Power transmission line condition monitoring system based on smart sensors

Abstract

The application discloses a power transmission line state monitoring system based on an intelligent sensor, which establishes an energy boundary constraint matrix by converting the physical energy levels of inductive energy current, battery temperature and pressure state and vibration envelope signals, maps the residual energy constraint to the maximum available computing power upper limit of the end side reasoning network, carries out multi-objective exit node optimization in combination with the inherent physical complexity of the input dynamics tensor, so that the algorithm adaptively cuts the network reasoning depth according to the signal feature saliency, and on this basis, generates an optimized state report suitable for narrowband communication through confidence degradation evaluation and redundant field compression coding. In this way, the key transient detection and monitoring blindness problems caused by energy and computing power imbalance of the power transmission line intelligent sensor under the extreme coupling working condition of low current energy taking and microclimate mutation are solved, and the effective monitoring life cycle of the equipment under the passive supply state is significantly extended.

Description

Technical Field

[0001] This application relates to the field of intelligent monitoring, and more specifically, to a power transmission line condition monitoring system based on intelligent sensors. Background Technology

[0002] As the core carrier of energy transmission in the power system, the operating status of transmission lines directly affects the security and continuity of power supply. Under the long-term influence of complex terrain and variable weather conditions, transmission lines face multiple threats, including increasing icing, wind-induced galloping, and fatigue loosening of hardware. Any abnormal condition that is not detected in time can evolve into catastrophic accidents such as line breaks and tower collapses, causing large-scale power outages and huge economic losses. Therefore, constructing an online sensing system capable of real-time and accurate monitoring of the operating status of transmission lines is of significant engineering importance for ensuring the safe operation of the power grid and improving operational efficiency.

[0003] Existing power transmission line condition monitoring solutions typically employ a technical architecture that involves installing sensor nodes on conductors or towers to periodically collect physical quantities such as vibration, tilt, and temperature, and then transmitting these data wirelessly to a cloud-based master station for centralized analysis. Some solutions introduce edge inference capabilities based on neural networks at the end-point, attempting to perform preliminary condition classification and anomaly diagnosis locally on the sensors to reduce communication transmission load and shorten alarm response latency. However, such solutions have revealed significant technical shortcomings in actual deployment: in extreme coupled scenarios where power transmission lines operate under low current-carrying conditions and simultaneously encounter micro-meteorological changes (such as a sudden increase in icing or a strong wind), the energy margin of sensors relying on inductive energy harvesting is extremely limited. Existing systems lack a coordinated scheduling mechanism for energy budget and inference computing power, making it impossible to support high-frequency continuous sampling and feature extraction calculations of complex AI models at the end-point under limited energy constraints. This results in severe missed detections and data gaps during critical transient processes such as the initial stage of conductor galloping and the sudden increase in icing. Meanwhile, existing edge-side inference architectures generally adopt a fixed-depth network execution strategy, which fails to dynamically adjust the inference level according to the physical complexity of the input signal. When faced with strong vibrations with significant characteristics, they still blindly call deep computing resources, resulting in unnecessary consumption of limited power during disasters. Ultimately, this leads to the sensors going completely down in the later stages of a disaster when continuous monitoring is most needed, due to premature battery depletion, causing serious consequences of large-scale monitoring blindness.

[0004] Therefore, there is a need for an optimized transmission line condition monitoring system based on smart sensors. Summary of the Invention

[0005] To address the aforementioned technical problems, this application provides a power transmission line condition monitoring system based on intelligent sensors.

[0006] According to one aspect of this application, a power transmission line condition monitoring system based on intelligent sensors is provided, comprising: The state perception and trigger judgment module is used to perform physical energy level conversion and peak threshold comparison on the acquired induced current value, battery temperature and pressure vector and simulated vibration envelope signal to obtain the energy boundary constraint matrix and transient trigger flag. The high-frequency transient data acquisition and energy budget module is used to perform high-frequency sampling of the conductor motion state within the maximum allowable time window based on the available energy Joule values ​​contained in the energy boundary constraint matrix in response to the transient triggering flag, so as to obtain high-frequency dynamic sequence data and the remaining energy constraint matrix; The computing resource adaptation module is used to perform frequency domain filtering and tensor dimension reshaping on high-frequency dynamic sequence data, and to perform computing power upper limit matching and conversion in combination with the remaining energy constraint matrix to obtain the normalized dynamic tensor and the target exit node identifier. The restricted reasoning and preliminary anomaly diagnosis module is used to perform feedforward blocking and local classifier reasoning on the normalized dynamic tensor based on the target exit node identifier to obtain the line anomaly state vector and information entropy score. The report optimization and wireless transmission adaptation module is used to perform confidence downgrading assessment and redundant field compression encoding on the line abnormal status vector and information entropy score to generate an optimized status report that matches the communication bandwidth and is used for scheduling by the cloud master station.

[0007] Compared with existing technologies, this application provides a power transmission line condition monitoring system based on intelligent sensors. It establishes an energy boundary constraint matrix by converting the physical energy levels of induced energy harvesting current, battery temperature and pressure status, and vibration envelope signals. The remaining energy constraints are mapped to the maximum available computing power upper limit of the end-side inference network. Combined with the inherent physical complexity of the input dynamic tensor, multi-objective exit node optimization is performed, allowing the algorithm to adaptively prune the network inference depth based on the significance of signal features. Furthermore, an optimized status report adapted for narrowband communication is generated through confidence degradation assessment and redundant field compression encoding. This approach solves the problem of critical transient missed detections and monitoring blindness caused by energy and computing power imbalances in power transmission line intelligent sensors under extreme coupled conditions of low current harvesting and micro-meteorological changes, significantly extending the effective monitoring lifespan of the equipment in passive supply states. Attached Figure Description

[0008] The above and other objects, features, and advantages of this application will become more apparent from the more detailed description of the embodiments of this application in conjunction with the accompanying drawings. The drawings are provided to further illustrate the embodiments of this application and form part of the specification. They are used together with the embodiments of this application to explain this application and do not constitute a limitation thereof. In the drawings, the same reference numerals generally represent the same components or steps.

[0009] Figure 1 This is a block diagram of a power transmission line condition monitoring system based on smart sensors according to an embodiment of this application; Figure 2 This is a schematic diagram of data flow in a power transmission line condition monitoring system based on smart sensors according to an embodiment of this application; Figure 3 This is a block diagram of the computing resource adaptation module in the intelligent sensor-based transmission line condition monitoring system according to an embodiment of this application; Figure 4 This is a block diagram of the reporting optimization and wireless transmission adaptation module in a smart sensor-based power transmission line condition monitoring system according to an embodiment of this application. Detailed Implementation

[0010] Hereinafter, exemplary embodiments according to this application will be described in detail with reference to the accompanying drawings. Obviously, the described embodiments are merely some embodiments of this application, and not all embodiments of this application. It should be understood that this application is not limited to the exemplary embodiments described herein.

[0011] As indicated in this application and claims, unless the context clearly indicates otherwise, the words "a," "an," "an," and / or "the" are not specifically singular and may include plural forms. Generally speaking, the terms "comprising" and "including" only indicate the inclusion of explicitly identified steps and elements, which do not constitute an exclusive list, and the method or apparatus may also include other steps or elements.

[0012] While this application makes various references to certain modules of the systems according to embodiments of this application, any number of different modules can be used and run on user terminals and / or servers. The modules described are merely illustrative, and different aspects of the systems and methods may use different modules.

[0013] Flowcharts are used in this application to illustrate the operations performed by the system according to embodiments of this application. It should be understood that the preceding or following operations are not necessarily performed in exact order. Instead, various steps can be processed in reverse order or simultaneously as needed. Furthermore, other operations can be added to these processes, or one or more steps can be removed from them.

