A method for failure prediction and health management in a humanoid robot training field

Abstract

This invention provides a method for fault prediction and health management in a humanoid robot training field, relating to the field of robot health management technology. The method includes: collecting and preprocessing robot health data, controller logs, and interaction detection data between robots; conducting an initial health assessment; constructing a dynamic interactive health graph and inferring health decay trends using a time-series graph neural network; constructing and correcting a health-fault causal graph by combining historical fault data and historical health status data; adjusting alarm levels and verifying abnormal correlations; constructing and updating a multimodal health embedding model; obtaining the root causes of faults and health impact chains, and generating maintenance suggestions, health maintenance plans, and risk assessment reports. This invention captures the correlation characteristics of multi-robot groups through a dynamic interactive health graph, integrates a time-series graph neural network with a corrected health-fault causal graph, significantly improving the accuracy and timeliness of fault probability prediction, and achieving early warning of faults.

Description

Technical Field

[0001] This invention relates to the field of robot health management technology, and in particular to a method for fault prediction and health management in a humanoid robot training field. Background Technology

[0002] Humanoid robots, as typical complex mechatronic systems, aim to mimic human appearance and behavior. They integrate technologies from multiple fields, including mechanical design, dynamics control, and sensory interaction. Their development can be traced back to the 1960s, from Waseda University's WABOT-1 achieving initial bipedal walking, to Honda's ASIMO completing dynamic balance and basic interaction, to Boston Dynamics' Atlas achieving complex terrain maneuvers, and Tesla's Optimus leveraging autonomous driving technology to advance industrialization. These advancements have gradually formed technical characteristics such as high-degree-of-freedom coupling, adaptation to unstructured environments, and multi-sensor integration. With the deepening of global industrial deployment, humanoid robots have gradually moved from the laboratory to practical applications such as industrial production and rescue services, leading to continuously increasing technical requirements for operational stability and continuous operation capabilities.

[0003] In the field of humanoid robots, the integrated application of fault prediction and health management technologies is based on two major technological foundations: First, the maturity of dynamic simulation technology. Real-time physics engines such as MuJoCo and Gazebo can realize virtual modeling of complex scenarios such as multi-rigid-body coupling and contact impact, providing a virtual testing environment for fault simulation and diagnostic model verification. Second, the improvement of multi-physics data fusion capabilities. By collecting operational data through various types of devices such as encoders and current sensors with high-frequency sampling above 1kHz, combined with efficient algorithms such as recursive Newton-Euler, the real-time fault diagnosis needs can be supported.

[0004] Existing fault prediction methods in humanoid robot training environments generally focus on assessing the health status of individual robots, neglecting the interactive and interconnected characteristics in multi-robot collaborative training. This makes it impossible to accurately characterize the impact of group health transmission, resulting in one-sided fault probability predictions, insufficient timeliness and accuracy, and difficulty in achieving early warning. Summary of the Invention

[0005] This invention provides a method for fault prediction and health management in a humanoid robot training field, which addresses the shortcomings of existing technologies that generally focus on assessing the health status of individual robots while neglecting the interactive and interconnected characteristics in multi-robot collaborative training.

[0006] On one hand, the present invention provides a method for fault prediction and health management in a humanoid robot training field, comprising:

[0007] S1: Collect robot health data, controller logs, and robot interaction detection data, and preprocess them to obtain robot health feature data and interaction health data;

[0008] S2: Based on ontological health feature data, interactive health data, and interactive requirements, conduct an initial health assessment to obtain a multi-robot diagnostic set and initial health assessment results;

[0009] S3: Based on the multi-robot diagnostic set and initial health assessment results, construct a dynamic interactive health graph, infer the health decline trend through a time-series graph neural network; combine historical fault data and historical health status data to construct and correct a health-fault causal graph, and obtain the corrected fault probability.

[0010] S4: Adjust the alarm level based on the corrected fault probability and health degradation trend, verify the abnormal correlation, and obtain the final alarm level, evidence data and health risk level as needed.

[0011] S5: Construct a multimodal health embedding model and update the multimodal health embedding model using the final alarm level, evidence data, and health risk level;

[0012] S6: Use interpretable algorithms to obtain the root cause of the fault and the health impact chain, and generate maintenance suggestions, health maintenance plans and risk assessment reports.

[0013] According to the fault prediction and health management method for a humanoid robot training field provided by the present invention, step S2, the step of conducting an initial health assessment, includes:

[0014] S21: Construct basic health assessment rules based on ontological health feature data and interactive health data;

[0015] S22: Compare and match the ontological health feature data and interactive health data feature by feature with the basic health assessment rules, mark abnormal features that exceed the normal value range, and record the degree of deviation and the time of occurrence of each abnormal feature.

[0016] S23: Input the labeled abnormal features into a locally deployed lightweight deep learning edge model. The model analyzes the correlation between abnormal features, eliminates mislabeling caused by isolated noise points, outputs the health status level, and generates a preliminary health assessment result document.

[0017] According to the fault prediction and health management method for a humanoid robot training field provided by the present invention, step S21, the step of constructing basic health assessment rules, includes:

[0018] S211: Define the dimensions of health assessment;

[0019] S212: Based on the standard parameters of the humanoid robot at the factory, the rated operating parameters of the training ground, and historical fault-free data, the normal value range of each dimension of the feature is determined through statistical analysis to form basic threshold rules;

[0020] S213: Based on the interaction requirements, set feature association constraint rules and clarify the linkage, matching and response matching thresholds of features between the same or different humanoid robots;

[0021] S214: Set the environment adaptation coefficient, and correct the basic threshold rules according to the training environment parameters to obtain the threshold rules after environment adaptation.

[0022] S215: Based on the threshold rules and feature association constraint rules after environmental adaptation, an initial basic health assessment rule is formed; through verification and optimization using historical failure cases, until the preset performance requirements are met, the rule construction is completed.

[0023] According to the fault prediction and health management method for a humanoid robot training field provided by the present invention, step S3, the step of constructing a dynamic interactive health map includes:

[0024] S31: Set the robot as a node, and set the interaction strength, frequency and health correlation as edge weights;

[0025] S32: Based on the multi-robot diagnostic set and initial health assessment results, update the node attributes and edge weights;

[0026] S33: Use graph construction algorithms to generate dynamic interactive health graphs.

[0027] According to the fault prediction and health management method for a humanoid robot training field provided by the present invention, step S3, the step of obtaining the corrected fault probability includes:

[0028] S34: Input the features of the dynamic interactive health graph into the time-series graph neural network and output the initial fault probability;

[0029] S35: Construct and correct the health-failure cause-effect graph by combining historical fault data and historical health status data;

[0030] S36: The causal relationship strength of the health-failure cause-effect graph is incorporated as a weight into the initial failure probability calculation to obtain the corrected failure probability.

[0031] According to the fault prediction and health management method for a humanoid robot training field provided by the present invention, step S34, the step of outputting the initial fault probability includes:

[0032] S341: Perform dimension alignment and normalization on the node and edge features of the dynamic interactive health graph to obtain standardized graph features;

[0033] S342: Divide the standardized graph features into continuous feature sequences according to the time series, input them into the temporal convolutional layer of the temporal graph neural network, and extract the temporal dependencies of the feature sequences;

[0034] S343: Input the output of the temporal convolutional layer into the graph convolutional layer to capture the spatial correlation features between nodes and obtain a feature vector that integrates spatiotemporal information;

[0035] S344: Input the feature vector that integrates spatiotemporal information into the fully connected layer, calculate the initial failure probability of each robot through the activation function, and output the initial failure probability.

