A residual neural network-based power distribution cabinet sealing diagnosis method and system

Abstract

The present application relates to power distribution equipment state monitoring technical field, disclose a kind of residual neural network-based switchboard sealing diagnosis method and system, utilize multi-path time division encoding pneumatic excitation to generate spatial differentiation pressure response waveform with, and using adaptive sampling technique extracts pressure attenuation characteristics, using physical model inversion preliminary leakage equivalent diameter, simultaneously using residual neural network compensates non-linear residual;At the same time, system introduces the shift detection mechanism based on maximum mean difference, when model mismatch, forced execution safety retreat logic, ensure the robustness of diagnosis under the scene of electric power industry.The present application improves the diagnostic accuracy, spatial positioning ability and long-term operation reliability without relying on additional sensors, has important practical value.

Description

Technical Field

[0001] This invention relates to the field of power distribution equipment condition monitoring technology, and in particular to a method and system for diagnosing the sealing of power distribution cabinets based on residual neural networks. Background Technology

[0002] The normal operation of internal components in complete sets of electrical equipment such as distribution cabinets and switchgear highly depends on the good sealing and protection performance of the cabinet to prevent external dust, moisture, and harmful media from entering and causing insulation breakdown or short circuit faults. Currently, online monitoring of the sealing performance and leakage source diagnosis of sealed cavity structures such as distribution cabinets have become a research hotspot in the field of smart grid condition monitoring.

[0003] In cabinet seal diagnostics, the most commonly used method is the pressure decay method (i.e., pressure drop method). This method involves filling the chamber under test with a slightly positive pressure gas, cutting off the gas supply, monitoring the pressure decay curve of the chamber over time using a pressure sensor, and calculating the total leakage rate based on the ideal gas law. However, existing technologies have the following significant drawbacks when converting transient pressure decay into an accurate physical leakage equivalent diameter: 1. The interior of a distribution cabinet is not an ideal, regular cavity, but rather densely packed with irregular electrical equipment such as circuit breakers, busbars, and transformers. Existing technologies typically use the geometric volume of the distribution cabinet's outer shell as the equivalent volume of the cavity for leakage rate inversion, ignoring the significant physical space occupied by the internal equipment. This leads to issues with the required volume input parameters for the inversion. The deviation is huge, which seriously affects the accuracy of the physical main link leakage rate calculation; 2. During long-term field operation, distribution cabinets have inherent static background leakage rates at locations such as mechanical overlaps of cabinet doors and hinge joints that cannot be completely eliminated by temporary sealing. Existing calculation models lack algebraic compensation terms for this background leakage rate. The precise introduction of [the technology] leads to the calculated equivalent diameter. These background interferences cannot be eliminated, making it difficult to accurately reflect sudden faults and leaks; 3. Equivalent orifice flow coefficient used in classical equivalent orifice flow calculation This flow coefficient is usually set as a fixed constant, but in actual multi-physics coupling and complex cabinet flow environments, it exhibits strong nonlinear dynamic changes due to the influence of Reynolds number, flow boundary and compression effect. Using a fixed constant for calculation will inevitably introduce obvious physical simplification systematic errors.

[0004] In terms of leak location, existing technologies often employ the method of deploying distributed acoustic sensor arrays (such as distributed ultrasonic probes or acoustic vibration sensor networks) inside the distribution cabinet or in leak-prone areas. While this method can achieve location, deploying multiple physical sensors inside the distribution cabinet cavity, which is characterized by strong electromagnetic fields and high voltage, makes them highly susceptible to high-frequency electromagnetic interference from the large current interruption of arcs by internal electrical switches. Furthermore, it significantly increases hardware configuration costs, wiring complexity, and the physical failure rate of the sensors themselves.

[0005] To reduce hardware costs, the industry has attempted to use a single pressure sensor combined with data-driven models such as machine learning to predict leak points. However, because the static pressure field of the distribution cabinet cavity reaches isotropy in a very short time under a single air intake path, the pressure decay scalar curve collected by a single measuring point lacks the ability to distinguish leak sources in different spatial locations. This severely limits the spatial observability of the "single sensor-single excitation channel" configuration, and the algorithm model is prone to false alarms or over-localization when it forcibly predicts a unique location coordinate. Although time-division excitation through multiple air intake branches can introduce transient pressure transmission differences, in actual deployment, drastic fluctuations in ambient temperature and humidity and sensor aging can lead to severe "concept drift." When existing adaptive update algorithms detect data distribution drift, due to the lack of online real labels, they often directly use the model's own historical prediction results as pseudo-labels for self-training updates. This is very likely to cause "confirmation bias" and lead to the rapid collapse of the decision boundary of the self-learning classifier. Meanwhile, existing algorithms lack a protective control mechanism to safely retreat the model to classical physical main link computation when unlabeled concept drift is detected, which cannot ensure the reliability of diagnostic conclusions and system robustness under abnormal drift conditions. Summary of the Invention

[0006] The present invention aims to solve at least one of the technical problems existing in related technologies. To this end, the present invention provides a method and system for diagnosing the sealing of power distribution cabinets based on residual neural networks.

[0007] The first technical solution provided by this invention is as follows.

