Novel servo robot health degree detection algorithm

Abstract

The invention discloses a novel servo robot health degree detection algorithm. The algorithm comprises the following steps of S1, performing multilevel data processing and auxiliary inference based on expert experience; s2, a robust multi-classification SVM (Support Vector Machine) model of Group lasso variable selection and truncation loss design is fused; s3, unsupervised abnormal distribution detection based on a VAE (variational auto-encoder); s4, constructing a multi-mode health degree detection and fault identification framework through organic combination; according to the method, normal and abnormal distribution can be identified by using an unsupervised deep learning model under the condition of no annotated data, and accurate diagnosis and health degree evaluation of different fault types can be realized by combining expert experience and a multi-classification statistical model with variable selection and robust loss design under the scene of annotated data; the defects in the prior art are overcome, the precision and efficiency of health degree detection of the servo robot are improved, and the urgent requirements of the modern automobile manufacturing industry for high-precision, high-efficiency and intelligent fault diagnosis are met.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of servo robot health detection, in particular to a novel servo robot health detection algorithm. BACKGROUND

[0002] In recent years, with the rapid development of industrial big data, artificial intelligence technology and sensor technology, data-driven methods have been gradually applied to the field of servo robot health detection. Such methods use multi-source sensing signals such as vibration data, temperature data, current and voltage data, force and torque data to achieve the recognition and diagnosis of robot faults through statistical modeling or deep learning models. Compared with purely rule-based or threshold-based methods, data-driven methods have higher adaptability and better generalization ability. Currently, the mainstream technology route for servo robot health detection generally includes fault rule judgment based on expert experience and pure data-driven machine learning or deep learning models. The judgment method based on expert experience relies on pre-defined diagnostic procedures or thresholds, which can quickly identify common faults of known types in specific scenarios. The method is intuitive and easy to understand. Deep learning models, such as unsupervised algorithms based on autoencoders (AE) or variational autoencoders (VAE), can learn the feature distribution of servo robots under normal operating conditions, thereby detecting abnormal data. In the diagnosis process of servo robot health, some solutions attempt to combine expert experience and machine learning methods by using expert rules for variable selection or abnormal data elimination in the preprocessing stage, and then using machine learning models to classify or cluster fault patterns.

[0003] The current traditional servo robot health detection relies on artificial experience or single threshold judgment, and is difficult to adapt to diversified fault types. In the automobile manufacturing process, the servo robot undertakes key operations such as welding, spraying and carrying. Once a fault occurs, it will seriously affect the production efficiency and the quality of parts. The existing method uses artificial experience or single threshold to judge the health state of the robot, which is difficult to make accurate diagnosis for complex and multiple fault forms, and is prone to missed detection or misjudgment. In the health detection, the number of positive samples (fault samples) is too small and the distribution is uneven, which makes it difficult for the model to fully learn abnormal features. For servo robots, fault states often involve multiple subsystems or abnormal sensor signals. However, in the real factory environment, the number of fault samples is usually limited and the difference is huge, which makes it difficult to fully learn some fault types. The traditional method lacks robustness when facing different fault mechanisms, lacks systematic use of expert experience and efficient data-driven algorithm means, and is difficult to build an accurate and scalable health diagnosis system. Although a large amount of expert experience and rules have been accumulated in the automobile industry, how to combine statistical or deep learning methods to realize comprehensive and accurate health detection of time domain, frequency domain and other multi-source sensor signals is still lacking of mature solutions, which is difficult to meet the development needs of modern automobile manufacturing for high efficiency, stability and intelligence. SUMMARY

[0004] The present application provides a novel servo robot health detection algorithm, which can effectively solve the problems raised in the above background art.

[0005] To achieve the above purpose, the present application provides the following technical scheme: a novel servo robot health detection algorithm, comprising the following steps:

[0006] S1, multi-level data processing and auxiliary inference based on expert experience;

[0007] S2, robust multi-class SVM model integrating Group lasso variable selection and truncated loss design;

[0008] S3, unsupervised anomaly distribution detection based on variational autoencoder VAE;

[0009] S4, constructing a multi-mode health detection and fault recognition framework by organic combination;

[0010] The S1 comprises the following steps:

[0011] The S101 comprises the following steps:

[0012] The S102 comprises the following steps:

[0013] According to the technical solution, the S101 is used for health degree diagnosis of the servo robot in the automobile manufacturing factory, and a large number of technical personnel have fault troubleshooting experience, which is usually embodied as rule type knowledge, process type knowledge and abnormal working condition disposal.

