Remote real-time monitoring and fault early warning method for power supply operation state

Abstract

The invention discloses a remote real-time monitoring and fault early warning method for a power supply operation state, and relates to the technical field of power electronics and energy management, and the method comprises the steps: collecting the temperature data of a key part of a power supply through an optical fiber Bragg grating sensor array, and analyzing the wavelength offset; electrical parameters are collected through a differential MEMS current / voltage sensor, and a permalloy / ferrite composite shielding layer is arranged in the sensor; collecting a mechanical vibration signal through a piezoelectric ceramic vibration sensor; for device data from different manufacturers, Modbus and CAN heterogeneous protocols are matched and analyzed through a regular expression, and the Modbus and CAN heterogeneous protocols are converted into a unified RDF triple format; inputting temperature, current and vibration data into a space-time alignment module, and constructing a semantic mapping relation by adopting an ontology; calculating a sensor confidence coefficient weight, and performing weighted fusion to generate an equipment health index; operating a lightweight LSTM model at an edge node, and detecting current harmonic and vibration resonance frequency band abnormity in real time; cOMSOL multi-physical field simulation is combined at the cloud end, and the evaporation rate of the capacitor electrolyte is predicted.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of power electronics and energy management, in particular to a remote real-time monitoring and fault early warning method for power supply operating state. BACKGROUND

[0002] Power supply failure in the fields of industry, medical treatment, data center, etc. can cause significant economic loss. Large industrial facilities in the United States lose more than 100 billion US dollars per year due to power quality problems, and the relevant loss in the European Union reaches 156 billion euros. The stable operation of core equipment such as extra-high voltage transformers directly affects the safety of the power grid, and traditional manual inspection and single-parameter threshold alarm cannot meet the demand.

[0003] The popularity of new energy power generation (such as solar energy and wind energy) increases the complexity of the power grid, and it is necessary to monitor the power fluctuation and grid stability of distributed power supply in real time. For example, a solar inverter optimizes power generation efficiency through an MPPT algorithm, which relies on a high-precision current sensor to provide data support.

[0004] The Internet of Things technology reduces the deployment cost of remote monitoring, and the low delay characteristics of the 5G network support real-time data transmission. For example, a UPS power supply realizes unattended operation through a cloud platform, and an administrator can obtain the on-off state and fault alarm in real time through a mobile phone APP.

[0005] However, the existing technology still has problems to be solved: sensor and data acquisition bottleneck, including insufficient reliability, traditional sensors are easily affected by electromagnetic interference, leading to data jump or misjudgment; difficulty in multi-source data fusion, the data format and communication protocol of devices from different manufacturers are not unified, and flexible adaptive interfaces need to be developed. SUMMARY

[0006] To solve the above technical problems, a remote real-time monitoring and fault early warning method for power supply operating state is provided, which solves the problems of insufficient implicit feature extraction capability and difficulty in multi-source data fusion.

[0007] To achieve the above purpose, the technical scheme adopted by the present application is:

[0008] The remote real-time monitoring and fault early warning method for power supply operating state comprises:

[0009] S1, collecting temperature data of key parts of the power supply through a fiber Bragg grating sensor array, and using wavelength division multiplexing technology to analyze the wavelength shift;

[0010] S2, collecting electrical parameters through a differential MEMS current / voltage sensor, and the sensor has a built-in permalloy / ferriet composite shielding layer;

[0011] S3, collecting mechanical vibration signals through a piezoelectric ceramic vibration sensor;

[0012] S4, parse Modbus, CAN heterogeneous protocols through regular expression matching for equipment data from different manufacturers, and convert to unified RDF triple format; timestamp multiple source data based on IEEE1588v2 protocol;

[0013] S5, input temperature, current, vibration data into space-time alignment module, construct semantic mapping relationship using ontology; calculate sensor confidence weight, and generate device health index through weighted fusion;

[0014] S6, run lightweight LSTM model on edge node to detect current harmonic and vibration resonance frequency band anomalies in real time; combine COMSOL multi-physical field simulation on cloud to predict evaporation rate of capacitor electrolyte, and dynamically adjust warning threshold.

[0015] Preferably, S1 specifically comprises:

[0016] Sensor array deployment design, FBG sensor nodes are arranged in ring array topology at three core areas of IGBT module heat dissipation substrate surface, DC bus capacitor shell, and transformer winding lead-out line inside power supply system;

[0017] Each ring array contains multiple FBG sensing units, which are fixed to the measured surface by high-temperature resistant ceramic glue with a predetermined unit spacing;

[0018] Anti-electromagnetic interference physical structure is set, the sensing optical fiber is protected by polyimide coating and stainless steel armor; the lead path avoids strong electromagnetic areas and is provided with a metal braided shielding conduit

[0019] Wavelength division multiplexing signal analysis, multi-channel spectral encoding, each FBG sensor is preset with an independent Bragg wavelength and a predetermined wavelength interval; broadband light source is incident on the optical fiber, and reflected light is split to multiple channels by arrayed waveguide grating;

[0020] Wavelength shift, high-resolution spectrometer is used to monitor the peak wavelength of each channel; Gaussian fitting algorithm is used.

[0021] Preferably, S1 specifically comprises:

[0022] Temperature field reconstruction, three-dimensional thermal field modeling, based on ring array data, IGBT surface temperature distribution is reconstructed by inverse distance weighted interpolation method, and temperature cloud picture is output:

[0023]

[0024] In the formula, T(x, y) is the estimated temperature at point (x, y), which is calculated by inverse distance weighted interpolation method according to the temperature values of surrounding sensors; T iThe temperature value measured for the i-th sensor represents a temperature data point at a known position in space;d i The Euclidean distance of the interpolation point (x, y) to the i-th sensor is used to measure the spatial distance between the interpolation point and the known temperature measurement point; p is the attenuation coefficient, which is 2, used to control the degree of influence of distance on weight; n is the number of sensors, representing the total number of all sensor data points used to calculate the interpolated temperature;

[0025] Fiber breakage monitoring, real-time detection of optical power in each channel, and triggering of alarm when attenuation is greater than a predetermined threshold; temperature drift compensation, embedding of reference FBG in the sensor array to correct the influence of environmental temperature; dynamic calibration, automatic comparison of MEMS temperature probe every 24 hours, and re-calibration when the error is within a predetermined temperature.

[0026] Preferably, S2 specifically comprises:

[0027] Composite electromagnetic shielding layer structure, permalloy layer forms low-frequency magnetic field, absorbs passageway, and is filled with nanocrystalline strip material at the slotted part;

[0028] Ferrite layer, inhibits high-frequency interference, laminated on the outer layer of permalloy, combined with hot isostatic pressing process;

[0029] The current detection of the differential MEMS sensing unit adopts the Lorentz force principle, when the current-carrying wire passes through the chip, the vertical magnetic field makes the comb electrode deviate; differential capacitance detection, the movable electrode and the fixed electrode form a Wheatstone bridge, and the capacitance change:

[0030]

[0031] In the formula, ΔC is the capacitance change, which is the change of capacitance value caused by the deviation of the comb electrode due to the action of the magnetic field in the differential capacitance detection, reflecting the size of the current; ε0 is the vacuum permittivity, indicating the permittivity in vacuum, used to calculate the size of the capacitance in the electric field; ε r is the relative permittivity, which is the ratio of the dielectric constant of the material to the vacuum permittivity, used to describe the polarization characteristics of the material, different materials have different relative permittivities; A is the electrode area, which is the opposite area of the movable electrode and the fixed electrode, with the unit of square meter, the larger the electrode area, the larger the capacitance value; d0 is the initial spacing, which is the distance between the movable electrode and the fixed electrode without the action of the magnetic field, with the unit of meter, it affects the initial value of the capacitance; δ is the displacement caused by the magnetic field, which is proportional to the current, when the current passes through the current-carrying wire and is in the vertical magnetic field, the Lorentz force will make the movable electrode deviate, the larger the displacement, the larger the capacitance change;

[0032] Voltage detection based on electrostatic induction, high-voltage electrode induces charge on the MEMS cantilever beam, which is converted into voltage by a charge amplifier and input impedance.