[0008] A method for diagnosing the sealing of a distribution cabinet based on a residual neural network includes the following steps: S1, obtain the equivalent diagnostic volume and background leakage rate compensation term of the distribution cabinet; S2, according to the preset coding sequence, perform multi-path time-division pneumatic excitation on the power distribution cabinet; S3, acquire the pressure response waveform in the distribution cabinet under the time-division pneumatic excitation of the multi-path, and extract the pressure attenuation feature from the pressure response waveform; S4, input the equivalent diagnostic volume, the background leakage rate compensation term and the pressure attenuation characteristic into the preset physical main link calculation model to calculate the physical leakage equivalent diameter; S5, input the pressure attenuation feature into the residual neural network, and output the residual correction amount for the simplification error of the physical model through the residual neural network; S6. Combining the physical leakage equivalent diameter and the residual correction amount, the final leakage diagnosis result of the distribution cabinet is obtained by fusion calculation.

[0009] Furthermore, in S1, the equivalent diagnostic volume is obtained as follows: Obtain the geometric volume of the power distribution cabinet casing; Obtain the total volume of the internal physical components of the power distribution cabinet from its design documents; The difference between the geometric volume and the sum of the volumes is the equivalent diagnostic volume; The process of obtaining the background leakage rate compensation term includes: Monitor the pressure drift within the distribution cabinet when the multi-path time-division pneumatic excitation is not executed; Identify the static leakage rate caused by the inherent gaps in the cabinet and label it as the background leakage rate compensation item.

[0010] Furthermore, in S2, the multi-path time-division aerodynamic excitation is achieved in the following manner: Control at least two air inlets on the power distribution cabinet to open or close sequentially at set time intervals; The pressure response waveform is constructed by generating aerodynamic pressure responses with spatial differences using air inlets at different locations.

[0011] Furthermore, in S3, the step of extracting the pressure attenuation feature includes: Real-time monitoring of the pressure attenuation rate within the power distribution cabinet; The sampling frequency of the pressure sensor is dynamically adjusted according to the pressure attenuation rate. When the pressure decay rate falls below the threshold, causing the sampling frequency to switch to the lower limit sampling rate, the sparse sampling data is supplemented by an interpolation algorithm.

[0012] Furthermore, in S5, the parameters input to the residual neural network also include: the index number of the currently open air inlet, and the real-time temperature data inside the power distribution cabinet.

[0013] Further, in step S6, the fusion calculation is obtained using the following formula: in, The leakage equivalent diameter is the final output of the system and is used for failure alarm classification. The physical leakage equivalent diameter, This refers to the equivalent diameter nonlinear residual correction value output by the residual neural network surrogate model. The fusion gating coefficients are used to correct the neural network.

[0014] Furthermore, it also includes a step of reliability monitoring of the residual neural network: The maximum mean difference algorithm is used to measure the statistical distance between the distribution of pressure decay features extracted from S3 and the distribution of the benchmark sample set. When the statistical distance exceeds the preset safety threshold, it is determined that concept drift has occurred, the fusion gating coefficient of the neural network correction is set to zero, the safety backoff mechanism is executed, and the physical leakage equivalent diameter is output as the final diagnostic result.

[0015] Furthermore, it also includes leak location steps: Based on the pressure attenuation features extracted in S3, and combined with the prior probability of the geometric distribution of the internal structure of the distribution cabinet, the fusion likelihood probability of each candidate leakage region is calculated. The prior probability of the geometric distribution is obtained by calculating the spatial geometric distance of each candidate leak location relative to the air inlet and the sensor. The leak location step also includes: Calculate the absolute difference between the maximum probability value and the second largest probability value in the fusion likelihood probability distribution; When the absolute difference is lower than the preset positioning threshold, the output of three-dimensional spatial coordinates stops, and the system degenerates into only outputting the two candidate fault regions with the highest probability.

[0016] Based on the first technical solution, a second technical solution was provided.

[0017] A distribution cabinet sealing diagnostic system based on residual neural networks executes a distribution cabinet sealing diagnostic method based on residual neural networks, including: The parameter calibration module is used to obtain the equivalent diagnostic volume and background leak rate compensation term; A multi-path excitation mechanism is used to perform time-division aerodynamic excitation according to a preset sequence; The intelligent acquisition module is used to acquire the response waveform and extract pressure attenuation characteristics; The physical calculation module is used to calculate the physical leakage equivalent diameter using the parameters obtained by the parameter calibration module; The residual correction module is used to process the pressure attenuation characteristics through a residual neural network to output a residual correction amount; The fusion diagnostic module is used to combine the outputs of the physical calculation module and the residual correction module to generate the final diagnostic result.

