A Comprehensive Method and System for Detecting Electrical Tools Based on Industrial Robots

Abstract

This application provides a method and system for the full-process inspection of electrical tools based on industrial robots, relating to the field of industrial robot technology. The method includes: transporting the electrical tool to be tested to a pre-inspection station according to the inspection task; collecting multi-dimensional physical features of the electrical tool at the pre-inspection station; inputting the multi-dimensional physical features into an autoencoder model to extract the corresponding low-dimensional latent feature vectors and determine the cluster category of the electrical tool; querying the detection parameter set from a detection parameter knowledge graph based on the cluster category; moving the electrical tool to be tested to an automated testing platform using a ground-rail robotic arm; applying high voltage for testing according to the test voltage; and classifying and returning the electrical tool to be tested to a smart shelf based on the test results. This method achieves personalized adaptation of detection parameters and full-process automation.

Description

Technical Field

[0001] This invention relates to the field of industrial robot technology, and in particular to a method and system for the whole-process testing of electrical tools based on industrial robots. Background Technology

[0002] In the daily operation and maintenance of power systems, electrical tools are the first line of defense for ensuring the personal safety of workers and the safe operation of power grid equipment. According to relevant industry regulations, these tools must undergo regular preventive testing to verify whether their electrical insulation performance and mechanical strength meet safety standards.

[0003] However, with the continuous expansion of the power grid and the constant improvement of safety production standards, the demand for testing power tools has experienced explosive growth. Currently, most power testing institutions in China still mainly rely on the traditional manual testing mode: testing personnel manually move the tools to be tested, select appropriate test parameters based on experience, manually connect high-voltage test lines, and manually record and sort the data after the test is completed.

[0004] This traditional model is facing increasingly severe challenges in dealing with the current massive testing tasks: on the one hand, manual operation has subjective differences, making it difficult to ensure a high degree of uniformity in testing standards; on the other hand, the high-pressure testing environment poses a potential safety threat to operators; in addition, facing the testing needs of tens of thousands of tools and instruments, relying solely on manpower is no longer sufficient to meet the actual operational requirements of high frequency and high efficiency.

[0005] Therefore, traditional detection methods are inadequate in terms of reliability, response speed, and intelligence, and new technologies are urgently needed to improve quality and efficiency. Summary of the Invention

[0006] This invention addresses the technical problems in the existing technology where the inspection of electrical tools has long relied on manual operation, resulting in low efficiency when facing massive inspection demands, inconsistent implementation of inspection standards, and serious safety hazards caused by manual close-range wiring under high-voltage testing environments. It provides a full-process inspection method and system for electrical tools based on industrial robots.

[0007] The technical solution of the present invention to solve the above-mentioned technical problems is as follows: In a first aspect, the present invention provides a method for the whole-process testing of electrical tools based on industrial robots, including: According to the testing task, the automated guided vehicle is dispatched to move the electrical equipment to be tested from the smart shelf to the pre-inspection station; The multidimensional physical characteristics of the electrical equipment under test are collected at the pre-inspection station. These multidimensional physical characteristics include geometric features, material features, and surface features. The multidimensional physical features are input into a pre-trained autoencoder model to extract the corresponding low-dimensional latent feature vectors. The clustering category of the electrical equipment under test is determined based on the low-dimensional latent feature vector, and the detection parameter set is obtained by querying the detection parameter knowledge graph according to the clustering category. The detection parameter set includes insulation distance, test voltage and fixture type. The ground rail robotic arm grasps the electrical tool under test according to the type of clamp and moves it to the automated test platform. The vision sensor at the end of the ground rail robotic arm guides the ground rail robotic arm to complete the automatic docking between the test high voltage line and the electrode of the electrical tool under test. After docking, the ground rail robotic arm is controlled to retreat to a safe area according to the insulation distance. Then, a high voltage is applied according to the test voltage to carry out the test. Based on the test results, the automated guided vehicle is dispatched to classify and return the electrical tools under test to the smart shelf.

[0008] Secondly, this invention provides a full-process electrical tool testing system based on industrial robots, comprising: The outbound scheduling module is used to schedule automated guided vehicles to move the electrical tools to be tested from the smart shelf to the pre-inspection station according to the testing task; The feature acquisition module is used to acquire the multi-dimensional physical features of the electrical tool under test at the pre-inspection station. The multi-dimensional physical features include geometric features, material features, and surface features. The feature encoding module is used to input the multidimensional physical features into a pre-trained autoencoder model and extract the corresponding low-dimensional latent feature vectors. The parameter matching module is used to determine the cluster category of the electrical appliance under test based on the low-dimensional potential feature vector, and to query the detection parameter set from the detection parameter knowledge graph according to the cluster category. The detection parameter set includes insulation distance, test voltage and fixture type. An automatic testing module is used for a ground-rail robotic arm to grasp the electrical appliance under test according to the type of clamp and move it to an automated testing platform. The vision sensor at the end of the ground-rail robotic arm guides the robotic arm to complete the automatic docking of the test high-voltage line with the electrode of the electrical appliance under test. After docking, the ground-rail robotic arm is controlled to retreat to a safe area according to the insulation distance. Then, a high voltage is applied according to the test voltage to perform the test. The inbound scheduling module is used to schedule the automated guided vehicle to classify and return the electrical tools under test to the smart shelf based on the test results.

[0009] The beneficial effects of this invention are: Compared to existing technologies, this application first uses an automated guided vehicle to transport the electrical equipment under test from the smart shelf to the pre-inspection station, and collects the geometric, material, and surface features of the electrical equipment under test at the pre-inspection station, providing a comprehensive data foundation for subsequent intelligent identification. Then, the multi-dimensional physical features are input into a pre-trained autoencoder model to extract low-dimensional latent feature vectors. Based on these feature vectors, the cluster category of the electrical equipment under test is determined, and the insulation distance, test voltage, and fixture type matching the category are obtained from the detection parameter knowledge graph. This achieves intelligent and accurate mapping from the physical properties of the equipment to the detection parameters, solving the problem of poor parameter adaptability caused by traditional methods that only rely on fixed table lookups based on voltage levels.

[0010] Based on this, the ground-rail robotic arm automatically grabs the electrical tools to be tested according to the type of gripper and moves them to the automated testing platform. The end vision sensor guides the test high-voltage line to automatically connect with the electrode of the tool. After the connection is completed, the ground-rail robotic arm is controlled to retreat to a safe area according to the insulation distance. Then, the automated testing platform applies high voltage according to the test voltage to conduct the test. Finally, according to the test results, the automated guided vehicle sorts and sends the tools back to the smart shelf, forming an unmanned closed-loop detection system covering the entire process of handling, identification, matching, testing and sorting.

[0011] Through the above technical solutions, this application improves testing efficiency and safety while achieving personalized adaptation of testing parameters and full-process automation, effectively solving the prominent problems of low efficiency, inconsistent standards, and significant safety hazards in traditional manual testing modes. Attached Figure Description

[0012] Figure 1 A flowchart illustrating the end-to-end electrical tool testing method based on industrial robots provided by this invention; Figure 2 This is a schematic diagram of the structure of the full-process electrical tool testing system based on industrial robots provided by the present invention.

[0013] In the attached diagram, the components represented by each number are as follows: Outbound scheduling module 11, feature acquisition module 12, feature encoding module 13, parameter matching module 14, automatic testing module 15, and inbound scheduling module 16. Detailed Implementation

[0014] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0015] In the description of this invention, the terms "first" and "second" are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of indicated technical features. Thus, a feature defined as "first" or "second" may explicitly or implicitly include one or more of the stated features. In the description of this invention, "a plurality of" means two or more, unless otherwise explicitly specified.

[0016] In the description of this invention, the term "for example" is used to mean "used as an example, illustration, or description." Any embodiment described as "for example" in this invention is not necessarily to be construed as being more preferred or advantageous than other embodiments. The following description is provided to enable any person skilled in the art to make and use the invention. Details are set forth in the following description for purposes of explanation. It should be understood that those skilled in the art will recognize that the invention can be made without using these specific details. In other instances, well-known structures and processes will not be described in detail to avoid obscuring the description of the invention with unnecessary detail. Therefore, the invention is not intended to be limited to the embodiments shown, but is consistent with the broadest scope of the principles and features disclosed herein.

[0017] Example 1, as Figure 1 As shown, this embodiment of the invention provides a full-process electrical tool testing method based on industrial robots, including: S10: Based on the testing task, dispatch automated guided vehicles to move the electrical tools to be tested from the smart shelf to the pre-inspection station.

[0018] In this embodiment, at the electrical equipment testing center, the electrical equipment to be tested is typically stored in smart shelves, categorized by type. Each piece of electrical equipment is equipped with a unique electronic tag to identify its identity. Based on the testing task, a task instruction is generated and a dispatch instruction is sent to an Automated Guided Vehicle (AGV) via the Industrial Internet of Things (IIoT). An idle AGV is then dispatched to the designated smart shelf location. The AGV identifies the electronic tag of the electrical equipment to be tested on the smart shelf using its onboard barcode reader, picks the equipment from the shelf, and transports it to the pre-inspection station. The pre-inspection station is equipped with a dedicated positioning marker; upon arrival, the AGV automatically aligns and places the electrical equipment to be tested, completing the transport task.

[0019] Among them, automated guided vehicles refer to industrial mobile equipment with autonomous navigation and obstacle avoidance functions, which can realize automatic material handling; intelligent shelves are intelligent storage devices used for classifying and storing electrical tools, such as classifying and storing insulating gloves, voltage detectors, grounding wires, etc., and with location positioning and inventory management functions.

[0020] For example, if the inspection task is "routine inspection of 10kV insulating gloves", the automated guided vehicle is dispatched to pick up the insulating gloves with the number JG-2024001 from the No. 2 storage location on the 3rd floor of the smart shelf and transport them to the pre-inspection station along the preset navigation path.

[0021] S20: Collect multi-dimensional physical characteristics of the electrical tool under test at the pre-inspection station. The multi-dimensional physical characteristics include geometric features, material features, and surface features.

[0022] In the testing of electrical tools, the physical properties of the tool under test, such as its geometric dimensions, material composition, and surface condition, are the core basis for determining its insulation performance, mechanical strength, and the required testing parameters.

[0023] However, traditional detection methods usually rely on manual visual inspection or single sensor measurement, which has problems such as limited data dimensions, inconsistent collection standards, and difficulty in quantitative analysis, and cannot provide accurate and multi-dimensional data support for subsequent intelligent identification and parameter matching.

[0024] To address the aforementioned issues, this application collects multidimensional physical features of the electrical equipment under test at the pre-inspection station. These multidimensional physical features include geometric features, material features, and surface features. This lays a reliable data foundation for subsequent intelligent feature extraction, clustering classification, and adaptive matching of detection parameters based on an autoencoder model.