[0033] Preferably, S2 specifically includes:

[0034] Common-mode noise suppression, using a third-order differential amplifier, setting the common-mode rejection ratio; temperature drift compensation, integrated platinum resistance temperature sensor, real-time correction of gain error; active shielding drive, collecting external common-mode voltage feedback to the shielding shell, controlling the sensor floating ground potential difference;

[0035] Self-calibration process, zero drift compensation, timing cut-off measured current, measuring zero output, storing drift curve, predicting compensation value through polynomial fitting; range calibration, built-in reference source injects 1mA / 1V standard signal, dynamically adjusts ADC gain coefficient.

[0036] Preferably, S3 specifically includes:

[0037] Piezoelectric ceramic vibration sensor uses lead zirconate titanate piezoelectric ceramic sheet, sandwich structure, silver electrode on the upper and lower surfaces of the ceramic sheet, bonded to the titanium alloy base through conductive epoxy resin, the base resonance frequency is greater than the predetermined frequency;

[0038] Targeted installation strategy, monitoring point positioning, including power fan bearing seat, transformer fastening bolt, IGBT heat sink substrate; installation method, rigid connection by bolt, contact surface coated with thermal silicone grease; the sensor axis is aligned with the main direction of vibration;

[0039] Wideband signal capture, set frequency response range, covering power mechanical fault characteristic frequency band; anti-interference design, electromagnetic shielding, shell made of μ-Metal alloy; double shielding of signal line using copper braid and aluminum foil, the end is connected to the acquisition circuit through a feedthrough filter; temperature compensation, built-in NTC thermistor, real-time correction of piezoelectric coefficient temperature drift.

[0040] Preferably, S4 specifically includes:

[0041] Regular expression matching mechanism, protocol feature library construction, pre-set Modbus / TCP message template, pre-set CAN2.0B extended frame template;

[0042] Private protocol compatibility strategy, dynamic learning through regular capture group, supporting fuzzy matching, when the protocol structure is unknown, identifying the payload according to byte entropy value;

[0043] Finite state machine analysis process receives raw messages, detects frame headers, determines protocol types through regular matching, extracts payloads, retransmits abnormally after checksum verification;

[0044] Unified RDF triple conversion, semantic mapping rules; using context enhancement technology, ontological reasoning, building power domain ontology library, automatically inferring semantic relationships;

[0045] Space-time reference binding, position information adopts WGS-84 coordinates, and the time reference is synchronized through IEEE1588v2.

[0046] Preferably, the S5 specifically comprises:

[0047] Multi-source data space-time reference unification, time dimension alignment, synchronization of each sensor clock based on IEEE1588v2, cubic spline interpolation for asynchronous data, and resampling to a given reference frequency;

[0048] Space coordinate binding, establishment of a three-dimensional coordinate system of the device, mapping of the FBG temperature point and the vibration sensor to a unified space grid;

[0049] Cross-modal data correlation rules are that the temperature is related to the IGBT junction temperature, the current harmonic is related to the change of the capacitor ESR, and the vibration energy is related to the fan bearing wear;

[0050] Ontology-driven semantic mapping, construction of a power domain ontology library, and core ontology elements include classes, definition of device entities, attributes, binding of physical quantities, and relationships, establishment of fault logic chains.

[0051] Preferably, the S5 specifically comprises:

[0052] Confidence dynamic weighted fusion, calculation of a basic weight formula, dynamic correction factors including an environmental compensation coefficient and a life attenuation factor, multi-modal feature fusion including a temperature gradient, a current THD, and a vibration kurtosis, and output of a final health index;

[0053] Health index grading early warning, state threshold setting, division into health, sub-health, early warning, and failure, dynamic threshold optimization, and load rate self-adaptation.

[0054] Preferably, the S6 specifically comprises:

[0055] Topology structure of an edge lightweight LSTM model, divided into a single-layer LSTM and a fully connected layer, input dimension is a current waveform and a vibration FFT amplitude spectrum, and the lightweight strategy is to quantize weights and prune optimization; hardware acceleration through TensorRT deployment to NVIDIA;

[0056] Abnormality detection logic, based on a multiple harmonic THD change rate to detect current harmonic distortion, based on a wavelet packet energy entropy within a given range to detect a vibration resonance frequency band, and based on a harmonic phase and a vibration envelope correlation to detect a mechanical-electrical coupling abnormality;

[0057] Cloud COMSOL multi-physical field simulation engine, based on a capacitor electrolyte evaporation rate model; multi-physical field coupling, including in turn an electro-thermal field, a Joule heat, a temperature field, a thermal expansion field, a stress field, a sealing deformation, and a permeability change;

[0058] Cloud edge collaborative working mechanism, edge to cloud data flow, current harmonic anomaly marking; edge node to cloud, vibration resonance energy overrun alarm; cloud to COMSOL, trigger capacitance evaporation simulation;

[0059] Cloud to edge instruction flow, threshold dynamic adjustment, cloud computing new threshold and encrypted to edge node, edge node updates LSTM decision threshold; model incremental update, when the simulation result and the measured deviation is greater than the predetermined proportion, then automatically generate a new training set, and issue a lightweight model incremental package every month.

[0060] Compared with the prior art, the beneficial effects of the present application are that:

[0061] The present application solves the problem that the traditional sensor is easy to be interfered by electromagnetic interference and fails in nuclear power plants and extra-high voltage scenes by the synergistic design of the permalloy-iron oxide composite shielding layer and the mu-metal alloy packaging, which still maintains the stable sensing accuracy in a 30kA / 10kV strong electromagnetic field.

[0062] The cross-modal correlation model of current harmonic-vibration energy-temperature gradient is established, the complementary mechanism of Lorentz force displacement sensing and electrostatic induction charge conversion is used, and the synchronous and accurate tracking of IGBT junction temperature, capacitor ESR and bearing wear is realized.

[0063] Based on the cloud edge collaborative architecture, the edge end realizes real-time anomaly detection through the quantized LSTM model, and the cloud relies on the COMSOL multi-physical field simulation to accurately deduce the capacitance evaporation dynamics process, forming a "perception-simulation-optimization" closed loop. BRIEF DESCRIPTION OF DRAWINGS

[0064] Figure 1 The flowchart of the remote real-time monitoring and fault warning method of the power supply operating state. DETAILED DESCRIPTION

[0065] The following description is used to disclose the present application so that those skilled in the art can implement the present application. The preferred embodiments in the following description are only as examples, and other obvious modifications can be thought of by those skilled in the art.

[0066] REFERENCE Figure 1 As shown in the figure, the remote real-time monitoring and fault warning method of the power supply operating state comprises:

[0067] S1, collecting temperature data of key parts of the power supply through a fiber Bragg grating sensor array, and analyzing wavelength shift using wavelength division multiplexing technology;

[0068] S2, collecting electrical parameters through a differential MEMS current / voltage sensor, and the sensor is provided with a permalloy / iron oxide composite shielding layer;

[0069] S3, collect mechanical vibration signals through piezoelectric ceramic vibration sensors;

[0070] S4, for equipment data from different manufacturers, parse Modbus and CAN heterogeneous protocols through regular expression matching and convert them into unified RDF triple format; timestamp multiple source data based on IEEE1588v2 protocol;

[0071] S5, input temperature, current and vibration data into the space-time alignment module, and use ontology to construct semantic mapping relationship; calculate sensor confidence weight and generate device health index through weighted fusion;

[0072] S6, run a lightweight LSTM model on the edge node to detect current harmonic and vibration resonance frequency band anomalies in real time; combine COMSOL multi-physics simulation in the cloud to predict the evaporation rate of the electrolyte, and dynamically adjust the warning threshold.

