Complex illumination low-voltage electricity meter frequency sensing detection method, system and device and medium

Abstract

The invention relates to the technical field of intelligent detection of power equipment, in particular to a complex illumination low-voltage electricity meter frequency sensing detection method, system and equipment and a medium, and aims to solve the problems of high-frequency noise interference, intra-class feature difference, insufficient edge end energy efficiency and the like existing in electricity meter detection in a high-brightness environment. Proposing a dynamic frequency domain compensation mechanism to suppress light noise and recover signal details, and optimizing dense scene feature consistency in combination with adaptive filtering and a multi-scale sampling strategy; the boundary fuzzy problem is solved through dynamic feature alignment and an edge compensation mechanism, and parallel acceleration of frequency domain decomposition and dynamic resampling is realized by adopting a heterogeneous hardware architecture; a multi-modal data training framework and a cross-domain transfer learning strategy are fused to enhance the generalization ability of the model; according to the invention, collaborative optimization of detection precision, real-time performance and low power consumption is realized in a complex illumination scene, and high-robustness technical support is provided for intelligent power grid inspection.

Description

Technical Field

[0001] The present invention relates to the technical field of intelligent detection of electric power equipment, and in particular to a method, system, equipment and medium for detecting frequency induction of a complex illumination low-voltage electric meter. Background Art

[0002] The current field of intelligent detection of power equipment mainly relies on deep learning visual inspection technology. Models based on spatial domain feature extraction have achieved certain application results in conventional environments. Frequency domain analysis methods, which suppress noise through filtering operations, are gradually gaining attention, but their synergy with deep learning frameworks has not yet been broken through. Although existing dynamic sampling mechanisms and feature fusion methods have made progress in improving multi-scale detection capabilities, they still face systemic bottlenecks in complex industrial scenarios. Core issues such as restoring high-frequency details in images in high-brightness environments, optimizing intra-class consistency of densely arranged targets, and improving edge positioning accuracy have not yet formed effective solutions. Defects such as the disconnect between frequency domain features and semantic understanding and insufficient hardware adaptability have restricted the implementation of the technology.

[0003] Existing technologies struggle to effectively recover meter surface texture and character details in high-reflectivity scenarios. Traditional convolution operations and downsampling processes result in continuous attenuation of high-frequency information, and spatial attention mechanisms lack the ability to separate frequency bands. In densely populated scenarios, adjacent meters experience shifts in feature distribution within a class due to occlusion and illumination differences. Existing feature pyramid structures are unable to simultaneously achieve multi-scale fusion and intra-class noise suppression. Edge localization accuracy is limited by boundary blurring introduced by low-frequency interference, and dynamic sampling methods lack frequency domain prior constraints, resulting in insufficient high-frequency edge recovery. Frequency domain enhancement techniques lack end-to-end synergy with deep learning models, manually designed filter parameters struggle to adapt to dynamic illumination changes, and the high computational complexity of frequency domain operations hinders real-time deployment. Data-driven approaches are limited by high annotation costs and scene generalization capabilities. Traditional data augmentation techniques cannot simulate real-world physical reflectivity characteristics, and the insufficient fidelity of high-frequency details in synthetic data further degrades model performance. These shortcomings collectively make existing solutions unable to meet the comprehensive requirements for detection accuracy, real-time performance, and environmental adaptability in complex industrial environments. Summary of the Invention

[0004] In view of the above existing problems, the present invention is proposed.

[0005] Therefore, the present invention provides a frequency-sensing detection method and system for low-voltage electric meters under complex lighting conditions to solve core problems such as high-frequency signal aliasing, intra-class feature differences of dense electric meters, blurred boundaries, and insufficient real-time performance at the edge in high-brightness environments. The dynamic frequency domain compensation mechanism is used to suppress optical noise interference and restore high-frequency details, and the adaptive feature fusion strategy is combined to optimize intra-class consistency in dense scenarios; the dynamic feature alignment strategy and edge compensation mechanism are innovatively introduced to significantly improve boundary positioning accuracy; based on the heterogeneous hardware collaborative architecture, parallel acceleration of frequency domain decomposition and dynamic resampling is achieved to optimize edge computing load and energy efficiency. At the same time, the multimodal data training framework and cross-domain transfer learning strategy are integrated to enhance the model's generalization ability for complex lighting environments. This method provides a highly robust, low-power industrial-grade detection solution for power inspections in high-brightness scenarios such as airports and commercial buildings, effectively supporting the efficient operation and maintenance needs of smart grids.

[0006] In order to solve the above technical problems, the present invention provides the following technical solutions:

[0007] In a first aspect, the present invention provides a method for detecting frequency induction of a low-voltage electric meter under complex illumination, comprising:

[0008] Through the filter generation inversion mechanism, compensation enhancement and suppression under complex lighting conditions can be achieved;

[0009] Through the adaptive filtering and sampling synergy mechanism, the distribution differences of target features in the scene are suppressed;

[0010] Dynamically adjust features based on convolution and compensation strategies to solve boundary blur and detail loss;

[0011] Fusion of output features, parallel processing, and adjustment strategies to optimize edge energy efficiency;

[0012] Fusion parameters and features for training, combining data sets and functions to enhance model generalization capabilities.

[0013] As a preferred solution of the method for frequency-sensing detection of low-voltage electric meters under complex illumination of the present invention, wherein: a filter is used to generate an inversion mechanism to achieve compensation enhancement and suppression under complex illumination, including:

[0014] Dynamically generate filters and perform inverse compensation to achieve joint enhancement and suppression;

[0015] Build a framework to distribute interference and signals by weights;

[0016] Leverage dynamic resource adaptation to optimize suppression boundaries.

[0017] By dynamically generating reversible filters, it achieves precise separation of high-frequency noise and effective signals, and adaptively compensates for detail loss in strong light environments; constructs an intelligent weight allocation framework to distinguish interference from target features, suppressing noise interference in overexposed areas while enhancing texture information in shadow areas; and dynamically optimizes resource allocation strategies based on edge hardware characteristics, significantly improving feature extraction accuracy under complex lighting conditions while ensuring a balance between real-time processing performance and energy efficiency of edge devices.

