A battery energy recycling method for underwater robots

Abstract

The invention discloses a battery energy recovery and utilization method for an underwater robot. The method comprises the following steps: integrating the equipment status and environmental data of the underwater robot, screening data by using a particle filter algorithm, capturing the depth information of the underwater environment in real time, collecting the battery material properties and energy transmission efficiency parameters, identifying the parameters by using a k-means clustering algorithm, constructing the constraint conditions of the energy recovery network, constructing the energy recovery network topology model according to the energy loss information and the node layout, obtaining the optimal path by using the Eulerian path algorithm, determining the arc curvature by using a non-dominated sorting genetic algorithm, dynamically updating the topology model parameters, obtaining the optimized energy recovery arc, constructing an arc optimization mechanism, scoring the arc by using a logistic regression algorithm, identifying the efficient nodes, simplifying the network topology by using an Apriori association rule mining algorithm, calculating the structural difference and updating the model, monitoring the network status and energy loss in real time by using a fitness function algorithm, and evaluating the transmission efficiency.

Description

Technical Field

[0001] The invention relates to the field of information technology, and in particular to a battery energy recovery and utilization method for an underwater robot. Background Art

[0002] In the topological design of the battery energy recovery system, arc optimization refers to the process of accurately designing and adjusting the curve shape in the energy transmission path through energy transmission theory and fluid dynamics to reduce energy loss during transmission and improve energy recovery efficiency. However, there is a key technical contradiction in the topological design of the energy recovery system, that is, the balance between arc optimization and structural complexity. On the one hand, arc optimization plays a vital role in the energy transmission process. By carefully designing and optimizing the arc in the transmission path, energy loss can be significantly reduced and the energy recovery efficiency of the entire system can be improved. The optimized arc can minimize the dissipation of electrical energy during transmission and ensure that more energy is effectively recovered and utilized.

[0003] On the other hand, excessive pursuit of arc optimization leads to a sharp increase in the complexity of the topological structure. In order to achieve an ideal arc, more components and connections are often required, which inevitably increases the complexity of the topological structure. The complex topological structure not only increases the difficulty of system implementation, but also leads to an increase in manufacturing costs and a decrease in reliability. At the same time, an overly complex structure also brings challenges to system maintenance and fault diagnosis.

[0004] Therefore, in topology design, how to find an appropriate balance between arc optimization and structural simplicity has become a technical problem that needs to be solved urgently. It is necessary to comprehensively weigh the energy recovery efficiency and implementation complexity. Through innovative design methods and ingenious structural layout, the complexity of the topological structure can be reduced as much as possible while ensuring a sufficiently high energy recovery efficiency. Summary of the invention

[0005] In order to solve the above-mentioned technical problems, the present invention provides a battery energy recovery and utilization method for an underwater robot.

[0006] The technical solution of the present invention is implemented as follows: a battery energy recycling method for an underwater robot, comprising:

[0007] Collect equipment status and environmental data from underwater robots and integrate the data to form a sample data set; use a particle filter algorithm to filter the data to capture and represent real-time depth information of the underwater environment; at the same time, collect battery material properties and energy transmission efficiency parameters, identify the property parameters through the k-means clustering algorithm, and construct the constraints of the energy recovery network;

[0008] According to the energy loss information and node layout information of the nodes in the energy recovery network, a preliminary energy recovery network topology model is constructed. The Eulerian path algorithm is applied to obtain the optimal path in the energy recovery process by combining real-time depth information and network constraints. According to the optimal path, the non-dominated sorting genetic algorithm is used to calculate the arc curvature corresponding to each node in the network and determine the preliminary energy recovery arc.

[0009] According to the network constraints, the objective function of maximizing the energy transmission efficiency is set, and the BP network algorithm is used to iteratively train the arc of the energy recovery network. The energy recovery network topology model parameters are dynamically updated by calculating the loss and efficiency of energy flowing through the arc nodes to obtain the optimized energy recovery arc.

[0010] Construct an arc optimization mechanism, use the logistic regression algorithm to score the complexity of the optimized energy recovery arc, identify nodes with simple structure and efficiency greater than the preset threshold, build an efficient topological node set, use the April association rule mining algorithm to analyze the causal relationship between nodes, guide the generation of a simplified network topology, calculate the structural difference and update the model to ensure that the energy recovery efficiency is maintained or improved while simplifying the network;

[0011] The fitness function algorithm is used to monitor the network status and energy loss in real time and evaluate the energy transmission efficiency. If the monitoring results show that the transmission efficiency decreases or the structural complexity increases, the energy transmission efficiency target and structural complexity of the current network are updated to the fitness function parameters, and the arc optimization mechanism is re-executed to obtain the optimized arc path;

[0012] An adaptive weight allocation algorithm is used to monitor the motion state and energy demand of the underwater robot in real time, and dynamically adjust the energy allocation weight of each node in the energy recovery network. When it is monitored that the network transmission efficiency decreases or the structural complexity increases, the arc optimization mechanism is triggered to generate a new optimal transmission arc.

[0013] Beneficial Effects

[0014] The present invention mainly solves the energy efficiency and network structure optimization problems of underwater robots in complex underwater environments. By integrating the device status and environmental data from the underwater robot, the present invention effectively captures the real-time depth information of the underwater environment and further optimizes the design and performance of the energy recovery network.

[0015] The present invention significantly improves the efficiency and reliability of the underwater robot energy recovery system through a series of advanced algorithms and real-time optimization mechanisms, especially in complex and dynamically changing underwater environments, effectively supporting the underwater robot's long-term operation and continuous performance maintenance. This is of great significance for improving the safety and economic benefits of underwater operations. BRIEF DESCRIPTION OF THE DRAWINGS

[0016] Figure 1 This is a structural block diagram of a battery energy recycling method for an underwater robot in an embodiment of the present invention;

[0017] Figure 2 The present invention is a flowchart of a battery energy recycling method for an underwater robot according to an embodiment of the present invention. DETAILED DESCRIPTION

[0018] In order to make the objects and advantages of the present invention more clearly understood, the present invention is further described below in conjunction with embodiments; it should be understood that the specific embodiments described herein are only used to explain the present invention and are not used to limit the present invention.

[0019] The preferred implementation methods of the present invention are described below with reference to the accompanying drawings. It should be understood by those skilled in the art that these implementation methods are only used to explain the technical principles of the present invention and are not intended to limit the protection scope of the present invention.