[0073] It should be noted that S1-S3 constructs an electrical-thermal-mechanical triple sensing network:

[0074] Fiber Bragg grating (FBG) realizes micro-level spatial resolution temperature field monitoring, solves the pain point of traditional thermocouple electromagnetic interference; differential MEMS sensor through permalloy / ferroelectric composite shielding, still maintains 0.1 level current measurement accuracy under 50kA / m strong magnetic field; piezoelectric ceramic vibration sensor captures mechanical fault characteristics with 0.5Hz-12kHz wide frequency response.

[0075] Synergistic effect, when IGBT overheats, current harmonic distortion and cooling fan vibration intensify to form coupling characteristics, significantly improving fault recognition confidence.

[0076] S4 establishes a protocol-time-space three-dimensional standardized framework:

[0077] Regular expression dynamic analysis eliminates device protocol heterogeneity (compatible with 12 types of industrial protocols);

[0078] IEEE1588v2 protocol realizes cross-node μs-level time synchronization, ensuring the phase alignment of vibration and current signals;

[0079] RDF triple gives data machine-readable semantics, such as <Capacitor C101> <Temperature> 62.3℃, laying the foundation for knowledge reasoning.

[0080] S5 proposes ontology and confidence weighted fusion mechanism:

[0081] Based on the power domain ontology library to construct physical constraint rules; introduce real-time sensor confidence weight, such as vibration sensor weight automatically reduces in high temperature environment; health index (HI) fusion temperature gradient, current THD, vibration kurtosis and other cross-domain features.

[0082] S6 realizes edge lightweight and cloud deep computing dual-engine:

[0083] Edge LSTM model (<350KB) captures harmonic / resonance anomaly with 10ms level delay, such as 5th harmonic surge 22%; Cloud COMSOL solves electrolyte evaporation partial differential equation to predict remaining life of capacitor; Dynamic threshold adjustment mechanism upgrades early warning logic from fixed threshold to life-driven, such as threshold automatically tightens as remaining life shortens.

[0084] The S1 specifically includes:

[0085] Sensor array deployment design, FBG sensor nodes are arranged in ring array topology at three core areas: IGBT module heat dissipation substrate surface, DC bus capacitor shell, and transformer winding lead-out line inside the power supply system;

[0086] Each ring array contains multiple FBG sensing units, set with a certain unit spacing, and fixed to the measured surface by high-temperature resistant ceramic glue;

[0087] Anti-electromagnetic interference physical structure is set, the sensing optical fiber adopts polyimide coating and stainless steel armored double-layer protection; the lead path avoids strong electromagnetic areas and is provided with a metal braided shielding conduit

[0088] Wavelength division multiplexing signal analysis, multi-channel spectral encoding, each FBG sensor is preset with an independent Bragg wavelength and a certain wavelength interval; broadband light source is incident on the optical fiber, and the reflected light is split to multiple channels by arrayed waveguide grating;

[0089] Wavelength shift, high-resolution spectrometer is used to monitor the peak wavelength of each channel; Gaussian fitting algorithm is adopted;

[0090] Temperature field reconstruction, three-dimensional thermal field modeling, based on ring array data, inverse distance weighted interpolation method is used to reconstruct IGBT surface temperature distribution, and temperature cloud picture is output:

[0091]

[0092] In the formula, T(x, y) is the estimated temperature at point (x, y), which is calculated by inverse distance weighted interpolation method according to the temperature values of the surrounding sensors; T i is the temperature value measured by the ith sensor, which represents the temperature data point at a known position in space; d i is the Euclidean distance from the interpolation point (x, y) to the ith sensor, which is used to measure the spatial distance between the interpolation point and the known temperature measurement point; p is the attenuation coefficient, which is 2, used to control the degree of influence of distance on weight; n is the number of sensors, representing the total number of sensor data points used to calculate the interpolated temperature;

[0093] Self-diagnosis and error control, fiber breakage monitoring, real-time detection of each channel optical power, and alarm triggered when attenuation is greater than the specified threshold; temperature drift compensation, reference FBG buried in the sensor array to correct the influence of environmental temperature; dynamic calibration, automatic comparison of MEMS temperature probe every 24 hours, and re-calibration when the error is within the specified temperature.

[0094] It should be noted that the sensor array deployment design:

[0095] IGBT heat dissipation substrate, using ring array topology, diameter 30mm, 6 FBG nodes evenly distributed, accurately capturing the radial temperature gradient from the chip to the heat dissipation boundary;

[0096] DC bus capacitor, double ring array arranged on the top / side of the aluminum shell, spacing 15mm, monitoring the heat diffusion path of the electrolyte, avoiding local hot spot detection;

[0097] Transformer winding, axial spiral array attached along the lead-out wire axis, 1 FBG per winding, real-time tracking of temperature rise distribution caused by copper loss.

[0098] Anti-interference packaging technology, double-layer protection structure:

[0099] Inner layer, polyimide coating, thickness 50μm, resistant to 300℃ instantaneous high temperature (IGBT switch transient);

[0100] Outer layer, 316L stainless steel armor, wall thickness 0.3mm, shielding >100kA / m electromagnetic field, while providing 20N compression strength.

[0101] Lead path planning, avoiding directly above the busbar (magnetic field >1T area), along the ground copper bar; metal woven shielding conduit, coverage >90%, 360° ring connection to ground at both ends, shifting induced current.

[0102] Wavelength division multiplexing signal analysis:

[0103] Independent wavelength allocation, each FBG preset Bragg wavelength, 1529nm-1565nm C-band, interval 0.8nm, avoiding AWG channel crosstalk;

[0104] Optical splitter selection, arrayed waveguide grating (AWG) channel number = sensor number x 1.2 (redundancy design), insertion loss <3dB;

[0105] High-resolution spectrometer, 2048-pixel InGaAs array selected, wavelength resolution 0.1pm, corresponding to 0.01℃ temperature sensitivity.

[0106] Gaussian fitting algorithm:

[0107] Raw data pre-processing, background noise (light source fluctuation + dark current) is deducted; Savitzky-Golay filter is applied to smooth the spectral curve.

[0108] Peak position optimization:

[0109]

[0110] In the formula, λ B is the peak wavelength, representing the central wavelength position under certain conditions; argmax λ is to find the value of λ that makes the function take the maximum value, which is used here to determine the peak wavelength; I(λ k ) is the light intensity at wavelength λ k , which is the intensity value in the spectral data; λ k is the kth wavelength point, which is the wavelength position in the spectral data; μ is the central wavelength, which is the center position parameter of the Gaussian function; σ is the standard deviation of the Gaussian function, which is used to control the width of the Gaussian curve.

[0111] Temperature field reconstruction model:

[0112] Adaptive improvement of inverse distance weighted interpolation (IDW), spatial weight function:

[0113] In the formula, w i is the spatial weight of the ith sensor, reflecting the importance of the sensor in the interpolation calculation, and the size of the weight determines the influence of the sensor data on the temperature estimation of the interpolation point; d i is the Euclidean distance from the interpolation point to the ith sensor, the closer the sensor, the greater the influence on the interpolation point, so its weight is higher; p is the attenuation coefficient, taking the value of 2, which strengthens the contribution of the neighboring points and conforms to the inverse square law of heat conduction; The distance threshold d max = 20mm is introduced, which is not involved in the interpolation beyond the range, to avoid distortion;

[0114] Boundary condition processing, supplementing virtual temperature points at the edge of the heat sink to suppress Gibbs oscillation;

[0115] Engineering mapping of three-dimensional thermal field modeling:

[0116] Output tomographic temperature cloud picture, axial layering generates isotherms every 0.1mm on the substrate surface; hotspot positioning, marking >125℃ area as red warning area (silicon chip limit temperature).