[0014] The technical solution of this application proposes a power transmission line condition monitoring system based on intelligent sensors. Figure 1 This is a block diagram of a power transmission line condition monitoring system based on smart sensors according to an embodiment of this application. Figure 2 This is a schematic diagram of data flow in a power transmission line condition monitoring system based on smart sensors according to an embodiment of this application. Figure 1 and Figure 2 As shown, the power transmission line condition monitoring system 300 based on intelligent sensors according to an embodiment of this application includes: a condition perception and trigger judgment module 310, used to perform physical energy level conversion and peak threshold comparison on the acquired induced current value, battery temperature and pressure vector, and simulated vibration envelope signal to obtain an energy boundary constraint matrix and a transient trigger flag; a high-frequency transient data acquisition and energy budgeting module 320, used to, in response to the transient trigger flag, perform high-frequency sampling of the conductor motion state within the maximum allowable time window based on the available energy Joule value contained in the energy boundary constraint matrix to obtain high-frequency dynamic sequence data and a remaining energy constraint matrix; and a computing resource adaptation module 3. 30 is used to perform frequency domain filtering and tensor dimension reshaping on high-frequency dynamic sequence data, and to perform upper limit matching and conversion of computing power in combination with the residual energy constraint matrix to obtain the normalized dynamic tensor and the target exit node identifier; the restricted reasoning and anomaly preliminary diagnosis module 340 is used to perform feedforward blocking and local classifier reasoning on the normalized dynamic tensor based on the target exit node identifier to obtain the line abnormal state vector and information entropy score; the report optimization and wireless transmission adaptation module 350 is used to perform confidence downgrading evaluation and redundant field compression encoding on the line abnormal state vector and information entropy score to generate an optimized status report that matches the communication bandwidth and is used for cloud master station scheduling.

[0015] Specifically, the state perception and trigger judgment module 310 is used to perform physical energy level conversion and peak threshold comparison on the acquired induced current value, battery temperature and pressure vector, and simulated vibration envelope signal to obtain the energy boundary constraint matrix and transient trigger flag. It should be understood that intelligent sensors for transmission lines are deployed long-term on the surface of outdoor towers or conductors, and their energy source is highly dependent on the coordinated supply of energy from conductor induction and onboard batteries. In extreme coupled scenarios of low current-carrying operation and micro-meteorological changes (such as sudden increases in icing or strong winds), the sensor energy margin is extremely limited. If the system cannot accurately grasp the current energy boundary and establish a constraint framework based on it before initiating high-frequency sampling and end-side inference, subsequent transient data acquisition, computing power allocation, and inference decisions will lose their physical feasibility basis, and may even lead to premature battery depletion and complete sensor shutdown due to blindly waking up high-power modules. Therefore, in the technical solution of this application, the current available energy reserve is quantified by physical energy level conversion of the induced current value and the battery temperature and pressure vector. At the same time, the peak threshold of the simulated vibration envelope signal is compared to determine whether the conductor has entered an abnormal motion state. This provides a decision basis for all subsequent processing procedures in terms of both energy constraints and event triggering, ensuring that the system only starts subsequent high-energy-consuming operations when it is confirmed that the energy can support the system and that there is indeed an abnormal transient state. This achieves a front-end closed-loop coupling of energy perception and state perception.

[0016] In practice, the process begins by calculating the energy and charging slope of the induced current and battery temperature-pressure vector to generate initial energy state parameters. First, the induced current value is read. This value, obtained through inductive coupling from a current transformer mounted on the transmission line, reflects the instantaneous power the sensor can extract from the electromagnetic field at the current current level of the conductor. Specifically, this induced current value is combined with the equivalent load impedance and rectification efficiency parameters of the inductive energy harvesting circuit. The instantaneous output power of the current energy harvesting circuit is calculated according to the electromagnetic power conversion relationship. Then, based on the sampling period, it is converted into the energy increment value obtainable within a single period, characterizing the instantaneous rate of external energy replenishment in joules. Simultaneously, the battery terminal voltage component and battery surface temperature component in the battery temperature-pressure vector are read. Based on the battery electrochemical discharge model, the terminal voltage is mapped to an estimated value of the battery's current remaining state of charge. Temperature correction compensation is then applied to the actual release capacity of the battery at the current ambient temperature, as the battery's internal resistance increases and usable capacity decreases in low-temperature environments. This compensation ensures that the energy assessment is not overly optimistic under extremely cold conditions. Based on this, the temperature-corrected remaining usable energy of the battery is superimposed and integrated with the single-cycle energy increment of inductive energy harvesting. Simultaneously, the charging slope of the inductive energy harvesting power over time is calculated. This slope reflects whether the external energy supply is on an upward trend (e.g., increased current carrying capacity in the conductor) or a downward trend (e.g., during a low load period). Both together constitute the initial energy state parameters. These initial energy state parameters, in the form of a parameter set, include the absolute value of the current usable energy (joules) and the slope of the energy supply change (joules / second), providing a physically unified energy assessment foundation for subsequent standardized coding and constraint matrix construction.

[0017] Next, based on the exponential smoothing penalty function, the initial energy state parameters are standardized and adaptively calculated to obtain the energy boundary constraint matrix and the dynamic correction threshold. In this process, firstly, the initial energy state parameters are standardized to map the absolute value of available energy and the charging slope to a unified normalized interval, eliminating scale differences between different physical dimensions and ensuring numerical stability and comparability in subsequent adaptive calculations. Next, an exponential smoothing penalty function is introduced to adaptively calculate the encoded energy parameters. The working mechanism of this exponential smoothing penalty function is as follows: the function integrates the standardized energy parameters of the current period with the smoothed cumulative values ​​of historical periods in the form of an exponentially weighted moving average, so that the energy assessment has the ability to suppress short-term noise fluctuations and track long-term trends. At the same time, the function has an embedded penalty term mechanism. When a sharp drop in energy parameters is detected (such as a cliff drop in inductive power due to a sudden drop in conductor current, or a rapid drop in battery voltage due to a sudden increase in internal resistance caused by extreme cold), the penalty term will apply an additional conservative offset to the smoothed output, actively suppressing the estimated value of available energy, forcing the system to enter a conservative operating mode in the early stage of energy state deterioration, thereby reserving more power margin for the later stages of disaster. Furthermore, after processing with an exponential smoothing penalty function, the smoothed and corrected multidimensional energy parameters are assembled and arranged according to a predefined matrix structure to generate an energy boundary constraint matrix. This matrix encodes multidimensional constraint boundary information such as the available energy Joule value of the current sensor system, the smoothed energy replenishment rate, the conservative upper limit estimate of energy consumption, and the energy margin confidence interval in a structured row and column form. This provides a complete physical energy constraint matrix for determining the maximum allowable time window for high-frequency sampling and converting the upper limit of inference computing power in subsequent steps. Simultaneously, a dynamic correction threshold is calculated based on the energy margin information encoded in the matrix. The generation logic of the dynamic correction threshold reflects the adaptive coupling between energy margin and trigger sensitivity—when the energy boundary constraint matrix shows sufficient current energy margin, the dynamic correction threshold is set at a relatively low level, making the system more sensitive to conductor vibration and able to capture weak abnormal vibration signals; when the energy boundary constraint matrix shows insufficient energy margin, the dynamic correction threshold is raised accordingly, increasing the trigger threshold to avoid frequent responses to minor vibrations that consume valuable remaining power, thereby achieving a dynamic balance between trigger sensitivity and energy protection under different energy states.

[0018] Then, the peak amplitude of the simulated vibration envelope signal is extracted and compared in hard real-time with a dynamic correction threshold to obtain a transient trigger flag. In this process, firstly, the simulated vibration envelope signal output by the vibration sensor is received. This signal undergoes envelope detection processing by the front-end analog signal conditioning circuit, and its waveform profile reflects the time-varying characteristics of the macroscopic vibration intensity of the conductor under external excitations such as wind load and icing load. Specifically, a peak amplitude extraction operation is performed on this simulated vibration envelope signal. Specifically, within the current detection window time range, the sampling sequence of the vibration envelope signal is scanned, and its maximum instantaneous amplitude is captured as the peak amplitude feature of the current period. This peak amplitude feature represents the maximum intensity level of conductor vibration within the current detection period in the form of a single scalar value. Then, the extracted peak amplitude feature is compared in hard real-time with the dynamic correction threshold output from the second sub-step. This means that this comparison operation must be completed within a strict deterministic time limit, allowing no non-deterministic delays, to ensure the real-time response of the system to sudden transient events. The comparison logic is as follows: if the current peak amplitude is greater than or equal to the dynamic correction threshold, the conductor is determined to have entered an abnormal motion state, and the transient trigger flag is set to valid (logical true); if the current peak amplitude is less than the dynamic correction threshold, the conductor is determined to be within the normal vibration range, and the transient trigger flag remains invalid (logical false). It is worth noting that when the transient trigger flag is valid, subsequent modules will respond to the flag and initiate high-frequency sampling of the conductor's motion state based on the available energy Joule value in the energy boundary constraint matrix; when the transient trigger flag is invalid, the system maintains a low-power standby state and does not perform high-frequency sampling operations, thereby maximizing the protection of the remaining battery power.

