Multi-component mechanical system competition risk health assessment method and system based on GMM-LSTM-DeepHit

Abstract

The invention provides a multi-component mechanical system competition risk health assessment method and system based on GMM-LSTM-DeepHit, and the method comprises the steps: carrying out the unsupervised clustering analysis of a two-dimensional feature time sequence matrix sample set through employing a Gaussian mixture model GMM, and obtaining a three-dimensional feature time sequence matrix sample set; an LSTM-DeepHit model is constructed, and the survival probability of the sample is obtained; and based on the sample survival probability, calculating a risk score mean value of the whole life cycle of the component, dividing a failure risk level of the component, converting the risk score mean value into a component service life prediction value, and combining the risk contribution degree of each feature to complete health assessment of the competition risk of the multi-component mechanical system. According to the method, automatic identification of potential risk events, accurate capture of time sequence degradation characteristics and quantitative modeling of risk conduction relations among components can be realized, and finally, the accuracy and engineering practicability of health assessment of a complex mechanical system are improved.

Description

Technical Field

[0001] This invention belongs to the field of mechanical system health monitoring and fault prediction technology, specifically relating to a method and system for assessing the competitive risk health of multi-component mechanical systems based on GMM-LSTM-DeepHit. Background Technology

[0002] In the field of mechanical engineering, complex mechanical systems such as wind turbine gearboxes, aero engines, and industrial machine tools are composed of multiple core components working together. The failure modes of each component (such as bearing wear, gear tooth breakage, and shaft misalignment) exhibit a "competitive risk" characteristic—the failure of any component not only directly affects the system's operating state but may also alter the failure path of other components through physical effects such as load transmission and vibration coupling. For example, abnormal wear of the bearing inner ring can lead to increased vibration of the gearbox input shaft, thereby accelerating fatigue damage to the gear tooth surface. Existing health assessment methods have significant limitations: First, risk events rely on manual definition, making it difficult to cover complex failures (such as "bearing wear + insufficient lubrication") or unknown failure modes in actual operation, resulting in subjectivity and limitations in risk modeling; second, the ability to capture temporal characteristics is insufficient, as sensor signals of mechanical systems (such as vibration and temperature) have significant time dependence, and existing models struggle to effectively extract risk characteristics in long-term degradation trends; third, the correlation between components is not explicitly modeled, often assuming that the risks of each component are independent, making it impossible to quantify the transmission impact of a component's failure on the risks of other components, ultimately leading to insufficient accuracy in health assessments and difficulty in supporting precise predictive maintenance decisions. Therefore, there is an urgent need for a method and system for assessing the competitive risk health of multi-component mechanical systems that can automatically identify potential risk events, effectively capture temporal characteristics, and model component relationships. Summary of the Invention

[0003] To address the problems in existing mechanical system health assessments, such as "risk events relying on manual definition, insufficient capture of temporal characteristics, and lack of component correlation modeling," this invention provides a method and system for competitive risk health assessment of multi-component mechanical systems based on GMM-LSTM-DeepHit. The aim is to achieve automatic identification of potential risk events, accurate capture of temporal degradation characteristics, and quantitative modeling of risk transmission relationships between components, ultimately improving the accuracy and engineering practicality of health assessments for complex mechanical systems.

[0004] To achieve the above objectives, the present invention provides the following solution:

[0005] A competitive risk health assessment method for multi-component mechanical systems based on GMM-LSTM-DeepHit includes:

[0006] The system uses sensors to collect operating status data of each core component of the mechanical system, obtains multi-source operating status data sequences, performs correlation analysis on the multi-source operating status data sequences, and constructs a two-dimensional feature time series matrix sample set.

[0007] Unsupervised clustering analysis was performed on the two-dimensional feature time series matrix sample set using a Gaussian mixture model (GMM) to obtain a three-dimensional feature time series matrix sample set.

[0008] By combining the temporal feature extraction subnetwork LSTM with the survival network model DeepHit, an LSTM-DeepHit model is constructed. The three-dimensional feature temporal matrix sample set is then input into the LSTM-DeepHit model to obtain the sample survival probability. The sample survival probability represents the probability that the sample will experience the k-th type of failure risk event in different time intervals.

