Iot platform data real-time processing and intelligent analysis system based on edge computing

Abstract

This invention relates to the field of edge computing technology and discloses a real-time data processing and intelligent analysis system for an IoT platform based on edge computing. The system includes: acquiring multi-source sensor data from a distribution transformer and performing time-scale normalization; performing frequency domain transformation on electrical parameter waveforms to extract multiple types of heat generation power and superimposing them to generate a total heat source; combining the total heat source and component metadata to recursively calculate the heat-retaining particle state quantity reflecting heat transfer hysteresis; obtaining the temperature distribution of hidden hotspots through a swarm optimization algorithm; generating insulation aging and damp heat activation indicators based on the extreme temperatures and locations of the hotspots, combined with moisture activity; calculating the sensitivity of each data point to the hotspots to allocate edge computing power and upload bandwidth; updating the device parameter library based on algorithm evolution and generating asset maintenance ranking scores. This invention effectively reduces the spatiotemporal lag of data under conditions of limited computing power and bandwidth, realizing dynamic analysis and adaptive resource allocation of hidden thermal states.

Description

Technical Field

[0001] This invention relates to the field of edge computing technology, and more specifically, to a real-time data processing and intelligent analysis system for an Internet of Things (IoT) platform based on edge computing. Background Technology

[0002] As a core component of the power grid, the operating status of distribution transformers directly affects the safety and stability of power supply. Currently, the industry widely adopts IoT gateways combined with external sensing units (such as oil temperature probes, voltage transformers, current transformers, and vibration probes) for online monitoring of equipment. However, due to limitations in the deployment conditions of the transformer substation, the limited computing power of the edge gateway, and the limited bandwidth of the communication link, conventional solutions often can only collect and upload superficial data. Heat transfer within the transformer exhibits complex hysteresis effects. After heat is generated in the internal windings, it must pass through multiple stages, including the flow of insulating oil, the obstruction of gaskets, and the cooling of the radiator, before being conducted to external sensors. This results in the real-time data received by the platform showing significant spatial misalignment, time lag, and thermal inertia superposition when reflecting the actual internal high-heat areas and insulation aging status. Traditional monitoring solutions mostly only set fixed alarm thresholds for surface oil temperature or current amplitude, making it difficult to effectively reflect the actual dynamic high-heat distribution inside, easily leading to missed early latent hazards. Furthermore, the indiscriminate uploading of massive amounts of data also wastes limited bandwidth and causes computational congestion. Summary of the Invention

[0003] This invention provides a real-time data processing and intelligent analysis system for an IoT platform based on edge computing, which solves the technical problems mentioned in the background.

[0004] This invention provides a real-time data processing and intelligent analysis system for an IoT platform based on edge computing, applied to a transformer monitoring system including edge gateways and multi-source sensing units, configured to execute: The multimodal raw acquisition data of the multi-source sensing unit is acquired, the arrival time interval of the raw acquisition data is extracted and the median absolute deviation is calculated, and an adaptive time smoothing scale is generated to perform time-scale alignment on the raw acquisition data to obtain a unified time-scaled stream data frame that excludes asynchronous sampling. The electrical parameter waveforms in the unified time-stamped stream data frame are subjected to discrete Fourier transform to separate the frequency domain components. The fundamental copper loss power of the winding, the additional heating power of harmonics and the heating power of the core loss are extracted and superimposed to generate the total heat source power that characterizes the overall heating degree. Based on the total heat source power and the pre-acquired transformer component metadata, the heat transfer hysteresis effect is introduced by recursive calculation to calculate the oil passage heat retention accumulation particle state quantity characterizing the oil passage heat transfer characteristics. The state variables of the accumulated heat in the oil passage are input into a preset particle swarm algorithm model for temperature field mapping. The individual fitness and the optimal fitness of all particles are calculated using the actual observation residuals. Based on this, the particle positions are iteratively updated, and a hidden hot spot field representing the true internal temperature is generated. The extreme temperatures and axial positions of the hot spots are extracted from the hidden hot spot field. The thermal field gradient is evaluated by combining the pre-extracted water activity of the oil paper, and an insulation thermal aging factor and a damp heat activation risk index characterizing the insulation state are generated. Calculate the disturbance sensitivity of the hidden hotspot field to each original collected data, allocate computing power share and communication bandwidth budget to each data based on the disturbance sensitivity to eliminate redundant interference, and aggregate to generate feature stream data to be uploaded; The model evolution step size is extracted based on the improvement of the optimal fitness of all particles within adjacent time scales to update the device-level thermal parameter library in the edge gateway. The asset health ranking score for multi-device scheduling decision is generated by fusion calculation using the time-series derivatives of the insulation thermal aging factor, the damp heat activation risk index, and the hot spot extreme temperature.

[0005] The beneficial effects of this invention are as follows: by constructing particle state variables that record the heat conduction hysteresis process at the edge and using particle swarm optimization algorithm to invert the internal hidden temperature field, dynamic tracking of latent hotspots and insulation aging trends is achieved without relying on fixed alarm thresholds; at the same time, based on the sensitivity of sensing data to hotspot extreme values, edge computing power and upload bandwidth are adaptively allocated, which not only reduces the communication pressure of invalid redundant data, but also improves the overall platform resource utilization efficiency, providing an objective basis for operation and maintenance scheduling for large-scale device clusters. Attached Figure Description

[0006] Figure 1 This is a flowchart of the real-time data processing and intelligent analysis system for the IoT platform based on edge computing, as described in this invention. Detailed Implementation

[0007] The subject matter described herein will now be discussed with reference to exemplary embodiments. It should be understood that these embodiments are discussed only to enable those skilled in the art to better understand and implement the subject matter described herein, and changes may be made to the function and arrangement of the elements discussed without departing from the scope of this specification. Various processes or components may be omitted, substituted, or added as needed in the examples. Furthermore, features described in some examples may be combined in other examples.

[0008] like Figure 1 As shown, an IoT platform data real-time processing and intelligent analysis system based on edge computing is applied to a transformer monitoring system that includes edge gateways and multi-source sensing units, and is configured to execute: The multimodal raw acquisition data of the multi-source sensing unit is acquired, the arrival time interval of the raw acquisition data is extracted and the median absolute deviation is calculated, and an adaptive time smoothing scale is generated to perform time-scale alignment on the raw acquisition data to obtain a unified time-scaled stream data frame that excludes asynchronous sampling. The electrical parameter waveforms in the unified time-stamped stream data frame are subjected to discrete Fourier transform to separate the frequency domain components. The fundamental copper loss power of the winding, the additional heating power of harmonics and the heating power of the core loss are extracted and superimposed to generate the total heat source power that characterizes the overall heating degree. Based on the total heat source power and the pre-acquired transformer component metadata, the heat transfer hysteresis effect is introduced by recursive calculation to calculate the oil passage heat retention accumulation particle state quantity characterizing the oil passage heat transfer characteristics. The state variables of the accumulated heat in the oil passage are input into a preset particle swarm algorithm model for temperature field mapping. The individual fitness and the optimal fitness of all particles are calculated using the actual observation residuals. Based on this, the particle positions are iteratively updated, and a hidden hot spot field representing the true internal temperature is generated. The extreme temperatures and axial positions of the hot spots are extracted from the hidden hot spot field. The thermal field gradient is evaluated by combining the pre-extracted water activity of the oil paper, and an insulation thermal aging factor and a damp heat activation risk index characterizing the insulation state are generated. Calculate the disturbance sensitivity of the hidden hotspot field to each original collected data, allocate computing power share and communication bandwidth budget to each data based on the disturbance sensitivity to eliminate redundant interference, and aggregate to generate feature stream data to be uploaded; The model evolution step size is extracted based on the improvement of the optimal fitness of all particles within adjacent time scales to update the device-level thermal parameter library in the edge gateway. The asset health ranking score for multi-device scheduling decision is generated by fusion calculation using the time-series derivatives of the insulation thermal aging factor, the damp heat activation risk index, and the hot spot extreme temperature.

[0009] The edge computing-based IoT platform data real-time processing and intelligent analysis system provided in this embodiment is applied to a transformer monitoring system that includes an edge gateway and multi-source sensing units. The edge gateway includes an industrial-grade edge computing terminal, a smart fusion terminal for the distribution area, and an IoT gateway with computing power support. The minimum hardware configuration of the edge gateway is a dual-core ARM Cortex-A53 or higher processor, with at least 512MB of memory and at least 4GB of storage capacity, supporting common industrial data transmission protocols such as Modbus, OPCUA, and HTTP / HTTPS. The multi-source sensing units include voltage transformers, current transformers, platinum resistance oil temperature sensors, ambient temperature sensors, and tank surface temperature sensors, deployed at corresponding monitoring points of the distribution transformer to collect multi-modal operating data during transformer operation. The sampling frequency of the voltage and current transformers is at least 2.5kHz, and the sampling frequency of the temperature sensors is at least 1Hz.

[0010] The system's entire execution sequence is triggered at a fixed period, consistent with the temperature sensor's sampling period, ranging from 1 to 60 seconds, and adjustable according to on-site monitoring accuracy requirements. The entire process steps are executed sequentially according to a serial dependency relationship, with the output of the previous step serving as the sole input to the next. The strong dependency between steps is: multi-source data timescale alignment processing. Total heat source power calculation Calculation of cumulative particle state variables of heat retention in oil passages Hidden hotspot field inversion Calculation of insulation condition assessment index Adaptive resource allocation and feature flow generation Equipment parameter library updates and asset health ranking score calculation. During system cold start, the initial loading of the equipment-level thermal parameter library is performed first, and then the full-process calculation is performed according to the above dependencies. When the transformer is shut down, the total heat source power and subsequent temperature field iteration calculations are terminated, and only the ambient temperature data acquisition and parameter library persistence operations are performed.

[0011] The first step is to acquire the multimodal raw data from the multi-source sensing units, perform time-stamp alignment processing, and obtain a unified time-stamped stream data frame that excludes asynchronous sampling. The specific steps are as follows: S101: Obtain the arrival time interval of adjacent samples of the same multi-source sensing unit, calculate the median of their absolute deviations and add a preset anti-zero bias to generate an adaptive time smoothing scale specific to the sensing unit.