[0117] Triple fault-tolerant mechanism:

[0118]

[0119] In the formula, Δλ refReference wavelength drift, refers to the amount of wavelength change caused by temperature change; α T Temperature drift coefficient, indicates the rate of wavelength change with temperature.

[0120] Reference FBG deployment:

[0121] Position selection, buried in the center of the sensor array (thermal equilibrium point), with the same optical fiber link as the measured FBG;

[0122] Environmental isolation, titanium alloy sealed cavity filled with boron nitride thermal insulation material (thermal conductivity 1.2 W / mK), blocking the heat conduction of the measured body.

[0123] S2 specifically includes:

[0124] Composite electromagnetic shielding layer structure, permalloy layer forms low frequency magnetic field, absorbs access, coated on the periphery of the sensor chip, and the slotted area is filled with nanocrystalline strip;

[0125] Ferrite layer, suppress high frequency interference, laminated on the outer layer of permalloy, combined with hot isostatic pressing process;

[0126] The current detection of the differential MEMS sensing unit adopts the Lorentz force principle. When the current-carrying wire passes through the chip, the vertical magnetic field makes the comb electrode deviate. The differential capacitance detection is to form a Wheatstone bridge with the movable electrode and the fixed electrode, and the capacitance change:

[0127]

[0128] In the formula, ΔC is the capacitance change, which is the change of capacitance value caused by the deviation of the comb electrode due to the action of the magnetic field in the differential capacitance detection, reflecting the size of the current; ε0 is the vacuum permittivity, indicating the permittivity in vacuum, used to calculate the size of the capacitance in the electric field; ε r The relative dielectric constant is the ratio of the dielectric constant of the material to the vacuum dielectric constant, which is used to describe the polarization characteristics of the material. Different materials have different relative dielectric constants; A is the electrode area, which refers to the opposite area of the movable electrode and the fixed electrode, with the unit of square meter. The larger the electrode area, the larger the capacitance value; d0 is the initial spacing, which refers to the distance between the movable electrode and the fixed electrode without the action of the magnetic field, with the unit of meter. It affects the initial value of the capacitance; δ is the displacement caused by the magnetic field, which is proportional to the current. When the current passes through the current-carrying wire and is in the vertical magnetic field, the Lorentz force will cause the displacement of the movable electrode. The larger the displacement, the larger the capacitance change;

[0129] Voltage detection, based on electrostatic induction, high voltage electrode induces charge on MEMS cantilever beam, which is converted to voltage by charge amplifier and input impedance;

[0130] Common-mode noise rejection, third-order differential amplifier is adopted to set common-mode rejection ratio; temperature drift compensation, integrated platinum resistance temperature sensor is adopted to correct gain error in real time; active shield drive, external common-mode voltage is collected and fed back to shield shell to control sensor floating ground potential difference;

[0131] Self-calibration process, zero-point drift compensation, measured current is cut off at a certain time, zero-point output is measured, drift curve is stored, and compensation value is predicted through polynomial fitting; range calibration, 1mA / 1V standard signal is injected through built-in reference source, and ADC gain coefficient is dynamically adjusted.

[0132] It should be noted that the composite electromagnetic shielding layer:

[0133] Low-frequency magnetic field suppression of the permalloy layer, magnetic flux path design, 82% nickel content permalloy (initial permeability > 20,000) is adopted to form a closed magnetic ring outside the sensor chip, and 50Hz-1kHz power frequency magnetic field is limited in the alloy layer to reduce the penetration of magnetic force lines;

[0134] Slotted filling technology, nanocrystalline strip (Fe-Si-B system) is filled in the slotted magnetic ring (width 0.5mm), and the high saturation magnetic induction (1.2T) is used to block the magnetic flux leakage, so that the low-frequency magnetic field attenuation is > 40dB;

[0135] High-frequency noise absorption of ferrite layer, laminated structure, Mn-Zn ferrite (thickness 0.8mm) is hot-pressed and compounded outside the permalloy, and the resistivity is > 10 5 Ω·m, which effectively suppresses 10kHz-1MHz switching noise eddy current loss;

[0136] Hot isostatic pressing process, double-metal layer atomic bonding is realized under the condition of 650℃ / 100MPa, the magnetic resistance jump caused by air gap is eliminated, and the shielding integrity is ensured.

[0137] MEMS sensing unit:

[0138] Current detection, Lorentz force-capacitance conversion model, current-carrying wire optimization, U-shaped copper rail with a cross section of 0.1x0.2mm 2 passes through the chip center, and the through-flow capacity reaches 100A (continuous) / 1kA (transient), and the joule heat temperature rise is <3℃;

[0139] Physical correlation of displacement (Δx):

[0140]

[0141] In the formula, B is the magnetic field strength of the permanent magnet (0.5T neodymium iron boron); k is the stiffness coefficient of the folded beam (design value 80N / m); I is the current flowing through the wire, unit is ampere; L is the length of the comb tooth arm (1.2mm).

[0142] Sensitivity of capacitance change (ΔC) is enhanced by interdigital array, 120 teeth, area A = 0.15 mm 2 Sensitivity is enhanced to 20 fF / A; initial gap d0 is reduced to 2 μm (SOI silicon wafer deep etching process), making ΔC linearly related to Δx.

[0143] Voltage detection, electrostatic induction-charge conversion mechanism:

[0144] Cantilever beam design, titanium alloy beam (size 200 x 50 x 5 μm 3 ) supports high voltage electrode (10 kV resistance); induced charge Q = C eq · V, where equivalent capacitance C eq = 1.5 pF;

[0145] Impedance matching, charge amplifier input impedance > 10 15 Ω (using JFET input stage), leakage current < 1 pA; feedback capacitance C_f = 10 pF ceramic capacitor (NP0 material), temperature drift < ± 30 ppm / °C.

[0146] Anti-interference and self-calibration:

[0147] Common-mode noise suppression technology, third-order differential amplifier:

[0148] First stage: instrument amplifier (INA823, CMRR = 120 dB @ 60 Hz);

[0149] Second stage: active filter (cutoff frequency 1 kHz, attenuation -40 dB / dec);

[0150] Third stage: notch filter (center frequency 150 kHz, eliminates IGBT switching noise).

[0151] Active shielding drive, real-time acquisition of common-mode voltage V_cm→ operational amplifier driving shield shell→ making the potential difference between sensor and ground < 10 mV, eliminating capacitive coupling interference.

