An object detection algorithm and system for a power robot

Abstract

The present invention discloses a target detection algorithm and system for a power robot. The method includes: S1: Equip with multiple sensors, a local processor, an edge server, and provide a data transmission channel between the power robot and the edge server; S2: Use a generative adversarial network to generate realistic synthetic data to expand the training data set; S3: Load a pre-trained convolutional neural network model; S4: Adaptively adjust the parameters and structure of the deep learning model to optimize the parameters and strategies for target detection; S5: Update the parameters of the ontology model; S6: Achieve the obstacle avoidance function by planning a safe path; S7: Continuously optimize the reinforcement learning strategy of the deep learning model. The present invention generates realistic synthetic data through a generative adversarial network, significantly expanding the training data set, solving the problems of high cost of obtaining real data and difficulty in annotation, and also improving the diversity and representativeness of the data by introducing conditional inputs to generate synthetic data that conforms to the characteristics of a specific environment.

Description

Technical Field

[0001] The present invention relates to the technical field of power robots, and in particular, to a power robot target detection algorithm and its system. Background Art

[0002] With the rapid development of the power industry, power robots are playing an increasingly important role in tasks such as power grid inspection, equipment maintenance, and fault troubleshooting. However, when performing these tasks, power robots face complex and changing environmental challenges, such as light changes, occlusions, and bad weather, which pose higher requirements for the target detection ability of power robots.

[0003] Traditional target detection algorithms mainly rely on a large amount of labeled real data for training. However, the acquisition cost of real data in the power field is high, the labeling difficulty is large, and there are often problems such as data imbalance and single features, resulting in limited generalization ability of the trained model and difficulty in adapting to complex and changing environments. In addition, when performing tasks, power robots need to process a large amount of sensor data in real time and perform operations such as target detection, tracking, and obstacle avoidance, which poses a severe challenge to the computing power and real-time performance of power robots.

[0004] In recent years, with the continuous development of deep learning technology, especially the rise of technologies such as generative adversarial networks (GANs) and reinforcement learning, new ideas have been provided for the innovation of power robot target detection algorithms. GANs can expand the training dataset by generating realistic synthetic data to solve the problem of difficult acquisition of real data; while reinforcement learning can dynamically adjust model parameters and strategies according to environmental feedback to improve the adaptability and robustness of the model.

[0005] However, most current power robot target detection algorithms on the market still have the following problems: First, the way of expanding the dataset is single, and it is difficult to generate synthetic data that conforms to the characteristics of a specific environment; second, there is a lack of an adaptive optimization mechanism and it is unable to dynamically adjust model parameters and strategies according to environmental changes; third, the obstacle avoidance function is not accurate enough to ensure the safety and stability of power robots during task execution. Summary of the Invention

[0006] The purpose of this part is to outline some aspects of the embodiments of the present invention and briefly introduce some preferred embodiments. Some simplifications or omissions may be made in this part, as well as in the abstract and title of the present application, to avoid obscuring the purpose of this part, the abstract, and the title, and such simplifications or omissions shall not be used to limit the scope of the present invention.

[0007] To solve the above technical problems, the present invention provides the following technical solution: A power robot target detection algorithm, including the following steps:

[0008] S1: Equip with multiple sensors, a local processor, an edge server, and provide a data transmission channel between the power robot and the edge server;

[0009] S2: Collect environmental data from multiple sensors, synchronize the data in time, and then use a generative adversarial network to generate realistic synthetic data to expand the training dataset;

[0010] S3: Load a pre-trained convolutional neural network model for preliminary object detection, and then run the preliminary object detection algorithm on the local processor to detect objects in the environment through the convolutional neural network and generate preliminary results;

[0011] S4: Extract environmental features from the sensor data, adaptively adjust the parameters and structure of the deep learning model according to the environmental features, and then optimize the parameters and strategies of object detection through a reinforcement learning algorithm;

[0012] S5: Assign the preliminary detection results and complex computing tasks to the edge server for processing. The edge server returns the processing results to the local processor and updates the ontology model parameters;

[0013] S6: Use an adaptive filtering algorithm to filter environmental noise, utilize particle filtering technology to achieve real-time tracking of the target, combine sensor data and detection results to identify obstacles in the environment, and realize the obstacle avoidance function by planning a safe path;

[0014] S7: Real-time monitor the detection performance of the algorithm, and continuously optimize the reinforcement learning strategy of the deep learning model according to the feedback data.

