Abstract:In order to reduce the false detection rate of the electric inspection robot in identifying the pointer meter in the transformer substation and improve the accuracy of meter reading identification, a pointer meter detection method based on deep learning is proposed. By adding a residual module to the YOLOv4-tiny network to improve the robustness of the model and improvements to the Hourglass network, precise identification of pointer meter readings is achieved. In order to verify the effectiveness of the proposed method, the method is tested with the image data of the transformer substation and the test results are compared with other methods. The experimental results show that the missing rate of the proposed approach is 1. 25%, the localization accuracy is less than 1. 125%, the overall detection time was less than 0. 5 s. Compare with Hough line detection with ORB or U-NET, the average error of reading recognition is reduced by 70. 8% and 58. 8%. The method provides new ideas for meter reading identification of transformer substations.

Abstract:In this paper, the YOLOv5 target detection algorithm was used as the base algorithm to improve accuracy for PCB defect detection. Firstly, an appropriate data augmentation method is selected through experiments. For the problem of small PCB defect size, the P2 detection head was added to the original three detection heads. A new PANet multi-feature fusion structure was designed to realize efficient two-way cross-scale connection and weighted feature layer fusion. For the problem of the complex PCB background, the CBAM attention module is introduced to enhance image information, the Transformer module is introduced to enhance the algorithm’s ability to capture PCB defect information at different locations. Finally, through these improvements, the mAP accuracy of the algorithm increased by 11. 3% while the FPS droped by only 7. 2.

Abstract:Based on the development demand of intelligent water area monitoring system, a personnel recognition algorithm for monitored water area is proposed. After data collection of the water area scene, data cleaning and labeling, a personnel category dataset YZ-Water4 under the monitored water area scene was independently constructed, with a total of 8 092 images and 24 011 tags. Based on the performance of the object detection algorithm YOLO v7 and the characteristics of the water area scene, object detection algorithm YOLOWA (you only look once-water area) for water environment is proposed. First, the FReLU activation function which is proposed for visual tasks is used to replace the activation function in YOLO v7 algorithm. Secondly, the attention mechanism is integrated into algorithm to improve the feature extraction ability of the algorithm. Finally, SIOU loss function is chosen to replace CIOU loss function in YOLO v7 algorithm to optimize the training process. The experimental results show that compared with the original algorithm, YOLO-WA has increased the precision rate from 82. 3% to 86. 9%, recall rate from 92. 0% to 92. 8%, mean average precision from 88. 4% to 90. 6%, and the processing speed is 85 frame per second, meeting the accuracy and speed requirements of real-time run.

Abstract:In order to meet the requirements of high detection accuracy, fast inspection speed and easy to be embedded in mobile devices, a lightweight object detection structure DE-YOLO is designed for mobile terminal devices. Firstly, combining depth-separable convolution, point-by-point convolution and ECA attention mechanism, the feature extraction module NewC3 is proposed, which is responsible for significantly reducing network parameters and strengthening the ability of network to extract effective insulator information. Then, the lightweight module DC-SE is designed with the help of channel number multiplication strategy and channel attention mechanism SE. It is used to weaken the interference of complex background to insulator fault, extract the subtle features of insulator complementary, and then enhance the extraction ability of target feature information of shallow network. Experiments show that the GFLOPs of DE-YOLO network on the expanded Homemade insulator data dataset are reduced by 45%, the running parameters are reduced by 42%, and the detection accuracy of self-exploding defects is up to 93. 2%. NewC3 and DC-SE can ensure the lightweight of DE-YOLO and meet the requirements of real-time detection of self-exploding insulator defects.

Abstract:To address the problems of large defect size span, similar characteristics, difficulty in recognition of small targets, and missed objects in chip defect detection, an improved method based on YOLOv5 is proposed. To solve missed and false detection of small targets, we presented a new small target feature detector ( S-Detector) to improve the learning capability of the model. For the large defect size span and similar characteristics, efficient attention feature pyramid networks (EA-FPNs) with highly active focus learning ability are proposed to improve the ability to detect different sizes of defects. The bounding box fusion algorithm (BFA) is developed to reduce the redundant boxes and time overhead in prediction. The experimental results show that the detection accuracy of this method is enhanced by 1. 2% and the accuracy of minor target defects is improved by 1. 6%; while using BFA to eliminate the redundant boxes, the detection time of a single image is 26. 8 μs, which is decreased by 5. 2 μs before BFA. The proposed method has good performance and efficiency in chip defect detection. Keywords:chip defect detection; deep l