[0152] Real-time compensation for temperature drift, platinum resistance temperature sensor integrated on the back of the MEMS chip (0.3 mm from the sensitive area); resistance-temperature relationship: R(T) = R0[1 + α(T-T0)], α = 0.00385 / °C

[0153] In the formula, R(T) is the resistance value at temperature T, with units of ohms, which reflects the resistance change of the platinum resistance temperature sensor at different temperatures; R0 is the initial resistance value, that is, the resistance value at the reference temperature T0; a is the resistance temperature coefficient, with units of per degree Celsius; T is the current temperature, with units of degrees Celsius; T0 is the reference temperature, which is usually the benchmark temperature for sensor calibration and measurement;

[0154] Dynamic correction strategy:

[0155] Gain error: G corr = G0 / [1 + β(T - 25)]

[0156] In the formula, G corr is the corrected gain, used to compensate for the influence of temperature changes on the system gain; G0 is the initial gain, which is the gain value at the reference temperature (25°C); β is the gain temperature coefficient, indicating the rate of gain change with temperature, which is -0.02% / °C here; T is the current temperature, with units of degrees Celsius;

[0157] Zero drift: V offset (T) = aT 2 + bT + c

[0158] In the formula, V offset (T) is a, b, and c are the coefficients of the second-order polynomial fit, used to describe the nonlinear relationship of zero drift with temperature change;

[0159] The S3 specifically includes:

[0160] The piezoelectric ceramic vibration sensor uses a lead zirconate titanate piezoelectric ceramic sheet in a sandwich structure, with silver electrodes plated on the upper and lower surfaces of the ceramic sheet, which is bonded to a titanium alloy base through conductive epoxy resin, and the base has a resonance frequency greater than a specified frequency;

[0161] The target installation strategy and monitoring point positioning include the power fan bearing seat, transformer fastening bolt, and IGBT heat sink base plate; the installation method is rigid connection by bolts, and the contact surface is coated with thermal conductive silicone grease; the sensor axis is aligned with the main direction of vibration;

[0162] The wideband signal capture sets the frequency response range to cover the power mechanical fault characteristic frequency band; the anti-interference design uses electromagnetic shielding, and the shell is made of μ-Metal alloy; the signal line is double-shielded with copper braid and aluminum foil, and the end is connected to the acquisition circuit through a feedthrough filter; temperature compensation is achieved by built-in NTC thermistors, which correct the piezoelectric coefficient temperature drift in real time.

[0163] It should be noted that the lead zirconate titanate (PZT) piezoelectric ceramic core uses PZT-5H ceramic (zirconium / titanium ratio 52 / 48), with a piezoelectric constant d 33 = 650 pC / N, ensuring high sensitivity capture of weak vibrations;

[0164] Sandwich structure:

[0165] Upper and lower silver electrode layers (0.1 μm thickness) are bonded by conductive epoxy (silver content > 80%) to achieve efficient charge extraction;

[0166] Titanium alloy base (Ti-6Al-4V) provides > 25 kHz resonance frequency (much higher than power supply fault band), avoiding self-resonance interference.

[0167] Temperature stability enhancement design, NTC thermistor compensation mechanism; base thermal stress isolation, titanium alloy and ceramic sheet add molybdenum foil transition layer (CTE = 5.5 x 10 -6 / ℃), inhibiting thermal expansion mismatch.

[0168] Targeted installation strategy:

[0169] Failure physical correlation of monitoring point positioning:

[0170]

[0171] Installation process, rigid connection with bolts, M4 titanium alloy bolts (tensile strength > 900 MPa), apply 10 ± 0.5 Nm pre-tightening torque (prevent loosening or overloading); contact surface heat-conducting silicone grease (containing boron nitride filler) fills micro gaps, thermal resistance < 0.05 K / W, while transmitting high-frequency vibrations;

[0172] Axis calibration requirements, laser alignment ensures that the sensor Z-axis deviates from the main vibration direction by < 3°, avoiding cosine error leading to sensitivity loss > 5%.

[0173] Wideband signal capture and anti-interference:

[0174] Frequency response range, low frequency lower limit 0.5 Hz (covering transformer DC bias vibration); high frequency upper limit 12 kHz (capturing early bearing damage characteristics);

[0175] Resonance peak control, through base mass (4.5 g) and PZT stiffness (k = 10 8 N / m) matching, resonance frequency > 25 kHz (guaranteeing 0.5-12 kHz flatness ± 3 dB);

[0176] Electromagnetic compatibility (EMC) triple protection:

[0177] Shell shielding, 2 mm thick μ-Metal alloy (initial magnetic permeability > 100,000), attenuating > 60 dB of magnetic field below 30 MHz;

[0178] Cable double shielding, inner layer, 100% coverage of aluminum foil (reflect high frequency interference); outer layer, tinned copper braid (coverage > 95%, guide low frequency eddy current);

[0179] Feedthrough filter, through-hole capacitor (C = 1nF) + ferrite bead (100Ω@100MHz), suppress conducted interference.

[0180] Dynamic temperature compensation verification, NTC real-time correction sensitivity in the range of -40℃ ~ 120℃:

[0181] Uncompensated temperature drift up to -12% (120℃);

[0182] After compensation, the error is less than ± 3% (full temperature range).

[0183] The S4 specifically includes:

[0184] Regular expression matching mechanism, protocol feature library construction, pre-set Modbus / TCP message template, pre-set CAN2.0B extended frame template;

[0185] Private protocol compatibility strategy, dynamic learning through regular capture group, supporting fuzzy matching, when the protocol structure is unknown, identifying the payload according to byte entropy value;

[0186] Finite state machine analysis process receives raw messages, detects frame headers, determines protocol types through regular matching, extracts payloads, and retransmits exceptions after checksum verification;

[0187] Unified RDF triple conversion, semantic mapping rules; using context enhancement technology, ontological reasoning, constructing power domain ontology library, automatically inferring semantic relationships;

[0188] Space-time reference binding, location information using WGS-84 coordinates, time reference synchronized through IEEE1588v2.

[0189] It should be noted that the protocol analysis engine:

[0190] Regular expression matching mechanism:

[0191] Protocol feature library construction, Modbus / TCP template, pre-set function code regular expression, supporting variable length message truncation (over 1024 bytes automatically framed); CAN2.0B extended frame template, 29-bit ID matching rule [0-9A-F]{8}, dynamic analysis of data field (DLC = 8, reorganized according to little-endian sequence);

[0192] Private protocol compatibility strategy:

[0193] Byte entropy value identification, calculating the Shannon entropy of the payload segment:

[0194] H = -∑p(x i )log2p(x i )

[0195] In the formula, H is the Shannon entropy, used to measure the uncertainty or randomness of data; p(x i ) is the probability of each byte x i occurrence, reflecting the relative frequency of the byte in the data set; log2 is the logarithm with base 2, used to calculate the amount of information; when H>6.5H>6.5, it is determined that the effective load (encrypted / compressed data entropy value feature);

[0196] The fuzzy matching engine uses the Levenshtein distance algorithm to find templates with a similarity of >85% in the protocol library as the analysis benchmark.

[0197] Fault-tolerant implementation of finite state machine (FSM):

[0198] Receive raw packet → frame header detection (0xA5 synchronization word) → protocol type determination (regular matching confidence >90%) → payload extraction (length field dynamic interception) → checksum verification (CRC32 / sum check dual mode) → exception handling: trigger NACK retransmission (up to 3 times) when verification fails

[0199] Anti-interference enhancement, still able to recover the original data through the Viterbi algorithm under 50% bit error rate; real-time guarantee, single packet parsing delay <200μs.

[0200] Precise binding of space-time reference:

[0201] Time synchronization mechanism, IEEE1588v2 precise time service, master clock uses GPS tamed rubidium atomic clock; slave clock synchronization deviation <±100ns (industrial switch delay compensation); time keeping ability, when the master clock fails, the local OCXO maintains <1μs drift for 24 hours.

[0202] Spatial location calibration, WGS-84 coordinate binding, equipment room coordinates are measured by RTK-GPS; indoor equipment uses UWB positioning tags;

[0203] Space-time correlation database, each data is attached with a four-dimensional label (longitude, latitude, altitude, PTP timestamp).

