level of real-time event backward inference includes: The equivalent time of sensor observation is obtained by the following formula: T*obs(m)=Tsample(m)−τresponse(m)−τtransmission(m) When the location of the event source or the candidate propagation path is known or can be estimated, the time of occurrence of the event source is obtained by the following formula: Tsource(m)=T*obs(m)−τpropagation(m) Where Tsample(m) is the sampling timestamp of sensor m, τresponse(m) is the sensor response delay, τtransmission(m) is the data transmission delay, T*obs(m) is the sensor observation equivalent time, τpropagation(m) is the physical propagation delay, and Tsource(m) is the event source occurrence time; the τpropagation(m) includes at least one of the following: sound propagation delay, thermal diffusion delay, water flow propagation delay, and gas propagation delay. The gas propagation delay is calculated using a three-segment model, introducing the effective wind direction component veff = vwind·cos(θ−θwind). When veff > vmin, τgas = d / veff; when |veff| ≤ vmin, τgas = d. 2 / (2Deff), when veff<−vmin τgas=[d 2 / (2Deff)]×[1+k|cos(θ−θwind)|], where vwind is the wind speed, θ is the direction angle from the event source to the sensor, θwind is the wind direction angle, vmin is the minimum effective wind speed threshold, d is the distance from the event source to the sensor, Deff is the equivalent turbulence diffusion coefficient, and k is the headwind attenuation coefficient. The fusion uncertainty is calculated using the following formula: U(t)=clip(β1·Var^(αi)+β2·(1−Sconsistency)+β3·Var^(Ri),0,1) Where αi is the adaptive attention weight of each sensor, Var^(αi) is the normalized attention weight variance, Sconsistency is the spatiotemporal consistency score, Ri is the sensor reliability, Var^(Ri) is the normalized reliability variance, β1, β2, and β3 are weight coefficients, and clip(·,0,1) indicates that the result is truncated to the interval between 0 and 1.

3. The method according to claim 1, characterized in that, The four-layer causal graph in the second level strictly follows the following causal directions: Venv→Vevent, Venv→Vobs, Venent→Vobs, Vsys→Vobs, Vsys→Vevent Among them, Venv is the environmental background layer, Venvent is the event root cause layer, Vobs is the observation manifestation layer, and Vsys is the system state layer; the NOTEARS algorithm with hierarchical constraints is used in the learning process, and the edge weights that violate the hierarchical direction are reset to zero; The counterfactual intervention baseline is determined according to the variable type: for binary event variables, it is set to 0; for continuous environmental background variables, the priority is automatically selected according to the median of the same period in the past 30 days in the same spatial grid > the historical mean of the same type of region in the same season > the national or industry standard threshold > the expert prior value, and automatically downgraded according to data availability; for system state variables, the median of the normal working interval is taken; for discrete state variables, the state with the highest frequency of occurrence in the historical normal state is taken.

4. The method according to claim 1, characterized in that, The second-level root cause type determination rule is as follows: The positive intervention intensity ICS+i=max(ICSi,0) is sorted in descending order as ICS+1≥ICS+2≥…≥ICS+n; When the number of positive candidate root causes is less than 2, let ICS+2=0; When (ICS+1+ε) / (ICS+2+ε)>1.5 and ICS+1>0.3, it is determined to be a single type corresponding to the level where ICS+1 is located; When (ICS+1+ε) / (ICS+2+ε)≤1.5 and ICS+1 and ICS+2 belong to different levels, and both ICS+1 and ICS+2 exceed the preset threshold, it is determined to be a compound cause, and multiple candidate root causes with intervention intensity exceeding the preset threshold are output. In other cases, the condition is determined to be uncertain and cross-device verification is triggered. Where ICSi is the intervention intensity of candidate root cause Ci, and ε is a very small positive number to prevent division by zero; The third level determines the handling rules based on the root cause type: when the root cause type is a real event, deep inference and broadcast alarms are initiated; when the root cause type is an abnormal system state or sensor failure, self-calibration and redundant sensor takeover are initiated and alarms are suppressed; when the root cause type is environmental interference, the threshold is adjusted or recovery is waited for and alarms are suppressed; when the root cause type is a composite cause or uncertain, cross-terminal verification is triggered.

5. The method according to claim 1, characterized in that, The fourth-level objective function is: maxΣt[Risk(t)·Reward(Task(t))]−λ·Penalty(SOCminviolation) The constraints include: The SOC state equation is SOC(t+1)=SOC(t)+[η·Psolar(t)·Δt−E(Task(t))] / Cbat, the SOC safe interval constraint is SOCmin≤SOC(t)≤SOCmax, and the hourly energy budget constraint is Σt∈hE(Task(t))≤B(h). Where Risk(t) is the risk level at time t, Reward(Task(t)) is the task reward, λ is the penalty coefficient, η is the photovoltaic conversion efficiency, Psolar(t) is the light power at time t, Δt is the time step, E(Task(t)) is the task energy consumption, Cbat is the battery capacity, and B(h) is the energy budget for hour h. The Penalty (SOCminviolation) is defined as Σtmax(SOCmin−SOC(t),0), which is used to suppress low SOC schemes during the candidate action evaluation stage of dynamic programming; the final output resource allocation scheme satisfies the hard constraint that SOCmin≤SOC(t)≤SOCmax.

