Automatic detecting and cleaning system for cooling fan

Abstract

The invention relates to the technical field of automatic detection data processing, in particular to a cooling fan automatic detection cleaning system which comprises an electromagnetic compatibility strengthening sensing module, a multi-physical field feature decoupling module, a layered fault diagnosis module, a dynamic cleaning strategy generation module, an anti-corona cleaning execution module and a closed loop feedback optimization module. Temperature field distribution, mechanical vibration signals and partial discharge characteristics are collected in real time through a distributed sensing network and are subjected to tensor fusion with SVG operation parameters to generate a multi-physical field joint characteristic space, and a closed-loop feedback module associates cleaning efficiency data with fault characteristics, reversely corrects model parameters and optimizes a cleaning strategy. According to the invention, traditional threshold limitation is broken through, and closed-loop control of electromagnetic interference suppression, unknown fault dynamic diagnosis and cleaning accurate response is realized.

Description

Technical Field

[0001] The present invention relates to the technical field of automatic detection data processing, and particularly to an automatic detection and cleaning system for a cooling fan. Background Art

[0002] The automatic detection and cleaning system for a cooling fan is a comprehensive device that combines sensing technology, intelligent control, and mechanical execution. Its core principle is to use an embedded sensor array to monitor key indicators such as the thickness of attachments on the fan blade surface, the air flow resistance coefficient, and the operating vibration spectrum in real time. When the detected data exceeds the preset threshold, the microcontroller will trigger a multi-level cleaning program. First, a high-pressure air flow backwashing module is used to directionally remove loose particulate matter, and then an axially adjustable flexible cleaning brush is used to rotate and wipe the blades. At the same time, a negative pressure recovery device is used to separate and collect pollutants. The self-learning algorithm built into the system can optimize the working cycle based on historical cleaning data and dynamically adjust the cleaning strategy in combination with the ambient temperature and humidity sensors, thereby reducing the risk of mechanical wear while maintaining the heat dissipation efficiency.

[0003] The existing system uses an embedded sensor array to monitor indicators such as the thickness of attachments, the air flow resistance coefficient, and the vibration spectrum in real time, and triggers a cleaning program based on a preset threshold. However, this mechanism relies on data modeling of known failure modes and lacks the ability to dynamically diagnose abnormal working conditions that are not included in the training set. When new mechanical structure damages or atypical pollutant depositions occur in the fan assembly, the fixed threshold judgment and the limited feature library lead to an increased misjudgment rate of fault features. Moreover, the self-learning algorithm is limited by the coverage of historical data samples and it is difficult to construct an association model between unknown faults and cleaning requirements in real time. Such problems prevent the system from accurately distinguishing between occasional vibration anomalies and persistent performance degradation, ultimately affecting the accuracy of fault diagnosis and maintenance decisions. Summary of the Invention

[0004] Aiming at the deficiencies of the prior art, the present invention provides an automatic detection and cleaning system for a cooling fan, and the present invention solves the technical problems of high misjudgment rate of abnormal working conditions and low cleaning response accuracy caused by the limitations of preset thresholds and the dependence on historical data.

[0005] To solve the above technical problems, the specific technical solution of the present invention is as follows: An automatic detection and cleaning system for a cooling fan provided by the present invention includes: An electromagnetic compatibility enhanced sensing module, which is used to collect the temperature field distribution, mechanical vibration signals, and partial discharge characteristics of the heat dissipation area of the SVG device in real time, and perform tensor fusion on the temperature field distribution, mechanical vibration signals, and partial discharge characteristics with the SVG operating parameters to generate a multi-physical field joint feature space; A multi-physical field feature decoupling module, connected to the electromagnetic compatibility enhanced sensing module, is used to separate the electromagnetic-mechanical features of the vibration signal in the multi-physical field joint feature space, and outputs the abnormal features of the mechanical structure after removing the electromagnetic interference components through an adaptive filtering algorithm; A hierarchical fault diagnosis module, connected to the multi-physical field feature decoupling module, receives the abnormal features of the mechanical structure and performs the following operations: identifying known fault modes through a pre-trained convolutional neural network classification channel, detecting unknown abnormal signals through an unsupervised clustering anomaly detection channel, and triggering an online incremental learning mechanism when unknown abnormal signals are detected, generating a new fault feature library based on a semi-supervised annotation strategy; A dynamic cleaning strategy generation module, connected to the hierarchical fault diagnosis module, is used to generate an optimized combination strategy of cleaning mode, intensity, and path according to the fault type and severity level output by the hierarchical fault diagnosis module; An anti-corona cleaning execution module, connected to the dynamic cleaning strategy generation module, includes a segmented conductive fiber bundle brush and a linear-rotary composite drive mechanism, and is used to perform a multi-mode collaborative cleaning operation matching the optimized combination strategy; A closed-loop feedback optimization module, connected to the anti-corona cleaning execution module and the hierarchical fault diagnosis module, is used to collect the temperature field distribution, mechanical vibration signal, and SVG operation parameters after cleaning, locate the root cause of unqualified cleaning through multi-source correlation analysis, and feedback the corrected feature extraction weights to the hierarchical fault diagnosis module, and at the same time adjust the parameters of the dynamic cleaning strategy generation module.

[0006] Furthermore, for the automatic detection and cleaning system of the cooling fan of the present invention, the electromagnetic compatibility enhanced sensing module includes: Miniaturized infrared thermal imaging probes deployed in a distributed star topology are arranged on the surface of the IGBT module radiator and are used to obtain the junction temperature distribution of power devices in real time; A magnetic saturation resistant vibration sensor is arranged at the inlet and bending points of the cooling air duct, and captures the coupled signal of turbulence and mechanical vibration through three-axis synchronous sampling; An optical fiber clock synchronization circuit is used to achieve microsecond-level time alignment between the infrared thermal imaging probe and the vibration sensor, and a common mode choke and a multi-layer shielding grounding structure are integrated in the signal conditioning module to suppress high-frequency harmonic interference.

[0007] Furthermore, for the automatic detection and cleaning system of the cooling fan of the present invention, the multi-physical field feature decoupling module includes: A load current harmonic analysis unit is used to generate a reference cancellation signal corresponding to the signal collected by the vibration sensor according to the SVG operating conditions; An adaptive filtering unit, connected to the harmonic analysis unit, for stripping the electromagnetic excitation component from the vibration signal and outputting the mechanical wear characteristics; A Gram angle field transformation unit, for encoding the sequential temperature data collected by the infrared thermal imaging probe into a two-dimensional thermal distribution texture map and inputting it into the convolutional neural network channel of the hierarchical fault diagnosis module.

[0008] Furthermore, for the automatic detection and cleaning system of the cooling fan of the present invention, the hierarchical fault diagnosis module includes: A spatial attention mechanism unit, embedded in the pre-trained convolutional neural network, for focusing on the thermal anomaly region in the two-dimensional thermal distribution texture map output by the Gram angle field transformation unit; A density peak clustering unit, for performing time continuity constraint analysis on the mechanical wear characteristics output by the adaptive filtering unit to distinguish instantaneous electromagnetic disturbances from persistent mechanical faults; An incremental learning trigger unit, connected to the density peak clustering unit, for starting a semi-supervised annotation process when an unmatched anomaly is detected and uploading the newly annotated fault characteristics to the cloud fault feature library.