[0012] Multimodal raw data refers to transformer operating data collected by multi-source sensing units according to a preset sampling period, including three-phase voltage, three-phase current, top oil temperature, ambient temperature, and tank surface temperature. Each type of sensing unit has a unique identifier j, where j is a positive integer. The arrival time interval between adjacent samples refers to the difference in timestamps between two consecutive data uploads from the same sensing unit j, denoted as . ,in Let n be the arrival timestamp of the nth sample. This is the arrival timestamp of the (n-1)th sample, in seconds.

[0013] The median absolute deviation (MAD) refers to the absolute deviation of each value in a sequence of consecutive arrival time intervals for the same sensing unit j from the median of the sequence, and then the median of this absolute deviation sequence is calculated. It is used to characterize the dispersion of sampling arrival times and can effectively suppress outlier interference caused by sudden communication delays. The preset zero-bias setting is a preset minimum positive real number, denoted as... The range of values ​​is In this embodiment, we take This is used to avoid errors caused by dividing by zero in subsequent calculations.

[0014] The formula for calculating the adaptive time smoothing scale is: in, For the k-th processing time scale, the adaptive time smoothing scale corresponding to the sensing unit identified as j, in seconds.

[0015] S102: Obtain the current processing time stamp of the system and the corresponding original acquisition time, and calculate the absolute time difference between the two to quantify the degree of data lag.

[0016] The current processing timestamp refers to the current moment when the edge gateway is performing data processing operations, denoted as . The original acquisition time refers to the moment when the sensing unit acquires the corresponding original data, denoted as . All units are seconds. The absolute time difference is the absolute value of the difference between the current processing timescale and the original acquisition time, denoted as . This value is used to characterize the degree of lag between the original data and the current processing time; the larger the value, the worse the timeliness of the data.

[0017] S103: Divide the absolute time difference by the adaptive time smoothing scale and take the opposite number, perform natural exponential decay operation, and generate a time decay weight that characterizes the validity of the data at the current moment.

[0018] The time-delay weight is used to assign weights to raw data collected at different times. Older data with higher lag levels are assigned lower weights to mitigate computational errors caused by asynchronous sampling and timing misalignment. The formula for calculating the time-delay weight is: in, For the k-th processing time scale, the sensing unit identified as j... The time-effect decay weight corresponding to the raw data collected at each moment is a dimensionless parameter with a value range of (0,1). When the raw collection time is exactly the same as the current processing time scale, the time-effect decay weight takes the maximum value of 1; as the absolute time difference increases, the time-effect decay weight decays exponentially.

[0019] S104: The raw data of each time node of the same sensing unit is weighted and accumulated with the corresponding time decay weight, and then normalized by dividing by the sum of the time decay weights to obtain normalized streaming data that eliminates timing misalignment.

[0020] The raw data refers to the sensing unit identified as j in The raw values ​​collected at each moment are denoted as The units are consistent with the physical units of the corresponding sensing quantities, with three-phase voltage in volts, three-phase current in amperes, and temperature data in degrees Celsius. Weighted accumulation refers to summing all raw data within a preset time window by multiplying them by their corresponding time-degradation weights. The preset time window is a continuous time interval ending at the current processing time point, with an interval length set to 3 to 10 times the sampling period of the corresponding sensing unit. In this embodiment, the time window length for electrical parameter data is set to 5 times the sampling period, and the time window length for temperature data is set to 3 times the sampling period. During system cold start, if the amount of effective sampled data within the time window is less than a preset threshold (the threshold is set to 50% of the number of sampling points corresponding to the time window length), then the time-degradation weight of all effective data is uniformly set to 1, and an equal-weighted average calculation is performed.

[0021] The formula for calculating normalized streaming data is: in, This is the normalized streaming data corresponding to sensor unit j at the k-th processing time scale, with units consistent with the physical units of the corresponding original data. The denominator is the sum of all time-degradation weights within a preset time window, used to normalize and correct the weighted accumulation result, eliminating numerical deviations caused by fluctuations in the total weight sum. When the calculated denominator is less than... At that time, the normalized streaming data directly takes the latest valid raw sample value within the time window.

[0022] S105: Time-series splicing of normalized stream data from various multi-source sensing units to generate a unified time-scaled stream data frame containing variables of three-phase voltage, three-phase current, oil temperature, and ambient temperature.

[0023] Timing splicing refers to combining the normalized stream data corresponding to all sensing units under the same processing time scale k according to the preset variable order to form a structured data stream. The unified time scale stream data frame is the set of all normalized stream data under the same processing time scale, denoted as S(k), and adopts a fixed-length binary frame structure. The frame structure is divided into four parts in sequence: frame header segment, time scale segment, data segment, and check segment. The frame header segment is 2 bytes long and fixed at 0xAA55, used for frame synchronization identification; the time stamp segment is 8 bytes long, using a 64-bit unsigned integer to store the Unix timestamp corresponding to the current processing time stamp, with a precision of milliseconds; the data segment is 56 bytes long, storing eight 32-bit single-precision floating-point variables in sequence: phase A voltage, phase B voltage, phase C voltage, phase A current, phase B current, phase C current, top oil temperature, ambient temperature, and enclosure surface temperature; the check segment is 2 bytes long, using the CRC16-Modbus check algorithm, checking all bytes of the frame header segment, time stamp segment, and data segment, used to check for errors during data transmission and storage. When a sensing unit has no valid normalized stream data, the corresponding data segment position is filled with a preset invalid value 0x7FC00000, i.e., a single-precision floating-point NaN value.

[0024] The second step involves performing a Discrete Fourier Transform on the electrical parameter waveforms in the unified time-scaled data frame to separate the frequency domain components, extracting the fundamental copper loss power of the winding, the additional harmonic heating power, and the core loss heating power, and then superimposing them to generate the total heat source power characterizing the overall heating level. The specific steps are as follows: S201: Extract the fundamental frequency using the built-in synchronous phase-locked loop of the edge gateway, divide the current sampling frequency by twice the fundamental frequency and round down to filter out high-frequency noise, and determine the highest harmonic analysis order that meets the Nyquist limit.

[0025] The synchronous phase-locked loop (PLL) is a digital PLL module built into the edge gateway. It adopts a second-order generalized integrator structure and is used to extract the fundamental frequency of the power grid from the three-phase voltage waveform. The fundamental frequency is denoted as . The unit is Hertz. The current sampling frequency refers to the sampling frequency at which the multi-source sensing unit acquires the electrical parameter waveform, denoted as... The unit is Hertz. The sampling frequency must satisfy the Nyquist sampling theorem, that is, the sampling frequency is greater than twice the frequency of the highest signal to be analyzed. In this embodiment, the sampling frequency of the electrical parameters is fixed at 256 times the rated fundamental frequency, which is 50 Hertz.

[0026] The formula for calculating the highest harmonic order is: in, The highest harmonic analysis order under the k-th processing time scale is a positive integer; This is for floor function. The highest harmonic analysis order determined by this formula can effectively filter out high-frequency noise exceeding the Nyquist limit, avoiding calculation errors caused by spectral aliasing.

[0027] S202: Perform windowed short-time discrete Fourier transform on the electrical parameter waveforms in the unified time-stamped stream data frame to extract the amplitude of harmonic current and voltage of each phase and order.

[0028] Electrical parameter waveforms refer to the three-phase voltage waveforms and three-phase current waveforms contained in a unified time-scaled data frame. Specifically, at the k-th processing time scale, the voltage waveform of the p-th phase includes... Let there be discrete sampling points, denoted as... The current waveform of phase p is denoted as p is the phase sequence identifier, with values ​​of 1, 2, and 3, corresponding to phases A, B, and C respectively; r is the index of the discrete sampling point, with a value range of... , The window length for the Fourier transform is set to the number of sampling points per period corresponding to the current fundamental frequency, i.e. The short-time discrete Fourier transform uses a Hanning window to window the time-domain waveform. The expression for the window function of the Hanning window is as follows: Where r is the sampling point index, the window function is multiplied by the original time-domain waveform before performing a Fourier transform to suppress spectral leakage. The overlap rate of the window functions for two adjacent processing time scales is set to 50% to ensure the temporal continuity of the frequency domain analysis.

[0029] The short-time discrete Fourier transform is used to convert time-domain electrical parameter waveforms into frequency-domain components, extracting the amplitude of each harmonic order. The calculation formula is as follows: in, The value of the harmonic current of the p-th phase and h-th order under the k-th processing time scale is expressed in amperes. The harmonic voltage amplitude of the p-th phase and h-th order at the k-th processing time scale is given in volts; h is the harmonic order, with a range of values ​​of... j is the imaginary unit; Let π be the mathematical constant pi. When h=1, the corresponding amplitudes are the fundamental current amplitude and the fundamental voltage amplitude, denoted as π / 2 and π / 2, respectively. and .

[0030] S203: Obtain the phase winding dynamic resistance after temperature correction based on the previous time scale, multiply it by the square of the fundamental current amplitude, and accumulate it on the three phases to separate the purely resistive heating characteristics, thereby obtaining the fundamental copper loss power of the winding.

[0031] The phase winding dynamic resistance refers to the DC resistance value of the transformer winding at a corresponding temperature, denoted as . The unit is ohms, and its value changes with the winding temperature. The correction formula follows the linear relationship between the resistance of the copper conductor and the temperature change, and the expression is: ,in The rated DC resistance of the winding at 75 degrees Celsius can be obtained from the transformer's nameplate specifications. This is the average winding temperature obtained from the previous timescale inversion, in degrees Celsius. During system cold start, the initial value of the average winding temperature is set to the current ambient temperature.