[0036] According to the fault prediction and health management method in a humanoid robot training field provided by the present invention, step S35, the step of correcting the health-fault cause-effect graph includes:

[0037] S351: Based on historical fault data and historical health status data, an initial health-fault causal graph is constructed using a Bayesian network, where nodes represent health features and fault types, edges represent the causal relationships between the two, and edge weights represent the initial causal relationship strength.

[0038] S352: Based on the failure case data and the corresponding health status evolution data, calculate the confidence level of each causal relationship and update the initial causal relationship strength; if there is a new relationship between health features and failure types, add it to the health-failure causal graph to obtain the updated health-failure causal graph.

[0039] S353: Verify the updated health-failure cause-effect graph, remove unreasonable causal relationships, adjust the strength of abnormal causal relationships, and obtain the corrected health-failure cause-effect graph.

[0040] S354: Calculate the goodness-of-fit index of the corrected health-failure cause-effect graph. If the goodness-of-fit index reaches the preset threshold, the correction is completed; if not, repeat steps S352-S353 until the goodness-of-fit index meets the requirements.

[0041] According to the fault prediction and health management method for a humanoid robot training field provided by the present invention, step S4, the steps of adjusting the alarm level and verifying the correlation of anomalies, include:

[0042] S41: Preset three-level alarm thresholds, and obtain the initial alarm level by matching and correcting the fault probability;

[0043] S42: Calculate the collaborative confidence level by combining the number of abnormal associations in the multi-robot diagnostic set and the rate of change of the health decay trend;

[0044] S43: Adjust the initial alarm level based on the collaborative confidence level, and output the adjusted alarm level and the corresponding range of abnormal robots that need to be checked;

[0045] S44: Based on the scope of abnormal robots that need to be checked, extract the voiceprint-vibration features of the corresponding abnormal robots in the multi-robot diagnostic set, compare them with the preset normal feature library, and obtain the voiceprint-vibration feature matching degree and the corresponding abnormal feature information.

[0046] S45: Based on the abnormal feature information corresponding to the voiceprint-vibration feature matching degree, collect visual evidence of the abnormal robot including its running posture and trajectory;

[0047] S46: Analyze the matching degree of voiceprint-vibration features and the feature synchronicity and correlation with visual evidence to verify the effectiveness of abnormal correlation.

[0048] According to the fault prediction and health management method for humanoid robot training ground provided by the present invention, step S5, the step of constructing a multimodal health embedding model includes:

[0049] S51: Construct a multimodal fusion network, input ontology health feature data, interaction health data and simulation parameters as training samples, and obtain a formatted training dataset;

[0050] S52: Using health status level as a label, train the network on the formatted training dataset using the cross-entropy loss function to obtain the network parameters to be optimized;

[0051] S53: Iteratively optimize the network parameters to obtain a multimodal health embedding model that outputs a unified dimension of health feature embedding vectors.

[0052] According to the fault prediction and health management method for a humanoid robot training field provided by the present invention, step S6, which uses an interpretable algorithm to obtain the root cause of the fault and the health impact chain, includes:

[0053] S61: Use an interpretable algorithm to analyze the health-failure cause-effect graph and filter out key nodes with high contribution to failure and node association information;

[0054] S62: By combining the final alarm level, evidence data, key nodes and node association information, trace the causal relationship path between each key node;

[0055] S63: Sort the causal paths according to their degree of influence to form the root cause chain of the failure, deduce the scope of health impact and transmission path, and obtain the health impact chain.

[0056] This invention provides a fault prediction and health management method for humanoid robot training fields. It captures the correlation characteristics of multi-robot groups through a dynamic interactive health graph, integrates a temporal graph neural network with a modified health-fault causal graph, significantly improving the accuracy and timeliness of fault probability prediction and enabling early warning of faults. It verifies abnormal correlations using multi-source data from voiceprint, vibration, and vision, ensuring the reliability of alarm level adjustments and effectively reducing the risk of false alarms and missed alarms. Furthermore, it strengthens the health status representation capability through the construction and updating of a multimodal health embedding model, providing more robust model support for assessment and prediction. Attached Figure Description

[0057] To more clearly illustrate the technical solutions in this invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of this invention. For those skilled in the art, other drawings can be obtained from these drawings without creative effort.

[0058] Figure 1 This is a flowchart of a fault prediction and health management method in a humanoid robot training field provided by an embodiment of the present invention;

[0059] Figure 2 This is a flowchart of the output initial fault probability in an embodiment of the present invention;

[0060] Figure 3 This is a flowchart of adjusting alarm levels and verifying abnormal correlations in an embodiment of the present invention. Detailed Implementation

[0061] To make the objectives, technical solutions, and advantages of this invention clearer, the technical solutions of this invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of this invention. All other embodiments obtained by those skilled in the art based on the embodiments of this invention without creative effort are within the scope of protection of this invention.

[0062] Example:

[0063] The following is combined with Figures 1-3 This invention describes a method for fault prediction and health management in a humanoid robot training field.

[0064] like Figure 1 As shown in the figure, an embodiment of the present invention provides a method for fault prediction and health management in a humanoid robot training field, comprising:

[0065] S1: Collect robot health data, controller logs, and interaction detection data between robots. Preprocess these to obtain robot health feature data and interaction health data. Robot health data is fundamental data reflecting the operating status of the humanoid robot's hardware and core components, and is the core basis for assessing the robot's individual health level. The main acquisition method is real-time collection via temperature sensors (joint motors, battery pack), pressure sensors (foot soles, joints), displacement sensors (joint range of motion), and current sensors (drive circuits) mounted on the robot. The sampling frequency is set to 10Hz, and the data is transmitted to the edge computing node via the robot's local bus. Controller logs are automatically recorded by the robot's main control unit (MCU), including instruction issuance time, execution results, error codes, CPU utilization, memory usage, etc., and are exported in batches by minute to local storage. This includes text and numerical data such as operation instructions, execution status, and error codes automatically recorded during the robot control system's operation, allowing for traceability of the control link's operational trajectory. Interaction detection data is collected via LiDAR, high-definition cameras, millimeter-wave radar, and the robot's built-in UWB positioning module deployed in the training area. The LiDAR and camera sampling frequency is 5Hz, and the UWB positioning data sampling frequency is 20Hz. The data is aggregated by the training area edge gateway and then transmitted to the analysis node. Data such as relative position, motion coordination, and information interaction latency obtained between robots through sensors and communication modules reflect the stability of group interactions. Preprocessing steps include:

[0066] Outliers caused by sensor malfunctions, missing values ​​generated during data transmission, and duplicate data are removed. For missing values, nearest neighbor interpolation is used to fill in the gaps; for outliers, the 3σ criterion is used to identify and remove them, i.e., if a data point x satisfies |x-μ|>3σ, where μ is the data mean and σ is the standard deviation, it is determined to be an outlier.