[0018] Furthermore, the system also includes a drift monitoring module; The drift monitoring module calculates the statistical distance between the real-time sampling distribution and the benchmark distribution. When an abnormal distribution is detected, the module sets the output gating coefficient of the neural network in the fusion diagnostic module to zero. The multi-path excitation mechanism includes multiple electromagnetic air intake valves disposed on different sides of the cabinet, and each electromagnetic air intake valve is controlled by a preset coding sequence. The intelligent acquisition module includes a high-precision differential pressure sensor and an edge processing unit. The edge processing unit dynamically adjusts the sampling step size according to the real-time rate of change of pressure. The physical calculation module has a built-in database of physical occupancy volumes associated with the internal CAD model of the power distribution cabinet. The fusion diagnostic module also includes a virtual sensor localization submodule, which is used to fuse aerodynamic features and geometric prior information to locate the leak source; It also includes an alarm module, which is used to issue an audible and visual alarm and display potential leakage areas when the final diagnostic result exceeds a safety threshold; The residual neural network adopts a deep residual architecture, and its training set is constructed based on various typical leakage conditions of the power distribution cabinet. The parameter calibration module has a temperature drift compensation function, which is used to correct the static zero-point drift of the sensor in real time. The system uploads diagnostic results to the substation monitoring backend via a wireless or wired communication interface.

[0019] The above-described one or more technical solutions in the embodiments of the present invention have at least one of the following technical effects: This invention breaks away from the idealized assumption of treating the cabinet as a "cavity" in the traditional pressure drop method. It compensates for the volume occupied by internal components through CAD geometric correction and introduces an algebraic compensation term for background leakage rate. Combined with residual neural network to correct residual errors in the simplification of the physical model, it effectively solves the calculation bias caused by nonlinear turbulence.

[0020] To address the potential "decision collapse" issue that neural networks may experience under unknown data distributions, this solution uniquely introduces concept drift detection based on maximum mean difference (MMD). When drastic environmental changes occur or sensors become unreliable due to aging, the system can forcibly retreat to interpretable physical main link calculations, ensuring the baseline security of power equipment monitoring.

[0021] By employing multipath-coded aerodynamic excitation technology, an asymmetric aerodynamic response characteristic is artificially constructed. This allows the system to determine the spatial location of the leak source without relying on expensive ultrasonic sensor arrays or acoustic cameras, significantly reducing hardware deployment costs and installation complexity.

[0022] Additional aspects and advantages of the invention will be set forth in part in the description which follows, and in part will be obvious from the description, or may be learned by practice of the invention. Attached Figure Description

[0023] To more clearly illustrate the technical solutions in this invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of this invention. For those skilled in the art, other drawings can be obtained from these drawings without creative effort.

[0024] Figure 1 This is a flowchart of the main steps of a power distribution cabinet sealing diagnosis method based on residual neural networks according to the present invention. Detailed Implementation

[0025] To make the objectives, technical solutions, and advantages of this invention clearer, the technical solutions of this invention will be clearly and completely described below. Obviously, the described embodiments are only some, not all, of the embodiments of this invention. Based on the embodiments of this invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this invention. The following embodiments are used to illustrate this invention but cannot be used to limit the scope of this invention.

[0026] Figure 1 The main steps of a power distribution cabinet sealing diagnosis method based on residual neural networks according to the present invention are shown.

[0027] The specific technical solution of the present invention is described in detail below.

[0028] A1: The edge computing unit first performs a closed-loop self-test on the diagnostic hardware, including the cabinet, micro positive pressure excitation components, individual differential pressure sensors, temperature sensors, and absolute pressure sensors, while the power distribution cabinet is in a closed state.

[0029] If the cabinet door is not fully closed, the pressure relief channel feedback is abnormal, or the status of the critical sealing mechanism is unknown, the system will immediately issue a hardware abnormality alarm and terminate the diagnostic process.

[0030] Assuming the self-test passes, the edge computing unit collects and records the static output of the differential pressure sensor on the time axis when no micro-positive pressure excitation is applied, and calculates its average fluctuation value as the zero-point drift correction value to implement dynamic zero-point elimination in subsequent measurements. Simultaneously, the gas rate lost under standard differential pressure from minute seams in the cabinet structure that cannot be completely sealed due to process errors or aging is calibrated as the system's inherent background leakage rate compensation item. This is used for algebraic deduction in subsequent physical flow calculations.

[0031] A2: The branch control module sends control commands to the micro-positive pressure excitation component according to the preset safety pressure upper limit, controlling the micro-positive pressure excitation component to sequentially inject time-division coded micro-positive pressure gas into the distribution cabinet through at least two different air inlets. Because the distribution cabinet is densely packed with components such as circuit breakers, busbars, and transformers, an asymmetric physical flow impedance network is formed. When transient pressure pneumatic waves are injected from different air inlets, even if the leak point is located at the same position, the pressure wave will propagate to the single differential pressure sensor measuring point through different flow impedance paths, resulting in propagation delays, transient rise edge curvature distortion, and reflected wave superposition in the time domain with differences ranging from microseconds to milliseconds. This greatly enriches the geometric characteristics of the transient pressure response and improves the distinguishability and spatial observability of a single pressure measuring point for spatial leak locations.

[0032] Before formal inflation, the system first performs a safety check by feedback from the pressure relief valve position, and applies a short transient purging pressure wave lasting less than 0.5 seconds to monitor whether the response slope of the pressure sensor is within the preset pipeline unobstructed range, in order to rule out pipeline bends or physical blockages of the sensor.