[0018] As a preferred solution of the complex illumination low-voltage meter frequency detection method described in the present invention, wherein: adaptive filtering and sampling collaborative mechanism is used to suppress the distribution difference of target features in the scene, including:

[0019] Dynamically generate filter kernels and optimize intra-class feature smoothing and distribution consistency through feature fusion strategies;

[0020] Build a collaborative framework, combined with a detail-preserving mechanism;

[0021] Perform feature energy difference analysis to match intra-class interference.

[0022] By dynamically generating spatially adaptive filter kernels, intra-class feature smoothing of densely packed electricity meter targets is achieved, effectively reducing the feature distribution offset caused by illumination differences. A collaborative framework for filtering and sampling is constructed to preserve high-frequency detail textures while maintaining the consistency of low-frequency main features. Dynamic matching interference patterns are analyzed based on feature energy differences, significantly reducing the problem of mismatching of adjacent electricity meters and improving detection robustness in complex scenarios.

[0023] As a preferred solution of the complex illumination low-voltage meter frequency detection method described in the present invention, the method dynamically adjusts features based on convolution and compensation strategies to solve boundary blurring and detail loss, including:

[0024] Establish a feature alignment strategy to achieve feature sharpening in boundary areas through prediction and control optimization;

[0025] Construct a feature compensation mechanism to suppress detail degradation through feature superposition and weight distribution;

[0026] Perform joint feature optimization to combine constraints and spatial domain semantic information to enhance boundary localization accuracy.

[0027] Through dynamic offset prediction and deformable convolution, precise alignment of boundary features is achieved, effectively restoring the sharpness of the meter contour under halo interference; a high-frequency detail compensation mechanism is constructed to suppress information loss in the downsampling process and ensure the integrity of key features such as character textures; and the frequency domain prior constraints and semantic information are integrated for joint optimization to significantly improve the boundary positioning accuracy and detection stability in complex lighting scenes.

[0028] As a preferred solution of the complex illumination low-voltage meter frequency detection method described in the present invention, the output features are integrated, parallel processing is performed, and the edge energy efficiency is optimized in combination with the adjustment strategy, including:

[0029] Utilize feature fusion to achieve feature collaborative optimization and calculation through dynamic weight allocation;

[0030] Perform parallel processing and collaboratively perform frequency domain decomposition, feature resampling, and smoothing tasks;

[0031] Implement dynamic adjustment strategies, combining computing resource awareness with dynamic voltage adaptation to optimize edge energy efficiency.

[0032] As a preferred solution of the complex illumination low-voltage meter frequency detection method described in the present invention, the following methods are used to integrate parameters and features for training, combine data sets and functions, and enhance the generalization ability of the model:

[0033] Construct a multimodal training dataset covering dynamic lighting conditions, integrating time-frequency spectrum features with spatial domain image data;

[0034] Design a composite loss function to collaboratively optimize positioning accuracy and category distribution balance, and suppress overfitting and sample imbalance through dynamic parameter weighting;

[0035] Implement cross-modal transfer learning strategies, enhance the adaptability to complex scenarios by migrating pre-trained model parameters, and achieve generalized mapping from low-frequency data to high-frequency features.

[0036] As a preferred solution of the complex illumination low-voltage meter frequency detection method of the present invention, the training of the fusion parameters and features includes:

[0037] Construct a joint time-frequency-spatial training dataset covering multiple illumination intensities, and achieve collaborative representation of high-frequency noise and low-frequency semantics through cross-modal feature fusion;

[0038] Design a dynamic weighted composite loss function, combining positioning error and category distribution entropy to optimize the model parameter update direction;

[0039] A cross-modal parameter transfer strategy is implemented to constrain the spatial domain detection network parameter initialization based on pre-training spectral features.

[0040] In a second aspect, the present invention provides a complex illumination low-voltage meter frequency sensing detection system, comprising:

[0041] High-frequency restoration module, used to compensate for high-frequency detail loss and enhance image edge and texture information through adaptive high-pass filter generation and kernel inversion mechanism;

[0042] The intra-class smoothing module optimizes the intra-class feature consistency of dense meter scenarios through spatially adaptive low-pass filtering and PixelShuffle upsampling;

[0043] Dynamic adaptive resampling module, which achieves feature alignment based on deformable convolution and boundary compensation mechanism, suppressing the over-smoothing effect of the ICS module;

[0044] Multimodal feature fusion module, which uses bidirectional feature pyramid (BiFPN) to fuse HFR, ICS, and DAR output features;

[0045] The edge computing module includes an FPGA-accelerated frequency domain decomposition unit, a dynamic resampling accelerator, and an intra-class smoothing coprocessor.

[0046] In a third aspect, the present invention provides an electronic device, comprising:

[0047] memory and processor;

[0048] The memory is used to store computer-executable instructions, and the processor is used to execute the computer-executable instructions. When the computer-executable instructions are executed by the processor, the steps of the complex light-induced low-voltage meter frequency sensing detection method are implemented.

[0049] In a fourth aspect, the present invention provides a computer-readable storage medium storing computer-executable instructions, which, when executed by a processor, implement the steps of the complex illumination low-voltage meter frequency sensing detection method.

[0050] Compared with the existing technology, the present invention has the following beneficial effects: the present invention suppresses high-frequency noise interference under complex lighting conditions and restores signal details through a dynamic frequency domain compensation mechanism, significantly improving detection sensitivity in strong light environments; optimizes intra-class feature consistency in dense scenes through adaptive filtering and multi-scale feature fusion strategies, effectively reducing the risk of target mismatching; combines dynamic feature alignment mechanisms with edge compensation strategies to solve boundary fuzziness problems and enhance positioning accuracy. In the intermediate steps, parallel processing of frequency domain decomposition and dynamic resampling is achieved through a heterogeneous hardware collaborative architecture to optimize edge computing efficiency and energy efficiency; finally, based on multimodal data fusion and cross-domain transfer learning strategies, the model's generalization and adaptability to complex lighting conditions is enhanced. Overall, this method achieves the coordinated optimization of robustness, real-time performance, and low power consumption for low-voltage meter detection in high-brightness scenes, providing reliable technical support for complex environment inspections of smart grids. BRIEF DESCRIPTION OF THE DRAWINGS

[0051] In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the following briefly introduces the drawings required for use in the description of the embodiments. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative work.