[0019] Specifically, the high-frequency transient data acquisition and energy budget module 320 is used to, in response to the transient trigger flag, perform high-frequency sampling of the conductor motion state within the maximum allowable time window based on the available energy Joule values ​​contained in the energy boundary constraint matrix, to obtain high-frequency dynamic sequence data and the remaining energy constraint matrix. It should be understood that when the state perception and trigger judgment module in the previous step determines that the conductor has entered an abnormal motion state and sets the transient trigger flag to valid, the system needs to initiate high-frequency data acquisition of the conductor motion state within the shortest response delay to obtain sufficient time-resolution dynamic information for subsequent end-side anomaly diagnosis inference. However, the energy source of the transmission line smart sensor is highly limited by the coordinated supply of inductive energy harvesting and battery energy storage, especially under low current-carrying conditions and extreme micro-meteorological coupling scenarios, where the available energy margin of the sensor is extremely scarce. Compared to the conventional low-power standby state, the system power consumption in the high-frequency sampling mode will increase by an order of magnitude—the continuous operation of the high-sampling-rate analog-to-digital converter, data cache writing, and clock frequency increases will all significantly consume the remaining battery power. If strict energy constraints are not imposed on the duration of high-frequency sampling, the system is highly likely to exhaust its battery power during the transient data acquisition phase. This would prevent subsequent critical steps such as computing resource adaptation, edge-side inference diagnosis, and report transmission from executing due to energy depletion, ultimately leading to a complete sensor shutdown in the later stages of a disaster. Therefore, in the technical solution of this application, under the premise of confirming the occurrence of a transient event, the available energy Joule value encoded in the energy boundary constraint matrix is ​​used as a physical hard constraint to accurately calculate the maximum allowable high-frequency sampling time window that the system can support under the current energy conditions. Within this window, high-frequency data acquisition of the conductor's motion state is completed to obtain high-frequency dynamic sequence data. Simultaneously, the actual energy consumption during the sampling process is tracked in real time, and the updated remaining energy constraint matrix is ​​reconstructed accordingly. This provides a precisely reduced energy budget basis for all subsequent processing steps, achieving an optimal balance between transient data acquisition quality and system energy endurance.

[0020] In practice, firstly, based on the system's dynamic power consumption parameters, the available energy extracted from the energy boundary constraint matrix and the transient trigger flag are used for duration estimation and multiplication blocking calculation to obtain the maximum sampling time window. During this process, the transient trigger flag transmitted from the previous step is received, and its validity is checked. When the transient trigger flag is invalid (logical false), the module performs multiplication blocking calculation—multiplying the duration estimation result with the zero value of the invalid flag, forcing the calculation result of the maximum sampling time window to zero and blocking it. The system does not initiate any high-frequency sampling operations, maintaining a low-power standby state, thus completely preventing false wake-ups of the high-frequency sampling module in the absence of abnormal events. When the transient trigger flag is valid (logical true), the flag value in the multiplication blocking calculation is one, having no blocking effect on the duration estimation result, and the module continues to execute the subsequent duration estimation process.

[0021] Under the premise that the transient trigger flag is valid, the available energy joules are extracted from the energy boundary constraint matrix. This available energy joule represents the total energy reserve available for consumption by the current sensor system after correction by the exponential smoothing penalty function. However, not all of this available energy is exclusively available for high-frequency sampling; the system must reserve a necessary share for subsequent frequency domain filtering and tensor dimension reshaping, edge inference and anomaly diagnosis, report encoding, and wireless transmission. Therefore, based on a preset energy allocation strategy, the estimated energy requirements of each subsequent processing stage are subtracted from the available energy joules to obtain the net available energy value that can be allocated to the high-frequency sampling stage. On this basis, a system dynamic power consumption parameter is introduced. This parameter, in watts, characterizes the comprehensive instantaneous power consumption of the sensor in high-frequency sampling mode, covering the power consumption of the ADC converter at high conversion frequencies, the read / write power consumption of the data buffer, the preprocessing power consumption of the digital signal processing front-end, and the static and dynamic power consumption of related peripheral interface circuits. Then, the net available energy value allocated to high-frequency sampling is divided by the system dynamic power consumption parameter to obtain the theoretical maximum time span that high-frequency sampling can sustain under the current energy constraints. This process can be expressed by the following formula: in, The maximum sampling time window calculated is in seconds; The available energy Joule value extracted from the energy boundary constraint matrix; This is the energy reserve that the system makes for subsequent reasoning, diagnosis, and communication processes. This refers to the system's dynamic power consumption parameters, i.e., the comprehensive instantaneous power consumption under high-frequency sampling mode; The logical value of the transient trigger flag (1 when valid, 0 when invalid) is used to implement the multiplication blocking function—when the transient trigger flag is invalid, If the value is zero, the entire product result is forced to zero, the maximum sampling time window is zero, and the system does not perform any high-frequency sampling operations. After the above calculations, the maximum sampling time window is obtained. This parameter constrains the subsequent high-frequency data acquisition process with a clear time upper limit, ensuring that the sampling duration does not exceed the maximum duration allowed by the energy safety boundary.

[0022] Next, data acquisition and splicing, along with dynamic power consumption accumulation, are performed within the maximum sampling time window to obtain high-frequency dynamic sequence data and an estimate of actual energy consumption. In this process, the high-frequency sampling mode is immediately initiated after obtaining the maximum sampling time window. Specifically, the module wakes up the high-speed ADC converter and data buffer from their low-power sleep state, configures the sampling frequency to a preset high-frequency mode (e.g., in the kilohertz range), and uses the maximum sampling time window as a hard upper limit for the sampling duration, initiating continuous data acquisition of the conductor's motion state. During acquisition, vibration sensors (such as triaxial accelerometers) output instantaneous acceleration values ​​of the conductor in each degree of freedom direction at a high-frequency sampling rate. The ADC converter converts the analog signal into digital sample values, and the data buffer sequentially stores the digitized data from each sampling moment according to time sequence. As the sampling process progresses, the module performs time-domain splicing on the discrete sampled data from each continuously acquired time segment, seamlessly connecting the beginning and end of each data segment according to the ascending order of the sampling timestamps. This ensures that the spliced ​​data sequence has complete continuity on the time axis, without any data discontinuities or timestamp jumps. When the sampling duration reaches the upper limit specified by the maximum sampling time window, the module immediately terminates the high-frequency sampling operation, shuts down the ADC converter and related high-power circuits, and outputs the spliced ​​complete time-domain data sequence as the high-frequency dynamic sequence data. The high-frequency dynamic sequence data completely records the time-domain evolution trajectory of vibration acceleration during the abnormal motion of the conductor within the maximum sampling time window. Its high temporal resolution ensures the integrity of the spectral information required for subsequent frequency domain analysis and feature extraction.

[0023] Meanwhile, within each time segment of high-frequency sampling, the module calculates the instantaneous energy consumption increment based on the current system dynamic power consumption parameters, and gradually accumulates the energy consumption increments of each time segment. Dynamic power consumption accumulation considers the dynamic fluctuations in power consumption that may occur during sampling due to factors such as changes in ambient temperature (e.g., power consumption drift caused by extreme cold) and the continuous decrease in battery voltage during discharge (leading to changes in the efficiency of the voltage regulator circuit), ensuring that the accumulated energy consumption is as close as possible to the actual physical consumption. When high-frequency sampling terminates, the final accumulated result of dynamic power consumption is the estimated value of actual energy consumption, accurately recording the total energy actually consumed in this high-frequency sampling operation in joules. This accumulation process can be expressed by the formula: in, This is an estimated value of actual energy consumption, expressed in joules. This represents the total number of time segments divided during the high-frequency sampling process. For the first The instantaneous power consumption value within a time segment, in watts, reflects the actual dynamic power consumption of the system within that segment; For the first The duration of a time segment, in seconds.