[0009] Based on the survival probability of the samples, the mean risk score of the entire life cycle of the components is calculated, the equipment failure risk level is divided, and the mean risk score is converted into the predicted value of the component's service life. Combining the risk contribution of each feature, a health assessment of the competitive risk of multi-component mechanical systems is completed.

[0010] Preferably, the method for constructing a two-dimensional feature time series matrix sample set includes:

[0011] Correlation analysis is performed on the multi-source operating status data sequence, and F data sequences that meet the preset discrimination index are selected from the multi-source operating status data sequence based on the calculated value of Pearson correlation coefficient.

[0012] Z-score normalization is used to normalize F data sequences that meet the preset discrimination criteria, and the sliding window method is used to convert the normalized F data sequences that meet the preset discrimination criteria into a two-dimensional feature time series matrix sample set.

[0013] Preferably, the method for obtaining a sample set of three-dimensional feature time series matrices includes:

[0014] The Gaussian Mixture Model (GMM) is used to perform unsupervised clustering analysis on the sample set of the two-dimensional feature time series matrix. The optimal number of mixed categories k is determined according to the Akaike Information Criterion and the Bayesian Information Criterion. Each sample is assigned a pseudo-label, and each pseudo-label corresponds to a failure risk event.

[0015] The two-dimensional feature time series matrix sample set is concatenated with the pseudo-label to form a new three-dimensional feature time series matrix sample set, which is then divided into a training set and a test dataset.

[0016] Preferably, the structure of the LSTM-DeepHit model includes:

[0017] The shared subnetwork, consisting of a single-layer LSTM network with residual connections and a multilayer perceptron, is used to extract global temporal degradation features from a three-dimensional feature temporal matrix sample set.

[0018] The event-specific subnetwork, consisting of a risk LSTM layer and a multilayer perceptron for capturing specific temporal patterns of k types of failure risk events, is used to learn the cumulative occurrence rate of each type of failure risk event.

[0019] Preferably, the joint loss function of the LSTM-DeepHit model includes ranking loss, event classification cross-entropy loss, and L2 regularization term;

[0020] The ranking loss is used to penalize sample pairs whose predicted risk ranking does not match the actual survival time ranking.

[0021] The event classification cross-entropy loss is used to optimize the model's accuracy in identifying the specific event type that ultimately leads to system failure among multiple competing risk events.

[0022] The L2 regularization term is used to suppress overfitting by penalizing the model parameters.

[0023] This invention also provides a competitive risk health assessment system for multi-component mechanical systems based on GMM-LSTM-DeepHit, used to implement the method, comprising:

[0024] The sample set construction module is used to collect the operating status data of each core component of the mechanical system using sensors, obtain multi-source operating status data sequences, perform correlation analysis on the multi-source operating status data sequences, and construct a two-dimensional feature time series matrix sample set.

[0025] The unsupervised clustering analysis module is used to perform unsupervised clustering analysis on the two-dimensional feature time series matrix sample set using a Gaussian mixture model (GMM) to obtain a three-dimensional feature time series matrix sample set.

[0026] The survival probability acquisition module is used to combine the temporal feature extraction subnetwork LSTM and the survival network model DeepHit model to construct an LSTM-DeepHit model, and input the three-dimensional feature temporal matrix sample set into the LSTM-DeepHit model to obtain the sample survival probability; wherein, the sample survival probability represents the probability that the sample will experience the k-th type of failure risk event in different time intervals;

[0027] The lifespan prediction module is used to calculate the average risk score of the entire life cycle of the component based on the survival probability of the sample, classify the equipment failure risk level, convert the average risk score into the predicted lifespan value of the component, and combine the risk contribution of each feature to complete the health assessment of the competitive risk of the multi-component mechanical system.

[0028] Preferably, the unsupervised clustering analysis module includes:

[0029] The pseudo-label generation unit is used to perform unsupervised clustering analysis on the two-dimensional feature time-series matrix sample set using Gaussian Mixture Model (GMM), determine the optimal number of mixed categories k based on the Akaike Information Criterion and Bayesian Information Criterion, assign pseudo-labels to each sample, and each pseudo-label corresponds to a failure risk event.

[0030] The 3D sample construction unit is used to concatenate the 2D feature time series matrix sample set with the pseudo-label to form a new 3D feature time series matrix sample set, and divide it into training set and test dataset.