[0020] It should be noted that when an element is considered to be "connected" to another element, it can be directly connected to the other element, or connected to the other element through an intermediate element. In addition, the "connection" in the following embodiments should be understood as "electrical connection", "communication connection", etc. if there is transmission of electrical signals or data between the connected objects.

[0021] When used herein, the singular forms "a", "an", and "said / the" may also include plural forms, unless the context clearly indicates otherwise. It should also be understood that the terms "include / comprise" or "have" and the like specify the presence of stated features, wholes, steps, operations, components, parts, or combinations thereof, but do not exclude the possibility of the presence or addition of one or more other features, wholes, steps, operations, components, parts, or combinations thereof. At the same time, the term "and / or" used in this specification includes any and all combinations of the relevant listed items.

[0022] See also Figure 1-2 As shown, a battery energy recycling method for an underwater robot includes:

[0023] Step S101, collect equipment status and environmental data from the underwater robot, integrate the data to form a sample data set; use the particle filter algorithm to filter the data to capture and represent the real-time depth information of the underwater environment; at the same time, collect the battery material properties and energy transmission efficiency parameters, identify the property parameters through the k-means clustering algorithm, and construct the constraints of the energy recovery network.

[0024] Specifically, multi-dimensional equipment status data and environmental parameters are collected by underwater robots, and big data technology is used to clean, annotate and integrate the collected data to construct a sample data set that comprehensively reflects the characteristics of the underwater environment. Based on the constructed sample data set, a particle filter algorithm is used to screen and optimize the data, extract key characteristic parameters that can accurately characterize the real-time depth information of the underwater environment, and establish a digital model of the underwater environment depth information. The physicochemical property parameters of battery materials and the efficiency parameters of the energy transmission process are obtained through experimental tests and literature research. The k-means clustering algorithm is used to extract features and classify the property parameters, explore the intrinsic correlation between the parameters, and construct the constraints of the energy recovery network. The underwater environment depth information model and the energy recovery network constraints are comprehensively analyzed. The support vector machine (SVM) and neural network machine learning algorithm are used to establish the mapping relationship between depth information and energy recovery efficiency. The network structure and parameters are dynamically optimized through a genetic algorithm, and the optimization effect is evaluated using a fitness function. The population is iteratively updated until convergence is achieved to achieve structural and parameter configuration optimization of the energy recovery network.

[0025] In one embodiment, when the underwater robot collects equipment status and environmental parameter data, multiple sensors are used, including temperature sensors, pressure sensors, and depth sensors. The sampling frequency is set to 10 Hz, and 100 samples are collected for each parameter. The collected data is processed using big data technology, and the MapReduce algorithm in the Hadoop framework is used to clean and annotate the data. Through data quality assessment, abnormal data with an error of more than 5% is eliminated, and labels are set according to data characteristics. When constructing a digital model of underwater environmental depth information, a particle filter algorithm is used to initialize 100 particles, each particle containing depth, temperature, and pressure state variables. Through prediction and update steps, the real-time depth information of the underwater environment is estimated, and the Kalman filter algorithm is used to smooth the depth information. The filter parameter is set to 0.8. When obtaining battery material property parameters, the energy density and cycle life of the battery are measured through experimental tests. Parameters, and through literature research, the conductivity, dielectric constant and physical and chemical properties of different materials are obtained, and the k-means clustering algorithm is used to classify the attribute parameters. The number of clusters k is set to 5, and the number of iterations is 100. By calculating the distance from each sample to the cluster center, the samples are divided into the nearest clusters, and the cluster centers are continuously updated until the clustering results are stable. When establishing the mapping relationship between depth information and energy recovery efficiency, the support vector machine (SVM) algorithm is used, and the Gaussian kernel function is selected. The penalty factor C=10, the algorithm parameters are optimized through grid search and cross-validation, the number of training samples is 1000, the number of test samples is 200, and the average error of the predicted energy recovery efficiency is controlled within 5%. The genetic algorithm is used to optimize the energy recovery network, and the population size is set to 50, the crossover probability is 0.8, the mutation probability is 0.1, the fitness function is energy recovery efficiency, and it is iterated 500 times. The individual with the highest fitness is selected as the optimal solution.

[0026] Step S102, constructing a preliminary energy recovery network topology model based on the energy loss information and node layout information of the nodes in the energy recovery network; applying the Eulerian path algorithm, combined with real-time depth information and network constraints, to obtain the optimal path in the energy recovery process; based on the optimal path, using the non-dominated sorting genetic algorithm, calculating the arc curvature corresponding to each node in the network, and determining the preliminary energy recovery arc.

[0027] Specifically, by analyzing the energy loss information and node layout information of the nodes in the energy recovery network, comprehensively considering the node energy consumption characteristics and network topology structure, the Kruskal algorithm is used to model the network. Through the construction method of the minimum spanning tree, the total weight of the network is minimized while ensuring the connectivity of the network, so as to construct a preliminary energy recovery network topology model. According to the constructed energy recovery network topology model, the Eulerian path algorithm is used to traverse and optimize the network. Combined with the real-time underwater environment depth information and network operation constraints, the optimal path in the energy recovery process is solved by the dynamic programming method. The state is defined as the current node and the set of visited nodes. The state transfer equation is the transfer cost from the current node to the next node. The boundary conditions are the starting node and the ending node. The time complexity of the algorithm is O(n^2), where n is the number of network nodes. On the basis of determining the optimal energy recovery path, the improved non-dominated sorting genetic algorithm is used to further optimize the path. Through binary coding, tournament selection, single-point crossover and bit flip mutation operations, the population size, crossover probability, mutation probability and number of iterations are set, and the arc curvature parameters corresponding to each node in the network are calculated to determine the preliminary energy recovery arc.