[0014] The rule type knowledge is explicit threshold value and logical judgment.

[0015] The process type knowledge is a relatively fuzzy and semi-explicit judgment process.

[0016] The abnormal working condition disposal refers to an abnormal working condition example and a disposal method based on fault logs and maintenance reports of specific production events.

[0017] In order to fully utilize the knowledge in the algorithm system, an expert experience database is established, and the database can be stored in a structured manner according to the dimensions of variable-threshold value-rule priority-associated fault type, thereby providing support for subsequent data processing and model inference.

[0018] According to the technical solution, the S102 is used for processing tens to hundreds of measurement variables of multi-source sensing data of the servo robot, including vibration, temperature, current and voltage, force and torque. In this process, a part of the variables has limited effect on fault detection or even brings noise. Through the expert experience database, a key variable list is refined, including explicit variable exclusion, threshold directional screening, expert-defined threshold, statistical model and expert rule fusion.

[0019] The explicit variable exclusion is that some sensors change very little under normal working conditions of the robot, and are not helpful for identifying various faults in actual historical analysis. Therefore, the expert experience is marked to exclude and reduce the weight of such variables.

[0020] The threshold directional screening is used for including relevant variables in the key monitoring list when the temperature exceeds the threshold value and the force feedback exceeds a certain safety range, and giving higher weight to the variables in subsequent model training and prediction.

[0021] In actual implementation, the following formal expression can be used:

[0022] Ω={x i ∣ExpertRule(x i )=TRUE};

[0023] Wherein, Ω represents a variable set reserved after screening by the expert experience database, ExpertRule represents a series of screening logic or threshold value judgment functions based on expert experience, which can be a Boolean output or a series of scoring functions corresponding to different importance levels.

[0024] In the factory environment, noise data and sensor fault data will cause abnormal outliers. The detection rules of outliers are set by expert experience, and the outliers are removed and corrected;

[0025] Expert-defined threshold: for known high-risk sensors and fault types, if the measurement value violates the artificially set device limit indicator, it is directly determined as dirty data;

[0026] Statistical model and expert rule fusion: using the 3σ principle based on historical data distribution, combined with the upper and lower limit correction coefficient defined by experts, the data is filtered in multiple layers;

[0027] It can be formalized as follows: if there is a certain sensor variable x i Satisfies:

[0028] |x i -μ i |>k·σ i ;

[0029] Where μ i and σ i are the mean and standard deviation of the variable, respectively, and k is the correction coefficient set by the expert. When the above formula is satisfied, some expert logic is also satisfied, and the abnormal value is determined and removed.

[0030] According to the above technical solution, the S2, when the data in the factory field has a certain amount of labeling, a supervised learning model capable of distinguishing common multiple fault types is constructed. Traditional SVM is good at processing high-dimensional data and has strong generalization ability, but in the scene of a small amount of labeled data, redundant variables and large differences between fault samples, pure SVM often has overfitting and unstable model problems;

[0031] In the SVM multi-classification strategy, the variable selection mechanism of Group lasso is fused, and the truncated loss function with anti-outlier ability is designed. Specifically, it can be divided into three key points of model objective function, variable grouping and regularization term, and truncated loss design.

[0032] Let the training data set be Where x i ∈R d represents the feature vector of the i-th sample, y i ∈{1,2,…,C} is the fault class label;

[0033] Using one-to-many or one-to-one strategy can be extended to multi-classification SVM. Using one-to-many scheme, C groups of decision functions w c ,b c are needed to learn for each classification c. The loss is defined as follows:

[0034]

[0035] where, is the truncated loss function, R({w c}) is the regularization term, and λ is the regularization coefficient.

[0036] According to the technical scheme, in actual production, the measurement variables can be naturally divided into different groups, divided according to sensor types or physical parts, and Grouplasso is introduced to select and remove some groups as a whole, which is beneficial to enhance the interpretability and dimensionality reduction effect of the model.

[0037] If d features are divided into G groups, according to expert experience, the variable index set of the gth group is denoted as I g , and the weight w c of each classifier can be defined as:

[0038]

[0039] In the multi-classification scenario, the same grouping sparsity is shared by all classifiers, and then the grouping constraint is performed on all w c to obtain the following regularization term:

[0040]

[0041] where, α g is the weight coefficient of each group, and the weight coefficient is set by experts according to prior knowledge. When α g is large, the gth group of variables is more likely to be compressed or even set to zero. If α g is small, the group is more likely to be retained.

[0042] By controlling different α g , Group lasso with expert information can be realized, and the fusion of data-driven and expert experience is considered.