[0009] Furthermore, for the automatic detection and cleaning system of the cooling fan of the present invention, the hierarchical fault diagnosis module further includes: a cloud fault feature library, for receiving and standardizing and encoding the new fault characteristics uploaded by the incremental learning trigger unit, as well as the partial discharge pattern and the heat dissipation efficiency decay curve across SVG sites; a feature distillation unit, connected to the cloud fault feature library, for extracting common fault patterns and generating a lightweight migration model; A federated learning unit, for dynamically updating the migration model parameters to the convolutional neural network classification channel and the unsupervised clustering anomaly detection channel of the pre-trained convolutional neural network.

[0010] Furthermore, for the automatic detection and cleaning system of the cooling fan of the present invention, the dynamic cleaning strategy generation module includes: A thermal simulation model mapping unit, for dividing the cleaning area according to the heat dissipation priority of the SVG device and inputting the division result into the cleaning parameter optimization unit; A cleaning parameter optimization unit, for associating the cleaning trigger threshold with the temperature gradient map output by the Gram angle field transformation unit and matching the collaborative operation mode of high-voltage pulsed air flow, conductive brushes and negative pressure adsorption; An ant colony algorithm path planning unit, for dynamically optimizing the cleaning trajectory based on the temperature gradient map generated by the Gram angle field transformation unit.

[0011] Furthermore, for the automatic detection and cleaning system of the cooling fan of the present invention, the anti-corona cleaning execution module includes: The segmented conductive fiber bundle brush, where each fiber bundle is independently grounded and integrated with an overcurrent protection circuit, is used to execute the cleaning mode output by the dynamic cleaning strategy generation module; The magnetic grating encoding closed-loop control unit, connected to the linear-rotary compound drive mechanism, is used to adjust the axial feed and circumferential swing accuracy of the brush according to the cleaning trajectory instruction; The negative pressure adsorption following unit is used to recover the peeled metal dust in real time in the path planning result of the ant colony algorithm.

[0012] Furthermore, for the automatic detection and cleaning system of the cooling fan of the present invention, the closed-loop feedback optimization module includes: an efficiency evaluation matrix unit, which is used to define a comprehensive evaluation index of the junction temperature drop slope, vibration attenuation rate, and waveform distortion improvement degree; The entropy weight dynamic allocation unit, connected to the efficiency evaluation matrix unit, is used to adjust the weights of each index according to the real-time working condition data collected by the sensing module; The fault tree traceability unit is used to perform correlation analysis on the air duct pressure difference data collected by the vibration sensor and the fault characteristics output by the incremental learning trigger unit, and locate the root cause of radiator dust accumulation or bearing lubrication deterioration.

[0013] Furthermore, for the automatic detection and cleaning system of the cooling fan of the present invention, the closed-loop feedback optimization module further includes: The cross-site data anonymization unit, based on differential privacy technology, performs desensitization processing on the new fault characteristics uploaded by the incremental learning trigger unit and inputs them into the cloud fault feature library; The transfer learning optimization unit is used to update the lightweight transfer model parameters to the pre-trained convolutional neural network classification channel.

[0014] Advantages of the present invention; Through the dual-channel cooperation mechanism of the hierarchical fault diagnosis module, the present invention uses the spatial attention focusing of the pre-trained convolutional neural network and the temporal constraint analysis of unsupervised clustering to achieve accurate classification of known faults and dynamic detection of unknown anomalies. Combining the incremental learning trigger mechanism and the federated learning parameter update, it breaks through the dependence of traditional systems on historical data samples and significantly reduces the misjudgment rate of abnormal working conditions; the multi-physical field feature decoupling module uses electromagnetic-mechanical feature separation technology, eliminates high electromagnetic interference through harmonic reference signal generation and adaptive filtering algorithms, and combines the Gram angular field transformation to enhance the thermal anomaly characterization ability and improve the accuracy of fault feature extraction; the closed-loop feedback optimization module establishes a closed-loop logic for cleaning efficiency evaluation and strategy parameter correction through dynamic entropy weight allocation and fault tree traceability mechanism, drives the adaptive iteration of the diagnostic model and the cleaning strategy, and finally realizes the precise coordination of the cooling fan anomaly detection and cleaning response, effectively extending the equipment life and reducing the maintenance cost. Description of the drawings

[0015] To more clearly illustrate the technical solution of the present invention, the following will briefly introduce the drawings required in the embodiments. Obviously, for those of ordinary skill in the art, without creative efforts, other drawings can also be obtained according to the drawings.

[0016] Figure 1 It is a system architecture diagram of an automatic detection and cleaning system for a cooling fan provided by an embodiment of the present invention. Specific embodiments

[0017] To make the objectives, technical solutions and advantages of the present invention clearer, the following will clearly and completely describe the technical solutions of the present invention in conjunction with the specific embodiments and corresponding drawings of the present invention. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative efforts belong to the scope of protection of the present invention. The following will detail the technical solutions provided by each embodiment of the present invention in conjunction with the drawings. To better understand the objectives of the present invention, the following will further describe the present invention in detail.

[0018] Please refer to Figure 1 , an automatic detection and cleaning system for a cooling fan provided by the present invention, includes: An electromagnetic compatibility enhanced sensing module, configured to collect the temperature field distribution, mechanical vibration signals, and partial discharge characteristics in the heat dissipation area of the SVG device in real time, and perform tensor fusion on the temperature field distribution, mechanical vibration signals, and partial discharge characteristics with the SVG operation parameters to generate a multi-physical field joint feature space; A multi-physical field feature decoupling module, connected to the electromagnetic compatibility enhanced sensing module, configured to separate the electromagnetic-mechanical characteristics of the vibration signals in the multi-physical field joint feature space, and output mechanical structure anomaly characteristics after removing the electromagnetic interference components through an adaptive filtering algorithm; A hierarchical fault diagnosis module, connected to the multi-physical field feature decoupling module, receiving the mechanical structure anomaly characteristics and performing the following operations: identifying known fault modes through a pre-trained convolutional neural network classification channel, detecting unknown anomaly signals through an unsupervised clustering anomaly detection channel, and triggering an online incremental learning mechanism when unknown anomaly signals are detected, and generating a new fault feature library based on a semi-supervised annotation strategy; A dynamic cleaning strategy generation module, connected to the hierarchical fault diagnosis module, configured to generate an optimized combination strategy of cleaning mode, intensity, and path according to the fault type and severity level output by the hierarchical fault diagnosis module; The anti-corona cleaning execution module, connected to the dynamic cleaning strategy generation module, includes a segmented conductive fiber bundle brush and a linear-rotary composite drive mechanism, and is used to execute multi-mode collaborative cleaning operations matching the optimized combination strategy; The closed-loop feedback optimization module, connected to the anti-corona cleaning execution module and the hierarchical fault diagnosis module, is used to collect the temperature field distribution, mechanical vibration signals, and SVG operating parameters after cleaning, locate the root cause of unqualified cleaning through multi-source correlation analysis, and feedback the corrected feature extraction weights to the hierarchical fault diagnosis module, while adjusting the parameters of the dynamic cleaning strategy generation module.

[0019] This system uses the electromagnetic compatibility enhanced sensing module to collect the temperature field distribution, mechanical vibration signals, and partial discharge characteristics of the heat dissipation area of the SVG device in real time. The sensing module uses a distributed star topology to deploy miniaturized infrared thermal imaging probes, which are arranged on the surface of the IGBT module radiator to capture the junction temperature distribution of the power devices. At the same time, anti-magnetic saturation vibration sensors are installed at key nodes of the heat dissipation air duct to obtain the coupled signal of turbulence and mechanical vibration through three-axis synchronous sampling, and a fiber optic clock synchronization circuit is used to achieve microsecond-level time alignment of multiple sensors. A common mode choke and a multi-layer shielding grounding structure are integrated in the signal conditioning module to suppress the conduction interference of SVG high-frequency harmonics on the sensing signal. The collected temperature field, vibration signals, and partial discharge characteristics and SVG operating parameters generate a multi-physical field joint feature space through tensor fusion, providing a data basis for subsequent feature decoupling.