[0024] Furthermore, the energy boundary constraint matrix and the estimated actual energy consumption are subjected to energy deduction and reconstruction of the remaining constraint matrix to obtain the remaining energy constraint matrix. In this process, firstly, the estimated actual energy consumption is subtracted from the available energy Joules encoded in the energy boundary constraint matrix to obtain the remaining available energy Joules of the system after high-frequency sampling. Next, the remaining constraint parameters in the energy boundary constraint matrix are updated and reconstructed to comprehensively reflect the changes in the system's energy state after high-frequency sampling. Specifically, based on the observed battery voltage drop trend and real-time changes in inductive power harvesting during high-frequency sampling, the energy replenishment rate parameter is reassessed—if the battery voltage drop rate exceeds expectations or the inductive power harvesting remains low during sampling, the energy replenishment rate is corrected downwards. In addition, based on the deducted remaining available energy Joules, a conservative upper limit estimate of energy consumption is recalculated, tightening the energy limit available for subsequent stages; simultaneously, the energy margin confidence interval is updated to reflect the uncertainty range of the energy state after sampling. Finally, all the energy parameters after reduction and update are reassembled and arranged according to the same matrix structure as the original energy boundary constraint matrix to complete the reconstruction of the remaining constraint matrix and generate the remaining energy constraint matrix.

[0025] Specifically, the computing resource adaptation module 330 is used to perform frequency domain filtering and tensor dimension reshaping on the high-frequency dynamic sequence data, and to perform upper limit matching and conversion of computing power in combination with the remaining energy constraint matrix to obtain the normalized dynamic tensor and the target exit node identifier. It should be understood that the high-frequency dynamic sequence data records the multi-channel motion characteristics of the conductor during transient events in the form of the original time-domain waveform. However, this original data inevitably contains frequency components unrelated to the actual abnormal state of the conductor, such as environmental noise, power frequency interference, and sensor background noise. If the unfiltered original data is directly fed into the subsequent edge-side neural network inference module, it will not only lead to a low input signal-to-noise ratio of the classifier and a significant decrease in diagnostic accuracy, but also cause unnecessary waste of computing power due to excessive redundant data. Furthermore, the dimensional structure of the original high-frequency dynamic sequence data is a one-dimensional time series, while the input format required by the edge-side neural network model is usually a multi-dimensional tensor structure. There is a mismatch in dimensional form between the two, requiring tensor dimension reshaping operations to fold and map the one-dimensional sequence data into a multi-dimensional tensor that conforms to the network input specifications. Meanwhile, after the energy consumption of the high-frequency sampling phase, the system's remaining energy has further narrowed, and the computing resources available for subsequent edge inference are subject to strict physical energy constraints. The edge neural network adopts a cascaded architecture with multiple early exit points. Different exit nodes correspond to different inference depths and computing power consumption requirements. The system must determine the optimal inference termination position within the upper limit of computing power supported by the current remaining energy. Therefore, in the technical solution of this application, firstly, the original high-frequency dynamic sequence data is transformed into a high-quality normalized dynamic tensor adapted to the neural network input specifications through frequency domain filtering and tensor dimension reshaping; then, by performing computing power upper limit matching and conversion on the remaining energy constraint matrix and combining the multi-exit network topology to optimize the exit node, the target exit node identifier at which level the inference computation should terminate under the hard energy constraint is determined. This provides the subsequent restricted inference and anomaly preliminary diagnosis modules with preprocessed high-quality input data and inference depth control instructions matched with energy constraints.

[0026] Figure 3 This is a block diagram of the computational resource adaptation module in a power transmission line condition monitoring system based on smart sensors according to an embodiment of this application. Figure 3As shown, in the first embodiment of this application, the computing resource adaptation module 330 includes: a frequency domain filtering and tensor dimension folding and reshaping unit 331, used to perform frequency domain filtering and tensor dimension folding and reshaping on high-frequency dynamic sequence data to obtain a normalized dynamic tensor; an energy efficiency conversion and network topology extraction unit 332, used to perform energy efficiency conversion and network topology extraction on the remaining energy constraint matrix based on basic energy consumption indicators and scheduling overhead parameters to obtain the maximum available computing power limit and a multi-exit network configuration table; and a target exit node deep optimization unit 333, used to perform multi-target computing power matching and exit node deep optimization on the multi-exit network configuration table and the maximum available computing power limit to obtain the target exit node identifier.

[0027] Specifically, the frequency domain filtering and tensor dimension folding and reshaping unit 331 is used to perform frequency domain filtering and tensor dimension folding and reshaping on the high-frequency dynamic sequence data to obtain a normalized dynamic tensor. In this process, firstly, the high-frequency dynamic sequence data is transformed from the time domain to the frequency domain. Specifically, the high-frequency dynamic sequence data contains high-resolution records of the acceleration components of the conductor in multiple spatial degrees of freedom over time. By performing a discrete Fourier transform on the time series of each channel, it is converted from an amplitude-time representation in the time domain to an amplitude-frequency representation in the frequency domain, obtaining the spectral distribution of the signal in each channel. In the frequency domain, selective processing of spectral components is performed according to a preset frequency band preservation strategy: low-frequency components below the preset lower cutoff frequency (these components mainly correspond to low-frequency interference unrelated to abnormal conductor movement, such as slow drift of the sensor mounting base and zero-point shift caused by temperature changes); high-frequency components above the preset upper cutoff frequency (these components mainly correspond to stray signals such as sensor background electronic noise and electromagnetic interference); and effective frequency band components located between the upper and lower cutoff frequencies (this frequency band covers the typical characteristic frequency range of abnormal movement modes such as conductor galloping, light wind vibration, and ice-breaking jumps). After selective preservation and suppression of the spectral components, an inverse discrete Fourier transform is performed on the processed frequency domain data to transform it back from the frequency domain to the time domain, obtaining frequency-domain filtered and denoised dynamic sequence data. This filtered data retains the characteristic information of abnormal conductor movement in the time domain waveform while effectively removing noise interference components, significantly improving the signal-to-noise ratio compared to the original data. Subsequently, dimensionality transformation and normalization processing are performed on the frequency-domain filtered dynamic sequence data to adapt it to the tensor format required by the input layer of the multi-exit neural network at the end side. Specifically, the long-term time series is first folded along the time axis into segments of fixed window length, transforming the originally continuous one-dimensional time series into multiple time window segments of equal length. Then, these time window segments are stacked along a newly added batch dimension, while retaining the time step dimension and feature channel dimension within each segment. This reshapes the original two-dimensional data into a three-dimensional tensor structure with batch dimension × time step dimension × feature channel dimension. After dimensionality folding and reshaping, the unit performs normalization scaling on the reshaped tensor data, mapping the numerical range of each feature channel to a standardized interval with zero mean and unit variance. This eliminates the inconsistency in numerical scale between different channels (such as different axial accelerations) caused by differences in physical dimensions and ranges, ensuring that the weights of each layer in the network pay balanced attention to the information of each channel during feature extraction, ultimately yielding a normalized dynamic tensor.

[0028] Specifically, the energy efficiency conversion and network topology extraction unit 332 is used to perform energy efficiency conversion and network topology extraction on the remaining energy constraint matrix based on basic energy consumption indicators and scheduling overhead parameters to obtain the maximum available computing power limit and multi-exit network configuration table. In this process, firstly, the remaining available energy joules are extracted from the remaining energy constraint matrix, and the basic energy consumption indicators and scheduling overhead parameters are read. These parameters describe the additional energy consumption generated by non-computational operations such as data transfer between network layers, intermediate feature map cache management, multi-exit node branch judgment, and post-processing of inference results during the inference process. Next, the remaining available energy joules are subtracted from the system-level overhead energy estimate represented by the scheduling overhead parameters to obtain the net available energy that can be purely used for neural network floating-point operations. Then, this net available energy is divided by the basic energy consumption indicators to obtain the maximum total floating-point operations that the current remaining energy can support, i.e., the maximum available computing power limit. Subsequently, the model configuration file of the pre-deployed edge-side multi-exit neural network is read from the sensor's onboard storage. All pre-defined early exit nodes in the network architecture are traversed, and the structured configuration information of each candidate exit node is extracted. This includes the node's depth parameters in the network, the theoretical floating-point computation required to feed forward from the input layer to the node, the structural parameters of the local classifier and regressor associated with the node, and the expected classification performance metrics obtained by the node during the training phase. The above configuration information of all candidate exit nodes is arranged and assembled in order of network depth from shallow to deep to generate a multi-exit network configuration table. This table, in a structured tabular form, completely records the topology of the edge-side multi-exit network and the attribute parameters of each exit node, providing a full attribute description of the candidate nodes for subsequent optimization of the target exit node's depth.