[0204] The S5 specifically includes:

[0205] Unified space-time reference of multi-source data, time dimension alignment, synchronize each sensor clock based on IEEE1588v2, use cubic spline interpolation for asynchronous data, and resample to the specified reference frequency;

[0206] Space coordinate binding, establish device three-dimensional coordinate system, map FBG temperature point, vibration sensor to unified space grid;

[0207] Cross-modal data correlation rule: temperature correlates with IGBT junction temperature, current harmonic correlates with capacitor ESR change, vibration energy correlates with fan bearing wear;

[0208] Ontology-driven semantic mapping, build power domain ontology library, core ontology elements include class, define device entity; attribute, bind physical quantity; relationship, establish fault logic chain;

[0209] Confidence dynamic weighted fusion, calculate basic weight formula; dynamic correction factor, including environmental compensation coefficient, life attenuation factor; multi-modal feature fusion includes temperature gradient, current THD, vibration kurtosis, output final health index;

[0210] Health index grading early warning, state threshold setting, divided into healthy, sub-healthy, early warning, fault, dynamic threshold optimization, load rate self-adaptation.

[0211] It should be noted that space coordinate mapping, device three-dimensional coordinate system construction:

[0212] Origin, geometric center of transformer; coordinate axis, X-axis parallel to busbar row, Y-axis perpendicular to installation base surface, Z-axis along the direction of cooling air flow;

[0213] Sensor space binding, FBG temperature point, mapped to IGBT chip junction temperature hot spot; vibration sensor, vibration vector converted to unified coordinate system through direction cosine matrix.

[0214] Fault logic chain derivation of power domain ontology library:

[0215] IGBT junction temperature ↑→ cooling demand ↑→ fan speed ↑→ bearing wear acceleration ↑→ vibration energy; cross-modal correlation rule:

[0216] Correlation pairs Physical mechanisms Temperature-current harmonics Increased nonlinearity of turn-on voltage drop due to elevated IGBT junction temperature Vibration-capacitance ESR Mechanical vibration accelerates electrolyte stratification and oxide film rupture Temperature gradient-fan wear Uneven heat dissipation leads to bearing thermal deformation

[0217] Confidence dynamic weighted fusion:

[0218] Basic weight distribution:

[0219] FBG temperature sensor: weight w T = 0.9 (strong anti-interference / long life);

[0220] MEMS current sensor: weight w I = 0.85 (easily disturbed by strong magnetic field);

[0221] Vibration sensor: weight w V = 0.8 (risk of loose installation).

[0222] Environmental compensation coefficient:

[0223] In the formula, C env is an environmental compensation coefficient, used to compensate the system according to environmental humidity and temperature; RH is relative humidity, indicating the ratio of water vapor content in the air to the saturated water vapor content at the same temperature; T is the environmental temperature, in degrees Celsius; when the humidity RH is greater than 80%, the vibration sensor weight is automatically reduced;

[0224] Lifetime attenuation factor: λ = e -0.0001·t , t is the running hours;

[0225] After 5 years of service, the FBG weight is attenuated to 90% of the initial value.

[0226] Multi-modal feature fusion formula:

[0227] In the formula, HI is the health index, indicating the comprehensive health status of the equipment; w T is the temperature gradient weight, indicating the weight of the temperature gradient feature in the health index; w I is the total harmonic distortion (THD) weight, indicating the weight of the harmonic distortion feature in the health index; w V is the kurtosis weight, indicating the weight of the kurtosis feature of the vibration signal in the health index; ΔT grad is the temperature gradient, indicating the rate of change of the IGBT substrate radial temperature; THD is the total harmonic distortion, indicating the harmonic content of the current or voltage waveform; Kurtosis is the kurtosis, which measures the strength of the impulse component in the vibration signal; C env is the environmental compensation coefficient, used to adjust the health index according to environmental conditions; λ is the lifetime attenuation factor, reflecting the effect of equipment running time on its health status;

[0228] Health index grading early warning mechanism:

[0229] Four-state threshold setting:

[0230]

[0231] The S6 specifically includes:

[0232] The topology structure of the edge lightweight LSTM model is divided into a single-layer LSTM and a fully connected layer; the input dimension is the current waveform and the vibration FFT amplitude spectrum; the lightweight strategy is to quantize the weight, and the pruning optimization is performed; hardware acceleration is performed through TensorRT deployment to NVIDIA;

[0233] Abnormality detection logic, based on the rate of change of the multiple harmonic THD, detects current harmonic distortion; based on the wavelet packet energy entropy within a certain range, detects the vibration resonance frequency band; based on the correlation between harmonic phase and vibration envelope, detects the abnormality of electromechanical coupling;

[0234] Cloud COMSOL multi-physics simulation engine, based on the evaporation rate model of the electrolyte in the capacitor; multi-physics coupling, including in turn electric-thermal field, Joule heat, temperature field, thermal expansion field, stress field, sealing deformation, permeability change;

[0235] Cloud edge collaborative working mechanism, edge to cloud data flow, edge node to cloud, current harmonic abnormality marking; edge node to cloud, vibration resonance energy out-of-limit alarm; cloud to COMSOL, trigger capacitor evaporation simulation;

[0236] Cloud to edge instruction flow, threshold dynamic adjustment, cloud computing new threshold and encrypted to edge node, edge node updates LSTM decision threshold; model incremental update, when the simulation result deviates from the measured value by more than a certain proportion, a new training set is automatically generated, and a lightweight model incremental package is issued every month.

[0237] It should be noted that the edge lightweight LSTM model:

[0238] Input feature dimension compression, current waveform is 128 points per cycle sampling→extract 1-13 harmonic amplitude and phase angle through FFT (dimension 26);

[0239] Vibration signal is 12kHz sampling→512-point FFT amplitude spectrum→aggregated into 1 / 3 octave energy (dimension 10); dimension reduction effect: input layer is compressed from 1400 dimensions to 36 dimensions, calculation amount is reduced by 97%;

[0240] Multi-modal anomaly detection:

[0241] Accurate capture of current harmonic distortion, dynamic threshold of THD change rate:

[0242]

[0243] THD t is the total harmonic distortion at the current time, reflecting the content of harmonic components in the current power signal; THD t-10s is the total harmonic distortion before 0 seconds, used for comparison with the current THD value to calculate the change rate of THD; when ΔTHD exceeds the standard for 30 seconds, it is determined that the ESR of the capacitor has suddenly risen (a precursor to electrolyte drying);

[0244] Phase correlation diagnosis, calculate the standard deviation of the phase difference of 5 / 7 harmonics, if σ(Δφ 5-7 )>15°, it indicates that the rectifier bridge contact is poor; σ is the standard deviation, Δφ 5-7Phase difference between 5th and 7th harmonics;

[0245] Vibration resonance band energy entropy monitoring:

[0246] Wavelet packet decomposition tree, original signal→8 layers of decomposition→256 sub-bands;

[0247] Extract 3.2kHz-4.8kHz frequency band(capacitor shell resonance area)

[0248] Energy entropy E judgment basis, normal state: E∈[1.8, 2.2]; over-limit alarm: E>2.5(predicting mechanical loosening).