[0020] The multi-physical field feature decoupling module separates the electromagnetic-mechanical features of the vibration signals in the joint feature space, generates a reference cancellation signal matching the SVG operating conditions through the load current harmonic analysis unit, uses an adaptive filtering algorithm to strip the electromagnetic excitation component from the vibration signals, and retains mechanical structure abnormality features such as bearing wear and blade imbalance. At the same time, the Gramian angular field transformation unit is used to encode the time-series temperature data into a two-dimensional thermal distribution texture map to enhance the recognition sensitivity of the convolutional neural network to local overheating areas. The decoupled mechanical abnormality features and thermal distribution texture map are used as the input data sources of the hierarchical fault diagnosis module.

[0021] The hierarchical fault diagnosis module adopts a dual-channel diagnosis mechanism. The pre-trained convolutional neural network classification channel embeds a spatial attention mechanism, focusing on the thermal anomaly regions of the IGBT heat dissipation substrate and the capacitor bank in the two-dimensional thermal distribution texture map to identify known fault modes. The unsupervised clustering anomaly detection channel analyzes the time continuity constraint of mechanical wear characteristics through the density peak clustering algorithm to distinguish instantaneous electromagnetic disturbances from persistent mechanical faults. When an unmatched anomaly signal is detected, the incremental learning mechanism is triggered to start the semi-supervised annotation process, upload the newly annotated fault features to the cloud fault feature library, and dynamically update the edge-side diagnosis model parameters in combination with the federated learning framework to achieve progressive learning of unknown faults.

[0022] The dynamic cleaning strategy generation module generates an optimized combination strategy according to the fault type and severity level output by the diagnosis module. It divides the cleaning area priority through the thermal simulation model mapping unit, dynamically associates the cleaning trigger threshold with the device junction temperature rise rate, and matches the cooperative operation modes of high-voltage pulsed air flow, conductive brushes, and negative pressure adsorption. The ant colony algorithm path planning unit dynamically generates a cleaning trajectory based on the temperature gradient map, preferentially covering the areas where the temperature rise exceeds the standard. At the same time, a Z-shaped reciprocating trajectory is adopted in the cleaning of the high-voltage busbar, and the negative pressure adsorption port follows the movement to recover the peeled dust in real time.

[0023] The anti-corona cleaning execution module uses a segmented conductive fiber bundle brush. Each fiber bundle is independently grounded and integrated with an overcurrent protection circuit. The axial feed and circumferential swing accuracy of the linear-rotary composite drive mechanism are adjusted through the magnetic grid encoding closed-loop control unit to ensure the precise matching of the cleaning action and the strategy instruction. The negative pressure adsorption following unit adjusts the position of the adsorption port in real time according to the cleaning path to prevent the secondary diffusion of metal dust.

[0024] The closed-loop feedback optimization module collects the temperature field distribution, vibration signals, and SVG operating parameters after cleaning, constructs an effectiveness evaluation matrix of the junction temperature drop slope, vibration attenuation rate, and waveform distortion improvement degree, and dynamically allocates the index weights through the entropy weight method to quantify the cleaning effect. If the performance after cleaning does not meet the expectations, the fault tree tracing unit analyzes the changes in the surface conductivity of the insulator and the air duct pressure difference data in an associated manner, locates the root causes such as dust accumulation in the radiator or deteriorated bearing lubrication, reversely corrects the feature extraction weights of the diagnosis module and the parameter mapping table of the cleaning strategy generation module, and realizes the collaborative optimization of multi-source data through the cross-site data anonymization unit and the transfer learning optimization unit to form a technical closed-loop from data collection to strategy iteration.

[0025] Specifically, for the automatic detection and cleaning system of the cooling fan described in the present invention, the electromagnetic compatibility enhancement sensing module includes: Miniature infrared thermal imaging probes deployed in a distributed star topology are arranged on the surface of the IGBT module radiator to obtain the junction temperature distribution of power devices in real time; The anti-magnetic saturation vibration sensor is deployed at the inlet and bending points of the heat dissipation air duct, and captures the coupled signals of turbulence and mechanical vibration through three-axis synchronous sampling; The fiber optic clock synchronization circuit is used to achieve microsecond-level time alignment between the infrared thermal imaging probe and the vibration sensor, and a common mode choke and a multi-layer shielding and grounding structure are integrated in the signal conditioning module to suppress high-frequency harmonic interference.

[0026] The electromagnetic compatibility enhanced sensing module of this system adopts a distributed star topology to deploy miniaturized infrared thermal imaging probes. The probes are designed based on vanadium oxide uncooled focal plane arrays, integrated with sapphire protection windows, with a working band covering 8-14μm, suitable for installation in the narrow space inside the SVG cabinet. By adopting a multi-node distributed layout, the risk of field of view occlusion of a single probe is reduced, the junction temperature distribution of the power devices on the surface of the IGBT module radiator is captured in real time, and the true temperature field data is obtained by penetrating dust interference. The anti-magnetic saturation vibration sensor is deployed at the inlet and bending points of the heat dissipation air duct, adopts a three-axis synchronous sampling mode, and the axial displacement spectrum analysis covers the X / Y / Z three-dimensional vibration components. By optimizing the saturation characteristics of the magnetic core material, the magnetic flux distortion caused by the high-frequency harmonics of the SVG is suppressed, the coupled signals of turbulence and mechanical vibration are effectively collected, and data distortion caused by electromagnetic interference is avoided.

[0027] The fiber optic clock synchronization circuit adopts a time division multiplexing mechanism, transmits synchronous trigger pulse signals through a single-mode optical fiber, realizes microsecond-level time alignment between the infrared thermal imaging probe and the vibration sensor, and eliminates the timing deviation of multi-source data acquisition. The signal conditioning module integrates a common mode choke and a multi-layer shielding and grounding structure. The common mode choke performs common mode impedance matching for the conducted interference in the frequency band of 10kHz-1MHz. Combining the copper foil shielding layer and the grounding loop design of the metal cabinet, the high-frequency harmonic interference is attenuated to below -60dB, ensuring the electromagnetic compatibility of the original sensing signal. The temperature field distribution, vibration signal and partial discharge characteristics output by this module generate a joint feature space through tensor fusion, providing high signal-to-noise ratio data input for subsequent electromagnetic-mechanical feature decoupling.

[0028] Specifically, for the automatic detection and cleaning system of the cooling fan described in the present invention, the multi-physical field feature decoupling module includes: The load current harmonic analysis unit is used to generate a reference cancellation signal corresponding to the signal collected by the vibration sensor according to the operating conditions of the SVG; The adaptive filtering unit is connected to the harmonic analysis unit and is used to strip the electromagnetic excitation component from the vibration signal and output the mechanical wear characteristics; The Gramian angular field transformation unit is used to encode the sequential temperature data collected by the infrared thermal imaging probe into a two-dimensional thermal distribution texture map and input it into the convolutional neural network channel of the hierarchical fault diagnosis module.