[0052] Figure 1 This is an overall architecture diagram of a frequency sensing detection method for a complex illumination low-voltage electric meter according to an embodiment of the present invention.

[0053] Figure 2 This is an algorithm flow chart of a frequency sensing detection method for a complex illumination low-voltage electric meter according to an embodiment of the present invention.

[0054] Figure 3 This is a hardware architecture diagram of a complex illumination low-voltage electric meter frequency sensing detection method according to an embodiment of the present invention.

[0055] Figure 4 This is a comparison chart of the detection effects of the complex illumination low-voltage meter frequency sensing detection method described in one embodiment of the present invention. DETAILED DESCRIPTION

[0056] To make the above-mentioned objects, features, and advantages of the present invention more clearly understood, the following detailed description of the specific embodiments of the present invention is given in conjunction with the accompanying drawings. It is obvious that the described embodiments are only part of the embodiments of the present invention, not all of them. Based on the embodiments of the present invention, all other embodiments obtained by ordinary persons in this field without creative work should fall within the scope of protection of the present invention.

[0057] Example 1, with reference to Figure 1 , as one embodiment of the present invention, provides a complex illumination low-voltage meter frequency sensing detection method, comprising:

[0058] S1: Generates an inversion mechanism through filters to achieve compensation enhancement and suppression under complex lighting conditions;

[0059] S2: Suppress the distribution differences of target features in the scene through the adaptive filtering and sampling synergy mechanism;

[0060] S3: Dynamically adjust features based on convolution and compensation strategies to solve boundary blur and detail loss;

[0061] S4: Fusion of output features, parallel processing, and optimization of edge energy efficiency using adjustment strategies;

[0062] S5: Fusion parameters and features for training, combining datasets and functions to enhance model generalization capabilities.

[0063] It should be noted that existing low-voltage meter detection technology faces core issues in complex lighting environments, including noise interference in high-frequency signals, which reduces the signal-to-noise ratio; discrepancies in target feature distribution in dense scenes, which lead to false detections; blurred boundaries and loss of detail that affect positioning accuracy; limited edge computing resources that make it difficult to meet real-time requirements; and insufficient model adaptability to dynamic lighting conditions. Traditional methods rely on manually designed filters or single-modal feature fusion, which cannot effectively balance noise suppression and detail preservation. They also lack collaborative optimization mechanisms for hardware energy efficiency, making them difficult to meet the needs of industrial detection in high-brightness scenarios.

[0064] Therefore, to address the aforementioned technical bottlenecks, a systematic end-to-end solution was constructed through steps S1-S5: S1 uses a dynamic frequency domain compensation mechanism to suppress optical noise interference and enhance signal details, addressing high-frequency aliasing. S2 employs an adaptive filtering and multi-scale sampling collaborative strategy to optimize intra-class feature consistency in dense scenes and reduce false detection rates. S3 combines dynamic convolution with a compensation strategy to reconstruct the feature space and improve boundary positioning accuracy. S4 optimizes edge computing efficiency through a heterogeneous parallel architecture and dynamic resource scheduling strategies to achieve real-time processing. S5 integrates a multimodal data training framework with a cross-domain transfer learning mechanism to enhance the model's generalization capabilities for complex lighting environments. The synergistic effect of these steps ultimately enables high-precision, low-power, and highly robust low-voltage meter detection under complex lighting conditions, providing reliable technical support for smart grid inspections.

[0065] Example 2, reference Figure 1-Figure 3 , which is an embodiment of the present invention, provides a complex illumination low-voltage meter frequency sensing detection method based on the above embodiment.

[0066] In the implementation manner of the present application, in step S1, a filter is generated through an inversion mechanism to achieve compensation enhancement and suppression under complex illumination. Through the principle of dynamic frequency domain feature decoupling, an initial filter kernel is generated through a convolution operation, and a low-pass filter kernel is obtained through normalization processing. Then, an adaptive high-pass filter kernel is generated through a matrix inverse operation; the high-pass filter is applied to the input feature map to extract high-frequency edge and texture information and suppress light noise interference; at the same time, a cross-modal attention fusion framework is constructed, and the original features and enhanced features are fused through a dynamic weight allocation strategy, invalid high-frequency noise is suppressed in the overexposed area, and detail retention is enhanced in the shadow area, thereby achieving coordinated optimization control of illumination interference and signal details.

[0067] In an optional embodiment, in step S1, a filter is used to generate an inversion mechanism to achieve compensatory enhancement and suppression under complex lighting conditions. A multi-scale frequency band decomposition framework of wavelet transform can be used in combination with a dynamic threshold adjustment strategy for light intensity to generate an adaptive band-stop filter kernel group, and noise interference in specific light bands can be suppressed by multi-band weight fusion.

[0068] In another optional embodiment, in step S1, an inversion mechanism is generated by a filter to achieve compensation enhancement and suppression under complex lighting. It is also possible to dynamically estimate the light noise distribution parameters by introducing a Bayesian probability model, combine the Markov random field to construct spatial constraints, generate an adaptive bandpass filter kernel related to the light intensity, and achieve joint sparse representation optimization of noise and signal through the frequency domain mask matrix.

[0069] In the embodiment of the present application, the filter generation inversion mechanism in step S1 is used to achieve compensation enhancement and suppression under complex lighting conditions, and further includes:

[0070] High frequency recovery module: such as Figure 2 As shown in (a), the HFR module is designed to enhance edge details and compensate for the high-frequency information lost during the downsampling process. It takes the initially fused feature map as input and ultimately generates a high-pass filter. Specifically:

[0071] The input is first processed by a 3×3 convolutional layer to generate a low-pass filter kernel, followed by a Softmax activation and filter inversion operation.

