Modular intelligent power distribution box heat hazard multi-mode real-time early warning method

Abstract

This invention discloses a multi-mode real-time early warning method for thermal hazards in modular intelligent distribution boxes, specifically relating to the field of intelligent power distribution thermal monitoring. It addresses the temporal misalignment and feature distortion of heterogeneous thermal data under unsteady-state impacts. The method extracts the unsteady-state load mutation timescale by collecting high-frequency transient current signals from the incoming line of the intelligent distribution box. Using this timescale, it simultaneously acquires the non-contact infrared surface temperature discrete sequence and the contact-type internal thermistor continuous curve of the target monitoring area. A thermal inertia compensation function is then constructed, and forward phase differential reconstruction is performed on the contact-type internal thermistor continuous curve to obtain a reference virtual internal thermal radiation sequence. A spatiotemporal decoupling cross-mapping matrix is ​​then constructed to extract the transient thermophysical tearing feature tensor. Simultaneously, the isothermal envelope boundary contraction rate and the Fourier second peak difference of the thermal waveform are extracted. Dynamic decision coefficients are generated using an isolated forest anomaly detection model, and threshold judgment is completed, outputting an anti-breakdown early warning action command.

Description

Technical Field

[0001] This invention relates to the field of intelligent power distribution thermal monitoring, and more specifically, to a multi-mode real-time early warning method for thermal hazards in modular intelligent power distribution boxes. Background Technology

[0002] In the smart grid architecture, smart distribution boxes, as core nodes, are crucial for accurate early warning of thermal hazards. Existing technologies often employ concurrent sampling using infrared thermal imaging visual sensors and contact thermistors to construct multimodal thermal early warning models. Under normal, stable operating conditions, this data fusion scheme can effectively assess the thermal health status of equipment. However, actual distribution substations frequently face unsteady-state shock conditions such as heavy-load startup of high-power motors, leading to drastic transient temperature rises in critical current-carrying components such as internal busbar contacts. This forces the monitoring system to simultaneously acquire heterogeneous multi-source signals within a rapidly changing unsteady electromagnetic thermal flux field.

[0003] For the aforementioned high-frequency, unsteady-state shock scenarios, existing solutions face severe problems of heterogeneous data temporal misalignment and feature distortion. Current technologies propose using dynamic time warping algorithms to align data from various sensors. However, this technology fails to overcome the physical layer decoupling challenge under transient conditions. The fundamental reason is that contact sensors, limited by the thermal inertia of the encapsulation medium, naturally exhibit smooth and lagging low-frequency physical responses; while non-contact infrared sensors, although capable of instantly capturing sudden changes in surface radiation, are limited by fixed discrete scanning periods. When encountering rapidly changing thermal shocks, the low-frequency, lagging contact response and the high-frequency discrete infrared array data inevitably undergo irreversible phase decoupling in both time and spatial mapping. During this evolution, when conventional fusion algorithms attempt to align these multidimensional tensors with vastly different physical delay characteristics, severe phase misalignment leads to misclassification of real transient thermal spikes as asynchronous high-frequency noise, which is then forcibly smoothed out, resulting in severe distortion of the feature vectors. The actual harm caused by this mechanism defect is that the system will continuously output a false stable fitting curve during the critical window period when load impact is most likely to cause insulation breakdown; when transient heat accumulation breaks through the critical point and causes a hidden arc or even a fire, the early warning system often triggers a severely delayed alarm only after the disaster has occurred, completely losing the core value of real-time early warning, which can easily lead to the risk of unplanned large-scale power outages in the entire power supply area.

[0004] To address the aforementioned problems, a technical solution is provided. Summary of the Invention

[0005] To overcome the aforementioned deficiencies of the prior art, embodiments of the present invention provide a modular intelligent distribution box thermal hazard multi-mode real-time early warning method. This method extracts the unsteady-state load mutation timescale by collecting high-frequency transient current signals from the incoming line of the intelligent distribution box. Using this timescale, it simultaneously acquires the non-contact infrared surface temperature discrete sequence and the contact-type internal thermosensitive continuous curve of the target monitoring area. A thermal inertia compensation function is then constructed, and forward phase differential reconstruction is performed on the contact-type internal thermosensitive continuous curve to obtain a reference virtual internal thermal radiation sequence. A spatiotemporal decoupling cross-mapping matrix is ​​then constructed to extract the transient thermophysical tearing feature tensor. Simultaneously, the isothermal envelope boundary contraction rate and the Fourier second peak difference of the thermal waveform are extracted. Dynamic decision coefficients are generated using an isolated forest anomaly detection model, and threshold judgment is completed. Finally, an anti-breakdown early warning action command is output to solve the problems mentioned in the background art.

[0006] To achieve the above objectives, the present invention provides the following technical solution: S1: Real-time acquisition of high-frequency transient current signals at the incoming end of the intelligent distribution box, extraction of the time scale of unsteady load change, and using the time scale of unsteady load change as the hardware trigger reference, synchronous acquisition of the non-contact infrared surface temperature discrete sequence and the contact internal thermosensitive continuous curve of the target monitoring area. S2: Based on the physical mechanism of material thermal conduction hysteresis, a thermal inertia compensation function is constructed. Taking the time scale of the unsteady load change as the time zero point, the forward phase differential reconstruction is performed on the continuous curve of the contact internal thermodynamics to obtain a benchmark virtual internal thermal radiation sequence that is aligned with the time domain of the discrete sequence of the non-contact infrared surface temperature. S3: Construct a spatiotemporal decoupling cross-mapping matrix, substitute the non-contact infrared surface temperature discrete sequence and the benchmark virtual internal thermal radiation sequence into the spatiotemporal decoupling cross-mapping matrix to perform cross-modal frame-by-frame differential calculation, and extract the transient thermophysical tearing feature tensor. S4: The Otsu method is used to extract the geometric contour polygon of the highest temperature connected domain and calculate the isothermal envelope boundary shrinkage rate. Fast Fourier Transform is performed on the contact internal thermosensitive continuous curve to extract the Fourier secondary peak difference of the thermal waveform. The two are analyzed together to obtain the dynamic decision coefficient. The threshold judgment is performed by combining the maximum eigenvalue of the transient thermophysical tearing characteristic tensor and outputting the anti-breakdown warning action command.

[0007] Furthermore, in step S1, the static steady-state sample segment of the high-frequency transient current signal immediately preceding the current moment within the current detection window is taken as the baseline sample segment. Discrete wavelet transform is performed on the high-frequency transient current signal to extract high-frequency detail coefficients, and first-order difference is performed on adjacent sampling points to form a current slope sequence. When the absolute value of the high-frequency detail coefficient and the absolute value of the current slope corresponding to the current sampling point continuously exceed the judgment boundary determined by the baseline sample segment and remain at two sampling points, the sampling moment that first meets the condition is determined as the time scale of the non-steady-state load change.