[0029] The multi-physical field feature decoupling module of this system analyzes the characteristic harmonic components in the SVG operating conditions through the load current harmonic analysis unit, uses the fast Fourier transform to identify the amplitude and phase information of the 5th and 7th harmonics, and generates a reference cancellation signal that matches the spectrum of the signal collected by the vibration sensor. The harmonic analysis unit uses the SVG reactive power and the DC bus voltage ripple rate as the operating condition labels, and dynamically adjusts the reference signal generation parameters to synchronize the modeling accuracy of the electromagnetic excitation component with the real-time load fluctuation. The adaptive filtering unit constructs a transversal filter based on the least mean square error criterion, performs a convolution operation on the reference cancellation signal and the original vibration signal, and realizes the adaptive cancellation of the electromagnetic interference component by iteratively adjusting the filter coefficients, and outputs the mechanical wear characteristics after stripping the electromagnetic noise, including abnormal signals such as bearing raceway spalling and blade dynamic balance deterioration.

[0030] The Gramian angular field transformation unit receives the sequential temperature data collected by the infrared thermal imaging probe, maps the one-dimensional temperature sequence into a Gramian angular difference matrix through polar coordinate transformation, and retains the spatio-temporal correlation characteristics of the time series. The transformation unit uses a sliding time window to intercept the temperature fluctuation segment, calculates the cosine angle between the temperature gradients at adjacent time points, and generates a two-dimensional texture map reflecting the heat conduction path on the radiator surface. The encoded thermal distribution texture map is input into the convolutional neural network channel of the hierarchical fault diagnosis module, and the spatial topological features of the local overheating area are extracted through the convolutional kernel, forming a multi-modal input with the mechanical wear characteristics output by the adaptive filtering unit to support the joint reasoning of the subsequent fault diagnosis module. During the electromagnetic-mechanical feature separation process, the reference signal generation, filter parameter iteration, and heat map encoding form a closed-loop coupling to eliminate the cross-coupling effect of multi-physical field data in a high electromagnetic interference environment.

[0031] Specifically, for the automatic detection and cleaning system of the cooling fan described in the present invention, the hierarchical fault diagnosis module includes: A spatial attention mechanism unit, embedded in the pre-trained convolutional neural network, for focusing on the thermal anomaly area in the two-dimensional thermal distribution texture map output by the Gramian angular field transformation unit; A density peak clustering unit, used to perform a time continuity constraint analysis on the mechanical wear characteristics output by the adaptive filtering unit to distinguish instantaneous electromagnetic disturbances from persistent mechanical faults; An incremental learning trigger unit, connected to the density peak clustering unit, used to start a semi-supervised annotation process when an unmatched anomaly is detected, and upload the newly annotated fault characteristics to the cloud fault feature library.

[0032] The hierarchical fault diagnosis module of this system enhances the thermal anomaly recognition ability of the pre-trained convolutional neural network through the spatial attention mechanism unit. This mechanism introduces channel attention weights at the feature map level of the neural network, calculates the correlation between each pixel region in the two-dimensional thermal distribution texture map output by the Gramian angular field transformation unit and the fault label, dynamically assigns the focusing weights of the convolution kernels, preferentially strengthens the gradient response in the edge region of the IGBT heat dissipation substrate, and suppresses the interference of background noise on the classification result. The focused thermal anomaly features and mechanical wear features are fused through the cross-modal feature splicing layer to form a joint input vector to support subsequent fault reasoning.

[0033] The density peak clustering unit conducts a time continuity constraint analysis on the mechanical wear features output by the adaptive filtering unit. It intercepts the time-domain statistics of the vibration signal segments using a sliding time window, calculates the similarity metric of adjacent time window signals through the dynamic time warping algorithm, and constructs a density-based clustering decision graph. This unit screens the clustering centers by setting local density thresholds and distance thresholds, and combines the vibration event duration determination rule to classify the short-term high-density clusters corresponding to instantaneous electromagnetic disturbances as interference signals, and marks the long-term low-density clusters corresponding to persistent mechanical faults as effective abnormal events, realizing the preliminary separation of fault types.

[0034] When the incremental learning trigger unit detects an unmatched abnormal event, it starts a semi-supervised annotation process. Based on the abnormal confidence score output by the density peak clustering and the classification probability of the thermal anomaly features, it generates a candidate annotation set, and pushes the samples with high uncertainty to the operation and maintenance end annotation platform through the manual review interface. After being confirmed by experts, the time-frequency domain key descriptors of the fault features are extracted. The newly annotated features use hash coding technology to compress the storage structure, upload them to the distributed nodes of the cloud fault feature library, and trigger the model fine-tuning under the federated learning framework, synchronize the updated diagnostic model parameters to the edge-side hierarchical fault diagnosis module, forming a closed-loop optimization process from anomaly detection to knowledge transfer.

[0035] Specifically, for the automatic detection and cleaning system of the cooling fan described in the present invention, the hierarchical fault diagnosis module further includes: a cloud fault feature library, which is used to receive and standardize the encoding of the new fault features uploaded by the incremental learning trigger unit, as well as the partial discharge patterns and heat dissipation efficiency decay curves across SVG sites; a feature distillation unit, connected to the cloud fault feature library, which is used to extract common fault patterns and generate a lightweight migration model; A federated learning unit, which is used to dynamically update the migration model parameters to the classification channels of the pre-trained convolutional neural network and the unsupervised clustering anomaly detection channels.

[0036] The cloud-based fault feature library of the hierarchical fault diagnosis module of this system adopts a distributed storage architecture. It receives the new fault feature hash codes uploaded by the incremental learning trigger unit, and integrates the partial discharge pulse current waveform parameters across SVG sites and the temperature gradient parameters of the heat dissipation efficiency decay curve. Through the unified feature encoding protocol, it converts multi-dimensional heterogeneous data into a standardized tensor format, and constructs retrievable feature entries by attaching timestamps, site identifiers, and device model metadata. The feature distillation unit mines common patterns from the standardized feature library based on the K-means clustering algorithm, extracts the time-frequency domain key descriptors of high-frequency faults such as IGBT overheating and bearing wear, and combines knowledge distillation technology to transfer the decision boundary of the complex fault classification model to a lightweight convolutional neural network, generating a transfer model suitable for edge deployment, compressing the model parameter quantity while retaining the generalization recognition ability of cross-site fault patterns.

[0037] The federated learning unit uses homomorphic encryption technology to securely aggregate the transfer model parameters. It collects the gradient updates uploaded by each edge node through the coordination server, calculates the global model parameter mean and synchronizes it to the convolution kernel weight matrix of the pre-trained convolutional neural network classification channel. At the same time, it adjusts the density peak threshold parameter of the unsupervised clustering anomaly detection channel, enabling the diagnostic module to integrate the fault recognition experience across sites. During the model update process, the federated learning unit retains the physical isolation of the data at each site, only exchanges the encrypted parameter increments, and achieves a balance between collaborative optimization of diagnostic capabilities and privacy protection. The updated classification channel enhances the spatial attention focusing accuracy for partial discharge patterns, and the clustering channel optimizes the temporal detection sensitivity for heat dissipation efficiency decay events, forming a dynamic adaptive diagnostic ability iteration mechanism.

[0038] Specifically, for the automatic detection and cleaning system of the cooling fan described in this invention, the dynamic cleaning strategy generation module includes: A thermal simulation model mapping unit, which is used to divide the cleaning area according to the heat dissipation priority of the SVG device and input the division result into the cleaning parameter optimization unit; A cleaning parameter optimization unit, which is used to associate the cleaning trigger threshold with the temperature gradient map output by the Gramian angular field transformation unit and match the collaborative operation mode of high-voltage pulsed air flow, conductive brushes, and negative pressure adsorption; An ant colony algorithm path planning unit, which dynamically optimizes the cleaning trajectory based on the temperature gradient map generated by the Gramian angular field transformation unit.