Abstract:Electric shovel is a large mechanical excavation equipment widely used in surface mining. During the excavation process, the prolonged direct impact of the shovel teeth against the ore can cause the shovel teeth to loosen or even break prematurely, resulting in unplanned downtime and lost productivity of the shovel. To solve this problem, an electric shovel tooth break detection method based on the improved YOLOX is proposed. This method is based on YOLOX. Firstly, for the problem of poor detection effect caused by uneven illumination, the dilated convolution attention mechanism is added to the feature pyramid network to enhance the saliency of the target in the complex background; Secondly, the corner efficient intersection over union(CEIOU) loss function is used to replace the original network loss function to optimize the network training process, thereby improving the detection accuracy of the target; Finally, considering the computing power of the embedded device itself, the model compression strategy is used to tailor the redundant channels in the network to reduce the model volume and improve the detection speed. The performance test is carried out on the self-built 4 200 WK- 10 electric shovel data set. The experimental results show that compared with the YOLOX network model, the average detection accuracy of the improved model reaches 95. 37%, which is 1. 95% higher, the detection speed is 46. 1 fps, an increase of 8. 4 fps, and the model size is 31. 74 MB, which is reduced to 32. 9% of the original. Compared with many other existing methods, the designed target detection algorithm has the advantages of high precision, small size and fast speed.

Abstract:In order to solve the problem of false detection and missing detection in new-energy vehicle battery collector disk due to disarranged target defect distribution, large size span and fuzzy features, a YOLOv5 method based on multi-scale deformations convolution (YOLOv5s-4Scale-DCN) was proposed for defect detection of vehicle battery collector disk. Firstly, for defect targets of different scales, a new detection layer is added based on the YOLOv5 model. By capturing defect features of different scales and integrating semantic features of different depths, the detection rate of defect targets of different scales is improved. Secondly, deformable convolution is introduced to enlarge the receptive field of the feature map, which makes the extracted feature discrimination stronger and effectively improves the defect recognition ability of the model. Experimental results show that the proposed YOLOv5s-4Scale-DCN algorithm can effectively detect the defects of new-energy vehicle battery collection panel, with mAP up to 91%, 2. 5% higher than that of the original algorithm, and the FPS reaches 113. 6. There are two types of defects, severe defects and uncovered defects. The detection and recall rate reached 100%, meeting the requirements of real-time detection of the defects of the battery collecting disk of new energy vehicles.

Abstract:To address the low accuracy of deep learning based fault diagnosis under distributed scenarios that caused by limited sample of single node and unbalanced distribution of working conditions of multiple nodes, et al, a multi-wavelet coefficients enhanced dynamic aggregation federal deep network ( MWCE-FedDWA) is proposed for fault diagnosis under multiple conditions with distributed small samples. A framework for fault diagnosis using MWCE-FedDWA is proposed, wavelet coefficient features are extracted by each terminal node from its local samples, a method based on multi-wavelet coefficient fusion in deep network is proposed for feature enhancement, each local model utilizes a set of diversified wavelet coefficients to extract more discriminative fault features. A global federal deep network model is constructed in aggregation node by aggregating the local models from multiple terminal nodes, and then adopted for fault diagnosis under multiple conditions. To reduce the influence of non-independent and identically distributed data among multiple nodes, a federated dynamic weighted aggregation algorithm is proposed to balance the contribution of local models. The results on bearing vibration data show that the proposed method can achieve high-precision diagnosis under multiple conditions with distributed small samples.

Abstract:In order to realize the effective detection and localization of the unexploded ordnance, an electromagnetic detection system with small transmitting coil and array receiving coils has been developed. The primary field interference is removed by using the eccentric configuration between the transmitting coil and array receiving coils. Because of the array receiving coils, the detection efficiency of the system is effectively improved. The horizontal location algorithm based on differential comparison of the response of array receiving coils is proposed, and the feasibility is verified by experiments. The research results show that the detection depth of the detection system for the typical annular metal cylinder reaches 1 m, and the location error between the calculated value and the actual value is 3. 76 cm. The system and method are able to realize the accurate detection and location of the unexploded ordnance.