[0032] The fundamental copper loss power of the transformer winding is the active power loss generated by the transformer winding under the action of the fundamental current, which is entirely converted into heat. The calculation formula is as follows: in, The fundamental copper loss power of the winding at the kth processing time scale is expressed in kilowatts. is the dimension conversion coefficient used to convert watts to kilowatts; the summation operation traverses the three-phase windings to summarize the three-phase fundamental copper losses.

[0033] S204: Extract the preset equipment harmonic eddy current heat dissipation coefficient and core loss coefficient, multiply them by the square of the harmonic current amplitude and the square of the harmonic voltage amplitude of the corresponding order, and accumulate them in the frequency dimension within the range from the second to the highest harmonic analysis order to generate the harmonic additional heating power and core loss heating power, respectively.

[0034] The equipment harmonic eddy current heat dissipation coefficient is a preset coefficient based on the transformer model and design parameters, denoted as . The unit is ohms per order, used to characterize the heating characteristics of winding eddy current losses at the corresponding harmonic order. Its value increases with increasing harmonic order, and its range is [value missing]. Where h is the harmonic order, This is the rated DC resistance of the winding. The core loss factor is a preset coefficient based on the transformer core material and structural parameters, denoted as... The unit is Siemens per order, used to characterize the heating characteristics of core hysteresis loss and eddy current loss at the corresponding harmonic order, with a value range of [value missing]. , where h is the harmonic order.

[0035] The additional heating power due to harmonics is the additional heating power of the winding caused by the harmonic current generated by the nonlinear load. The calculation formula is as follows: in, The additional heating power for the harmonics at the kth processing time scale, in kilowatts; Dimensional conversion factor, used to convert watts to kilowatts; double summation operation traverses the three-phase windings and phases 2 to 3 respectively. The order of harmonics is used to achieve the sum of additional heating from all phases and all orders of harmonics.

[0036] The heat generated by the core loss is the active power loss produced by the transformer core under the action of an alternating magnetic field. The calculation formula is as follows: in, The core loss heating power at the kth processing time scale is expressed in kilowatts. The dimension conversion factor is used to convert watts to kilowatts; the double summation operation iterates through the three phases and from 1 to... Harmonics of all orders are used to summarize the losses of the core in all phases and orders.

[0037] S205: The fundamental copper loss power of the winding, the additional heating power of harmonics, and the heating power of the core loss are linearly summed to combine multiple heat generation mechanisms and generate a total heat source power.

[0038] The total heat source power is the total heat generated by all heat-generating mechanisms inside the transformer. It is the core input for subsequent temperature field calculations, and the calculation formula is as follows: in, This represents the total heat source power at the k-th processing time scale, in kilowatts. When the transformer is shut down, the total heat source power is set to 0, terminating subsequent temperature field iteration calculations.

[0039] The third step involves using the total heat source power and pre-acquired transformer component metadata to introduce a heat transfer hysteresis effect through recursive calculations, and then calculating the oil passage heat retention accumulation particle state quantity, which characterizes the oil passage heat transfer characteristics. Specifically, the following steps are performed: S301: Parse the axial height characteristics and equivalent physical length of each oil circuit unit from the transformer component metadata, and use the pre-built local oil velocity state dictionary to match the local oil velocity estimation value corresponding to each algorithm particle.

[0040] The transformer component metadata is pre-stored within the edge gateway, corresponding to the transformer's design and structural parameters, including winding axial height, number and size of oil channels, cooling oil circuit structure, winding material parameters, rated capacity, and voltage level. An oil circuit unit is the smallest calculated unit for dividing the transformer's cooling oil circuits along the axial direction. Each oil circuit unit has a unique identifier l, where l is a positive integer. The number of oil circuit units is set to 10 to 20, adjustable according to the winding axial height. The axial height characteristic refers to the winding axial height corresponding to oil circuit unit l, denoted as... The unit is meters; the equivalent physical length refers to the equivalent flow path length of the cooling oil within the oil circuit unit l, denoted as... The unit is meters, and it is consistent with the axial height of the oil circuit unit.

[0041] The pre-built local oil velocity state dictionary is a mapping dictionary stored in the edge gateway, using oil circuit unit identifier, transformer load rate, and top oil temperature as keys. It is used to match the flow velocity of cooling oil within the oil circuit unit under corresponding operating conditions. The values ​​in the dictionary are pre-calibrated based on the fluid dynamics simulation results of the transformer cooling type. The estimated local oil velocity refers to the flow velocity of cooling oil within the oil circuit unit l corresponding to the i-th algorithm particle at the k-th processing time scale, denoted as... The unit is meters per second, and i is the unique identifier of the algorithm particle, with a value range from 1 to the total number of particle swarms. The total number of particle swarms is set to 30 to 100, and in this embodiment it is set to 50.

[0042] S302: Divide the equivalent physical length by the sum of the absolute value of the local oil velocity estimate and the zero-offset to avoid static singularities, and calculate the oil passage residence time that characterizes the delay in cooling medium flow.

[0043] Oil passage residence time refers to the time required for cooling oil to flow through the corresponding oil passage unit. It is used to characterize the hysteresis effect of heat transfer in the oil passage. The calculation formula is: in, The oil passage dwell time of the oil passage unit l corresponding to the i-th algorithm particle under the k-th processing time scale is in seconds; A preset zero-offset value is used to avoid division by zero errors when the local oil speed estimation value is 0.

[0044] 303: Perform natural exponential decay calculation by dividing the current time step by the oil passage residence time to simulate the physical decay process of heat loss over time and generate a heat retention factor characterizing the thermal inertia strength.

[0045] The current time step size refers to the time interval between two adjacent processing time steps, denoted as... The unit is seconds, consistent with the trigger cycle of the entire system execution. The heat retention factor is used to characterize the proportion of heat retained by the cooling oil during its residence within the oil circuit unit. A higher value indicates stronger thermal inertia and less heat attenuation.

[0046] The formula for calculating the heat retention factor is: in, Let be the heat retention factor of the oil circuit unit l corresponding to the i-th algorithm particle at the k-th processing time scale. It is a dimensionless parameter with a value range of (0,1). When the oil circuit residence time is much greater than the time scale step, the heat retention factor approaches 1, indicating that there is almost no heat attenuation; when the oil circuit residence time is much less than the time scale step, the heat retention factor approaches 0, indicating that the heat is almost completely lost.

[0047] S304: The total heat source power is spatially non-uniformly distributed using the prior weight of the end leakage magnetic field to restore the leakage magnetic heating characteristics, and the heat exchange loss calculated based on the temperature difference between oil and the environment is superimposed to generate the transient new thermal excitation for the current time scale.

[0048] The prior weight of end leakage flux is used to characterize the spatial distribution characteristics of heat generation caused by leakage flux at the ends of transformer windings. The leakage flux density is higher at the winding ends, resulting in more significant heat generation; therefore, the weight corresponding to the ends is larger. The calculation process for the prior weight of end leakage flux is as follows: First, the axial height of the oil circuit unit l is normalized, and the calculation formula is as follows: in, Let be the normalized axial height of the oil circuit unit l, which is a dimensionless parameter with a value range of [-1, 1]. This represents the maximum axial height of the winding. The minimum axial height of the winding is given in meters. This is a preset zero-offset value. When the number of oil circuit units is 1, the normalized axial height is directly set to 0.

[0049] The prior weight of end leakage magnetic flux is calculated based on the normalized axial height, and the calculation formula is as follows: in, Let be the a priori weight of the end leakage magnetic flux corresponding to oil circuit unit l. This is a dimensionless parameter with a value range of (0,1), and the sum of all weights is 1. The summation operation iterates through all oil circuit units m to normalize the weights. For asymmetric winding structures, asymmetric weight distribution can be achieved by adjusting the offset of the normalized axial height.

[0050] Heat loss through heat exchange refers to the heat lost between the cooling oil within the oil circuit unit and the external environment through heat exchange. It is calculated based on the temperature difference between the oil and the environment and the surface heat transfer coefficient, and is expressed as follows: ,in The preset surface heat transfer coefficient is expressed in kilowatts per degree Celsius, and its value range is [value range missing]. kilowatts per degree Celsius; The normalized top oil temperature, The ambient temperature is normalized, and the unit is degrees Celsius.

[0051] The transient increase in thermal excitation is the amount of new heat input in oil circuit unit l at the current time scale, and the calculation formula is: in, For the k-th processing time scale, the transient thermal excitation of the oil circuit unit l corresponding to the i-th algorithm particle is added, with the unit being kilowatts.

[0052] S305: The heat retention factor is used to proportionally attenuate the heat retained by the upstream oil circuit unit in the previous time scale. The product of the transient new thermal excitation and the time scale step is added to complete the dimensional transformation, and the spatiotemporal heat transfer process is integrated to generate the current particle's oil circuit heat retention cumulative particle state quantity.

[0053] The upstream oil circuit unit refers to the adjacent oil circuit unit located upstream of the current oil circuit unit l in the direction of cooling oil flow, denoted as The heat retention of the upstream oil passage unit at the previous time scale is the cumulative particle state quantity of the heat retention in the oil passage of the upstream oil passage unit calculated at the previous time scale, denoted as... The unit is kilojoules. During system cold start, the initial value of the accumulated particle state quantity of oil passage heat in all oil circuit units is set to 0. The initial state quantity is calculated using ambient temperature and rated heat capacity parameters for the first processing time scale after startup.

[0054] The cumulative state quantity of heat retention in the oil passage is used to characterize the total heat retained in the cooling oil within the oil passage unit. It is a core state parameter reflecting the heat transfer characteristics and heat transfer hysteresis effect of the oil passage. The recursive calculation formula is as follows: in, At the kth processing time point, the oil passage heat accumulation particle state quantity of the oil passage unit l corresponding to the i-th algorithm particle, in kilojoules; The time step is in seconds, used to convert kilowatts to kilojoules for dimensional matching. The first term in the formula is the residual heat transferred from the upstream oil circuit unit after attenuation by the heat retention factor, and the second term is the total energy corresponding to the new thermal excitation at the current time step. The fusion of the two achieves accurate simulation of the spatiotemporal heat transfer process and hysteresis effect.