[0072] Let K' denote the kernel size of the high-pass filter. To ensure that the final kernel performs high-pass filtering, this module first generates a low-pass filter by applying a kernel-wise Softmax.

[0073] The high-pass filter is then obtained by subtracting the low-pass filter from the identity kernel.

[0074] By applying this filter inversion mechanism, the HFR module effectively restores high-frequency details and enhances the model's sensitivity to edges and fine-grained information in high-glare environments. To address the problem of high-frequency electrical signal loss (such as harmonic distortion and voltage transients) caused by photosensor saturation in strong light environments, the filter inversion mechanism compensates for the loss of high-frequency details as follows:

[0075] Adaptive high-pass filter generation:

[0076] The input feature map is convolved by 3×3 to generate the initial filter kernel, and the normalized low-pass filter kernel (K) is generated by the Softmax function. low ), and use the kernel inversion technique to generate the adaptive high-pass filter kernel:

[0077] K high =IK low

[0078] Where I is the unit kernel matrix, and the elements satisfy Represents an all-pass filter, preserving all frequency components. K lowThe low-pass filter kernel is generated by 3×3 convolution and normalized by Softmax K low =Softmax(W conv *F input +b) Suppress high-frequency noise and retain low-frequency main features. high The high-pass filter kernel is generated by matrix subtraction The filter kernel size K′=5×5 is experimentally verified to be sensitive to high-frequency features. The softmax temperature coefficient τ=0.1 controls the concentration of the kernel weight distribution and normalizes the constraint ∑ i,j K low (i, j) = 1. Matrix subtraction is used to invert the spectrum from low-pass to high-pass.

[0079] High-frequency feature enhancement:

[0080] Apply the high-pass filter to the original feature map to extract edge and texture information. The formula is:

[0081] F enhanced =F input *K high

[0082] Among them, * represents the depth-separable convolution operation, the input feature map With nuclear K high Convolution reduces the amount of computation and preserves cross-channel relationships, F enhanced Output dimension The enhanced feature map retains the high-frequency details of the meter signal and highlights the harmonic features and edge textures (such as harmonic distortion waveforms).

[0083] Dynamic Weight Fusion:

[0084] Through the cross-channel attention mechanism (formula: α=σ(W T [F input ,F enhanced ])) Fusion of original and enhanced features to suppress optical noise interference, where W T is the learnable weight matrix W∈R 2C×C Perform cross-channel feature interaction, σ(·) is the Sigmoid activation function to generate [0, 1] dynamic weights, α is the fusion weight coefficient Suppress noise in overexposed areas. When the ambient light intensity is >5000 lux, α approaches 0 in the overexposed area, suppressing invalid high-frequency noise. In the shadow area, α→1, preserving the enhanced details.

[0085] Software core collaborative workflow, such as Figure 1 As shown:

[0086] The input image is passed through the backbone network to extract multi-level features.

[0087] The input image is processed through an improved YOLOv8-Tiny backbone network (Backbone) to extract multi-level spatial-frequency domain hybrid features. The backbone network uses depthwise separable convolution and frequency decomposition layers in parallel to output three levels of feature maps:

[0088] Shallow features: resolution 640×640, including high-frequency edge and texture information;

[0089] Mid-level features: resolution 320×320, integrating spatial domain semantics and frequency domain energy distribution;

[0090] Deep features: resolution 160×160, representing global context and harmonic patterns;

[0091] High frequency recovery module (HFR);

[0092] Input: shallow feature map (640×640×256);

[0093] Low-pass filter kernel generation: Generate the initial filter kernel through 3×3 convolution, and generate the normalized low-pass kernel (K) through Softmax activation. low );

[0094] Spectral inversion: Perform kernel inversion operation K high =IK low Where I is the unit kernel matrix, generating an adaptive high-pass filter;

[0095] High-frequency enhancement: Apply the high-pass kernel to the input feature map and calculate the formula F enhanced =F input *K high Where * represents the convolution operation;

[0096] Cross-channel attention fusion: Dynamically fuse original features and enhanced features through a gating mechanism to suppress optical noise interference;

[0097] Output: 640×640×256 feature map (after high-frequency details are restored).

[0098] In the implementation manner of the present application, in step S2, an adaptive filtering and sampling collaborative mechanism is used to suppress the distribution differences of target features in the scene. A spatially adaptive low-pass filter kernel is generated through dynamic light intensity perception, and local smoothing is performed on the high-level feature map to reduce the intra-class feature variance; combined with a multi-scale feature fusion strategy, the correlation between low-frequency main features and high-frequency details is retained through PixelShuffle upsampling; a cross-meter harmonic energy difference analysis model is simultaneously constructed, the frequency domain energy distribution differences of adjacent meter areas are compared, and the filter kernel weight allocation strategy is dynamically adjusted to suppress mismatch interference; finally, the adaptive sampling interval control and the feature alignment mechanism are collaboratively optimized to achieve a balanced control of target feature distribution consistency and detail retention in dense scenes.

[0099] In an optional embodiment, in step S2, the distribution differences of target features in the scene can be suppressed through the collaborative mechanism of adaptive filtering and sampling. Alternatively, a topological feature propagation framework based on a graph neural network can be constructed, and an adaptive neighborhood filter kernel group can be generated using the spatial distribution relationship of the electricity meters. The filtering weights can be dynamically allocated in combination with the node attention mechanism to achieve collaborative optimization of cross-meter feature correlation and intra-class consistency.

[0100] In another optional embodiment, in step S2, the adaptive filtering and sampling collaborative mechanism is used to suppress the distribution differences of target features in the scene. The adversarial generative network framework can also be introduced to jointly train the generator and discriminator models. The generator outputs the filter kernel parameters that uniformly distribute the features, and the discriminator constrains the consistency of the feature space with the actual meter distribution. At the same time, the contrastive learning strategy is combined to enhance the intra-class feature similarity measurement.