Abstract:A squirrel search algorithm with adaptive mutation chaos (AMCSSA) is proposed in this paper to improve the learning rate of LSTM and the prediction pattern of its decay factor. This is to solve issues of the tendency of local optimal for optimized parameters and the decay of LSTM prediction efficiency during the optimization of LSTM with SSA. By calculating its time complexity, AMCSSA is proved to be able to increase the searching efficiency without increasing the complexity of the algorithm. AMCSSA adopts Chebyshev chaotic map to generate the chaotic initial population, and switches to use a nonlinear decreasing for the predator probability. The positional greedy selection strategy is used to continuously update and keep individuals of more advantages during the iteration of algorithms, then the adaptive T mutant is introduced to improve SSA's exploration capabilities in searching space. AMCSSA optimizes the parameters for the learning rate of LSTM and their decay factors, thus the predictive ability of LSTM is further improved. This is verified by experiments on the remaining useful life of rolling bearing. The results show that the prediction accuracy of LSTM optimized by AMCSSA increased by 1. 05%, 7. 61%, 8. 4%, and 7. 73%, respectively, compared to those optimized by traditional SSA, particle swarm optimization (PSO), bat algorithm (BAT), and firefly algorithm (FA). With the proposed algorithm, the number of iteration required for the optimized LSTM to complete the convergence is also reduced, so that the prediction efficiency is increased.

Abstract:The use of transient electromagnetic method and harmonious electromagnetic method instrument to observe the underground space data and perform fusion processing can improve the detection imaging ability, but most of the existing transient electromagnetic method and harmonious electromagnetic method instruments are independent sets and must be equipped with multiple sets to separate collection. A small size, large bandwidth and high sensitivity electromagnetism transceiver coil that can meet the use of both transient / harmonic modes is designed to lay the foundation for the development of portable, shallow and high resolution transient / harmonic integrated electromagnetic detection system. First, the equivalent circuit model of the air-core coil sensor is established, and the sensitivity and noise characteristics of the coil-core sensor are analyzed, the test results of the air-core sensor show that at 1 125 Hz, the sensitivity is 23. 12 mV/ nT, the equivalent magnetic noise is 9. 27 pT / √Hz and the bandwidth is 1 Hz~ 100 kHz. Then, the formula and finite element analysis software are used to calculate the transceiver coupling relation when the sensor coils are placed eccentrically, and the optimal eccentric location of the combination of transceiver coil is determined, the calculation and simulation results differ from actual location by 0. 21% and 0. 17% respectively, which effectively eliminated the launch magnetic field (primary field). Finally, the time-domain and frequency-domain electromagnetic method systems are constructed, and the test results show that the designed transient / harmonic integrated transceiver coil can meet the needs of time-domain and frequency-domain electromagnetic detection modes.

Abstract:Network communication performance will be affected when permanent fault occurs, most of the existing fault tolerant methods use re-routing strategies, the uncertainty of the bypass selection can bring longer delays or even form hotspots around the failed node leading to deadlocks. In this paper, a new router architecture, DRRA, is proposed to address various failure cases in 2D mesh NoC. The different input and output ports are connected by the added components and three different connect methods are defined, when packets encounter a faulty node, it will select the appropriate method to directly bypass the faulty node based on the specific fault location and routing information to ensure the connectivity of the network. The experimental results show that compared with other fault-tolerant schemes, the proposed method has good performance and reliability in the presence of multiple failed nodes in the network and does not bring excessive hardware overhead. In the hotspot traffic mode, the proposed scheme in this paper can reduce the average packet transmission delay by 57. 4% compared with the ReRS scheme and 38. 9% compared with MiCoF.

Abstract:In view of the problems of small test size space, poor low-frequency response, and internal and environmental reflection affecting measurement accuracy when the existing specular monocone TEM cell is used as the calibration of pulsed electric field sensors, the standard device of pulsed electric field based on ultra-large low-reflection specular monocone is designed and developed on the basis of simulation optimization analysis by using ultra-large precision single-cone structure, terminal impedance loading and low-reflection special-shaped anechoic chamber. The generatrix of the monocone is 2 m long and can adapt to pulse signals with a pulse width greater than 5 ns, and the amplitude of the standard pulse electric field generated by the device is 50~ 150 V/ m, the rise time of the pulse is less than 100 ps, the pulse width is greater than 5 ns. This standard device meets the needs of large space, wide frequency band and high accuracy calibration equipped with test pulse electric field sensors.