[0015] As a preferred solution of the object detection algorithm for the power robot described in the present invention, wherein: the sensors include a camera, an infrared sensor, and a lidar for obtaining environmental data.

[0016] As a preferred solution of the object detection algorithm for the power robot described in the present invention, wherein: the specific steps for generating the realistic synthetic data are as follows:

[0017] S21: Collect environmental data from multiple sensors, and clean and preprocess the collected data;

[0018] S22: Design a generator and a discriminator network structure. The generator network takes a random noise vector as input and generates realistic synthetic data. The discriminator network inputs real data and the generated data and outputs the authenticity probability of the data;

[0019] S23: Through an adversarial training method, jointly train the generator and the discriminator, use the cross-entropy loss function to optimize the generator and the discriminator respectively, and use the Adam optimization algorithm for model training;

[0020] S24: Introduce conditional inputs into the generator and discriminator. By combining environmental features with random noise vectors, generate synthetic data that meets specific conditions;

[0021] S25: Use the structural similarity index and peak signal-to-noise ratio to evaluate the quality of the generated data;

[0022] S26: Integrate the generated synthetic data with real data to expand the training dataset;

[0023] S27: Use the expanded dataset to train and validate the object detection model, evaluate the performance of the model under different environmental conditions, and adjust the parameters and architecture of the generative adversarial network according to the validation results.

[0024] As a preferred solution of the object detection algorithm for the power robot described in the present invention, wherein: the joint training of the generator and discriminator includes the following steps:

[0025] S231: Data preparation

[0026] Randomly extract a batch of real data from the real dataset P data (x);

[0027] Randomly extract a batch of noise vectors from the random noise distribution P z (Z);

[0028] S232: Generate data

[0029] Use the generator to generate a batch of synthetic data

[0030] S233: Calculate the discriminator loss

[0031] For each real data x i , calculate the discriminator output D(xi ) ;

[0032] For each generated data Calculate the discriminator output

[0033] Calculate the loss L of the discriminator D ;

[0034] S234: Update the discriminator

[0035] Backpropagate to calculate the gradient of the discriminator and update the parameters of the discriminator using the Adam optimization algorithm;

[0036] S235: Calculate the generator loss

[0037] For each generated data Calculate the discriminator output

[0038] Calculate the loss L of the generator G ;

[0039] S236: Update the generator

[0040] Backpropagate to calculate the gradient of the generator and update the parameters of the generator using the Adam optimization algorithm;

[0041] S237: Repeat steps S231 to S236 until the generator and discriminator reach the desired performance.

[0042] As a preferred solution of the power robot target detection algorithm described in the present invention, wherein: the loss function of the generator is:

[0043]

[0044] Wherein:

[0045] represents the number of data in a batch;

[0046] x i represents the i-th real data sample;

[0047] z i represents the i-th noise vector;

[0048] G(z i ) represents the i-th synthetic data generated by the generator;

[0049] D(x i ) represents the output of the discriminator for the real data sample x i and represents the probability that the sample is discriminated as real data;

[0050] D(G(z i )) represents the output of the discriminator for the generated data G(z i ) and represents the probability that the sample is discriminated as real data;

[0051] logD(x i ) represents the log probability that the real data is discriminated as real data;

[0052] log(1 - D(G(z i ))) represents the log probability that the generated data is discriminated as fake data;

[0053] The loss function of the generator is:

[0054]

[0055] Wherein:

[0056] m represents the number of data in a batch;

[0057] z i represents the i-th noise vector;

[0058] G(z i ) represents the i-th synthetic data generated by the generator;

[0059] D(G(z i )) represents the output of the discriminator for the generated data G(z i ), representing the probability that the sample is judged as real data;

[0060] log(D(G(z i ))) represents the log probability that the generated data is judged as real data.

[0061] As a preferred solution of the power robot target detection algorithm described in the present invention, wherein: the data time synchronization is performed by means of timestamp alignment and data interpolation to ensure the spatio-temporal consistency of the data of all sensors at the same time point.

[0062] As a preferred solution of the power robot target detection algorithm described in the present invention, wherein: the generator and discriminator network structures of the generative adversarial network are multi-layer convolutional neural networks. The generator network includes a transposed convolution layer and a batch normalization layer, and the discriminator network includes a convolution layer and a LeakyReLU activation function.