[0055] The fourth step involves inputting the state variables of the accumulated heat in the oil duct into a pre-defined particle swarm optimization model for temperature field mapping. The individual fitness and the optimal fitness of all particles are calculated using actual observation residuals. Based on this, the particle positions are iteratively updated, and a hidden hotspot field representing the true internal temperature is generated. The specific steps are as follows: S401: Based on the preset thermal capacity mapping matrix, the state variables of the accumulated heat particles in the oil passage are combined with the current ambient temperature, and a hidden segmented temperature field representing the internal heating mode of the winding is generated through affine transformation operation.

[0056] The heat capacity mapping matrix is ​​a parameter matrix preset based on the heat capacity characteristics of the transformer winding, containing constant terms. and ,in This is the heat-to-temperature conversion coefficient, measured in degrees Celsius per kilojoule, with a value range of [value missing]. Celsius / kilojoules; is the temperature difference correction coefficient, and is a dimensionless parameter with a value range of 0 to 1. Both are pre-set based on the structural parameters and material heat capacity characteristics of the transformer, and are the variables to be optimized in the particle swarm algorithm.

[0057] Affine transformation is used to convert the residual heat in the oil circuit into the corresponding temperature of the winding. The formula for calculating the hidden piecewise temperature field is as follows: in, For the k-th processing time scale, the hidden segment temperature of the oil circuit unit l corresponding to the i-th algorithm particle is expressed in degrees Celsius. The hidden segment temperatures of all oil circuit units are combined to form a hidden segment temperature field, which is used to characterize the temperature distribution along the axial direction inside the winding.

[0058] S402: Using the equipment surface observation mapping coefficient, the hidden segmented temperature field is extrapolated to the outer surface, and the observed oil temperature fitting value and the box temperature fitting value containing the influence of fluid convection are synthesized.

[0059] The surface observation mapping coefficients of the equipment are preset coefficients based on the thermal conductivity characteristics of the transformer, including the oil temperature mapping coefficient. Mapping coefficient with enclosure temperature , are dimensionless parameters, ranging from 0 to 1, used to characterize the proportional relationship between the internal temperature of the winding and the external oil temperature and the surface temperature of the housing, and are the variables to be optimized in the particle swarm optimization algorithm.

[0060] The observed oil temperature fitting value is the transformer top-level oil temperature fitting value obtained based on the hidden piecewise temperature field deduction, denoted as The unit is Celsius, and the calculation formula is: The summation operation iterates through all oil circuit units l.

[0061] The fitted value of the transformer tank temperature is the fitted value of the transformer tank surface temperature obtained based on the hidden piecewise temperature field derivation, denoted as . The unit is Celsius, and the calculation formula is: The summation operation iterates through all oil circuit units l.

[0062] S403: Calculate the deviation between the fitted value and the actual observed value in the unified time-scaled stream data frame, and construct a dynamic confidence weight by combining the median absolute deviation of the residuals to suppress anomalous noise.

[0063] Actual observed values ​​refer to the measured values ​​of top-layer oil temperature and tank surface temperature collected by the sensing unit in the unified time-scaled data frame, denoted as... , where q is the identifier of the observed variable, and the range of q covers all observed variables, including top oil temperature and tank surface temperature, all in degrees Celsius. The residual between the fitted value and the actual observed value is . ,in Let q be the fitted value of the observed variable q corresponding to the i-th algorithm particle.

[0064] The median absolute deviation of the residuals is the median of the absolute deviations of the residuals of all observed variables, used to characterize the dispersion of the residuals. The dynamic confidence weights are constructed based on the median absolute deviation of the residuals. For outliers whose absolute residual value is greater than 3 times the median absolute deviation, the weight is set to 0; for normal values ​​whose absolute residual value is less than or equal to 3 times the median absolute deviation, the weight is set to 1, thus effectively suppressing the interference of outlier sampling values ​​on model fitting. The dynamic confidence weights are denoted as... , is a dimensionless parameter with a value range of [0,1].

[0065] S404: Multiply the dynamic confidence weights by the normalized squared residuals, add the spatial temperature jump penalty calculated using the adjacent segment smoothness constraint term to correct for non-physical jumps, and generate the individual fitness that characterizes the model fit.

[0066] The normalized squared residual is the square of the residual divided by the historical variance of the corresponding observed variable. It is used to eliminate the influence of dimensions and achieve dimensionless processing of the residual. The expression is: ,in The variance of the observed variable q within a preset historical time window, This is a preset zero-bias amount. The adjacent segment smoothness constraint term is a preset weight coefficient, denoted as... , a dimensionless positive number ranging from 0.1 to 10, is used to constrain temperature jumps between adjacent oil circuit units, preventing abrupt temperature changes that violate the physical laws of heat conduction. The spatial temperature jump penalty is the product of the square of the hidden segment temperature difference between adjacent oil circuit units and the smoothness constraint term, calculated using the following formula: The summation operation iterates through all oil circuit units.

[0067] Individual fitness characterizes the degree of fit between the temperature field model corresponding to a single algorithm particle and actual observed data. A smaller value indicates a higher good fit and a model that more closely approximates the true temperature distribution. It is a dimensionless scalar. The formula for calculating individual fitness is: in, Let be the individual fitness of the i-th algorithm particle at the k-th processing time scale; Q is the set of all observed variables; The axial height difference between adjacent oil circuit units is expressed in meters and is used to eliminate the dimensional influence of the spatial step size. The summation operation iterates through all observed variables and all oil circuit units respectively.

[0068] S405: Initialize the variables to be optimized, hyperparameters and initial particle states of the particle swarm optimization algorithm. Using the particle swarm iteration rule, extract the convergence gradient based on the individual's historical best position and the overall optimal fitness of all particles. Adaptively update the optimization speed and spatial coordinates of each particle. Output the final converged optimal coordinates of all particles as a hidden hotspot field.

[0069] The complete list of variables to be optimized in the particle swarm optimization algorithm includes: the heat-to-temperature conversion coefficients corresponding to each oil circuit unit. Temperature difference correction factor Oil temperature mapping coefficient Box temperature mapping coefficient Surface heat transfer coefficient Local oil speed estimation value All variables to be optimized are given physically reasonable upper and lower limits. The upper and lower limits for dimensionless parameters are [0,1], and the upper and lower limits for dimensional parameters are set based on ±50% of the transformer's rated parameters to avoid non-physical parameter values ​​during the iteration process.

[0070] The core hyperparameters of the particle swarm optimization algorithm are set as follows: population size is set to 50, maximum number of iterations is set to 100, and cognitive learning factor is set to... Set to 1.49445, social learning factor The value is set to 1.49445. The upper and lower limits of the particle velocity clamping are set to ±10% of the corresponding variable's value range. The upper and lower limits of the particle position clamping are consistent with the upper and lower limits of the physical values ​​of the variable to be optimized.

[0071] The initial particle state generation rule is as follows: the initial position of the first particle is set to the nominal value corresponding to the rated parameter of the transformer, and the initial positions of the remaining 49 particles are randomly generated in a uniform distribution within the upper and lower limits of the value of the variable to be optimized. The velocity of all initial particles is initialized to 0.

[0072] The adaptive inertia weight is generated based on the statistical characteristics of individual fitness and is used to balance the algorithm's global search capability and local convergence capability. The calculation formula is as follows: in, Let be the adaptive inertia weight corresponding to the iterth iteration, and be a dimensionless parameter with a value range of [0.4, 0.9]. Let be the standard deviation of the fitness of all individual particles in the iterth iteration; This represents the average fitness of all individual particles in the iterth iteration; This is a preset zero-bias setting.

[0073] The individual historical optimal position refers to the spatial coordinates of the i-th algorithm particle when it achieves the minimum individual fitness in the iteration history, denoted as . The optimal position for the entire swarm refers to the spatial coordinates corresponding to the minimum individual fitness achieved by the entire particle swarm in its iteration history, denoted as . During the iteration process, after each iteration, the individual historical best position of each particle and the group best position of the entire particle swarm are updated.

[0074] Introduce uniformly distributed independent random numbers with values ​​ranging from [0,1]. and This is used to increase the randomness of the algorithm and avoid getting trapped in local optima. The update formula for the particle optimization speed is: in, Let be the optimization speed of the i-th algorithm particle in the (iter+1)-th iteration; Let be the optimization speed of the i-th algorithm particle in the iter-th iteration; Let be the spatial coordinates of the i-th algorithm particle during the iter-th iteration. After the velocity update is completed, a velocity clamping mechanism is used to limit velocities exceeding the upper and lower limits within the clamping range.

[0075] The formula for updating particle spatial coordinates is: After the coordinates are updated, a position clamping mechanism is used to restrict coordinates that exceed the upper and lower limits of the physical values ​​to within the range of values.

[0076] The algorithm convergence conditions include: the number of iterations reaches the preset maximum number of iterations (100); the optimal fitness of the swarm does not decrease for 20 consecutive iterations; and the standard deviation of the fitness of individual particles is less than the preset convergence threshold. The iteration terminates when any one of the conditions is met. If the algorithm fails to converge after reaching the maximum number of iterations, the position corresponding to the optimal fitness of the population during the iteration process is taken as the final result. After the iteration terminates, the optimal position of the population is... The corresponding hidden segmented temperature field output is the hidden hot spot field, which is the actual temperature distribution along the axial direction inside the transformer winding, and can accurately locate the high-temperature hot spots lurking inside.

[0077] The fifth step involves extracting the extreme temperatures and axial positions of the hotspots from the hidden hotspot field. This, combined with the pre-extracted moisture activity of the oil paper, is used to assess the thermal field gradient and generate an insulation thermal aging factor and a damp-heat activation risk index characterizing the insulation state. The specific steps are as follows: S501: Compare the temperatures of each spatial node in the hidden hotspot field, select the highest temperature extreme value as the hotspot extreme value temperature, and extract its corresponding spatial height feature sequence to confirm the hotspot axial position.

