CN-121744006-B - Fill many trouble intelligent detection of electric pile system based on improve YoloV11

CN121744006BCN 121744006 BCN121744006 BCN 121744006BCN-121744006-B

Abstract

The invention discloses a charging pile multi-fault intelligent detection system based on an improvement YoloV (compact disc) 11, which relates to the technical field of intelligent operation and maintenance of charging piles, wherein an input layer acquires four-mode original data, a preprocessing module respectively processes the four-mode original data to output a unified size feature map, a main network performs feature extraction and enhancement through a lightweight cross-channel convolution unit and a dynamic local feature module, a REPRESNET-18 module is additionally arranged for visible light mode data to perform feature enhancement, a cross-mode dynamic gating enhancement fault association area is used, a neck network realizes cross-mode multi-level fusion and multi-scale layering, multi-scale features of a detection head are detected, boundary frame prediction, fault category prediction and confidence prediction are performed, and finally a prediction result is output through an output layer. The invention realizes breakthrough in detection integrity, robustness and efficiency, can be directly applied to intelligent monitoring of the edge end of the charging pile, and obviously reduces operation and maintenance cost.

Inventors

LI XINZHI
LIU QINGXUE
WANG XIA
JIA YOUDONG
SONG KE
SHEN FUYU
Yuan Zhangmin

Assignees

昆明学院

Dates

Publication Date: 20260508
Application Date: 20260226

Claims (8)

1. The intelligent charging pile multi-fault detection system based on the improvement YoloV comprises an input layer, a preprocessing module, a main network, a neck network, a detection head and an output layer, and is characterized in that the input layer acquires four-mode original data through an ultraviolet imaging unit, a visible light imaging unit, an infrared thermal imaging unit and a vibration sensing unit, and the preprocessing module respectively processes the four-mode original data to output a unified size characteristic diagram; The main network convolves ultraviolet data, infrared data and vibration data by a lightweight cross-channel convolution unit to obtain characteristic distribution of different modes, enhances small target details and edge characteristics by a dynamic local characteristic module, and adds REPRESNET-18 modules to visible light mode data to perform characteristic enhancement; the neck network sequentially realizes the cross-mode multi-level fusion and multi-scale layering through an early fusion module, a mid-stage level characteristic distillation module, a multi-scale cross-mode attention module and a cross-mode pyramid attention module to finally obtain multi-scale cross-mode characteristics; The detection head carries out boundary frame prediction, fault class prediction and confidence prediction based on multi-scale characteristics output by the neck network, and finally outputs a prediction result through an output layer; the boundary box prediction specifically comprises the following steps: Bounding box parameters: Wherein, the Is the center coordinates of the bounding box; Is wide and high for the bounding box; is the bounding box rotation angle; The loss function is Wherein, the Is a prediction bounding box; Is a true bounding box; is the distance intersection ratio of the two points, , Is the cross-over ratio of the two adjacent layers, Is the square of the euclidean distance at the center of the two frames, The square of the diagonal line of the minimum circumscribed rectangle of the two frames; for the predicted/true angle.
2. The intelligent detection system for multiple faults of a charging pile based on a modification YoloV as claimed in claim 1, wherein the vibration sensing unit comprises a vibration sensor arranged on the surface of a charging pile main controller module, a vibration sensor arranged at a charging gun interface and a distributed vibration sensor distributed along a cable.
3. The intelligent detection system for multiple faults of the charging pile based on the improvement YoloV as set forth in claim 1, wherein the specific processing flow of the preprocessing module on the four-mode raw data is as follows: For the data collected by the ultraviolet imaging unit, extracting a corona discharge candidate feature point set by adopting an ORB algorithm, and carrying out Gaussian interpolation on the feature point set to output a corona pseudo image; removing environmental noise from the time domain vibration signals acquired by the vibration sensing unit through self-adaptive band-pass filtering, and applying short-time Fourier transform to output a vibration time-frequency diagram; for thermal imaging data acquired by an infrared thermal imaging unit, eliminating temperature measurement deviation through a dynamic temperature compensation algorithm, converting the temperature measurement deviation into visual characteristics through pseudo-color mapping, and outputting a hot spot characteristic diagram; performing a mild gaussian denoising of the image acquired by the visible light imaging unit does not change the original structural features.
4. The improved YoloV-11-based multi-fault intelligent detection system for charging piles of claim 1, wherein the lightweight cross-channel convolution unit design comprises a double-branch structure of a main branch and an inexpensive branch, the main branch is used for channel dimension reduction and feature extraction, the inexpensive branch is used for space detail enhancement, and the double-branch features are spliced according to channel dimensions through a fusion layer; The dynamic local feature module comprises a dynamic depth separable convolution, a lightweight Swin attention and a feature fusion module, wherein the dynamic depth separable convolution dynamically generates convolution kernel parameters according to input features and self-adaptively adjusts convolution kernel weights, the lightweight Swin attention reduces the calculation complexity of global attention through window division, long-distance dependent modeling in a local area is kept, a channel and space aligned double-branch feature is obtained through a dimension alignment module, and the feature fusion module integrates the double-branch feature through element-by-element multiplication to strengthen a fault sensitive area.
5. The intelligent detection system for the multi-fault of the charging pile based on the improvement YoloV is characterized in that a dual-mode regulation mechanism is adopted in the cross-mode dynamic gating, global mode weight calculation is firstly carried out, global trend of each mode is extracted, four-mode weight prediction is carried out, local space attention generation is secondly carried out, local fault characteristics of each mode are focused, and a fault area map is generated.
6. The intelligent detection system for multiple faults of charging piles based on improvement YoloV as claimed in claim 1 is characterized in that the REPRESNET-18 module adopts a multi-branch parallel mode in a training stage, simultaneously operates a dimension-reducing branch, a main convolution branch and an identity mapping branch, combines the results of the three branches by element-by-element addition, outputs the results through normalization and ReLU activation, combines the multiple branches of the training stage into a single 3×3 convolution by a reparameterization technology in an inference stage, and outputs the characteristics before activation.
7. The intelligent detection system for multiple faults of charging piles based on improvement YoloV as claimed in claim 1, wherein said early fusion module judges importance of each mode based on global features, dynamically adjusts importance weights of each mode, generates spatial attention masks in modes through a spatial attention enhancing unit, fuses global mode weights, local spatial masks and own features of each mode data to output early fusion results ; The mid-stage characteristic distillation module allows for the distillation loss Fine grain characteristic information of a multi-mode middle layer is learned, characteristic expression of small faults is enhanced, and distillation enhancement characteristics are output ; The multi-scale cross-mode attention module takes a single-scale feature of middle-term fusion as input, converts the middle-term fusion feature into three-scale features of small fault, middle fault and large fault through scale adaptation and downsampling, four mode heads in each scale work in parallel, attention weight dynamically deviates to an advantage mode of the scale, and three scale feature maps of deep fusion are output; The cross-modal pyramid attention module is based on YoloV feature pyramid architecture, embeds a cross-modal attention mechanism, distributes three scale feature graphs fused in depth to corresponding pyramid levels, realizes feature fusion among different levels through up/down sampling, and selectively fuses the advantage features of each mode according to fault scale characteristics.
8. The charging pile multi-fault intelligent detection system based on the improvement YoloV11 is characterized in that pyramid level design in the cross-mode pyramid attention module is specifically that a P3 layer is used for small fault fine granularity detection, a P4 layer is used for medium fault balance detection, a P5 layer is used for large fault global detection, a P6 layer is used for fault correlation layer, a P5 layer and a P4 layer are fused, fault correlation characteristics are output, a P7 layer is used for global context layer, a P6 layer and a P5 layer are fused, and global context is output.