[0028] In one embodiment, when the Kruska l algorithm is used to construct an energy recovery network topology model, the network nodes are first abstracted as vertices of the graph, and the connection relationship between the nodes is abstracted as the edge of the graph. The weight of the edge is determined according to the energy loss and distance factors, including the energy loss weight of 0.6 and the distance weight of 0.4. Then, the edge weights are selected in order from small to large. If adding the edge will not form a loop, it will be added to the minimum spanning tree until all nodes are connected. The final minimum spanning tree is the optimized topology of the energy recovery network. When applying Euler When the ian path algorithm solves the optimal path, the network nodes are abstracted as states, and the transfer relationship between nodes is abstracted as state transfer. The transfer cost takes into account energy loss, distance, and node residual energy factors, including energy loss cost of 0.1J per unit distance, distance cost of 0.01m per unit distance, and node residual energy cost of 0.05J per unit energy. Then, the dynamic programming algorithm is used to define the state dp[i][j] to represent the minimum cost from the starting point to node i and through the node set j. The state transfer equation is dp[i][j]=min(dp[k][j-{i}]+cost[k][i]), where k represents the predecessor node of node i, cost[k][i] represents the cost of transferring from node k to node i, and the optimal state dp[end][V ] is the total cost of the optimal path. The optimal path is obtained by backtracking the state transfer process. In the improved non-dominated sorting genetic algorithm, real number coding is used to encode the arc curvature parameters as chromosomes, including the arc starting point coordinates (10.5, 20.3), the arc end point coordinates (50.2, 80.1), and the arc control point coordinates (30.7, 40.6). The chromosomes are optimized through selection, crossover, and mutation operations. The binary tournament selection strategy is used to select two individuals from the population with a probability of 0.8, and the individual with higher fitness is added to the next generation population; the arithmetic crossover strategy is used to linearly combine a certain gene segment of the two individuals with a probability of 0.6 to generate a new individual; the Gaussian mutation strategy is used to slightly perturb a certain gene of the individual with a probability of 0.1 to introduce a new search area.

[0029] Step S103, according to the network constraints, set the objective function of maximizing the energy transmission efficiency, use the BP network algorithm to iteratively train the arcs of the energy recovery network, dynamically update the energy recovery network topology model parameters by calculating the loss and efficiency when the energy flows through the arc nodes, and obtain the optimized energy recovery arc.

[0030] Specifically, the actual operating constraints of the energy recovery network are comprehensively analyzed, including node layout, energy transmission distance, and energy loss rate. By establishing a multi-objective optimization model, multiple objectives such as energy transmission efficiency, network cost, and reliability are considered, and an optimization function with the main goal of maximizing energy transmission efficiency is constructed. As a quantitative indicator for subsequent network training and optimization, the energy recovery network is abstracted into a weighted directed graph model, and the Dijkstra shortest path algorithm is used to encode and represent network nodes and arcs. The network topology information is stored through the adjacency matrix. The time complexity of the algorithm is O(n^2), which is suitable for the shortest path calculation of dense graphs, providing a basis for subsequent network analysis and optimization. For the energy recovery network topology structure model, a machine learning algorithm based on BP neural network is constructed. A three-layer feedforward neural network is constructed, including an input layer, a hidden layer, and an output layer. The number of nodes in the hidden layer is set to twice that of the input layer. The Sigmad activation function and Xavier activation function are used. er weight initialization method, using momentum method and adaptive learning rate adjustment strategy, iteratively trains and optimizes network parameters through back propagation algorithm, learns the complex nonlinear relationship of energy transmission and conversion in the network, and in the BP network training process, according to the real-time monitoring data of energy flowing through each node on the network arc, the energy loss and transmission efficiency data are obtained at a frequency of sampling each node at intervals, which are used as sample data sets for network training. Each sample includes node ID, timestamp, energy loss rate, and transmission efficiency attributes. The sample data is cleaned and converted through data preprocessing and feature engineering, and the network parameters are dynamically updated using small batch stochastic gradient descent method. 32 samples are randomly selected as a batch each time, and the learning rate is set to 0.01, the momentum factor is 0.9, and the regularization coefficient is 0.001, by setting early stopping method and L2 regularization to avoid overfitting, continuous iterative optimization, so that the average absolute error of the network model on the training set is less than 1%, and the average absolute error on the validation set is less than 5%. By training and tuning the BP network model, a network model that can accurately predict and optimize the energy transmission efficiency is obtained. The model is used to exhaustively analyze all arc combinations in the energy recovery network. The energy transmission efficiency prediction value of each arc is calculated based on the forward propagation of the network model. At the same time, the energy loss, transmission distance, and node reliability of the arc are comprehensively considered to construct a weighted comprehensive evaluation index. All arc combinations are scored and sorted. Finally, the arc combination with a ranking greater than the preset threshold is selected as the candidate optimal arc. The candidate optimal arc is remapped to the original energy recovery network topology model. By comparing the candidate arc and the original network topology, the redundant arcs that need to be deleted and the potential arcs that need to be added are automatically identified. The network topology is adjusted and updated by the maximum spanning tree algorithm. Under the premise of ensuring network connectivity and coverage, the total network cost is minimized, and a more reasonable structure and higher efficiency energy recovery network topology model is obtained. .

[0031] In one embodiment, when constructing a multi-objective optimization model for maximizing energy transmission efficiency, a weighted summation method is used to convert multiple objectives into a single-objective problem. The weight of energy transmission efficiency is set to 0.6, the weight of network cost is set to 0.3, and the weight of reliability is set to 0.1. The overall optimization objective function is formed through linear combination. For the encoding of the network topology structure, an adjacency matrix representation method is used. For a network with n nodes, an n×n two-dimensional matrix A is used. If there is an arc connecting node i and node j, then A[i][j]=1, otherwise A[i][j]=0. Through Di The jkstra algorithm calculates the shortest path between any two nodes with a time complexity of O(n^2). In the construction of the BP neural network, the number of input layer nodes is the number of nodes in the energy network, the number of output layer nodes is 1, indicating the energy transmission efficiency, and the number of hidden layer nodes is set to twice the number of input layer nodes, that is, 2n nodes. The Sigmoid activation function is used, and the initial weights are randomly sampled from a uniform distribution of [-0.1, 0.1]. When training the BP network, each sample contains node ID, timestamp, energy loss rate, and transmission efficiency attributes. The energy loss rate and transmission efficiency are used as input features, and the corresponding true efficiency values ​​are used as labels. Every 10 minutes Sampling is performed once, 144 samples are collected from each node every day, and 144n samples are collected from the entire network every day. The small batch stochastic gradient descent method is used, 32 samples are randomly selected each time, the learning rate is set to 0.01, the momentum factor is set to 0.9, the regularization coefficient is set to 0.001, and the iteration is performed 10,000 times. The iteration is stopped when the mean absolute error on the training set is less than 1% and the mean absolute error on the validation set is less than 5%. For the screening of candidate optimal arcs, the comprehensive evaluation index E = 0.5 × efficiency + 0.3 × (1-loss rate) + 0.2 × (1 / distance) is set, the E value is calculated for all arc combinations, and the top 10% with the highest E value are selected as candidate arcs.

