Rotor bearing reliability analysis method based on deep attention and belief fusion

Abstract

The application discloses a rotor bearing reliability analysis method based on deep attention and belief fusion, comprising the following steps: collecting multi-sensor time series based on a sensor group installed on a helicopter rotor, constructing a state monitoring data set, and preprocessing samples in the state monitoring data set; constructing and training a deep belief fusion analysis model, including a parallel deep feature extraction branch, a belief rule reasoning branch and an evidence fusion module; in the application stage, real-time multi-sensor time series are acquired, and after being divided by a sliding window and preprocessed, the real-time multi-sensor time series are input into the trained deep belief fusion analysis model to obtain a health state evaluation value; meanwhile, when preset threshold conditions and confidence conditions are met, an alarm is given, and online monitoring of rotor bearing reliability is realized. The method has high prediction performance of deep learning, and also has interpretability and credibility, and can be widely applied to health management of aero-engines, rotor systems and unmanned aerial vehicle power transmission components.

Description

Technical Field

[0001] This invention belongs to the field of health management and reliability analysis of aerospace equipment, and in particular relates to a reliability assessment method for helicopter propeller bearings that combines deep temporal attention and belief rule base-evidential reasoning (BRB-ER). Background Technology

[0002] Helicopter rotor systems are critical power components of aircraft. Their main bearings are subjected to complex loads and high-frequency vibration environments over long periods, including various stress sources such as centrifugal force, aerodynamic loads, torque impacts, and temperature fluctuations. Failure of these bearings directly affects flight safety and mission reliability; therefore, monitoring the health status and predicting the remaining life of rotor bearings is of great significance.

[0003] Traditional reliability analysis of aerospace bearings largely relies on statistical or empirical models, such as fatigue life curves, oil analysis, and spectral diagnostics. However, these methods suffer from limitations such as feature extraction dependence on experience and difficulty in handling nonlinear degradation patterns. In recent years, deep learning (especially CNN, LSTM, and Transformer) has demonstrated superior performance in rotating machinery health monitoring, automatically learning temporal features to achieve higher accuracy in life prediction. However, its insufficient interpretability and confidence level limit its direct application in the field of aerospace safety.

[0004] On the other hand, belief rule bases and evidence-based reasoning methods have advantages in uncertainty modeling and expert knowledge fusion, and can express reliability distributions in the form of probabilistic beliefs. However, traditional BRB methods have static parameters and manually defined rules, making it difficult to adaptively fuse with dynamic time-series data.

[0005] Therefore, there is an urgent need for an intelligent reliability analysis method that has both deep temporal feature modeling capabilities and belief fusion and confidence assessment mechanisms, in order to achieve high-precision, interpretable and reliable health status assessment of helicopter propeller bearings. Summary of the Invention

[0006] The purpose of this invention is to propose a rotor bearing reliability analysis method based on the fusion of deep attention and belief, so as to realize intelligent analysis, reliability prediction and confidence output of multi-source state signals of propeller bearings.

[0007] To achieve the above objectives, the present invention employs the following technical solution: A rotor bearing reliability analysis method based on the fusion of deep attention and belief includes: A condition monitoring dataset is constructed by collecting multi-sensor time series data based on a sensor array installed on the helicopter rotor, and the samples in the condition monitoring dataset are preprocessed. A deep belief fusion analysis model is constructed and trained, comprising parallel deep feature extraction branches, belief rule inference branches, and evidence fusion modules. The deep feature extraction branch takes a normalized sample as input and outputs the probability distribution of the sample belonging to each health state. The rule base inference branch takes a normalized key statistical feature vector of the sample as input and outputs the belief degree distribution of the sample belonging to each health state. The evidence fusion module performs decision fusion on the outputs of the deep feature extraction branch and the belief rule inference branch to obtain the comprehensive health score corresponding to the sample. In the application phase, the time series data from multiple sensors are acquired in real time, and after sliding window partitioning and preprocessing, they are input into a trained deep belief fusion analysis model to obtain a health status assessment value. At the same time, an alarm is triggered when preset threshold conditions and confidence conditions are met, thereby realizing online monitoring of the reliability of the rotor bearing.

[0008] Furthermore, the multi-sensor time series is divided into multiple segments using a sliding window, with each segment serving as a sample, and the health status of the helicopter rotor bearing corresponding to the segment serving as the label of the sample; The multi-sensor time series of each sample is subjected to max-min normalization to obtain the normalized sample; key statistical features are extracted for each sample to construct a key statistical feature vector; the key statistical features include vibration kurtosis, vibration root mean square value and lubricating oil temperature; For each key statistical feature of all samples, a max-min normalization process is performed to obtain a normalized key statistical feature vector.

[0009] Furthermore, the deep feature extraction branch includes an embedding layer, a positional encoding layer, a self-attention layer, and an output layer, wherein: The embedding layer is used to perform linear mapping on the normalized samples to obtain high-dimensional feature representations; the embedding layer is a fully connected layer. The positional coding layer is used to perform temporal labeling on the high-dimensional feature representation to obtain a feature matrix with temporal position information; the positional coding layer uses sine and cosine functions to generate absolute position codes. The self-attention layer is used to perform multi-head scaling dot product attention calculation on the feature matrix to obtain context features containing degenerate dependencies; the self-attention layer is followed by a Dropout layer, which is used to randomly discard some neuron outputs during the training phase; The output layer is used to perform nonlinear transformation and normalization on the context features output by the self-attention layer to obtain the probability distribution of the samples output by the deep feature extraction branch belonging to each health state. The output layer specifically includes a feedforward network and a Softmax normalization function. The output of the feedforward network is followed by a Dropout layer. The deep feature vectors processed by the feedforward network activation and Dropout layer are extracted and used as the input of the deep collaboration mechanism part in the belief rule base inference branch.