[0078] First, the hidden segment temperatures of all algorithm particles are weighted and fused to generate a deterministic global weighted segment temperature field. The weights for the weighted fusion are normalized particle contribution weights calculated based on the individual particle fitness, and the calculation formula is as follows: in, Let be the normalized particle contribution weight of the i-th algorithm particle at the k-th processing time scale. This is a dimensionless parameter with a value range of (0,1), and the sum of all weights is 1. The summation operation iterates through all algorithm particles m. When the individual fitness of all particles is greater than a preset threshold, the contribution weight of all particles is uniformly set to the average value 1 / N, where N is the total size of the particle swarm, to avoid the calculation error of the denominator approaching 0.

[0079] The formula for calculating the weighted piecewise temperature field is: in, For the k-th processing time scale, the weighted segmented temperature corresponding to oil circuit unit l is in degrees Celsius; the summation operation iterates through all algorithm particles.

[0080] Numerical comparisons are performed on the temperatures of all spatial nodes in the weighted piecewise temperature field, and the maximum value is selected as the hotspot extreme temperature. The calculation formula is as follows: in, Let be the hotspot extreme temperature at the k-th processing time scale, in degrees Celsius. The normalized axial height sequence of the oil circuit unit corresponding to the hotspot extreme temperature is marked as the hotspot axial position, denoted as . , which is a dimensionless parameter. When multiple identical maximum temperature extremes appear in the hidden hotspot field, the axial positions corresponding to all hotspots are retained, the evaluation index of each hotspot is calculated separately, and the maximum value is taken as the final output result.

[0081] S502: By subtracting the reciprocal of the sum of the hot spot extreme temperature and the thermodynamic conversion constant from the reciprocal of the thermodynamic reference constant, a natural exponential scaling operation is performed through the Arrhenius equation to generate an insulation thermal aging factor that quantifies insulation degradation.

[0082] The Arrhenius equation is a classic equation describing the relationship between the thermal aging rate of insulating paper and temperature. The thermal aging rate of insulating paper increases exponentially with increasing temperature. The thermodynamic reference constant is the ratio of the activation energy of thermal aging of ordinary cellulose insulating paper to the ideal gas constant, with a value of 15000 in Kelvin, corresponding to an activation energy of 124.7 kJ / mol, which is the industry-standard activation energy parameter for thermal aging of cellulose insulating paper. The reference temperature constant is 110 in degrees Celsius, representing the rated operating reference temperature of transformer windings. The thermodynamic conversion constant is 273 in Kelvin, used to convert Celsius temperature to thermodynamic temperature.

[0083] The insulation thermal aging factor is used to quantify the relative thermal aging rate of transformer insulation paper. The calculation formula is as follows: in, The insulation thermal aging factor is a dimensionless parameter for the k-th processing timescale. When the extreme hotspot temperature equals the reference temperature of 110 degrees Celsius, the insulation thermal aging factor is 1; when the extreme hotspot temperature is higher than the reference temperature, the insulation thermal aging factor is greater than 1 and increases exponentially with increasing temperature, used to quantify the accelerated rate of thermal aging of the insulation paper. The cumulative thermal aging amount can be obtained by integrating the insulation thermal aging factor over time, used to assess the remaining life of the insulation paper.

[0084] S503: The absolute value of the axial temperature gradient at the hot spot extreme temperature at the hot spot axial position is calculated using the central difference method, and then normalized by dividing by the global gradient deviation to highlight the degree of local heat concentration, thus obtaining the local hot spot concentration.

[0085] The central difference method is a commonly used method for numerically calculating the first derivative. It is used to calculate the temperature field gradient along the axial direction and can effectively reduce numerical calculation errors. The absolute value of the axial temperature gradient is the absolute value of the first derivative of the temperature change with axial height at the axial location of the hot spot. It is used to characterize the drasticness of local temperature changes, and the calculation formula is: in, This is the identifier for the oil circuit unit corresponding to the axial position of the hot spot; and This represents the normalized axial height of adjacent oil passage units; This is a preset zero-bias setting.

[0086] The global gradient deviation is the median absolute deviation of the axial gradient across the entire temperature field. It is used to normalize the absolute value of the gradient, eliminating the influence of overall gradient fluctuations under different operating conditions and highlighting the degree of local heat concentration. The formula for calculating the concentration of local hot spots is: in, Let be the local hotspot concentration at the kth processing time scale, which is a dimensionless parameter; This represents the median absolute deviation of the axial gradient across the entire temperature field. To prevent errors caused by dividing by zero, a preset zero bias value is used. A higher local hotspot concentration value indicates a greater degree of local heat concentration at the hotspot location, and a greater risk of insulation degradation.

[0087] S504: Extract the preset water migration activation energy parameter, combine it with the pre-extracted water activity of oil paper, and construct a thermal kinetic energy release term characterizing the ability to remove water at high temperatures through exponential calculation.

[0088] The water activity of oil-paper insulation refers to the relative activity of water in the transformer oil-paper insulation system, denoted as . , is a dimensionless parameter with a value range of [0,1]. It can be calculated using the oil-paper insulation moisture monitoring model based on the top oil temperature, ambient humidity, and transformer operating years, or directly collected by the corresponding moisture sensing unit. The moisture migration activation energy parameter is the Gibbs free energy change required for moisture migration in the oil-paper, denoted as . The value is taken as 40 kJ / mol, with the unit being joules / mol; the ideal gas constant is denoted as... The value is 8.314 joules / mole / Kelvin.

[0089] The thermal kinetic energy release term characterizes the ability of moisture to escape from the oiled paper at high temperatures. The higher the temperature, the larger the value of the thermal kinetic energy release term, and the higher the risk of moisture escape. The calculation formula is as follows: in, Let be the thermal kinetic energy release term at the k-th processing time scale, which is a dimensionless parameter.

[0090] S505: The thermal kinetic energy release term is cross-producted with the local hot spot concentration to couple the physical interaction between the temperature gradient and moisture movement, generating a humid heat activation risk index for early warning of bubble formation risk.

[0091] The damp heat activation risk index is used to couple the synergistic effect of thermal aging and moisture migration, providing early warning of the risk of bubble formation in oil-paper insulation. Bubble formation significantly reduces insulation strength, triggering partial discharge or even insulation breakdown. The formula for calculating the damp heat activation risk index is: in, , is the damp heat activation risk index at the k-th treatment timescale, and is a dimensionless parameter. This index is coupled with the moisture activity of the oil paper, the high-temperature thermal kinetic energy release effect, and the degree of local heat concentration. The larger the value, the higher the risk of bubble formation in the oil paper insulation and the greater the probability of insulation failure.

[0092] Step 6: Calculate the sensitivity of the hidden hotspot field to the disturbance of each original data set. Based on the disturbance sensitivity, allocate computing power and communication bandwidth budget to each data set to eliminate redundant interference. Aggregate the data to generate the feature stream data to be uploaded. The specific steps are as follows: S601: The partial derivatives of the extreme temperatures of hot spots with respect to each flow data variable are calculated using the numerical perturbation method. After dimensionless processing, the derivatives are divided by the historical statistical fluctuation range of the corresponding variable to eliminate dimensional differences, thereby generating a perturbation sensitivity that measures the contribution value of single-point data.

[0093] Numerical perturbation is a commonly used method for numerically calculating partial derivatives. It involves applying a small perturbation to the input variable and calculating the ratio of the change in the output variable to the perturbation value to obtain the partial derivative. The perturbation is set to 0.1% of the historical statistical standard deviation of the corresponding streaming data variable, ensuring that the output change caused by the perturbation is effectively detected without altering the system's operating state. For the streaming data variable labeled j... The value of the variable after applying the perturbation is ,in As a perturbation, the hotspot extreme temperature is recalculated to obtain the perturbed hotspot extreme temperature. .

[0094] The formula for calculating the partial derivative of the hotspot extreme temperature with respect to the stream data variable is: The unit of the partial derivative is degrees Celsius / [s_j], where [s_j] is the physical dimension of the corresponding flow data variable.

[0095] The partial derivatives are dimensionless to eliminate the dimensional differences between different variables. The formula for calculating the dimensionless partial derivatives is as follows: in, For streaming data variables The average value within a preset historical time window; The extreme temperature of the hotspot is the average value within a preset historical time window, and the dimensionless partial derivative is a dimensionless parameter.

[0096] Historical statistical fluctuation amplitude is the median absolute deviation of the numerical fluctuation of the corresponding streaming data variable within a preset historical time window, denoted as . This is used to further eliminate the impact of differences in variable fluctuation amplitude. The formula for calculating disturbance sensitivity is: The formula for calculating disturbance sensitivity is: in, For the k-th processing timescale, the perturbation sensitivity of the streaming data variable corresponding to sensor unit j is a dimensionless parameter. A larger perturbation sensitivity value indicates a higher contribution of the streaming data variable to the hotspot temperature calculation and a greater impact on the final analysis results.

[0097] S602: Obtain the preset baseline CPU time of the edge computing algorithm module, divide the perturbation sensitivity of various data by the baseline CPU time and take the square root, and allocate the computing power share representing the computing power tilt weight through global normalization operation.

[0098] The preset baseline CPU time refers to the baseline CPU time required for a single execution of the computation and processing task for the corresponding streaming data variable within the edge gateway, denoted as... The unit is milliseconds, where r is the identifier of the edge computing task, corresponding one-to-one with the streaming data variable, and can be pre-calibrated through the performance testing module of the edge gateway. The baseline CPU time is used to characterize the computing power consumption of the corresponding computing task; the longer the time, the greater the computing power consumption.

[0099] The computing power share is the proportion of CPU computing resources allocated to a corresponding computing task out of the total available computing power. It is used to achieve adaptive allocation of computing resources, allocating more computing resources to tasks with high contribution value and less to tasks with redundancy or interference, thereby improving computing power utilization efficiency. The formula for calculating the computing power share is: in, For the k-th processing time scale, the computing power share corresponding to the computing task identified as r is a dimensionless parameter with a value range of (0,1). The sum of the computing power shares of all computing tasks is 1. The summation operation traverses all edge computing tasks m. The minimum computing power share of each computing task is set to 1% to avoid the situation where the computing power allocation is 0, ensuring the normal execution of basic computing functions.