[0033] A3: Within the transient pressure response window triggered by the inflation of each intake branch, a single differential pressure sensor, absolute pressure sensor, and temperature sensor simultaneously collect the pressure difference, absolute pressure, and absolute temperature inside the cabinet.

[0034] The edge computing unit dynamically adjusts the physical sampling frequency based on the calculated differential pressure decay rate, ambient temperature fluctuation amplitude, and the sensor's inherent noise floor. During the high decay phase of rapid differential pressure decrease, the sampling frequency is adaptively increased to the upper limit to fully capture the rising edge curvature and dynamic overshoot of the transient pressure waveform. During the slow decay phase of decreasing differential pressure, the sampling frequency is adaptively reduced to the lower limit to reduce the storage and computational load at the edge. In the low sampling rate phase, if the sparse pressure response sequence is resampled or morphologically interpolated to unify the feature dimensions of the residual neural network input, the virtual signal values ​​generated by the interpolation are only temporarily stored in the feature alignment module in memory. It is strictly prohibited to write these values ​​as real sampling points into the underlying historical physical database to prevent the introduction of artificially created high-frequency numerical noise during interpolation calculations.

[0035] A4: Pressure attenuation calculation module is based on the equivalent diagnostic volume of the distribution cabinet. The absolute pressure, absolute temperature, and test time span are collected, and the pressure decay leakage rate is calculated using the ideal gas law. The equivalent diagnostic volume... The determination method is as follows: The total internal cavity volume of the distribution cabinet is obtained through a 3D geometric CAD assembly model. The total physical volume of the circuit breakers, busbars, transformers, and secondary relay protection equipment compartments installed inside the cabinet is then precisely deducted to obtain the basic air volume. This basic value is then calibrated and corrected using an offline inflation and deflation standard cycle experiment. The leakage equivalent diameter is then calculated. At that time, the pressure decay leakage rate is deducted from the system's inherent background leakage rate compensation item recorded in step S1 using the algebraic subtraction method. To obtain the net leakage rate of the leak after eliminating background interference such as cabinet door seams, it is then substituted into the equivalent orifice flow rate model containing a complete square root operator for inversion calculation, where the equivalent orifice flow rate coefficient is... Dynamic interpolation updates are performed based on the Reynolds number and real-time differential pressure.

[0036] A5: The residual neural network (physically serving as a proxy residual correction model for the physical main link of the equivalent orifice flow, used specifically to fit and compensate for theoretical simplification errors caused by neglecting local turbulence and temperature and pressure unsteady-state conduction in the classical steady-state one-dimensional flow model) receives short-time window pressure difference features extracted by the pressure decay window and outputs the equivalent diameter correction. And model credibility.

[0037] Before input, the short-window feature needs to be constrained by the data signal-to-noise ratio quality label calculated by weighting the real-time temperature and humidity fluctuations of the environment, the static zero-point jitter of the absolute pressure sensor, and the high-frequency electromagnetic interference noise level on site, so as to automatically suppress unqualified noisy input features. After the residual neural network is deployed, its online incremental learning training labels are only allowed to come from the physical calibration of external standard leaks on site, the true value of offline airtightness test, or the maintenance records confirmed by manual verification. It is prohibited to use the predicted value output by the network itself as the true value label for self-training, so as to eliminate the risk of self-reinforcement collapse due to "confirmation bias" under unsupervised drift.

[0038] A6: The virtual sensor proxy localization algorithm takes the transient pressure response characteristics collected under multiple air intake branches as input, calculates the likelihood probability of candidate leakage areas, and integrates the prior probability of the spatial geometric distribution of the overlapping gaps in the internal structure of the distribution cabinet, and finally outputs the spatial fusion probability of each candidate leakage area.

[0039] The transient pressure response characteristics specifically include the transient pressure rise time delay constant, transient overshoot peak attenuation rate, and frequency response impedance characteristics generated when time-division encoded positive pressure pulses are sequentially injected from different air inlets and propagate to the differential pressure sensor measuring point in the internal asymmetric air resistance channel.

[0040] 1. Time delay constant of transient pressure rise time Meaning: This feature reflects the mapping relationship between the propagation speed of the pneumatic excitation signal and the physical distance in the complex flow resistance network inside the distribution cabinet. Due to the asymmetrical layout of components inside the distribution cabinet, the path length and obstacles (resistance) of the gas injected from different air inlets to the sensor measurement point are different, resulting in differences in the arrival time of the pressure wave.

[0041] How to obtain: Record the moment the solenoid valve opens. ; The sensor detects the point at which the pressure begins to deviate from the baseline, or when the first peak value is reached. The moment ; Calculate the difference between the two. That is, the time delay constant.

[0042] Function: It serves as the primary basis for locating the source of a leak, and is used to distinguish from which excitation path the gas leaked or propagated.

[0043] 2. Transient overshoot peak decay rate Meaning: When a high-pressure gas pulse suddenly enters a relatively sealed enclosure, the pressure will instantaneously exceed the static equilibrium value due to the momentum effect and compressibility of the gas (resulting in "overshoot"). This characteristic describes the speed at which the pressure drops from its instantaneous peak value back to the quasi-steady-state decay curve.