[0031] Preferably, in the survival probability acquisition module, the structure of the LSTM-DeepHit model includes:

[0032] The shared subnetwork, consisting of a single-layer LSTM network with residual connections and a multilayer perceptron, is used to extract global temporal degradation features from a three-dimensional feature temporal matrix sample set.

[0033] The event-specific subnetwork, consisting of a risk LSTM layer and a multilayer perceptron for capturing specific temporal patterns of k types of failure risk events, is used to learn the cumulative occurrence rate of each type of failure risk event.

[0034] Compared with existing technologies, the beneficial effects of this invention are as follows: By improving the output layer design of the DeepHit model and combining Gaussian Mixture Model (GMM) event recognition with LSTM temporal feature capture, more accurate risk stratification is achieved—the low, medium, and high risk groups form significant intervals on the Kaplan-Meier survival curves, and the Log-rank test shows that the survival differences among the groups are statistically significant (p < 0.05); in remaining useful life (RUL) prediction, the improved model's RMSE is significantly lower than that of the basic DeepHit model, especially for high-risk equipment, the RUL prediction error can be controlled within 5%, and the prediction accuracy is significantly improved; SurvSHAP feature importance analysis further enhances the model's interpretability, revealing the dynamic impact of key features at different time points, providing data support for failure mechanism analysis; from an engineering application perspective, the risk level and maintenance recommendations output by the model can directly meet industrial needs, prioritizing maintenance for high-risk groups can reduce unplanned downtime losses, and extending the maintenance cycle and maintenance costs for low-risk groups, combining theoretical innovation with practical application value. Attached Figure Description

[0035] To more clearly illustrate the technical solution of the present invention, the drawings used in the embodiments are briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0036] Figure 1 This is a flowchart of a method for assessing the competitive risk health of a multi-component mechanical system based on GMM-LSTM-DeepHit, according to an embodiment of the present invention. Detailed Implementation

[0037] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0038] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0039] Example 1:

[0040] like Figure 1 As shown, a competitive risk health assessment method for multi-component mechanical systems based on GMM-LSTM-DeepHit includes:

[0041] S1: Utilizing sensors to collect operational status data from the core components of the mechanical system, multi-source operational status data sequences are obtained. Correlation analysis is then performed on these sequences to construct a two-dimensional feature time-series matrix sample set. The relationship between multi-component failures and risk competition is explained as follows: Sometimes, even if one component fails, the equipment can still function normally; however, if another component fails, the equipment will cease to function properly even in the early stages of damage. Data-driven fault identification requires a large amount of fault data. Due to the aforementioned situation, the "fault data" is imbalanced, and models trained directly using this data will exhibit a significant "bias." Risk competition analysis aims to learn and represent this "bias."

[0042] A further implementation method involves constructing a two-dimensional feature time-series matrix sample set, including:

[0043] Correlation analysis is performed on the multi-source operational status data sequences, and F data sequences that meet the preset discrimination index are selected from the multi-source operational status data sequences based on the calculated Pearson correlation coefficient as the main feature references;

[0044] Z-score normalization is used to normalize F data sequences that meet the preset discrimination criteria, eliminating differences in the dimensions of the sensor data. Then, a sliding window method is used to convert the normalized F data sequences that meet the preset discrimination criteria into a two-dimensional feature time-series matrix sample set. Specifically, the sliding window method is used to convert the F sets of one-dimensional feature long sequences contained in each sample into two-dimensional time-series matrices. , where N is the total number of samples; T is the number of segments, which divides the total running time (and corresponding sensor data) of the samples into T segments. The value of T needs to be selected based on the statistical characteristics of each sensor data and the total number of samples. It is required to retain enough truncation to show the gradual and continuous degradation trend of the component, while avoiding insufficient sample size due to an excessively large window or cumbersome subsequent calculations and overfitting due to an excessively small window; F represents the number of channels in the sensor data sequence;

[0045] S2: Unsupervised clustering analysis is performed on the two-dimensional feature time series matrix sample set using a Gaussian Mixture Model (GMM) to obtain a three-dimensional feature time series matrix sample set. A further implementation method for obtaining the three-dimensional feature time series matrix sample set includes:

[0046] Gaussian Mixture Model (GMM) is used to perform unsupervised clustering analysis on the two-dimensional feature time series matrix sample set. The optimal number of mixture categories k is determined based on the Akaike Information Criterion and the Bayesian Information Criterion. Each sample is assigned a pseudo-label, and each pseudo-label corresponds to a failure risk event. The two-dimensional feature time series matrix sample set is concatenated with the pseudo-label to form a new three-dimensional feature time series matrix sample set, which is then divided into a training set and a test dataset.