[0063] As a preferred solution of the power robot target detection algorithm described in the present invention, wherein: the method for adaptively adjusting the parameters and structure of the deep learning model includes using a genetic algorithm or a Bayesian optimization method to adaptively select the best network structure and parameter combination according to the environmental characteristics.

[0064] As a preferred solution of the power robot target detection algorithm described in the present invention, wherein: the detection performance of the real-time monitoring algorithm is adjusted by setting multiple performance indicators and dynamically adjusting the training strategy and parameters of the deep learning model according to these indicators.

[0065] A system using the above power robot target detection algorithm includes:

[0066] A sensor module, which is used to obtain multi-dimensional data in the environment through a variety of sensors and transmit the data to a local processor;

[0067] A data processing module, which is used to synchronize, clean, and preprocess the collected data to ensure the spatio-temporal consistency of the data;

[0068] A generative adversarial network module, which is used to generate synthetic data through a generative adversarial network and fuse it with real data to expand the training dataset and improve the robustness of the model;

[0069] A target detection module, which is used to load a pre-trained convolutional neural network model, perform preliminary target detection on a local processor, and generate detection results;

[0070] An adaptive adjustment module, which is used to adjust the parameters and structure of the deep learning model according to environmental features, and optimize the parameters and strategies of target detection through a reinforcement learning algorithm;

[0071] An edge computing module, which is used to allocate preliminary detection results and complex computing tasks to an edge server for processing, and return the processing results to the local processor to update the model parameters;

[0072] A real-time tracking module, which uses adaptive filtering and particle filtering techniques to perform real-time tracking on the target, combines sensor data to identify obstacles, and realizes obstacle avoidance through path planning;

[0073] A performance monitoring module, which is used to monitor the detection performance of the algorithm in real time, optimize the reinforcement learning strategy of the deep learning model according to the feedback data, and improve the overall performance of the system.

[0074] Advantages of the present invention:

[0075] 1. The present invention generates realistic synthetic data through a generative adversarial network (GAN), significantly expanding the training dataset. This method not only solves the problems of high cost and difficult annotation of real data acquisition, but also generates synthetic data that conforms to specific environmental features by introducing conditional inputs, improving the diversity and representativeness of the data. At the same time, the quality of the generated data is evaluated using the structural similarity index and peak signal-to-noise ratio, ensuring the high quality of the training data, thereby improving the generalization ability and accuracy of the target detection model.

[0076] 2. In the algorithm of the present invention, a local processor is used for preliminary target detection, and complex computing tasks are allocated to an edge server for processing. This distributed processing architecture effectively reduces the computing burden of the power robot, improves the real-time performance of target detection. At the same time, the powerful computing ability of the edge server can quickly process complex tasks and feedback the processing results to the local processor in a timely manner, realizing efficient target detection and obstacle avoidance functions.

[0077] 3. In the present invention, the algorithm can adaptively adjust the parameters and structure of the deep learning model according to environmental characteristics, and optimize the parameters and strategies of object detection through the reinforcement learning algorithm. This adaptive optimization mechanism enables the power robot to better adapt to complex and changing environments, improving the accuracy and robustness of object detection. At the same time, the detection performance of the real-time monitoring algorithm is monitored, and the reinforcement learning strategy of the deep learning model is continuously optimized according to the feedback data, further enhancing the overall performance of the algorithm.

[0078] 4. The present invention can extract environmental characteristics from sensor data and adaptively adjust the parameters and structure of the deep learning model. Additionally, by optimizing the parameters and strategies of object detection through the reinforcement learning algorithm, it not only improves the detection efficiency and accuracy but also can dynamically adjust according to environmental changes to ensure continuous and efficient detection performance.

[0079] 5. The present invention uses an adaptive filtering algorithm and particle filtering technology to achieve real-time tracking of the target, combines sensor data and detection results to identify obstacles in the environment, and realizes the obstacle avoidance function of the power robot by planning a safe path. This precise real-time tracking and obstacle avoidance ability ensure the safety and stability of the power robot during task execution. BRIEF DESCRIPTION OF THE DRAWINGS

[0080] In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the following will briefly introduce the drawings required for the description of the embodiments. Obviously, the following drawings are only some embodiments of the present invention. For those of ordinary skill in the art, other drawings can be obtained based on these drawings without creative efforts. Among them:

[0081] Figure 1 It is a flow chart of an object detection algorithm for a power robot according to the present invention. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0082] To make the above objects, features, and advantages of the present invention more apparent and understandable, the following will provide a detailed description of the specific embodiments of the present invention with reference to the accompanying drawings of the specification.