[0008] Furthermore, in step S1, the outer surface projection area corresponding to the placement position of the contact thermal element is determined as the target monitoring area, and a hardware trigger pulse is output at the time scale of the unsteady load change. The same hardware trigger pulse simultaneously starts the non-contact infrared acquisition and the contact thermal acquisition. The start time of each frame of the non-contact infrared surface temperature discrete sequence and the sampling point time of each sampling point of the contact internal thermal continuous curve are synchronously bound to the time scale of the unsteady load change.

[0009] Furthermore, in step S2, the initial heating segment after the timescale of the unsteady load change is extracted from the continuous curve of the contact internal thermosensitive element. The interval where the temperature slope first continuously increases is determined as the compensation identification segment. The temperature increments of each sampling point in the compensation identification segment relative to the timescale of the unsteady load change are used to form a normalized thermal response curve. The normalized thermal response curve is fitted using a first-order lumped heat conduction model to obtain the material thermal conduction hysteresis time constant and heat transfer attenuation coefficient, and a thermal inertia compensation function is generated accordingly.

[0010] Furthermore, in step S2, along the relative time sequence of the contact-type internal thermal continuous curve, continuous sampling points are selected in the positive time direction for each sampling moment, and the forward temperature rise slope is calculated using a five-point forward differential algorithm. Then, the corresponding forward compensation factor is read according to the thermal inertia compensation function, and the forward temperature rise slope and the forward compensation factor are combined to form a phase compensation amount and superimposed on the current temperature value to generate a virtual internal thermal radiation value. Then, sampling is performed according to the frame time marker of the non-contact infrared surface temperature discrete sequence to form a reference virtual internal thermal radiation sequence.

[0011] Furthermore, in step S3, the surface temperature matrix of the target monitoring area corresponding to the non-contact infrared surface temperature discrete sequence is read frame by frame, and each frame matrix is ​​cropped into a grid matrix with a uniform row and column size; the spatial mapping origin is taken as the center of the projection of the installation point of the contact thermal element on the outer surface of the target monitoring area, and the spatial projection coefficient is calculated according to the distance of each grid unit to the spatial mapping origin to form a spatial projection matrix. Then, a time index matrix is ​​established based on the relative time mark of the non-contact infrared surface temperature discrete sequence and the reference virtual internal thermal radiation sequence, and the two are combined to form a spatiotemporal decoupling cross mapping matrix.

[0012] Furthermore, in step S3, the non-contact infrared surface temperature matrix and the reference virtual internal thermal radiation sequence frame value of the same frame number are selected by the spatiotemporal decoupling cross-mapping matrix, and the frame value is expanded into an internal thermal radiation reference matrix of the same size as the non-contact infrared surface temperature matrix; the two matrices are subtracted element by element to form an amplitude difference matrix, the gradient amplitude of the two matrices is calculated respectively and then subtracted element by element to form a gradient difference matrix, and the two types of difference results are expanded and the outer product operation is performed to generate the tear sub-matrices of each frame, which are stacked along the time dimension to form the transient thermophysical tear feature tensor.

[0013] Furthermore, in step S4, linear gray-scale mapping is performed on the surface temperature matrix of each frame of the non-contact infrared surface temperature discrete sequence to obtain a single-frame gray-scale matrix, and the Otsu method is used to complete the high gray-scale region segmentation to obtain candidate connected components; the candidate connected component with the highest average temperature in the region is determined as the highest temperature connected component, and boundary tracking and recursive broken line approximation are performed on it to form a geometric contour polygon, and then the isothermal envelope boundary shrinkage rate is generated based on the normal displacement of the geometric contour polygons of two adjacent frames and the inter-frame time interval.

[0014] Furthermore, in step S4, the contact-type internal thermosensitive continuous curve within a predetermined judgment window after the time scale of the unsteady load change is extracted. The temperature value corresponding to the time scale of the unsteady load change is used as the reference temperature to perform baseline subtraction. After applying a Hanning window to the obtained zero reference temperature rise curve, a fast Fourier transform is performed to generate a frequency domain spectral density distribution map. The local peaks of the frequency domain spectral density distribution map are scanned along the frequency axis. The local peak with the largest amplitude is determined as the main frequency peak, and the next local peak that is closest to it on its high-frequency side is determined as the first sidelobe secondary peak.

[0015] Furthermore, in step S4, a two-dimensional coordinate point is constructed with the isothermal envelope boundary contraction rate as the abscissa and the difference of the second peak value of the thermal waveform Fourier transform as the ordinate. The two-dimensional coordinate point is then input into the isolated forest anomaly detection model. The number of split layers from the two-dimensional coordinate point to the leaf node is recorded in all random decision trees, and the average path length is taken. Then, dynamic decision coefficients are generated based on the theoretical expected depth corresponding to the training sample size. Eigenvalue decomposition is performed on each time slice of the transient thermophysical tearing feature tensor, and the maximum eigenvalue is taken. Together with the dynamic decision coefficients, a comprehensive risk characterization value is formed. This value is then compared with the physical insulation breakdown critical threshold, and an anti-breakdown warning action command is output.

[0016] The technical effects and advantages of the modular intelligent distribution box thermal hazard multi-mode real-time early warning method of the present invention are as follows: This invention uses the timescale of non-steady-state load abrupt change as a unified time starting point, enabling a clear correspondence between the discrete sequence of non-contact infrared surface temperature and the continuous curve of contact internal thermometry during the acquisition stage. Subsequent processing no longer relies on extensive post-processing alignment, thereby allowing the surface thermal information and internal thermal information of the target monitoring area to enter the continuous processing chain at the same time coordinate, reducing object reference deviation caused by inconsistent sampling starting points.

[0017] At the thermal signal processing level, this invention converts the contact-type internal thermal continuous curve into a reference virtual internal thermal radiation sequence through thermal inertia compensation function and forward phase differential reconstruction. Then, by using a spatiotemporal decoupling cross-mapping matrix to perform cross-modal frame-by-frame differential, a transient thermophysical tearing feature tensor is formed, so that the misalignment distortion between heterogeneous thermal flow fields can be expressed with a unified intermediate object, and the input objects, intermediate results and subsequent calling relationships in the processing chain remain clear.