[0043] According to the technical scheme, in the actual scenario, some samples of the servo robot have large noise and extreme outliers. The truncated loss is introduced, and when the misclassification degree exceeds a certain threshold, the loss value is no longer increased, so as to prevent the extreme samples from excessively affecting the overall decision. Taking the binary classification scenario as an example, the truncated Hinge loss can be defined as:

[0044] l trunc (y i ,f(x i ))=min(L max ,max(0,1-y i ·f(x i )));

[0045] where L max is a truncation threshold, when the misclassification severity exceeds a certain range, the loss is fixed at a constant, which not only preserves the idea of hinge loss interval maximization, but also prevents the adverse effects of a small number of extreme samples.

[0046] According to the above technical scheme, S3, in the servo robot health detection task, it is difficult to accurately detect new types of faults purely relying on a supervised learning model, and a variational autoencoder VAE is introduced to learn the multi-source data distribution of the robot in a normal state, so as to identify abnormal data;

[0047] The VAE includes an encoder and a decoder, and for an input sample x, the encoder maps it to a latent space z, and the decoder generates a reconstructed output from z The goal of VAE is to minimize the reconstruction error and follow the prior distribution in the latent space;

[0048] The prior distribution of the latent variable z is defined as: The encoder outputs the conditional distribution q φ (z| x), and the decoder outputs p θ (x| z), and the training objective of VAE can be represented as maximizing the evidence lower bound ELBO:

[0049]

[0050] where D KL is the Kullback-Leibler divergence, which is used to constrain the distribution output by the encoder not to deviate too much from the Gaussian prior. Through training on a large amount of normal operating data, the encoder learned by VAE can map the distribution of normal data to a compact latent space, and the decoder can also reconstruct the original data with small error.

[0051] According to the above technical scheme, S3, after obtaining the trained VAE, the new sensor data x (new) is input into the model, and its reconstruction error can be calculated as an anomaly score, specifically:

[0052]

[0053] where, is the reconstruction output of VAE for input x (new) If the error is significantly greater than the statistical threshold learned on normal data, it can be considered that the data corresponds to an abnormal state, in addition, rules set by experts are introduced for secondary judgment, and higher weights are applied to the reconstruction error of the corresponding variables, specifically:

[0054]

[0055] wherein the weight ω j The determination of the importance of the variable from the expert experience library can also partially utilize the expert experience in the unsupervised detection stage, and improve the accuracy of detection.

[0056] According to the technical solution, the S4 organically combines the expert experience system, the truncated multi-classification SVM with Grouplasso, and the VAE model, constructs a multi-mode health degree detection and fault identification framework, and the overall process can be divided into the following steps:

[0057] a, data acquisition and preliminary cleaning;

[0058] b, expert experience driven deep preprocessing;

[0059] c, model training and inference, supervised SVM, unsupervised VAE;

[0060] d, result fusion and decision output;

[0061] In addition, a fault handling suggestion based on a large language model is adopted, historical treatment methods are formed into formatted texts, and Lora fine-tuning is performed on the basis of llama3. When the model outputs a fault, the fault handling suggestion is directly output. The model will be enhanced in performance as the historical data increases.

[0062] According to the technical solution, the S4 is configured to integrate the results of supervised and unsupervised determination and embody expert experience, and provide a configurable multi-level fusion strategy, which is as follows:

[0063] If the SVM is highly certain and consistent with the expert rules, the fault type is directly output, and the corresponding treatment suggestion is given;

[0064] If the SVM is not highly confident and the VAE analysis shows that the abnormal score is extremely high, it is determined as a possible new fault type, and manual or expert system is submitted for further inspection, and the sample and its fault label can be included in subsequent training, and the SVM confidence is determined by the margin of the decision function distance from the classification boundary;

[0065] If the SVM shows normal or low risk, but the VAE shows that the abnormal score is relatively high, the semi-explicit rules for specific sensor data in the expert experience are combined, and when the rules are triggered, it is considered as a suspicious state and the detection frequency and alarm triggering are increased;

[0066] If the VAE determines normal and the SVM determines fault, it belongs to model overfitting and improper design of expert system threshold, and the model parameters need to be updated according to subsequent labeling, and the expert threshold is corrected.

[0067] Compared with the prior art, the present application has the following advantages:

[0068] 1. Overall improve detection accuracy and robustness: existing technology usually relies on a single machine learning or simple threshold-based rules, vulnerable to sample size and data noise, compared with the deep integration of expert experience and multi-source data processing strategy, from the source to filter out interference information, combined with Group lasso and robust multi-class SVM model with truncated loss, significantly enhance the detection ability and robustness of different fault types.