[0025] Specifically, step S20 in the method includes: The electrical tool under test is scanned by a three-dimensional laser profilometer to obtain three-dimensional point cloud data of the electrical tool under test, and geometric features are extracted from the three-dimensional point cloud data, wherein the geometric features include at least length, diameter, taper and curvature. The surface spectral information of the electrical appliance under test is acquired by a hyperspectral camera, and material features are extracted from the surface spectral information, wherein the material features include at least the infrared spectral absorption peaks of insulating materials; The surface image of the electrical appliance under test is acquired by a vision camera, and surface features are extracted from the surface image, wherein the surface features include at least cracks, scratches and signs of aging.

[0026] In this embodiment, a three-dimensional laser profilometer is first used to scan the electrical tool under test to obtain its three-dimensional point cloud data, and then geometric features are extracted from the three-dimensional point cloud data. Specifically, the three-dimensional laser profilometer scans the surface of the electrical tool under test by emitting a laser beam, receives the reflected light signal, and generates the three-dimensional point cloud data. The three-dimensional point cloud data is a set of discrete point coordinates on the surface of the electrical tool under test in space. By performing point cloud registration, feature extraction, and other processing on the three-dimensional point cloud data, the geometric features of the electrical tool under test can be calculated. These geometric features include at least length, diameter, taper, and curvature: length refers to the dimension of the electrical tool under test along its main axis; diameter refers to the width of the cross-section of the electrical tool under test, which is particularly important for rod-shaped tools such as insulating rods; taper refers to the rate of change of the diameter of the electrical tool under test along its length, used to determine whether the tool is uniform; curvature refers to the degree to which the main axis of the electrical tool deviates from an ideal straight line, and excessive curvature will affect insulation performance.

[0027] For example, for an insulating rod, the three-dimensional point cloud data obtained after scanning by a three-dimensional laser profilometer can be processed to obtain: a length of 2.5 meters, a diameter of 35 millimeters, a taper of 0.02, and a curvature of 1.5 millimeters.

[0028] Secondly, the surface spectral information of the electrical appliance under test is acquired using a hyperspectral camera, and material characteristics are extracted from this information. These material characteristics include at least the infrared absorption peaks of the insulating material. Specifically, the hyperspectral camera can acquire spectral information from the surface of the electrical appliance within a continuous spectral band. Each pixel corresponds to a continuous spectral curve. Furthermore, different insulating materials exhibit specific absorption characteristics for different wavelengths of light, forming characteristic absorption peaks. By analyzing the acquired surface spectral information, identifying the position and intensity of the infrared absorption peaks, and comparing them with a standard material spectral library, the material type and aging degree of the electrical appliance under test can be determined.

[0029] Among them, the infrared absorption peak refers to the position where a material strongly absorbs infrared light of a specific wavelength within the infrared band. The position and intensity of the absorption peak are different for different materials, and it can be used as a fingerprint feature to identify the material.

[0030] For example, a hyperspectral camera is used to collect the surface spectral information of the insulating glove. If a characteristic absorption peak of silicone rubber appears at a wavelength of 3.4 μm and the absorption peak intensity meets the standard, it can be determined that the material of the insulating glove is qualified silicone rubber. If the absorption peak intensity weakens or the peak position shifts, it indicates that the material is aging.

[0031] Finally, a visual camera is used to acquire surface images of the electrical equipment under test, and surface features are extracted from these images. Specifically, the visual camera acquires surface images of the electrical equipment under test, and then image processing algorithms are used to analyze the surface images and extract surface features. These surface features include at least cracks, scratches, and signs of aging: cracks refer to linear fracture lines appearing on the surface of the equipment, which may penetrate into the material; scratches refer to linear damage to the surface of the equipment caused by friction or impact; signs of aging refer to surface discoloration, powdering, cracking, etc., caused by long-term use or environmental factors. These surface features directly affect the insulation performance and mechanical strength of the equipment and are important criteria for determining whether the equipment needs to be scrapped or repaired.

[0032] For example, the Canny edge detection algorithm combined with morphological operations can be used to analyze surface images to extract crack features. The specific implementation steps are as follows: First, convert the surface image to a grayscale image, then perform noise reduction using a Gaussian filter (e.g., kernel size 5×5); next, use the Canny edge detection algorithm, setting a low threshold of 50 and a high threshold of 150, an aperture size of 3 for the Sobel operator, and using the L2 norm for gradient calculation; after edge detection, a binary edge image is obtained, and then the broken edge segments are connected using morphological closing operations (e.g., using 5×5 elliptical structuring elements); finally, linear regions with a length greater than a preset threshold (e.g., 20 pixels) are filtered out through connected component analysis and marked as cracks. For scratch detection, a method combining directional filtering and morphological gradients can be used, employing a Gabor filter bank to enhance linear features in a specific direction, where the wavelength is set to 8 pixels and the direction to 0°, 30°, 60°, 90°, 120°, and 150°, and then the scratch region is extracted through threshold segmentation. For signs of aging, gray-level co-occurrence matrix analysis can be used to analyze texture features. The distance parameter is set to 1 pixel, and the angle parameters are 0°, 45°, 90°, and 135°. Texture feature values ​​such as contrast, energy, and entropy are calculated and compared with the feature values ​​of standard aging samples to determine the degree of aging.

[0033] For example, analysis of the surface image of an insulating glove revealed a crack approximately 15 millimeters long in the palm area, along with several minor scratches and localized discoloration and aging marks on the surface.

[0034] In summary, compared to existing technologies, this application collects multi-dimensional physical characteristics of the electrical tool under test at the pre-inspection station. These multi-dimensional physical characteristics include geometric features, material features, and surface features. This achieves comprehensive digitization and precise quantification of the tool's physical properties, providing a high-quality, multi-dimensional data foundation for subsequent steps.

[0035] S30: Input the multidimensional physical features into a pre-trained autoencoder model to extract the corresponding low-dimensional latent feature vectors.

[0036] In the testing of electrical tools and equipment, the multidimensional physical features acquired by multispectral sensor arrays are characterized by high dimensionality, diverse dimensions, and information redundancy. Directly using the original high-dimensional features for cluster analysis and parameter matching not only results in high computational complexity but is also susceptible to noise interference and feature redundancy, leading to poor clustering results and decreased matching accuracy.

[0037] To address the aforementioned issues, this application utilizes the nonlinear dimensionality reduction capability of its encoder portion to input the multidimensional physical features into a pre-trained autoencoder model, extracting the corresponding low-dimensional latent feature vectors. This process removes redundant information while preserving the essential physical properties of the tools, providing a compact and efficient feature representation for subsequent accurate clustering category determination and detection parameter matching.

[0038] Specifically, step S30 in the method includes: Collect a multi-dimensional physical feature sample set of historical electrical tools, wherein each multi-dimensional physical feature sample of historical electrical tools includes geometric features, material features and surface features. The historical electrical tools include tools with known insulation distance labels, tools with known voltage level labels and tools without labels. Normalize each sample in the multidimensional physical feature sample set to obtain a normalized multidimensional physical feature sample set, and save the statistical parameters used in the normalization process. Construct an autoencoder neural network, wherein the autoencoder neural network includes an encoder structure and a decoder structure; The normalized multidimensional physical feature sample set is input into the autoencoder neural network, and the autoencoder neural network is trained with the goal of minimizing the comprehensive loss function. The comprehensive loss function is a weighted sum of reconstruction loss, insulation distance proximity loss and voltage level classification loss. After training is complete, the network parameters of the encoder structure are saved to obtain the pre-trained autoencoder model.

[0039] In this embodiment, a multi-dimensional physical feature sample set of historical electrical tools is first collected. Each historical electrical tool's multi-dimensional physical feature sample includes geometric features, material features, and surface features. These historical electrical tools include tools with known insulation distance labels, known voltage level labels, and unlabeled tools. Specifically, the collection scope of the multi-dimensional physical feature sample set can cover electrical tools of different types, specifications, and usage states, such as insulating gloves, insulating rods, voltage detectors, and grounding wires. Each multi-dimensional physical feature sample is collected with complete geometric features, material features, and surface features according to the method in step S20. Tools with known insulation distance labels and known voltage level labels are used to construct the insulation distance proximity loss and voltage level classification loss during the training process; unlabeled tools are used for unsupervised training to improve the model's feature extraction capability and reduce dependence on labeled samples.

[0040] For example, 1,000 historical electrical tools samples were collected, of which 600 samples were labeled with insulation distance tags, such as 0.2m, 0.3m, etc., and voltage level tags, such as 10kV, 35kV, etc., and 400 samples were unlabeled samples, which together constituted a multi-dimensional physical feature sample set.

[0041] Secondly, each sample in the multidimensional physical feature sample set is normalized to obtain a normalized multidimensional physical feature sample set, and the statistical parameters used in the normalization process are saved. Specifically, the normalization process can use the Z-score normalization method to calculate the mean and standard deviation of each feature dimension (e.g., length, infrared spectral absorption peak) in the multidimensional physical feature sample set, converting the feature value of each sample into a value under a standard normal distribution, eliminating the influence of dimensions, and ensuring the stability of model training.

[0042] The statistical parameters include the mean (μ) and standard deviation (σ) of each feature dimension. Saving the statistical parameters is to use the same standard as the training samples when normalizing the features of the electrical equipment under test in the future, so as to avoid the normalization bias affecting the model output.

[0043] For example, the "length" feature in the multidimensional physical feature sample set is normalized, and the mean (μ) of the length of all samples is calculated to be 1.0m and the standard deviation (σ) is 0.2m. The length value of each sample is converted according to the formula (length of sample - μ) / σ to obtain the normalized length, and the two statistical parameters of mean 1.0m and standard deviation 0.2m are saved.

[0044] Next, an autoencoder neural network is constructed, which includes an encoder structure and a decoder structure. Specifically, the autoencoder neural network adopts a deep learning architecture. The encoder structure is responsible for reducing the dimensionality of high-dimensional input features to low-dimensional latent feature vectors, and the decoder structure is responsible for reconstructing the low-dimensional latent feature vectors into high-dimensional vectors with the same dimension as the input. Through the collaborative training of the encoder and decoder, accurate compression and reconstruction of high-dimensional features are achieved.

[0045] For example, the parameter settings for the encoder structure can be as follows: The encoder structure consists of multiple fully connected layers, with the number of neurons in the input layer matching the dimension of the high-dimensional input feature vector, for example, 128 dimensions; subsequently, three hidden layers are set, with the number of neurons being 64, 32, and 16 respectively, and each layer uses a linear rectified function as the activation function; the bottleneck layer (i.e., the low-dimensional latent feature vector layer) has 8 neurons, uses a linear activation function, and does not perform nonlinear transformations to ensure the linear interpretability of the features. Batch normalization is used between layers to accelerate training convergence, and a Dropout layer (e.g., a dropout rate of 0.2) is added to prevent overfitting.