[0032] Step S104, construct an arc optimization mechanism, use a logistic regression algorithm to score the complexity of the optimized energy recovery arc, identify nodes with simple structure and efficiency greater than a preset threshold, and build an efficient topological node set; use the April association rule mining algorithm to analyze the causal relationship between nodes, guide the generation of a simplified network topology, calculate the structural differences and update the model to ensure that the energy recovery efficiency is maintained or improved while simplifying the network.

[0033] Specifically, an arc optimization mechanism based on a logistic regression algorithm is constructed. Through feature engineering and data preprocessing, the optimized energy recovery arc is converted into a high-dimensional feature vector, including arc length, energy transmission efficiency, and node quantity attributes, and the features are normalized and standardized. A logistic regression model is constructed, and the model parameters are trained and optimized using the maximum likelihood estimation method. The specific implementation process includes defining the likelihood function of the logistic regression model, setting model parameters, taking the logarithm of the likelihood function to obtain the logarithmic likelihood function, converting the maximum likelihood function into the maximum logarithmic likelihood function, calculating the gradient of the logarithmic likelihood function, and obtaining the gradient vector. The gradient descent method or Newton's method optimization algorithm is used to iteratively update the parameters along the gradient direction until convergence or the maximum number of iterations is reached, and a probability prediction model of the arc complexity is obtained. The trained logistic regression model is used to estimate the energy recovery arc after optimization. Carry out complexity scoring, divide the arcs into two categories: complex arcs and simple arcs according to the predicted probability value, set the complexity threshold, if the predicted probability is greater than the threshold, it is determined to be a complex arc, otherwise it is a simple arc, and control the granularity and accuracy of the simplified network by adjusting the threshold. For the simple arcs after scoring, extract their corresponding nodes as candidate efficient topological nodes, and use the SDNE graph embedding method based on the autoencoder to learn the low-dimensional vector representation of the node in an unsupervised manner to capture the structural similarity of the node. The loss function of SDNE includes reconstruction loss and regularization loss. The Adam optimization algorithm is used for training, and finally the low-dimensional embedding vector of the node is obtained. The K-Means clustering algorithm is used to cluster the nodes in the embedding space. Each cluster represents an efficient topological node set. The node structure in the cluster is simple and the energy transmission efficiency is high. On the basis of the efficient topological node set, the April association rule mining algorithm is used to analyze the causal relationship and frequent co-occurrence pattern between nodes, and the grid search or cross-validation method is used to select the optimal parameter combination;Select the parameter combination with moderate number of generated frequent item sets and association rules and high coverage. The mined association rules guide the optimization and simplification of network topology. The node association rules mined by the i algorithm are used to automatically generate a simplified network topology structure. Heuristic rules or constraint-based optimization methods are used for graphical processing and optimization. The edges are weighted according to the confidence of the association rules. The higher the confidence, the greater the weight of the edge. When the same node appears in multiple association rules, the one with the highest confidence is selected as the directed edge. The loops in the graph are removed by topological sorting or minimum spanning tree algorithm to ensure that the generated DAG structure is a DAG structure. For nodes with higher degrees, edges are pruned or merged to control the sparsity of the graph, and a simplified network topology with simple structure and high energy transmission efficiency is obtained. The difference between the simplified network topology and the original network topology is evaluated. The graph convolutional neural network (GCN) is used to extract features and learn representations of the two network topologies. The network structure of GCN consists of an input layer, a convolutional layer, and an output layer. A semi-supervised node classification task is used. The loss function is the cross entropy loss. The optimization algorithm is Adam. The model parameters are updated by backpropagation. After obtaining the embedded representation of the node, the two networks are calculated. The Euclidean distance between the embedded matrices is used as the structural difference. At the same time, the energy recovery efficiency of the simplified network is evaluated, and the efficiency improvement ratio before and after simplification is calculated. According to the network topology difference and the energy recovery efficiency improvement ratio, the model is dynamically updated. By setting the difference threshold and the efficiency improvement threshold, it is judged whether the current simplified network meets the optimization requirements. The threshold is set by combining historical data analysis and expert experience. The distribution of the difference and efficiency improvement ratio is obtained, and the quantile is selected as the threshold. Experts in the field are invited to evaluate and adjust the threshold. The network scale, energy transmission characteristics, and optimization target factors are comprehensively considered to give a recommended threshold range. In the actual optimization process, the threshold is dynamically adjusted according to the performance of the simplified network. If the structural difference of the simplified network is large but the efficiency improvement is not obvious, the difference threshold is appropriately increased or the efficiency improvement threshold is lowered. Otherwise, the difference threshold is appropriately lowered or the efficiency improvement threshold is increased. If the threshold requirements are met, the simplified network is used as the new baseline network, and the logistic regression model and Apr are updated. ior i association rules, enter the next round of iterative optimization; if not satisfied, restore the original network topology, adjust the simplification strategy and threshold, re-perform arc scoring and node mining, until the optimal simplified network is found. ;

[0034] In one embodiment, when constructing a logistic regression model, arc length, energy transmission efficiency, number of nodes, average node degree, and 10 features related to arc complexity are selected, continuous features are subjected to minimum-maximum normalization, discrete features are subjected to one-hot encoding, feature vectors are input into the Sigma function, model parameters are calculated by maximum likelihood estimation, the L-BFGS optimization algorithm is used to iterate 100 times, and the iteration is stopped when the change in the log-likelihood function is less than 0.001. When scoring the arc complexity, arcs with a predicted probability greater than 0.7 are classified as complex arcs, arcs with a predicted probability less than 0.3 are classified as simple arcs, and arcs between 0.3 and 0.7 are classified as moderately complex arcs. For clustering of candidate efficient topological nodes, the SDNE model is used, the node adjacency matrix is ​​used as input, and the nodes are embedded into 12 In the 8-dimensional vector space, the number of hidden layer nodes of the encoder and decoder is 512 and 256 respectively, the activation function is ReLU, the Adam optimizer is used, the learning rate is 0.01, and 500 epochs are trained to finally obtain the embedding vector of the node. Then the K-Means algorithm is used for clustering. The number of clusters is determined by the elbow method or the silhouette coefficient, and the value is generally between 5 and 20. When mining association rules, the transaction database is first processed, each arc is regarded as a transaction, and the nodes on the arc are regarded as items in the transaction, and then the Apr The ior i algorithm generates frequent item sets and association rules. The support threshold is set to 0.05 and the confidence threshold is set to 0.8. The number and quality of association rules are controlled by adjusting the thresholds. When generating a simplified network topology, the association rules are sorted from high to low according to the confidence, and the top 20% of the rules with the highest confidence are selected to construct directed edges. For nodes that appear in multiple rules, the rule with the highest confidence is selected. The dag_longest_path function in the NetworkX library is used to solve the longest path of the DAG and obtain the simplified network topology. When evaluating the differences in network structures, a 2-layer GCN model is used. The convolution kernel size of the first layer is 64, the convolution kernel size of the second layer is 32, the activation function is ReLU, the cross entropy loss function and the Adam optimizer are used, the learning rate is 0.01, and 200 epochs are trained. After obtaining the embedded representation of the node, the Froben i of the two network embedding matrices is calculated. The us norm is used as the structural difference. When setting the threshold, the historical data is statistically analyzed. For example, the difference and efficiency improvement ratio of 100 historical optimization cases are statistically analyzed. It is found that the median of the difference is 0.15 and the 75% percentile of the efficiency improvement ratio is 10%. The difference threshold is set to 0.15 and the efficiency improvement threshold is set to 10%. At the same time, three experts in the field are invited to evaluate the thresholds and the recommended range is 0.10.2 for the difference threshold and 8% to 12% for the efficiency improvement threshold. In actual optimization, if the difference of the simplified network exceeds 0.2, but the efficiency improvement is less than 8%, then the difference threshold is increased to 0.25, the efficiency improvement threshold is reduced to 5%, and the network is simplified and evaluated again.