[0067] To map raw data of different dimensions to a unified interval and avoid interference from differences in magnitude to subsequent analysis, a min-max normalization method is used, expressed by the formula:

[0068]

[0069] Where, x norm The normalized feature values ​​range from [0,1]; x represents the original feature data; x min The minimum value of this feature; x max This is the maximum value of this feature.

[0070] Core features (such as average motor temperature, battery voltage fluctuation, joint motion accuracy, etc.) are extracted from the standardized body health data to form body health feature data; interactive collaboration features are extracted from the interactive detection data to form interactive health data.

[0071] S2: Based on ontology health feature data, interaction health data, and interaction requirements, a preliminary health assessment is performed, resulting in a multi-robot diagnostic set and preliminary health assessment results. The multi-robot diagnostic set includes records of abnormal features for each robot. The preliminary health assessment results include health status level and descriptions of abnormal features.

[0072] Step S2, the steps for conducting the initial health assessment include:

[0073] S21: Construct basic health assessment rules based on ontological health feature data and interactive health data. These basic health assessment rules are a set of rules constructed based on the robot's operational standards, environmental adaptability, and interactive collaboration requirements to determine whether health features are normal and whether the relationships between features are reasonable.

[0074] The steps to develop basic health assessment rules include:

[0075] S211: Defining Health Assessment Dimensions. The dimensional division criteria are based on the humanoid robot's operating mechanism and training ground collaboration requirements, dividing it into 3 primary dimensions and 8 secondary dimensions. The primary dimensions include the robot's performance dimension, interaction and collaboration dimension, and environmental adaptation dimension. The three primary dimensions are further subdivided into secondary dimensions, as detailed below:

[0076] The body performance dimension includes three sub-dimensions: motor operating status, battery performance, and sensor accuracy.

[0077] The interactive collaboration dimension includes three sub-dimensions: relative positioning accuracy, action coordination synchronization, and information interaction stability.

[0078] The environmental adaptability dimension includes two sub-dimensions: temperature and humidity adaptability and site flatness adaptability.

[0079] Each secondary dimension corresponds to several specific health characteristics, such as motor operating status corresponding to characteristics such as motor temperature, speed fluctuation, and current stability.

[0080] S212: Based on the humanoid robot's factory standard parameters, training range rated operating parameters, and historical fault-free data, statistical analysis is used to determine the normal value ranges for each dimension of characteristics, forming basic threshold rules. Factory standard parameters are obtained from the robot manufacturer's technical manual and include the rated operating parameters for each component. The training range rated operating parameters are determined based on the training range's design specifications and training task requirements. Historical fault-free data is extracted from the training range's historical operating database, filtering robot operation data with no fault records within the past 12 months, with a sample size of no less than 1000 sets. Statistical analysis uses a normal distribution statistical model to determine the normal value ranges for each characteristic. The normal value ranges for each dimension of characteristics are compiled into a basic threshold rule table, including fields such as characteristic name, corresponding dimension, normal range, and reference standard.

[0081] S213: Based on interaction requirements, set feature association constraint rules to clarify the linkage, matching, and response matching thresholds of features between the same or different humanoid robots. Feature association constraint rules are used to regulate the linkage relationships, matching accuracy, and response time between different features of the same robot and between similar features of different robots, ensuring coordinated operation of individual robots and collaborative interaction among groups.

[0082] The same robot feature linkage constraint sets linkage thresholds based on associated features within the ontology performance dimension. Different robot feature matching constraints set matching thresholds based on features within the interaction and collaboration dimension. Response matching constraints set response delay thresholds based on information interaction features. Finally, the above constraint relationships and corresponding thresholds are compiled into a feature association constraint rule table.

[0083] S214: Set the environment adaptation coefficient and correct the basic threshold rules based on the training environment parameters to obtain the environmentally adapted threshold rules. The environment adaptation coefficient is used to quantify the influence of training environment parameters (such as temperature, humidity, site flatness, light intensity, etc.) on the normal value range of robot health characteristics, realizing the environmental adaptive adjustment of the basic threshold rules. It is mainly achieved by acquiring environmental parameters in real time through environmental sensors deployed in the training field, with a sampling frequency of 1Hz, and transmitting the data to edge computing nodes for processing. The calculation method of the environment adaptation coefficient is expressed as follows:

[0084]

[0085]

[0086] Where ω is the environmental adaptation coefficient, ω T ω is the temperature adaptor coefficient. H ω is the humidity aptamer coefficient. S ω is the flatness fitter coefficient. L w is the aptamer coefficient for illumination. T w H w S w L Let w be the weight of each factor, satisfying w T +w H +w S +w L =1, determined using the Analytic Hierarchy Process (AHP), with values ​​of 0.4, 0.2, 0.3, and 0.1 respectively. T is the real-time monitored temperature, T0 is the lower limit of the optimal temperature, T1 is the upper limit of the optimal temperature, and k T The temperature influence coefficient is obtained through correlation analysis between historical fault data and temperature. The calculation logic for other sub-coefficients is similar, with corresponding formulas set according to the optimal range and degree of influence of each environmental factor.

[0087] The threshold after environment adaptation is expressed as:

[0088]

[0089] Among them, T norm_adj This represents the normal value range after environmental adaptation; T norm_high T norm_low These are the upper and lower limits of the normal range in the basic threshold rule.

[0090] S215: Based on the environment-adapted threshold rules and feature association constraint rules, integrate the environment-adapted threshold rules and feature association constraint rules to form initial basic health assessment rules. Optimize through historical failure cases until preset performance requirements are met, completing rule construction. Specifically, extract 500 typical failure cases from the training ground's historical failure database, input the health feature data before the failure into the initial rules, determine whether the rules can accurately identify abnormal features, and statistically analyze the rule's precision (P), recall (R), and F1 score.

[0091] S22: Compare and match the ontological health feature data and interactive health data feature by feature with the basic health assessment rules, marking abnormal features that exceed the normal value range. If a feature does not meet the rule requirements, it is marked as an abnormal feature, and the abnormality type is recorded. The degree of deviation of the abnormal features and the occurrence time are recorded using relative deviation metric. The calculation method for relative deviation is as follows:

[0092]

[0093] Where δ is the relative deviation; y is the actual value of the abnormal feature; y avg This is the mean value of the normal range for this characteristic.

[0094] Feature comparison and matching involves comparing the preprocessed ontology health feature data and interaction health data feature by feature with the basic health assessment rules, including the comparison of single features with thresholds after environment adaptation, and the comparison of associated features with feature association constraint rules.

[0095] S23: Input the labeled anomaly features into a locally deployed lightweight deep learning edge model. The model analyzes the correlation between anomaly features, eliminates mislabeling caused by isolated noise points, outputs the health status level, and generates a preliminary health assessment document. Specific steps include:

[0096] It adopts a lightweight structure of "convolutional layer + LSTM layer + fully connected layer". The input is an abnormal feature sequence (dimension is 1×m, where m is the number of abnormal features) and the output is a feature correlation score, ranging from 0 to 1. A score ≥0.6 is judged as an associated anomaly, and <0.6 is judged as an isolated noise point.

[0097] The model was trained using the cross-entropy loss function, with Adam as the optimizer, a learning rate of 0.001, and 50 training iterations. After training, the model was deployed to an edge computing node.