[0100] S603: Monitor the current network interface status of the edge gateway to obtain the available uplink communication bandwidth, allocate the available uplink communication bandwidth proportionally according to the disturbance sensitivity ratio of each data, and obtain the communication bandwidth budget that represents the transmission priority limit.

[0101] Available uplink bandwidth refers to the maximum bandwidth of the uplink communication link currently available to the edge gateway, denoted as . The unit is bytes per second, which can be obtained in real time through the network interface status monitoring module of the edge gateway. The communication bandwidth budget is the uplink transmission bandwidth allocated to the corresponding stream data variable. It is used to achieve adaptive allocation of bandwidth resources, allocating more transmission bandwidth to variables with high contribution value and less transmission bandwidth to variables with redundant interference, thereby reducing the transmission of invalid data.

[0102] The formula for calculating the communication bandwidth budget is: in, For the k-th processing time scale, the communication bandwidth budget for the streaming data variable corresponding to sensor unit j is given, in bytes per second; the summation operation iterates through all streaming data variables m. The minimum bandwidth budget for each streaming data variable is set to 0.5% of the available uplink bandwidth to avoid situations where bandwidth allocation is zero and to ensure the normal transmission of basic data.

[0103] S604: Divide the total number of buffered bytes of the data to be transmitted by the communication bandwidth budget and round down to calculate the downsampling interval. Based on this, perform a segmented smoothing mean operation on the original sequence to filter out high-frequency redundancy and output the feature stream data to be uploaded that is adapted to the limited bandwidth.

[0104] The total number of buffered bytes of data to be transmitted refers to the total number of bytes of data to be transmitted for the corresponding stream data variable within a preset transmission time window, denoted as . The unit is bytes, and the preset transmission time window is consistent with the system processing cycle. The downsampling interval refers to the number of interval points between two adjacent sampling points when downsampling the original sampling sequence. It is used to reduce the amount of data transmitted and adapt to the allocated bandwidth budget.

[0105] The formula for calculating the downsampling interval is: in, For the k-th processing time scale, the downsampling interval of the streaming data variable corresponding to the sensor unit identified as j is a positive integer; This is for rounding up; The system's time step size is in seconds. To prevent zero bias, a preset offset is used to avoid division by zero errors. The maximum downsampling interval is set to 100, and the minimum is set to 1 to prevent over-downsampling from causing data feature loss.

[0106] Segmented smoothing and averaging involves dividing the original sampled sequence into segments according to the calculated downsampling intervals, and taking the arithmetic mean of all sampled points within each segment as the output value for that segment. This effectively filters out high-frequency redundant noise while preserving the core trend characteristics of the data. The calculation formula for the feature stream data to be uploaded is: in, For the k-th processing time scale, the feature stream data to be uploaded in the q-th segment corresponding to the sensing unit identified as j, with the unit being consistent with the physical unit of the corresponding original data; the summation operation traverses all sampling points within the q-th segment.

[0107] The feature stream data to be uploaded is packaged using a fixed frame structure, with the packaging cycle matching the system processing cycle. Each data packet contains a frame header, timestamp, data segment, and checksum, and is uploaded to the cloud platform via HTTP / HTTPS protocol. When the uplink communication link is interrupted, the feature stream data to be uploaded is cached in the local storage of the edge gateway, with a maximum cache retention period of 7 days. After the link is restored, the data is retransmitted sequentially in chronological order to avoid data loss.

[0108] Step 7: Extract the model evolution step size based on the improvement of the optimal fitness of all particles within adjacent time scales to update the device-level thermal parameter library in the edge gateway. Then, use the time-series derivatives of insulation thermal aging factor, damp heat activation risk index, and hotspot extreme temperature to calculate and generate an asset health ranking score for multi-device scheduling decisions. The specific steps are as follows: S701: Calculate the relative difference between the optimal fitness of all particles at the current processing timescale and the previous timescale, and divide it by the smoothing factor consisting of the sum of their absolute values ​​and the minimum anti-zero bias, to generate the model evolution step size for controlling the update intensity of the parameters.

[0109] The optimal fitness of all particles refers to the fitness value corresponding to the optimal position of the particle swarm after iterative convergence, denoted as . , where is the optimal fitness of all particles at the k-th processing timescale. This represents the optimal fitness of all particles at the previous timescale. The absolute value of the relative difference characterizes the improvement in model fit; a larger improvement indicates a greater need to update the model parameters.

[0110] The model evolution step size is used to control the update magnitude of the device-level thermal parameter library, avoiding large jumps in parameters and improving model stability. The calculation formula is as follows: in, is the model evolution step size at the k-th processing time scale, and is a dimensionless parameter with a value range of [0,1]. To prevent errors caused by dividing by zero, a pre-set zero bias is used. When the model fit does not improve, the model evolution step size approaches 0, and the parameters are hardly updated; when the model fit improves significantly, the model evolution step size approaches 1, and the parameters are updated significantly.

[0111] S702: Read the initial device-level thermal parameter library stored in the edge cache at the previous time scale, multiply it by the feature deviation of the current hidden hot spot by the model evolution step size, perform negative feedback accumulation correction, and overwrite and save it as the device-level thermal parameter library at the current time scale.

[0112] The device-level thermal parameter library is a collection of all thermal characteristic parameters corresponding to the monitored transformer, stored in the edge gateway cache, denoted as... The parameters, including the coefficients of the heat capacity mapping matrix, surface observation mapping coefficients, harmonic loss coefficients, heat transfer coefficients, and local oil velocity calibration values, are the core fundamental parameters of the temperature field inversion model. The device-level thermal parameter library is stored in the non-volatile storage medium of the edge gateway, ensuring data integrity even after power failure. During system cold starts, the latest parameter library data is directly loaded. The update cycle of the parameter library is consistent with the system processing cycle; each update overwrites and saves historical versions, with a maximum retention limit of 100 historical versions for parameter backtracking and anomaly analysis.

[0113] The initial equipment-level thermal parameter library is the parameter library saved after the previous timescale processing is completed, denoted as . The feature bias of the current hidden hotspot field is the difference between the parameter set corresponding to the optimal position of the population obtained by the current iteration convergence and the initial device-level thermal parameter library, denoted as . Negative feedback cumulative correction is used to achieve adaptive updates of the parameter library, ensuring that the model parameters continuously adapt to changes in the transformer's operating state and aging characteristics. The update formula is: in, This is the updated device-level thermal parameter library for the current time scale. After the update is completed, it is overwritten and saved to the non-volatile storage of the edge gateway for data processing and model calculation in the next time scale.

[0114] S703: Perform first-order difference derivation on the time series of hotspot extreme temperatures and normalize it in combination with global bias to extract the time series derivative of hotspot extreme temperatures after filtering out steady-state fluctuations.

[0115] The time series of hotspot extreme temperatures is a sequence of hotspot extreme temperatures processed at continuously time scales. The first-order difference derivation involves performing a first-order difference operation on the time series to obtain the rate of change of the hotspot extreme temperatures over time, i.e., the time derivative, denoted as . The time derivative, measured in degrees Celsius per second, characterizes the rate of increase or decrease of hotspot temperature, reflecting the transient trend of temperature change. The formula for calculating the time derivative is: in, The time interval between two adjacent processing time markers is expressed in seconds.

[0116] The global bias is the median absolute deviation of the time series derivative within a preset historical time window, denoted as . This is used to normalize the time derivative, eliminating the influence of overall fluctuations in the rate of temperature change under different operating conditions and highlighting abnormal transient temperature changes. The formula for calculating the normalized time derivative is: in, To prevent zero bias by setting a preset value, avoid calculation errors when dividing by zero.

[0117] S704: The natural logarithmic enhancement term of the insulation thermal aging factor, the scale-normalized damp heat activation risk index, and the normalized time-series derivative are weighted and accumulated to integrate steady-state aging trends and transient deterioration risks, and to calculate and generate an asset health ranking score that guides the order of operation and maintenance intervention.

[0118] The natural logarithmic enhancement term is used to nonlinearly enhance the insulation thermal aging factor, amplifying the differences in the aging factor. The calculation formula is as follows: Where ln is the natural logarithm. Scale normalization refers to dividing the damp-heat activation risk index by its own sequence's median absolute deviation within a preset historical time window to eliminate the influence of dimensions. The calculation formula is: where The median absolute bias of the humid heat activation risk index series is given. This is a preset zero-bias setting.

[0119] The asset health ranking score is used to comprehensively quantify the insulation health status of transformers. A higher score indicates a worse health status, a higher risk of insulation failure, and a higher priority for maintenance intervention. The formula for calculating the asset health ranking score is: in, , where is the asset health ranking score under the kth processing time point, and is a dimensionless parameter; , , These are the weighting coefficients for the three sub-items, all of which are dimensionless positive numbers with values ​​of 0.4, 0.35, and 0.25 respectively. The total weight is 1, and the weight allocation can be adjusted according to the voltage level and importance of the transformer.

[0120] For multi-device cluster scenarios, asset health ranking scores need to be normalized across devices. The normalization method is to divide the score of a single device by the maximum score of the cluster of devices with the same voltage level and capacity, resulting in a normalized ranking score between 0 and 1. This normalized score is used for horizontal comparison and maintenance priority ranking between different devices. Maintenance priorities are divided from high to low according to the normalized ranking score. Devices with a score greater than 0.8 are classified as Level 1 warning devices, requiring immediate on-site inspection; devices with a score between 0.5 and 0.8 are classified as Level 2 warning devices, requiring special testing within 7 days; devices with a score less than 0.5 are classified as normal operating devices, requiring maintenance according to the regular cycle.