[0083] In the following description, many specific details are set forth to fully understand the present invention. However, the present invention can also be implemented in other ways different from those described herein. Those skilled in the art can make similar extensions without departing from the connotation of the present invention. Therefore, the present invention is not limited by the specific embodiments disclosed below.

[0084] Second, the "one embodiment" or "embodiment" mentioned herein refers to specific features, structures, or characteristics that may be included in at least one implementation of the present invention. The appearances of "in one embodiment" in different places in this specification do not all refer to the same embodiment, nor are they separate or alternative embodiments that are mutually exclusive of other embodiments.

[0085] Thirdly, the present invention is described in detail in conjunction with schematic diagrams. When detailing the embodiments of the present invention, for ease of explanation, the cross-sectional views showing the device structure will be enlarged locally not in accordance with the general scale, and the schematic diagrams are only examples and should not limit the scope of protection of the present invention herein. In addition, in actual production, three-dimensional spatial dimensions including length, width, and depth should be included.

[0086] Embodiment 1

[0087] Referring to Figure 1 , the first embodiment of the present invention provides a target detection algorithm for a power robot, including the following steps:

[0088] S1: Equip with multiple sensors, a local processor, an edge server, and provide a data transmission channel between the power robot and the edge server;

[0089] Among them, the multiple sensors include cameras, infrared sensors, lidar, etc., which are used to obtain environmental data;

[0090] The local processor is used to perform preliminary data processing and target detection, and the edge server is used to process complex computing tasks and model updates.

[0091] S2: Collect environmental data from multiple sensors, including images, infrared data, and lidar point cloud data, and synchronize the data in time to ensure the spatio-temporal consistency of the data. Then, use a generative adversarial network (GAN) to generate realistic synthetic data, expand the training dataset, and improve the robustness of the model;

[0092] Among them, data time synchronization is carried out by methods such as timestamp alignment and data interpolation to ensure that the data of all sensors has spatio-temporal consistency at the same time point.

[0093] The specific steps for generating the realistic synthetic data are as follows:

[0094] S21: Collect environmental data from multiple sensors (such as cameras, infrared sensors, lidar), and these data include images, infrared images, and point cloud data. Then, clean and preprocess the collected data, such as denoising, normalization, and data augmentation (such as rotation, scaling, and flipping).

[0095] S22: Generative adversarial network (GAN) model design

[0096] Generator Design: Design a generator network that takes a random noise vector as input and generates realistic synthetic data. For example, the generator can adopt a Convolutional Neural Network (CNN) architecture to map the noise vector to a high-dimensional image space.

[0097] Discriminator Design: Design a discriminator network that takes real data and generated data as input and outputs the probability of the data being real. The discriminator can also adopt a CNN architecture to distinguish between real data and synthetic data.

[0098] The generator and discriminator network structures of the Generative Adversarial Network are multi-layer convolutional neural networks. Through the multi-layer convolutional neural network structure, high-level feature representations can be extracted and generated layer by layer in the generator and discriminator. This structure helps to deepen the complexity of the network, improve the expressive ability of the model, and enhance the authenticity of the generated data.

[0099] The generator network includes deconvolution layers and batch normalization layers. The deconvolution layers in the generator can expand the low-dimensional random noise vector into high-dimensional synthetic data, which helps to generate high-resolution and detail-rich images. The batch normalization layers in the generator can normalize each batch of data, reduce internal covariate shift, stabilize and accelerate the training process, and improve the quality and diversity of the generated data.

[0100] The discriminator network includes convolutional layers and LeakyReLU activation functions. The convolutional layers in the discriminator can extract local features and hierarchical representations of the data, which are very important for identifying the authenticity of images and detecting targets. The LeakyReLU activation functions in the discriminator can avoid the "neuron death" problem of traditional ReLU (i.e., some neurons may completely become zero during training and thus no longer update), while retaining the ability of non-linear transformation. This can improve the stability and performance of the discriminator and enhance its ability to distinguish between generated data and real data.

[0101] S23: Through the adversarial training method, jointly train the generator and the discriminator. The generator continuously generates more realistic synthetic data, while the discriminator continuously improves its ability to distinguish between real data and synthetic data;

[0102] Use the cross-entropy loss function to optimize the generator and the discriminator respectively. The goal of the generator is to maximize the probability that the discriminator judges the generated data as real data, while the goal of the discriminator is to minimize the probability of misjudgment;

[0103] Use the Adam optimization algorithm for model training to ensure the stability and convergence of the training process.