[0018] At the result determination level, this invention incorporates the isothermal envelope boundary shrinkage rate, the second peak difference of the thermal waveform Fourier transform, the dynamic decision coefficient, and the maximum eigenvalue of the transient thermophysical tearing characteristic tensor into the same decision chain. This preserves both the geometric evolution information of the surface hot spot and the frequency domain information and cross-modal misalignment information of the internal thermal waveform, making the formation path of the comprehensive risk characterization value clear. The final output anti-breakdown warning action command has a clear object source and processing basis. Attached Figure Description

[0019] Figure 1 This is a flowchart of a multi-mode real-time early warning method for thermal hazards in a modular intelligent distribution box according to the present invention; Figure 2 This is a schematic diagram showing the relationship between the monitoring objects and sensor deployment of the intelligent power distribution box of the present invention; Figure 3 This is a timing diagram of the non-steady-state load change timescale extraction and dual-channel synchronous acquisition of the present invention; Figure 4 This is a diagram showing the generation of the thermal inertia compensation and reference virtual internal thermal radiation sequence of the present invention. Figure 5 This is the generation diagram of the spatiotemporal decoupling cross-mapping matrix and the transient thermophysical tearing feature tensor of this invention; Figure 6 This is a diagram illustrating the generation of integrated risk assessment and anti-penetration warning action commands for this invention. Detailed Implementation

[0020] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0021] Please see Figure 1 This invention provides a modular intelligent distribution box thermal hazard multi-mode real-time early warning method, including: S1: Real-time acquisition of high-frequency transient current signals at the incoming end of the intelligent distribution box, extraction of the time scale of unsteady load change, and using the time scale of unsteady load change as the hardware trigger reference, synchronous acquisition of the non-contact infrared surface temperature discrete sequence and the contact internal thermosensitive continuous curve of the target monitoring area. S2: Based on the physical mechanism of material thermal conduction hysteresis, a thermal inertia compensation function is constructed. Taking the time scale of the unsteady load change as the time zero point, the forward phase differential reconstruction is performed on the continuous curve of the contact internal thermodynamics to obtain a benchmark virtual internal thermal radiation sequence that is aligned with the time domain of the discrete sequence of the non-contact infrared surface temperature. S3: Construct a spatiotemporal decoupling cross-mapping matrix, substitute the non-contact infrared surface temperature discrete sequence and the benchmark virtual internal thermal radiation sequence into the spatiotemporal decoupling cross-mapping matrix to perform cross-modal frame-by-frame differential calculation, and extract the transient thermophysical tearing feature tensor. S4: The Otsu threshold segmentation method is used to extract the geometric contour polygon of the highest temperature connected domain and calculate the isothermal envelope boundary shrinkage rate. Fast Fourier Transform (FFT) is performed on the contact-type internal thermosensitive continuous curve to extract the secondary peak difference of the thermal waveform. The isothermal envelope boundary shrinkage rate and the secondary peak difference of the thermal waveform are used to construct two-dimensional coordinate points and input into the Isolation Forest anomaly detection model to generate dynamic decision coefficients. Then, the maximum eigenvalue of the transient thermophysical tear feature tensor is combined to perform threshold judgment and output the anti-breakdown warning action command.

[0022] Specifically, this invention focuses on processing the current signal, surface temperature information, and internal thermal information generated by an intelligent distribution box under unsteady load abrupt changes. First, a high-frequency transient current signal is used to locate the timescale of the unsteady load abrupt change, and this timescale is used as a unified hardware trigger reference to organize the synchronous acquisition and unified time base calibration of the non-contact infrared surface temperature discrete sequence and the contact internal thermal continuous curve. Then, based on the material thermal conduction hysteresis physical mechanism, a thermal inertia compensation function is constructed to reconstruct the contact internal thermal continuous curve through forward phase differentiation, forming a reference virtual internal thermal radiation sequence corresponding to the time domain of the non-contact infrared surface temperature discrete sequence. On this basis, a spatiotemporal decoupling cross-mapping matrix is ​​used to transform the two types of thermal information into a unified object that can be compared frame by frame, obtaining a transient thermophysical tearing feature tensor. Furthermore, discriminant quantities are extracted from the surface hotspot contour evolution and the frequency domain distribution of the internal thermal waveform. Combined with an isolated forest anomaly detection model, dynamic decision coefficients are output, ultimately completing the comprehensive risk characterization value judgment and outputting a breakdown warning action command.

[0023] During the equipment installation and commissioning phase, a mapping table is established based on the module number, the location of the current-carrying component, the location of the contact thermistor, the target monitoring area number, and the circuit breaker control channel number. In subsequent steps, the established correspondence between the contact thermistor and the target monitoring area, as well as the hardware binding relationship of the module to which the target monitoring area belongs, are all called according to this mapping table. In this embodiment, the predetermined judgment window starts with the time scale of the unsteady load change, and extends by one infrared frame interval as the later of the time when the first local temperature rise peak appears on the continuous curve of the contact internal thermistor and the time when the area of ​​the highest temperature connected domain in the discrete sequence of the non-contact infrared surface temperature reaches its maximum value.

[0024] When a smart distribution box is subjected to a sudden increase or decrease in high-power load, the physical starting point of the evolution of thermal hazards is first manifested as a high-frequency transient transition of the incoming current. Only by establishing a unified time reference around this transition starting point can the discrete sequence of non-contact infrared surface temperature and the continuous curve of contact internal thermometry be comparable. However, the existing concurrent sampling method is difficult to lock the same zero point in the unsteady window, which makes it difficult for subsequent thermal inertia compensation and cross-modal differential to have a reliable input basis. Therefore, step S1 first completes the extraction of the time scale of the unsteady load change and starts the synchronous acquisition of the two types of temperature data with this time scale.

[0025] S101: High-frequency transient current signal acquisition at the input end and extraction of time scale for unsteady load change.

[0026] Firstly, the current transition at the incoming line of the intelligent distribution box is used as the starting point for this thermal hazard detection. It is necessary to first locate the moment of the unsteady load abrupt change from the continuous current. To ensure that subsequent temperature data has a unified zero point that can be repeatedly retrieved, this sub-step first generates a timescale for the unsteady load abrupt change.

[0027] In practice, a high-frequency current acquisition unit is installed at the incoming end of the intelligent distribution box to continuously sample the current-carrying circuit and output high-frequency transient current signals arranged in the order of sampling. Within the current detection window, the static and stable sample segment immediately preceding the current moment is extracted as the baseline sample segment. The baseline sample segment is only used to construct the boundary for this sudden change judgment, so that the time scale of the non-steady load sudden change comes from the local current background under the same load condition.

[0028] After obtaining the high-frequency transient current signal, a Discrete Wavelet Transform (DWT) is performed on the signal to extract high-frequency detail coefficients characterizing the transient impact of the current. First-order difference is then performed on adjacent sampling points to obtain a current slope sequence characterizing the rate of current change. Subsequently, preset quantiles, such as the 90th quantile, of the absolute values ​​of the high-frequency detail coefficients and the absolute values ​​of the current slope within the baseline sample segment are statistically analyzed to form energy and slope judgment boundaries. When the absolute values ​​of the high-frequency detail coefficients and the absolute values ​​of the current slope corresponding to the current sampling point exceed their respective judgment boundaries simultaneously, and this out-of-bounds state is maintained for two consecutive sampling points, the sampling time that first satisfies the dual out-of-bounds conditions is determined as the time scale of the unsteady-state load change, and a hardware trigger pulse is output at this time scale for direct use by subsequent synchronous acquisition.

[0029] S102: Synchronous acquisition of two types of temperature data with the time scale of unsteady load change as the hardware trigger reference.

[0030] After obtaining the timescale of the unsteady load abrupt change, the acquisition process transitions to the same-source triggering stage. At this point, it is necessary to limit the acquisition of surface temperature and internal temperature to the same physical starting point. Based on this starting point, the two types of temperature data are no longer calculated from their respective independent cycles, but are instead developed together around this timescale.