[0047] In the “feature time series matrix” described in this invention, the meaning of “feature” is equivalent to “main or core”, that is: “F main features selected from the multi-source operating status data sequence based on Pearson correlation calculation”. These “main or core features” may be equipment operating status signals (different equipment will show different core feature data), such as time series signals such as current, voltage, temperature, and pressure, or they may be equipment operating condition signals, such as height, surface force, etc.

[0048] Specifically, a Gaussian Mixture Module (GMM) is used to fit the distribution pattern of the F-channel sensor data matrix samples, assuming... The probability density function is derived from Composed of Gaussian weighted components:

[0049]

[0050] in, (F is the feature number) represents the sensor feature vector of the mechanical system; For the parameter set of the probability model; For the first The weights of the Gaussian components satisfy the following conditions: For the first There are Gaussian components, which follow an M-dimensional Gaussian distribution. It is the mean vector. The covariance matrix describes the correlation between features;

[0051] GMM parameters The estimation is performed iteratively using the Expectation-Maximization (EM) algorithm, which consists of two alternating steps:

[0052] E-step (expectation): Given the current model parameters, calculate the "membership" of each sample to each Gaussian component. That is, sample Belongs to the The posterior probability of each Gaussian component:

[0053]

[0054] Indicates sample For components The membership degree of each degree is summed up to the number of degrees of membership. ;

[0055] M-step (maximization): Based on the membership degrees calculated in the E-step, update the model parameters to maximize the likelihood function. The formula for updating the new parameters is as follows:

[0056]

[0057]

[0058] The EM algorithm repeats the E-step and M-step until the log-likelihood function converges, thereby obtaining stable Gaussian components and their distribution statistics.

[0059] The optimal method is selected using the Akaike Information Criterion (AIC) and the Bayesian Information Criterion (BIC). value:

[0060]

[0061]

[0062] in, The value of the converged log-likelihood function; This represents the total number of model parameters. Number of samples. Optimal. The value is the one that minimizes both AIC and BIC.

[0063] GMM achieves automatic identification of strong correlations between sample risk events through data-driven clustering: based on the optimal number of mixed categories k of the determined sample set and the clustering analysis results, pseudo-labels (label=1,2,3,…k) are assigned to the feature time series matrix samples described in claim 1; each label corresponds to a pseudo failure risk event (each pseudo failure event does not necessarily correspond one-to-one with actual physical failures, such as bearing wear, gear tooth breakage, or compound mechanical failures, etc.), and the original data samples are spliced ​​together with pseudo-labels to form a new three-dimensional feature time series matrix, which serves as training and testing data samples;

[0064] S3: Combining the temporal feature extraction sub-network LSTM with the survival network model DeepHit, an LSTM-DeepHit model is constructed. The sample set of three-dimensional feature temporal matrix is ​​input into the LSTM-DeepHit model to obtain the sample survival probability; where the sample survival probability represents the probability of a sample experiencing a type k failure risk event in different time intervals. The temporal feature extraction network LSTM (Long Short-Term Memory) and the survival network DeepHit are fused to construct the LSTM-DeepHit network. This network adopts a two-branch structure: a shared layer extracts general degradation patterns, and multiple branch layers obtain event-specific risks.

[0065] For example, a device (system) consists of multiple interconnected components such as a fan and a compressor. The failure modes of these components (such as fan blade breakage, bearing wear, and broken gear teeth) interact through load transmission and vibration coupling, creating "competitive risks"—the initial failure of some components can lead to system failure and prevent other potential failures from occurring. For instance, abnormal vibration caused by wear on the inner ring of a bearing may accelerate gear tooth fatigue, while broken gear teeth can mask further bearing degradation.

[0066] In the example above, "fan blade breakage" and "bearing wear" are both (failure) risk events; while "bearing wear (hereinafter referred to as A)" and "gear tooth breakage (hereinafter referred to as B)" are competing risks.