[0104] S24: Introduce conditional inputs into the generator and discriminator, such as environmental features (e.g., lighting conditions, obstacle distribution, etc.), and generate synthetic data that meets specific conditions by combining environmental features with random noise vectors.

[0105] In the CGAN architecture, the generator and discriminator not only accept random noise or data as inputs, but also accept environmental features as additional inputs to generate more targeted synthetic data.

[0106] S25: Use metrics such as the Structural Similarity Index (SSIM) and Peak Signal-to-Noise Ratio (PSNR) to evaluate the quality of the generated data, ensuring that the generated data is visually similar to the real data and representative in the feature space.

[0107] Professional personnel can also be invited to subjectively evaluate the generated data to ensure that the generated data has high credibility under different environmental conditions.

[0108] S26: Fuse the generated synthetic data with the real data to expand the training dataset, and in this way, improve the robustness of the model under various environmental conditions.

[0109] Before fusing the generated synthetic data with the real data, it is necessary to annotate the generated synthetic data to ensure that its labels are consistent with the real data. For example, use semi-automated annotation tools to perform object annotation on the generated data to improve the annotation efficiency.

[0110] S27: Use the expanded dataset to train and validate the object detection model, evaluate the performance of the model under different environmental conditions, and adjust the parameters and architecture of the generative adversarial network according to the validation results to optimize the quality and diversity of the generated data.

[0111] Through the above steps, a large number of realistic synthetic data can be generated to expand the training dataset for power robot object detection. These data cover a variety of environmental conditions and improve the robustness of the deep learning model.

[0112] S3: Load a pre-trained Convolutional Neural Network (CNN) model for preliminary object detection, and then run the preliminary object detection algorithm on a local processor to detect objects in the environment through the convolutional neural network and generate preliminary results.

[0113] S4: Extract environmental features from the sensor data, and adaptively adjust the parameters and structure of the deep learning model according to the environmental features. For example, in the case of insufficient light, increase the weight of the infrared data; then optimize the parameters and strategies of object detection through a reinforcement learning algorithm (such as Deep Q-Network, DQN) to improve the detection efficiency and accuracy.

[0114] Among them, the method for adaptively adjusting the parameters and structure of the deep learning model includes using a genetic algorithm or a Bayesian optimization method to adaptively select the best network structure and parameter combination according to environmental characteristics.

[0115] S5: Assign the preliminary detection results and complex calculation tasks to the edge server for processing. The edge server runs high-performance computing tasks, such as complex inference and model update of the deep model. The edge server returns the processing results to the local processor and updates the ontology model parameters.

[0116] S6: Use an adaptive filtering algorithm (such as Kalman filtering) to filter environmental noise and improve the accuracy of target detection. Utilize techniques such as particle filtering to achieve real-time tracking of the target and ensure continuous detection of the target in a complex environment. In addition, combine sensor data and detection results to identify obstacles in the environment and achieve the obstacle avoidance function by planning a safe path.

[0117] S7: Real-time monitor the detection performance of the algorithm, such as accuracy, detection speed, etc. According to the feedback data, continuously optimize the reinforcement learning strategy of the deep learning model to improve the overall performance of the system. Among them, the detection performance of the real-time monitoring algorithm is set by multiple performance indicators, and the training strategy and parameters of the deep learning model are dynamically adjusted according to these indicators.

[0118] Specifically, the joint training of the generator and discriminator includes the following steps:

[0119] S231: Data preparation

[0120] Randomly extract a batch of real data from the real dataset P data (x)

[0121] Randomly extract a batch of noise vectors from the random noise distribution P z (Z) The noise vector is used as the input of the generator to generate new synthetic data;

[0122] S232: Generate data

[0123] Use the generator to generate a batch of synthetic data That is, use the generator to convert the noise vector into synthetic data for training the discriminator so that it can distinguish between generated data and real data;

[0124] S233: Calculate the discriminator loss

[0125] For each real data x i , calculate the discriminator output D(x i )

[0126] For each generated data Calculate the discriminator output

[0127] Calculate the discriminator loss L D , The discriminator loss is used to measure the performance of the discriminator in distinguishing real data and generated data. The goal is to maximize this loss so that the discriminator can accurately distinguish the two types of data;