[0069] 2. Stronger adaptability in the scene of limited samples and high fault heterogeneity: by introducing VAE unsupervised anomaly detection, it can effectively mine the feature distribution of normal operation data under a small amount of labeled data, and quickly detect new abnormal situations, at the same time, with the help of expert experience, it can accurately locate the possible key fault position, solve the pain point that diversified faults are difficult to be detected by traditional algorithms in complex scenarios.

[0070] 3. Both explainability and scalability: unlike traditional deep learning methods, expert rules that can be quantified or explicitly described are introduced in the stages of data preprocessing, variable selection and model training. This approach not only ensures the explainability of complex faults, but also makes subsequent model iteration and function expansion more flexible, allowing rules to be updated or models to be retrained at any time according to new fault knowledge or new type of data.

[0071] 4. Adhere to the actual industrial demand, easy to deploy and maintain: at the beginning of the design, it is aimed at the changing production environment in the automobile manufacturing process, fully considering the reality factors such as sensor cost, data acquisition convenience and accumulative update of expert knowledge, so it is easier to deploy and promote on the actual production line, and can continuously optimize the model with the help of perfect expert diagnosis mechanism, further improve the precision and efficiency of servo robot health degree detection

[0072] In summary, the present application can not only identify normal and abnormal distribution using unsupervised deep learning model in the absence of labeled data, but also combine expert experience and multi-class statistical model with variable selection and robust loss design in the presence of labeled data, realize accurate diagnosis and health degree evaluation of different fault types, overcome the shortcomings of existing technology, meet the urgent needs of modern automobile manufacturing industry for high precision, high efficiency and intelligent fault diagnosis. BRIEF DESCRIPTION OF DRAWINGS

[0073] The accompanying drawings are used to provide a further understanding of the present application, and constitute a part of the specification, together with embodiments of the present application, used to explain the present application, and do not constitute a limitation of the present application.

[0074] In the drawings:

[0075] Figure 1 is the overall flowchart of the health degree detection of the present application;

[0076] Figure 2 is a schematic diagram of the truncated loss function of the present application;

[0077] Figure 3 is a schematic diagram of the framework of the VAE model of the present application. DETAILED DESCRIPTION

[0078] The preferred embodiments of the present application are described below in conjunction with the accompanying drawings, it should be understood that the preferred embodiments described herein are only used to illustrate and explain the present application, and are not used to limit the present application.

[0079] Embodiment: As shown in the present application provides a technical solution, a new type of servo robot health detection algorithm, comprising the following steps: Figure 1

[0080] S1, multi-level data processing and auxiliary inference based on expert experience;

[0081] S2, robust multi-class SVM model fusing Group lasso variable selection and truncated loss design;

[0082] S3, unsupervised anomaly distribution detection based on variational autoencoder VAE;

[0083] S4, construct a multi-mode health detection and fault identification framework by organic combination.

[0084] Based on the above technical solution, S1, comprising the following steps:

[0085] S101, construction of expert experience database;

[0086] S102, data preprocessing and variable selection;

[0087] S101, in the automobile manufacturing factory, for the health diagnosis of servo robot, often has a lot of technical personnel's troubleshooting experience, these experiences usually embody as: rule type knowledge, process type knowledge and abnormal working condition disposal;

[0088] Rule type knowledge: is the explicit threshold and logical judgment, such as torque exceeding a certain critical value, it is determined that there is an over-collision fault;

[0089] Process type knowledge: is a more fuzzy and semi-explicit judgment process, such as detecting a part clamp force anomaly, it needs to check the servo motor torque first, and then make a comprehensive judgment combining with the welding current;

[0090] Abnormal working condition disposal: refers to abnormal working condition examples and their disposal methods, based on the fault log and repair report of specific production events;

[0091] ​In order to fully utilize these knowledge in the algorithm system, an expert experience library is established, and the database can be stored in the dimension of variable-threshold-rule priority-fault type, which provides support for subsequent data processing and model inference.