[0039] The dynamic cleaning strategy generation module of this system analyzes the convective heat transfer coefficient and heat flux density distribution of the SVG device's heat dissipation structure through the thermal simulation model mapping unit, divides the cleaning priorities of the IGBT module heat dissipation substrate, bus bar heat sink, and capacitor bank radiator based on the finite element simulation results, marks the high-temperature hot spot area as the primary cleaning target, and marks the medium- and low-temperature diffusion area as the secondary cleaning target, and outputs the area division matrix to the cleaning parameter optimization unit. The cleaning parameter optimization unit receives the multi-layer temperature gradient map generated by the Gramian angular field transformation unit, dynamically adjusts the cleaning trigger threshold according to the gradient amplitude. When the local temperature rise rate exceeds the preset safety slope, it activates the high-pressure pulsed air flow module for preliminary dust removal, and matches the contact pressure and negative pressure adsorption intensity parameters of the conductive brush according to the gradient direction, generating a multi-tool collaborative operation instruction set.

[0040] The ant colony algorithm path planning unit constructs a rasterized cleaning map based on the temperature gradient map, converts the temperature gradient amplitude into a pheromone concentration weight, and searches for the shortest cleaning coverage path through iterative search. In the algorithm initialization stage, it preferentially traverses the primary cleaning target area, releases virtual pheromones between path nodes to guide the movement direction of the cleaning mechanism, and at the same time introduces a dynamic evaporation factor to avoid the retention of local optimal solutions, and outputs a Z-shaped reciprocating trajectory or a spiral progressive trajectory to the anti-corona cleaning execution module. The path planning result and the cleaning parameter instruction set are sent down synchronously to form a closed-loop control link from thermodynamic simulation to execution actions.

[0041] Specifically, for the heat dissipation fan automatic detection and cleaning system of the present invention, the anti-corona cleaning execution module includes: A segmented conductive fiber bundle brush, each fiber bundle is independently grounded and integrated with an overcurrent protection circuit, and is used to execute the cleaning mode output by the dynamic cleaning strategy generation module; A magnetic grating encoding closed-loop control unit, connected to the linear-rotary composite drive mechanism, and is used to adjust the axial feed and circumferential swing accuracy of the brush according to the cleaning trajectory instruction; A negative pressure adsorption following unit, which is used to recover the peeled metal dust in real time in the ant colony algorithm path planning result.

[0042] The anti-corona cleaning execution module of this system uses a segmented conductive fiber bundle brush. The brush is woven into a bundle based on a carbon nanotube-doped nylon substrate, and a silver layer is plated on the surface to reduce the contact resistance. The resistance value of a single bundle is ≤1Ω. Each bundle of fibers is connected to the SVG device's grounding bus through an independent grounding wire to eliminate the risk of static electricity accumulation during the cleaning operation, and an overcurrent protection circuit is integrated to monitor the brush current intensity. When an abnormal discharge current is detected, the fusing mechanism is triggered. The segmented design allows the brush unit to be dynamically disassembled and combined according to the shape of the cleaning area, adapting to the slit structure of the IGBT heat dissipation substrate and the cleaning requirements of the wide-width plane of the bus bar.

[0043] The magnetic grating encoding closed-loop control unit drives the linear-rotary composite mechanism to execute the cleaning trajectory instruction. The hollow shaft magnetic grating encoder is used to real-time feedback the axial displacement and rotation angle of the brush. The PID closed-loop algorithm is adopted to dynamically compensate for the motion errors caused by mechanical transmission clearance and load disturbance, so that the feeding accuracy of the brush is controlled within the range of ±0.01 mm, and the rotation angle deviation ≤ 0.1°. It meets the safety distance control requirements between the cleaning mechanism and the live components under high-voltage environment. The control unit is built-in with a harmonic suppression algorithm to eliminate the interference of the SVG high-frequency electromagnetic field on the encoder signal.

[0044] The negative pressure adsorption following unit synchronously adjusts the position of the adsorption port in the cleaning path planned by the ant colony algorithm. The adsorption port is flexibly connected to the negative pressure pipeline through a universal joint. The built-in wind speed sensor real-time monitors the negative pressure value of the adsorption port, and dynamically adjusts the speed parameters of the centrifugal fan according to the dust stripping intensity. The edge of the adsorption port adopts a silicone sealing lip design, which fits the surface of the radiator to form a local sealed space, improving the metal dust recovery efficiency. The recovered dust is separated into metal particles and insulating particles through a multi-stage filtration system. The metal particles are deposited in the magnetic collection bin, and the insulating particles are centrally processed through a cyclone separator to prevent secondary diffusion of dust from polluting the internal electrical components of the SVG cabinet.

[0045] Specifically, for the automatic detection and cleaning system of the cooling fan of the present invention, the closed-loop feedback optimization module includes: an efficiency evaluation matrix unit for defining a comprehensive evaluation index of the junction temperature drop slope, vibration attenuation rate, and waveform distortion improvement degree; An entropy weight dynamic allocation unit, connected to the efficiency evaluation matrix unit, for adjusting the weight of each index according to the real-time working condition data collected by the sensing module; A fault tree traceability unit for correlating and analyzing the air duct pressure difference data collected by the vibration sensor and the fault characteristics output by the incremental learning trigger unit to locate the root cause of radiator dust accumulation or bearing lubrication deterioration.

[0046] The closed-loop feedback optimization module of this system constructs a multi-dimensional cleaning efficiency evaluation system through the efficiency evaluation matrix unit. The junction temperature drop slope is calculated based on the temperature field data before and after cleaning collected by the infrared thermal imaging probe, and is obtained by fitting the temperature decay curve of the highest temperature point within the time window; the vibration attenuation rate is quantified according to the change rate of the integral value of the frequency domain signal energy collected by the anti-magnetic saturation vibration sensor; the waveform distortion improvement degree extracts the total harmonic distortion rate (THD) of the SVG output current as the evaluation index, comprehensively reflecting the improvement effect of the cleaning operation on the electrical performance. The entropy weight dynamic allocation unit uses the information entropy theory to analyze the dispersion degree of each index data. When the environmental particulate matter concentration increases, it automatically increases the weight coefficient of the junction temperature drop slope, and focuses on the contribution degree of the vibration attenuation rate when the electromagnetic interference increases, dynamically generating a comprehensive scoring model adapted to the current working condition.

[0047] The fault tree traceability unit receives the new fault feature codes uploaded by the incremental learning trigger unit, aligns them with the time series data of the air duct pressure difference collected by the vibration sensor, calculates the correlation between the pressure difference fluctuation and the vibration signal energy change through the Pearson correlation coefficient. When the pressure difference rises accompanied by an increase in high-frequency vibration energy, it is determined that the dust accumulation in the radiator is the main cause; if the pressure difference is stable while the vibration at the bearing characteristic frequency continues to increase, it is classified as a problem of bearing lubrication deterioration. The traceability results are used to reverse-correct the feature extraction weights of the hierarchical fault diagnosis module, enhance the convolution kernel response intensity of the temperature gradient features related to dust accumulation, and at the same time adjust the trigger threshold of the dynamic cleaning strategy generation module, optimize the mapping relationship between the cleaning frequency and the operation intensity parameters, forming a closed-loop optimization link from performance evaluation to strategy iteration.

[0048] Specifically, for the automatic detection and cleaning system of the cooling fan described in the present invention, the closed-loop feedback optimization module further includes: The cross-site data anonymization unit performs desensitization processing on the new fault features uploaded by the incremental learning trigger unit based on differential privacy technology and inputs them into the cloud fault feature library; The transfer learning optimization unit is used to update the lightweight transfer model parameters to the pre-trained convolutional neural network classification channels.