[0098] The labeled anomalous features are input into the model, and the model outputs the correlation score of each anomalous feature, removing isolated noise points with a score <0.6.

[0099] Based on the number of remaining abnormal features, the degree of deviation, and the scope of correlation, the health status is divided into 5 levels. The specific classification of the levels is shown in Table 1.

[0100] Table 1:

[0101]

[0102] Organize information such as health status level, list of valid abnormal features, and abnormality type to generate a preliminary health assessment result document in JSON format.

[0103] S3: Based on the multi-robot diagnostic set and initial health assessment results, a dynamic interactive health graph is constructed, and the health degradation trend is inferred through a time-series graph neural network. A health-failure causal graph is constructed and corrected by combining historical fault data and historical health status data to obtain the corrected fault probability. The dynamic interactive health graph is a time-series dynamic graph structure with robots as nodes and the interaction relationships between robots as edges. The edge weights quantify the interaction intensity, frequency, and health correlation, and can reflect the health correlation status in multi-robot collaborative training in real time.

[0104] Step S3, the steps for constructing the dynamic interactive health graph include:

[0105] S31: Set the robots as nodes, and the interaction strength, frequency, and health correlation as edge weights. Each robot is a node, and its attributes include the robot's health status level, core abnormal characteristics, and basic hardware parameters (such as model and service life), represented as follows:

[0106]

[0107] Among them, V i For node attributes, L i δ represents the health status level of robot i. ij Let P be the deviation of robot i from the j-th anomalous feature. ik Let be the k-th hardware parameter of robot i.

[0108] The edge weight is calculated as follows:

[0109]

[0110] Among them, W ijW represents the edge weight. int Interaction strength, W freq Interaction frequency, W health For health-related degrees.

[0111] Interaction intensity W int Based on the relative position deviation d between robots i and j ij The calculation is expressed as:

[0112]

[0113] Where, d max The maximum allowable relative distance for the training task is set to 5m, with a value range of [0,1]. ij The smaller the value, the stronger the interaction.

[0114] The interaction frequency is the number of interactions f between robots i and j per unit time. ij Normalization yields the following, expressed as:

[0115]

[0116] Among them, f max The maximum interaction frequency is set to 30 times / minute, with a value range of [0,1].

[0117] Health correlation is calculated based on the similarity of the health status levels of robots i and j, and is expressed as:

[0118]

[0119] Among them, L i L j The health status level is 1-5, with a value range of [0,1]. The closer the levels are, the higher the health correlation.

[0120] S32: Based on the multi-robot diagnostic set and initial health assessment results, update the node attributes and edge weights. Specifically, update the node attributes and edge weights every 5 minutes to ensure that the graph structure reflects the latest health and interaction status.

[0121] S33: A graph construction algorithm is used to generate a dynamic interactive health graph. Specifically, a dynamic graph construction algorithm based on the adjacency matrix is ​​used, including constructing an initial adjacency matrix; updating the adjacency matrix according to the updated edge weights; and generating a dynamic interactive health graph based on the updated adjacency matrix and node attributes.

[0122] The initial adjacency matrix is ​​constructed as follows: the adjacency matrix is ​​initialized as an N×N square matrix, where N is the number of robots in the training field, and the diagonal elements A ii =0, off-diagonal element A ij(i≠j) is assigned the edge weight W between robot i and robot j calculated earlier. ij This forms the initial adjacency matrix A0.

[0123] The adjacency matrix is ​​updated by recalculating the edge weights W between robots based on the latest collected interaction data, according to the 5-minute update cycle set in S32. ij The new value is used to update only the off-diagonal elements in the adjacency matrix, resulting in an updated adjacency matrix A(t) that matches the current time.

[0124] The updated adjacency matrix A(t) and the concurrently updated node attribute matrix V(t) (containing the health status level, abnormal feature deviation, hardware parameters, etc. of each robot at time t) are used as inputs. A dynamic interactive health graph G(t) is constructed through a graph structure encapsulation algorithm, where t is the timestamp of the current update cycle, used to mark the time dimension information corresponding to the graph structure. The final generated G(t) contains four core elements: "node set - edge set - edge weight - timestamp", which can be directly used as feature input for subsequent temporal graph neural networks.

[0125] Step S3, the steps for obtaining the corrected fault probability include:

[0126] S34: Input the features of the dynamic interactive health graph into the temporal graph neural network and output the initial failure probability. The neural network model that integrates the temporal convolutional network (TCN) and the graph convolutional network (GCN) can simultaneously capture temporal dynamic features and spatial correlation features, which can be used to infer the robot's health decay trend and initial failure probability.

[0127] like Figure 2 As shown, the steps for outputting the initial fault probability include:

[0128] S341: Perform dimensional alignment and normalization on the node and edge features of the dynamic interactive health graph to obtain standardized graph features. Dimension alignment uses zero-padding or feature filtering to ensure that all node and edge features have consistent dimensions. Node features are normalized using min-max normalization, and edge features are normalized using Z-score normalization.

[0129] S342: Divide the standardized graph features into continuous feature sequences according to the time series, as follows:

[0130]

[0131] Among them, G seq For a continuous characteristic sequence, G norm To standardize the graph features, τ is the length of the time window, set to 12, which is 60 minutes of historical data, and t is the current time.

[0132] The temporal convolutional layer of the input temporal graph neural network extracts the temporal dependencies of the feature sequence. The temporal convolutional network uses 1D convolutional kernels and ReLU activation function, extracting dependencies at different time scales through multi-scale dilated convolution, and outputting a temporal feature sequence F. temporal , represented as Where m' is the temporal feature dimension, R is the set of real numbers, τ is the time window length, and N is the number of robots in the training field.

[0133] S343: Input the output of the temporal convolutional layer into the graph convolutional layer to capture the spatial relationship features between nodes. The formula is as follows:

[0134]

[0135]

[0136] Among them, F spatial (t) represents the spatial correlation features at time t, which is the output of the graph convolutional layer and is used to characterize the spatial correlation relationships between robot nodes at that time. , where m'' is the spatial feature dimension. Represented as the normalized adjacency matrix, W g b is the graph convolution weight matrix; g Let be the bias term of the graph convolutional layer, D be the degree matrix, and be a diagonal matrix where the diagonal elements are the sum of the edge weights of the corresponding nodes. Let A be the original adjacency matrix, where each element represents the combined edge weight W between robots. ij I is the identity matrix, used to preserve the characteristics of the nodes themselves.

[0137] By concatenating temporal and spatial features, a feature vector F∈R that integrates spatiotemporal information is obtained. N×τ×m'' .

[0138] S344: The feature vectors fused with spatiotemporal information are input into the fully connected layer. The initial failure probability of each robot is calculated using the sigmoid activation function, and the initial failure probability is output. The formula is as follows:

[0139]

[0140] Among them, P init Let W be the initial failure probability of a single robot. f b is the weight matrix of the fully connected layer; f Here, γ is the bias term, and γ is the sigmoid activation function γ(a)=1 / (1+e -a ), where a is the linear output of the fully connected layer. The final output initial failure probability is represented by the set {P}. init1 ,P init2 ,...,P initN}

[0141] S35: Construct and refine the health-failure causal graph by combining historical fault data and historical health status data. The health-failure causal graph is a directed graph built on a Bayesian network. Nodes represent health characteristics and fault types, edges represent the causal relationship between the two, and edge weights represent the strength of the causal relationship, used to quantify the impact of abnormal health characteristics on the occurrence of faults. Historical fault data comes from the training field fault record database, containing 800 fault cases from the past two years, including fault type, health characteristic data before the fault, fault occurrence time, environmental parameters, etc.; historical health status data consists of robot health status records from the past two years, with a sample size of 5000 sets.