[0031] In specific implementation, the outer surface projection area corresponding to the placement position of the contact thermistor inside the current-carrying component is first determined as the target monitoring area. When the hardware trigger pulse output by S101 arrives, the same trigger pulse simultaneously drives the non-contact infrared acquisition end and the contact thermistor acquisition end to enter the sampling state. The non-contact infrared acquisition end outputs the surface temperature matrix of the target monitoring area frame by frame. Each frame is arranged in time order to form a discrete sequence of non-contact infrared surface temperature. The contact thermistor acquisition end continuously outputs internal temperature sampling points along the same clock. Each sampling point is connected in time order to form a continuous curve of contact internal thermistor.

[0032] During the synchronous acquisition process, the start time of the first frame of the non-contact infrared surface temperature discrete sequence and the first sampling point of the contact internal thermosensitive continuous curve are simultaneously bound to the non-steady-state load change time scale. This allows the surface thermal response of the target monitoring area to share the same trigger source, the same zero point of time, and the same time sequence direction with the internal thermal response of the same current-carrying component. Subsequent steps can then directly carry out thermal inertia compensation and time-domain alignment processing around this time scale.

[0033] S103: Time base labeling of synchronous sampling results and construction of output from step S1.

[0034] After the two types of temperature data are triggered by the same hardware pulse, the original sampling results need to be processed into output objects that can be directly called in subsequent steps. This processing step revolves around the time base, without changing the individual sampled values ​​themselves, but rather completing the time base connection and fixing the object correspondence.

[0035] In practice, the time stamp of the unsteady load abrupt change is subtracted from the acquisition start time of each frame in the non-contact infrared surface temperature discrete sequence to obtain the relative time stamp for each frame; similarly, the time stamp of the unsteady load abrupt change is subtracted from the sampling time of each sampling point in the contact internal thermistor continuous curve to obtain the relative time stamp for each sampling point. These relative time stamps are attached to their respective sequences in ascending order, thus organizing the non-contact infrared surface temperature discrete sequence into a discrete time sequence with the unsteady load abrupt change time stamp as the zero point, and organizing the contact internal thermistor continuous curve into a continuous time curve with the unsteady load abrupt change time stamp as the zero point.

[0036] After completing the time base annotation, step S1 outputs three directly transferable results: a non-steady-state load abrupt change time scale as the time zero point, a non-contact infrared surface temperature discrete sequence corresponding to the target monitoring area and carrying a relative time mark, and a contact internal thermosensitive continuous curve corresponding to the internal position of the same current-carrying component and carrying a relative time mark. The non-contact infrared surface temperature discrete sequence and the contact internal thermosensitive continuous curve establish a one-to-one correspondence for the same triggering event. Based on this, step S2 can place the contact internal thermosensitive continuous curve under the physical time coordinate defined by the time scale to perform thermal inertia compensation function call and forward phase differential reconstruction.

[0037] Through step S1, the non-steady-state load change timescale has been extracted from the high-frequency transient current signal at the incoming end of the intelligent distribution box. Under the trigger of this timescale, the non-contact infrared surface temperature discrete sequence and the contact internal thermosensitive continuous curve corresponding to the target monitoring area are obtained. Both types of temperature data have completed the time base labeling relative to this timescale and established the correspondence under the same triggering event, thereby forming the time zero point, surface temperature discrete input and internal thermosensitive continuous input that can be directly called in step S2.

[0038] Step S1 has already output the non-steady-state load change time scale, the non-contact infrared surface temperature discrete sequence with relative time stamp, and the contact internal thermistor continuous curve. At this time, although the two types of temperature data share the same triggering starting point, the contact internal thermistor continuous curve still retains the response hysteresis caused by the thermal conduction hysteresis of the encapsulation medium. However, before performing cross-modal frame-by-frame differential calculation in step S3, hysteresis compensation and phase forward must be completed under the physical time coordinate defined by this time scale to reconstruct the contact internal thermistor continuous curve into a reference virtual internal thermal radiation sequence that corresponds to the non-contact infrared surface temperature discrete sequence frame by frame.

[0039] S201: Construction of thermal inertia compensation function.

[0040] After establishing a unified time base in step S1, the continuous curve of the contact-type internal thermistor already has a relative time marker with the time scale of the unsteady-state load abrupt change as zero. This sub-step extracts the thermal conduction hysteresis parameter from this curve and constructs a thermal inertia compensation function. The input starting point of the thermal inertia compensation function is the continuous curve of the contact-type internal thermistor and the time scale of the unsteady-state load abrupt change, and the output result is the time compensation rule called by the subsequent forward phase differential reconstruction.

[0041] In practice, the initial temperature rise segment after the unsteady-state load abrupt change timescale is first extracted from the continuous curve of the contact-type internal thermistor. The first sampling interval where the temperature difference between three consecutive sampling intervals is greater than 0 is determined as the compensation identification segment. Then, the temperature increments of each sampling point within the compensation identification segment relative to the unsteady-state load abrupt change timescale are arranged in chronological order to form a normalized thermal response curve. Subsequently, a first-order lumped heat conduction model is used to perform least-squares fitting on the normalized thermal response curve. In this model, the internal heat source of the current-carrying component, the encapsulation medium, and the contact-type thermistor are equivalent to a first-order thermal inertia network composed of a single heat capacity and a single heat conduction channel, with the normalized thermal response curve being used as the basis for the fitting. The line is used as the target curve for fitting; the range of candidate parameters is determined based on the thickness of the encapsulation medium, thermal conductivity, and the encapsulation specifications of the contact thermistor, or obtained from step heating tests of calibration samples with the same structure; then the sum of squared residuals between the theoretical response curve and the normalized thermal response curve corresponding to each combination of candidate parameters is compared, and the parameter combination with the smallest sum of squared residuals is taken as the material thermal conductivity hysteresis time constant and the heat transfer attenuation coefficient. The material thermal conductivity hysteresis time constant is used to characterize the average delay length of heat transfer from the inside of the current-carrying component to the sensitive end of the contact thermistor, and the heat transfer attenuation coefficient is used to characterize the degree of attenuation of the transient temperature rise amplitude by the encapsulation medium.

[0042] After obtaining the material's thermal hysteresis time constant and heat transfer attenuation coefficient, a thermal inertia compensation function is generated with the unsteady-state load abrupt change time scale as the time zero point. Specifically, the time range from the unsteady-state load abrupt change time scale to three times the material's thermal hysteresis time constant is first divided into continuous time intervals, and a compensation lookup table is formed by pre-generating basic compensation values ​​for each time interval. For each sampling moment, the time interval to which it belongs is first located based on its relative time, and then the basic compensation value is corrected based on the heat transfer attenuation coefficient. When the relative time is between the boundaries of two adjacent time intervals, linear interpolation is performed on the two adjacent basic compensation values ​​to obtain the forward compensation factor for that sampling moment. This ensures that each sampling point in the contact-type internal thermosensitive continuous curve obtains a compensation weight corresponding to its relative time position, which is then used for forward phase differential reconstruction in the next sub-step.

[0043] S202: Forward phase differential reconstruction of contact-type internal thermal continuous curve.