[0128] S234: Update the discriminator

[0129] Calculate the gradient of the discriminator through backpropagation and update the parameters of the discriminator using the Adam optimization algorithm. Updating the parameters of the discriminator through backpropagation and the optimization algorithm can improve its ability to distinguish real data and generated data;

[0130] S235: Calculate the generator loss

[0131] For each generated data Calculate the discriminator output

[0132] Calculate the generator loss L G , The generator loss is used to measure the performance of the generator in deceiving the discriminator. The goal is to minimize this loss so that the generated data is closer to the real data, thus deceiving the discriminator;

[0133] S236: Update the generator

[0134] Calculate the gradient of the generator through backpropagation and update the parameters of the generator using the Adam optimization algorithm. Updating the parameters of the generator through backpropagation and the optimization algorithm improves its ability to generate high-quality synthetic data, making it difficult for the discriminator to distinguish between generated data and real data;

[0135] S237: Repeat steps S231 to S236 until the generator and discriminator reach the desired performance.

[0136] During the process of jointly training the generator and the discriminator, by repeatedly iterating these steps, continuously optimizing the generator and the discriminator, the generator can generate more realistic data, while the discriminator can more accurately distinguish between real data and generated data. This adversarial training method ultimately improves the quality of the generated data and the robustness of the system.

[0137] As a preferred embodiment of the power robot target detection algorithm described in the present invention, wherein: The loss function of the generator is:

[0138]

[0139] Wherein:

[0140] m represents the number of data in a batch;

[0141] x i represents the i-th real data sample;

[0142] z i represents the i-th noise vector;

[0143] G(z i ) represents the i-th synthetic data generated by the generator;

[0144] D(x i ) represents the output of the discriminator for the real data sample x i , indicating the probability that the sample is judged as real data;

[0145] D(G(z i )) represents the output of the discriminator for the generated data G(z i ), indicating the probability that the sample is judged as real data;

[0146] log D(x i ) represents the log probability that the real data is judged as real data. The discriminator should output a value close to 1, so the value of this term should be as large as possible.

[0147] log(1 - D(G(z i ))) represents the log probability that the generated data is judged as fake data. The discriminator should output a value close to 0, so the value of this term should be as large as possible.

[0148] The goal of the discriminator is to maximize this loss function to improve its ability to distinguish real data and generated data.

[0149] The loss function of the generator is as follows:

[0150]

[0151] where:

[0152] m represents the number of data in a batch;

[0153] z i represents the i-th noise vector;

[0154] G(z i ) represents the i-th synthetic data generated by the generator;

[0155] D(G(z i )) represents the output of the discriminator for the generated data G(z i ), indicating the probability that the sample is judged as real data;

[0156] log(D(G(z i))) represents the log probability that the generated data is judged to be real data. The generator hopes that the discriminator believes that the generated data is real, so the value of this item should be as large as possible.

[0157] The goal of the generator is to minimize this loss function to improve the authenticity of the data it generates, so that the discriminator cannot distinguish between the generated data and the real data.

[0158] A system using the above-mentioned power robot target detection algorithm includes:

[0159] A sensor module, which is used to obtain multi-dimensional data in the environment through multiple sensors and transmit the data to the local processor;

[0160] A data processing module, which is used to synchronize, clean and preprocess the collected data in terms of time to ensure the spatio-temporal consistency of the data;

[0161] A generative adversarial network module, which is used to generate synthetic data through a generative adversarial network, fuse it with real data, expand the training data set, and improve the robustness of the model;

[0162] A target detection module, which is used to load a pre-trained convolutional neural network model, perform preliminary target detection on the local processor, and generate detection results;

[0163] An adaptive adjustment module, which is used to adjust the parameters and structure of the deep learning model according to environmental characteristics, and optimize the parameters and strategies of target detection through a reinforcement learning algorithm;

[0164] An edge computing module, which is used to allocate the preliminary detection results and complex calculation tasks to the edge server for processing, and return the processing results to the local processor to update the model parameters;

[0165] A real-time tracking module, which uses adaptive filtering and particle filtering techniques to perform real-time tracking on the target, combines sensor data to identify obstacles, and realizes the obstacle avoidance function through path planning;

[0166] A performance monitoring module, which is used to monitor the detection performance of the algorithm in real time, optimize the reinforcement learning strategy of the deep learning model according to the feedback data, and improve the overall performance of the system.