[0044] How to obtain: Identify the highest pressure point of instantaneous fluctuation in the pressure waveform. and their corresponding times ; Identify the point where the overshoot fluctuation ends and the normal pressure drop curve begins. and their corresponding times ; Calculate the attenuation coefficient .

[0045] Function: This feature relates to the free space volume (equivalent volume) inside the cabinet. The geometry of the leakage hole is highly correlated with the residual, and is a key input for residual neural networks to perform residual correction.

[0046] 3. Frequency response impedance characteristics Technical implications: The internal space of the distribution cabinet is considered as a low-pressure pneumatic filter network. The air inlet is the input end, and the sensor is the output end. The geometry of the distribution cabinet, the obstruction of components, and the location of leaks collectively determine the "aerodynamic impedance" of the system, that is, the attenuation and phase shift characteristics of aerodynamic disturbances of different frequencies.

[0047] How to obtain: The frequency domain features are obtained by performing a Fast Fourier Transform (FFT) on the acquired transient pressure response time-domain signal. Extract the amplitude-frequency characteristics (gain attenuation of different frequency components) and phase-frequency characteristics (phase delay of different frequency components). Calculate the amplitude ratio at a specific frequency point (such as the low-frequency resonance point) as the impedance characteristic vector.

[0048] Transient pressure response time-domain signal refers to the original pressure change sequence data formed by the differential pressure sensor arranged in the distribution cabinet continuously collecting data according to the adaptive sampling step size during the time-division aerodynamic excitation of multiple paths and recording it on the time axis.

[0049] Function: Used to describe the "pneumatic fingerprint" of the internal structure of the power distribution cabinet, providing a more robust diagnostic basis when the time-domain characteristics are not obvious due to environmental interference.

[0050] The prior probability of spatial geometric distribution is determined by the reciprocal normalization of the three-dimensional spatial geometric distances of each preset leak-prone candidate area, such as the cabinet door sealing strip, cable threading sealing hole, top exhaust valve, and side panel splicing seam, relative to each air inlet and the single differential pressure sensor in the three-dimensional model inside the distribution cabinet. The positioning algorithm completely shields and does not input any measured acoustic, ultrasonic, or vibration signals.

[0051] When the difference between the calculated maximum fusion probability and the second-highest fusion probability is lower than the set location confidence threshold, the edge computing unit forcibly shuts down the output of the unique three-dimensional coordinates of the leakage source, degenerating to only outputting the top two candidate fault areas with the highest probability and manual flaw detection verification prompts, to prevent over-localization under single sensor configuration.

[0052] A7: The drift detection module built into the edge computing unit uses maximum mean difference (MMD) to unsupervisedly measure the statistical distance between the online feature distribution collected in real time within the current sliding window and the factory baseline feature distribution.

[0053] When the measurement value If the set drift threshold is exceeded, and the edge computing unit fails to detect any externally input physical standard tags (such as standard leak retest tags, manual on-site verification tags, etc.) within a preset monitoring sliding cycle, the system determines that a tagless concept drift has occurred. At this time, the edge computing unit immediately initiates protective safety backoff control, forcibly reducing the fusion gating coefficient of the residual neural network correction. Setting it to zero completely degrades the final output leakage equivalent diameter to the physical main link calculation value of the equivalent orifice. At the same time, it outputs to the terminal the algorithm's online self-updating temporary inactivation warning, the current highest probability candidate region, and a manual calibration and verification prompt.

[0054] If no concept drift is detected, the gating coefficient is dynamically adjusted using model credibility. The physical calculation equivalent diameter is weighted and fused with the neural network correction value to output the final leakage equivalent diameter. It also outputs the corresponding sealing failure warning level based on the positioning results.

[0055] The technical solution will be further illustrated below through specific embodiments.

[0056] Example 1: Before the diagnostic process begins, the edge computing unit first reads the closure status of the distribution cabinet door closure sensor, active ventilation valve, and spare cable hole. The micro-positive pressure excitation component sequentially controls the conduction of different solenoid valve branches, applying coded micro-positive pressure pulses into the cabinet.

[0057] The differential pressure decay rate is calculated using the following formula: Equation (1) in, For the first The rate of pressure drop decay under each window ( ), and These are the pressure difference values ​​measured by the differential pressure sensor at the current moment and the previous comparison moment, respectively. ), For a defined comparison time interval ( ).

[0058] The sampling rate is adjusted adaptively with a bounded margin based on the differential pressure decay rate, and the adjustment formula is as follows: Equation (2) in, The calculated adaptive transient sampling rate ( ), and These are the minimum and maximum sampling frequencies allowed by the differential pressure sensor and edge acquisition module hardware, respectively. ), The sampling rate is adaptively adjusted based on the offline calibration dataset. .

[0059] When the lower sampling rate is activated, the decay curve appears as a sparse time series. In order to unify the input dimension of the residual neural network, only the geometric shape of the decay curve is interpolated and aligned. Furthermore, the virtual data points generated by this interpolation are only allowed to reside in the feature alignment module in memory. It is absolutely forbidden to write the non-real sampling points generated by the interpolation into the underlying historical physical database to prevent the introduction of fake numerical noise by the interpolation calculation.