[0142] In step S35, the steps for correcting the health-failure cause-effect graph include:

[0143] S351: Based on historical fault data and historical health status data, an initial health-fault causal graph is constructed using a Bayesian network, where nodes represent health features and fault types, edges represent the causal relationships between them, and edge weights represent the initial causal relationship strength. The steps include:

[0144] The node set is determined, containing 20 health characteristic nodes and 10 fault type nodes;

[0145] Based on domain knowledge and historical data, determine the initial causal relationships between nodes;

[0146] The initial causal relationship strength (edge ​​weight) is calculated using the conditional probability P(B|A), where A is the cause node and B is the result node, i.e., the probability that B occurs when A occurs. This is obtained through historical data statistics, and the formula is as follows:

[0147]

[0148] Where, N A∩B N represents the number of cases where A and B occur simultaneously. A Let A be the number of cases that occurred.

[0149] S352: Based on failure case data and corresponding health status evolution data, calculate the confidence level of each causal relationship and update the initial causal relationship strength. Specifically, based on 500 sets of the latest failure case data and corresponding health status evolution data, calculate the confidence level of each causal relationship. The formula is as follows:

[0150]

[0151] Where, N A∩B,new N represents the number of times A and B occur simultaneously in the latest case. A,newα represents the number of occurrences of A in the latest cases; α is the weighting coefficient, usually set to 0.6, used to balance the influence of new data and historical data. The larger α is, the more significant the influence of the latest cases on the confidence level.

[0152] If a new correlation exists between health features and fault types, the corresponding node and edge are added, their confidence level is calculated as the edge weight, and the node and edge are added to the health-fault causal graph to obtain an updated health-fault causal graph.

[0153] S353: Verify the updated health-failure cause-effect graph and eliminate unreasonable causal relationships. Specifically, this is done through a combination of expert review and data consistency testing. The expert review is conducted by three experts in the field of robot fault diagnosis, who score the reasonableness of the causal relationships (0-10 points, ≥6 points is reasonable). The data consistency test calculates the support of the relationship (a frequency of occurrence ≥5 times is considered valid) and eliminates unreasonable causal relationships with an expert score <6 points or a support <5 times.

[0154] Adjust the strength of abnormal causal relationships. For retained causal relationships, if their confidence level exceeds the normal range (0.1~0.9), adjust them to a reasonable range by combining domain knowledge to obtain the corrected health-failure causal graph.

[0155] S354: Calculate the goodness-of-fit index of the corrected health-failure cause-effect graph. If the goodness-of-fit index reaches a preset threshold, the correction is complete. If not, repeat steps S352-S353 until the goodness-of-fit index meets the requirements. The Akaike Information Criterion (AIC) is used as the goodness-of-fit index to evaluate the fitting effect of the corrected cause-effect graph on historical fault data, expressed as:

[0156]

[0157] Where AIC is the goodness-of-fit index, k is the number of parameters in the cause-effect graph (i.e., the number of edges), and L is the likelihood function value of the cause-effect graph, calculated based on historical fault data. The smaller the AIC value, the better the fit. The preset AIC threshold is 100. If the AIC value of the corrected cause-effect graph is ≤100, the correction is complete; otherwise, repeat steps S352-S353 until the AIC value meets the requirements.

[0158] S36: The strength of causal associations in the health-failure cause-effect graph is incorporated as a weight into the initial failure probability calculation to obtain the corrected failure probability, expressed by the formula:

[0159]

[0160] Among them, P corr To correct for the failure probability, M represents the number of causal relationships associated with the robot; b sLet be the strength weight of the s-th causal relationship, satisfying ;P init Let P be the initial failure probability of a single robot. cause The failure probability is derived based on the health-failure causal graph. Specifically, it is obtained by tracing the corresponding "failure node" in the corrected health-failure causal graph through the "abnormal health feature node" of the robot, and then deriving the failure probability, which reflects the direct causal impact of abnormal health features on failure.

[0161] S4: Adjust the alarm level based on the corrected fault probability and health degradation trend, verify the abnormal correlation, and obtain the final alarm level, evidence data and health risk level as needed.

[0162] like Figure 3 As shown, step S4, which involves adjusting the alarm level and verifying the correlation of anomalies, includes:

[0163] S41: Preset three-level alarm thresholds, match and correct the fault probability to obtain the initial alarm level, and the specific correspondence is shown in Table 2.

[0164] Table 2:

[0165]

[0166] S42: Calculate the collaborative confidence score by combining the number of anomaly associations and the rate of change of health degradation trends in the multi-robot diagnostic set. The collaborative confidence score is an indicator that quantifies the reliability and severity of anomalies by comprehensively considering the number of anomaly associations and the rate of change of health degradation trends. The specific calculation method is as follows:

[0167]

[0168] Among them, C synergy For collaborative confidence, N max The maximum possible number of abnormal associations is typically set to 10; N assoc v represents the number of abnormal associations. decay The rate of change of health decline trend is defined as the decrease in health status level per unit time, and its normalized value range is [0,1].

[0169] S43: Adjust the initial alarm level based on the collaborative confidence level, and output the adjusted alarm level and the corresponding range of abnormal robots requiring key verification. The alarm level adjustment strategy is: if C synergy If ≥0.8, the alarm level is increased by 1 level; if 0.3≤C synergy If C <0.8, the alarm level remains unchanged; synergyIf the value is less than 0.3, the alarm level is reduced by 1 level. After adjustment, the final alarm level and the range of abnormal robots that need to be checked are output, i.e., robots with a collaboration confidence level ≥ 0.5.

[0170] S44: Based on the scope of abnormal robots requiring key verification, extract the voiceprint-vibration features of the corresponding abnormal robots from the multi-robot diagnostic set, compare them with the preset normal feature library, and obtain the voiceprint-vibration feature matching degree and the corresponding abnormal feature information. Voiceprint-vibration features are characteristic combinations of sound signals and vibration signals generated by components such as motors and joints during robot operation, and are an important basis for judging whether components are faulty. Voiceprint data is collected through the microphone mounted on the robot body, with a sampling frequency of 16kHz. Vibration data is collected through vibration sensors mounted on the joints and motors, with a sampling frequency of 1kHz. The preset normal feature library contains 100 sets of voiceprint-vibration feature samples from healthy robots, constructed as a feature library and stored at the edge node. The matching method for the voiceprint-vibration feature matching degree is expressed as follows:

[0171]

[0172] Among them, F audio-vib To obtain the voiceprint-vibration feature vector, the voiceprint features (12-dimensional) are extracted using Mel-frequency cepstral coefficients, and the vibration features (16-dimensional) are extracted using wavelet transform. These are then concatenated to obtain the voiceprint-vibration feature vector F. audio-vib ∈R 28 F normal Let S be the sample feature vector in the normal feature library, and ||·|| be the L2 norm. match If the value is less than 0.7, it is determined to be an abnormal voiceprint-vibration feature, and the corresponding abnormal feature information is recorded.