[0167] In summary, the present invention generates realistic synthetic data through a generative adversarial network (GAN), significantly expanding the training dataset. This method not only solves the problems of high cost and difficult annotation of real data acquisition, but also generates synthetic data that conforms to the characteristics of a specific environment by introducing conditional inputs, improving the diversity and representativeness of the data. At the same time, the quality of the generated data is evaluated using the structural similarity index and peak signal-to-noise ratio, ensuring the high quality of the training data, thereby enhancing the generalization ability and accuracy of the object detection model. In the algorithm of the present invention, a local processor is used for preliminary object detection, and complex computing tasks are assigned to an edge server for processing. This distributed processing architecture effectively reduces the computing burden of the power robot and improves the real-time performance of object detection. At the same time, the powerful computing ability of the edge server can quickly process complex tasks and timely feedback the processing results to the local processor, realizing efficient object detection and obstacle avoidance functions. The algorithm in the present invention can adaptively adjust the parameters and structure of the deep learning model according to the environmental characteristics, and optimize the parameters and strategies of object detection through a reinforcement learning algorithm. This adaptive optimization mechanism enables the power robot to better adapt to complex and changing environments, improving the accuracy and robustness of object detection. At the same time, the detection performance of the real-time monitoring algorithm is monitored, and the reinforcement learning strategy of the deep learning model is continuously optimized according to the feedback data, further enhancing the overall performance of the algorithm. The present invention can extract environmental characteristics from sensor data and adaptively adjust the parameters and structure of the deep learning model. In addition, by optimizing the parameters and strategies of object detection through a reinforcement learning algorithm, it not only improves the detection efficiency and accuracy, but also can dynamically adjust according to environmental changes, ensuring continuous and efficient detection performance. The present invention uses an adaptive filtering algorithm and particle filtering technology to achieve real-time tracking of the target, and combines sensor data and detection results to identify obstacles in the environment. By planning a safe path, the obstacle avoidance function of the power robot is realized. This precise real-time tracking and obstacle avoidance ability ensures the safety and stability of the power robot during the task execution process.

[0168] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and not to limit them. Although the present invention has been described in detail with reference to the preferred embodiments, those of ordinary skill in the art should understand that the technical solutions of the present invention can be modified or equivalently replaced without departing from the spirit and scope of the technical solutions of the present invention, and they should all be covered by the scope of the claims of the present invention.

Claims

1. A target detection algorithm for a power robot, characterized in that, It includes the following steps: S1: Equip with multiple sensors, a local processor, an edge server, and provide a data transmission channel between the power robot and the edge server; S2: Collect environmental data from multiple sensors, perform time synchronization on the data, and then use a generative adversarial network to generate realistic synthetic data to expand the training dataset; S21: Collect environmental data from multiple sensors, and clean and preprocess the collected data; S22: Design a generator and a discriminator network structure. The generator network takes a random noise vector as input and generates realistic synthetic data. The discriminator network inputs real data and generated data and outputs the authenticity probability of the data; S23: Through an adversarial training method, jointly train the generator and the discriminator, use the cross-entropy loss function to optimize the generator and the discriminator respectively, and use the Adam optimization algorithm for model training; S24: Introduce conditional inputs into the generator and the discriminator. By combining environmental features with the random noise vector, generate synthetic data that meets specific conditions; S25: Use the structural similarity index and the peak signal-to-noise ratio to evaluate the quality of the generated data; S26: Fuse the generated synthetic data with the real data to expand the training dataset; S27: Use the expanded dataset to train and validate the object detection model, evaluate the performance of the model under different environmental conditions, and adjust the parameters and architecture of the generative adversarial network according to the validation results; S3: Load a pre-trained convolutional neural network model for preliminary object detection, and then run the preliminary object detection algorithm on the local processor to detect objects in the environment through the convolutional neural network and generate preliminary results; S4: Extract environmental features from the sensor data, adaptively adjust the parameters and structure of the deep learning model according to the environmental features, and then optimize the parameters and strategies of object detection through a reinforcement learning algorithm; S5: Assign the preliminary detection results and complex calculation tasks to the edge server for processing. The edge server returns the processing results to the local processor and updates the ontology model parameters; S6: Use an adaptive filtering algorithm to filter environmental noise, use particle filtering technology to achieve real-time tracking of the target, combine sensor data and detection results, identify obstacles in the environment, and achieve the obstacle avoidance function by planning a safe path; S7: Real-time monitor the detection performance of the algorithm, and continuously optimize the reinforcement learning strategy of the deep learning model according to the feedback data.