[0029] Specifically, the target exit node depth optimization unit 333 is used to perform multi-target computing power matching and exit node depth optimization on the multi-exit network configuration table and the maximum available computing power limit to obtain the target exit node identifier. During this process, all candidate exit nodes are traversed from the multi-exit network configuration table, and the estimated power consumption requirement of each candidate node (i.e., the theoretical floating-point operation required from the input layer feedforward to the node) is extracted. This estimated power consumption is compared with the maximum available computing power limit. Specifically, the nodes are checked sequentially according to the network depth from the deepest to the shallowest: for the candidate exit node with the greatest network depth, if its estimated power consumption requirement does not exceed the maximum available computing power limit, the deepest node is directly identified as the target exit node, and the system will invest all available computing power into the feedforward calculation of the deepest network to obtain the most sufficient feature extraction depth; if the estimated power consumption of the deepest node exceeds the maximum available computing power limit, the next deepest node is checked, and so on, until the first candidate node whose estimated power consumption does not exceed the maximum available computing power limit is found, and it is identified as the target exit node. In this way, while satisfying the hard energy constraint (estimated power consumption does not exceed the maximum available computing power limit), the exit node with the largest network depth is always selected as the inference termination point, maximizing the use of currently available computing power for the deepest possible feature extraction. This process can be expressed by the following formula: in, The final output is the target exit node identifier. This refers to the set of all candidate early exit nodes included in the multi-exit network configuration table. For the first Network depth parameters of each candidate node For the first The estimated floating-point computation cost of each candidate node. This represents the maximum available computing power limit. The physical meaning of this formula is: among all candidate exit nodes, select a subset of legal nodes whose estimated floating-point operation volume does not exceed the maximum available computing power limit, and then select the node with the largest network depth in this legal subset as the target exit node.

[0030] In particular, the study found that the multi-objective computing power matching and exit node deep optimization mechanism of the first embodiment adopts an extremely rigid greedy exhaustion strategy at the underlying logic level. Under severe operating conditions such as extremely cold microclimates or severe icing causing a precipitous drop in power extraction from transmission lines, this mechanism exposes a fatal blind spot in physical boundaries.

[0031] Specifically, the first embodiment relies solely on the computing power limit calculated from the remaining power as the single criterion. Once it determines that the remaining energy is sufficient, it unreservedly feeds computing resources to the deepest nodes of the neural network. This unidirectional triggering logic, where the computing power limit determines the inference depth, completely severs the extremely unique nonlinear relationship between the inherent physical complexity of the input signal and the marginal benefit of network feature extraction. In industrial settings, when conductors violently dance or jump off ice, their dynamic signal characteristics exhibit extremely strong salience. Shallow neural network nodes can easily achieve extremely high state classification confidence. At this time, the state warning information gain brought by forcibly calling deep networks is almost zero. If the first embodiment is used, the system will still blindly mobilize huge deep computing resources in the face of easily identifiable strong feature signals, wasting extremely precious remaining power during disasters. This causes sensors to completely shut down due to premature battery depletion in the later stages of ice disasters when continuous monitoring is most needed, leading to serious accidents of monitoring blindness.

[0032] To address the aforementioned deficiencies, this application proposes a second embodiment.

[0033] Specifically, firstly, the inherent complexity of the normalized dynamic tensor is dynamically evaluated to obtain the tensor complexity index. That is, for the normalized dynamic tensor of this acquisition period, the rate of change of its L2 gradient along the continuous time step and spatial axis is extracted, and the global environmental background variance of the tensor is deeply integrated to quantify the intensity of the current physical signal characteristics under the interplay of light wind turbulence and snow cover. Substituting the parameters of the above dimensions into a nonlinear attenuation model, the tensor complexity index, reflecting the ease of identifying the current physical state of the line, is calculated. Through the nonlinear combination of the gradient norm and the global variance, the complexity of the fusion between environmental noise and the actual fault state under wind and rain conditions is accurately characterized. A larger index indicates a weaker and more chaotic signal, while a smaller index indicates extremely obvious disaster characteristics. The calculation process of the tensor complexity index can be expressed by the following formula: in, The tensor complexity index is calculated. This represents the total number of sequence steps of the normalized tensor within the time window; For the tensor in the first... The high-dimensional gradient vector at each time step; Let L2 norm be the tensor used to measure the amplitude of fluctuations; This represents the global background variance of the entire normalized tensor within the current period; The temperature correction attenuation constant is used to reflect the sensor's background noise.

[0034] Next, the tensor complexity index, the multi-exit network configuration table, and the maximum available computing power limit are reconstructed using a joint reward matrix of node marginal utility and energy conservation to obtain the node optimization benefit matrix. In other words, after understanding the characteristics of the input signal, the optimal computing power deployment target needs to be found under multi-dimensional constraints. Specifically, the maximum available computing power limit and the multi-exit network configuration table are introduced simultaneously. Each candidate exit node in the configuration table is traversed, and its corresponding network depth parameters and theoretical computing power consumption requirements are precisely extracted. During execution, a joint evaluation system is constructed for all candidate nodes in the network, combining the newly generated tensor complexity index. The hyperbolic tangent function is used to fit the diminishing marginal utility of information gain under the coupling of node depth and signal complexity, while the exponential decay law is used to measure the energy conservation penalty when the node's computing power consumption approaches the maximum available computing power limit. By weighted fusion of information gain utility and energy conservation penalty, a comprehensive optimization score is assigned to each candidate network level, thereby integrating and generating a multi-dimensional node optimization benefit matrix, and collaboratively transmitting the maximum available computing power limit and the multi-exit network configuration table. This reconstruction process mathematically endows the system with a frugal yet astute survival strategy. When faced with easily identifiable large fluctuations, shallow depths can quickly saturate the utility terms, and the system will automatically discard deep nodes to complete strategic energy reserves. The calculation process for the joint reward score of nodes can be expressed by the following formula: in, In the constructed node optimization reward matrix, the first... The combined reward score for each exit node; A dynamic weighting factor is used to balance information gain and energy conservation. The hyperbolic tangent function is used to simulate the nonlinear saturation characteristics of marginal returns; This represents the sensitivity coefficient of network depth to complex feature extraction. Let be the network depth of the i-th candidate node; The tensor complexity index is the one passed through. The estimated floating-point computation cost for the i-th candidate node; This is the maximum available computing power limit calculated for the preceding stages.

[0035] Furthermore, the maximum utility node is locked under hard energy constraints on the node optimization benefit matrix, multi-exit network configuration table, and maximum available computing power limit to obtain the target exit node identifier. Specifically, a hard review is initiated on all candidate nodes in the multi-exit network configuration table, forcibly intercepting and eliminating illegal nodes whose estimated power consumption exceeds the maximum available computing power limit, thus establishing an absolute bottom line of physical energy isolation. In the legal node clusters that cross the hard constraint threshold, the node optimization benefit matrix is ​​compared horizontally, the maximization solution operator is activated, and the level that achieves the highest joint reward score is locked as the optimal feature extraction exit. The unique identifier of this optimal matching level is extracted and output as the target exit node identifier, thereby completely consuming the benefit matrix and configuration table data. This process, while adhering to the ironclad rule of hardware not downtime, incorporates soft dynamic utility optimization to ensure the best balance between status diagnosis and power endurance under disaster conditions. The process of determining the target exit node identifier can be expressed by the formula: In the formula, The final output is the target exit node identifier; Solver for the set of independent variables that maximizes the objective condition; This is the set of all candidate early exit nodes included in the multi-exit network configuration table; Find the score element in the revenue matrix for the node; This refers to the computing power of the corresponding node. This represents the maximum available computing power.

[0036] Specifically, the deep optimization mechanism proposed in the second embodiment achieves a perfect balance between the resilient survival and accurate alarm of the intelligent monitoring device at the transmission line end-side under extreme weather conditions. It breaks away from the rigid inference mode of traditional end-side AI architectures, which are simply limited by computing power bottlenecks, and endows sensors with the intelligent game-theoretic ability to determine network depth based on the transient intensity of the physical scene. In the face of high-frequency, high-characteristic disaster emergencies, the system can spontaneously resist the temptation of deep network computing power, completing high-confidence early warnings with extremely low energy consumption, and storing all remaining Joule-level power in supercapacitors; while in the weak signal dormancy period at the beginning of a disaster, it can precisely squeeze computing power to the physical limit to avoid missed alarms. Ultimately, it fundamentally cures the industry problem of large-scale offline and inactive transmission line IoT devices under severe conditions such as power outages due to ice storms and rainstorms, greatly extending the effective monitoring lifespan of the equipment in a passive supply state.