[0035] Step S105, use the fitness function algorithm to monitor the network status and energy loss in real time to evaluate the energy transmission efficiency; if the monitoring results show that the transmission efficiency decreases or the structural complexity increases, the energy transmission efficiency target and structural complexity of the current network are updated to the fitness function parameters, and the arc optimization mechanism is re-executed to obtain the optimized arc path.

[0036] Specifically, by deploying intelligent sensors at key nodes of the energy recovery network, the energy loss data, transmission delay data, node load data network status parameters are collected in real time, and the collected data is transmitted to the central controller for summary analysis to form a network status feature vector containing multiple dimensions of energy consumption, delay, and load. As the input of the fitness function, an energy transmission efficiency evaluation model is constructed. Taking into account multiple performance indicators such as energy loss, transmission delay, and network throughput, the analytic hierarchy process (AHP) is used to determine the weight coefficient of each indicator. Domain experts compare the indicators pairwise, construct a judgment matrix, calculate the relative importance of the indicators, and perform a consistency test to finally obtain the weight of the energy transmission efficiency. , the weight of delay, the weight of throughput, and quantify each indicator into a comprehensive efficiency score by weighted summation, and set the threshold of the efficiency score. When the efficiency score calculated in real time is lower than the threshold, it is judged that the transmission efficiency has decreased, triggering the arc optimization mechanism. By defining the average node degree, average path length, and clustering coefficient topological structure parameters of the network, a mapping relationship between network complexity and topological parameters is established. For a directed graph, it is converted into an undirected graph, that is, the directed edges are replaced by undirected edges, and then the local clustering coefficient and the global clustering coefficient are calculated. The local clustering coefficient measures the degree of aggregation of a single node and is defined as the ratio of the actual number of edges between the neighbors of the node to the number of all possible edges. The global clustering coefficient measures the degree of aggregation of a single node. The degree of aggregation of the entire network is measured, which is defined as the average value of the local clustering coefficients of all nodes. When the global clustering coefficient exceeds the preset threshold, it is judged that the network structure complexity is high. When the topological structure of the network is monitored to change, the complexity index of the network is calculated in real time, and the threshold of the complexity index is set. When the complexity index exceeds the threshold, it is judged that the structural complexity increases, triggering the arc optimization mechanism. According to the judgment results of the decrease in transmission efficiency and the increase in structural complexity, the target parameters and constraints in the fitness function are dynamically adjusted. The energy transmission efficiency target and structural complexity parameters of the current network are used as the optimization target and constraint conditions of the fitness function. L1 regularization or L2 regularization is used to add structure to the fitness function. The penalty term of structural complexity and energy consumption is added, and the penalty term coefficient is adjusted according to the focus of network optimization. The fitness function is expressed as, f(x) = efficiency index-penalty term coefficient of structural complexity*complexity index-penalty term coefficient of energy consumption*energy consumption index. The goal is to maximize the fitness function value. By increasing the penalty term, the network optimization algorithm is guided to evolve in the direction of reducing energy consumption and simplifying the structure. The arc optimization mechanism is re-executed and implemented using the Python-based particle swarm optimization algorithm. Multiple particles are initialized, and each particle contains the three-dimensional coordinates of the starting point and the end point. The fitness function value is used as the fitness of the particle, the number of iterations is set, the upper limit of the particle speed is set, and the global optimal solution is selected as the arc path through the threshold method.

[0037] In one embodiment, various types of intelligent sensors such as temperature, humidity, and vibration are deployed on key nodes of the energy recovery network to monitor the network status in real time. The sampling frequency of the sensor is set to 100 Hz, and 1000 data samples are collected in each monitoring cycle. The data is preprocessed by the edge computing device to extract key features such as average temperature, maximum humidity, and energy loss rate to form a 20-dimensional network status feature vector. The indicator weights in the energy transmission efficiency evaluation model are quantitatively analyzed, a 5th-order judgment matrix is ​​constructed, and the 1-9 scale is used to compare the importance of indicators in pairs. The eigenvector corresponding to the maximum eigenvalue is calculated, and a consistency check is performed. If the consistency ratio is less than 0.1, the weights of efficiency, delay, and throughput are determined to be 0.5, 0.3, and 0.2 respectively through the hierarchical analysis method, and are normalized as weighted summation coefficients. When the comprehensive efficiency score is lower than 0.8, the arc optimization mechanism is triggered. When evaluating the network complexity, Newo The rkX toolkit converts a directed graph into an undirected graph, calculates the clustering coefficient of the node based on the adjacency matrix, and automatically calculates the clustering coefficient through a script. When the global clustering coefficient exceeds 0.6, it is judged to be high complexity. The complexity assessment service is deployed in a container to achieve real-time monitoring of complexity indicators. When optimizing the fitness function, the L1 regularization method is used, and the penalty coefficients of structural complexity and energy consumption are set to 0.15 and 0.25 respectively. The penalty coefficient is adaptively adjusted by the gradient descent method, and the learning rate is set to 0.01. When the fitness function value increase is lower than 0.05, the update of the penalty coefficient is stopped. For the arc path optimization problem, a Python-based particle swarm optimization algorithm is used to initialize 50 particles, each of which contains the three-dimensional coordinates of the starting point and the end point. The fitness function value is used as the fitness of the particle. It is iterated 100 times, and the particle speed upper limit is set to 15% of the coordinate range. The global optimal solution is selected as the arc path by the threshold method.