[0067] When equipment fails, it's often impossible to definitively say whether the cause is A or B. Machine learning algorithms typically identify whether samples contain features of A or B and provide maintenance suggestions based on past data and work experience. Specifically, in this process, because A might be "more competitive" than B, the training data might not include diverse samples of "A and B occurring" or "B occurring, but A not occurring." Without a sufficiently balanced sample set, the algorithm is prone to misjudgment when faced with test samples it hasn't learned from or reasoned about. The solution described in this invention aims to: analyze and indicate that A might be "more competitive" than B, thereby providing a more reasonable and interpretable judgment on the fault state corresponding to any combination of A and B.

[0068] A further implementation method is that the structure of the LSTM-DeepHit model includes:

[0069] The shared subnetwork, consisting of a single-layer LSTM network with residual connections and a multilayer perceptron (MLP), is used to extract global temporal degradation features from a 3D feature temporal matrix sample set. Specifically, the shared subnetwork's global temporal degradation feature extraction is composed of a single-layer LSTM network with residual connections and a multilayer perceptron (MLP), with an input dimension of... The hidden layer has 64 units, forming a shared LSTM layer whose input follows the format of "number of samples - number of time steps - number of features". This layer processes data step by step. The output is a hidden variable. To capture the temporal degradation patterns common to all components (samples); to extract the variables generated within the last time step of the LSTM output. As the feature extraction output, it condenses the degradation information of the entire time window, avoiding the impact of redundant time steps on model efficiency; at the same time, residual connections are introduced to alleviate feature decay in deep networks. in This is the output of the LSTM at the last time step. To input the features of the last time step, This network ultimately shares features. Its core value lies in capturing long-term temporal dependencies through LSTM and preserving early weak degradation signals through residual connections, providing fundamental features for subsequent event-specific modeling.

[0070] The event-specific subnetwork, consisting of a risk LSTM layer and a multilayer perceptron for capturing specific temporal patterns of k types of failure risk events, is used to learn the cumulative occurrence rate of each type of failure risk event. Specifically, it is designed for GMM identification. For degenerate events, construct k parallel event-specific learning subnetworks ( Figure 1Each subnetwork independently learns the cumulative incidence rate (CIF) of a certain type of event, achieving explicit quantification of multi-event competition. Each subnetwork is configured to capture the first... The system consists of risk LSTM and MLP layers representing event-specific temporal patterns, ensuring that shared features are aligned with the input dimensions of the sub-networks. It also ensures the association between "global degradation" and "event-specific risk" through risk residual connections.

[0071] The outputs of the k event subnetworks are aggregated through an integration function. Its core function is to integrate the predictions of the k independent subnetworks into a unified tensor based on the "event type" dimension, ultimately yielding the model output. ,in Indicates "the The sample at the th The first time interval occurred within the first time interval "Probability of a class of events", that is, the probability of the first class of events Cumulative Incidence Rate (CIF) of Class Events.

[0072] A further implementation method is that the joint loss function of the LSTM-DeepHit model includes ranking loss. Event classification cross-entropy loss and L2 regularization term To ensure both accuracy and temporal consistency in competitive risk analysis, the joint loss function mainly includes the following three components:

[0073]

[0074] Ranking loss is used to penalize sample pairs whose predicted risk ranking does not match the actual survival time ranking; its core idea is that if one device fails earlier than another, then the model's predicted risk score for the former should be higher than that for the latter. This ensures that the risk score output by the model can reflect the physical law that "risk increases monotonically over time", which is the key to achieving accurate RUL prediction.

[0075]

[0076] The event classification cross-entropy loss is used to optimize the model's accuracy in identifying the specific event type that ultimately leads to system failure among multiple competing risk events. This loss function ensures that the model can accurately predict which competing event is most likely to cause system failure. This is crucial for achieving "fault diagnosis" and is related to... Together, they constitute a joint prediction of "when it will fail" and "why it will fail".

[0077]

[0078] The L2 regularization term is used to suppress overfitting by penalizing the model parameters, thus enhancing the model's ability to generalize to new data.

[0079]

[0080] , As a weighting factor, its value should be chosen to strike a balance between "accuracy of risk ranking" and "accuracy of event classification" in the model, while effectively suppressing overfitting. For example, taking... =0.2、 =0.1.