[0092] Based on the above technical solution, S102, for the multi-source sensing data of the servo robot, including vibration, temperature, current voltage, force and torque, dozens to hundreds of measurement variables need to be processed. In this process, a part of the variables has limited effect on fault detection or even brings noise. Through the expert experience library, the key variable list is refined, including explicit variable exclusion, threshold directional screening, expert-defined threshold, statistical model and expert rule fusion;

[0093] Explicit variable exclusion is that some sensors change very little under normal working conditions of the robot, and have no help for the identification of multiple faults in actual historical analysis. Therefore, such variables are excluded and given a lower weight through expert experience marking

[0094] Threshold directional screening is to include relevant variables in the key monitoring list when the temperature exceeds the threshold and the force feedback exceeds a certain safety range, and to give higher weight in subsequent model training and prediction;

[0095] In actual implementation, it can be expressed in the following form:

[0096] Ω={x i ∣ExpertRule(x i )=TRUE};

[0097] Wherein, Ω represents the variable set retained after screening by the expert experience library, ExpertRule represents a series of screening logic or threshold judgment functions based on expert experience, which can be a Boolean output or a series of scoring functions corresponding to different importance levels;

[0098] In the factory environment, noise data and sensor fault data can cause abnormal outliers. By setting the detection rule of the abnormal value through expert experience, the abnormal value is removed and corrected;

[0099] Expert-defined threshold: for sensors and fault types with known high risk, if the measured value violates the artificially set equipment limit indicator, it is directly determined as dirty data;

[0100] Statistical model and expert rule fusion: use the 3σ principle based on historical data distribution, including mean μ and standard deviation σ, combined with the upper and lower limit correction coefficients defined by experts, to perform multi-layer filtering on the data;

[0101] It can be formalized as follows: if a certain sensing variable x i satisfies:

[0102] |x i -μ i |>k·σ i ;

[0103] where, μ i and σ i are the mean and standard deviation of the variable respectively, k is the correction coefficient set by the expert, when the above formula is met, some expert logic is also met, such as the key variable jumps in a short time, then further verify the sensor state, determine it as an abnormal value and eliminate it, otherwise do other processing, such as interpolation correction.

[0104] Based on the above technical solution, S2, when the data on the factory site has a certain amount of annotation, such as fault type classification label, a supervised learning model capable of distinguishing common multiple fault types is constructed. Traditional Support Vector Machine is good at processing high-dimensional data and has strong generalization ability, but in the scene of small amount of labeled data, redundant variables and large difference of fault samples, pure SVM often appears overfitting and unstable model problems;

[0105] On the SVM multi-classification strategy, the variable selection mechanism of Group lasso is fused, and the truncated loss function with anti-outlier ability is fused. Specifically, it can be divided into three key points of model objective function, variable grouping and regularization term, and truncated loss design;

[0106] Let the training data set be where x i ∈R d represents the feature vector of the i-th sample, containing multi-dimensional sensor data in time domain and frequency domain, y i ∈{1,2,…,C} is the fault class label, including the normal class;

[0107] One-vs-Rest or One-vs-One strategy can be extended to multi-class SVM. Using the One-vs-Rest scheme, C groups of decision functions w c ,b c are needed for each classification c. The loss is defined as follows:

[0108]

[0109] where, is the truncated loss function, R({w c}) is the regularization term, and λ is the regularization coefficient. The truncated loss function is shown in Figure 2 .

[0110] Based on the above technical scheme, S2, in actual production, the measurement variable can be naturally divided into different groups Group, divided by sensor type or by physical part, the Group lasso is introduced to make some groups be selected and removed as a whole, which is beneficial to enhance the interpretability and dimensionality reduction effect of the model;

[0111] If d features are divided into G groups, set by expert experience, record the variable index set of the gth group as I g Then the weight w c of each classifier can be defined as:

[0112]

[0113] In the multi-classification scenario, the same grouping sparsity is shared by all classifiers, and then the grouping constraint can be performed on all w c , and the following regularization term is obtained:

[0114]

[0115] Where, alpha g is the weight coefficient of each group, and the weight coefficient is set by experts according to prior knowledge. When alpha g is large, the gth group of variables is more likely to be compressed or even set to zero. If alpha g is small, the group is more likely to be retained.

[0116] By controlling different alpha g , Group lasso with expert information can be realized, and the integration of data-driven and expert experience is considered.

[0117] Based on the above technical scheme, S2, in actual scenarios, some samples of the servo robot have large noise and extreme outliers. The truncated loss is introduced, and when the misclassification degree exceeds a certain threshold, the loss value is no longer increased, so as to prevent the excessive influence of extreme samples on the overall decision. Taking the binary classification scenario as an example, the truncated Hinge loss can be defined as:

[0118] l trunc (y i ,f(x i ))=min(L max ,max(0,1-y i ·f(x i )));

[0119] Where, L max is the truncation threshold. When the misclassification severity exceeds a certain range, the loss is fixed at a constant, which not only preserves the idea of interval maximization of Hinge loss, but also prevents the adverse effects of a small number of extreme samples.