[0046] For example, the parameter settings of the decoder structure can be referenced as follows: the decoder structure is symmetrical to the encoder structure. Starting from the 8-dimensional input of the bottleneck layer, three hidden layers are set with the number of neurons being 16, 32, and 64 respectively. Each layer uses a linear rectified function as the activation function. The number of neurons in the output layer is consistent with the dimension of the input layer (e.g., 128 dimensions in this example). The Sigmoid activation function is used to map the output value to the [0,1] interval so as to calculate the reconstruction error with the normalized input features.

[0047] Furthermore, the normalized multidimensional physical feature sample set is input into the autoencoder neural network, and the network is trained with the goal of minimizing the comprehensive loss function. The comprehensive loss function is a weighted sum of the reconstruction loss, insulation distance proximity loss, and voltage level classification loss. Specifically, during training, the normalized multidimensional physical feature sample set is input into the autoencoder neural network. The encoder outputs a low-dimensional latent feature vector, the decoder reconstructs this vector into a high-dimensional vector, the error between the model output and the real samples is calculated using the comprehensive loss function, and the model parameters are updated using a gradient descent algorithm until the loss function converges.

[0048] The comprehensive loss function is a weighted sum of three losses, and the weights can be dynamically set according to training requirements. For example, the weight of the reconstruction loss can be set to 0.5, the weight of the insulation distance proximity loss to 0.3, and the weight of the voltage level classification loss to 0.2. Those skilled in the art can dynamically adjust the specific values ​​of the weights according to the actual application scenario. Specifically, the reconstruction loss measures the difference between the reconstructed vector output by the decoder structure and the original input sample, ensuring that the low-dimensional latent feature vector can retain the core information of the original features; the insulation distance proximity loss constrains samples with similar insulation distances to be close in the latent feature space, so that the low-dimensional latent feature vector contains semantic information about insulation distance; and the voltage level classification loss improves the model's ability to recognize voltage levels, ensuring that the low-dimensional latent feature vector can effectively reflect the voltage level differences of tools and equipment.

[0049] For example, the training process of an autoencoder neural network can refer to the following configuration: The Adam optimizer is used for parameter updates, with an initial learning rate set to 0.001. An exponential decay strategy is used for the learning rate, multiplying by 0.9 every 20 training epochs. The training epochs are set to 200, and the batch size is set to 64. During training, the normalized multidimensional physical feature sample set is randomly divided into a training set and a validation set in a 7:3 ratio. After each training epoch, the comprehensive loss function value is calculated using the validation set. Training is stopped early when the validation set loss no longer decreases for 10 consecutive epochs to prevent overfitting. Model parameters are initialized using the Xavier initialization method, and batch normalization layers are added between hidden layers to accelerate convergence.

[0050] Finally, after training, the network parameters of the encoder structure are saved, resulting in a pre-trained autoencoder model. Specifically, during training, after each training epoch, the comprehensive loss function value is calculated using the validation set. When the comprehensive loss function value of the validation set no longer decreases for 10 consecutive epochs, and the change in the comprehensive loss function value of the training set is less than 0.001, the model training is considered to have reached convergence, and training is stopped. Furthermore, since subsequent practical applications only require the encoder structure to extract low-dimensional latent feature vectors from the electrical equipment under test, without the need for a decoder structure for data reconstruction, only the network parameters of the encoder structure are saved, while the parameters of the decoder structure are discarded, forming the pre-trained autoencoder model. The pre-trained autoencoder model can be directly deployed in the detection system for low-dimensional latent feature vector extraction.

[0051] Specifically, the multidimensional physical features are input into a pre-trained autoencoder model to extract the corresponding low-dimensional latent feature vectors, including: Using the saved statistical parameters, the geometric features, material features, and surface features are normalized and then concatenated to form a high-dimensional input feature vector; The high-dimensional input feature vector is input into a pre-trained autoencoder model. Forward propagation is performed through the encoder part of the autoencoder model to output the activation value of the bottleneck layer of the encoder part. The activation value is used as a low-dimensional latent feature vector.

[0052] In this embodiment, the geometric features, material features, and surface features are first normalized using stored statistical parameters and then concatenated to form a high-dimensional input feature vector. Specifically, since the geometric features, material features, and surface features have different dimensions and numerical ranges, direct concatenation would lead to unstable model training. Therefore, it is necessary to use the statistical parameters stored during the model training phase in the aforementioned steps for normalization. These statistical parameters include the mean and standard deviation of each feature dimension, which are calculated from a multi-dimensional physical feature sample set of historical power equipment during the autoencoder model training phase. The normalized geometric features, material features, and surface features are then concatenated in a fixed order to form a one-dimensional high-dimensional input feature vector, the dimension of which is equal to the sum of all feature dimensions.

[0053] For example, the four geometric features, one material feature, and three surface features of the insulating rod are normalized and then concatenated into an 8-dimensional high-dimensional input feature vector, which is used to input the autoencoder model.

[0054] Secondly, the high-dimensional input feature vector is fed into a pre-trained autoencoder model. Forward propagation is performed through the encoder part of the autoencoder model, outputting the activation value of the bottleneck layer of the encoder part. This activation value is then used as a low-dimensional latent feature vector. Specifically, the pre-trained autoencoder model has saved the parameters of the trained encoder network. After the high-dimensional input feature vector is input into it, it is processed only by the encoder part. The encoder performs progressive dimensionality reduction and feature fusion on the high-dimensional features through forward propagation of a multi-layer neural network. Finally, the activation value is output at the bottleneck layer (i.e., the last layer of the encoder, with the lowest dimension). This activation value is the low-dimensional latent feature vector, which can retain the core information of the original multi-dimensional physical features to the greatest extent while eliminating redundant information.

[0055] Among them, forward propagation computation refers to the computation process from the input layer of the neural network, through the hidden layer, to the output layer; the bottleneck layer is the smallest dimension layer in the encoder, and its output activation value is the core compressed representation of high-dimensional features.

[0056] For example, an 8-dimensional high-dimensional input feature vector is input into an autoencoder model. After forward propagation calculation by the encoder, the bottleneck layer outputs a 3-dimensional activation value, which is the low-dimensional latent feature vector that can be used for subsequent clustering and classification.

[0057] Furthermore, the process of constructing the insulation distance proximity loss and the voltage level classification loss includes: The first training batch is constructed by selecting tool samples with known insulation distance labels from the multidimensional physical feature sample set of historical power tools; The samples in the first training batch are input into the encoder structure to obtain the first low-dimensional latent feature vector corresponding to the first training batch. Calculate the pairwise Euclidean distance between the first low-dimensional latent feature vectors corresponding to the first training batch, and construct the first distance matrix; Based on the known insulation distance labels, calculate the absolute difference between the insulation distances of the tool samples in the first training batch, and construct a second distance matrix; An insulation distance proximity loss is constructed with the objective of minimizing the difference between the first distance matrix and the second distance matrix; A second training batch is constructed by selecting tool samples with known voltage level labels from a multi-dimensional physical feature sample set of historical power tools. A fully connected classification layer is connected after the encoder structure to form a voltage level classifier. The samples in the second training batch are input into the encoder structure to obtain the second low-dimensional latent feature vector corresponding to the second training batch. The second low-dimensional latent feature vector corresponding to the second training batch is input into the voltage level classifier, and the predicted voltage level is output. A voltage level classification loss is constructed with the objective of minimizing the cross-entropy loss between the predicted voltage level and the actual voltage level of the tool samples in the second training batch.

[0058] In this embodiment, samples with known insulation distance labels are first selected from a set of historical power equipment multidimensional physical feature samples to construct a first training batch. Specifically, all samples labeled with insulation distance are selected from the set of historical power equipment multidimensional physical feature samples and divided into multiple first training batches according to batch size (e.g., 32 samples / batch) for training the insulation distance proximity loss. Samples with known insulation distance labels are selected because the core of this loss is to constrain the correlation between the low-dimensional features of the samples and the insulation distance, which requires relying on label information to construct constraints.

[0059] For example, 32 tool samples labeled with insulation distance tags, such as 0.2m, 0.3m, 0.5m, etc., are selected from the multi-dimensional physical feature sample set of historical power tools to form a first training batch.

[0060] Secondly, the samples from the first training batch are input into the encoder structure to obtain the first low-dimensional latent feature vector corresponding to the first training batch. Specifically, the normalized feature vector of each sample in the first training batch is input into the encoder structure of the autoencoder. Through forward propagation, the low-dimensional latent feature vector corresponding to each sample is output, i.e., the first low-dimensional latent feature vector. The dimension of the first low-dimensional latent feature vector is consistent with the dimension of the encoder bottleneck layer, and it is used to subsequently calculate the similarity between samples, thereby constructing the insulation distance proximity loss.

[0061] Next, calculate the pairwise Euclidean distances between the first low-dimensional latent feature vectors corresponding to the first training batch, and construct the first distance matrix. Specifically, Euclidean distance refers to the straight-line distance between two vectors, used to measure the similarity between two first low-dimensional latent feature vectors. The smaller the Euclidean distance, the higher the feature similarity. Calculate the pairwise Euclidean distances between the first low-dimensional latent feature vectors of all samples in the first training batch, and arrange these distance values ​​in matrix form to form the first distance matrix. The number of rows and columns of the matrix are equal to the number of samples in the first training batch, and the matrix element (i,j) represents the Euclidean distance between the i-th sample and the j-th sample.

[0062] Furthermore, based on the known insulation distance labels, the absolute differences between the insulation distances of the tool samples in the first training batch are calculated, and a second distance matrix is ​​constructed. Specifically, for each sample in the first training batch, based on its known insulation distance labels, the absolute difference in insulation distance between any two samples is calculated, and these differences are arranged in matrix form to form the second distance matrix. The number of rows and columns of the second distance matrix is ​​the same as that of the first distance matrix, and the matrix element (i,j) represents the absolute difference in insulation distance between the i-th sample and the j-th sample, used to reflect the true difference in insulation distance between samples.

[0063] Furthermore, an insulation distance proximity loss is constructed with the goal of minimizing the difference between the first and second distance matrices. Specifically, the core logic of the insulation distance proximity loss is that samples with closer insulation distances should also have closer Euclidean distances in their low-dimensional latent feature vectors, meaning that the first and second distance matrices should be as similar as possible.

[0064] For example, mean squared error (MSE) can be used to calculate the difference between the first distance matrix and the second distance matrix, and this difference can be used as the insulation distance proximity loss. During training, by minimizing the insulation distance proximity loss, the model can learn the mapping relationship between insulation distance similarity and low-dimensional feature similarity.