[0101] In the embodiment of the present application, step S2 suppresses the distribution difference of target features in the scene through the adaptive filtering and sampling collaborative mechanism, and further includes:

[0102] Intra-class smoothing module: such as Figure 2 The ICS module shown in (b) is designed to predict an adaptive low-pass filter to smooth high-level features, thereby alleviating the problem of feature inconsistency. At the same time, it upsamples high-level features to maintain feature details and consistency during the fusion process. Specifically:

[0103] First, the input feature map is processed by a 3×3 convolutional layer, followed by a Softmax operation to generate a spatially adaptive low-pass filter.

[0104] By applying the ICS module, a spatially-variant adaptive low-pass filter is generated, which effectively smooths high-level features and reduces feature inconsistency.

[0105] PixelShuffle is used to upsample the feature maps, enhancing the representation of fine details. This effectively alleviates the inconsistency of intra-class features, resulting in smoother and more consistent feature representations, thereby improving the accuracy of the model in detecting dense objects.

[0106] To address intra-class feature differences in densely populated metering scenarios (such as false detections caused by similar harmonic distributions), we improve detection consistency through spatial adaptive filtering and multi-scale feature alignment. The details are as follows:

[0107] Feature smoothing:

[0108] Generate a spatially adaptive low-pass filter kernel K on a high-level feature map (such as the output of a frequency-domain deep belief network) smooth , the formula is:

[0109] K smooth =Softmax(W conv *F in )

[0110] Where W conv To achieve learnable convolution parameters, the convolution layer is implemented as a 3×3 depthwise convolution + a 1×1 pointwise convolution, reducing the number of parameters while maintaining the receptive field. Softmax ensures kernel weight normalization to generate a locally adaptive low-pass kernel. In areas with dense electricity meters (such as distribution cabinets), ICS reduces intra-class feature variance by 30%, and evens out harmonic pattern differences caused by occlusion, keeping the difference in smoothing weight coefficients to less than 0.05.

[0111] Detail Preservation:

[0112] Use PixelShuffle upsampling (Formula: F up =PixelShuffle(F smooth )) Replaces traditional interpolation to reduce edge blur caused by downsampling. PixelShuffle converts channel dimensions into spatial resolution input C×H×W→output (C / r 2 )×rH×rW, the upsampling rate is r=2, and the resolution is increased from 160×160 to 320×320 using bilinear interpolation pre-calculated LUT acceleration.

[0113] Cross-meter feature matching:

[0114] By comparing the harmonic energy distribution differences between adjacent meters (referring to the cross-meter network characteristic matching algorithm), abnormal harmonic patterns (such as specific frequency offsets in electricity theft) can be identified.

[0115] Intra-class smoothing module (ICS):

[0116] Input: deep feature map (160×160×512)

[0117] Spatial adaptive filtering: Generate a low-pass filter kernel (K) through 3×3 convolution smooth ), the feature map is smoothed after Softmax normalization.

[0118] Detail-preserving upsampling: PixelShuffle operation (upsampling rate 2×2) is used to increase the resolution to 320×320, expressed as:

[0119] F up =PixelShuffle(F smooth )

[0120] Cross-meter feature alignment: Calculate the harmonic energy difference (HED) between adjacent meter areas to suppress mismatching;

[0121] Output: 320×320×256 feature map (after optimizing for intra-class consistency).

[0122] In the implementation mode of the present application, in step S3, features are dynamically adjusted based on convolution and compensation strategies to solve boundary blur and detail loss. A dual-branch offset prediction network is constructed through a deformable convolution layer to generate feature sampling direction offsets and amplitude coefficients respectively; the feature map sampling position is adjusted through a dynamic resampling mechanism, and the boundary area features are sharpened in combination with a bilinear interpolation function; Canny edge detection is simultaneously introduced to locate the boundary area of ​​the meter, and the edge features output by the high-frequency recovery module are superimposed to compensate for the detail degradation caused by excessive smoothing; further, a cross-modal feature fusion strategy is used to jointly optimize the frequency domain prior constraints and spatial domain semantic information, and dynamic weight distribution is used to suppress noise interference and enhance boundary positioning accuracy, ultimately achieving high-fidelity reconstruction and precise detection of meter features under complex lighting conditions.

[0123] In an optional embodiment, in step S3, features are dynamically adjusted based on convolution and compensation strategies to solve boundary blur and detail loss. A feature calibration network based on a self-attention mechanism is constructed, and a dynamic feature offset field is generated in combination with a local gradient sensitive function. Non-uniform sampling and feature sharpening of boundary sensitive areas are achieved through a multi-scale feature pyramid structure. At the same time, an adversarial edge generation module is introduced to enhance the reconstruction of high-frequency details.

[0124] In another optional embodiment, in step S3, features are dynamically adjusted based on convolution and compensation strategies to solve boundary blur and detail loss. The optical flow estimation network and frequency domain feature constraints are integrated, boundary motion trajectories are predicted based on the timing information of adjacent frames, and dynamic convolution kernel parameters and compensation coefficients are jointly optimized. The random noise interference of single-frame detection is suppressed through the spatiotemporal consistency loss function.

[0125] In the embodiment of the present application, step S3 dynamically adjusts features based on the convolution and compensation strategy to solve the boundary blur and detail loss, and further includes:

[0126] Dynamic adaptive resampling module: such as Figure 2 As shown in (c), the DAR module is designed to overcome the limitations of low-pass filters in smoothing features, especially in processing large areas of inconsistent features and preserving fine details or boundary features. While low-pass filters effectively smooth large inconsistent areas, they tend to blur thin areas and boundary regions, resulting in loss of edge information. To address this issue, DAR generates offset values ​​to guide the sampling direction towards feature areas with higher intra-category consistency.

[0127] This makes up for the shortcomings of ICS in handling large areas or border areas.

[0128] In this module, D' represents the offset direction and A' represents the offset scale, both of which are predicted by two 3×3 convolutional layers.

[0129] Resampling the feature map according to the offset enhances the consistency of features through DAR, sharpens boundary details, and improves detection accuracy in complex environments.