[0173] S45: Based on the abnormal feature information corresponding to the voiceprint-vibration feature matching degree, visual evidence such as the robot's running posture and trajectory is collected. The visual data is collected by high-definition cameras (30fps) deployed in the training field to collect the robot's running posture and trajectory data.

[0174] S46: Analyze the synchronicity and correlation between the voiceprint-vibration feature matching degree and visual evidence features to verify the effectiveness of the anomaly correlation. Feature synchronicity refers to the degree of synchronization between abnormal voiceprint-vibration features and visual anomaly features in the time dimension. The synchronicity analysis method is to calculate the time difference Δt between the occurrence time of abnormal voiceprint-vibration features and the occurrence time of visual anomaly features. sync If Δt sync If the time is ≤100ms, it is considered time synchronization. Correlation analysis involves calculating the correlation coefficient r between the two abnormal features. If r ≥ 0.6, it is considered feature correlation. If both time synchronization and feature correlation are satisfied, the abnormal correlation is verified as valid; otherwise, it is considered invalid correlation, and the alarm level is readjusted.

[0175] S5: Construct a multimodal health embedding model and update the multimodal health embedding model using the final alarm level, evidence data, and health risk level.

[0176] Step S5, the steps for constructing the multimodal health embedding model include:

[0177] S51: Construct a multimodal fusion network, inputting ontology health feature data, interaction health data, and simulation parameters as training samples to obtain a formatted training dataset. To address the heterogeneity of different modal data, preprocessing and modality alignment are performed to ensure data consistency. The specific steps are as follows:

[0178] Outliers were removed from the simulation parameter data, and missing values ​​were filled in. The ontology health feature and interaction health feature data were verified a second time to ensure that the data format was consistent with the data format after S1 preprocessing.

[0179] Based on the sampling timestamp of the ontology health data, timestamp matching is performed on the interactive health data and simulation parameters, and a resampling method is used to unify the sampling frequency to 10Hz. Through feature filtering and zero padding, the ontology health feature dimension is fixed to 15 dimensions, the interactive health feature dimension is fixed to 12 dimensions, and the simulation parameter dimension is fixed to 8 dimensions.

[0180] The ontological health features and interactive health features follow the min-max normalization of S1; the simulation parameters are normalized using Z-score to ensure that the range of values ​​for each modality data is uniformly [0,1] or a standard normal distribution, so as to avoid the impact of magnitude differences on model training.

[0181] S52: Using health status level as a label, train the network on the formatted training dataset using the cross-entropy loss function to obtain the network parameters to be optimized.

[0182] The ontology health feature extraction branch adopts a "convolutional layer + pooling layer" structure to adapt to the local temporal correlation of ontology health features. The specific structure is as follows: input layer, 1D convolutional layer 1, max pooling layer 1, 1D convolutional layer 2, global average pooling layer, and output ontology feature vector F∈R.

[0183] The interactive health feature extraction branch adopts a graph convolutional layer (GCN) structure to capture the spatial correlation characteristics of multi-robot interactions. The specific structure consists of an input layer, graph convolutional layer 1, graph convolutional layer 2, node feature aggregation, and output interactive feature vector F∈R.

[0184] The simulation parameter extraction branch adopts a fully connected layer structure to adapt to the high-dimensional numerical characteristics of simulation parameters. The specific structure consists of an input layer, a fully connected layer 1, a Dropout layer, a fully connected layer 2, and an output simulation feature vector F∈R.

[0185] S53: Iteratively optimize the network parameters to obtain a multimodal health embedding model that outputs a unified-dimensional health feature embedding vector. The health feature embedding vector is represented as:

[0186]

[0187]

[0188] Among them, E health Embedded vectors for 32-dimensional health features; W embed This is the embedding layer weight matrix, with dimensions 64×32; b embed For the embedding layer bias term, dimension 1×32; F fusion To fuse the feature vectors, three modalities of data—including ontological health feature data, interactive health feature data, and simulation parameters—were preprocessed and aligned. These data were collected from the robot's own sensors, training environment sensors, and robot communication module, and then generated by the robot simulation platform. body w inter w sim These are the attention weights for ontological health features, interaction health features, and simulation parameters, respectively, satisfying W. body +w inter +w sim =1, obtained through adaptive learning during model training, with initial values ​​all set to 1 / 3. F body F inter F sim These represent ontological health features, interactive health features, and simulation parameters, respectively. σ is the sigmoid activation function, which maps the embedding vector values ​​to the interval [0,1].

[0189] S6: An interpretable algorithm is used to obtain the root cause of the fault and the health impact chain, generating maintenance suggestions, health maintenance plans, and risk assessment reports to realize the practical application of fault prediction results. For the interpretable algorithm, the "Bayesian interpretability analysis + SHAP value (SHapley Additive exPlanations) fusion algorithm" is selected. Bayesian interpretability analysis is adapted to the probabilistic reasoning characteristics of the health-fault cause-effect graph, and SHAP value can quantify the contribution of each node to the fault result. The combination of the two achieves dual analysis of "causal logic + contribution quantification".

[0190] Step S6, which involves using an interpretable algorithm to obtain the root cause of the fault and the health impact links, includes:

[0191] S61: An interpretable algorithm is used to analyze the health-failure causal graph, identifying key nodes with high fault contribution and their association information. Key nodes with high fault contribution refer to healthy characteristic nodes or intermediate fault nodes whose contribution to the probability of fault occurrence meets a preset threshold; they are core candidate nodes for the root cause of the fault. Node association information includes quantitative information such as the causal direction, association strength, and propagation delay between nodes.

[0192] The steps for identifying critical nodes with high contribution to fault detection include:

[0193] Initialize the Bayesian interpretability analysis module and the SHAP value calculation module, and set the core parameters. The contribution screening threshold θ=0.2, which is verified through historical failure cases to balance screening precision and recall. The sample size for SHAP value calculation is 1000 groups, taken from historical failure data that match the current failure type. The confidence level of Bayesian inference is 0.95.

[0194] The corrected health-failure cause-effect graph is input into the Bayesian interpretability analysis module. Using the failure type nodes determined in S4 as the target output nodes, the posterior probabilities of each predecessor node (cause node) are inferred, and the causal effect values ​​between nodes are calculated. These causal effect values ​​reflect the degree of influence of the cause node on the failure node.

[0195] The node characteristics and corresponding failure probabilities of the health-failure causal graph are input into the SHAP value calculation module to obtain the SHAP value of each node, where a positive value indicates that the node increases the failure probability, and a negative value indicates that the failure probability is suppressed. The absolute value of the SHAP value is taken as the contribution metric. The comprehensive contribution is calculated by fusing the Bayesian causal effect value and the SHAP value, and the calculation method is expressed as follows:

[0196]

[0197] Among them, SHAP abs SHAP is the absolute value of the SHAP value, normalized to [0,1]; CEV is the Bayesian causal effect value, normalized to [0,1]; 0.6 and 0.4 are the absolute values ​​of the SHAP values. abs The weighting coefficients for CEV were determined through expert review, with priority given to the accuracy of the quantitative assessment of contribution.