Abstract:In order to solve the problem that the filtering accuracy is poor or even divergent when the inertial navigation system is affected by model errors and measurement outlier errors, a multi-sensor fusion algorithm based on square-root cubature Kalman filter(SRCKF)wdetection is proposed. The square-root cubature Kalman filter innovation sequence is adaptively adjusted by the covariance matching method, and the adjusted innovation will compensate the measurement noise variance matrix and reduce the influence of model error, then use the adjusted innovation to detect the error and improve the w- detection accuracy. And construct the observation value replacement criterion to replace the error observation value, so as to solve the influence of the measurement abnormal value error. Finally, the attitude fusion is carried out by using SRCKF, the attitude of the gyroscope is taken as the state equation, and the attitude of the accelerometer and magnetometer replaced by detection is taken as the measurement equation. The experimental results show that the proposed algorithm can accurately estimate the system attitude, and the average solution accuracy can be improved by 62. 43% compared with the traditional algorithm. Under different conditions, the overall performance of the algorithm can be greatly improved, and the attitude can be solved quickly to ensure the solution accuracy.

Abstract:According to the current trend of miniaturization and broadband development of wireless communication systems, a broadband dipole with multiple resonant modes is designed. First of all, the resonant modes of the inversed T-shaped conductor are analyzed by the theory of characteristic mode and the first two resonant modes are chosen; then a parasitic resonant dipole element is introduced without affecting the resonant characteristics of the inversed T-shaped conductor; finally, a lumped port is placed at the place that the three resonant modes have common maximum current value, these three resonant modes are excited simultaneously, and a wideband resonant antenna with three resonant modes is obtained. The simulated and measured results are in good agreement, with the final operating bandwidth of the antenna ranging from 1. 8 to 2. 56 GHz (34. 9% relative bandwidth). As a result, the designed antenna is not only simple in structure, but also good in performance and can be widely used in wireless communication systems.

Abstract:The optical fiber pressure sensor needs temperature compensation because its performance is greatly affected by temperature. To solve this problem, a software compensation scheme combining grey wolf optimization and least squares support vector machine (GWO-LSSVM) algorithm is proposed. The penalty factor ζ and kernel parameter σ of least squares support vector machine are iteratively optimized by grey wolf optimization algorithm within the specified range to construct the compensation algorithm model. In different temperature situations, the input and output data of the sensor are measured by calibration test and are divided into test set and training set. By taking the root mean square error which is calculated from the predicted values of the test set as the fitness function, the temperature compensation problem is transformed into a convex quadratic optimization problem with constraints. The results show that compared with previous compensation, the sensitivity temperature coefficient of the fiber optic pressure sensor after temperature compensation is increased from 9. 405 × 10 -3 / ℃ to 1. 201 6 × 10 -4 / ℃ , and the relative value of the additional temperature error is increased from 28. 215% to 0. 481%. The temperature stability of the sensor is greatly improved.

Abstract:In the electric vehicle wireless power transfer ( WPT) system, the magnetic shielding effect is often at the expense of transmission efficiency. Reducing magnetic leakage and improving transmission efficiency are the difficult problems. Therefore, a new strong coupling magnetic shielding structure is proposed to reduce the magnetic leakage of the system. Firstly, a strong coupling magnetic shielding coil structure is proposed. Based on the proposed shielding coil structure, the characteristics of system magnetic leakage and transmission efficiency on the impedance of shielding coil are analyzed. Secondly, a magnetic flux leakage optimization method is given. Using the proposed optimization method, the coil parameters that meet the design requirements are obtained. Finally, according to the obtained coil parameters, a wireless charging system based on electric vehicle with magnetic shielding structure is developed. The effectiveness of the proposed structure and method is verified by simulation and experiment. The results show that the proposed new strong coupling magnetic shielding coil not only effectively reduces the magnetic induction intensity of the system by 28%, but also improves the transmission efficiency by nearly 4%.

Abstract:In order to solve the problem of inaccurate measurement results due to environmental factors while measuring the sediment concentration of rivers with the capacitance method, a fusion model based on Kalman filtering and LSTM (Kalman-LSTM) is proposed. Firstly, the Kalman filtering is used for filtering to reduce the random error of sensor measurement. Then, the LSTM was used to integrate multi-sensor data of sediment content information and environmental content information, to reduce the influence of environmental factors on sediment content measurement by the capacitance method. Finally, a Kalman-LSTM fusion model for measuring sediment concentration by capacitance method was developed. To verify the fusion effect of the Kalman-LSTM fusion model, the root mean square error, maximum absolute error, average absolute error, and average relative error of each model are compared with the BP model, the RBF model, and the LSTM model. The experimental results show that the average relative error of the Kalman-LSTM fusion model is 2. 54% and the root mean square error is 2. 47 kg / m 3 . The fusion model can effectively reduce the influence of environmental factors on sediment concentration measurement and improve the accuracy of sediment concentration measurement by the capacitance method.