[0037] Specifically, the restricted reasoning and preliminary anomaly diagnosis module 340 is used to perform feedforward blocking and local classifier reasoning based on the target exit node identifier on the normalized dynamic tensor to obtain the line anomaly state vector and information entropy score. It should be understood that the normalized dynamic tensor carries the full-process dynamic characteristic information of the conductor during the transient event in a high-quality, multi-dimensional structured form, and the target exit node identifier clearly indicates at which level the end-side neural network reasoning should terminate under the current remaining energy constraint. However, the completion of data preprocessing and resource allocation is not equivalent to the actual diagnosis of the conductor's anomaly state—the physical characteristic information contained in the normalized dynamic tensor has not yet undergone layer-by-layer feature extraction and nonlinear mapping by the neural network. At this point, the system still cannot determine what kind of anomaly state the conductor is currently in (such as icing growth, wind-induced galloping, ice-breaking jumps, hardware loosening, etc.), nor can it quantify the reliability of the diagnostic conclusion. Therefore, in the technical solution of this application, the target exit node identifier is used as the inference depth control instruction to drive the end-side cascaded neural network to perform limited-depth feedforward inference on the normalized dynamic tensor. At the target exit node, the feedforward propagation of the network is terminated in advance and the feature representation of that level is truncated. The truncated features are mapped to specific abnormal state classification results and physical parameter estimates through the local classifier and regressor attached to the node. At the same time, the information entropy score is calculated based on the probability distribution of the classification prediction to quantify the confidence level of the diagnostic conclusion.

[0038] In practice, firstly, the normalized dynamics tensor is processed by computational allocation, blocking, and activation extraction using the target exit node identifier to obtain a truncated feature map. During this process, the target exit node identifier is first read to confirm which level of the cascaded network the current inference process should terminate at. Next, the normalized dynamics tensor is fed as input into the input layer of the end-side cascaded neural network to initiate the feedforward inference process. Starting from the network input layer, the normalized dynamics tensor sequentially undergoes standard feedforward computation steps such as convolution operations, batch normalization operations, nonlinear activation function processing, and pooling operations at each network level, performing feature extraction and dimensionality transformation layer by layer. As the feedforward inference progresses deeper layer by layer, the module continuously monitors the network level identifier reached by the current inference and compares it with the target exit node identifier in real time. When the feedforward inference reaches the network layer specified by the target exit node identifier, the module immediately performs a computing power allocation blocking operation—forcibly interrupting the feedforward propagation calculation of subsequent deeper layers, cutting off the processor's allocation of floating-point operation resources to all network layers deeper than the target exit node, so that the data flow no longer continues to be transmitted to deeper layers of the network.

[0039] Simultaneously, activation extraction is performed at the output of the target exit node level. Specifically, the module reads the activation values ​​generated after the feedforward computation of that level from the output buffer of the target exit node level. This activation value is the intermediate feature representation output by that level after all feedforward computations from the input layer to the current layer on the normalized dynamics tensor. This activation value exists in the form of a multi-dimensional feature tensor, where each dimension corresponds to network structure parameters such as the number of feature channels and the feature map spatial dimension. It encapsulates the hierarchical feature information accumulated from the network input layer through layer-by-layer convolution and nonlinear transformations up to the depth of the target exit node. The module outputs the extracted activation values ​​as a truncated feature map.

[0040] Next, the local classifiers and regressors, awakened based on the target exit node identifier, perform fully connected mapping and physical parameter concatenation on the truncated feature map to obtain the probability distribution array and the line anomaly state vector. In this process, firstly, the local classifiers and regressors associated with the exit node are awakened based on the target exit node identifier. Each candidate exit node in the edge-side cascaded neural network is pre-configured with an independent local classifier and regressor. These classifiers and regressors have undergone specialized parameter optimization for the feature depth of their respective layers during network training, enabling them to output meaningful classification and regression results based on the truncated feature map of the corresponding layer. In normal low-power standby mode, the local classifiers and regressors of all exit nodes are in a dormant state to save memory and computing resources. When the target exit node identifier is determined, the module only awakens the local classifiers and regressors associated with the exit node corresponding to that identifier, loading their parameters from the firmware storage area into the processor's running memory. The local classifiers and regressors of other exit nodes remain in a dormant state and are not loaded, thereby minimizing unnecessary memory usage and energy consumption. Furthermore, the local classifier first flattens the multidimensional tensor of the truncated feature map into a one-dimensional feature vector, and then feeds this one-dimensional feature vector into a fully connected neural network layer. In the fully connected layer, each element of the feature vector establishes a weighted connection with each neuron in the output layer. A linear transformation is completed by matrix multiplication of the weight matrix and the feature vector, followed by the superposition of the bias vector. After processing by a non-linear activation function, the result is output to the next fully connected layer or the final output layer. The original score value output by the fully connected layer is converted into a probability distribution form by the Softmax normalization function, generating a probability distribution array. The probability distribution array fully describes the membership distribution of the current truncated feature map in each abnormal state category. The category with the highest predicted probability value is the most likely abnormal state type determined by the local classifier.

[0041] Simultaneously, a fully connected mapping is performed on the truncated feature map received after wake-up via a regressor. The physical parameter estimates output by the regressor include key physical quantities related to the current abnormal state, such as conductor galloping amplitude (in meters), dominant vibration frequency (in Hertz), and estimated icing thickness (in millimeters), which are continuous physical parameters that provide quantitative physical support for assessing the severity of the abnormal state. Subsequently, the probability distribution array (representing the discrete predicted probability distribution of each abnormality category) and the physical parameter estimates (representing the continuous physical quantity estimates of the abnormal state) are concatenated end-to-end in the vector dimension to form a unified composite vector, namely the line abnormal state vector.

[0042] Then, discrete entropy integral calculation and confidence inverse quantization evaluation are performed on the classification prediction probability values ​​contained in the probability distribution array to obtain the information entropy score. In this process, firstly, discrete entropy integral calculation is performed on all classification prediction probability values ​​contained in the probability distribution array. The discrete entropy integral calculation is based on the Shannon entropy formula in information theory, performing a weighted logarithmic summation operation on the prediction probability value of each abnormal state category in the probability distribution array. This process can be expressed by the formula: in, The calculated discrete entropy value; the calculated discrete entropy value; the total number of predefined line abnormal state categories; The probability distribution array is the first... The classification prediction probability value of each abnormal state category; This refers to a logarithmic operation with base 2. Here, the physical meaning of discrete entropy is as follows: When the probability distribution array is highly concentrated in a specific category (i.e., the predicted probability value of one category is close to 1, and the predicted probability values ​​of the other categories are close to 0), the dispersion of the probability distribution is lowest, and the discrete entropy value approaches 0, indicating that the classifier has extremely high certainty in judging the type of the current abnormal state; when the probability distribution array is evenly distributed among multiple categories (i.e., the predicted probability values ​​of each category are similar and all much less than 1), the dispersion of the probability distribution is highest, and the discrete entropy value approaches its theoretical maximum value, indicating that the classifier cannot make a clear judgment and distinction among multiple candidate categories, and the uncertainty of the diagnostic conclusion is extremely high.

[0043] Furthermore, a confidence-inverse quantification evaluation operation is performed on the discrete entropy values ​​to map the continuous discrete entropy values ​​to a standardized information entropy score. This mapping process reflects the inverse relationship between information entropy and diagnostic confidence. Specifically, the discrete entropy value is normalized by dividing it by its theoretical maximum value, so that the information entropy score is mapped to the standardized interval [0,1]. A normalized information entropy score of 0 indicates that the diagnostic conclusion has the highest confidence (the probability distribution is completely concentrated in a single category), while an information entropy score of 1 indicates that the diagnostic conclusion has the lowest confidence (the probability distribution is completely uniformly dispersed across all categories).