[0130] The ICS module optimizes the problem of over-smoothing boundary details, especially in handling inconsistent features in large areas and preserving fine details or boundary features. Feature alignment capabilities are enhanced through dynamic offset prediction and deformable convolution.

[0131] Offset prediction:

[0132] The dual-branch 3×3 convolutional layer predicts the feature offset direction D' and offset amplitude A' respectively:

[0133] D', A'=Conv dir (F in ),Conv amp (F in )

[0134] in Indicates the offset direction range limit of each pixel point ||D'|| ∞ ≤1, convolution followed by Tanh activation, To normalize the amplitude coefficient, the range is limited to A'=Sigmoid(A), and the output of the independent convolution branch is predicted by two 3×3 convolutional layers.

[0135] Dynamic resampling:

[0136] Adjust the feature sampling position according to the offset. The formula is:

[0137]

[0138] Where N(p) is the neighborhood of pixel p, defined within a 3×3 grid, using sparse sampling with deformable convolution, and δ(·) is a bilinear interpolation function.

[0139] Boundary compensation mechanism:

[0140] In the boundary area of ​​the meter (positioned by Canny edge detection F final =F out +λ·Canny(F HFR ), λ = 0.3 is a learnable parameter) and superimposes high-frequency recovery features (HFR output) to suppress over-smoothing.

[0141] Dynamic Adaptive Resampling Module (DAR):

[0142] Input: mid-level feature map (320×320×384) and ICS output features;

[0143] Offset prediction: dual-branch 3×3 convolution predicts the direction offset separately and amplitude coefficient

[0144] Boundary Compensated Resampling: Performs dynamic feature alignment: F out (p)=∑ q∈N(p) F in (q)·A'(q)·δ(pq-D'(q));

[0145] High-frequency feature superposition: HFR output features are superimposed on the boundary area located by Canny edge detection;

[0146] Output: 320×320×384 feature map (after edge sharpening).

[0147] In the embodiment of the present application, step S4 integrates the output features, processes in parallel, and optimizes the edge energy efficiency in combination with the adjustment strategy, including:

[0148] Multimodal feature fusion and detection output:

[0149] Cross-scale feature aggregation: HFR, ICS, and DAR output features are fused through a bidirectional feature pyramid (BiFPN)

[0150] Detection Head: It uses a decoupled head structure to predict bounding box regression based on CIoULoss optimized positioning.

[0151] FocalLoss handles class imbalance probability:

[0152] Non-maximum suppression (NMS): The threshold is set to 0.5 to filter overlapping detection boxes.

[0153] Hardware core collaborative workflow, such as Figure 3 As shown:

[0154] Hardware architecture composition:

[0155] This system adopts a three-level distributed hardware architecture (such as Figure 3 shown), including:

[0156] Perception layer:

[0157] High dynamic range CMOS camera (IMX686 sensor, dynamic range 120dB);

[0158] Multispectral photoelectric sensor array (covering the 400-1100nm band);

[0159] Ambient light intensity monitoring module (0-10000 lux measurement accuracy);

[0160] Edge computing layer:

[0161] Customized edge gateway (RK3588 SoC + Xilinx Artix-7 FPGA);

[0162] The FPGA acceleration module includes:

[0163] Frequency domain decomposition unit: realizes real-time FFT / IFFT conversion (<2ms delay);

[0164] Dynamic resampling accelerator: supports parallel calculation of offsets of DAR module;

[0165] Intra-class smoothing coprocessor: dedicated matrix convolution kernel (supports 5x5 ICS filtering);

[0166] Communication layer:

[0167] Dual-mode communication module (4G Cat.1+LoRaWAN);

[0168] Power protocol conversion gateway (supports Modbus-RTU / DL / T645 protocol conversion);

[0169] In the embodiment of the present application, step S5 integrates parameters and features for training, combines data sets and functions, and enhances the generalization ability of the model, including:

[0170] The perception layer synchronously collects meter images (30fps@1080P) and ambient light intensity parameters;

[0171] The raw data is transmitted to the edge gateway via the RS485 bus;

[0172] FPGA modules execute in parallel:

[0173] High-frequency recovery operation (5x5 kernel convolution acceleration of the HFR module);

[0174] Dynamic resampling coordinate offset calculation (gradient back propagation acceleration of DAR module);

[0175] Intra-class feature smoothing (accelerated by matrix normalization of the ICS module);

[0176] The test results are uploaded to the monitoring center through an encrypted channel;

[0177] The power management module implements dynamic voltage regulation (0.8-1.2V dynamic range);

[0178] Key performance parameters are shown in Table 1:

[0179] Table 1 Key performance parameters

[0180] Module Performance indicators Test conditions FPGA acceleration unit 128GOPS computing power 500MHz main frequency Image processing pipeline 22ms end-to-end latency 1080P input Environmental adaptability -20℃~70℃ operating temperature 95% humidity Energy efficiency ratio 3.8TOPS / W Typical workloads

[0181] This architecture achieves real-time computing of HFR / ICS / DAR modules through FPGA hardware acceleration, increasing the inference speed by 5.2 times compared to pure CPU solutions, meeting millisecond-level response requirements in industrial scenarios.

[0182] In summary, the present invention proposes a frequency-sensing detection method and system for low-voltage electric meters under complex lighting conditions, which suppresses optical noise interference in high-brightness environments through a dynamic frequency domain compensation mechanism, optimizes intra-class feature consistency in dense scenes by combining adaptive filtering and multi-scale sampling strategies, and improves boundary positioning accuracy based on deformable convolution and edge compensation mechanisms; constructs a heterogeneous hardware architecture to achieve parallel acceleration of frequency domain decomposition, dynamic resampling, and intra-class smoothing, and optimizes edge energy efficiency by combining dynamic voltage regulation strategies; integrates multimodal training data with a cross-domain transfer learning framework to enhance the model's generalization ability for complex lighting conditions. Through frequency domain feature decoupling, space-frequency domain joint optimization, hardware acceleration, and data-driven collaborative innovation, the technical bottlenecks of traditional methods in noise suppression, feature consistency, real-time performance, and generalization are systematically solved, providing a highly robust, low-power industrial-grade solution for power equipment detection in high-brightness scenes such as airports and commercial buildings, significantly improving the intelligence level and operation and maintenance efficiency of smart grid inspections.