[0060] Pressure decay leakage rate is calculated using the following formula: Equation (3) in, For pressure decay leakage rate (under standard conditions, ), and These are the absolute temperatures under standard conditions ( ) and standard atmospheric pressure ( ); and The absolute pressure measured by the absolute pressure sensor inside the distribution cabinet at the previous and next times ( ), and These are the absolute temperatures sampled by the temperature sensor at the corresponding times ( ).

[0061] in the formula For the equivalent diagnostic cavity volume of the distribution cabinet ( The method for determining this is as follows: by retrieving the 3D CAD general drawing of the distribution cabinet, calculate the total internal cavity volume of the distribution cabinet shell. It also accurately calculates the total volume occupied by the internally installed circuit breaker body, busbar body, current / voltage transformer body, and secondary relay protection compartment. ,Depend on The baseline value was calculated, and then calibrated and corrected by combining it with the offline standard inflation cycle release residual calibration experiment.

[0062] The leakage equivalent diameter is calculated using the following formula: Equation (4) in, The physical leakage equivalent diameter derived from the physical formula ( ), The average value of the static output fluctuation of the differential pressure sensor in the zero-excitation state in step S1 is combined with the system's inherent background leakage rate compensation term measured by the airtightness calibration experiment of the non-sealed cabinet structure mechanical joints. ); The equivalent orifice flow coefficient obtained by interpolation calibration through offline standard calibration orifice experiments is used in diagnostics as a dynamic adaptive coefficient with respect to real-time differential pressure and Reynolds number, rather than a fixed constant, in order to overcome the systematic distortion of the one-dimensional simplified flow model under multi-resistance networks. The calculated pressure difference across the leakage orifice ( Here, the average dynamic pressure difference between the inside and outside of the cabinet, measured by the differential pressure sensor, is used to determine the value. The physical density of the internal humid air at the current temperature and absolute pressure ( ).

[0063] Example 2: The residual neural network (or surrogate neural network) receives short-time window features generated by the pressure decay window, including , Temperature compensation item, intake branch number, and curve morphology characteristics.

[0064] Network output equivalent diameter correction And model credibility, in which the residual neural network is essentially a proxy model for the main link of the classical equivalent orifice physical calculation, specifically used to fit and compensate for nonlinear residual errors under unsteady aerodynamic transmission in multi-physics fields.

[0065] Virtual sensor proxy model with multi-intake branch response characteristics Using the input, calculate candidate leakage regions. The likelihood probability. Here, the response features The physical dimensions specifically include the time delay constant of the pressure rise time generated when time-division encoded positive pressure pulses are sequentially injected from different air inlets, propagating through the flow resistance channels of the internal irregular components to the measuring point of a single differential pressure sensor. The transient pressure overshoot peak attenuation rate and the air resistance frequency response characteristics after pseudo-random coding demodulation are used to construct anisotropic transient pressure transmission maps under different intake paths.

[0066] The likelihood function of the candidate region is calculated using the following formula: Equation (5) in, To obtain the characteristic observations of multiple air intakes Next, the Candidate leak areas The probability likelihood, The total number of candidate areas pre-divided based on the locations of potential leaks in the distribution cabinet; The observed feature vector With the The Mahalanobis distance between candidate regions and the feature vectors of the historical calibration dictionary, or the classification residual distance output by the surrogate model.

[0067] The denominator in the formula is for all to The index terms corresponding to each candidate region are summed.

[0068] After incorporating the prior probabilities of spatial geometric distribution, the fusion probability of each candidate leakage region is calculated using the following formula: Equation (6) in, For the first The final spatial fusion probability of each candidate region; The prior probability of the spatial geometric distribution of the power distribution cabinet is determined by: obtaining the three-dimensional spatial geometric distance of each candidate leakage point (cabinet door sealing strip, cable threading sealing hole, top exhaust valve, side panel splice seam) relative to each air inlet and differential pressure sensor through the three-dimensional CAD design file inside the power distribution cabinet; normalizing the weighted average of the inverse of the geometric distance as the prior probability of the structural geometric space; thereby completely shielding the input end and eliminating dependence on the measured signals of any physical sound wave, vibration or ultrasonic sensor.

[0069] When the absolute difference between the probability of the largest candidate region and the probability of the second largest candidate region in the calculated probability distribution is lower than the preset positioning threshold. When the edge computing unit determines that the spatial observability is limited due to the single measurement point configuration, the system automatically shuts down the output of the unique three-dimensional spatial coordinates of the leakage source, and forces it to degenerate into only outputting the two candidate fault areas with the highest probability and providing a verification prompt, in order to prevent over-localization under the single sensor configuration.

[0070] Example 3: Concept drift detection can be measured using the kernel-based unsupervised maximum mean difference (MMD), calculated as follows: Equation (7) in, For the reason A set of source domain reference samples collected under factory standard calibration or offline health airtightness test, excluding leakage conditions; For those collected online within the current sliding window A set of feature samples for each target domain; is a Gaussian kernel function used to project the sample distribution onto the regenerated Hilbert space (RKHS) to calculate the statistical distance between the centroids of the distributions.