[0065] Furthermore, a second training batch is constructed by selecting equipment samples with known voltage level labels from a set of historical multidimensional physical feature samples of electrical equipment. Specifically, the fully connected classification layer is a classification module composed of multiple fully connected neural networks. Its input is the low-dimensional latent feature vector output by the encoder structure, and its output is the predicted probability that the electrical equipment under test belongs to each preset voltage level. The fully connected classification layer is connected to the encoder structure to form a voltage level classifier, which is used to map the low-dimensional latent feature vector to the voltage level prediction result, and then calculate the voltage level classification loss based on the prediction result and the true voltage level label.

[0066] In this case, the number of neurons in the output layer of the fully connected classification layer is equal to the total number of voltage level categories. For example, if the preset voltage levels include three categories: 10kV, 35kV, and 110kV, then the number of neurons in the output layer is 3, corresponding to the prediction probabilities of the three voltage levels.

[0067] For example, the construction process of the voltage level classifier can be referenced as follows: The fully connected classification layer adopts a two-layer fully connected neural network structure. The first layer is a hidden layer with 32 neurons, using the Rectified Linear Unit (ReLU) activation function, and adding a batch normalization layer and a Dropout layer with a dropout rate of 0.3 to enhance the model's generalization ability and prevent overfitting. The second layer is the output layer with the same number of neurons as the number of voltage level categories (e.g., 3 categories), using the Softmax activation function to normalize the output values ​​into a probability distribution form, such that the sum of the probabilities of all output nodes is 1. The weights are initialized using the He initialization method, and the bias term is initialized to 0.

[0068] For example, during training, the fully connected classification layer and the encoder structure jointly participate in the forward propagation computation: the encoder structure receives normalized multi-dimensional physical feature sample input and outputs a low-dimensional latent feature vector; the fully connected classification layer receives this low-dimensional latent feature vector and outputs the probability distribution of the predicted voltage level. The difference between the predicted probability distribution and the true voltage level label (one-hot encoded form) is calculated using the cross-entropy loss function. This cross-entropy loss is used as the voltage level classification loss and incorporated into the comprehensive loss function to participate in gradient backpropagation, synchronously updating the network parameters of the encoder structure and the fully connected classification layer.

[0069] Furthermore, the samples from the second training batch are input into the encoder structure to obtain the second low-dimensional latent feature vector corresponding to the second training batch. Specifically, consistent with the method of obtaining the first low-dimensional latent feature vector, the normalized multidimensional physical feature vector of each sample in the second training batch is input into the encoder structure. Through forward propagation calculation, the low-dimensional latent feature vector corresponding to each sample is output. These vectors are the second low-dimensional latent feature vectors, which will be used as the input to the voltage level classifier.

[0070] Furthermore, the second low-dimensional latent feature vector corresponding to the second training batch is input into the voltage level classifier, and the predicted voltage level is output. Specifically, after the second low-dimensional latent feature vector is input into the voltage level classifier, it is calculated through layer-by-layer forward propagation of the fully connected classification layer. The output layer uses the Softmax activation function to generate the probability distribution of each sample belonging to each preset voltage level, and the voltage level with the highest probability value is selected as the predicted voltage level of the sample.

[0071] For example, after a sample is calculated by a voltage level classifier, the predicted probability distribution is 0.85 for 10kV, 0.12 for 35kV, and 0.03 for 110kV. Then, the 10kV with the highest predicted probability is selected as the predicted voltage level of the sample.

[0072] Finally, a voltage level classification loss is constructed with the objective of minimizing the cross-entropy loss between the predicted voltage level and the true voltage level of the tool samples in the second training batch. Specifically, the cross-entropy loss is a loss function used in classification tasks to measure the difference between the predicted probability distribution and the true label distribution, and its calculation formula is: Cross-entropy loss = Where C represents the total number of voltage level categories. One-hot encoding for the actual voltage level. To predict the probability that an output sample belongs to class c, substitute the predicted probability distribution of each sample and its true voltage level label into the cross-entropy loss formula to calculate the classification loss for that sample. The average of the classification losses for all samples is the voltage level classification loss for the current training batch.

[0073] For example, during training, by minimizing the voltage level classification loss, gradient backpropagation synchronously updates the network parameters of the encoder structure and the fully connected classification layer, thereby improving the model's accuracy in recognizing voltage levels and ensuring that the low-dimensional latent feature vector output by the encoder structure can effectively retain discriminative information related to voltage levels.

[0074] In summary, compared to existing technologies, this application inputs the multidimensional physical features into a pre-trained autoencoder model to extract the corresponding low-dimensional latent feature vectors. In this way, the high-dimensional multidimensional physical features are reduced to low-dimensional latent feature vectors, preserving the essential physical properties of the tools while removing data redundancy and noise interference. This provides a compact and effective feature representation for subsequent accurate clustering category determination and detection parameter matching.

[0075] S40: Determine the cluster category of the electrical appliance under test based on the low-dimensional potential feature vector, and obtain the detection parameter set from the detection parameter knowledge graph according to the cluster category. The detection parameter set includes insulation distance, test voltage and fixture type.

[0076] In the testing of electrical tools, there is a complex nonlinear relationship between the physical properties of different tools and the required testing parameters. Traditional methods rely solely on fixed table lookup matching based on voltage level, which is difficult to adapt to the parameter differences of tools with different materials and aging levels at the same voltage level. This results in poor adaptability of testing parameters and increased safety hazards.

[0077] To address the aforementioned issues, this application determines the clustering category of the electrical equipment under test based on the low-dimensional latent feature vector, and obtains the detection parameter set from the detection parameter knowledge graph according to the clustering category. The detection parameter set includes insulation distance, test voltage, and fixture type, thereby realizing intelligent and precise mapping from the physical properties of the equipment to the detection parameters.

[0078] Specifically, step S40 in the method includes: Calculate the Euclidean distance between the low-dimensional latent feature vector and a plurality of pre-defined cluster centers; The electrical equipment under test is classified into the cluster category corresponding to the cluster center with the smallest Euclidean distance to the low-dimensional potential feature vector; Based on the clustering category, search for matching detection parameter records in the detection parameter knowledge graph; The insulation distance, test voltage, and fixture type corresponding to the cluster category are read from the detection parameter record to form a detection parameter set; The process of constructing the detection parameter knowledge graph includes: determining the insulation distance, test voltage, and fixture type corresponding to each cluster category based on the known detection parameters of historical power equipment in each cluster category, establishing the mapping relationship between cluster categories and detection parameters, and forming the detection parameter knowledge graph.

[0079] In this embodiment, the Euclidean distance between the low-dimensional latent feature vector and a set of pre-defined cluster centers is first calculated. Specifically, the pre-defined cluster centers are obtained using a constrained clustering algorithm based on the low-dimensional latent features of historical samples, and each cluster center corresponds to a cluster category. The Euclidean distance between the low-dimensional latent feature vector of the electrical appliance under test and each cluster center is calculated. The magnitude of the Euclidean distance reflects the similarity between the appliance under test and that cluster category; the smaller the distance, the higher the similarity.

[0080] Secondly, the electrical equipment under test is categorized into the cluster category corresponding to the cluster center with the smallest Euclidean distance to the low-dimensional latent feature vector. Specifically, by comparing the Euclidean distances between the low-dimensional latent feature vector of the equipment under test and all cluster centers, the cluster center with the smallest distance is selected, and the equipment under test is assigned to the cluster category corresponding to that cluster center. This classification method ensures that the equipment under test has similar physical characteristics and testing requirements to historical equipment in the same cluster category.

[0081] For example, if the electrical tool to be tested is an insulating glove, and its low-dimensional latent feature vector has an Euclidean distance of 0.12 from the cluster center of "10kV insulating gloves" and an Euclidean distance of 0.85 from the cluster center of "35kV insulating gloves", then it is classified into the "10kV insulating gloves" cluster category.

[0082] Next, based on the cluster category, a matching detection parameter record is searched in the detection parameter knowledge graph. Specifically, the detection parameter knowledge graph stores the mapping relationship between each cluster category and the corresponding detection parameter. This mapping relationship is established based on historical testing experience and standard specifications of the tools under test. Based on the cluster category of the tool under test, the corresponding detection parameter record for that cluster category is found in the detection parameter knowledge graph, ensuring the suitability of the detection parameters.

[0083] Furthermore, the insulation distance, test voltage, and fixture type corresponding to the cluster category are read from the test parameter records to form a test parameter set. Specifically, each test parameter record contains the insulation distance, test voltage, and fixture type corresponding to that cluster category. The insulation distance refers to the minimum air gap that needs to be maintained between the high-voltage electrode and the grounding body during high-voltage testing, ensuring no discharge occurs during the test and protecting equipment and personnel safety. The test voltage refers to the voltage value applied during electrical performance testing, determined based on the voltage level of the tool and the test standard. The fixture type refers to the adaptive fixture specifications required when gripping this type of tool, guiding the ground-rail robotic arm to retrieve the correct fixture from the tool library. These three parameters are integrated to form a test parameter set for the electrical tool under test, guiding subsequent gripping operations of the ground-rail robotic arm, safety distance control, and test parameter settings for the automated test platform.

[0084] Finally, based on the known testing parameters of historical power equipment in each cluster, the insulation distance, test voltage, and fixture type corresponding to each cluster are determined, establishing a mapping relationship between clusters and testing parameters, forming a testing parameter knowledge graph. Specifically, the construction of the testing parameter knowledge graph is based on historical data. First, the known testing parameters (at least insulation distance, test voltage, and fixture type) of all historical power equipment in each cluster are statistically analyzed. By taking the average and selecting standard values, the standard testing parameters corresponding to that cluster are determined. For example, the standard testing parameters for the "10kV insulating gloves" cluster are: insulation distance 0.3m, test voltage 20kV, and fixture type: special fixture for insulating gloves. Subsequently, a one-to-one mapping relationship between clusters and standard testing parameters is established, stored using a graph structure. For example, nodes represent clusters and testing parameters, and edges represent mapping relationships, forming a testing parameter knowledge graph for easy and rapid subsequent lookup.

[0085] Furthermore, the steps for setting cluster centers include: Collect a multi-dimensional physical feature sample set of historical power equipment, and input the multi-dimensional physical feature sample set of historical power equipment into a pre-trained autoencoder model to obtain the corresponding historical low-dimensional latent feature vector set; Using the historical low-dimensional latent feature vectors as clustering objects and the insulation distance of the historical power equipment as clustering constraints, a constrained clustering algorithm is used to divide the historical power equipment into K cluster categories. Calculate the centroid of each cluster category and use the centroid as the predefined cluster center.