[0081] The joint optimization strategy, through multi-objective optimization, can ensure a certain balance between "risk ranking", "event classification" and "model generalization" in the model, which is the key to achieving accurate competitive risk assessment.

[0082] S4: Based on the sample survival probability, calculate the mean risk score of the entire life cycle of the component, classify the component failure risk level, convert the mean risk score into the predicted value of the component's service life, and combine the risk contribution of each feature to complete the health assessment of the competitive risk of the multi-component mechanical system.

[0083] Methods for risk stratification (classifying component failure risk levels) include:

[0084] The survival probabilities of components output by the LSTM-DeepHit network model are a matrix. Elements in the matrix Representing the The components in the first The probability of survival within a given time interval. The cumulative occurrence rate of each competing risk event is converted into a risk score:

[0085]

[0086] The average risk score of each device is obtained by calculating the average risk score across all time intervals throughout its entire lifecycle. :

[0087]

[0088] Based on the average risk score of components Risk levels are determined by the distribution characteristics, with risk scores divided into three intervals based on percentiles, corresponding to three risk levels: low risk group, ... These components exhibit low overall degradation, strong survivability, and long remaining lifespan; the medium-risk group... These components are in a state of mild to moderate degradation, requiring enhanced monitoring but no urgent maintenance; the high-risk group, namely... These components are prone to severe degradation, have a high risk of failure, and a short remaining lifespan, so they should be prioritized for maintenance.

[0089] The effectiveness of risk stratification was validated using a combination of Kaplan-Meier (KM) survival curves and the Log-rank test. Specifically, the KM survival curves estimated the survival probability of components at different time points using a nonparametric method.

[0090] Assuming the sample data involves and contains Each piece of equipment has an actual effective lifespan of [number] days. The group of devices in time The survival probability at KM is estimated as follows:

[0091]

[0092] in For the first Equipment failure time of the platform =1 indicates that the device has failed. For time The number of devices still in operation.

[0093] The intuitive differences in the KM curves were further quantified and statistically represented using the Log-rank test. The sample sizes for the high-risk group (Group A) and the low-risk group (Group B) were respectively... and The two groups are in the time interval The number of failed devices within are respectively and The number of exposed devices (the number of devices still operating within the interval) are respectively and Then the Log-rank test statistic is:

[0094]

[0095] in For group A in the first... The expected number of failures within a time interval The variance is given. This statistic follows a set of variances with 1 degree of freedom. Distribution, if the test result This indicates that the difference between the two survival curves is statistically significant.

[0096] Methods for obtaining effective useful life prediction (i.e., life prediction value) include:

[0097] The average risk score is converted into the predicted effective service life of the equipment:

[0098]

[0099] Among them, PVUL (Predicted Valid Useful Life) is the predicted effective lifespan of the equipment (the recommended replacement time for components). Representing the The equipment in the first Survival probability within a time interval This represents the survival probability threshold determined through further statistical analysis based on the actual lifespan time series of the device (training dataset) and the corresponding survival probability sequence (LSTM-DeepHit network output). In other words, for a device or component... When it predicts the probability of survival First time falling below the threshold When this point in time is considered the upper limit of the equipment's effective lifespan (PVUL), it indicates that maintenance or replacement is required; if If the value does not fall below the threshold, it means that the device is still usable.

[0100] This embodiment also provides a method for obtaining the contribution of risk characteristics, the specific steps of which include:

[0101] Feature contribution value, that is, for a feature Its Shapley value in this prediction scenario is defined as:

[0102]

[0103] in, For the LSTM-DeepHit network model, the first The equipment in the first Under the category of events, the first CIF forecast values ​​for each time interval, Not included Any subset of features; To use only a subset of features Model for Predicted values; coefficients For feature subset The weights are determined to ensure that all subsets are considered fairly.

[0104] Shapley value The physical meaning is: characteristics Regarding "the first The equipment in the first The first time interval occurred within the first time interval The marginal contribution of the probability of a class of events — This indicates that as the eigenvalue increases, the probability of the event occurring increases (accelerates degradation). This indicates that as the eigenvalue increases, the probability of the event occurring decreases (suppressing degradation); the larger the absolute value, the more significant the influence of the feature on the prediction result.