[0071] The final leakage equivalent diameter is output by linearly fusing the physical main link calculation value and the neural network residual correction value, as shown in the following formula: Equation (8) in, The leakage equivalent diameter is the final output of the system and is used for failure alarm classification. ), The physical leakage equivalent diameter calculated according to equation (4) is ( ), The equivalent diameter nonlinear residual correction amount output by the residual neural network surrogate model ( ), The fusion gate coefficients for neural network corrections (dimensionless, ranging from 0 to 1). When the drift metric value Below the preset drift warning threshold At that time, the gating coefficient The activation is maintained, and its magnitude is determined by the model confidence level output by the neural network; the model confidence level is a value between 0 and 1, used to dynamically adjust the gating coefficient. .

[0072] When the drift metric value Greater than or equal to the preset drift warning threshold When the system determines that a conceptual drift has occurred due to large-scale temperature and humidity fluctuations, electromagnetic high-frequency coupling interference, or sensor aging, it first searches the external input interface. If, within the current detection sliding cycle, no externally input physical standard leak retest tag, maintenance personnel infrared / acoustic imaging on-site measurement confirmation tag, or maintenance record physical confirmation tag is detected, the system is determined to be in a tagless protective drift state. To completely eliminate the risk of self-training online crashes and cascaded prediction errors caused by "confirmation bias" in the deep learning model, the edge computing unit initiates protective safety backoff control, forcibly adjusting the gating coefficient. When set to zero, the final leakage equivalent diameter completely degenerates into the physical calculation value of the main hydrodynamic link. At the same time, it outputs a warning of temporary inactivation of the model self-learning function to the operation and maintenance terminal, outputs the candidate fault area with the highest probability, and generates maintenance prompts that suggest manual on-site calibration and manual input confirmation labels.

[0073] Example 4 The diagnostic system that implements the diagnostic method of the present invention is given below: The system includes: The parameter calibration module is used to execute S1 to obtain the equivalent diagnostic volume and background leak rate compensation term. The parameter calibration module has a temperature drift compensation function, which is used to correct the static zero drift of the sensor in real time.

[0074] A multi-path excitation mechanism is used to execute S2, which performs time-division pneumatic excitation according to a preset sequence. The multi-path excitation mechanism includes multiple electromagnetic air intake valves disposed on different sides of the cabinet, and each of the electromagnetic air intake valves is controlled by a preset coded timing sequence.

[0075] The intelligent acquisition module is used to execute S3, acquire the response waveform and extract the pressure decay characteristics; the intelligent acquisition module includes a high-precision differential pressure sensor and an edge processing unit, which dynamically adjusts the sampling step size according to the real-time rate of change of pressure.

[0076] The physical calculation module is used to execute S4 and calculate the physical leakage equivalent diameter using the parameters obtained by the parameter calibration module; the physical calculation module has a built-in database of physical occupancy volumes associated with the internal CAD model of the distribution cabinet.

[0077] The residual correction module is used to execute S5, which processes the pressure decay characteristics through a residual neural network to output the residual correction amount; the residual neural network adopts a deep residual architecture, and its training set is constructed based on various typical leakage conditions of the distribution cabinet.

[0078] The fusion diagnostic module is used to execute S6, combining the outputs of the physical calculation module and the residual correction module to generate the final diagnostic result.

[0079] The fusion diagnostic module also includes a virtual sensor positioning submodule, which is used to fuse aerodynamic features and geometric prior information to locate the leak source; and a drift monitoring module, which calculates the statistical distance between the real-time sampling distribution and the reference distribution, and sets the gating coefficient of the neural network output in the fusion diagnostic module to zero when an abnormal distribution is detected.

[0080] An alarm module is used to issue an audible and visual alarm and display potential leak areas when the final diagnostic result exceeds a safety threshold.

[0081] The system uploads diagnostic results to the substation monitoring backend via wireless or wired communication interfaces.

[0082] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method for diagnosing the sealing of a power distribution cabinet based on a residual neural network, characterized in that, Includes the following steps: S1, obtain the equivalent diagnostic volume and background leakage rate compensation term of the distribution cabinet; S2, according to the preset coding sequence, perform multi-path time-division pneumatic excitation on the power distribution cabinet; S3, acquire the pressure response waveform in the distribution cabinet under the time-division pneumatic excitation of the multi-path, and extract the pressure attenuation feature from the pressure response waveform; S4, input the equivalent diagnostic volume, the background leakage rate compensation term and the pressure attenuation characteristic into the preset physical main link calculation model to calculate the physical leakage equivalent diameter; S5, input the pressure attenuation feature into the residual neural network, and output the residual correction amount for the simplification error of the physical model through the residual neural network; S6. Combining the physical leakage equivalent diameter and the residual correction amount, the final leakage diagnosis result of the distribution cabinet is obtained by fusion calculation; In S2, the time-division aerodynamic excitation of the multipath is achieved in the following way: Control at least two air inlets on the power distribution cabinet to open or close sequentially at set time intervals; The pressure response waveform is constructed by generating aerodynamic pressure responses with spatial differences using air inlets at different locations. It also includes leak location steps: Based on the pressure attenuation features extracted in S3, and combined with the prior probability of the geometric distribution of the internal structure of the distribution cabinet, the fusion likelihood probability of each candidate leakage region is calculated. The prior probability of the geometric distribution is obtained by calculating the spatial geometric distance of each candidate leak location relative to the air inlet and the sensor. The leak location step also includes: Calculate the absolute difference between the maximum probability value and the second largest probability value in the fusion likelihood probability distribution; When the absolute difference is lower than the preset positioning threshold, the output of three-dimensional spatial coordinates stops, and the system degenerates into only outputting the two candidate fault regions with the highest probability.