Description

Fill many trouble intelligent detection of electric pile system based on improve YoloV11 Technical Field The invention relates to the technical field of intelligent operation and maintenance of charging piles, in particular to a charging pile multi-fault intelligent detection system based on an improvement YoloV. Background Along with the popularization of electric automobiles, the charging pile is used as a key infrastructure, the operation reliability and the safety of the charging pile are of great importance, and the accurate detection of potential faults is of great significance to the prevention of accidents, the reduction of operation and maintenance cost and the improvement of user experience. At present, the mainstream fault detection technology relies on electrical parameter monitoring of a preset threshold value or single vision sensor image analysis, is difficult to adapt to complex deployment environment and fine operation and maintenance requirements, and has three significant defects: First, the single-mode detection has inherent limitation, and the hidden fault omission rate is high. The existing scheme adopts an isolated sensing mode, lacks multi-dimensional cross validation and complementation, has insufficient identification capability on compound type and early-stage slow-changing faults, and has potential safety hazards. Secondly, the environment anti-interference capability is weak, and the scene generalization performance is poor. The existing algorithm depends on limited scene training data, and the performance is reduced when facing to dynamic interference of different regions, seasons and weather, and the false detection rate is high. Third, architecture redundancy results in significant power consumption and difficulty in adapting edge deployments. The traditional scheme needs to independently deploy a plurality of special models, has serious internal consumption of resources, not only increases extra energy consumption and hardware cost, but also prevents the large-scale application of the stock charging piles and the low-power consumption new piles. Disclosure of Invention In order to overcome the problems in the prior art, the invention provides a charging pile multi-fault intelligent detection system based on an improvement YoloV. The technical scheme adopted for solving the technical problems is that the charging pile multi-fault intelligent detection system based on the improvement YoloV comprises an input layer, a preprocessing module, a main network, a neck network, a detection head and an output layer, wherein the input layer acquires four-mode original data through an ultraviolet imaging unit, a visible light imaging unit, an infrared thermal imaging unit and a vibration sensing unit, and the preprocessing module respectively processes the four-mode original data to output a unified size characteristic diagram; The main network convolves ultraviolet data, infrared data and vibration data by a lightweight cross-channel convolution unit to obtain characteristic distribution of different modes, enhances small target details and edge characteristics by a dynamic local characteristic module, and adds REPRESNET-18 modules to visible light mode data to perform characteristic enhancement; the neck network sequentially realizes the cross-mode multi-level fusion and multi-scale layering through an early fusion module, a mid-stage level characteristic distillation module, a multi-scale cross-mode attention module and a cross-mode pyramid attention module to finally obtain multi-scale cross-mode characteristics; and the detection head carries out boundary frame prediction, fault category prediction and confidence prediction based on the multi-scale characteristics output by the neck network, and finally outputs a prediction result through an output layer. The intelligent detection system for the multiple faults of the charging pile based on the improvement YoloV is characterized in that the vibration sensing unit comprises a vibration sensor arranged on the surface of a main controller module of the charging pile, a vibration sensor arranged at an interface of a charging gun and a distributed vibration sensor distributed along a cable. According to the charging pile multi-fault intelligent detection system based on the improvement YoloV11, the specific processing flow of the preprocessing module on the four-mode original data is as follows: For the data collected by the ultraviolet imaging unit, extracting a corona discharge candidate feature point set by adopting an ORB algorithm, and carrying out Gaussian interpolation on the feature point set to output a corona pseudo image; removing environmental noise from the time domain vibration signals acquired by the vibration sensing unit through self-adaptive band-pass filtering, and applying short-time Fourier transform to output a vibration time-frequency diagram; for thermal imaging data acquired by an infrared thermal imaging unit, e