[0044] Specifically, the report optimization and wireless transmission adaptation module 350 is used to perform confidence downgrading assessment and redundant field compression encoding on the line abnormal state vector and information entropy score to generate an optimized state report that matches the communication bandwidth and is used for scheduling by the cloud master station. It should be understood that the completion of end-side inference does not mean that the diagnostic mission of the monitoring system has been accomplished—the diagnostic conclusion must be reported to the cloud master station via the wireless communication link to truly be incorporated into the power grid dispatch decision-making system and trigger corresponding operation and maintenance responses. Transmission line smart sensors are deployed on outdoor towers, and their wireless communication typically relies on low-power wide-area network (LPWAN) protocols (such as LoRa, NB-IoT, etc.). These protocols achieve long-distance coverage with extremely low power consumption, but at the cost of extremely limited communication bandwidth and strictly constrained maximum load capacity for a single transmission. If the line abnormal state vector is directly transmitted wirelessly in its original floating-point form, not only may the data volume exceed the single-frame load limit of the communication protocol, but it will also consume a large amount of communication energy due to excessive transmission time, accelerating battery depletion under extreme energy-scarce conditions. Furthermore, due to the early termination of target exit nodes, the confidence level of edge-side inference is not always high. If diagnostic conclusions with high information entropy scores are reported to the master station without annotation, it may mislead scheduling decisions and lead to resource misallocation. Therefore, in the technical solution of this application, firstly, the confidence level of the diagnostic conclusions is downgraded and alarm labels are determined based on the information entropy score, and a downgrade mark is added to the low-confidence results to remind the master station to treat them with caution; secondly, the data volume is compressed to the minimum binary message that matches the communication bandwidth through spatiotemporal information fusion and differential compression coding; finally, the communication header is encapsulated based on the low-power wide area network protocol and the adaptive sleep time is calculated by combining the negative exponential decay function to generate an optimized status report that simultaneously carries diagnostic content and sensor scheduling instructions, thereby achieving multi-objective optimization of diagnostic information integrity, communication efficiency, and energy saving.

[0045] Figure 4 This is a block diagram of the reporting optimization and wireless transmission adaptation module in a smart sensor-based power transmission line condition monitoring system according to an embodiment of this application. Figure 4As shown, the report optimization and wireless transmission adaptation module 350 includes: an alarm tag determination and matrix splicing unit 351, used to determine alarm tags and perform matrix dimension direct sum splicing processing on the information entropy score and line abnormal state vector through a preset tolerance threshold and indicator function logic to obtain a labeled state vector; a spatiotemporal fusion and differential coding unit 352, used to perform spatiotemporal information fusion and differential compression coding on the labeled state vector to obtain a compressed binary message; and an adaptive sleep time estimation unit 353, used to perform communication header encapsulation and adaptive sleep time estimation on the compressed binary message based on the low power wide area network protocol and negative exponential decay function to obtain an optimized status report.

[0046] Specifically, the alarm label determination and matrix concatenation unit 351 is used to perform alarm label determination and matrix dimension direct sum concatenation processing on the information entropy score and the line abnormal state vector through a preset tolerance threshold and indicator function logic to obtain the labeled state vector. In this process, firstly, an alarm label determination operation based on the preset tolerance threshold and indicator function logic is performed on the information entropy score. The preset tolerance threshold is a benchmark value for information entropy score determination pre-set by the system during the deployment phase. This threshold is set based on the offline calibration of the classification performance statistical characteristics of the peer-side inference model at different exit depths, reflecting the system's maximum tolerance for diagnostic uncertainty. Specifically, the information entropy score is numerically compared with the preset tolerance threshold, and alarm labels are generated through indicator function logic. The specific implementation of the indicator function logic is as follows: An indicator function is defined. When the information entropy score is strictly greater than a preset tolerance threshold, the indicator function outputs a value of 1, and the alarm label is marked as "requires downgrading," indicating that the uncertainty of the current diagnostic conclusion exceeds the system's tolerance range. This conclusion, after being reported to the cloud master station, should be considered a low-confidence result, reminding the master station to treat it with caution and potentially triggering supplementary testing or manual review. When the information entropy score is less than or equal to the preset tolerance threshold, the indicator function outputs a value of 0, and the alarm label is marked as "credible," indicating that the uncertainty of the current diagnostic conclusion is within the system's tolerance range. This conclusion can be directly incorporated into the master station's scheduling decision. Furthermore, the alarm label and information entropy score are added as new dimension elements and appended to the end of the line abnormal state vector to form a dimensionally expanded labeled state vector. Here, the direct concatenation operation adds the information entropy score and alarm label as independent new dimension elements to the end of the vector without changing the original data structure and values ​​of each component of the line abnormal state vector. This allows the attached state vector to simultaneously carry the diagnostic conclusion content (predicted probability of each abnormal category and estimated physical parameter value) and confidence assessment information (information entropy score and alarm label) in a single data object.

[0047] Specifically, the spatiotemporal fusion and differential coding unit 352 is used to perform spatiotemporal information fusion and differential compression coding on the tag state vector to obtain a compressed binary message. In this process, firstly, the precise timestamp of the current data acquisition time provided by the onboard real-time clock module of the sensor board is read, along with the spatial location identifier pre-written into the firmware of the sensor node during the deployment phase. Next, the timestamp and spatial location identifier are converted into a compact binary representation according to a predefined encoding format, and then concatenated and fused with the binary encoded representation of the tag state vector at the bitstream level to form a complete information payload containing diagnostic content, confidence assessment, acquisition time, and spatial location. Then, differential compression coding is performed on the fused information payload. Differential compression coding utilizes the correlation between data in adjacent transmission cycles to compress redundant information. Specifically, a copy of the tag state vector successfully transmitted last time is retained in the local cache as a reference baseline. The tag state vector of the current cycle is differentially calculated element-wise with this reference baseline to obtain the difference vector (differential vector) between the current cycle and the previous cycle. For elements in the label state vector with small numerical changes (e.g., when the conductor state does not change drastically during a continuous monitoring period, the probability distribution and the change in physical parameters are close to zero), the difference value approaches zero. These near-zero difference values ​​can be encoded using very few bits, significantly compressing the data volume. The unit employs a variable-length encoding strategy to encode the difference vector in binary: elements with difference values ​​of zero or close to zero are assigned shorter codewords, while elements with larger difference values ​​are assigned longer codewords. This ensures that the total number of bits after encoding is far lower than the number of bits required for direct fixed-length encoding of absolute values ​​when the data change is small. For the first transmission (i.e., when there is no reference baseline from the previous period) or scenarios where the difference between the current period's data and the previous period's data is too large (e.g., when the conductor abruptly changes from a normal state to a violent galloping state, causing the difference values ​​to deviate significantly from zero), the unit automatically reverts to full absolute value encoding mode to ensure that the integrity of diagnostic information is not lost due to differential encoding. After differential compression encoding, the final output is a compressed binary message. This message carries complete tag state vector information and spatiotemporal metadata with a minimized bit length. Its data volume has been compressed to a range that matches the single-frame payload capacity of low-power wide area network protocols.

[0048] Specifically, the adaptive sleep time estimation unit 353 is used to perform communication header encapsulation and adaptive sleep time estimation on the compressed binary message based on the Low Power Wide Area Network (LPWAN) protocol and a negative exponential decay function to obtain an optimized status report. In this process, firstly, the compressed binary message is encapsulated with a communication header based on the LPWAN protocol. Specifically, according to the frame format specification of the LPWAN protocol (such as LoRa or NB-IoT) adopted by the system, a complete protocol frame structure is constructed for the compressed binary message. A synchronization word (used for frame synchronization detection at the receiver), a device address field (containing a unique device identifier for the sensor node, enabling the base station and master station to identify the data source), a frame type identifier field (indicating that the current frame is a status monitoring report type), and a frame payload length field (recording the number of bytes in the compressed binary message for correct parsing by the receiver) are added sequentially to the front of the compressed binary message. Simultaneously, a Cyclic Redundancy Check (CRC) code is added to the rear of the compressed binary message for the receiver to verify the integrity of the frame data, ensuring that possible bit errors during wireless transmission can be detected. After encapsulation with the communication header, the compressed binary message is wrapped into a complete communication frame conforming to the transmission format requirements of the Low Power Wide Area Network (LPWAN) protocol, enabling it to be correctly received, identified, and parsed by the base station and the cloud master station. Simultaneously, a severity index is extracted from the attached state vector. This index is based on a comprehensive assessment of the maximum predicted probability value of the probability distribution array in the line anomaly state vector and its corresponding anomaly category severity level. A higher severity index indicates a more critical current line anomaly state. The current severity index is then substituted into a negative exponential decay function to calculate the sleep time. This process can be expressed by the formula: in, Minimum forced sleep interval, Basic dormancy constant, This represents the severity index of the current situation. This is the event severity attenuation coefficient. This is the base of the natural logarithm. Furthermore, the complete communication frame, encapsulated in the communication header, is integrated and packaged with the adaptive sleep time parameter to generate an optimized status report. The optimized status report, as the final output of the entire intelligent sensor-based transmission line status monitoring system, contains two core components: First, a compressed binary message encapsulated in the communication header, carrying complete diagnostic conclusions of abnormal line conditions (predicted probability distributions and physical parameter estimates for each abnormality category), confidence assessment information (information entropy score and alarm tags), and spatiotemporal metadata of data acquisition (timestamps and spatial location identifiers). This content is sent to the cloud master station via a low-power wide-area network wireless link. After receiving the message, the master station performs frame de-framing, CRC verification, and payload parsing according to the protocol frame format to restore all information fields in the attached status vector, thereby making line safety situation assessment and operation and maintenance scheduling decisions. Second, an adaptive sleep time parameter, which is written into the sensor's local power management module. After the report is sent, the sensor enters a low-power sleep mode of the corresponding duration, shutting down the power supply to the high-frequency sampling subsystem, inference processor, and wireless communication module, retaining only the minimum maintenance power of the real-time clock and low-power wake-up circuit. After the sleep time is reached, the wake-up circuit triggers the system to re-enter the status perception and trigger judgment process, starting the next complete monitoring cycle.