2. The power robot target detection algorithm according to claim 1, characterized in that: The sensors include cameras, infrared sensors, and lidar for obtaining environmental data.

3. The power robot target detection algorithm according to claim 1, wherein: The joint training of the generator and the discriminator includes the following steps: S231: Data preparation Randomly extract a batch of real data from the real dataset P data (x) Randomly sample a batch of noise vectors from the random noise distribution P z (Z) S232: Generate data Generate a batch of synthetic data using a generator S233: Calculate the discriminator loss For each real data x i , compute the discriminator output D(x i ); For each generated data Calculate the discriminator output Calculate the discriminator loss L D ; S234: Update the discriminator Backpropagate to calculate the gradient of the discriminator and use the Adam optimization algorithm to update the parameters of the discriminator; S235: Calculate the generator loss For each generated data Calculate the discriminator output Calculate the loss L of the generator G ; S236: Update the generator Backpropagate to calculate the gradient of the generator and use the Adam optimization algorithm to update the parameters of the generator; S237: Repeat steps S231 to S236 until the generator and the discriminator reach the expected performance.

4. The power robot target detection algorithm according to claim 3, wherein: The loss function of the discriminator is as follows: Where: m represents the number of data in a batch; x i represents the i-th true data sample; z i represents the i-th noise vector; G(z i ) represents the i-th synthetic data generated by the generator; D(x i ) represents the output of the discriminator for the real data sample x i , indicating the probability that the sample is discriminated as real data; D(G(z i )) represents the output of the discriminator for the generated data G(z i ), indicating the probability that the sample is judged as real data; logD(x i ) represents the logarithmic probability that the true data is judged as true data; log(1 - D(G(z i ))) represents the logarithmic probability that the generated data is judged as fake data; The loss function of the generator is as follows: Where: m represents the number of data in a batch; z i represents the i-th noise vector; G(z i ) represents the i-th synthetic data generated by the generator; D(G(z i )) represents the output of the discriminator for the generated data G(z i ), indicating the probability that the sample is discriminated as real data; log(D(G(z i ))) represents the log probability that the generated data is judged to be real data.

5. The power robot target detection algorithm according to claim 1, wherein: The data time synchronization is carried out by means of timestamp alignment and data interpolation to ensure the spatio-temporal consistency of the data of all sensors at the same time point.

6. The power robot target detection algorithm according to claim 1, characterized in that: The generator and discriminator network structures of the generative adversarial network are multi-layer convolutional neural networks. The generator network includes deconvolution layers and batch normalization layers, and the discriminator network includes convolutional layers and LeakyReLU activation functions.

7. The power robot target detection algorithm according to claim 1, characterized in that: The method for adaptively adjusting the parameters and structure of the deep learning model includes using genetic algorithms or Bayesian optimization methods to adaptively select the best network structure and parameter combination according to environmental characteristics.

8. The target detection algorithm of the power robot according to claim 1, characterized in that: The detection performance of the real-time monitoring algorithm is adjusted by setting multiple performance metrics and dynamically adjusting the training strategy and parameters of the deep learning model according to these metrics.

9. The system of the power robot target detection algorithm according to any one of claims 1-8, characterized in that, Including: A sensor module, which is used to obtain multi-dimensional data in the environment through multiple sensors and transmit the data to the local processor; A data processing module, which is used to perform time synchronization, cleaning and preprocessing on the collected data to ensure the spatio-temporal consistency of the data; A generative adversarial network module, which is used to generate synthetic data through a generative adversarial network and fuse it with real data to expand the training dataset and improve the robustness of the model; A target detection module, which is used to load a pre-trained convolutional neural network model, perform preliminary target detection on the local processor, and generate detection results; An adaptive adjustment module, which is used to adjust the parameters and structure of the deep learning model according to environmental characteristics and optimize the parameters and strategies of target detection through reinforcement learning algorithms; An edge computing module, which is used to allocate the preliminary detection results and complex calculation tasks to the edge server for processing, and return the processing results to the local processor to update the model parameters; A real-time tracking module, which uses adaptive filtering and particle filtering techniques to track the target in real time, combines sensor data to identify obstacles, and realizes the obstacle avoidance function through path planning; A performance monitoring module, which is used to monitor the detection performance of the real-time monitoring algorithm, optimize the reinforcement learning strategy of the deep learning model according to the feedback data, and improve the overall performance of the system.

Images