[0044] After the thermal inertia compensation function is constructed, the processing chain moves to the phase advance stage of the contact-type internal thermistor continuous curve. At this point, the thermal inertia compensation function obtained in S201 is directly used, and the contact-type internal thermistor continuous curve is reconstructed point by point along the relative time sequence formed in step S1.

[0045] In practice, at each sampling moment of the contact-type internal thermistor continuous curve, several sampling points are continuously selected from the current sampling point in the positive time direction. The forward temperature rise slope of the current sampling point is calculated using a five-point forward difference algorithm to obtain a local micro-component characterizing the short-term trend of internal temperature changes. When the current sampling moment is located at the tail of the contact-type internal thermistor continuous curve and there are not enough to obtain 5 continuous sampling points, a three-point forward difference is performed using continuous sampling points that can be obtained after the current sampling moment. Then, the forward compensation factor in the thermal inertia compensation function at that sampling moment is read, and the forward temperature rise slope is multiplied by the forward compensation factor to obtain the phase compensation amount used to offset the thermal conduction hysteresis. This phase compensation amount is then superimposed on the current temperature value to generate the virtual internal thermal radiation value at the current sampling moment.

[0046] To ensure that the reconstruction results can be integrated into the same thermal characterization scale as the discrete sequence of non-contact infrared surface temperature, after generating virtual internal thermal radiation values, they are converted into equivalent quantities of internal thermal radiation according to the established correspondence between the contact thermal sensing element and the target monitoring area, and arranged in ascending order according to the relative time stamp to form a continuous trajectory of virtual internal thermal radiation. This continuous trajectory retains the continuous sampling characteristics of the continuous curve of the contact internal thermal sensing element, while pushing the original hysteresis response forward to the physical evolution beat defined by the time stamp of the non-steady-state load change, thereby providing continuous input for frame-level temporal domain alignment.

[0047] S203: Frame-level alignment generation of the baseline virtual internal thermal radiation sequence.

[0048] After completing the virtual internal thermal radiation continuous trajectory, the continuous result needs to be transformed into a discrete result with the same time sampling structure as the non-contact infrared surface temperature discrete sequence. Since step S3 calls the two types of data in a frame-by-frame manner, this sub-step uses the frame time marker of the non-contact infrared surface temperature discrete sequence as the unique sampling index.

[0049] In practice, the relative time markers of the discrete sequence of non-contact infrared surface temperature are read frame by frame, and the same or adjacent time positions are searched in the continuous trajectory of virtual internal thermal radiation. When the infrared frame time marker coincides with the reconstructed sampling time in the continuous trajectory, the corresponding virtual internal thermal radiation value is directly read. When the infrared frame time marker is located between two adjacent reconstructed sampling times, the virtual internal thermal radiation value at that time is obtained by using the piecewise cubic Hermite interpolation algorithm. When the infrared frame time marker is located near the start or end of the continuous trajectory of virtual internal thermal radiation but is insufficient to construct the neighboring points required by the piecewise cubic Hermite interpolation algorithm, the virtual internal thermal radiation value corresponding to the reconstructed sampling time closest to the infrared frame time marker is directly read. In this way, a unique corresponding equivalent value of internal thermal radiation is generated for each frame of the discrete sequence of non-contact infrared surface temperature.

[0050] According to the frame order of the non-contact infrared surface temperature discrete sequence, the virtual internal thermal radiation values ​​corresponding to each frame are arranged sequentially and bound in the same order with the relative time stamp of the corresponding infrared frame to obtain the reference virtual internal thermal radiation sequence. At this point, the reference virtual internal thermal radiation sequence and the non-contact infrared surface temperature discrete sequence form a frame-by-frame correspondence in terms of time zero point, frame-level index and time domain position. In step S3, the two can be directly substituted into the spatiotemporal decoupling cross-mapping matrix to perform cross-modal frame-by-frame differential calculation.

[0051] Through step S2, a thermal inertia compensation function has been constructed based on the contact-type internal thermosensitive continuous curve under the physical time coordinate defined by the time scale of the non-steady-state load change. The forward phase differential reconstruction is completed using the thermal inertia compensation function. Furthermore, a reference virtual internal thermal radiation sequence is generated according to the frame time marker of the non-contact infrared surface temperature discrete sequence. This sequence establishes a frame-by-frame time domain alignment relationship with the non-contact infrared surface temperature discrete sequence and can be used as the internal thermal field input of the spatiotemporal decoupling cross-mapping matrix in step S3.

[0052] Step S2 has output a reference virtual internal thermal radiation sequence that is frame-by-frame time-domain aligned with the non-contact infrared surface temperature discrete sequence. At this point, the two types of data correspond on the time axis, but still represent the surface temperature field and internal thermal radiation field of the target monitoring area, respectively. However, the decision calculation in step S4 requires direct access to a unified result that can quantify the degree of misalignment and distortion between the two. Therefore, step S3 constructs a spatiotemporal decoupling cross-mapping matrix on the same frame index and the same spatial grid, and extracts the transient thermophysical tearing feature tensor accordingly.

[0053] S301: Construction of spatiotemporal decoupling cross-mapping matrix.

[0054] After completing temporal alignment in step S2, the processing chain enters the stage of establishing spatial correspondence. At this point, it is necessary to organize the frame matrix in the non-contact infrared surface temperature discrete sequence and the frame values ​​in the reference virtual internal thermal radiation sequence into the same computational grid. Based on this input, this sub-step first generates a spatiotemporal decoupling cross-mapping matrix.

[0055] In practice, the surface temperature matrix of the target monitoring area in the discrete sequence of non-contact infrared surface temperature is read frame by frame, and each frame matrix is ​​cropped into a grid matrix with a uniform row and column size. At the same time, the frame value of the reference virtual internal thermal radiation sequence corresponding to the relative time mark of each frame is read, and the projection center of the outer surface of the corresponding installation point of the contact thermal element on the target monitoring area is determined as the origin of spatial mapping. Then, the Euclidean distance between the center of each grid cell and the origin of spatial mapping is calculated, and each distance is converted into a non-negative projection weight according to the monotonically decaying rule that the closer the distance, the greater the projection contribution. Subsequently, all projection weights of the same frame are normalized so that the sum of all projection weights is 1, and arranged in grid order to form a spatial projection matrix.

[0056] After obtaining the spatial projection matrix, a time index matrix is ​​established based on the relative time markers of the non-contact infrared surface temperature discrete sequence and the reference virtual internal thermal radiation sequence. The time index matrix retains only the correspondence between the same frame number and the same relative time position. Then, the time index matrix and the spatial projection matrix are combined in the order of first time pairing and then spatial expansion to form a spatiotemporal decoupling cross-mapping matrix. The time index matrix is ​​used to lock the cross-modal input object at the same physical moment, and the spatial projection matrix is ​​used to expand the single frame value into an internal thermal radiation reference grid of the same size as the surface temperature matrix. Thus, subsequent frame-by-frame differential calculations have a unified time entry and spatial entry.

[0057] S302: Cross-modal frame-by-frame differential computation and tear submatrix generation.