[0049] As described above, the intelligent sensor-based transmission line condition monitoring system 300 according to the embodiments of this application can be implemented in various wireless terminals, such as servers with intelligent sensor-based transmission line condition monitoring algorithms. In one possible implementation, the intelligent sensor-based transmission line condition monitoring system 300 according to the embodiments of this application can be integrated into the wireless terminal as a software module and / or hardware module. For example, the intelligent sensor-based transmission line condition monitoring system 300 can be a software module in the operating system of the wireless terminal, or it can be an application developed for the wireless terminal; of course, the intelligent sensor-based transmission line condition monitoring system 300 can also be one of many hardware modules of the wireless terminal.

[0050] Alternatively, in another example, the smart sensor-based transmission line condition monitoring system 300 and the wireless terminal can also be separate devices, and the smart sensor-based transmission line condition monitoring system 300 can be connected to the wireless terminal via wired and / or wireless networks, and transmit interactive information in accordance with an agreed data format.

[0051] The various embodiments of this disclosure have been described above. These descriptions are exemplary and not exhaustive, nor are they limited to the disclosed embodiments. Many modifications and variations will be apparent to those skilled in the art without departing from the scope and spirit of the described embodiments. The terminology used herein is chosen to best explain the principles, practical application, or improvement of the technology in the market, or to enable others skilled in the art to understand the embodiments disclosed herein.

Claims

1. A power transmission line condition monitoring system based on intelligent sensors, characterized in that, include: The state perception and trigger judgment module is used to perform physical energy level conversion and peak threshold comparison on the acquired induced current value, battery temperature and pressure vector and simulated vibration envelope signal to obtain the energy boundary constraint matrix and transient trigger flag. The high-frequency transient data acquisition and energy budget module is used to perform high-frequency sampling of the conductor motion state within the maximum allowable time window based on the available energy Joule values ​​contained in the energy boundary constraint matrix in response to the transient triggering flag, so as to obtain high-frequency dynamic sequence data and the remaining energy constraint matrix; The computing resource adaptation module is used to perform frequency domain filtering and tensor dimension reshaping on high-frequency dynamic sequence data, and to perform computing power upper limit matching and conversion in combination with the remaining energy constraint matrix to obtain the normalized dynamic tensor and the target exit node identifier. The restricted reasoning and preliminary anomaly diagnosis module is used to perform feedforward blocking and local classifier reasoning on the normalized dynamic tensor based on the target exit node identifier to obtain the line anomaly state vector and information entropy score. The report optimization and wireless transmission adaptation module is used to perform confidence downgrading assessment and redundant field compression encoding on the line abnormal status vector and information entropy score to generate an optimized status report that matches the communication bandwidth and is used for scheduling by the cloud master station.

2. The power transmission line condition monitoring system based on intelligent sensors according to claim 1, characterized in that, The state awareness and trigger judgment module is used for: Energy and charging slope are calculated based on the induced current value and the battery temperature-pressure vector to generate initial energy state parameters; Based on the exponential smoothing penalty function, the initial energy state parameters are standardized and adaptively calculated to obtain the energy boundary constraint matrix and dynamic correction threshold. The peak amplitude of the simulated vibration envelope signal is extracted and compared with the dynamic correction threshold in real time to obtain the transient trigger flag.

3. The power transmission line condition monitoring system based on intelligent sensors according to claim 1, characterized in that, The high-frequency transient data acquisition and energy budgeting module is used for: Based on the system's dynamic power consumption parameters, the available energy extracted from the energy boundary constraint matrix and the transient trigger flag are estimated for duration and multiplied to calculate the blocking effect, so as to obtain the maximum sampling time window; Data acquisition and splicing within the maximum sampling time window, along with dynamic power consumption accumulation, are performed to obtain high-frequency dynamic sequence data and an estimated value of actual energy consumption. The energy boundary constraint matrix and the estimated actual energy consumption are subjected to energy reduction and reconstruction of the remaining constraint matrix to obtain the remaining energy constraint matrix.

4. The power transmission line condition monitoring system based on intelligent sensors according to claim 1, characterized in that, The computing resource adaptation module includes: The frequency domain filtering and tensor dimension folding and reshaping unit is used to perform frequency domain filtering and tensor dimension folding and reshaping on high-frequency dynamic sequence data to obtain normalized dynamic tensors. The energy efficiency conversion and network topology extraction unit is used to perform energy efficiency conversion and network topology extraction on the remaining energy constraint matrix based on basic energy consumption indicators and scheduling overhead parameters, so as to obtain the maximum available computing power limit and multi-exit network configuration table. The target exit node deep optimization unit is used to perform multi-target computing power matching and exit node deep optimization on the multi-exit network configuration table and the maximum available computing power limit to obtain the target exit node identifier.

5. The power transmission line condition monitoring system based on intelligent sensors according to claim 1, characterized in that, The restricted reasoning and preliminary anomaly diagnosis module is used for: The normalized dynamic tensor is processed by computational power allocation, blocking and activation extraction using the target exit node identifier to obtain a truncated feature map. Based on the local classifier and regressor awakened by the target exit node identifier, the truncated feature map is fully connected and concatenated with physical parameters to obtain the probability distribution array and the line abnormal state vector. Discrete entropy integral calculation and confidence inverse quantization evaluation are performed on the classification prediction probability values ​​contained in the probability distribution array to obtain the information entropy score.

6. The power transmission line condition monitoring system based on intelligent sensors according to claim 1, characterized in that, The report optimization and wireless transmission adaptation module includes: The alarm label determination and matrix splicing unit is used to perform alarm label determination and matrix dimension splicing on the information entropy score and line abnormal state vector through preset tolerance threshold and indicator function logic, so as to obtain the labeled state vector. The spatiotemporal fusion and differential coding unit is used to perform spatiotemporal information fusion and differential compression coding on the label state vector to obtain a compressed binary message; The adaptive sleep time estimation unit is used to perform communication header encapsulation and adaptive sleep time estimation on compressed binary messages based on low-power wide area network protocol and negative exponential decay function to obtain an optimized status report.

7. The power transmission line condition monitoring system based on intelligent sensors according to claim 6, characterized in that, The adaptive sleep time calculation unit is used to: adaptively calculate the sleep time of compressed binary messages using the following formula: in, Minimum forced sleep interval, Basic dormancy constant, This represents the severity index of the current situation. This is the event severity attenuation coefficient. is the base of the natural logarithm.

8. The power transmission line condition monitoring system based on intelligent sensors according to claim 4, characterized in that, The target exit node depth optimization unit is used for: The inherent complexity of the normalized dynamic tensor is dynamically evaluated to obtain the tensor complexity index; The node optimization reward matrix is ​​reconstructed by combining the node marginal utility and energy conservation joint reward matrix with the tensor complexity index, the multi-exit network configuration table and the maximum available computing power limit; The maximum utility node is locked under the hard energy constraint of the node optimization benefit matrix, the multi-exit network configuration table and the maximum available computing power limit to obtain the target exit node identifier.

Images