[0049] The cross-site data anonymization unit of the closed-loop feedback optimization module of this system adopts the Laplace noise injection mechanism to perform desensitization processing on the new fault features uploaded by the incremental learning trigger unit, superimpose random noise that meets the differential privacy budget in the numerical fields of the feature vectors, and at the same time retain the statistical distribution characteristics of the feature vectors, so that the fault data of a single site cannot be restored through reverse engineering. The desensitized feature data is converted into a tensor structure of a unified dimension through standardized coding, and after adding a timestamp and a device model label, it is stored in the isolated partition of the cloud fault feature library to meet the compliance requirements for cross-border data transmission.

[0050] The transfer learning optimization unit extracts the common fault patterns of multiple sites from the cloud fault feature library, uses the knowledge distillation technology to transfer the decision boundary of the complex diagnosis model to the lightweight convolutional neural network, adjusts the softening degree of the classification probability distribution through temperature scaling, and retains the generalization recognition ability of cross-scenario fault features. The lightweight model parameters are updated to the pre-trained convolutional neural network classification channels through the secure aggregation protocol under the federated learning framework. During the local training process of the edge node, only the model gradient increment is uploaded and the homomorphic encryption technology is used to protect data privacy. The aggregation server calculates the global gradient mean and then distributes it to each node for synchronous update to achieve the knowledge sharing and collaborative optimization of the diagnosis model. The updated classification channels enhance the classification confidence in the desensitized fault features, and at the same time maintain the recognition accuracy of the original model for the local known fault patterns, forming a dual technical effect of privacy protection and improvement of diagnosis ability.

[0051] In the specific implementation of this system, distributed star - shaped topology miniaturized infrared thermal imaging probes are deployed in the heat dissipation area of the SVG device through an electromagnetic compatibility - enhanced sensing module. The probes are designed based on a vanadium oxide uncooled focal plane array, with a working band covering 8 - 14μm. They are arranged on the surface of the IGBT module radiator to capture the junction temperature distribution of power devices in real - time, penetrate dust interference to obtain real - temperature field data; anti - magnetic - saturation vibration sensors are installed at the inlet and bends of the heat dissipation air duct, and a three - axis synchronous sampling mode is used to capture the coupled signals of turbulence and mechanical vibration, and high - frequency harmonic interference is suppressed through the magnetic - core saturation characteristic; the fiber - optic clock synchronization circuit transmits trigger pulse signals through single - mode fibers to achieve micro - second - level time alignment of multiple sensors. Combined with a common - mode choke and a multi - layer shielding and grounding structure, the high - frequency harmonic interference is attenuated to below - 60dB, and a multi - physical - field joint feature space integrating temperature field, vibration signal, and partial - discharge characteristics is output.

[0052] The multi - physical - field feature decoupling module analyzes the 5th and 7th harmonic components in the SVG operating conditions through the load - current harmonic analysis unit, generates a reference cancellation signal that matches the vibration - signal spectrum, uses an adaptive filtering algorithm to strip the electromagnetic excitation component, and extracts mechanical anomaly features such as bearing wear and blade imbalance; the Gram - angle field transformation unit encodes the time - series temperature data into a two - dimensional thermal - distribution texture map, calculates the cosine angle of adjacent temperature gradients through a sliding time window, generates texture features reflecting the heat - dissipation path, and inputs them into the hierarchical fault - diagnosis module to support multi - modal analysis.

[0053] The hierarchical fault - diagnosis module uses a pre - trained convolutional neural network classification channel to embed a spatial - attention mechanism, focuses on the thermal - anomaly area of the IGBT heat - dissipation substrate in the thermal - distribution texture map, combines the density - peak clustering unit to perform a time - continuity analysis on mechanical - wear features, and distinguishes instantaneous electromagnetic disturbances from persistent mechanical faults; the incremental - learning trigger unit starts a semi - supervised annotation process when detecting unmatched anomalies. After expert review, new fault - feature labels are generated, uploaded to the cloud feature library, and the edge - end model parameters are updated through the federated - learning framework to achieve dynamic modeling of unknown faults.

[0054] The dynamic cleaning - strategy generation module divides the cleaning priorities of the IGBT heat - dissipation substrate, bus bars, and capacitor banks based on a thermal - simulation model, dynamically associates the temperature - gradient map with the cleaning - trigger threshold, and matches the cooperative - operation parameters of high - voltage pulsed air flow, conductive brushes, and negative - pressure adsorption; the ant - colony - algorithm path - planning unit constructs a pheromone - concentration weight according to the temperature - gradient amplitude, generates Z - type or spiral - trajectory instructions, preferentially cleans the area where the temperature rise exceeds the standard, and synchronously issues them to the anti - corona execution module.

[0055] The anti-corona cleaning execution module uses a segmented conductive fiber bundle brush. Each fiber bundle is independently grounded and integrated with an overcurrent protection circuit. The axial feed accuracy of the linear-rotary composite drive mechanism is adjusted to ±0.01 mm through a magnetic grating encoding closed-loop control unit. The negative pressure adsorption following unit adjusts the position of the adsorption port in real time according to the cleaning path, combines with a silicone sealing lip to form a local sealed space, separates and recovers metal and insulating dust, and prevents secondary pollution.

[0056] The closed-loop feedback optimization module constructs an effectiveness evaluation matrix for the temperature drop slope, vibration attenuation rate, and waveform distortion improvement degree. The index weights are adjusted through dynamic entropy weight allocation. The fault tree tracing unit correlates the air duct pressure difference data with the incremental learning features, locates the root causes of dust accumulation or lubrication deterioration, and reversely corrects the response intensity of the convolution kernel and the cleaning trigger threshold of the diagnostic model. The cross-site data anonymization unit uses Laplace noise injection to desensitize new fault features, and the transfer learning optimization unit extracts common patterns in the cloud to update the lightweight model, realizing the closed-loop iteration of diagnostic capabilities and cleaning strategies, and improving the system's self-adaptability and maintenance accuracy.

[0057] Explanation of the technical feature terms of the present invention: Electromagnetic compatibility enhanced sensing module: Miniature infrared thermal imaging probe: A temperature measurement device using a vanadium oxide uncooled focal plane array, with a working band of 8-14 μm, used for non-contact monitoring of the surface temperature distribution of power devices, and obtaining real thermal field data by penetrating dust interference.

[0058] Anti-magnetic saturation vibration sensor: A three-axis vibration sensor optimized with the magnetic core saturation characteristic, suppressing SVG high-frequency harmonic interference, and collecting the coupled signals of mechanical vibration and turbulence in the heat dissipation air duct.

[0059] Optical fiber clock synchronization circuit: Transmits synchronous trigger pulses through a single-mode optical fiber to achieve microsecond-level time alignment of multiple sensors and eliminate the timing deviation of data acquisition.

[0060] Multi-physical field feature decoupling module: Load current harmonic analysis unit: Analyzes the 5 / 7th harmonic components in the SVG operating current, generates a reference signal that matches the vibration signal spectrum, and is used for electromagnetic interference cancellation.

[0061] Adaptive filtering unit: A dynamic filter based on the least mean square error criterion, which strips the electromagnetic excitation component from the vibration signal and outputs pure mechanical wear characteristics (such as bearing spalling, blade imbalance).

[0062] Gramian angular field transformation unit: Maps the time-series temperature data into a two-dimensional thermal distribution texture map, and encodes the heat conduction path characteristics through the cosine angle between the temperature gradients at adjacent time points.

[0063] Hierarchical fault diagnosis module: Spatial attention mechanism: A weight allocation module embedded in a convolutional neural network that enhances the model's focusing ability on the edge region of the IGBT heat dissipation substrate in the thermal distribution texture map.