[0198] The key node identification method is as follows: nodes with a comprehensive contribution greater than or equal to a preset screening threshold are selected as high-contribution key nodes for the fault. Their core function is to filter out nodes with weak impact on the fault, focusing on key influencing factors and providing a precise set of nodes for subsequent causal path tracing and root cause link construction. The preset screening threshold is typically set to 0.2, validated through historical fault cases to balance screening precision and recall. The association information for each key node includes the associated node ID and causal direction, association strength, propagation delay, and matching evidence data features. Association strength corresponds to the confidence level in the health-fault causal graph. Propagation delay is the average time from the abnormality of the cause node to the abnormality of the result node, obtained from historical data statistics. Matching evidence data features include, for example, "abnormal motor temperature" corresponding to "abnormal vibration amplitude of 100-200Hz" in the soundprint-vibration features.

[0199] S62: Combining the final alarm level, evidence data, key nodes, and node association information, trace the causal relationship paths between key nodes. Using the final fault node as the endpoint, conduct bidirectional tracing based on the node association information in the health-failure cause-effect graph. Root cause localization starts from the fault node and traces upstream key nodes along the cause-effect direction, recording the association strength and propagation delay of each node to form multiple potential root cause paths. Impact range prediction starts from the initial cause node obtained through reverse tracing and traces subsequent impact nodes along the cause-effect direction, recording the propagation relationships between nodes to lay the foundation for constructing subsequent health impact chains.

[0200] The completeness of the potential root cause paths traced is verified. If there are logical gaps, such as the "battery voltage fluctuation → motor failure" missing the intermediate node "abnormal motor current", the missing intermediate key nodes are supplemented by combining domain knowledge and historical failure cases to ensure the causal logic of the path is coherent.

[0201] S63: The causal pathways are ranked according to their degree of impact to form the root cause chain of the failure, and the scope and transmission path of health impacts are derived to obtain the health impact chain. The ranking of the root cause paths prioritizes the completed potential root cause paths by setting three ranking indicators: path contribution, path confidence, and path transmission efficiency. The Analytic Hierarchy Process (AHP) is used in conjunction with expert review and historical data verification to determine the indicator weights, typically set to 0.5 for path contribution, 0.3 for path confidence, and 0.2 for path transmission efficiency. The comprehensive ranking score for each path is calculated using the following formula:

[0202]

[0203] Among them, S sum C represents the path contribution. prod For path confidence, T total This represents the total propagation delay along the path.

[0204] The top three paths with the highest comprehensive ranking scores were selected as the core root cause paths of the fault; if there were fewer than three, all paths were retained. Paths were then sorted from highest to lowest score to identify the primary and secondary root cause paths. Based on the root cause localization results, and combined with the multi-robot dynamic interaction health map, the scope of health impact was defined, including the individual impact scope and the group impact scope. The individual impact scope refers to the components within the faulty robot affected by the root cause node, such as a motor fault affecting the joint drive system or control system. The group impact scope refers to other robots affected by the interaction with the faulty robot; for example, in collaborative training, the motion deviation of the faulty robot affects the positioning accuracy of adjacent robots. Robots with a collaborative confidence level ≥ 0.5 were identified as the core affected objects.

[0205] The construction of the health impact chain integrates the scope of health impact with the transmission path of positive tracing. It constructs the health impact chain according to the logic of "initial root cause node → intermediate impact node → terminal impact node", and marks the degree of impact and transmission path type of each node.

[0206] Based on the priority and key node characteristics of the root cause chain, tiered maintenance recommendations are generated, including emergency maintenance recommendations, routine maintenance recommendations, and preventative maintenance recommendations. Emergency maintenance recommendations primarily target the initial root cause node in the primary root cause chain, specifying emergency handling measures, operating procedures, required tools, and safety precautions. Routine maintenance recommendations primarily target the key nodes in the secondary root cause chain, developing a 24-hour maintenance plan that specifies maintenance steps, component replacement lists, and acceptance criteria. Preventative maintenance recommendations primarily target intermediate nodes in the health impact chain, proposing regular inspection and maintenance measures to prevent the fault from escalating.

[0207] By combining the health status assessment results of the multimodal health embedding model with historical maintenance data, short-term, medium-term, and long-term full-cycle health maintenance plans are formulated. Short-term maintenance focuses on key components corresponding to current abnormal characteristics, develops high-frequency inspection plans, sets inspection indicators and anomaly thresholds, and clarifies data recording and reporting processes. Medium-term maintenance expands the maintenance scope to all components covered by the health impact chain, conducts in-depth inspections, and develops component maintenance plans. Long-term maintenance, based on robot service life, hardware aging patterns, and trend prediction results from the multimodal health embedding model, develops component replacement and system upgrade plans.

[0208] In summary, this method for fault prediction and health management in a humanoid robot training field captures the correlation characteristics of multi-robot groups through a dynamic interactive health graph. It integrates a temporal graph neural network and a modified health-fault causal graph, significantly improving the accuracy and timeliness of fault probability prediction and enabling early warning of faults. By verifying the correlation of anomalies through multi-source data of voiceprint, vibration, and vision, it ensures the reliability of alarm level adjustments and effectively reduces the risk of false alarms and missed alarms. Furthermore, the construction and updating of a multimodal health embedding model further strengthens the ability to represent health status, providing stronger model support for assessment and prediction. Through interpretable algorithms, it clarifies the root causes of faults and the health impact chain, generating graded maintenance suggestions, full-cycle health maintenance plans, and quantitative risk assessment reports. This not only achieves accurate fault tracing and targeted handling, reducing operation and maintenance costs and downtime losses, but also helps the training field achieve full-cycle health management of robots, ensuring the continuity and safety of collaborative training, and improving the overall efficiency and intelligence level of operation and maintenance management.

[0209] Through the above description of the embodiments, those skilled in the art can clearly understand that each embodiment can be implemented by means of software plus necessary general-purpose hardware platforms, and of course, it can also be implemented by hardware. Based on this understanding, the above technical solutions, in essence or the part that contributes to the prior art, can be embodied in the form of a software product. This computer software product can be stored in a computer-readable storage medium, such as ROM / RAM, magnetic disk, optical disk, etc., and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute the methods described in the various embodiments or some parts of the embodiments.

[0210] Finally, 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 foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method for fault prediction and health management in a humanoid robot training field, characterized in that, include: S1: Collect robot health data, controller logs, and robot interaction detection data, and preprocess them to obtain robot health feature data and interaction health data; S2: Based on the ontological health feature data, the interactive health data, and the interactive requirements, a preliminary health assessment is performed to obtain a multi-robot diagnostic set and the preliminary health assessment results; S3: Based on the multi-robot diagnostic set and the initial health assessment results, construct a dynamic interactive health graph and infer the health decline trend through a time-series graph neural network; By combining historical fault data and historical health status data, a health-fault cause-effect graph is constructed and corrected to obtain the corrected fault probability. S4: Adjust the alarm level according to the corrected fault probability and the health decay trend, verify the abnormal correlation, and obtain the final alarm level, evidence data and health risk level as needed. S5: Based on the ontology health feature data, interaction health data, and simulation parameters, construct a formatted training dataset; A multimodal health embedding model is constructed, and a health status assessment result is output; the multimodal health embedding model is then updated using the final alarm level, evidence data, and health risk level. S6: Based on the health status assessment results and historical maintenance data, formulate short-term, medium-term, and long-term full-cycle health maintenance plans; An interpretable algorithm is used to obtain the root cause of the failure and the health impact chain, and to generate maintenance recommendations, health maintenance plans and risk assessment reports.