[0105] The component health assessment results include component risk stratification level, component effective lifetime prediction, and the dynamic impact of key characteristics in different time intervals (the characteristic contribution of failure, i.e., Shapley value).

[0106] Example 2

[0107] This invention also provides a multi-component mechanical system competition risk health assessment system based on GMM-LSTM-DeepHit, for implementing the method of Embodiment 1, including:

[0108] The sample set construction module is used to collect the operating status data of each core component of the mechanical system using sensors, obtain multi-source operating status data sequences, perform correlation analysis on the multi-source operating status data sequences, and construct a two-dimensional feature time series matrix sample set.

[0109] The unsupervised clustering analysis module is used to perform unsupervised clustering analysis on a two-dimensional feature time series matrix sample set using a Gaussian mixture model (GMM) to obtain a three-dimensional feature time series matrix sample set.

[0110] The survival probability acquisition module is used to combine the temporal feature extraction subnetwork LSTM and the survival network model DeepHit model to construct the LSTM-DeepHit model, and input the three-dimensional feature temporal matrix sample set into the LSTM-DeepHit model to obtain the sample survival probability; where the sample survival probability represents the probability that the sample will experience the k-th type of failure risk event in different time intervals.

[0111] The lifespan prediction module is used to calculate the average risk score of the entire life cycle of the components based on the sample survival probability, classify the equipment failure risk level, convert the average risk score into the predicted value of the component lifespan, and combine the risk contribution of each feature to complete the health assessment of the competitive risk of multi-component mechanical systems.

[0112] A further implementation method includes an unsupervised clustering analysis module comprising:

[0113] The pseudo-label generation unit is used to perform unsupervised clustering analysis on the sample set of two-dimensional feature time matrix using Gaussian mixture model (GMM). It determines the optimal number of mixed categories k based on the Akaike information criterion and Bayesian information criterion, and assigns a pseudo-label to each sample. Each pseudo-label corresponds to a failure risk event.

[0114] The 3D sample construction unit is used to concatenate the 2D feature time series matrix sample set with the pseudo-label to form a new 3D feature time series matrix sample set, and divide it into training set and test dataset.

[0115] A further implementation method involves the following: In the survival probability acquisition module, the structure of the LSTM-DeepHit model includes:

[0116] The shared subnetwork, consisting of a single-layer LSTM network with residual connections and a multilayer perceptron, is used to extract global temporal degradation features from a three-dimensional feature temporal matrix sample set.

[0117] The event-specific subnetwork, consisting of a risk LSTM layer and a multilayer perceptron for capturing specific temporal patterns of k types of failure risk events, is used to learn the cumulative occurrence rate of each type of failure risk event.

[0118] The embodiments described above are merely preferred embodiments of the present invention and are not intended to limit the scope of the present invention. Various modifications and improvements made to the technical solutions of the present invention by those skilled in the art without departing from the spirit of the present invention should fall within the protection scope defined by the claims of the present invention.

Claims

1. A method for competitive risk health assessment of multi-component mechanical systems based on GMM-LSTM-DeepHit, characterized in that, include: The system uses sensors to collect operating status data of each core component of the mechanical system, obtains multi-source operating status data sequences, performs correlation analysis on the multi-source operating status data sequences, and constructs a two-dimensional feature time series matrix sample set. Unsupervised clustering analysis was performed on the two-dimensional feature time series matrix sample set using a Gaussian mixture model (GMM) to obtain a three-dimensional feature time series matrix sample set. By combining the temporal feature extraction subnetwork LSTM with the survival network model DeepHit, an LSTM-DeepHit model is constructed. The three-dimensional feature temporal matrix sample set is then input into the LSTM-DeepHit model to obtain the sample survival probability. The sample survival probability represents the probability that the sample will experience the k-th type of failure risk event in different time intervals. Based on the survival probability of the samples, the mean risk score of the entire life cycle of the components is calculated, the component failure risk level is divided, and the mean risk score is converted into the predicted value of the component's service life. Combining the risk contribution of each feature, a health assessment of the competitive risk of multi-component mechanical systems is completed.