[0058] After the spatiotemporal decoupling cross-mapping matrix is ​​constructed, the processing chain transitions to the same-frame differential stage. At this point, a unique reference virtual internal thermal radiation sequence frame value can be found for each frame of the non-contact infrared surface temperature matrix. This sub-step directly calls the spatiotemporal decoupling cross-mapping matrix to perform cross-modal frame-by-frame differential calculations on the two types of inputs.

[0059] In practice, the k-th frame non-contact infrared surface temperature matrix and the k-th frame reference virtual internal thermal radiation sequence frame value are first selected by the time index matrix. Then, the frame value is expanded into an internal thermal radiation reference matrix of the same size as the k-th frame non-contact infrared surface temperature matrix by the spatial projection matrix. Subsequently, the two matrices are subtracted cell by cell at the same grid position to obtain the k-th frame amplitude difference matrix. The amplitude difference matrix is ​​used to characterize the degree of direct deviation of the surface temperature field from the internal thermal radiation field at the same moment.

[0060] After obtaining the amplitude difference matrix of the k-th frame, the gradient amplitude matrices of the non-contact infrared surface temperature matrix and the internal thermal radiation reference matrix of the k-th frame are calculated using the Sobel gradient operator. Then, the two gradient amplitude matrices are subtracted element-wise to obtain the gradient difference matrix of the k-th frame. Subsequently, the amplitude difference matrix and the gradient difference matrix of the k-th frame are expanded sequentially along the same grid and concatenated end-to-end to form the transient tear eigenvector of the k-th frame. Then, an outer product operation is performed on the transient tear eigenvector of the k-th frame to generate the tear sub-matrix of the k-th frame. Since the tear sub-matrix of the k-th frame is formed by the outer product of the same eigenvectors, it maintains a symmetrical structure and fully records the linkage between surface temperature deviation and spatial gradient deviation within that frame.

[0061] S303: Construction and output of transient thermophysical tearing feature tensor.

[0062] After the tear submatrices for each frame are generated, the processing chain enters the cross-temporal organization stage. At this point, the frame-by-frame difference results need to be organized into a unified object that can be directly called in step S4. This unified object must not only retain the misalignment distortion intensity of each frame, but also retain the continuous evolution order in the time dimension.

[0063] In practice, according to the frame order of the non-contact infrared surface temperature discrete sequence, all tear sub-matrices are stacked sequentially along the time dimension, and the relative time marker corresponding to each frame is synchronously attached to the corresponding time slice, thereby forming a transient thermophysical tear feature tensor. Each time slice of this tensor corresponds to a cross-modal frame-by-frame difference result, and each time slice contains a symmetric tear sub-matrix. Therefore, it not only preserves the spatial misalignment and distortion distribution of the target monitoring area within a single frame, but also preserves the sequential relationship of distortion evolution in each frame after the abrupt change of the non-steady-state load.

[0064] After forming the transient thermophysical tearing feature tensor, each time slice is registered in chronological order as a set of slices that can be decomposed by eigenvalues. When called in step S4, the eigenvalues ​​are calculated slice by slice and the maximum value is taken. Thus, the maximum eigenvalue of the transient thermophysical tearing feature tensor is used as an input for calculating the comprehensive risk characterization value. Thus, step S3 completes the closed transformation from the non-contact infrared surface temperature discrete sequence and the benchmark virtual internal thermal radiation sequence to the transient thermophysical tearing feature tensor.

[0065] In step S3, based on the frame-by-frame temporal alignment formed in step S2, a spatiotemporal decoupling cross-mapping matrix for unifying time pairing and spatial expansion is constructed. Based on this matrix, cross-modal frame-by-frame differential calculations are performed on the non-contact infrared surface temperature discrete sequence and the benchmark virtual internal thermal radiation sequence, ultimately obtaining a transient thermophysical tear feature tensor with time slice index. This tensor retains the frame-by-frame symmetric tear sub-matrix and its maximum eigenvalue extraction basis, and can be directly called in step S4 for the calculation of comprehensive risk characterization value.

[0066] Step S3 has formed a transient thermophysical tearing feature tensor with time slice index. At this point, the cross-modal misalignment distortion has been uniformly characterized. However, whether to initiate circuit breaker isolation still needs to be determined by combining the surface thermo-geometric contraction behavior and the frequency domain anomaly of the internal thermal waveform. However, these two criteria are derived from the non-contact infrared surface temperature discrete sequence and the contact internal thermosensitive continuous curve, respectively. Only by further extracting the isothermal envelope boundary contraction rate and the thermal waveform Fourier secondary peak difference, and linking them with the maximum eigenvalue of the transient thermophysical tearing feature tensor, can a breakdown warning action command that can be directly output be formed.

[0067] S401: Geometric contour extraction and isothermal envelope boundary shrinkage calculation of non-contact infrared surface temperature discrete sequence.

[0068] After completing the cross-modal misalignment distortion characterization in step S3, the processing chain first extracts the geometric shrinkage information of surface hot spots from the non-contact infrared surface temperature discrete sequence. The input starting point of this sub-step is the non-contact infrared surface temperature discrete sequence arranged in frames, and the output result is the isothermal envelope boundary shrinkage rate in millimeters per second.

[0069] In practice, the surface temperature matrix corresponding to the target monitoring area is read frame by frame, and linear grayscale mapping is performed according to the minimum and maximum temperature values ​​of that frame to obtain a single-frame grayscale matrix. Then, the Otsu thresholding method is used to calculate the binarization threshold of the single-frame grayscale matrix, and high-grayscale region segmentation is performed according to this threshold to obtain multiple candidate connected components for the current frame. Each candidate connected component is then mapped back to the surface temperature matrix of the current frame, and the regional average temperature of each candidate connected component is calculated. The connected component with the highest regional average temperature is taken as the highest temperature connected component. For the highest temperature connected component, an ordered set of boundary points is extracted using a boundary tracking algorithm, and then a closed geometric contour polygon is generated using a recursive polyline approximation algorithm, thus ensuring that each frame forms a unique corresponding geometric contour polygon.

[0070] After obtaining the geometric contour polygons in two adjacent frames, the pixel distance is first converted to millimeter distance according to the imaging calibration ratio of the target monitoring area. The imaging calibration ratio is obtained during the equipment installation and debugging phase by placing a calibration piece of known size in the target monitoring area and acquiring the corresponding infrared image. The conversion relationship between pixels and millimeters is established according to the pixel size of the calibration piece in the infrared image and its actual size. Then, the geometric contour polygons of the previous frame are resampled at equal intervals according to the boundary arc length to obtain a set of boundary sampling points. For each boundary sampling point, the normal direction pointing from the boundary segment to the inside of the polygon is calculated, and the intersection point of the geometric contour polygons of the next frame is searched along the normal direction to obtain the normal contraction distance of the sampling point between the two frames. The arithmetic mean of all normal contraction distances is used to form the normal displacement amount of the centroid contracting inward between adjacent frames. Then, the normal displacement amount is divided by the corresponding inter-frame time interval to obtain the isothermal envelope boundary contraction rate of the current frame pair, and the isothermal envelope boundary contraction rate sequence is generated in chronological order. When step S4 is called in a subsequent step, the maximum value of the isothermal envelope boundary contraction rate sequence within the predetermined judgment window after the unsteady load mutation is taken as the isothermal envelope boundary contraction rate input for the current target monitoring area.