[0121] Thermodynamic constant 15000 is the activation energy for thermal aging of ordinary cellulose insulating paper. With the ideal gas constant The ratio, i.e. The engineering significance is the sensitivity coefficient of the thermal aging rate of insulating paper to temperature; the dimension is Kelvin (K); the value is based on the standard value of the activation energy of thermal aging of cellulose insulating paper specified in IEC60076-7 Power Transformer Load Guidelines, corresponding to... , Calculated It is applicable to ordinary Kraft cellulose insulating paper, but not to high-temperature resistant Nomex insulating paper.

[0122] Activation energy of water migration In an oil-paper insulation system, this is the Gibbs free energy change required for bound water to escape from the cellulose molecular chain; its engineering significance is to characterize the ease with which water escapes from the oil paper; its dimension is joules per mole (J / mol); the value is 40000 J / mol (40 kJ / mol), with a positive sign, conforming to the thermodynamic law that water needs to absorb energy to escape from its binding; the value is based on the industry-standard empirical value in the IEC60422 standard for determining the moisture content of electrical insulating oils and oil-paper insulation systems.

[0123] Adjacent segment smoothness constraint term This is the regularization penalty coefficient for the second spatial derivative in the heat conduction equation; its engineering significance is to suppress non-physical spatial jumps that occur during the temperature field inversion process and ensure the spatial continuity of the temperature field; it is dimensionless; its value ranges from 0.1 to 10, and the value is determined based on the spatial continuity requirements of axial heat conduction in transformer windings. The longer the axial length of the winding, the larger the value.

[0124] Preset zero offset To avoid small quantities and singularities in numerical calculations; in engineering terms, to avoid numerical calculation errors such as division by zero or zero logarithmic independent variable; to set dimensions and values ​​according to the differences in calculation stages: for time-scale alignment stages... The power calculation stage takes Temperature calculation process The weight calculation process takes Dimensionless, adaptable to the numerical accuracy requirements of different stages.

[0125] The CRC16-Modbus check generator polynomial is a cyclic redundancy check generator polynomial used for error checking in data frame transmission; its engineering significance is to check for bit errors in data frames during transmission and storage; the generator polynomial is... The value corresponds to hexadecimal 0x8005, with an initial value of 0xFFFF and a result XORed with 0x0000. This is a standard verification rule for the Modbus-RTU protocol commonly used in the industrial control field.

[0126] The transfer function of the second-order generalized integrator phase-locked loop is the closed-loop transfer function of the synchronous phase-locked loop; its engineering significance is that it extracts the fundamental frequency and phase from the three-phase voltage without steady-state error, suppressing harmonic interference; the transfer function is... ,in The cutoff angular frequency, The rated fundamental angular frequency is suitable for power grid frequency fluctuations ranging from 45 to 55 Hz.

[0127] Particle swarm cognitive learning factor Social learning factors This is the step size control coefficient for the individual cognition term and the social cooperation term in the particle swarm optimization algorithm; its engineering significance is to balance the individual exploration ability and the group convergence ability of the algorithm; the value is 1.49445, which is the standard value of the convergence factor method commonly used in the industry, which can guarantee the convergence and global search ability of the algorithm.

[0128] The factory initialization process for the equipment-level thermal parameter library is as follows: Step 1: Extract the rated parameters from the transformer's factory test report, including the DC resistance of the winding at 75℃, rated capacity, voltage level, cooling method, winding axial height, and number of oil passages; Step 2: Based on the transformer's rated parameters, calculate the core parameters such as the nominal heat-to-temperature conversion coefficient, mapping coefficient, and heat transfer coefficient using the standard thermal circuit model in IEC60076-7, and use them as the initial nominal values ​​for the parameter library; Step 3: Write the initial nominal value to the non-volatile storage partition of the edge gateway, set it as a read-only factory backup area, and copy it to the read-write running area for subsequent adaptive updates; Step 4: Before putting the transformer into no-load operation, perform cold calibration, collect ambient temperature and winding DC resistance, perform temperature correction on the initial parameter library, and complete the initialization.

[0129] The fallback procedure for non-convergence in the particle swarm optimization algorithm is as follows: Step 1: When the algorithm reaches the maximum number of iterations (100) and the optimal fitness of the population has not decreased for 20 consecutive iterations, it is determined to be non-convergent. Step 2: Read the optimal position of the swarm converged at the previous processing time point stored in the edge gateway, use it as the initial particle swarm center for this iteration, regenerate the initial particles, and reduce the distribution range of the initial particles to 50% of the original range; Step 3: Re-execute the iteration, adjusting the maximum number of iterations to 50, and fixing the inertia weight at 0.7 to accelerate convergence; Step 4: If the second iteration still fails to converge, the optimal position of the population at the previous processing time point is directly used as the result for this iteration. At the same time, an error log is recorded, and an error parameter alarm is triggered.

[0130] The hardware synchronization calibration process for multiple sensing units is as follows: Step 1: The edge gateway sends a synchronization sampling trigger signal to all multi-source sensing units via the PPS second pulse synchronization signal, and the time synchronization error of the trigger signal does not exceed 1μs; Step 2: After all sensing units receive the trigger signal, they simultaneously perform sampling, package the sampled value and the timestamp of the sampling time, and upload them to the edge gateway; Step 3: The edge gateway clocks the sampling timestamps of all sensing units and uses the NTP network time protocol to synchronize the clocks of the gateway and the sensing units, with a synchronization error of no more than 1ms. Step 4: Perform a full-link synchronous calibration every 24 hours to correct clock drift in the sensing unit and ensure the timing consistency of the sampled data.

[0131] The underlying implementation process of edge gateway computing power scheduling is as follows: Step 1: The edge gateway uses the cgroups CPU resource isolation mechanism of the Linux system to allocate a corresponding share of computing power to each computing task; Step 2: Based on the calculated computing power share, set the CPU time slice weight for each task. The higher the computing power share, the greater the time slice weight, and the more CPU execution time is obtained. Step 3: Set the minimum CPU time slice weight to 1% to ensure the lowest execution priority for all basic computing tasks; Step 4: Update the computing power allocation once per processing cycle to dynamically adjust the CPU resource usage of each task and avoid computing power congestion.

[0132] The buffering and retransmission process for uplink communication link interruption is as follows: Step 1: When the edge gateway detects an uplink interruption, it writes the feature stream data to be uploaded into a local storage circular cache queue. The cache queue uses Flash storage and has a maximum capacity of 7 days of full data. Step 2: Check the link status every 10 seconds. When the link is restored, retransmit the cached data packets in the cache queue in order of timestamp from earliest to latest. Step 3: During the retransmission process, the newly generated feature stream data is written to the buffer first, and retransmission is performed in parallel to avoid data loss; Step 4: When the cache queue is full, overwrite the oldest historical data and record the cache overflow log.

[0133] The process for handling extreme operating conditions and boundary conditions is as follows: Sensor Failure Condition: When the edge gateway detects that no valid data is uploaded from any of the sensor units for more than 3 processing cycles, it terminates all temperature field inversion and insulation assessment calculations, keeps the device-level thermal parameter library from being updated, and triggers a hardware fault alarm. After the sensor units recover, it first performs 3 cycles of calibration calculations, and then resumes normal full-process operation.

[0134] Severe harmonic distortion conditions: When the total harmonic distortion rate of the grid voltage exceeds 10%, the highest harmonic analysis order is increased to twice the original order, and the window length of the Fourier transform is adjusted to two fundamental cycles to improve the accuracy of harmonic analysis.

[0135] Power outage recovery mode: When the edge gateway is powered on again, it first loads the latest device-level thermal parameter library from non-volatile storage, reads the last 10 sets of historical data before the power outage to perform initial value calibration, and then resumes normal cycle processing, while recording the power outage event log.

[0136] The embodiments of this example have been described above. However, this example is not limited to the specific implementation methods described above. The specific implementation methods described above are merely illustrative and not restrictive. Those skilled in the art can make many other forms based on the guidance of this example, and all of them are within the protection scope of this example.

Claims

1. A real-time data processing and intelligent analysis system for an IoT platform based on edge computing, applied to a transformer monitoring system including edge gateways and multi-source sensing units, characterized in that... Configured for execution: The multimodal raw acquisition data of the multi-source sensing unit is acquired, the arrival time interval of the raw acquisition data is extracted and the median absolute deviation is calculated, and an adaptive time smoothing scale is generated to perform time-scale alignment on the raw acquisition data to obtain a unified time-scaled stream data frame that excludes asynchronous sampling. The electrical parameter waveforms in the unified time-stamped stream data frame are subjected to discrete Fourier transform to separate the frequency domain components. The fundamental copper loss power of the winding, the additional heating power of harmonics and the heating power of the core loss are extracted and superimposed to generate the total heat source power that characterizes the overall heating degree. Based on the total heat source power and the pre-acquired transformer component metadata, the heat transfer hysteresis effect is introduced by recursive calculation to calculate the oil passage heat retention accumulation particle state quantity characterizing the oil passage heat transfer characteristics. The state variables of the accumulated heat in the oil passage are input into a preset particle swarm algorithm model for temperature field mapping. The individual fitness and the optimal fitness of all particles are calculated using the actual observation residuals. Based on this, the particle positions are iteratively updated, and a hidden hot spot field representing the true internal temperature is generated. The extreme temperatures and axial positions of the hot spots are extracted from the hidden hot spot field. The thermal field gradient is evaluated by combining the pre-extracted water activity of the oil paper, and an insulation thermal aging factor and a damp heat activation risk index characterizing the insulation state are generated. Calculate the disturbance sensitivity of the hidden hotspot field to each original collected data, allocate computing power share and communication bandwidth budget to each data based on the disturbance sensitivity to eliminate redundant interference, and aggregate to generate feature stream data to be uploaded; The model evolution step size is extracted based on the improvement of the optimal fitness of all particles within adjacent time scales to update the device-level thermal parameter library in the edge gateway. The asset health ranking score for multi-device scheduling decision is generated by fusion calculation using the time-series derivatives of the insulation thermal aging factor, the damp heat activation risk index, and the hot spot extreme temperature.