[0120] Based on the above technical scheme, S3, in the servo robot health detection task, another common problem is: fault samples, i.e. positive samples are extremely rare, or new fault types appear, which are beyond the existing annotation range. In this case, it is difficult to accurately detect new types of faults purely relying on supervised learning models. A variational autoencoder Variational Autoencoder is introduced to learn the multi-source data distribution of the robot in the normal state, so as to identify abnormal data. The structure of VAE is as shown in Figure 3 ;

[0121] VAE includes an encoder Encoder and a decoder Decoder. For an input sample x, the encoder maps it to a latent space z, and the decoder generates a reconstructed output from z The goal of VAE is to minimize the reconstruction error and follow the prior distribution in the latent space.

[0122] The prior distribution of the latent variable z is defined as: The encoder outputs the conditional distribution q φ (z| x), and the decoder outputs p θ (x| z). The training objective of VAE can be represented as maximizing the evidence lower bound:

[0123]

[0124] Where D KL is the Kullback-Leibler divergence, which is used to constrain the distribution output by the encoder not to deviate too much from the Gaussian prior. Through training on a large amount of normal working condition data, the encoder learned by VAE can map the distribution of normal data to a compact latent space, and the decoder can also reconstruct the original data with small error.

[0125] Based on the above technical scheme, S3, after obtaining the trained VAE, the new sensor data x (new) is input into the model, and its reconstruction error can be calculated as an anomaly score. Specifically:

[0126]

[0127] Where, is the reconstruction output of VAE for input x (new) If the error is significantly greater than the statistical threshold learned on normal data, it can be considered that the data corresponds to an abnormal state, i.e. a potential fault. In addition, expert-set rules are introduced for secondary judgment, and higher weights are applied to the reconstruction error of the corresponding variables. Specifically:

[0128]

[0129] where ω is the weight of the variable j The determination of the importance of the variable from the expert experience library can also partially utilize the expert experience in the unsupervised detection stage, thereby improving the accuracy of detection.

[0130] Based on the above technical solution, S4, the expert experience system, the truncated multi-class SVM with Group lasso, and the VAE model are organically combined to construct a multi-mode health degree detection and fault identification framework, and the overall process can be divided into the following steps:

[0131] a, data acquisition and preliminary cleaning;

[0132] b, expert experience driven deep preprocessing;

[0133] c, model training and inference, supervised SVM, unsupervised VAE;

[0134] d, result fusion and decision output;

[0135] In addition, a fault handling suggestion based on a large language model is adopted, historical treatment methods are formed into formatted text, and Lora fine-tuning is performed on the basis of llama3. When the model outputs a fault, the fault handling suggestion is directly outputted. The performance of the model will be enhanced with the increase of historical data.

[0136] Based on the above technical solution, S4, a configurable multi-level fusion strategy is provided to integrate the results of supervised and unsupervised determination and reflect expert experience, which is as follows:

[0137] If the SVM has high certainty and is consistent with the expert rules, the fault type is directly outputted, such as torque anomaly, and the corresponding treatment suggestion is given;

[0138] If the SVM has low confidence and the VAE analysis shows that the abnormal score is extremely high, it is determined as a possible new fault type, which is submitted to artificial or expert system for further inspection, and the sample and its fault label can be included in subsequent training. The SVM confidence is determined by the margin of the decision function distance from the classification boundary;

[0139] If the SVM shows normal or low risk, but the VAE shows that the abnormal score is high, combined with the semi-explicit rules of expert experience on specific sensor data, when the rules are triggered, it is considered as a suspicious state and the detection frequency and alarm triggering are increased;

[0140] If the VAE determines normal and the SVM determines fault, it belongs to model overfitting and improper design of expert system threshold, and the model parameters need to be updated according to subsequent labeling, and the expert threshold needs to be corrected.

[0141] Finally, it should be noted that the above only describes the preferred examples of the present application, and is not intended to limit the present application. Although the present application has been described in detail with reference to the foregoing embodiments, those skilled in the art will appreciate that modifications can be made to the technical solutions described in the foregoing embodiments, or some of the technical features thereof can be replaced equivalently, without departing from the spirit and principle of the present application. Any modifications, equivalent replacements, improvements, etc. made within the spirit and principle of the present application shall be included in the protection scope of the present application.