[0064] Density peak clustering: An unsupervised algorithm based on local density and distance threshold, combined with the time continuity constraint of vibration signals, to distinguish instantaneous electromagnetic disturbances and persistent mechanical faults.

[0065] Incremental learning trigger unit: When detecting unknown anomalies, it pushes high-uncertainty samples to the manual annotation platform to generate new fault labels and update the cloud feature library.

[0066] Dynamic cleaning strategy generation module: Thermal simulation model mapping: Based on finite element analysis, it divides the cleaning priorities of IGBT modules and bus bar heat sinks, and generates a high-temperature area marking matrix.

[0067] Ant colony algorithm path planning: It converts the temperature gradient amplitude into a pheromone concentration weight, dynamically generates a Z-shaped / spiral trajectory, and preferentially covers the area where the temperature rise exceeds the standard.

[0068] Anti-corona cleaning execution module: Segmented conductive fiber bundle brush: A grounding cleaning tool with a carbon nanotube-doped nylon substrate, with a single bundle resistance ≤ 1Ω, suitable for different radiator structures.

[0069] Magnetic grating encoding closed-loop control: It feeds back displacement / angle signals through a hollow shaft magnetic grating encoder, and the PID algorithm compensates for mechanical errors, with a control accuracy of ±0.01mm.

[0070] Closed-loop feedback optimization module: Effectiveness evaluation matrix: A comprehensive evaluation system of the junction temperature drop slope (temperature decay curve fitting), vibration decay rate (frequency domain energy integral change rate), and waveform distortion improvement degree (THD calculation).

[0071] Federated learning framework: It uses homomorphic encryption to aggregate the model gradients of edge nodes, and realizes the collaborative optimization of cross-site diagnostic models while protecting data privacy.

[0072] Pre-trained convolutional neural network classification channel: Function: Used to identify known fault modes (such as IGBT overheating, bearing wear). Technical implementation: Spatial attention mechanism: Embed an attention weight matrix in the convolutional layer, dynamically focus on the high-temperature anomaly area (such as the edge of the IGBT heat dissipation substrate) in the two-dimensional thermal distribution texture map generated by the Gram angular field transform, and suppress background noise interference.

[0073] Training data: Supervised training is carried out based on the thermal distribution maps and vibration feature samples in the historical fault library to learn the spatial correlation between fault features and temperature gradients. Application scenario: Real-time analysis of thermal distribution texture maps, outputting the classification probabilities of known faults to support hierarchical diagnosis decisions.

[0074] Density peak clustering unit (unsupervised clustering anomaly detection channel): Function: Detect unknown abnormal signals and distinguish instantaneous electromagnetic disturbances from persistent mechanical faults. Technical principle: Density peak algorithm: By calculating the local density and distance threshold of mechanical wear characteristics, screening high-density clustering centers, and combining the time continuity of vibration signals (such as continuous duration > 10 seconds) to determine the type of anomaly.

[0075] Dynamic time warping: Align the vibration signal segments in adjacent time windows, analyze the similarity of time-domain waveforms, and exclude occasional interferences. Application scenario: Unsupervised analysis of the mechanical wear characteristics output by adaptive filtering to mark unmatched abnormal events.

[0076] Incremental learning trigger mechanism and federated learning framework: Function: Dynamically expand the fault feature library and achieve collaborative optimization of cross-site models. Technical implementation: Semi-supervised annotation process: When the density peak clustering detects unmatched anomalies, push high-uncertainty samples to the manual annotation platform, and let experts confirm the fault type and extract time-frequency domain features.

[0077] Federated learning framework: Adopt homomorphic encryption technology to aggregate the model gradient increments of each edge node (SVG site), calculate the global parameter mean, and then synchronously update the pre-trained network to avoid the leakage of original data. Technical effect: Break through the limitations of historical data and incorporate new fault features (such as new pollutant deposition patterns) into the diagnosis system.

[0078] Ant colony algorithm path planning unit: Function: Dynamically generate the optimal cleaning trajectory to cover the area with excessive temperature rise. Technical principle: Pheromone mechanism: Map the temperature gradient amplitude to the pheromone concentration to guide the cleaning mechanism to preferentially traverse high-temperature areas (such as the heat dissipation substrate of the IGBT module).

[0079] Dynamic evaporation factor: Avoid the algorithm falling into local optimum, generate Z-shaped or spiral trajectories by iteratively adjusting the path weights. Application scenario: Combine the temperature gradient map output by the Gram angular field transformation to optimize the cleaning path efficiency.

[0080] Effectiveness evaluation matrix and entropy weight dynamic allocation: Function: Quantify the cleaning effect and dynamically adjust the weights of evaluation indicators. Technical implementation: Comprehensive evaluation indicators: Junction temperature drop slope: calculated by fitting the attenuation curve of the highest temperature point before and after cleaning.

[0081] Vibration attenuation rate: Evaluates the degree of mechanical state recovery based on the rate of change of the integral energy of the frequency domain signal.

[0082] Improvement in waveform distortion: The improvement in electrical performance is measured by the reduction in the harmonic distortion (THD) of the SVG output current.

[0083] Entropy weight method: Dynamically assign weights to each indicator based on real-time operating conditions such as ambient particulate matter concentration and electromagnetic interference intensity (e.g., focusing on the junction temperature drop slope in a high dust environment).

[0084] Federated learning framework and differential privacy technology: Function: Realize cross-site secure data sharing and model update. Technical principle: Differential privacy: Laplace noise is injected into the new fault features uploaded by the incremental learning trigger unit, making it impossible to reversely restore the data of a single site, thus meeting the cross-border data compliance requirements.

[0085] Model distillation and aggregation: Extract common failure modes from the cloud feature library, generate lightweight models through knowledge distillation, and federate and aggregate edge node gradients to update the global model.

[0086] Technical effect: Under the premise of protecting privacy, the model's ability to generalize and identify unknown cross-site failures is improved.

[0087] The present invention solves the technical problems of high misjudgment rate of abnormal working conditions and low cleaning response accuracy caused by the limitations of preset thresholds and historical data dependence through the following technical solutions: First, the hierarchical fault diagnosis module adopts a dual-channel diagnosis mechanism, and the pre-trained convolutional neural network classification channel focuses on the thermal abnormality area through the spatial attention mechanism, and combines the unsupervised clustering channel to analyze the temporal continuity of mechanical wear characteristics, and dynamically identify known and unknown fault modes. When an unmatched abnormal signal is detected, the incremental learning trigger unit starts the semi-supervised labeling process, generates new fault feature labels through expert review and uploads them to the cloud feature library, and uses the federated learning framework to update the edge model parameters, breaking through the coverage limitations of historical data, and realizing real-time modeling and diagnostic capability iteration of unknown faults.

[0088] Secondly, the multi-physical field feature decoupling module generates a reference cancellation signal through load current harmonic analysis, uses an adaptive filtering algorithm to strip the electromagnetic interference components from the vibration signal, and retains the pure mechanical anomaly features. At the same time, the Gramian angular field transformation unit encodes the time-series temperature data into a two-dimensional thermal distribution texture map, enhancing the sensitivity of the convolutional neural network to local overheating areas. Through electromagnetic-mechanical feature separation and multi-modal data fusion, the cross-coupling effect is eliminated in a high electromagnetic interference environment, the accuracy of fault feature extraction is improved, and misjudgment caused by fixed thresholds is avoided.