[0038] Step S106, using an adaptive weight allocation algorithm to monitor the motion state and energy demand of the underwater robot in real time, and dynamically adjust the energy allocation weight of each node in the energy recovery network; when it is monitored that the network transmission efficiency decreases or the structural complexity increases, the arc optimization mechanism is triggered to generate a new optimal transmission arc.

[0039] Specifically, an adaptive weight allocation algorithm is used for the motion state and energy demand of the underwater robot. By deploying multiple sensors on the key components of the robot, such as pressure sensors to measure underwater pressure to convert it into depth information, and inertial measurement units (IMUs) to measure the robot's three-axis angular velocity and acceleration to calculate the attitude angle and displacement, the robot's speed, acceleration, angular velocity, depth, attitude motion parameters, as well as battery voltage, current, and temperature energy parameters are collected in real time, and the collected data are transmitted to the central control unit for fusion analysis. A real-time monitoring model for the robot's motion state and energy demand is constructed. According to the robot's motion mode and task stage, the weight coefficient of each parameter in the model is dynamically adjusted to achieve an accurate assessment of the robot's state and demand. The real-time monitoring model for the underwater robot's motion state and energy demand is correlated with the node state of the energy recovery network. By deploying energy monitoring and communication modules on network nodes, the energy collection, storage, and consumption data of the nodes are regularly collected. Using ARI The MA (autoregressive integrated moving average) model is used to predict the trend of node energy data. The model parameters are estimated by the least squares method, and the model order with the smallest AIC (Akaike Information Criterion) is selected to make a rolling prediction of energy data in the future. For anomaly detection, iso lat The ionforest (isolation forest) algorithm constructs multiple isolated trees by randomly splitting data on attributes recursively. The average path length of abnormal points in the tree is short. The abnormal situation is judged according to the path length threshold, and a dynamic threshold is set to adapt to the changing trend of node energy data. When the energy state of the node changes significantly or deviates greatly from the predicted value, the adjustment mechanism of the energy allocation weight is triggered. An adaptive weight allocation algorithm based on reinforcement learning is constructed. The overall energy transmission efficiency of the energy recovery network and the energy balance of the node are optimized, and the energy allocation weight of each node is used as the decision variable. The state space includes the energy collection, storage, and consumption status of each node, as well as the robot's motion mode and task stage context information. The action space is the energy allocation weight of each node, and the value range is a continuous value between 0 and 1. The reward function comprehensively considers the energy transmission efficiency of the network and the energy balance of the node, and is constructed in the form of weighted sum. The weight is given by expert experience or automatically learned through hyperparameter optimization methods. The learning algorithm adopts the DDPG (deep deterministic policy gradient) algorithm, which combines DQN (deep Q network) and Actor-Cr Itic takes advantage of the interactive learning between the value network and the policy network to achieve policy optimization in the continuous action space, continuously tries and learns different weight allocation strategies, and updates the strategies according to the feedback efficiency and balance indicators, and finally converges to the optimal weight allocation strategy, achieving dynamic balance and adaptive scheduling of energy between network nodes, and building a global energy state monitoring model for the energy recovery network, comprehensively considering the motion state of the underwater robot, energy demand and energy allocation factors of the network nodes,The Graph Attention Network (GAT) model is used to adaptively assign weights to different nodes through the attention mechanism, aggregate the feature information of the nodes, and learn the topological structure and energy flow law of the network. The training process adopts a semi-supervised method, with the energy state of the node and the performance indicators of the network as labels. The network parameters are optimized through the cross entropy loss function and the back propagation algorithm. At the same time, L2 regularization or dropout is introduced to prevent overfitting. The early stopping mechanism is used to automatically determine the optimal number of training rounds based on the performance on the validation set, and output the key performance indicators of the network, such as transmission efficiency, energy utilization, and node failure rate. When the network transmission efficiency decreases or the node failure rate increases abnormally, the optimization mechanism is triggered in time. For the abnormal situation of decreased transmission efficiency or increased structural complexity, the arc optimization mechanism is automatically activated. The genetic algorithm is used for heuristic search optimization, and the network topology is encoded as a chromosome. Each node and edge corresponds to a gene. The real number encoding method is used, and the initial population is generated by random generation or heuristic construction method. The population size is set, the fitness function is constructed as a weighted combination of energy transmission efficiency and network complexity, a penalty term is introduced to penalize overly complex or inefficient topological structures, tournament selection, uniform crossover and Gaussian mutation are selected as genetic operators, the crossover probability is set, the mutation probability is set, and the termination condition is set as the maximum evolutionary generation or the convergence degree of the fitness value, such as the change in fitness value for multiple consecutive generations is less than a certain threshold, search and optimize in the topological structure space of the energy recovery network, evaluate the transmission efficiency and structural complexity of the alternative arcs through the fitness function, and select the transmission arc with the best comprehensive performance as the new energy. The transmission path is calculated, and the network topology is dynamically adjusted and updated according to the optimization results. In the process of arc optimization, the complexity and variability of the underwater environment and the heterogeneity of network nodes are fully considered, and an adaptive weight allocation mechanism is introduced. Different optimization strategies and weight allocation schemes are adopted for different types of nodes. According to the energy collection, storage, and transmission properties of the nodes, the nodes are clustered and analyzed to obtain different types of node clusters. For node clusters with surplus energy, their energy allocation weights are appropriately reduced to encourage them to transmit energy to other nodes; for node clusters with energy shortages, their energy allocation weights are appropriately increased to give priority to meeting their energy needs; for node clusters on critical transmission paths, higher transmission priorities are given to ensure the connectivity and reliability of the network; when allocating weights, the dynamic adjustment range and step size of the weights are set to avoid excessive or frequent changes in weights, which may cause network oscillations. For example, for nodes with strong energy collection capabilities but low transmission efficiency, their energy allocation weights are reduced and the priority of energy transmission is increased; for nodes with insufficient energy reserves but on critical transmission paths, their weights are dynamically increased to ensure the stability and continuity of the transmission paths, and a closed-loop feedback mechanism for arc optimization is constructed, and a network optimization controller is constructed.The input is the network status information (node ​​energy status, transmission efficiency) and performance indicators (energy utilization, average delay), and the output is the control strategy of weight allocation and arc optimization. The controller adopts the model-based predictive control (MPC) method. By collecting network status information in real time, updating the network model, predicting the network behavior in the future, and obtaining the optimal control strategy, the control strategy is applied to the actual network to form a closed-loop feedback. In the feedback process, an adaptive mechanism is introduced to dynamically adjust the controller parameters according to the changing trend of the network status, such as the prediction time domain, control time domain, and the weight of the optimization target, to achieve online learning and adaptive optimization of the control system, and combine the generation of the optimal transmission arc with the adjustment of the energy allocation weight to form a two-way interaction and dynamic balance of network optimization. By continuously monitoring the network status and robot needs, dynamically triggering arc optimization and weight adjustment, and continuously iterating and evolving, the adaptive optimization and intelligent regulation of the energy recovery network are achieved, and ultimately the goals of improving energy transmission efficiency, balancing node energy consumption, and ensuring the reliability of robot operation are achieved. ,