[0086] In this embodiment, a multi-dimensional physical feature sample set of historical power equipment is first collected. This sample set is then input into a pre-trained autoencoder model to obtain a corresponding set of historical low-dimensional latent feature vectors. Specifically, the collected multi-dimensional physical feature sample set of historical power equipment is input into the pre-trained autoencoder model. Through encoder processing, a historical low-dimensional latent feature vector corresponding to each historical power equipment is output. All historical low-dimensional latent feature vectors constitute a set of historical low-dimensional latent feature vectors for subsequent cluster analysis.

[0087] Secondly, using historical low-dimensional latent feature vectors as clustering objects and the insulation distance of historical power equipment as clustering constraints, a constrained clustering algorithm is employed to divide historical power equipment into K cluster categories. Specifically, the clustering object is the historical low-dimensional latent feature vector, and the clustering constraint is the insulation distance of historical power equipment. This requires that samples within the same cluster category not only have similar low-dimensional features but also similar insulation distances, avoiding clustering results with similar features but excessively large differences in insulation distance. The constrained clustering algorithm can be the constrained K-means algorithm, where the K value (i.e., the number of cluster categories) is set according to the type and quantity of historical power equipment samples. For example, based on voltage level and equipment type, K=5 can be set.

[0088] For example, the constrained K-means algorithm is used to divide 1000 historical samples into 5 clusters with historical low-dimensional latent feature vectors as clustering objects and insulation distance as a constraint. The insulation distance difference of samples in each cluster does not exceed 0.1m.

[0089] Finally, the centroid of each cluster is calculated and used as the pre-defined cluster centers. Specifically, the centroid of a cluster is the mean vector of all historical low-dimensional latent feature vectors within that cluster. Calculating the centroid of each cluster involves taking the mean of each dimension of all historical low-dimensional latent feature vectors within that cluster; the resulting vector is the centroid. The centroid of each cluster is calculated using the same method and used as multiple pre-defined cluster centers for subsequent clustering and classification of the electrical equipment under test, ensuring that the cluster centers represent the core features of that category.

[0090] In summary, compared to existing technologies, this application determines the cluster category of the electrical appliance under test based on the low-dimensional latent feature vector, and obtains the detection parameter set from the detection parameter knowledge graph according to the cluster category. The detection parameter set includes insulation distance, test voltage, and fixture type. Thus, through cluster matching based on low-dimensional latent feature vectors and knowledge graph querying, intelligent and accurate mapping from the physical attributes of the electrical appliance under test to the detection parameters is achieved, ensuring a high degree of fit between insulation distance, test voltage, and fixture type and the individual characteristics of the appliance.

[0091] S50: The ground rail robotic arm grabs the electrical appliance to be tested according to the type of clamp and moves it to the automated test platform. The vision sensor at the end of the ground rail robotic arm guides the ground rail robotic arm to complete the automatic docking of the test high voltage line with the electrode of the electrical appliance to be tested. After docking, the ground rail robotic arm is controlled to retreat to a safe area according to the insulation distance. Then, high voltage is applied according to the test voltage for testing.

[0092] In the traditional testing process for electrical tools, the connection and disconnection of high-voltage test lines usually rely on manual operation. Operators need to make close contact with the electrodes of the tools in a high-voltage environment to make connections, which poses problems such as high risk of electric shock, low operating efficiency, and poor wiring consistency.

[0093] Meanwhile, manual wiring often relies on experience and visual estimation to determine safe distances, lacking precise quantitative control. This can easily lead to equipment-to-equipment discharge during testing due to insufficient distance. Furthermore, different voltage levels of tools have varying requirements for test voltage and insulation distance, making precise adaptation difficult to achieve manually.

[0094] To address the aforementioned issues, this application utilizes a ground-rail robotic arm to grasp the electrical appliance under test according to the clamp type and move it to an automated testing platform. A vision sensor at the end of the robotic arm guides it to automatically connect the high-voltage test line to the electrodes of the electrical appliance under test. After connection, the robotic arm is controlled to retreat to a safe area based on the insulation distance. Subsequently, a high voltage is applied for testing. This achieves fully automated wiring operations, quantitative control of safe distances, and automatic adaptation of test parameters, fundamentally eliminating the safety hazards of manual operation and improving the accuracy and efficiency of testing.

[0095] Specifically, step S50 in the method includes: The ground-rail robotic arm calls an adaptive gripper that matches the gripper type from the tool library, wherein the adaptive gripper has a built-in pressure sensor; The ground-rail robotic arm carries the adaptive gripper to the picking point, grips the electrical tool to be tested according to its gripping position, and controls the gripping force through the pressure sensor. The electrical tool under test is transported to an automated testing platform, and the electrode position coordinates of the electrical tool under test are identified by a vision sensor at the end of the ground-rail robotic arm. The ground-rail robotic arm grasps the connector of the test high-voltage line and guides the connector of the test high-voltage line to be connected to the electrode of the electrical tool under test according to the electrode position coordinates to form an electrical connection. After the electrical connection is completed, the ground rail robotic arm loosens the connector of the test high voltage line, calculates the coordinates of the safe stopping point based on the insulation distance, plans the retreat path, and moves to the safe stopping point; After moving to the safe dwell point, the automated test platform performs electrical performance tests on the electrical equipment under test according to the test voltage; During the test, the positions of the ground-rail robotic arm and the automated guided vehicle are monitored in real time to ensure that the distance between all metal objects and the high-voltage electrodes of the electrical appliance under test is not less than the insulation distance.

[0096] In this embodiment, the ground-rail robotic arm first retrieves an adaptive fixture matching the fixture type from the tool library. The adaptive fixture has a built-in pressure sensor. Specifically, the tool library stores various types of specialized fixtures corresponding to different types of electrical tools, such as insulating glove fixtures, insulating rod fixtures, and voltage detector fixtures. The ground-rail robotic arm retrieves the corresponding adaptive fixture from the tool library via a mechanical interface based on the fixture type in the detection parameter set.

[0097] Among them, adaptive clamps refer to clamps that can automatically adjust the clamping range according to the size and shape of the tool. The built-in pressure sensor is used to detect the magnitude of the clamping force in real time to avoid damage to the tool due to excessive clamping force or detachment of the tool due to insufficient clamping force.

[0098] Secondly, the ground-rail robotic arm, carrying an adaptive gripper, moves to the pick-up point and grips the electrical tool under test according to its gripping position. The gripping force is controlled by feedback from a pressure sensor. Specifically, the pick-up point is the location of the electrical tool under test in the pre-inspection station. The vision sensor at the pre-inspection station locates the gripping position of the electrical tool under test, such as the middle of an insulating rod or the wrist of an insulating glove, and sends the position coordinates to the ground-rail robotic arm. The ground-rail robotic arm, carrying the adaptive gripper, moves to the pick-up point, aligns with the gripping position, and performs the gripping. Simultaneously, the pressure sensor provides real-time feedback of the gripping force signal, and the gripping force is adjusted based on the feedback signal to ensure stable gripping without damaging the electrical tool under test.

[0099] Next, the electrical equipment under test is transported to the automated testing platform, where a vision sensor at the end of a ground-rail robotic arm identifies the coordinates of its electrodes. Specifically, after the ground-rail robotic arm picks up the electrical equipment under test, it transports it to the testing station on the automated testing platform. The end of the ground-rail robotic arm integrates a vision sensor, which scans the electrical equipment under test and uses image recognition algorithms to locate the electrode positions. Then, the vision sensor outputs the precise coordinates (x, y, z) of the electrodes in three-dimensional space, providing navigation for subsequent automatic docking.

[0100] The visual sensor at the end of the ground-rail robotic arm can be a 3D vision camera, such as a structured light camera, a binocular vision camera, or a time-of-flight camera, or a high-resolution CCD industrial camera. The specific selection can be determined based on the detection accuracy requirements and the site environment.

[0101] Electrodes refer to the metal parts that need to be connected to the high-voltage test line during high-voltage testing, such as the metal hook at the end of the insulating rod, the contact electrode of the electroscope, and the conductive layer interface of the insulating glove.

[0102] For example, the image recognition algorithm in the above-mentioned "locating the electrode position of the electrical tool under test using an image recognition algorithm" can employ a deep learning-based object detection algorithm, such as YOLOv5, to identify the electrode position. Specific implementation steps are as follows: First, construct an electrode image dataset, collecting at least 1000 images containing electrodes under different lighting conditions and angles, and using an annotation tool to label the electrode regions in the images with rectangular boxes; divide the dataset into training and validation sets in an 8:2 ratio. Train the model using the YOLOv5 network structure, setting the input image size to 640×640 pixels, the batch size to 16, the initial learning rate to 0.01, using the SGD optimizer, the momentum factor to 0.937, and the weight decay coefficient to 0.0005. Perform 200 training epochs, employing data augmentation strategies (including random flipping, scaling, color adjustment, etc.). After training, deploy the model in the controller of the ground-rail robotic arm. During actual detection, a vision sensor acquires real-time images, inputs them into the model for inference, and outputs the three-dimensional position coordinates of the electrode in the robot's base coordinate system.

[0103] Furthermore, the ground-rail robotic arm grasps the connector of the test high-voltage line and guides it to connect with the electrode of the electrical appliance under test based on the electrode position coordinates, forming an electrical connection. Specifically, the ground-rail robotic arm grasps the connector of the test high-voltage line from its fixed support. The other end of the test high-voltage line is connected to the high-voltage output terminal of the automated testing platform. Based on the identified electrode position coordinates, the ground-rail robotic arm guides the connector to move near the electrode and dynamically adjusts its own posture through motion control algorithms to align the connector with the electrode and complete the connection, ensuring good contact and forming a stable electrical connection. This avoids problems such as poor contact and arcing during high-voltage testing.

[0104] For example, the motion control algorithm can employ a position control method based on visual servoing. Specific implementation steps are as follows: First, the images acquired in real-time by the visual sensor are correlated with the electrode position coordinates using a coordinate system unification. The electrode position in the image coordinate system is then converted to the target position in the robot's base coordinate system using a hand-eye calibration matrix. Second, a PID controller is used to calculate the target angles of each joint of the robot, with a control cycle of 10 milliseconds, a proportional coefficient Kp=0.8, an integral coefficient Ki=0.05, and a derivative coefficient Kd=0.1. During movement, the visual sensor continuously provides feedback on the relative positional deviation between the connector and the electrode, and the robot's movement is adjusted in real-time based on the deviation until the distance between the connector and the electrode is less than a preset threshold (a dynamically set parameter, e.g., 1 mm). Once the distance meets the requirement, linear interpolation motion is executed to slowly insert the connector into the electrode, completing the insertion.