[0089] Finally, the closed-loop feedback optimization module dynamically allocates the weights of the junction temperature drop slope, vibration attenuation rate, and waveform distortion improvement degree based on the effectiveness evaluation matrix, combines the fault tree tracing unit to correlate the air duct pressure difference data with the new fault features, locates the root cause of insufficient cleaning, and reversely corrects the feature weights of the diagnostic model. The transfer learning optimization unit extracts the common fault patterns across sites in the cloud, updates the pre-trained network through lightweight model parameters, realizes the dynamic adaptation of the diagnostic ability and cleaning strategy, forms a closed-loop logic link from data acquisition, anomaly detection to strategy optimization, and effectively improves the accuracy of cleaning response and the system self-adaptability.

Claims

1. A cooling fan automatic detection and cleaning system, characterized in that: include: The electromagnetic compatibility enhanced sensing module is used to collect the temperature field distribution, mechanical vibration signal and partial discharge characteristics of the heat dissipation area of ​​the SVG equipment in real time, and perform tensor fusion of the temperature field distribution, mechanical vibration signal, partial discharge characteristics and SVG operation parameters to generate a multi-physical field joint feature space; A multi-physical field feature decoupling module, connected to the electromagnetic compatibility enhanced sensing module, is used to perform electromagnetic-mechanical feature separation on the vibration signal in the multi-physical field joint feature space, and output mechanical structure abnormality features after removing the electromagnetic interference component through an adaptive filtering algorithm; a hierarchical fault diagnosis module, connected to the multi-physical field feature decoupling module, receiving the mechanical structure abnormality feature and performing the following operations: identifying known fault modes through a pre-trained convolutional neural network classification channel, detecting unknown abnormal signals through an unsupervised clustering abnormality detection channel, and triggering an online incremental learning mechanism when an unknown abnormal signal is detected, and generating a new fault feature library based on a semi-supervised annotation strategy; A dynamic cleaning strategy generation module, connected to the hierarchical fault diagnosis module, for generating an optimized combination strategy of cleaning mode, intensity and path according to the fault type and severity level output by the hierarchical fault diagnosis module; an anti-corona cleaning execution module, connected to the dynamic cleaning strategy generation module, comprising a segmented conductive fiber bundle brush and a linear-rotational composite drive mechanism, for executing a multi-mode collaborative cleaning operation matching the optimized combination strategy; A closed-loop feedback optimization module is connected to the anti-corona cleaning execution module and the hierarchical fault diagnosis module, and is used to collect the temperature field distribution, mechanical vibration signal and SVG operation parameters after cleaning, locate the root cause of non-compliance with cleaning standards through multi-source correlation analysis, and feed back the corrected feature extraction weights to the hierarchical fault diagnosis module, while adjusting the parameters of the dynamic cleaning strategy generation module.

2. The automatic detection and cleaning system for cooling fans according to claim 1, characterized in that: The electromagnetic compatibility enhanced sensing module comprises: Miniaturized infrared thermal imaging probes deployed in a distributed star topology are placed on the surface of the IGBT module heat sink to obtain the junction temperature distribution of power devices in real time; The anti-magnetic saturation vibration sensor is placed at the entrance and bend of the heat dissipation duct, and captures the coupling signal of turbulence and mechanical vibration through three-axis synchronous sampling; The optical fiber clock synchronization circuit is used to achieve microsecond time alignment between the infrared thermal imaging probe and the vibration sensor, and integrates a common mode choke and a multi-layer shielded grounding structure in the signal conditioning module to suppress high-frequency harmonic interference.

3. The automatic detection and cleaning system for cooling fans according to claim 2, characterized in that: The multi-physics field feature decoupling module includes: A load current harmonic analysis unit, used to generate a reference cancellation signal corresponding to the vibration sensor acquisition signal according to the SVG operating conditions; An adaptive filtering unit, connected to the harmonic analysis unit, for stripping the electromagnetic excitation component from the vibration signal and outputting mechanical wear characteristics; The Gram angular field transform unit is used to encode the time-series temperature data collected by the infrared thermal imaging probe into a two-dimensional thermal distribution texture map, which is input into the convolutional neural network channel of the hierarchical fault diagnosis module.

4. The automatic detection and cleaning system for cooling fans according to claim 1, characterized in that: The hierarchical fault diagnosis module comprises: A spatial attention mechanism unit, embedded in the pre-trained convolutional neural network, for focusing on thermal anomaly areas in the two-dimensional thermal distribution texture map output by the Gram angular field transform unit; The density peak clustering unit is used to perform time continuity constraint analysis on the mechanical wear characteristics output by the adaptive filter unit to distinguish between instantaneous electromagnetic disturbances and persistent mechanical failures; The incremental learning trigger unit is connected to the density peak clustering unit and is used to start the semi-supervised labeling process when an unmatched anomaly is detected, and upload the newly labeled fault features to the cloud fault feature library.

5. The automatic detection and cleaning system for cooling fans according to claim 4, characterized in that: The hierarchical fault diagnosis module further includes: a cloud fault feature library for receiving and standardizing new fault features uploaded by the encoding incremental learning trigger unit, as well as partial discharge patterns and heat dissipation efficiency attenuation curves across SVG sites; a feature distillation unit connected to the cloud fault feature library for extracting common fault modes and generating a lightweight migration model; A federated learning unit is used to dynamically update the migration model parameters to the pre-trained convolutional neural network classification channel and the unsupervised clustering anomaly detection channel.

6. The automatic detection and cleaning system for cooling fans according to claim 1, characterized in that: The dynamic cleaning strategy generation module includes: A thermal simulation model mapping unit, used for dividing the cleaning area according to the heat dissipation priority of the SVG equipment, and inputting the division result into the cleaning parameter optimization unit; A cleaning parameter optimization unit is used to associate the cleaning trigger threshold with the temperature gradient map output by the Gram angle field conversion unit, and match the collaborative operation mode of high-pressure pulse airflow, conductive brush and negative pressure adsorption; The ant colony algorithm path planning unit dynamically optimizes the cleaning trajectory based on the temperature gradient map generated by the Gram angular field transform unit.

7. The automatic detection and cleaning system for cooling fans according to claim 1, characterized in that: The anti-corona cleaning execution module comprises: Segmented conductive fiber bundle brush, each fiber bundle is independently grounded and integrated with overcurrent protection circuit, used to execute the cleaning mode output by the dynamic cleaning strategy generation module; A magnetic grating encoding closed-loop control unit, connected to the linear-rotation compound drive mechanism, for adjusting the axial feed and circumferential swing accuracy of the brush according to the cleaning trajectory instruction; Negative pressure adsorption following unit is used to recover the stripped metal dust in real time in the ant colony algorithm path planning results.

8. The automatic detection and cleaning system for cooling fans according to claim 1, characterized in that: The closed-loop feedback optimization module includes: a performance evaluation matrix unit, which is used to define comprehensive evaluation indicators of junction temperature drop slope, vibration attenuation rate and waveform distortion improvement; An entropy weight dynamic allocation unit, connected to the performance evaluation matrix unit, for adjusting the weight of each indicator according to the real-time working condition data collected by the sensor module; The fault tree tracing unit is used to correlate and analyze the air duct pressure difference data collected by the vibration sensor and the fault characteristics output by the incremental learning trigger unit to locate the root cause of radiator dust accumulation or bearing lubrication deterioration.

9. The automatic detection and cleaning system for cooling fans according to claim 8, characterized in that: The closed-loop feedback optimization module also includes: The cross-site data anonymization unit desensitizes the new fault features uploaded by the incremental learning trigger unit based on differential privacy technology and inputs them into the cloud fault feature library; Transfer learning optimization unit, used to update the lightweight transfer model parameters to the pre-trained convolutional neural network classification channel.

Images