Claims

1. A novel servo robot health detection algorithm, characterized in that: Includes the following steps: S1. Multi-level data processing and assisted inference based on expert experience; S2, a robust multi-class SVM model that integrates Group lasso variable selection and truncation loss design; S3. Unsupervised anomaly detection based on variational autoencoder (VAE); S4. Construct a multi-mode health detection and fault identification framework through organic integration; S1 includes the following steps: The construction of the expert experience base, as described in S101; S102, data preprocessing and variable selection.

2. The novel servo robot health detection algorithm according to claim 1, characterized in that: In the S101 section, in automobile manufacturing plants, the health diagnosis of servo robots often involves a large number of technicians with extensive experience in troubleshooting. This experience is typically reflected in: rule-based knowledge, process-based knowledge, and handling of abnormal operating conditions. Regular Knowledge: consists of clearly defined thresholds and logical judgments; Process type Knowledge: is a relatively vague and semi-explicit judgment process; Abnormal operating condition handling: refers to examples of abnormal operating conditions and their handling methods, based on fault logs and maintenance reports of specific production events; To fully utilize this knowledge within the algorithm system and establish an expert experience base, the database can be structured and stored according to the dimensions of variables, thresholds, rule priorities, and associated fault types, providing support for subsequent data processing and model inference.

3. The novel servo robot health detection algorithm according to claim 2, characterized in that: In step S102, for the multi-source sensor data of the servo robot, including vibration, temperature, current and voltage, force and torque, dozens to hundreds of measurement variables need to be processed. In this process, some variables have limited effect on fault detection or even introduce noise. Through the expert experience base, a list of key variables is extracted, including explicit variable exclusion, threshold-oriented screening, expert-defined thresholds, and fusion of statistical models and expert rules. Explicitly excluded variables are those whose sensors show minimal variation under normal robot operating conditions and are of no help in identifying various faults in actual historical analysis. These variables are then excluded and weighted less based on expert experience. Threshold-oriented screening involves adding relevant variables to a key monitoring list when the temperature exceeds a threshold or the force feedback exceeds a certain safe range, and then assigning them higher weights during subsequent model training and prediction. In practical implementation, it can be formally expressed as follows: Ω={x i ∣ExpertRule(x i )=TRUE}; Where Ω represents the set of variables retained after being filtered by the expert experience base, and ExpertRule represents a series of filtering logic or threshold judgment functions based on expert experience. It can be a Boolean output or a series of scoring functions, corresponding to different importance levels. In a factory environment, noise data and sensor failure data can lead to outliers. Outlier detection rules are set through expert experience, and outliers are removed and corrected. Expert-defined thresholds: For sensors and fault types known to be high-risk, if the measured value violates the manually set equipment limit indicators, it will be directly judged as dirty data; Fusion of statistical models and expert rules: Utilizing the 3σ principle based on historical data distribution, combined with expert-defined upper and lower limit correction coefficients, the data is filtered in multiple layers; This can be formalized as follows: If there is a certain sensing variable x i satisfy: |x i -m i |>k·s i ; Where, μ i and σ i These are the mean and standard deviation of the variable, respectively, and k is the correction coefficient set by the experts. When the above formula is satisfied, it also satisfies certain expert logic, and is judged as an outlier and removed.

4. The novel servo robot health detection algorithm according to claim 1, characterized in that: S2, when the data at the factory site has a certain amount of annotation, construct a supervised learning model that can distinguish between a variety of common fault types. Traditional SVM is good at processing high-dimensional data and has strong generalization ability, but in scenarios with a small amount of labeled data, a lot of redundant variables and large differences in fault samples, simple SVM often suffers from overfitting and model instability. In the SVM multi-class strategy, the variable selection mechanism of Group lasso and the cutoff loss function with the ability to resist outliers are integrated. Specifically, it can be divided into three key points: model objective function, variable grouping and regularization term, and cutoff loss design. Let the training dataset be Where, x i ∈R d Let y represent the feature vector of the i-th sample. i ∈{1,2,…,C} represents the fault category label; This can be extended to multi-class SVM using a one-to-many or one-to-one strategy. Using a one-to-many approach requires learning the C-group decision function w. c ,b c For each category c, the loss is defined as follows: in, For the truncation loss function, R({w c }) represents the regularization term, and λ is the regularization coefficient.

5. A novel servo robot health detection algorithm according to claim 4, characterized in that: In actual production, measured variables may naturally be divided into different groups, either by sensor type or by physical location. Introducing Grouplasso allows certain groups to be selected and eliminated as a whole, which helps to enhance the interpretability and dimensionality reduction effect of the model. If we divide the d features into G groups, determined by expert experience, and denote the variable index set of the g-th group as I... g Then the weight w for each classifier c The group norm can be defined as: In multi-class classification scenarios, sharing the same grouping sparsity across all classifiers allows for the classification of all w classes. c After applying grouping constraints, the following regularization term is obtained: Where, α g The weighting coefficients for each group are set by experts based on prior knowledge, when α g When α is large, the variables in the g-th group are more easily compressed or even set to zero. g If the size is smaller, the group is more likely to remain. By controlling different α g It enables Grouplasso with expert information, balancing data-driven approaches with expert experience.