2. The method for real-time data processing and intelligent analysis of an IoT platform based on edge computing according to claim 1, characterized in that, The multimodal raw acquisition data of the multi-source sensing unit is acquired, the arrival time interval of the raw acquisition data is extracted and the median absolute deviation is calculated, and an adaptive time smoothing scale is generated to perform time-stamp alignment on the raw acquisition data to obtain a unified time-stamped stream data frame that excludes asynchronous sampling, including: The arrival time interval of adjacent samples of the same multi-source sensing unit is obtained, the median of their absolute deviation is calculated and a preset anti-zero bias is added to eliminate the computational interference of sudden communication delays and generate the adaptive time smoothing scale specific to the sensing unit. Obtain the current processing time stamp of the system and the corresponding original acquisition time, and calculate the absolute time difference between the two to quantify the degree of data lag; Divide the absolute time difference by the adaptive time smoothing scale and take the opposite number, perform a natural exponential decay operation to weaken the weight of old data, and generate a time decay weight that characterizes the validity of the data at the current moment. The raw data of each time node of the same sensing unit are weighted and accumulated with the corresponding time decay weight, and then normalized by dividing by the sum of the time decay weights to obtain normalized streaming data that eliminates timing misalignment. The normalized stream data from various multi-source sensing units are spliced ​​together in time sequence to generate a unified time-scaled stream data frame containing variables of three-phase voltage, three-phase current, oil temperature, and ambient temperature.

3. The method for real-time data processing and intelligent analysis of an IoT platform based on edge computing according to claim 1, characterized in that, Perform a Discrete Fourier Transform on the electrical parameter waveforms in the unified time-stamped data frame to separate the frequency domain components, extract the fundamental copper loss power of the winding, the additional heating power of harmonics, and the heating power of the core loss, and superimpose them to generate the total heat source power characterizing the overall heating degree, including: The fundamental frequency is extracted using the synchronous phase-locked loop built into the edge gateway. The current sampling frequency is divided by twice the fundamental frequency and rounded down to filter out high-frequency noise, thus determining the highest harmonic analysis order that meets the Nyquist limit. The dynamic resistance of the phase winding after temperature correction at the previous time scale is obtained, multiplied by the square of the fundamental current amplitude, and accumulated on the three phases to separate the purely resistive heating characteristics, thereby obtaining the fundamental copper loss power of the winding. Extract the preset equipment harmonic eddy current heat dissipation coefficient and core loss coefficient, and multiply them by the square of the corresponding order harmonic current amplitude and the square of the harmonic voltage amplitude, respectively. The additional temperature rise caused by the nonlinear load is corrected by accumulating the frequency dimension within the range from the second to the highest harmonic analysis order, thereby generating the additional heating power of the harmonics and the heating power of the core loss. The fundamental copper loss power of the winding, the additional heating power of the harmonics, and the heating power of the core loss are linearly summed to combine multiple heat generation mechanisms and generate the total heat source power.

4. The method for real-time data processing and intelligent analysis of an IoT platform based on edge computing according to claim 1, characterized in that, Based on the total heat source power and pre-acquired transformer component metadata, a heat transfer hysteresis effect is introduced using recursive calculations to calculate the oil passage heat retention accumulation particle state quantities characterizing the oil passage heat transfer characteristics, including: The axial height characteristics and equivalent physical length of each oil circuit unit are parsed from the metadata of the transformer components, and the local oil velocity estimation value corresponding to each algorithm particle is matched using a pre-built local oil velocity state dictionary. Divide the equivalent physical length by the sum of the absolute value of the local oil velocity estimate and the zero-prevention offset to avoid static singularities, and calculate the oil passage residence time that characterizes the cooling medium flow delay. The natural exponential decay calculation is performed by dividing the current time step by the oil passage residence time to simulate the physical decay process of heat loss over time and generate a heat retention factor characterizing the thermal inertia intensity. The total heat source power is spatially non-uniformly distributed using the prior weight of end leakage magnetic flux to restore the leakage magnetic flux heating characteristics, and the heat exchange loss calculated based on the temperature difference between oil and the environment is superimposed to generate the transient new thermal excitation for the current time scale. The heat retention factor is used to proportionally attenuate the heat retained by the upstream oil circuit unit at the previous time scale, and the transient new thermal excitation is added to fuse the spatiotemporal heat transfer process to generate the current particle's accumulated state quantity of oil circuit heat retention.

5. The method for real-time data processing and intelligent analysis of an IoT platform based on edge computing according to claim 1, characterized in that, The state variables of the accumulated heat in the oil duct are input into a preset particle swarm optimization model for temperature field mapping. The individual fitness and the optimal fitness of all particles are calculated using actual observation residuals. Based on this, the particle positions are iteratively updated, and a hidden hotspot field representing the true internal temperature is generated, including: Based on the preset thermal capacity mapping matrix, the state variables of the accumulated heat particles in the oil passage are combined with the current ambient temperature, and a hidden segmented temperature field characterizing the internal heating mode of the winding is generated through affine transformation operation. The hidden segmented temperature field is extrapolated to the outer surface using the equipment surface observation mapping coefficient, and the observed oil temperature fitting value and the box temperature fitting value containing the influence of fluid convection are synthesized. Calculate the deviation between the fitted value and the actual observed value in the unified time-stamped stream data frame, and construct a dynamic confidence weight by combining the median absolute deviation of the residuals to suppress abnormal noise; The dynamic confidence weights are multiplied by the squared residuals, and a spatial temperature abrupt change penalty value calculated using adjacent segment smoothness constraints is added to correct for non-physical jumps, generating the individual fitness that characterizes the model fit. Using the particle swarm iteration law, the convergence gradient is extracted based on the individual's historical best position and the optimal fitness of all particles. The optimization speed and spatial coordinates of each particle are adaptively updated, and the optimal coordinates of all particles that have finally converged are output as the hidden hotspot field.

6. The method for real-time data processing and intelligent analysis of an IoT platform based on edge computing according to claim 1, characterized in that, The extreme temperatures and axial positions of the hotspots are extracted from the hidden hotspot field. Combined with the pre-extracted moisture activity of the oil paper to assess the thermal field gradient, an insulation thermal aging factor and a damp-heat activation risk index characterizing the insulation state are generated, including: The temperatures of each spatial node in the hidden hotspot field are numerically compared, and the highest temperature extreme value is selected as the hotspot extreme temperature. The corresponding spatial height feature sequence is extracted to confirm the axial position of the hotspot. The insulation thermal aging factor, which quantifies insulation degradation, is generated by subtracting the reciprocal of the sum of the hot spot extreme temperature and the thermodynamic conversion constant from the reciprocal of the thermodynamic reference constant and performing a natural exponential scaling operation through the Arrhenius equation. The absolute value of the axial temperature gradient at the axial position of the hot spot extreme temperature is calculated using the central difference method, and then normalized by dividing by the global gradient deviation to highlight the degree of local heat concentration, thus obtaining the local hot spot concentration. The preset water migration activation energy parameter is extracted and combined with the pre-extracted water activity of the oil paper. A thermal kinetic energy release term characterizing the ability to remove water at high temperatures is constructed through exponential calculation. The thermal kinetic energy release term is cross-producted with the local hot spot concentration to couple the physical interaction between temperature gradient and moisture movement, thereby generating the humid heat activation risk index for early warning of bubble formation risk.

7. The method for real-time data processing and intelligent analysis of an IoT platform based on edge computing according to claim 1, characterized in that, Calculate the disturbance sensitivity of the hidden hotspot field to each original collected data, allocate computing power share and communication bandwidth budget to each data based on the disturbance sensitivity to eliminate redundant interference, and aggregate to generate feature stream data to be uploaded, including: The partial derivatives of the extreme temperatures of the hotspots with respect to each flow data variable are calculated using the numerical perturbation method. These derivatives are then divided by the historical statistical fluctuation range of the corresponding variables to eliminate dimensional differences, thereby generating the perturbation sensitivity that measures the contribution value of single-point data. The preset baseline CPU time of the edge computing algorithm module is obtained, the perturbation sensitivity of various types of data is divided by the baseline CPU time and the square root is taken, and the computing power share representing the computing power tilt weight is allocated through global normalization operation. The available uplink communication bandwidth is obtained by monitoring the current network interface status of the edge gateway. The available uplink communication bandwidth is allocated proportionally according to the disturbance sensitivity ratio of each data point to obtain the communication bandwidth budget that characterizes the transmission priority limitation. The total number of buffered bytes of the data to be transmitted is divided by the communication bandwidth budget and rounded to calculate the downsampling interval. Based on this, a segmented smoothing mean operation is performed on the original sequence to filter out high-frequency redundancy, and the feature stream data to be uploaded is output to adapt to the limited bandwidth.

8. The method for real-time data processing and intelligent analysis of an IoT platform based on edge computing according to claim 1, characterized in that, The model evolution step size is extracted based on the improvement magnitude of the optimal fitness of all particles within adjacent time scales to update the device-level thermal parameter library in the edge gateway. Then, using the time-series derivatives of the insulation thermal aging factor, the damp heat activation risk index, and the hotspot extreme temperature, an asset health ranking score for multi-device scheduling decisions is generated, including: Calculate the relative difference between the optimal fitness of all particles at the current processing timescale and the previous timescale, and divide it by the smoothing factor consisting of the sum of their absolute values ​​and the minimum anti-zero bias, to generate the model evolution step size for controlling the update intensity of the parameters. Read the initial device-level thermal parameter library stored in the edge cache at the previous time point, multiply it by the feature deviation of the current hidden hot spot field by the model evolution step size, perform negative feedback accumulation correction, and overwrite the device-level thermal parameter library stored at the current time point. The time series of the hotspot extreme temperatures is subjected to first-order difference derivation and normalized by combining global bias to extract the time-series derivative of the hotspot extreme temperatures after filtering out steady-state fluctuations. The natural logarithmic enhancement term of the insulation thermal aging factor, the scale-normalized damp heat activation risk index, and the time-series derivative are weighted and accumulated to integrate steady-state aging trends and transient deterioration risks, and the asset health ranking score is calculated to guide the order of operation and maintenance interventions.

Images