[0040] In one embodiment, a high-precision pressure sensor and an IMU are installed on key components of the underwater robot. The measurement range of the pressure sensor is 0-30MPa, the accuracy is 0.1%FS, and the sampling frequency is 10Hz; the measurement range of the IMU is ±2000° / s (angular velocity), ±16g (acceleration), the accuracy is 0.1° / s (angular velocity), 0.01g (acceleration), and the sampling frequency is 100Hz. The Kalman filter algorithm is used to filter the IMU. MU data is filtered and fused to estimate the robot's attitude angle and displacement. The attitude and displacement information is updated every 0.1s. The robot's energy monitoring module uses high-precision voltage and current sensors with a measurement range of 0-30V, 0-50A, an accuracy of 0.1% FS, and a sampling frequency of 1kHz. The energy data is smoothed by a sliding average filtering algorithm, and the average voltage, current, and power values ​​are calculated every 1s. The real-time monitoring model of the robot's motion state and energy demand uses the support vector machine (SVM) algorithm. The optimal kernel function (Gaussian kernel) and model parameters (penalty coefficient C and kernel function parameter γ) are selected through grid search and cross-validation. The model is trained using training set data (including historical robot state and energy data), and the model performance is evaluated on the test set with an accuracy of more than 95%. The energy monitoring module is deployed on the nodes of the energy recovery network, the charge integration method is used to measure the energy collection amount of the node, and the coulomb meter method is used to measure the energy consumption of the node. The energy data is reported every 10s, and the time series model of the node energy data is fitted by the least squares method to select the AR I with the smallest AI C MA (p, d, q) model, where p is the autoregressive order, d is the difference order, and q is the moving average order. The model parameters are solved by the maximum likelihood estimation method, and the rolling prediction method is used to predict the energy data for the next 1 minute. Node energy anomaly detection is based on the isolation forest algorithm. By randomly selecting feature subsets, the data is randomly divided recursively to construct multiple isolated trees. The characteristic of anomalies is that the path length in most trees is short. The anomalies are judged by setting the path length threshold (less than 0.5 times the average path length). The threshold is automatically updated every 1 minute. The state space of the reinforcement learning model includes the energy collection rate, storage level, and transmission power of each node. The action space is the energy allocation weight of the node, which ranges from [0, 1]. The reward function is constructed as a weighted sum of R = 0.6*energy transmission efficiency + 0.4*node energy balance, where energy transmission efficiency = actual transmission power / maximum transmission power, node energy balance = 1-standard deviation (node ​​residual energy) / average value (node ​​residual energy), weights are obtained through grid search and cross-validation, the Actor network and Critic network of the DDPG algorithm both use a 3-layer fully connected neural network, the number of hidden layer nodes is 64 and 32 respectively, the activation function is ReLU, the output layer activation functions are tanh and linear functions respectively, the optimizer is Adam, the learning rate is 0.001, the batch size is 32, the discount factor is 0.95, the target network soft update coefficient is 0.01, the global energy state monitoring model of the energy recovery network uses a 5-layer GAT model, the input layer is the node feature (energy state, node degree), the hidden layer contains 64, 32, and 16 attention heads respectively, the attention mechanism adopts the scaled dot product form, and the output layer adopts a fully connected layer and si gmoi d activation function, predicting the energy state and key performance indicators of the node, the loss function is the weighted sum of cross entropy and mean square error, with weights of 0.7 and 0.3 respectively, the optimizer is Adam, the learning rate is 0.005, the regularization coefficient is 0.001, the dropout rate is 0.5, the tolerance number of early stopping method is 10 times, the arc optimization adopts genetic algorithm, the individual encoding is a real vector, indicating the coordinates of the start and end points of the arc, the initial population is randomly generated, the population size is 50, the fitness function is f = 0.8 * energy transmission efficiency - 0.2 * (network complexity + 1) ^ 2, the top 10% of individuals in fitness are selected by the tournament selection operator, the probability of uniform crossover and Gaussian mutation is The rates are 0.8 and 0.1 respectively. The iteration is terminated when the optimal fitness changes less than 1% after 100 generations or 20 consecutive generations. The network optimization controller is based on model predictive control (MPC). The objective function is to maximize the weighted sum of energy utilization and delay, with weights of 0.6 and 0.4 respectively. The constraints are node energy balance and link capacity limit. The prediction time domain is 10 steps and the control time domain is 5 steps. The sequential quadratic programming (SQP) algorithm is used to solve the optimal control sequence, and the control amount of the first 3 steps is applied to the actual system. Rolling optimization is performed. According to the network state feedback, the prediction time domain, control time domain and objective function weight of MPC are adaptively adjusted to achieve adaptive and real-time closed-loop optimization control. .

[0041] So far, the technical solutions of the present invention have been described in conjunction with the preferred embodiments shown in the accompanying drawings. However, it is easy for those skilled in the art to understand that the protection scope of the present invention is obviously not limited to these specific embodiments. Without departing from the principle of the present invention, those skilled in the art can make equivalent changes or substitutions to the relevant technical features, and the technical solutions after these changes or substitutions will fall within the protection scope of the present invention.

[0042] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. For those skilled in the art, the present invention may have various modifications and variations. Any modification, equivalent replacement, improvement, etc. made within the spirit and rules of the present invention shall be included in the protection scope of the present invention.