6. The method according to claim 1, characterized in that, The fifth-level migration decision inequality is: Gain−Costnorm>Θmigrate in: Gain=w1×ΔSurvival(A)+w2×Riskweight(T)+w3×Coveragegain(B) Costnorm=μ1×(Ecomm+Erelocate) / Eref+μ2×Latency / Ttol ΔSurvival(A)=clip(Esave / Esave,ref,0,1) Coveragegain(B)=Coverageoverlap(A,B)×clip((SOCB−SOCreserve) / (1−SOCreserve),0,1) Where Gain is the migration gain, Costnorm is the normalized cost, Θmigrate is the migration decision threshold, ΔSurvival(A) is the endurance improvement of the original terminal A, Riskweight(T) is the risk weight of the task to be migrated T, Coveragegain(B) is the coverage gain of the takeover terminal B, w1, w2, w3 are the gain weight coefficients, Econm is the communication energy consumption, Erelocate is the migration energy consumption, Eref is the reference energy consumption, Latency is the migration latency, Ttol is the tolerable latency, μ1, μ2 are the cost weight coefficients, Esave is the energy that the original terminal A can save per hour, Esave,ref is the reference energy saving scale, Coverageoverlap(A,B) is the coverage overlap rate between terminal A and terminal B, SOCB is the state of charge of the takeover terminal B, SOCreserve is the reserved state of charge, and clip(·,0,1) indicates that the result is truncated to the interval between 0 and 1. After the migration is completed, the sensor node that takes over terminal B is connected to the dynamic causal propagation graph G(t) of the original terminal A, forming a cross-terminal fusion closed loop.

7. A distributed resource management terminal based on multimodal perception, characterized in that, It includes an environmental protection and installation support system, a three-level energy system, a multimodal sensing integration module, a BeiDou spatiotemporal reference module, an edge computing module, a multimode communication module, and a regional broadcast alarm module; The three-level energy system includes photovoltaic modules, energy storage batteries and supercapacitors, which are used to supply power to the distributed resource management terminal and provide the edge computing module with state of charge (SOC), power generation and energy storage status data. The BeiDou time and space reference module provides a unified time synchronization, positioning and timekeeping reference for all sensors; The edge computing module includes a processor and a memory, wherein the memory stores a computer program, and the processor executes the computer program to implement the method according to any one of claims 1 to 6; The multi-mode communication module is used to perform event reporting, communication with neighboring terminals, cross-terminal task migration, and communication channel switching. The regional broadcast alarm module is used to send regional alarm information to nearby terminals, management platforms or on-site personnel when the root cause type is a real event, a compound cause or an uncertain cause and the alarm conditions are met. The terminal is also configured to communicate with a GIS visualization module to enable causal chain display, multimodal trajectory synchronous playback, and regional situation display.

8. The terminal according to claim 7, characterized in that, The environmental protection and installation support system includes at least one of the following: waterproof and dustproof shell, wind-resistant mounting bracket, lightning protection grounding unit, temperature control and heat dissipation unit, anti-condensation structure, and field fixing structure; The multimodal sensing integrated module includes at least three of the following: gas sensor array, water quality sensor array, infrared thermal imaging sensor, visible light camera, microphone array, vibration sensor, and meteorological sensor. The photovoltaic modules in the three-level energy system have a power of 100W to 1000W, the energy storage battery capacity is 1kWh to 20kWh, and the supercapacitor capacity is 10F to 500F.

9. The terminal according to claim 7, characterized in that, The BeiDou time and space reference module includes a BeiDou-3 multi-frequency receiver, a 1PPS timing interface, an RTK high-precision positioning antenna, and a crystal oscillator timing module. In open environments, the timing accuracy of the BeiDou time and space reference module is better than 20ns; Under RTK fixed solution conditions, the positioning accuracy of the BeiDou spatiotemporal reference module is better than 0.1m; When the BeiDou signal is briefly lost, the crystal oscillator timing module is used to maintain the continuity of the local clock and to perform time correction after the BeiDou signal is restored.

10. The terminal according to claim 7, characterized in that, The edge computing module includes at least one of an ARM architecture processor, an x86 architecture processor, a GPU acceleration unit, and an FPGA coprocessor, and is used to perform graph neural network fusion, causal reasoning, dynamic programming, and task migration decision-making. The multi-mode communication module includes at least two of the following: 4G / 5G cellular communication unit, LoRa long-distance communication unit, Beidou short message communication unit, and WiFi local area network communication unit, and supports dynamic switching of communication channels based on task priority, link quality, communication power consumption, and network status. The regional broadcast alarm module includes at least one of the following: an audible and visual alarm unit, a LoRa broadcast unit, a BeiDou short message alarm unit, and a cellular network alarm unit.

Images