[0105] Optionally, during the insertion process, the ground-rail robotic arm can employ a force / position hybrid control strategy to ensure smooth insertion and avoid rigid collisions that could damage the electrodes or connectors. For example, position control is performed in the insertion direction, while a six-dimensional force sensor integrated at the end of the ground-rail robotic arm monitors the contact force in real time. Force control is then performed perpendicular to the insertion direction to limit the contact force within a preset range (dynamically set parameters, e.g., 5-10 Newtons). When the contact force exceeds the threshold, the position feed is paused and the attitude is finely adjusted to eliminate lateral deviation, ensuring a smooth and damage-free insertion process.

[0106] Furthermore, the ground-rail robotic arm loosens the connector of the high-voltage test line, calculates the coordinates of the safe stopping point based on the insulation distance, plans a retreat path, and moves to the safe stopping point. Specifically, after the connection is completed, the ground-rail robotic arm loosens the connector of the high-voltage test line, leaving the connector on the electrode. At this point, the ground-rail robotic arm itself, as a metal object, if it remains near the high-voltage electrode, will shorten the air gap between the high-voltage electrode and the grounding body, potentially causing inter-equipment discharge. Therefore, the ground-rail robotic arm needs to calculate the coordinates of the safe stopping point based on the insulation distance and plan a retreat path from its current position to the safe stopping point, ensuring that it does not touch other equipment during the retreat process, and then moves along the planned path to the safe stopping point.

[0107] The safe stopping point refers to the position where the ground rail robotic arm can safely stop during the test. The distance between this position and the high voltage electrode is not less than the insulation distance plus the preset safety margin.

[0108] Furthermore, the automated testing platform performs electrical performance tests on the electrical equipment under test based on the test voltage. Specifically, after the ground-rail robotic arm moves to the safe stopping point, the automated testing platform receives the test voltage command and applies high voltage to the electrodes of the electrical equipment under test through the test high-voltage line to perform electrical performance tests.

[0109] Electrical performance testing includes at least insulation resistance testing, AC withstand voltage testing, and DC withstand voltage testing, with specific test items determined based on the type of equipment and testing standards. During the testing process, the automated testing platform collects parameters such as voltage and current in real time to determine whether the electrical equipment under test is qualified.

[0110] For example, a 20kV test voltage is applied to a 10kV insulating glove, and the test is conducted for 1 minute. Leakage current data is collected. If the leakage current is less than 10μA, the electrical performance is deemed qualified.

[0111] Finally, during the test, the positions of the rail-mounted robotic arm and the automated guided vehicle (AGV) are monitored in real time to ensure that the distance between any metal object and the high-voltage electrode of the electrical appliance under test is not less than the insulation distance. Specifically, during the test, the distance between the rail-mounted robotic arm and the AVT and the high-voltage electrode of the electrical appliance under test is calculated by monitoring their positions in real time. If any metal object (including at least the rail-mounted robotic arm, AVT, or other equipment) is less than the insulation distance from the high-voltage electrode, an alarm is immediately issued, an emergency stop is triggered, and the high-voltage output is cut off to ensure test safety.

[0112] Further, calculating the coordinates of the safe stopping point based on the insulation distance, planning the retreat path, and moving to the safe stopping point includes: A spherical restricted area is constructed with the high-voltage electrode position of the electrical appliance under test as the center and the sum of the insulation distance and the preset safety margin as the radius. Starting from the current position of the ground-rail robotic arm, search for reachable poses of the ground-rail robotic arm outside the spherical restricted area, and use the searched reachable poses as a set of candidate dwell points; Select the reachable pose closest to the starting point from the set of candidate dwell points as the coordinates of the safe dwell point; The retreat path is planned using the artificial potential field method, and a repulsive field is constructed at the location of the high-voltage electrode, so that the ground-rail robotic arm is subjected to an attractive force pointing towards the coordinates of the safe stopping point and a repulsive force moving away from the spherical restricted area during the retreat process.

[0113] In this embodiment, a spherical restricted area is first constructed with the high-voltage electrode location of the electrical appliance under test as the center and the sum of the insulation distance and the preset safety margin as the radius. Specifically, the high-voltage electrode location of the electrical appliance under test (i.e., the electrode location coordinates identified in the aforementioned steps) is first obtained. Then, with the high-voltage electrode location as the center and the sum of the insulation distance and the preset safety margin as the radius, a spherical region is constructed in three-dimensional space. This region is defined as the spherical restricted area. During the test, no metal object may enter this spherical restricted area, otherwise it may cause a discharge risk.

[0114] The preset safety margin is an extra distance added to increase safety. It can be set to a fixed value, such as 0.1 meters, depending on the test voltage level and environmental conditions.

[0115] Secondly, using the current position of the ground-rail manipulator as the starting point, the search is conducted outside the spherical restricted area to find reachable poses for the manipulator. The found reachable poses are then used as a candidate set of dwell points. Specifically, the current position of the ground-rail manipulator is the position at which the connection is completed. This position is typically very close to the high-voltage electrode, located within or near the boundary of the spherical restricted area, thus requiring a retreat to a safe zone. Using the current position of the ground-rail manipulator as the starting point for path planning, all poses within the manipulator's workspace that satisfy the following conditions are searched: the pose is outside the spherical restricted area, the manipulator can reach this pose through inverse kinematics (i.e., reachable poses), and all found reachable poses constitute a candidate set of dwell points.

[0116] The search process can employ sampling algorithms or grid search methods for spatial traversal. For example, a fast random search tree algorithm can be used to perform sampling searches in joint space. The specific implementation steps are as follows: First, initialize a search tree with the joint angle corresponding to the current position of the ground-rail robotic arm as the root node, setting the maximum number of iterations to 5000 and the expansion step size to 50 mm or the corresponding joint angle step size. In each iteration, randomly sample a target pose within the workspace (note that the target pose must be outside the spherical restricted area). Then, find the node closest to the sampling point in the search tree and expand it by one step towards the sampling point to obtain a new node. Verify whether the joint angle corresponding to the new node satisfies joint constraints and does not interfere with obstacles through inverse kinematics. If it does, add the new node to the search tree. When the new node is located in a safe area and sufficiently far from the starting point, record the pose corresponding to that node as a candidate dwell point. Preferably, to improve search efficiency, a target bias strategy can be introduced: directly set the target point as the sampling point with a probability of 0.1, guiding the search tree to grow towards the safe area. After the search is completed, the reachable poses of all records constitute a set of candidate dwell points.

[0117] Next, the reachable pose closest to the starting point is selected from the set of candidate dwell points as the coordinates of the safe dwell point. Specifically, in order to minimize the travel distance of the ground-rail robotic arm and improve testing efficiency, the reachable pose closest to the starting point (i.e., the current position) is selected from the set of candidate dwell points as the coordinates of the safe dwell point. This safe dwell point satisfies the safety distance requirement and is as close to the starting point as possible, facilitating rapid arrival and possible subsequent operations.

[0118] Finally, the artificial potential field method is used to plan the retreat path. A repulsive field is constructed at the location of the high-voltage electrode, so that the ground-rail robotic arm experiences both an attractive force pointing towards the safe stopping point coordinates and a repulsive force moving away from the spherical restricted area during the retreat process. Specifically, the artificial potential field method is a path planning method that simulates the motion of an object in a virtual force field. A repulsive field is constructed at the location of the high-voltage electrode, and the strength of the repulsive field is inversely proportional to the distance: the closer to the high-voltage electrode, the stronger the repulsive force, and the direction is away from the high-voltage electrode. At the same time, an attractive field is constructed at the safe stopping point coordinates, with the attractive force pointing towards the safe stopping point.

[0119] During the retraction process, the ground-rail robotic arm is subjected to both attractive and repulsive forces: the attractive force guides the robotic arm towards the safe stopping point, while the repulsive force keeps it away from the high-voltage electrode and the spherical restricted area. By adjusting the values ​​of the attractive and repulsive force coefficients, a smooth and safe retraction path can be planned, ensuring that the ground-rail robotic arm maintains a safe distance from the high-voltage electrode throughout the entire retraction process and ultimately reaches the safe stopping point accurately.

[0120] For example, the artificial potential field method can be used to plan a retreat path with the following parameter configuration: set the attraction coefficient to 0.5, the repulsion coefficient to 1.2, and the repulsion influence range to 1.5 times the sum of the insulation distance and the preset safety margin. For example, if the insulation distance is 1.0 meter and the safety margin is 0.1 meter, then the influence range is 1.65 meters. The path planning adopts an iterative method. Each step moves a small step according to the direction of the resultant force of the attraction and repulsion at the current position, with a step size of 0.05 meters, until a safe stopping point is reached.

[0121] In summary, compared to existing technologies, the present application's ground-rail robotic arm grasps the electrical appliance under test according to the clamp type and moves it to the automated testing platform. A vision sensor at the end of the robotic arm guides it to automatically connect the high-voltage test line to the electrodes of the electrical appliance under test. After connection, the robotic arm is controlled to retreat to a safe area based on the insulation distance. Subsequently, high voltage is applied for testing. This achieves unmanned operation of the entire high-voltage test wiring and testing process, eliminating the risk of electric shock to personnel, ensuring the safety of the testing process, and improving wiring accuracy and testing efficiency.

[0122] S60: Based on the test results, the automated guided vehicle is dispatched to classify and return the electrical tools under test to the smart shelf.

[0123] In this embodiment, after the test is completed, the test data uploaded by the automated testing platform is received, and the test data is compared with the preset pass / fail standards to generate a pass or fail test result. Next, the test result is bound to the electronic tag of the electrical tool under test, and the database of the intelligent shelf management system is updated. Then, an automated guided vehicle is dispatched to the material picking position of the automated testing platform to pick up the electrical tool under test.

[0124] For example, based on the test results, the automated guided vehicle (AGV) transports the electrical tools that pass the test to the qualified product storage area of ​​the smart shelf, and the electrical tools that fail the test to the unqualified product isolation area of ​​the smart shelf. The entrance to the unqualified product isolation area is equipped with an electrically locking door. After the AGV stores the unqualified electrical tools in the unqualified product isolation area, the electrically locking door automatically closes and locks, simultaneously disabling the electronic tag's exit permission for the unqualified electrical tools to prevent misuse.

[0125] In this way, a closed loop is completed from sending for inspection to sorting and returning to the warehouse.

[0126] In summary, the embodiments of this application have at least the following technical effects: Compared with existing technologies, this application first uses an automated guided vehicle to transport the electrical equipment to be tested from the smart shelf to the pre-inspection station according to the testing task schedule, thereby realizing the automated logistics scheduling of the testing task and replacing the traditional manual handling mode.

[0127] Secondly, this application collects multi-dimensional physical features of the electrical tool under test at the pre-inspection station. These multi-dimensional physical features include geometric features, material features, and surface features, achieving comprehensive digitization and precise quantification of the physical properties of the tool, and providing a multi-dimensional data foundation for subsequent intelligent identification and parameter matching.