Abstract:GNSS / INS integrated navigation system has been widely used. However, the loosely coupled system will degrade when the GNSS system is subject to environmental constraints. Meanwhile, the tightly coupled system of GNSS / INS will significantly decrease the positioning accuracy due to the frequent changes in constellation geometry layout. To solve this problem, the adaptive Kalman filter tight integrated navigation method based on new information is used in this paper to design a vehicle-mounted experiment scheme in an urban environment. It takes the errors of the pseudo-range and pseudo-range rate as the observation quantity to perform filtering estimation. At the same time, the adaptive factor is determined according to the innovation variance’s theoretical and actual values, and the system’s measurement noise is adjusted to correct the system errors further. Experimental results show that when the number of visible satellites is less than four, the 3D positioning error of this method is reduced by about 24. 2% compared with the traditional Kalman filtering method. In the environment of frequent changes in visible satellite conditions, the maximum positioning error is reduced by about 60%. The results showed that this method could adjust the observation noise of satellite pseudo-distance when the observation conditions are poor, reduce the state error, enhance the robustness of the filter, and provide more accurate position information.

Abstract:In milling machining engineering, chattering is generated, which seriously affects the machining accuracy and surface quality of products. In order to effectively avoid chattering during milling, a milling chattering monitoring and identification method based on Adaptive chirp mode pursuit (ACMP) is proposed. The method integrates the bandwidth and faint characteristics of vibration signals, and ACMP captures the signal modes one by one in a recursive framework. In this algorithm, we do not need to input the number of signal modes, but can learn them by evaluating the energy of the residual signal, so that we can avoid the problems of modal mixing or over-decomposition due to the uncertainty of the number of decomposition layers. Firstly, the algorithm is verified to have high recognition accuracy for chattering signals using simulated signals; then the method is demonstrated to be effective in recognizing chattering in time based on field milling experimental data; finally, the power spectrum entropy value is extracted from the ACMP processed signals as chattering recognition features. This method solves the modal mixing and pseudo-component problems of empirical mode decomposition (EMD) algorithm, and reduces the influence of unstable accuracy of variational mode decomposition (VMD), which can accurately and quickly identify the chattering and is of great significance to improve the machining quality.

Abstract:In order to solve the problem that the accuracy of life prediction model of multi-component equipment decreases due to the difference in the distribution of degraded data under different working conditions, a domain feature fusion network (DFF-Net) which can adapt to different working conditions is proposed in this paper. Firstly, the degraded data of different working conditions were input into the feature extraction network to obtain the cross-working conditions characteristics. Then, the domain feature fusion network (DFF-Net) was used to adjust the cross-working conditions characteristics. Finally, the adjusted data was input into the life prediction model to output the life prediction results of the equipment under different working conditions. Tests on public data sets show that the MAE and RMSE of the predicted results of the proposed model on the test set decrease by 6. 5% and 7. 4%, respectively, compared with the lifetime prediction model without adding the domain feature fusion network, which indicates that the proposed model can effectively improve the accuracy of cross-working condition equipment life prediction.

Abstract:QCA is regarded as one of the new technologies to replace CMOS due to its advantages of short delay time, low power consumption, and small footprint. In response to the high power consumption, capacitance parasitic, and crosstalk issues caused by the decreasing size of CMOS devices, this paper constructs a recursive box filter using QCA technology for the first time. Among them, a new QCA full adder has been proposed, which reduces the circuit area by 55% compared to the proposed QCA full adder. 56. 7% fewer cells were used. Quantum costs have also been reduced by more than 10%. Based on this, an efficient RCA and an efficient CSA were designed to form the addition unit of the box filter. The box filter constructed based on this saves 32. 6% of hardware resources compared to general QCA box filters. Reduce circuit running time by 20%. Reduced power consumption by 48. 7%. And simulated using QCA Designer, the results show that this design can completely replace the traditional box filter function, and significantly reduce efficiency, power consumption, circuit area, and resource occupation.