[0249] Implementation steps:

[0250] Step 1: Fiber Bragg grating temperature sensing array installation

[0251] Paste FBG sensors in the key heat areas of the power supply(IGBT heat dissipation base plate, capacitor group bus bar), analyze the wavelength shift of the optical signal through the wavelength division multiplexer, convert it into temperature value, use distributed optical cable topology, and connect up to 32 measuring points in series with a single optical fiber;

[0252] Step 2: Electromagnetic compatibility type electric parameter acquisition

[0253] Install differential MEMS sensors on AC and DC bus bars, the built-in permalloy / ferromagnetic composite shielding layer can attenuate more than 30dB magnetic field interference, and realize current / voltage synchronous sampling;

[0254] Step 3: Targeted installation of vibration sensors

[0255] Install piezoelectric ceramic vibration sensors through titanium alloy bolts at the fan bearing seat, transformer bolt mechanical weak points, with an axis deviation of <3° from the main direction of vibration, covering a wide frequency band of 0.5Hz-12kHz;

[0256] Step 4: Protocol adaptive analysis engine

[0257] Parse the original data of the device through the regular expression template library(preset Modbus function code, CAN ID rule); enable byte entropy value analysis for private protocols, dynamically extract the payload; convert the analysis result into RDF triple, and map it to unified semantics according to the power ontology library;

[0258] Step 5: Nanosecond-level space-time reference binding

[0259] Use IEEE 1588v2 precision time service, the master clock synchronizes the timestamp deviation of each node <100ns; enter the spatial database according to the WGS-84 standard, forming a "longitude-latitude-elevation" three-dimensional label;

[0260] Step 6: Temporal Alignment and Semantic Association

[0261] Perform three times of spline interpolation on asynchronous data, resample to 10kHz reference frequency; establish physical association chain based on ontology library;

[0262] Step 7: Dynamic Weighted Health Index Generation

[0263] Calculate sensor confidence weights (FBG temperature weight 0.9, vibration sensor weight 0.8); fuse temperature gradient, current THD, vibration kurtosis, etc. features, output 0-100 health index;

[0264] Step 8: Edge Real-time Anomaly Detection

[0265] Deploy lightweight LSTM model (INT8 quantization + pruning optimization) on NVIDIA Jetson edge node: on the current side, monitor THD change rate > 0.5% / s harmonic distortion; on the vibration side, analyze 3.2-4.8kHz frequency band wavelet packet energy entropy > 2.5 resonance features;

[0266] Step 9: Cloud Multi-physical Field Simulation Optimization

[0267] When the edge alarms, the cloud starts COMSOL electrolyte evaporation simulation: coupling electrothermal field, thermal expansion field, sealing stress field, predicting the remaining life of the capacitor; dynamically calculate the new warning threshold and encrypt it to the edge node;

[0268] Step 10: Model Self-evolution Mechanism

[0269] Compare simulation and measured data every month, generate incremental training set automatically when deviation > 15%; push <500KB LSTM model incremental package to edge node, update full connection layer weights;

[0270] Step 11: Hierarchical Warning and Operation Response

[0271] Health status four-level judgment

[0272] Healthy (HI > 85): green identification, routine inspection;

[0273] Sub-health (70-85): yellow warning, encrypted data collection frequency;

[0274] Warning (50-70): orange alert, prepare spare parts and plan shutdown;

[0275] Fault (HI < 50): red emergency stop, immediately cut off power and repair;

[0276] Step 12: Threshold Dynamic Optimization

[0277] Automatically adjust the reference threshold according to the load rate (HI threshold tightens to 72 under heavy load and relaxes to 92 under light load); combined with the device life attenuation coefficient, avoid false alarms during the aging period.

[0278] In summary, the advantages of the present application are:

[0279] Through optical fiber sensing and composite electromagnetic shielding double-drive, the difficulty of monitoring in strong electromagnetic environment is solved. The FBG temperature array uses the wavelength immune characteristic, and still maintains the precision of 0.1 ℃ under the impact of 30 kA short-circuit current; the MEMS sensor permalloy / ferromagnetic layer shielding attenuates the magnetic field interference by more than 30 dB; the vibration sensor adopts μ-Metal alloy shell, so that the high-frequency electromagnetic noise suppression ability is improved.

[0280] An electric-thermal-mechanical cross-modal coupling analysis model is established, the coherence function algorithm of current harmonic phase and vibration envelope is used to accurately capture the hidden faults such as IGBT solder layer debonding; based on the COMSOL simulation of electrolyte evaporation kinetics equation, the capacitor failure is early warned for 120 days; the spatiotemporal alignment ontology mapping reveals the cascading effect chain of harmonic distortion, junction temperature rise and bearing wear.

[0281] Through edge real-time detection-cloud simulation deduction-closed-loop dynamic optimization, the edge lightweight LSTM model realizes the millisecond response of harmonic mutation and resonance frequency band; the cloud multi-physics field coupling simulation dynamically adjusts the warning threshold to adapt to load fluctuation and device aging; the monthly incremental model updating mechanism enables the diagnostic algorithm to continuously evolve, and is free of on-site calibration for life.

[0282] The above shows and describes the basic principles, main features and advantages of the present application. Those skilled in the art should understand that the present application is not limited to the above embodiments, and the above embodiments and descriptions in the specification are only the principles of the present application. Without departing from the spirit and scope of the present application, various changes and improvements can be made to the present application, and these changes and improvements all fall within the scope of the claimed present application. The scope of protection claimed by the present application is defined by the appended claims and their equivalents.

Claims

1. A remote real-time monitoring and fault warning method for power supply operation status, characterized in that: include: S1, collects temperature data of key parts of the power supply through a fiber Bragg grating sensor array, and uses wavelength division multiplexing technology to analyze wavelength offset; S2, collects electrical parameters through differential MEMS current / voltage sensors with built-in Permalloy / ferrite composite shielding layer; S3, collecting mechanical vibration signals through piezoelectric ceramic vibration sensors; S4. Parse Modbus and CAN heterogeneous protocols using regular expression matching for device data from different manufacturers and convert them into a unified RDF triple format; timestamp multi-source data based on the IEEE1588v2 protocol; S5. Input the temperature, current, and vibration data into the spatiotemporal alignment module and construct a semantic mapping relationship using ontology; Calculate sensor confidence weights and perform weighted fusion to generate a device health index; S6. Run a lightweight LSTM model on edge nodes to detect current harmonics and vibration resonance frequency band anomalies in real time; Combined with COMSOL multi-physics field simulation in the cloud, the evaporation rate of capacitor electrolyte is predicted and the warning threshold is dynamically adjusted.

2. The remote real-time monitoring and fault warning method for power supply operation status according to claim 1 is characterized in that: The S4 specifically includes: Regular expression matching mechanism, protocol feature library construction, preset Modbus / TCP message template, preset CAN2.0B extended frame template; Private protocol compatibility strategy, through dynamic learning of regular capture groups, supports fuzzy matching, and identifies payload by byte entropy value when the protocol structure is unknown; The finite state machine parsing process is to receive the original message, perform frame header detection, determine the protocol type through regular matching, extract the payload, check and verify, and then perform abnormal retransmission; Unify RDF triple conversion and semantic mapping rules; use context enhancement technology and ontology reasoning to build a power domain ontology library and automatically infer semantic relationships; The space-time reference is bound, the location information uses WGS-84 coordinates, and the time reference is synchronized via IEEE1588v2.

3. The remote real-time monitoring and fault warning method for power supply operation status according to claim 2 is characterized in that: Said S1 specifically includes: The sensor array deployment design arranges FBG sensor nodes in a ring array topology at three core areas within the power system: the surface of the IGBT module heat sink substrate, the DC bus capacitor housing, and the transformer winding lead wires. Each ring array contains multiple FBG sensor units, set with a predetermined unit spacing, and fixed to the surface to be measured by high-temperature resistant ceramic glue; A physical structure is designed to resist electromagnetic interference, and the sensing optical fiber is protected by a double layer of polyimide coating and stainless steel armor. The lead path avoids strong electromagnetic areas and is passed through a metal braided shielding conduit. Wavelength division multiplexing signal analysis, multi-channel spectral encoding, each FBG sensor preset independent Bragg wavelength and established wavelength interval; broadband light source is incident on the optical fiber, and the reflected light is split into multiple channels by the arrayed waveguide grating; wavelength offset is resolved, and the peak wavelength of each channel is monitored using a high-resolution spectrometer; and a Gaussian fitting algorithm is used.