[0183] Example 3, reference Figure 4 , is an embodiment of the invention, which provides a dual-high-characteristic power transient stability analysis method. In order to verify the beneficial effects of the present invention, scientific demonstration is carried out through economic benefit calculation and simulation experiments.

[0184] Model training:

[0185] Dataset: 1500 images of low-voltage electricity meters, covering strong light simulation and densely packed scenes, annotated with bounding boxes and category labels.

[0186] Backbone network: Improved based on YOLOv8-Tiny, replacing its native feature fusion module with the module of the present invention.

[0187] Loss function: CIoU Loss is used to optimize bounding box regression, and Focal Loss is used to solve the category imbalance problem.

[0188] Module parameter configuration:

[0189] HFR module: filter kernel size K'=5, softmax temperature coefficient 0.1.

[0190] ICS module: low-pass filter kernel size 3×3, PixelShuffle upsampling rate 2×2.

[0191] DAR module: The offset prediction branch has 64 output channels and uses ReLU as the activation function.

[0192] 3. Reasoning process:

[0193] The input image is passed through the backbone network to extract multi-scale features.

[0194] The features pass through the HFR (high frequency enhancement) → ICS (intra-class smoothing) → DAR (boundary optimization) modules in sequence.

[0195] The output layer generates detection boxes and category probabilities, and the non-maximum suppression (NMS) threshold is set to 0.5.

[0196] Hardware implementation:

[0197] It uses an edge computing gateway with an integrated FPGA-accelerated frequency domain decomposition module, supporting parallel processing of 10 electricity meter signals.

[0198] The photosensitive sensor uses the same technology as the variable frequency multi-connected high-precision pressure sensor, with a dynamic range of 120dB.

[0199] Software Implementation:

[0200] Develop a signal processing library based on the iterative DFT algorithm, compatible with the Modbus-RTU protocol, to achieve seamless integration with existing electricity meter systems.

[0201] A transfer learning framework for deploying deep learning models. The pre-training dataset contains time-frequency spectrograms of electricity meters under illumination levels of 2000-10000 lux.

[0202] Technical effect:

[0203] Improved detection accuracy: In strong light environments (>5000 lux), the fundamental frequency detection error is ≤0.02Hz, a 50% improvement over traditional DFT methods.

[0204] Energy efficiency optimization: Dynamic resource allocation strategies reduce computing load by 30%, reducing edge power consumption to below 2W.

[0205] Improved detection accuracy: AP increased by 2.2% in strong light environments.

[0206] Edge-preserving optimization: The HFR module restores high-frequency details, and the boundary intersection-over-union ratio (AP75) is improved to 42.7%.

[0207] Dense target separation: The ICS+DAR module reduces the false detection rate in dense scenes and reduces the intra-class feature differences by 30%.

[0208] In high-brightness scenarios at airports (8000 lux), the HFR+DAR combination module increases the meter digital area detection recall rate to 96.2%.

[0209] The ICS module reduces cross-meter mismatching incidents by 70%, meeting the needs of dense deployment in commercial buildings.

[0210] The overall power consumption at the edge is ≤2.5W, and the computing load is reduced by 45% compared to pure cloud solutions.

[0211] Cross-scenario adaptability: supports high-brightness and complex environments such as airports and commercial buildings, with a false alarm rate of less than 0.5%.

[0212] Check visual performance:

[0213] Visual inspection results are as follows Figure 4 As shown in the figure, the proposed method achieves the most accurate detection results, closely matching the ground truth. In contrast, the other two representative methods suffer from missed detections, false detections, and inaccurate bounding box placement. Thanks to the ICS and DAR modules, the consistency of intra-category features is enhanced, significantly reducing missed detections and false positives. Furthermore, thanks to the HFR module, the model captures clear instrument edge information even in high-glare environments, enabling more precise bounding box positioning.

[0214] In Example 4, the above is a schematic scheme of a complex illumination low-voltage electricity meter frequency detection method. It should be noted that the technical solution of the complex illumination low-voltage electricity meter frequency detection system and the technical solution of the complex illumination low-voltage electricity meter frequency detection method described above are based on the same concept. For details not described in detail in the technical solution of the complex illumination low-voltage electricity meter frequency detection system in this embodiment, please refer to the description of the technical solution of the complex illumination low-voltage electricity meter frequency detection method described above.

[0215] This embodiment also provides a complex illumination low-voltage electric meter frequency sensing detection system, including:

[0216] High-frequency restoration module, used to compensate for high-frequency detail loss and enhance image edge and texture information through adaptive high-pass filter generation and kernel inversion mechanism;

[0217] The intra-class smoothing module optimizes the intra-class feature consistency of dense meter scenarios through spatially adaptive low-pass filtering and PixelShuffle upsampling;

[0218] Dynamic adaptive resampling module, which achieves feature alignment based on deformable convolution and boundary compensation mechanism, suppressing the over-smoothing effect of the ICS module;

[0219] Multimodal feature fusion module, which uses bidirectional feature pyramid (BiFPN) to fuse HFR, ICS, and DAR output features;

[0220] The edge computing module includes an FPGA-accelerated frequency domain decomposition unit, a dynamic resampling accelerator, and an intra-class smoothing coprocessor.

[0221] This embodiment also provides an electronic device suitable for the case of complex illumination low-voltage electricity meter frequency detection, including: a memory and a processor; the memory is used to store computer-executable instructions, and the processor is used to execute computer-executable instructions to implement the complex illumination low-voltage electricity meter frequency detection method proposed in the above embodiment.

[0222] This embodiment further provides a storage medium storing a computer program, which, when executed by a processor, implements the method for frequency sensing detection of a low-voltage electric meter under complex illumination as proposed in the above embodiment.