Claims

1. A battery energy recycling method for an underwater robot, characterized in that: Step S101, collecting equipment status and environmental data from the underwater robot, using a particle filter algorithm to filter the data, collecting battery material properties and energy transmission efficiency parameters, identifying the property parameters through a k-means clustering algorithm, and constructing constraints for the energy recovery network; Step S102, constructing an energy recovery network topology model according to the energy loss information and node layout information of the nodes in the energy recovery network; applying the Eulerian path algorithm to obtain the optimal path in the energy recovery process, and using the non-dominated sorting genetic algorithm to calculate the arc curvature corresponding to each node in the network; Step S103, according to the network constraints, set the objective function of maximizing the energy transmission efficiency, use the BP network algorithm to iteratively train the arc of the energy recovery network, dynamically update the energy recovery network topology model parameters by calculating the loss and efficiency when the energy flows through the arc nodes, and obtain the optimized energy recovery arc; Step S104, constructing an arc optimization mechanism, scoring the complexity of the optimized energy recovery arc through a logistic regression algorithm, and constructing a topological node set; using the Apriori association rule mining algorithm to analyze the causal relationship between nodes, calculate the structural difference and update the model; Step S105, using the fitness function algorithm to monitor the network status and energy loss in real time and evaluate the energy transmission efficiency; if the monitoring result shows that the transmission efficiency decreases or the structural complexity increases, the energy transmission efficiency target and structural complexity of the current network are updated to the fitness function parameters, and the arc optimization mechanism is re-executed to obtain the optimized arc path; Step S106, using an adaptive weight allocation algorithm to monitor the motion state and energy demand of the underwater robot in real time, and dynamically adjust the energy allocation weight of each node in the energy recovery network; When it is monitored that the network transmission efficiency decreases or the structural complexity increases, the arc optimization mechanism is triggered to generate a new optimal transmission arc.

2. The method for recycling battery energy for an underwater robot according to claim 1, characterized in that: The step S101 uses the MapReduce algorithm in the Hadoop framework to clean and annotate the data. When constructing a digital model of the underwater environment depth information, a particle filter algorithm is used to initialize 100 particles, calculate the real-time depth information of the underwater environment, and use the Kalman filter algorithm to smooth the depth information. The filter parameter is set to 0.8, and the k-means clustering algorithm is used to classify the attribute parameters. The number of clusters k is set to 5, the number of iterations is 100, and the distance from each sample to the cluster center is calculated to divide the sample into the nearest cluster.

3. The method for recycling battery energy for an underwater robot according to claim 1, characterized in that: When the Kruskal algorithm is used to construct the energy recovery network topology model in step S102, the network nodes are first abstracted as vertices of the graph, and the connection relationships between the nodes are abstracted as edges of the graph. The weights of the edges are determined according to energy loss and distance factors, including an energy loss weight of 0.6 and a distance weight of 0.

4. Then, the edges are selected in order from small to large in weight. If adding the edge will not form a loop, it is added to the minimum spanning tree until all nodes are connected. The minimum spanning tree finally obtained is the optimized topology structure of the energy recovery network.

4. The battery energy recycling method for an underwater robot according to claim 1, characterized in that: The step S103 adopts the weighted summation method to transform multiple objectives into a single objective problem; the adjacency matrix representation method is adopted, for a network with n nodes, an n×n two-dimensional matrix A is used, if there is an arc connecting node i and node j, then A[i][j]=1, otherwise A[i][j]=0, and the shortest path between any two nodes is calculated by the Dijkstra algorithm.

5. The battery energy recycling method for an underwater robot according to claim 1, characterized in that: The S104 step performs minimum-maximum normalization processing on the continuous features, performs one-hot encoding on the discrete features, inputs the feature vector into the Sigmoid function, calculates the model parameters through maximum likelihood estimation, uses the L-BFGS optimization algorithm to iterate 100 times, stops the iteration when the change of the log-likelihood function is less than 0.001, and when scoring the arc complexity, the arcs with a predicted probability greater than 0.7 are classified as complex arcs, those less than 0.3 are classified as simple arcs, and those between 0.3 and 0.7 are classified as medium-complex arcs. The SDNE model is used, and the adjacency matrix of the node is used as input. The node is embedded in a 128-dimensional vector space through a 3-layer autoencoder structure. The number of hidden layer nodes of the encoder and decoder are 512 and 256 respectively, the activation function is ReLU, the Adam optimizer is used, the learning rate is 0.01, and 500 epochs are trained to finally obtain the embedded vector of the node.

6. The method for recycling battery energy for an underwater robot according to claim 1, characterized in that: The sampling frequency of the sensor in step S105 is set to 100 Hz, and 1000 data samples are collected in each monitoring cycle. The data is preprocessed by the edge computing device to extract key features, including average temperature, maximum humidity, and energy loss rate, to form a 20-dimensional network state feature vector, and the indicator weights in the energy transmission efficiency evaluation model are quantitatively analyzed. A 5th-order judgment matrix is ​​constructed, and the importance of indicators is compared pairwise using a 1-9 scale. The eigenvector corresponding to the maximum eigenvalue is calculated, and a consistency check is performed. If the consistency ratio is less than 0.1, the weights of efficiency, delay, and throughput are determined to be 0.5, 0.3, and 0.2 respectively through the hierarchical analysis method, and are normalized as coefficients for weighted summation. When the comprehensive efficiency score is lower than 0.8, the arc optimization mechanism is triggered. When evaluating the network complexity, the directed graph is converted into an undirected graph through the NetworkX toolkit, and the clustering coefficient of the node is calculated based on the adjacency matrix. When the global clustering coefficient exceeds 0.6, it is judged to be high complexity.

7. The battery energy recycling method for an underwater robot according to claim 1, characterized in that: In the step S106, a high-precision pressure sensor and an IMU are installed. The measurement range of the pressure sensor is 0-30MPa, the accuracy is 0.1%FS, and the sampling frequency is 10Hz. The measurement range of the IMU is ±2000° / s, indicating angular velocity, ±16g, indicating acceleration, the accuracy is 0.1° / s, indicating angular velocity, 0.01g, indicating acceleration, and the sampling frequency is 100Hz. The IMU data is filtered and fused by the Kalman filter algorithm to estimate the attitude angle and displacement of the robot, and the attitude and displacement information is updated every 0.1s. The energy monitoring module of the robot adopts a high-precision voltage and current sensor with a measurement range of 0-30V, 0-50A, an accuracy of 0.1%FS, and a sampling frequency of 1kHz. The energy data is smoothed by a sliding average filter algorithm, and the average voltage, current and power values ​​are calculated every 1s.

Images