6. A novel servo robot health detection algorithm according to claim 4, characterized in that: In S2, in real-world scenarios, some samples from the servo robot contain significant noise and extreme outliers. Therefore, a truncated loss is introduced. When the misclassification exceeds a certain threshold, the loss value is no longer increased to prevent extreme samples from excessively influencing the overall decision. Taking a binary classification scenario as an example, the truncated Hinge loss can be defined as: Among them, L max To truncate the threshold, the loss is fixed at a constant when the severity of misclassification exceeds a certain range. This retains the idea of ​​maximizing the interval of Hinge loss while preventing the adverse effects of a few extreme samples.

7. A novel servo robot health detection algorithm according to claim 1, characterized in that: In the S3 section, in the servo robot health detection task, relying solely on the supervised learning model makes it difficult to accurately detect new types of faults. Therefore, a variational autoencoder (VAE) is introduced to learn the multi-source data distribution of the robot under normal conditions, thereby identifying abnormal data. The VAE consists of two parts: an encoder and a decoder. For an input sample x, the encoder maps it to the latent space z, and the decoder then generates a reconstructed output from z. The goal of VAE is to minimize the reconstruction error and follow the prior distribution in the latent space; Define the prior distribution of the latent variable z as: Encoder output conditional distribution q φ (z∣x), decoder output p θ (x|z), the training objective of VAE can be expressed as maximizing the lower bound of evidence ELBO: Among them, D KL Kullback-Leibler divergence is used to constrain the distribution of the encoder output to not deviate too much from the Gaussian prior. Through training on a large amount of normal operating data, the encoder learned by the VAE can map the distribution of normal data to a compact latent space, and the decoder can also reconstruct the original data with a small error.

8. A novel servo robot health detection algorithm according to claim 7, characterized in that: In step S3, after obtaining the trained VAE, new sensing data x (new) Input the model, and its reconstruction error can be calculated as an anomaly score. Specifically: in, For VAE input x (new) If the reconstructed output error is significantly greater than the statistical threshold learned from normal data, then the data can be considered to correspond to an abnormal state. Furthermore, expert-defined rules are introduced for a secondary judgment, assigning higher weight to the reconstruction error of the corresponding variables. Specifically: Wherein, weight ω j The determination of the importance of this variable from the expert experience base can partially utilize expert experience to improve the accuracy of the test, even in the unsupervised testing phase.

9. A novel servo robot health detection algorithm according to claim 1, characterized in that: S4 organically combines the expert experience system, the truncated multi-class SVM with Group lasso, and the VAE model to construct a multi-modal health detection and fault identification framework. The overall process can be divided into the following steps: a) Data collection and preliminary cleaning; b. Expert-driven deep preprocessing; c. Model training and inference: supervised SVM, unsupervised VAE; d, Results fusion and decision output; In addition, fault handling suggestions based on a large language model are adopted, and the historical handling methods are formed into formatted text. Based on llama3, LoRa is fine-tuned. When the model outputs a fault, the fault handling suggestions are directly output. The model's performance will improve as the amount of historical data increases.

10. A novel servo robot health detection algorithm according to claim 9, characterized in that: S4, in order to integrate the results of supervised and unsupervised judgments and reflect expert experience, provides a configurable multi-level fusion strategy, as follows: If the SVM has high determinism and is consistent with expert rules, it will directly output the fault type and provide corresponding handling suggestions. If the SVM confidence is low and the VAE analysis shows an extremely high anomaly score, it is determined to be a possible new fault type. It is submitted to a human or expert system for further verification, and the sample and its fault label can be included in subsequent training. The SVM confidence is determined by the margin of the decision function from the classification boundary. If the SVM shows normal or low risk, but the VAE shows an abnormal score that is too high, combined with the semi-explicit rules for specific sensor data based on expert experience, when the rules are triggered, it is considered a suspicious state and the detection frequency is increased and an alarm is triggered. The VAE is judged to be normal while the SVM is judged to be faulty. This is due to model overfitting and improper expert system threshold design. The model parameters need to be updated based on subsequent annotations, and the expert threshold needs to be corrected.

Images