[0128] Furthermore, this application inputs the multidimensional physical features into a pre-trained autoencoder model to extract the corresponding low-dimensional latent feature vectors, reducing the high-dimensional features to a compact low-dimensional representation. This removes data redundancy and noise interference while preserving the essential physical properties of the tools, providing effective feature support for subsequent clustering determination and parameter matching.

[0129] Furthermore, this application determines the clustering category of the electrical equipment under test based on the low-dimensional latent feature vector, and obtains the detection parameter set from the detection parameter knowledge graph according to the clustering category. The detection parameter set includes insulation distance, test voltage and fixture type, realizing intelligent and accurate mapping from the physical attributes of the equipment to the detection parameters, and ensuring a high degree of fit between the detection parameters and the individual characteristics of the equipment.

[0130] Furthermore, in this application, a ground-rail robotic arm grasps the electrical appliance under test according to the type of clamp and moves it to an automated testing platform. A vision sensor at the end of the ground-rail robotic arm guides the arm to automatically connect the high-voltage test line to the electrodes of the electrical appliance under test. After connection, the ground-rail robotic arm is controlled to retreat to a safe area based on the insulation distance. Subsequently, a high voltage is applied for testing, realizing unmanned operation of the entire high-voltage test wiring and testing process, eliminating the risk of electric shock to personnel, and improving wiring accuracy and testing efficiency.

[0131] Finally, based on the test results, this application dispatches the automated guided vehicle to classify and return the electrical tools under test to the intelligent shelf, forming a closed loop of the entire process from delivery for inspection to sorting and returning to the warehouse. This achieves automatic isolation between qualified and unqualified products and prevents unqualified tools from being misused.

[0132] Through the above technical solutions, this application has constructed an unmanned and intelligent power tool testing system covering the entire process of handling, identification, matching, testing and sorting. While improving testing efficiency and safety, it has also achieved personalized adaptation of testing parameters and automated control of the entire process.

[0133] Example 2, as Figure 2 As shown, based on the same inventive concept as the industrial robot-based end-to-end electrical tool testing method provided in Embodiment 1, this embodiment of the invention also provides an industrial robot-based end-to-end electrical tool testing system, including: The outbound scheduling module 11 is used to schedule automated guided vehicles to move the electrical tools to be tested from the smart shelf to the pre-inspection station according to the testing task; Feature acquisition module 12 is used to acquire multi-dimensional physical features of the electrical tool under test at the pre-inspection station. The multi-dimensional physical features include geometric features, material features and surface features. The feature encoding module 13 is used to input the multidimensional physical features into a pre-trained autoencoder model and extract the corresponding low-dimensional latent feature vectors. The parameter matching module 14 is used to determine the clustering category of the electrical appliance under test based on the low-dimensional potential feature vector, and to query the detection parameter set from the detection parameter knowledge graph according to the clustering category. The detection parameter set includes insulation distance, test voltage and fixture type. The automatic testing module 15 is used to have a ground rail robotic arm grasp the electrical tool under test according to the type of clamp and move it to the automated testing platform. The vision sensor at the end of the ground rail robotic arm guides the ground rail robotic arm to complete the automatic docking of the test high voltage line with the electrode of the electrical tool under test. After docking, the ground rail robotic arm is controlled to retreat to a safe area according to the insulation distance. Then, a high voltage is applied according to the test voltage to perform the test. The inbound scheduling module 16 is used to schedule the automated guided vehicle to classify and return the electrical tools under test to the smart shelf according to the test results.

[0134] Specifically, the feature acquisition module 12 is used for: The electrical tool under test is scanned by a three-dimensional laser profilometer to obtain three-dimensional point cloud data of the electrical tool under test, and geometric features are extracted from the three-dimensional point cloud data, wherein the geometric features include at least length, diameter, taper and curvature. The surface spectral information of the electrical appliance under test is acquired by a hyperspectral camera, and material features are extracted from the surface spectral information, wherein the material features include at least the infrared spectral absorption peaks of insulating materials; The surface image of the electrical appliance under test is acquired by a vision camera, and surface features are extracted from the surface image, wherein the surface features include at least cracks, scratches and signs of aging.

[0135] Specifically, the feature encoding module 13 is used for: Collect a multi-dimensional physical feature sample set of historical electrical tools, wherein each multi-dimensional physical feature sample of historical electrical tools includes geometric features, material features and surface features. The historical electrical tools include tools with known insulation distance labels, tools with known voltage level labels and tools without labels. Normalize each sample in the multidimensional physical feature sample set to obtain a normalized multidimensional physical feature sample set, and save the statistical parameters used in the normalization process. Construct an autoencoder neural network, wherein the autoencoder neural network includes an encoder structure and a decoder structure; The normalized multidimensional physical feature sample set is input into the autoencoder neural network, and the autoencoder neural network is trained with the goal of minimizing the comprehensive loss function. The comprehensive loss function is a weighted sum of reconstruction loss, insulation distance proximity loss and voltage level classification loss. After training is complete, the network parameters of the encoder structure are saved to obtain the pre-trained autoencoder model.

[0136] Specifically, the multidimensional physical features are input into a pre-trained autoencoder model to extract the corresponding low-dimensional latent feature vectors, including: Using the saved statistical parameters, the geometric features, material features, and surface features are normalized and then concatenated to form a high-dimensional input feature vector; The high-dimensional input feature vector is input into a pre-trained autoencoder model. Forward propagation is performed through the encoder part of the autoencoder model to output the activation value of the bottleneck layer of the encoder part. The activation value is used as a low-dimensional latent feature vector.

[0137] Furthermore, the process of constructing the insulation distance proximity loss and the voltage level classification loss includes: The first training batch is constructed by selecting tool samples with known insulation distance labels from the multidimensional physical feature sample set of historical power tools; The samples in the first training batch are input into the encoder structure to obtain the first low-dimensional latent feature vector corresponding to the first training batch. Calculate the pairwise Euclidean distance between the first low-dimensional latent feature vectors corresponding to the first training batch, and construct the first distance matrix; Based on the known insulation distance labels, calculate the absolute difference between the insulation distances of the tool samples in the first training batch, and construct a second distance matrix; An insulation distance proximity loss is constructed with the objective of minimizing the difference between the first distance matrix and the second distance matrix; A second training batch is constructed by selecting tool samples with known voltage level labels from a multi-dimensional physical feature sample set of historical power tools. A fully connected classification layer is connected after the encoder structure to form a voltage level classifier. The samples in the second training batch are input into the encoder structure to obtain the second low-dimensional latent feature vector corresponding to the second training batch. The second low-dimensional latent feature vector corresponding to the second training batch is input into the voltage level classifier, and the predicted voltage level is output. A voltage level classification loss is constructed with the objective of minimizing the cross-entropy loss between the predicted voltage level and the actual voltage level of the tool samples in the second training batch.

[0138] Specifically, the parameter matching module 14 is used for: Calculate the Euclidean distance between the low-dimensional latent feature vector and a plurality of pre-defined cluster centers; The electrical equipment under test is classified into the cluster category corresponding to the cluster center with the smallest Euclidean distance to the low-dimensional potential feature vector; Based on the clustering category, search for matching detection parameter records in the detection parameter knowledge graph; The insulation distance, test voltage, and fixture type corresponding to the cluster category are read from the detection parameter record to form a detection parameter set; The process of constructing the detection parameter knowledge graph includes: determining the insulation distance, test voltage, and fixture type corresponding to each cluster category based on the known detection parameters of historical power equipment in each cluster category, establishing the mapping relationship between cluster categories and detection parameters, and forming the detection parameter knowledge graph.

[0139] Furthermore, the steps for setting cluster centers include: Collect a multi-dimensional physical feature sample set of historical power equipment, and input the multi-dimensional physical feature sample set of historical power equipment into a pre-trained autoencoder model to obtain the corresponding historical low-dimensional latent feature vector set; Using the historical low-dimensional latent feature vectors as clustering objects and the insulation distance of the historical power equipment as clustering constraints, a constrained clustering algorithm is used to divide the historical power equipment into K cluster categories. Calculate the centroid of each cluster category and use the centroid as the predefined cluster center.

[0140] The automatic testing module 15 is specifically used for: The ground-rail robotic arm calls an adaptive gripper that matches the gripper type from the tool library, wherein the adaptive gripper has a built-in pressure sensor; The ground-rail robotic arm carries the adaptive gripper to the picking point, grips the electrical tool to be tested according to its gripping position, and controls the gripping force through the pressure sensor. The electrical tool under test is transported to an automated testing platform, and the electrode position coordinates of the electrical tool under test are identified by a vision sensor at the end of the ground-rail robotic arm. The ground-rail robotic arm grasps the connector of the test high-voltage line and guides the connector of the test high-voltage line to be connected to the electrode of the electrical tool under test according to the electrode position coordinates to form an electrical connection. After the electrical connection is completed, the ground rail robotic arm loosens the connector of the test high voltage line, calculates the coordinates of the safe stopping point based on the insulation distance, plans the retreat path, and moves to the safe stopping point; After moving to the safe dwell point, the automated test platform performs electrical performance tests on the electrical equipment under test according to the test voltage; During the test, the positions of the ground-rail robotic arm and the automated guided vehicle are monitored in real time to ensure that the distance between all metal objects and the high-voltage electrodes of the electrical appliance under test is not less than the insulation distance.

[0141] Specifically, calculating the coordinates of the safe stopping point based on the insulation distance, planning the retreat path, and moving to the safe stopping point includes: A spherical restricted area is constructed with the high-voltage electrode position of the electrical appliance under test as the center and the sum of the insulation distance and the preset safety margin as the radius. Starting from the current position of the ground-rail robotic arm, search for reachable poses of the ground-rail robotic arm outside the spherical restricted area, and use the searched reachable poses as a set of candidate dwell points; Select the reachable pose closest to the starting point from the set of candidate dwell points as the coordinates of the safe dwell point; The retreat path is planned using the artificial potential field method, and a repulsive field is constructed at the location of the high-voltage electrode, so that the ground-rail robotic arm is subjected to an attractive force pointing towards the coordinates of the safe stopping point and a repulsive force moving away from the spherical restricted area during the retreat process.

[0142] It should be noted that the descriptions of each embodiment in the above embodiments have different focuses. For parts that are not described in detail in a certain embodiment, please refer to the relevant descriptions in other embodiments.

[0143] Those skilled in the art will understand that embodiments of the present invention can be provided as methods, systems, or computer program products. Therefore, the present invention can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present invention can take the form of a computer program product embodied on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0144] This invention is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of the invention. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded computer, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart illustrations. Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0145] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0146] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0147] Although preferred embodiments of the invention have been described, those skilled in the art, once they have learned the basic inventive concept, can make other changes and modifications to these embodiments.