Abstract:In order to meet the requirement of ultra-high repetition frequency pulse generation in current ultra-fast pulse generation technology, this paper proposes a high repetition frequency reconfigurable pulse generation method based on solid state semiconductor switching technology, pulse power synthesis technology and picosecond high precision delay trigger control technology, and develops a prototype device of high repetition frequency reconfigurable pulse generator with kV/ MHz. The test results show that the pulse generator can stably output sub-nanosecond pulses with amplitude 1 kV, repetition frequency 1 MHz, rise time 700 ps. By precisely adjusting the trigger time sequence of each pulse, the rise time, pulse width and waveform can be flexibly changed, and the repetition frequency can be adjusted from 1 MHz to 50 MHz. It can effectively meet the application requirements of electronic countermeasures, ultra-wideband detection and communication industries.

Abstract:In order to improve the convergence speed and positioning accuracy of the pressure gas leakage source localization method based on acoustic detection under strong noise interference, this paper proposes a wideband direction finding sparse representation model (W-SpSF) based on weighted subspace fitting (WSF) criterion. The model is combined with the bilateral correlation transform algorithm (TCT), and the covariance matrix at the focusing frequency is obtained by focusing operation. As the sparse recovery data, the DOA positioning is finally realized and the location of the leakage source is obtained. In positioning process, weighted subspace fitting is used to reduce the sensitivity of noise. In this paper, the algorithm is tested in the simulation and laboratory simulation environment respectively. The simulation results show that under the condition of adding strong noise interference, the operation speed of the algorithm is increased by 50% and the positioning error is reduced by 20% compared with the similar algorithm. In the measured environment, this paper builds a sound source positioning system of 8-element acoustic sensor array, which can realize positioning for the simulated leakage of gas cylinders, and the leakage source positioning speed is fast, which proves the feasibility of the algorithm in the actual environment.

Abstract:In order to improve the accuracy of robot scene recognition in complex indoor environments, this paper proposes a robot scene recognition model that fuses convolutional neural network (CNN) and visual Transformer structure. The model uses CNN to extract local features of the scene. And the visual Transformer structure is used to capture the distant dependencies in the features. The proposed visual Transformer structure consists of three parts, they are a feature encoding structure (Attention Embedding), an Encoder structure, and a structure that converts high-level semantic features into pixel-level features (Attention Project). The robot scene recognition model studied in this paper uses CNN to improve the description ability of local detail features of the visual Transformer. Furthermore, the visual Transformer helps CNN to construct the dependencies of distant features, which can effectively characterize and utilize the visual features of the robot working scene images. Finally, the effectiveness of the model is verified by experimenting with the dataset collected by the robot in the actual working environment and the open source COLD dataset. The scene recognition accuracy of our model is higher.

Abstract:Nowadays, the combined attitude measurement method of inertial sensor and dual antenna RTK ( real time kinematic) is widely used in the field of vehicle attitude measurement. This method integrates the characteristics of rapid update rate of inertial sensor and the advantages of high accuracy of dual antenna RTK. When encountering high buildings or trees, the satellite signal will be lost, resulting in the dual antenna RTK unable to provide the available heading angle, and the inertial sensor cannot maintain the heading angle accuracy for a long time. This paper attempts to adopt an information fusion method based on micro inertial measurement unit miniature inertial measurement unit(MIMU), magnetic sensor and dual antenna RTK. The experimental analysis of the actual sports car is carried out. It is concluded that this method can ensure the stable output of high-precision vehicle attitude data by introducing magnetic sensors to provide heading information under the condition of satellite signal loss.

Abstract:Aiming at the problem that the existing insulator detection algorithm cannot detect insulators and their defects in an oriented manner, an aerial insulator identification and defect detection method improved by YOLOv5 algorithm is proposed. By orienting the aerial insulator pictures, the aerial insulator dataset and defective insulator dataset are formed. The lightweight attention mechanism module is introduced in the backbone feature extraction network of YOLOv5, and the improved spatial pyramid pooling structure is used in the feature fusion stage. By improving the head structure of the YOLOv5 network, the network can perform directional identification of insulators and add angular loss classification to the loss function. The experimental results show that under the premise that the detection time does not increase significantly from 0. 044 s to 0. 049 s per sheet, the value of mAP (mean average precision) on the test set of the improved algorithm is 95. 00%, which realizes directional identification of insulators and their leakage cap defects, and can also be applied to insulator video stream detection. This provides a good basis for the subsequent precise positioning of insulators and further fault detection.