[0223] The storage medium proposed in this embodiment and the method for realizing frequency sensing detection of complex illumination low-voltage electric meters proposed in the above embodiment belong to the same inventive concept. The technical details not fully described in this embodiment can be referred to the above embodiment, and this embodiment has the same beneficial effects as the above embodiment.

[0224] Through the above description of the implementation methods, those skilled in the art can clearly understand that the present invention can be implemented with the help of software and necessary general hardware, and of course can also be implemented by hardware. Based on this understanding, the technical solution of the present invention is essentially or the part that contributes to the prior art can be embodied in the form of a software product, which can be stored in a computer-readable storage medium, such as a computer floppy disk, read-only memory (ROM), random access memory (RAM), flash memory (FLASH), hard disk or optical disk, etc., including a number of instructions for enabling a computer device (which can be a personal computer, server, or network device, etc.) to execute the methods of various embodiments of the present invention.

[0225] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit the present invention. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that the technical solutions of the present invention may be modified or replaced by equivalents without departing from the spirit and scope of the technical solutions of the present invention, which should all be included in the scope of the claims of the present invention.

Claims

1. A frequency sensing detection method for low-voltage electric meters under complex illumination, characterized in that: include: Through the filter generation inversion mechanism, compensation enhancement and suppression under complex lighting conditions can be achieved; Through the adaptive filtering and sampling synergy mechanism, the distribution differences of target features in the scene are suppressed; Dynamically adjust features based on convolution and compensation strategies to solve boundary blur and detail loss; Fusion of output features, parallel processing, and adjustment strategies to optimize edge energy efficiency; Fusion parameters and features for training, combining data sets and functions to enhance model generalization capabilities.

2. The complex illumination low-voltage electric meter frequency sensing detection method according to claim 1, characterized in that: The filter generates an inversion mechanism to achieve compensation enhancement and suppression under complex lighting conditions, including: Dynamically generate filters and perform inverse compensation to achieve joint enhancement and suppression; Build a framework to distribute interference and signals by weights; Leverage dynamic resource adaptation to optimize suppression boundaries.

3. The complex illumination low-voltage electric meter frequency sensing detection method according to claim 2, characterized in that: Through the adaptive filtering and sampling synergy mechanism, the differences in target feature distribution in the scene are suppressed, including: Dynamically generate filter kernels and optimize intra-class feature smoothing and distribution consistency through feature fusion strategies; Build a collaborative framework, combined with a detail-preserving mechanism; Perform feature energy difference analysis to match intra-class interference.

4. The complex illumination low-voltage electric meter frequency sensing detection method according to claim 3, characterized in that: Dynamically adjust features based on convolution and compensation strategies to solve boundary blur and detail loss, including: Establish a feature alignment strategy to achieve feature sharpening in boundary areas through prediction and control optimization; Construct a feature compensation mechanism to suppress detail degradation through feature superposition and weight distribution; Perform joint feature optimization to combine constraints and spatial domain semantic information to enhance boundary localization accuracy.

5. The complex illumination low-voltage electric meter frequency sensing detection method according to claim 4, characterized in that: Fusion of output features, parallel processing, and adjustment strategies are combined to optimize edge energy efficiency, including: Utilize feature fusion to achieve feature collaborative optimization and calculation through dynamic weight allocation; Perform parallel processing and collaboratively perform frequency domain decomposition, feature resampling, and smoothing tasks; Implement dynamic adjustment strategies, combining computing resource awareness with dynamic voltage adaptation to optimize edge energy efficiency.

6. The complex illumination low-voltage electric meter frequency sensing detection method according to claim 5, characterized in that: Fusion parameters and features for training, combining datasets and functions, and enhancing model generalization capabilities, including: Construct a multimodal training dataset covering dynamic lighting conditions, integrating time-frequency spectrum features with spatial domain image data; Design a composite loss function to collaboratively optimize positioning accuracy and category distribution balance, and suppress overfitting and sample imbalance through dynamic parameter weighting; Implement cross-modal transfer learning strategies, enhance the adaptability to complex scenarios by migrating pre-trained model parameters, and achieve generalized mapping from low-frequency data to high-frequency features.

7. The complex illumination low-voltage electric meter frequency sensing detection method according to claim 6, characterized in that: The fusion parameters and features are trained, including: Construct a joint time-frequency-spatial training dataset covering multiple illumination intensities, and achieve collaborative representation of high-frequency noise and low-frequency semantics through cross-modal feature fusion; Design a dynamic weighted composite loss function, combining positioning error and category distribution entropy to optimize the model parameter update direction; A cross-modal parameter transfer strategy is implemented to constrain the spatial domain detection network parameter initialization based on pre-training spectral features.

8. A complex illumination low-voltage electric meter frequency sensing detection system, applying the method according to any one of claims 1 to 7, characterized in that: include: High-frequency restoration module, used to compensate for high-frequency detail loss and enhance image edge and texture information through adaptive high-pass filter generation and kernel inversion mechanism; The intra-class smoothing module optimizes the intra-class feature consistency of dense meter scenarios through spatially adaptive low-pass filtering and PixelShuffle upsampling; Dynamic adaptive resampling module, which achieves feature alignment based on deformable convolution and boundary compensation mechanism, suppressing the over-smoothing effect of the ICS module; Multimodal feature fusion module, which uses a bidirectional feature pyramid BiFPN to fuse HFR, ICS, and DAR output features; The edge computing module includes an FPGA-accelerated frequency domain decomposition unit, a dynamic resampling accelerator, and an intra-class smoothing coprocessor.

9. An electronic device comprising: memory and processor; The memory is used to store computer-executable instructions, and the processor is used to execute the computer-executable instructions. When the computer-executable instructions are executed by the processor, the steps of the complex light-irradiated low-voltage meter frequency sensing detection method according to any one of claims 1 to 7 are implemented.

10. A computer-readable storage medium storing computer-executable instructions, wherein the computer-executable instructions, when executed by a processor, implement the steps of the complex illumination low-voltage meter frequency sensing detection method according to any one of claims 1 to 7.

Images