2. The method for sealing diagnosis of distribution cabinets based on residual neural networks according to claim 1, characterized in that, In S1, the equivalent diagnostic volume is obtained as follows: Obtain the geometric volume of the power distribution cabinet casing; Obtain the total volume of the internal physical components of the power distribution cabinet from its design documents; The difference between the geometric volume and the sum of the volumes is the equivalent diagnostic volume; The process of obtaining the background leakage rate compensation term includes: Monitor the pressure drift within the distribution cabinet when the multi-path time-division pneumatic excitation is not executed; Identify the static leakage rate caused by the inherent gaps in the cabinet and label it as the background leakage rate compensation item.

3. The method for sealing diagnosis of distribution cabinets based on residual neural networks according to claim 1, characterized in that, In S3, the step of extracting the pressure attenuation feature includes: Real-time monitoring of the pressure attenuation rate within the power distribution cabinet; The sampling frequency of the pressure sensor is dynamically adjusted according to the pressure attenuation rate. When the pressure decay rate falls below the threshold, causing the sampling frequency to switch to the lower limit sampling rate, the sparse sampling data is supplemented by an interpolation algorithm.

4. The method for sealing diagnosis of distribution cabinets based on residual neural networks according to claim 1, characterized in that, In S5, the parameters input to the residual neural network also include: the index number of the currently open air inlet, and the real-time temperature data inside the power distribution cabinet.

5. The method for sealing diagnosis of distribution cabinets based on residual neural networks according to claim 1, characterized in that, In step S6, the fusion calculation is obtained using the following formula: in, The leakage equivalent diameter is the final output of the system and is used for failure alarm classification. The physical leakage equivalent diameter, This refers to the equivalent diameter nonlinear residual correction value output by the residual neural network surrogate model. The fusion gating coefficients are used to correct the neural network.

6. The method for sealing diagnosis of distribution cabinets based on residual neural networks according to claim 5, characterized in that, It also includes the step of monitoring the reliability of the residual neural network: The maximum mean difference algorithm is used to measure the statistical distance between the distribution of pressure decay features extracted from S3 and the distribution of the benchmark sample set. When the statistical distance exceeds the preset safety threshold, it is determined that concept drift has occurred, the fusion gating coefficient of the neural network correction is set to zero, the safety backoff mechanism is executed, and the physical leakage equivalent diameter is output as the final diagnostic result.

7. A distribution cabinet sealing diagnosis system based on residual neural networks, comprising performing the distribution cabinet sealing diagnosis method based on residual neural networks as described in any one of claims 1-6, characterized in that, include: The parameter calibration module is used to obtain the equivalent diagnostic volume and background leak rate compensation term; A multi-path excitation mechanism is used to perform time-division aerodynamic excitation according to a preset sequence; The intelligent acquisition module is used to acquire the response waveform and extract pressure attenuation characteristics; The physical calculation module is used to calculate the physical leakage equivalent diameter using the parameters obtained by the parameter calibration module; The residual correction module is used to process the pressure attenuation characteristics through a residual neural network to output a residual correction amount; The fusion diagnostic module is used to combine the outputs of the physical calculation module and the residual correction module to generate the final diagnostic result.

8. The distribution cabinet sealing diagnostic system based on residual neural network according to claim 7, characterized in that, The system also includes a drift monitoring module; The drift monitoring module calculates the statistical distance between the real-time sampling distribution and the benchmark distribution. When an abnormal distribution is detected, the module sets the output gating coefficient of the neural network in the fusion diagnostic module to zero. The multi-path excitation mechanism includes multiple electromagnetic air intake valves disposed on different sides of the cabinet, and each electromagnetic air intake valve is controlled by a preset coding sequence. The intelligent acquisition module includes a high-precision differential pressure sensor and an edge processing unit. The edge processing unit dynamically adjusts the sampling step size according to the real-time rate of change of pressure. The physical calculation module has a built-in database of physical occupancy volumes associated with the internal CAD model of the power distribution cabinet. The fusion diagnostic module also includes a virtual sensor localization submodule, which is used to fuse aerodynamic features and geometric prior information to locate the leak source; It also includes an alarm module, which is used to issue an audible and visual alarm and display potential leakage areas when the final diagnostic result exceeds a safety threshold; The residual neural network adopts a deep residual architecture, and its training set is constructed based on various typical leakage conditions of the power distribution cabinet. The parameter calibration module has a temperature drift compensation function, which is used to correct the static zero-point drift of the sensor in real time. The system uploads diagnostic results to the substation monitoring backend via a wireless or wired communication interface.

Images