[0071] S402: Frequency domain transformation of contact-type internal thermosensitive continuous curve and extraction of Fourier secondary peak difference of thermal waveform.

[0072] After extracting the surface thermal geometric shrinkage, the processing chain moves to the frequency domain interpretation stage of the internal thermal waveform. This sub-step directly uses the contact-type internal thermal continuous curve output from step S1, and the output result is the Fourier secondary peak difference of the thermal waveform in degrees Celsius per Hertz.

[0073] In practice, the contact-type internal thermistor continuous curve within a predetermined decision window after the timescale of the unsteady-state load abrupt change is first extracted. The temperature value at that timescale is used as the reference temperature, and baseline subtraction is performed on the curve to obtain a zero-reference temperature rise curve. Subsequently, a Hanning window is applied to the zero-reference temperature rise curve, and a Fast Fourier Transform (FFT) is performed according to the sampling order to convert the time-domain temperature rise curve into a frequency-domain spectral density distribution map. After the frequency-domain spectral density distribution map is formed, all local peaks are scanned along the frequency axis. The local peak with the largest amplitude is determined as the main frequency peak, and the next local peak located on the high-frequency side of the main frequency peak and closest to it is determined as the first sidelobe secondary peak.

[0074] After obtaining the main frequency peak and the first sidelobe secondary peak, the amplitudes of the two are read in the frequency domain spectral density distribution map, and the absolute difference between the two amplitudes is calculated to obtain the thermal waveform Fourier secondary peak difference. The larger this difference is, the more obvious the high-frequency distortion superimposed on the internal thermal response in addition to the main frequency disturbance is. When step S4 is called in a subsequent step, the thermal waveform Fourier secondary peak difference extracted within the same predetermined decision window is used as another dimension input paired with the isothermal envelope boundary contraction rate, thereby forming the two-dimensional coordinate points received by the isolated forest anomaly detection model.

[0075] S403: Dynamic decision coefficient generation, comprehensive risk characterization value calculation, and anti-breakdown early warning action command output.

[0076] After obtaining the isothermal envelope boundary contraction rate and the thermal waveform Fourier second peak difference, step S4 enters the final decision stage. At this point, the two-dimensional thermal anomaly coordinates need to be combined with the transient thermophysical tearing feature tensor output from step S3 to form a judgment result that can directly drive the circuit breaker to trip. The input starting point of this sub-step is the isothermal envelope boundary contraction rate, the thermal waveform Fourier second peak difference, and the transient thermophysical tearing feature tensor.

[0077] As an example, the pre-trained Isolation Forest anomaly detection model can construct a training set from normal samples, slightly anomalous samples, and pre-breakdown anomalous samples, with a total of 2400 samples, including 1800 normal samples, 400 slightly anomalous samples, and 200 pre-breakdown anomalous samples. Each sample retains two features: the isothermal envelope boundary contraction rate and the thermal waveform Fourier second peak difference. During model training, the number of isolated trees is set to 160, the number of subsamples per tree is set to 256, the maximum tree depth is set to 8 layers, the anomalous sample ratio is set to 0.12, and the feature extraction ratio is set to 1.0. As another example, 5-fold cross-validation is used to optimize the number of isolated trees, the number of subsamples, and the anomalous sample ratio. The final parameters are determined when the sum of the false positive rate and the false negative rate of the validation set is minimized. After optimization, the number of isolated trees is set to 180, the number of subsamples is set to 512, the anomalous sample ratio is set to 0.1, and the convergence condition is that the fluctuation of the validation results for three consecutive rounds does not exceed 0.02.

[0078] After completing the model training, a two-dimensional coordinate point of the current target monitoring area is constructed with the isothermal envelope boundary contraction rate as the x-axis and the thermal waveform Fourier second peak difference as the y-axis. The two-dimensional coordinate point is then input into the isolated forest anomaly detection model. The model records the number of split layers from the two-dimensional coordinate point to the leaf node in all random decision trees, and then averages all the split layers to obtain the average path length. At the same time, dynamic decision coefficients are generated based on the theoretical expected depth corresponding to the training sample size of the isolated forest anomaly detection model. Subsequently, eigenvalue decomposition is performed sequentially on each time slice of the transient thermophysical tear feature tensor. The maximum value among all time slice eigenvalues ​​is taken as the maximum eigenvalue of the transient thermophysical tear feature tensor. The dynamic decision coefficient is then multiplied by this maximum eigenvalue to obtain the comprehensive risk characterization value. The critical threshold for physical insulation breakdown is determined by taking the quartiles after sorting the comprehensive risk characterization values ​​of the labeled insulation breakdown samples in ascending order before the breakdown occurs. When the comprehensive risk characterization value of the current target monitoring area exceeds the critical threshold for physical insulation breakdown, a millisecond-level anti-breakdown warning action command is immediately output based on the hardware binding relationship of the module to which the target monitoring area belongs, to control the circuit breaker of the corresponding fault module to disconnect.

[0079] In step S4, the isothermal envelope boundary shrinkage rate and the Fourier second peak difference of the thermal waveform have been extracted from the non-contact infrared surface temperature discrete sequence and the contact internal thermosensitive continuous curve, respectively. The two are then used to construct two-dimensional coordinate points and input into the isolated forest anomaly detection model to generate dynamic decision coefficients. At the same time, the maximum eigenvalue is extracted from the transient thermophysical tear feature tensor, thereby completing the calculation of the comprehensive risk characterization value and comparing it with the physical insulation breakdown critical threshold. Finally, a millisecond-level breakdown warning action command is generated for the corresponding fault module circuit breaker.

[0080] The above embodiments are merely preferred embodiments of the present invention and are not intended to limit the scope of protection of the present invention. Any modifications, equivalent substitutions, or improvements made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A modular intelligent distribution box thermal hazard multi-mode real-time early warning method, characterized in that, Including the following steps: S1: Real-time acquisition of high-frequency transient current signals at the incoming end of the intelligent distribution box, extraction of the time scale of unsteady load change, and using the time scale of unsteady load change as the hardware trigger reference, synchronous acquisition of the non-contact infrared surface temperature discrete sequence and the contact internal thermosensitive continuous curve of the target monitoring area. S2: Based on the physical mechanism of material thermal conduction hysteresis, a thermal inertia compensation function is constructed. Taking the time scale of the unsteady load change as the time zero point, the forward phase differential reconstruction is performed on the continuous curve of the contact internal thermodynamics to obtain a benchmark virtual internal thermal radiation sequence that is aligned with the time domain of the discrete sequence of the non-contact infrared surface temperature. S3: Construct a spatiotemporal decoupling cross-mapping matrix, substitute the non-contact infrared surface temperature discrete sequence and the benchmark virtual internal thermal radiation sequence into the spatiotemporal decoupling cross-mapping matrix to perform cross-modal frame-by-frame differential calculation, and extract the transient thermophysical tearing feature tensor. S4: The Otsu method is used to extract the geometric contour polygon of the highest temperature connected domain and calculate the isothermal envelope boundary shrinkage rate. Fast Fourier Transform is performed on the contact internal thermosensitive continuous curve to extract the Fourier secondary peak difference of the thermal waveform. The two are analyzed together to obtain the dynamic decision coefficient. The threshold judgment is performed by combining the maximum eigenvalue of the transient thermophysical tearing characteristic tensor and outputting the anti-breakdown warning action command.