2. The method according to claim 1, characterized in that, Methods for constructing a two-dimensional feature time series matrix sample set include: Correlation analysis is performed on the multi-source operating status data sequence, and F data sequences that meet the preset discrimination index are selected from the multi-source operating status data sequence based on the calculated value of Pearson correlation coefficient. Z-score normalization is used to normalize F data sequences that meet the preset discrimination criteria, and the sliding window method is used to convert the normalized F data sequences that meet the preset discrimination criteria into a two-dimensional feature time series matrix sample set.

3. The method according to claim 1, characterized in that, Methods for obtaining a sample set of three-dimensional feature time series matrices include: The Gaussian Mixture Model (GMM) is used to perform unsupervised clustering analysis on the sample set of the two-dimensional feature time series matrix. The optimal number of mixed categories k is determined according to the Akaike Information Criterion and the Bayesian Information Criterion. Each sample is assigned a pseudo-label, and each pseudo-label corresponds to a failure risk event. The two-dimensional feature time series matrix sample set is concatenated with the pseudo-label to form a new three-dimensional feature time series matrix sample set, which is then divided into a training set and a test dataset.

4. The method according to claim 3, characterized in that, The structure of the LSTM-DeepHit model includes: The shared subnetwork, consisting of a single-layer LSTM network with residual connections and a multilayer perceptron, is used to extract global temporal degradation features from a three-dimensional feature temporal matrix sample set. The event-specific subnetwork, consisting of a risk LSTM layer and a multilayer perceptron for capturing specific temporal patterns of k types of failure risk events, is used to learn the cumulative occurrence rate of each type of failure risk event.

5. The method according to claim 4, characterized in that, The joint loss function of the LSTM-DeepHit model includes ranking loss, event classification cross-entropy loss, and L2 regularization term; The ranking loss is used to penalize sample pairs whose predicted risk ranking does not match the actual survival time ranking. The event classification cross-entropy loss is used to optimize the model's accuracy in identifying the specific event type that ultimately leads to system failure among multiple competing risk events. The L2 regularization term is used to suppress overfitting by penalizing the model parameters.

6. A competitive risk health assessment system for multi-component mechanical systems based on GMM-LSTM-DeepHit, used to implement the method described in any one of claims 1-5, characterized in that, include: The sample set construction module is used to collect the operating status data of each core component of the mechanical system using sensors, obtain multi-source operating status data sequences, perform correlation analysis on the multi-source operating status data sequences, and construct a two-dimensional feature time series matrix sample set. The unsupervised clustering analysis module is used to perform unsupervised clustering analysis on the two-dimensional feature time series matrix sample set using a Gaussian mixture model (GMM) to obtain a three-dimensional feature time series matrix sample set. The survival probability acquisition module is used to combine the temporal feature extraction subnetwork LSTM and the survival network model DeepHit model to construct an LSTM-DeepHit model, and input the three-dimensional feature temporal matrix sample set into the LSTM-DeepHit model to obtain the sample survival probability; wherein, the sample survival probability represents the probability that the sample will experience the k-th type of failure risk event in different time intervals; The lifespan prediction module is used to calculate the average risk score of the entire life cycle of the component based on the survival probability of the sample, classify the equipment failure risk level, convert the average risk score into the predicted lifespan value of the component, and combine the risk contribution of each feature to complete the health assessment of the competitive risk of the multi-component mechanical system.

7. The system according to claim 6, characterized in that, The unsupervised clustering analysis module includes: The pseudo-label generation unit is used to perform unsupervised clustering analysis on the two-dimensional feature time-series matrix sample set using Gaussian Mixture Model (GMM), determine the optimal number of mixed categories k based on the Akaike Information Criterion and Bayesian Information Criterion, assign pseudo-labels to each sample, and each pseudo-label corresponds to a failure risk event. The 3D sample construction unit is used to concatenate the 2D feature time series matrix sample set with the pseudo-label to form a new 3D feature time series matrix sample set, and divide it into training set and test dataset.

8. The system according to claim 7, characterized in that, In the survival probability acquisition module, the structure of the LSTM-DeepHit model includes: The shared subnetwork, consisting of a single-layer LSTM network with residual connections and a multilayer perceptron, is used to extract global temporal degradation features from a three-dimensional feature temporal matrix sample set. The event-specific subnetwork, consisting of a risk LSTM layer and a multilayer perceptron for capturing specific temporal patterns of k types of failure risk events, is used to learn the cumulative occurrence rate of each type of failure risk event.

Images