4. The remote real-time monitoring and fault warning method for power supply operation status according to claim 3 is characterized in that: Said S1 specifically includes: Temperature field reconstruction, 3D thermal field modeling, based on the ring array data, uses the inverse distance weighted interpolation method to reconstruct the IGBT surface temperature distribution and output the temperature cloud map: Where T(x,y) is the estimated temperature at point (x,y), which is calculated based on the temperature values ​​of surrounding sensors using the inverse distance weighted interpolation method; T i The temperature value measured by the i-th sensor represents the temperature data point at a known position in space; d i is the Euclidean distance from the interpolation point (x, y) to the i-th sensor, which is used to measure the spatial distance between the interpolation point and the known temperature measurement point; p is the attenuation coefficient, which is set to 2 and is used to control the influence of distance on the weight; n is the number of sensors, which represents the total number of all sensor data points used to calculate the interpolated temperature; Fiber break monitoring: Real-time detection of optical power in each channel. An alarm is triggered if the attenuation exceeds a predetermined threshold. Temperature drift compensation: Reference FBGs are embedded in the sensor array to correct for ambient temperature effects. Dynamic calibration: Regular and automatic comparison of MEMS temperature probes is performed, and recalibration is performed when the error reaches a predetermined temperature.

5. The remote real-time monitoring and fault warning method for power supply operation status according to claim 4 is characterized in that: The S2 specifically includes: Composite electromagnetic shielding layer structure, the Permalloy layer forms a low-frequency magnetic field, absorbs the path, covers the periphery of the sensor chip, and the gap is filled with nanocrystalline strips; The ferrite layer, which suppresses high-frequency interference, is laminated on the outer layer of Permalloy and combined with the hot isostatic pressing process; The current detection of the differential MEMS sensing unit uses the Lorentz force principle. When the current-carrying wire passes through the chip, the vertical magnetic field causes the comb-tooth electrodes to deflect. For differential capacitance detection, the movable electrode and the fixed electrode form a Wheatstone bridge. The capacitance change is: Where ΔC is the capacitance change, which refers to the change in capacitance caused by the offset of the comb electrode due to the magnetic field in differential capacitance detection, and reflects the magnitude of the current; ε0 is the vacuum dielectric constant, which represents the permittivity in a vacuum and is used to calculate the magnitude of the capacitance in an electric field; ε r is the relative dielectric constant, which is the ratio of the dielectric constant of the material to the dielectric constant of the vacuum, and is used to describe the polarization characteristics of the material. Different materials have different relative dielectric constants. A is the electrode area, which refers to the area directly opposite the movable electrode and the fixed electrode, and the unit is square meters. The larger the electrode area, the greater the capacitance value. d0 is the initial spacing, which refers to the distance between the movable electrode and the fixed electrode when there is no magnetic field, and the unit is meter. It affects the initial value of the capacitance. δ is the displacement caused by the magnetic field, which is proportional to the current. When the current passes through the current-carrying wire and is in a perpendicular magnetic field, the Lorentz force will cause the movable electrode to displace. The greater the displacement, the greater the change in capacitance. Voltage detection is based on electrostatic induction. The high-voltage electrode induces charge on the MEMS cantilever beam, which is converted into voltage through a charge amplifier and input into impedance.

6. The remote real-time monitoring and fault warning method for power supply operation status according to claim 5 is characterized in that: The S2 specifically includes: Common-mode noise suppression uses a third-order differential amplifier to set the common-mode rejection ratio; temperature drift compensation uses an integrated platinum resistance temperature sensor to correct gain errors in real time; active shield drive collects external common-mode voltage and feeds it back to the shield shell to control the sensor's floating ground potential difference; Self-calibration process, zero drift compensation, timed cut-off of the measured current, measurement of zero output, storage of drift curve, and prediction of compensation value through polynomial fitting; range calibration, built-in reference source injects 1mA / 1V standard signal, and dynamically adjusts the ADC gain coefficient.

7. The remote real-time monitoring and fault warning method for power supply operation status according to claim 6, characterized in that: The S5 specifically includes: The spatiotemporal benchmarks of multi-source data are unified, the time dimension is aligned, the clocks of each sensor are synchronized based on IEEE1588v2, and cubic spline interpolation is used for asynchronous data to resample to the established reference frequency; Spatial coordinate binding, establishing a three-dimensional coordinate system for the device, and mapping FBG temperature points and vibration sensors to a unified spatial grid; Cross-modal data association rules are: temperature is associated with IGBT junction temperature, current harmonics are associated with capacitor ESR changes, and vibration energy is associated with fan bearing wear; Ontology-driven semantic mapping builds a power domain ontology library. The core ontology elements include classes, which define device entities; attributes, which bind physical quantities; and relationships, which establish fault logic chains.

8. The remote real-time monitoring and fault warning method for power supply operation status according to claim 7 is characterized in that: The S5 specifically includes: Dynamic weighted fusion of confidence to calculate the basic weight formula; dynamic correction factors, including environmental compensation coefficient and life attenuation factor; multimodal feature fusion including temperature gradient, current THD, vibration kurtosis, and output of the final health index; Health index graded warning, status threshold setting, divided into healthy, sub-healthy, warning, fault, dynamic threshold optimization, load rate adaptation.

9. The remote real-time monitoring and fault warning method for power supply operation status according to claim 8, characterized in that: The S6 specifically includes: The topology of the edge lightweight LSTM model consists of a single LSTM layer and a fully connected layer. The input dimensions are the current waveform and the vibration FFT amplitude spectrum. The lightweight strategy involves weight quantization and pruning optimization. Hardware acceleration is implemented and deployed to NVIDIA via TensorRT. Abnormal detection logic detects current harmonic distortion based on the rate of change of multiple harmonics THD; detects vibration resonance frequency band based on the energy entropy of wavelet packets within a given range; and detects electromechanical coupling abnormalities based on the correlation between harmonic phase and vibration envelope. The cloud-based COMSOL multi-physics simulation engine is based on a capacitor electrolyte evaporation rate model. Multi-physics coupling includes electrothermal field, Joule heating, temperature field, thermal expansion field, stress field, seal deformation, and permeability change. Cloud-edge collaborative working mechanism, edge-to-cloud data flow, current harmonic anomaly marking; edge node to cloud, vibration resonance energy exceeding limit alarm; cloud to COMSOL, triggering capacitor evaporation simulation; The cloud-to-edge instruction flow and threshold are dynamically adjusted. The cloud calculates the new threshold and encrypts it and sends it to the edge node. The edge node updates the LSTM judgment threshold. The model is incrementally updated. When the deviation between the simulation result and the actual measurement is greater than the established ratio, a new training set is automatically generated, and a lightweight model incremental package is issued every month.

10. The remote real-time monitoring and fault warning method for power supply operation status according to claim 9, characterized in that: The S3 specifically includes: The piezoelectric ceramic vibration sensor uses lead zirconate titanate piezoelectric ceramic sheets with a sandwich structure. The upper and lower surfaces of the ceramic sheets are plated with silver electrodes and bonded to a titanium alloy base with conductive epoxy resin. The resonant frequency of the base is greater than the predetermined frequency. Targeted installation strategy and monitoring point positioning, including the power supply fan bearing seat, transformer fastening bolts, and IGBT heat sink baseplate; installation method, rigid bolt connection, contact surface coated with thermal grease; sensor axis aligned with the main vibration direction; Wideband signal capture with a set frequency response range covering the characteristic frequency band of power supply mechanical failures; anti-interference design, electromagnetic shielding, and a housing made of μ-Metal alloy; the signal line is double-shielded with copper braid and aluminum foil, and the end is connected to the acquisition circuit through a feed-through filter; temperature compensation, with a built-in NTC thermistor, corrects the piezoelectric coefficient temperature drift in real time.

Images