2. The modular intelligent distribution box thermal hazard multi-mode real-time early warning method according to claim 1, characterized in that, In step S1, the static steady-state sample segment of the high-frequency transient current signal immediately preceding the current time within the current detection window is taken as the baseline sample segment. Discrete wavelet transform is performed on the high-frequency transient current signal to extract high-frequency detail coefficients, and first-order difference is performed on adjacent sampling points to form a current slope sequence. When the absolute value of the high-frequency detail coefficient and the absolute value of the current slope corresponding to the current sampling point continuously exceed the judgment boundary determined by the baseline sample segment and remain at two sampling points, the sampling time that first meets the condition is determined as the time scale of the non-steady-state load change.

3. The modular intelligent distribution box thermal hazard multi-mode real-time early warning method according to claim 2, characterized in that, In step S1, the outer surface projection area corresponding to the placement of the contact thermal element is determined as the target monitoring area, and a hardware trigger pulse is output at the time of the unsteady load change. The same hardware trigger pulse simultaneously starts the non-contact infrared acquisition and the contact thermal acquisition. The start time of each frame of the non-contact infrared surface temperature discrete sequence and the sampling point time of each sampling point of the contact internal thermal continuous curve are synchronously bound to the time of the unsteady load change.

4. The modular intelligent distribution box thermal hazard multi-mode real-time early warning method according to claim 3, characterized in that, In step S2, the initial heating segment after the timescale of the unsteady load change is extracted from the continuous curve of the contact internal thermocouple. The interval where the temperature slope first increases continuously is determined as the compensation identification segment. The temperature increment of each sampling point in the compensation identification segment relative to the timescale of the unsteady load change is used to form a normalized thermal response curve. The normalized thermal response curve is fitted using a first-order lumped heat conduction model to obtain the material thermal conduction hysteresis time constant and heat transfer attenuation coefficient, and a thermal inertia compensation function is generated accordingly.

5. The modular intelligent distribution box thermal hazard multi-mode real-time early warning method according to claim 4, characterized in that, In step S2, along the relative time sequence of the contact-type internal thermal continuous curve, continuous sampling points are selected in the positive time direction for each sampling moment, and the forward temperature rise slope is calculated using a five-point forward differential algorithm. Then, the corresponding forward compensation factor is read according to the thermal inertia compensation function. The forward temperature rise slope and the forward compensation factor are combined to form a phase compensation amount and superimposed on the current temperature value to generate a virtual internal thermal radiation value. Then, sampling is performed according to the frame time marker of the non-contact infrared surface temperature discrete sequence to form a reference virtual internal thermal radiation sequence.

6. The modular intelligent distribution box thermal hazard multi-mode real-time early warning method according to claim 5, characterized in that, In step S3, the surface temperature matrix of the target monitoring area corresponding to the non-contact infrared surface temperature discrete sequence is read frame by frame, and each frame matrix is ​​cropped into a grid matrix with a uniform row and column size. Taking the center of the projection of the installation point of the contact thermal element on the outer surface of the target monitoring area as the origin of spatial mapping, the spatial projection coefficient is calculated according to the distance of each grid unit to the origin of spatial mapping to form a spatial projection matrix. Then, a time index matrix is ​​established based on the relative time mark of the non-contact infrared surface temperature discrete sequence and the reference virtual internal thermal radiation sequence, and the two matrices are combined to form a spatiotemporal decoupling cross-mapping matrix.

7. A modular intelligent distribution box thermal hazard multi-mode real-time early warning method according to claim 6, characterized in that, In step S3, the non-contact infrared surface temperature matrix and the reference virtual internal thermal radiation sequence frame value with the same frame number are selected by the spatiotemporal decoupling cross-mapping matrix, and the frame value is expanded into an internal thermal radiation reference matrix of the same size as the non-contact infrared surface temperature matrix. The two matrices are subtracted element by element to form an amplitude difference matrix. The gradient amplitudes of the two matrices are calculated and then subtracted element by element to form a gradient difference matrix. The two difference results are then expanded and an outer product operation is performed to generate tear sub-matrices for each frame. These are then stacked along the time dimension to form a transient thermophysical tear feature tensor.

8. The modular intelligent distribution box thermal hazard multi-mode real-time early warning method according to claim 1, characterized in that, In step S4, linear gray-level mapping is performed on the surface temperature matrix of each frame of the non-contact infrared surface temperature discrete sequence to obtain a single-frame gray-level matrix, and the Otsu method is used to complete the high gray-level region segmentation to obtain candidate connected components. The candidate connected component with the highest average temperature in the region is determined as the highest temperature connected component, and boundary tracking and recursive broken line approximation are performed on it to form a geometric contour polygon. Then, the isothermal envelope boundary shrinkage rate is generated based on the normal displacement of the geometric contour polygons of two adjacent frames and the inter-frame time interval.

9. A modular intelligent distribution box thermal hazard multi-mode real-time early warning method according to claim 8, characterized in that, In step S4, the contact-type internal thermosensitive continuous curve within a predetermined decision window after the time scale of the unsteady load change is extracted. The temperature value corresponding to the time scale of the unsteady load change is used as the reference temperature to perform baseline subtraction. After applying a Hanning window to the obtained zero reference temperature rise curve, a fast Fourier transform is performed to generate a frequency domain spectral density distribution map. The local peaks of the frequency domain spectral density distribution map are scanned along the frequency axis. The local peak with the largest amplitude is determined as the main frequency peak, and the next local peak that is closest to it on its high-frequency side is determined as the first sidelobe secondary peak.

10. A modular intelligent distribution box thermal hazard multi-mode real-time early warning method according to claim 9, characterized in that, In step S4, a two-dimensional coordinate point is constructed with the isothermal envelope boundary contraction rate as the abscissa and the second peak difference of the thermal waveform Fourier as the ordinate. The two-dimensional coordinate point is then input into the isolated forest anomaly detection model. The number of split layers from the two-dimensional coordinate point to the leaf node is recorded in all random decision trees, and the average path length is taken. Then, dynamic decision coefficients are generated based on the theoretical expected depth corresponding to the training sample size. Eigenvalue decomposition is performed on each time slice of the transient thermophysical tearing feature tensor, and the maximum eigenvalue is taken. Together with the dynamic decision coefficients, a comprehensive risk characterization value is formed. This value is then compared with the physical insulation breakdown critical threshold, and an anti-breakdown warning action command is output.

Images