[0148] Obviously, those skilled in the art can make various modifications and variations to this invention without departing from its spirit and scope. Therefore, if these modifications and variations fall within the scope of this invention and its equivalents, this invention also intends to include these modifications and variations.

Claims

1. A method for the whole-process testing of electrical tools based on industrial robots, characterized in that, The method includes: According to the testing task, the automated guided vehicle is dispatched to move the electrical equipment to be tested from the smart shelf to the pre-inspection station; The multidimensional physical characteristics of the electrical equipment under test are collected at the pre-inspection station. These multidimensional physical characteristics include geometric features, material features, and surface features. The multidimensional physical features are input into a pre-trained autoencoder model to extract the corresponding low-dimensional latent feature vectors. The clustering category of the electrical equipment under test is determined based on the low-dimensional latent feature vector, and the detection parameter set is obtained by querying the detection parameter knowledge graph according to the clustering category. The detection parameter set includes insulation distance, test voltage and fixture type. The ground rail robotic arm grasps the electrical tool under test according to the type of clamp and moves it to the automated test platform. The vision sensor at the end of the ground rail robotic arm guides the ground rail robotic arm to complete the automatic docking between the test high voltage line and the electrode of the electrical tool under test. After docking, the ground rail robotic arm is controlled to retreat to a safe area according to the insulation distance. Then, a high voltage is applied according to the test voltage to carry out the test. Based on the test results, the automated guided vehicle is dispatched to classify and return the electrical tools under test to the smart shelf.

2. The method for full-process testing of electrical tools based on industrial robots according to claim 1, characterized in that, The multidimensional physical characteristics of the electrical appliance under test are collected at the pre-inspection station, including: The electrical tool under test is scanned by a three-dimensional laser profilometer to obtain three-dimensional point cloud data of the electrical tool under test, and geometric features are extracted from the three-dimensional point cloud data, wherein the geometric features include at least length, diameter, taper and curvature. The surface spectral information of the electrical appliance under test is acquired by a hyperspectral camera, and material features are extracted from the surface spectral information, wherein the material features include at least the infrared spectral absorption peaks of insulating materials; The surface image of the electrical appliance under test is acquired by a vision camera, and surface features are extracted from the surface image, wherein the surface features include at least cracks, scratches and signs of aging.

3. The method for full-process testing of electrical tools based on industrial robots according to claim 1, characterized in that, The training process of an autoencoder model includes: Collect a multi-dimensional physical feature sample set of historical electrical tools, wherein each multi-dimensional physical feature sample of historical electrical tools includes geometric features, material features and surface features. The historical electrical tools include tools with known insulation distance labels, tools with known voltage level labels and tools without labels. Normalize each sample in the multidimensional physical feature sample set to obtain a normalized multidimensional physical feature sample set, and save the statistical parameters used in the normalization process. Construct an autoencoder neural network, wherein the autoencoder neural network includes an encoder structure and a decoder structure; The normalized multidimensional physical feature sample set is input into the autoencoder neural network, and the autoencoder neural network is trained with the goal of minimizing the comprehensive loss function. The comprehensive loss function is a weighted sum of reconstruction loss, insulation distance proximity loss and voltage level classification loss. After training is complete, the network parameters of the encoder structure are saved to obtain the pre-trained autoencoder model.

4. The method for full-process testing of electrical tools based on industrial robots according to claim 1, characterized in that, The multidimensional physical features are input into a pre-trained autoencoder model to extract the corresponding low-dimensional latent feature vectors, including: Using the saved statistical parameters, the geometric features, material features, and surface features are normalized and then concatenated to form a high-dimensional input feature vector; The high-dimensional input feature vector is input into a pre-trained autoencoder model. Forward propagation is performed through the encoder part of the autoencoder model to output the activation value of the bottleneck layer of the encoder part. The activation value is used as a low-dimensional latent feature vector.

5. The method for full-process testing of electrical tools based on industrial robots according to claim 3, characterized in that, The process of constructing the insulation distance proximity loss and the voltage level classification loss includes: The first training batch is constructed by selecting tool samples with known insulation distance labels from the multidimensional physical feature sample set of historical power tools; The samples in the first training batch are input into the encoder structure to obtain the first low-dimensional latent feature vector corresponding to the first training batch. Calculate the pairwise Euclidean distance between the first low-dimensional latent feature vectors corresponding to the first training batch, and construct the first distance matrix; Based on the known insulation distance labels, calculate the absolute difference between the insulation distances of the tool samples in the first training batch, and construct a second distance matrix; An insulation distance proximity loss is constructed with the goal of minimizing the difference between the first distance matrix and the second distance matrix; A second training batch is constructed by selecting tool samples with known voltage level labels from a multi-dimensional physical feature sample set of historical power tools. A fully connected classification layer is connected after the encoder structure to form a voltage level classifier. The samples in the second training batch are input into the encoder structure to obtain the second low-dimensional latent feature vector corresponding to the second training batch. The second low-dimensional latent feature vector corresponding to the second training batch is input into the voltage level classifier, and the predicted voltage level is output. A voltage level classification loss is constructed with the objective of minimizing the cross-entropy loss between the predicted voltage level and the actual voltage level of the tool samples in the second training batch.

6. The method for full-process testing of electrical tools based on industrial robots according to claim 1, characterized in that, Based on the low-dimensional latent feature vector, the cluster category of the electrical equipment under test is determined, and the detection parameter set is obtained from the detection parameter knowledge graph according to the cluster category, including: Calculate the Euclidean distance between the low-dimensional latent feature vector and a plurality of pre-defined cluster centers; The electrical equipment under test is classified into the cluster category corresponding to the cluster center with the smallest Euclidean distance to the low-dimensional potential feature vector; Based on the clustering category, search for matching detection parameter records in the detection parameter knowledge graph; The insulation distance, test voltage, and fixture type corresponding to the cluster category are read from the detection parameter record to form a detection parameter set; The process of constructing the detection parameter knowledge graph includes: determining the insulation distance, test voltage, and fixture type corresponding to each cluster category based on the known detection parameters of historical power equipment in each cluster category, establishing the mapping relationship between cluster categories and detection parameters, and forming the detection parameter knowledge graph.

7. The method for full-process testing of electrical tools based on industrial robots according to claim 6, characterized in that, The steps for setting up cluster centers include: Collect a multi-dimensional physical feature sample set of historical power equipment, and input the multi-dimensional physical feature sample set of historical power equipment into a pre-trained autoencoder model to obtain the corresponding historical low-dimensional latent feature vector set; Using the historical low-dimensional latent feature vectors as clustering objects and the insulation distance of the historical power equipment as clustering constraints, a constrained clustering algorithm is used to divide the historical power equipment into K cluster categories. Calculate the centroid of each cluster category and use the centroid as the predefined cluster center.

8. The method for full-process testing of electrical tools based on industrial robots according to claim 1, characterized in that, The ground-rail robotic arm grasps the electrical appliance under test according to the type of clamp and moves it to the automated testing platform. A vision sensor at the end of the robotic arm guides it to automatically connect the high-voltage test line to the electrodes of the electrical appliance under test. After connection, the robotic arm is controlled to retreat to a safe area based on the insulation distance. Subsequently, a high voltage is applied according to the test voltage for testing, including: The ground-rail robotic arm calls an adaptive gripper that matches the gripper type from the tool library, wherein the adaptive gripper has a built-in pressure sensor; The ground-rail robotic arm carries the adaptive gripper to the picking point, grips the electrical tool to be tested according to its gripping position, and controls the gripping force through the pressure sensor. The electrical tool under test is transported to an automated testing platform, and the electrode position coordinates of the electrical tool under test are identified by a vision sensor at the end of the ground-rail robotic arm. The ground-rail robotic arm grasps the connector of the test high-voltage line and guides the connector of the test high-voltage line to be connected to the electrode of the electrical tool under test according to the electrode position coordinates to form an electrical connection. After the electrical connection is completed, the ground rail robotic arm loosens the connector of the test high voltage line, calculates the coordinates of the safe stopping point based on the insulation distance, plans the retreat path, and moves to the safe stopping point; After moving to the safe dwell point, the automated test platform performs electrical performance tests on the electrical equipment under test according to the test voltage; During the test, the positions of the ground-rail robotic arm and the automated guided vehicle are monitored in real time to ensure that the distance between all metal objects and the high-voltage electrodes of the electrical appliance under test is not less than the insulation distance.

9. The method for full-process testing of electrical tools based on industrial robots according to claim 8, characterized in that, Calculate the coordinates of the safe stopping point based on the insulation distance, plan the retreat path and move to the safe stopping point, including: A spherical restricted area is constructed with the high-voltage electrode position of the electrical appliance under test as the center and the sum of the insulation distance and the preset safety margin as the radius. Starting from the current position of the ground-rail robotic arm, search for reachable poses of the ground-rail robotic arm outside the spherical restricted area, and use the searched reachable poses as a set of candidate dwell points; Select the reachable pose closest to the starting point from the set of candidate dwell points as the coordinates of the safe dwell point; The retreat path is planned using the artificial potential field method, and a repulsive field is constructed at the location of the high-voltage electrode, so that the ground-rail robotic arm is subjected to an attractive force pointing towards the coordinates of the safe stopping point and a repulsive force moving away from the spherical restricted area during the retreat process.

10. A full-process electrical tool testing system based on industrial robots, characterized in that, The method for performing the full-process electrical tool testing method based on industrial robots as described in any one of claims 1-9 includes: The outbound scheduling module is used to schedule automated guided vehicles to move the electrical tools to be tested from the smart shelf to the pre-inspection station according to the testing task; The feature acquisition module is used to acquire the multi-dimensional physical features of the electrical tool under test at the pre-inspection station. The multi-dimensional physical features include geometric features, material features, and surface features. The feature encoding module is used to input the multidimensional physical features into a pre-trained autoencoder model and extract the corresponding low-dimensional latent feature vectors. The parameter matching module is used to determine the cluster category of the electrical appliance under test based on the low-dimensional potential feature vector, and to query the detection parameter set from the detection parameter knowledge graph according to the cluster category. The detection parameter set includes insulation distance, test voltage and fixture type. An automatic testing module is used for a ground-rail robotic arm to grasp the electrical appliance under test according to the type of clamp and move it to an automated testing platform. The vision sensor at the end of the ground-rail robotic arm guides the robotic arm to complete the automatic docking of the test high-voltage line with the electrode of the electrical appliance under test. After docking, the ground-rail robotic arm is controlled to retreat to a safe area according to the insulation distance. Then, a high voltage is applied according to the test voltage to perform the test. The inbound scheduling module is used to schedule the automated guided vehicle to classify and return the electrical tools under test to the smart shelf based on the test results.

Images