2. The method for fault prediction and health management in a humanoid robot training field according to claim 1, characterized in that, Step S2, the steps for conducting the initial health assessment include: S21: Based on the ontological health feature data and the interactive health data, construct basic health assessment rules; S22: Compare and match the ontological health feature data and the interactive health data feature by feature with the basic health assessment rules, mark abnormal features that exceed the normal value range, and record the degree of deviation and occurrence time of each abnormal feature. S23: Input the labeled abnormal features into a locally deployed lightweight deep learning edge model. The model analyzes the correlation between abnormal features, eliminates mislabeling caused by isolated noise points, outputs the health status level, and generates a preliminary health assessment result document.

3. The method for fault prediction and health management in a humanoid robot training field according to claim 2, characterized in that, In step S21, the steps for constructing basic health assessment rules include: S211: Define the dimensions of health assessment; S212: Based on the standard parameters of the humanoid robot at the factory, the rated operating parameters of the training ground, and historical fault-free data, the normal value range of each dimension of the feature is determined through statistical analysis to form basic threshold rules; S213: Based on the aforementioned interaction requirements, set feature association constraint rules to clarify the linkage, matching, and response matching thresholds of features between the same or different humanoid robots; S214: Set the environment adaptation coefficient, and correct the basic threshold rule according to the training environment parameters to obtain the threshold rule after environment adaptation. S215: Based on the threshold rules adapted to the environment and the feature association constraint rules, an initial basic health assessment rule is formed; through verification and optimization using historical fault cases, until the preset performance requirements are met, the rule construction is completed.

4. The method for fault prediction and health management in a humanoid robot training field according to claim 1, characterized in that, Step S3, the steps for constructing the dynamic interactive health graph include: S31: Set the robot as a node, and set the interaction strength, frequency and health correlation as edge weights; S32: Based on the multi-robot diagnostic set and the initial health assessment results, update the attributes of the node and the edge weights; S33: The dynamic interactive health graph is generated using a graph construction algorithm.

5. The method for fault prediction and health management in a humanoid robot training field according to claim 1, characterized in that, Step S3, the steps for obtaining the corrected fault probability include: S34: Input the features of the dynamic interactive health graph into the time-series graph neural network and output the initial fault probability; S35: Construct and correct the health-failure cause-effect graph by combining historical fault data and historical health status data; S36: The causal correlation strength of the health-failure cause-effect graph is used as a weight to incorporate into the initial failure probability calculation to obtain the corrected failure probability.

6. The method for fault prediction and health management in a humanoid robot training field according to claim 5, characterized in that, Step S34, the step of outputting the initial fault probability, includes: S341: Perform dimension alignment and normalization on the node features and edge features of the dynamic interactive health graph to obtain standardized graph features; S342: Divide the standardized graph features into continuous feature sequences according to the time series, input them into the temporal convolutional layer of the temporal graph neural network, and extract the temporal dependencies of the feature sequences; S343: Input the output of the temporal convolutional layer into the graph convolutional layer to capture the spatial correlation features between nodes and obtain a feature vector that integrates spatiotemporal information; S344: Input the feature vector of the fused spatiotemporal information into the fully connected layer, calculate the initial failure probability of each robot through the activation function, and output the initial failure probability.

7. The method for fault prediction and health management in a humanoid robot training field according to claim 5, characterized in that, In step S35, the step of correcting the health-failure cause-effect graph includes: S351: Based on historical fault data and historical health status data, an initial health-fault causal graph is constructed using a Bayesian network, where nodes represent health features and fault types, edges represent the causal relationships between the two, and edge weights represent the initial causal relationship strength. S352: Based on the failure case data and the corresponding health status evolution data, calculate the confidence level of each causal relationship and update the initial causal relationship strength; if there is a new relationship between health features and failure types, add it to the health-failure causal graph to obtain the updated health-failure causal graph. S353: Verify the updated health-failure cause-effect graph, remove unreasonable causal relationships, adjust the strength of abnormal causal relationships, and obtain the corrected health-failure cause-effect graph; S354: Calculate the goodness-of-fit index of the corrected health-failure cause-effect graph. If the goodness-of-fit index reaches a preset threshold, the correction is completed. If it does not reach the threshold, repeat steps S352-S353 until the goodness-of-fit index meets the requirements.

8. The method for fault prediction and health management in a humanoid robot training field according to claim 1, characterized in that, Step S4, which involves adjusting the alarm level and verifying the correlation of anomalies, includes: S41: Preset three-level alarm thresholds, and match the corrected fault probability to obtain the initial alarm level; S42: Calculate the collaborative confidence level by combining the number of abnormal associations in the multi-robot diagnostic set and the rate of change of the health decline trend; S43: Adjust the initial alarm level according to the collaborative confidence level, and output the adjusted alarm level and the corresponding range of abnormal robots that need to be checked; S44: Based on the range of abnormal robots that need to be checked, extract the voiceprint-vibration features of the corresponding abnormal robots in the multi-robot diagnostic set, compare them with the preset normal feature library, and obtain the voiceprint-vibration feature matching degree and the corresponding abnormal feature information. S45: Based on the abnormal feature information corresponding to the voiceprint-vibration feature matching degree, collect visual evidence of the abnormal robot's running posture and trajectory; S46: Analyze the feature synchronization and correlation between the voiceprint-vibration feature matching degree and the visual evidence to verify the effectiveness of the abnormal correlation.

9. A method for fault prediction and health management in a humanoid robot training field according to claim 2, characterized in that, Step S5, the steps for constructing the multimodal health embedding model include: S51: Construct a multimodal fusion network, input the ontology health feature data, the interaction health data and simulation parameters as training samples, and construct a formatted training dataset; S52: Using the health status level as a label, train the network on the formatted training dataset using the cross-entropy loss function to obtain the network parameters to be optimized; S53: Iteratively optimize the network parameters to be optimized to obtain the multimodal health embedding model that outputs health status assessment results.

10. The method for fault prediction and health management in a humanoid robot training field according to claim 1, characterized in that, Step S6, which involves using an interpretable algorithm to obtain the root cause of the fault and the health impact links, includes: S61: Short-term maintenance focuses on key components corresponding to current abnormal characteristics, formulates high-frequency inspection plans, and sets inspection indicators and abnormal thresholds; medium-term maintenance expands the maintenance scope to all components covered by the health impact chain, conducts in-depth inspections, and formulates component maintenance plans; long-term maintenance formulates component replacement plans and system upgrade plans based on robot service life, hardware aging patterns, and health status assessment results of the multimodal health embedding model. S61: Use an interpretable algorithm to analyze the health-failure cause-effect graph and filter out key nodes with high contribution to failure and node association information; S62: Combining the final alarm level, the evidence data, key nodes and node association information, trace the causal relationship path between each of the key nodes; S63: Sort the causal paths according to their degree of influence to form the root cause chain of the failure, deduce the scope of health impact and transmission path, and obtain the health impact chain.

Images