[0010] Furthermore, the construction process of the reasoning branches of the belief rule base includes: First, define the prerequisite attributes for the rules: select the key statistical features contained in the normalized key statistical feature vector of the samples as the attributes for rule input; Secondly, define attribute reference values: based on the helicopter transmission system design manual and expert experience, set semantic reference levels for each attribute; Redefine the rule consequent: Set the health status assessment result set of the bearing, corresponding to the three health statuses: {normal status, early degradation, and severe failure}. Finally, a rule table is generated: combining the reference levels of all attributes forms a series of "IF-THEN" rules, thus forming a belief rule base; where the first... Rules The form is expressed as follows: ; in," " is the logical "AND" operator. Indicates the first In rule number 1, Reference level for each attribute; These correspond to impact characteristics, vibration energy characteristics, and environmental thermal stress characteristics, respectively. For the first The attribute weight of each attribute; For the first The rule determines the state to be healthy. The confidence level of the belief These correspond to normal state, early degradation, and serious failure, respectively. Indicates the first The rule weight of each rule .

[0011] Furthermore, the construction process of the reasoning branches of the belief rule base also includes: For the key statistical features contained in the normalized key statistical feature vector of the sample, the matching degree between them and the center of each reference value in the rule is calculated using the Gaussian membership function. Rule activation is calculated based on matching degree, and the first rule is calculated. Activation weight of the rule : ; in, , Indicates the first , No. The rule weight of each rule For the first The attribute weight of each attribute; The number of rules in the belief rule base. The total number of attributes; , For the current sample The attribute for the first Rule No. The degree of matching of the rules; Deep collaboration mechanism: The deep feature vector output by the deep feature extraction branch is passed through a fully connected layer to generate a correction factor, which is added to the belief confidence of the belief rule base inference branch to calibrate the belief confidence. Finally, the calibrated belief confidence scores for each rule are calculated according to their corresponding activation weights. The data is aggregated to generate a distribution of belief levels for each health status within the health status assessment set. .

[0012] Furthermore, the input to the evidence fusion module is the output of the sample after passing through the deep feature extraction branch and the belief rule reasoning branch; The evidence fusion calculation process of the evidence fusion module is as follows: The Dempster-Shafer evidence theory is used to fuse two independent sources of evidence: the deep feature extraction branch and the belief rule reasoning branch. and The propositional variables representing different health states, derived from the deep feature extraction branch and the belief rule inference branch, are denoted as the health state assessment result set. Any subset of; These correspond to normal condition, early degradation, and severe failure, respectively. ; First, calculate the conflict coefficient between the two types of propositions. : ; in The deep feature extraction branch supports the proposition. The probability distribution, Indicate the supporting proposition of the reasoning branch of the belief rule. The distribution of belief levels; Subsequently, the consistency evidence was orthogonally normalized and fused based on the conflict coefficient to obtain the final distribution of health status beliefs. : ; Combining health status belief distribution Compared with the preset level utility value Calculate the final comprehensive health score. The formula is: .

[0013] Furthermore, the state monitoring dataset is input into the deep belief fusion analysis model, and end-to-end training is performed using a cost-sensitive joint loss function; the loss function of the deep belief fusion analysis model includes MSE mean squared error, KL divergence, and cost term.

[0014] Furthermore, in practical applications, the Dropout layer in the deep feature extraction branch is forced to remain enabled; the multi-sensor time series to be tested is taken, and after sliding window partitioning and preprocessing, the sequence within one of the sliding windows is selected. This allows it to continuously undergo deep belief fusion analysis model. Second forward propagation; obtained Slightly different comprehensive health score sequences among groups ;in For the first The overall health score obtained from the first forward propagation; Calculate the predicted mean The formula for the final output health status assessment value is: ; in, For the first The overall health score is obtained from the forward propagation.

[0015] Furthermore, the prediction variance is calculated. Used to characterize the deep belief fusion analysis model for Uncertainty in judgment: ; in This is the model accuracy term, representing the inherent noise of the data; Finally, based on the preset confidence level, and combined with the prediction variance and prediction mean, the confidence interval of the prediction result is determined; an alarm is issued only when both of the following conditions are met simultaneously: The first is the threshold condition, which is that the predicted mean is lower than the preset health threshold. Second, the confidence condition, that is, the lower confidence limit of the confidence interval is lower than the preset risk threshold.

[0016] A terminal device includes a processor, a memory, and a computer program stored in the memory; when the processor executes the computer program, it implements the rotor bearing reliability analysis method based on deep attention and belief fusion.

[0017] A computer-readable storage medium storing a computer program; when executed by a processor, the computer program implements the rotor bearing reliability analysis method based on deep attention and belief fusion.

[0018] Compared with the prior art, the present invention has the following technical features: This invention not only achieves accurate modeling of helicopter rotor bearing degradation, but also outputs interpretable belief distributions and confidence levels. This method combines the high predictive performance of deep learning with the interpretability and reliability of BRB models, and can be widely applied to the health management of aero-engines, rotor systems, and UAV power transmission components. Attached Figure Description

[0019] Figure 1 This is a fusion structure diagram of the deep belief fusion analysis model of the present invention; Figure 2 This is a block diagram of the evidence fusion mechanism of the present invention; Figure 3 This is a structural diagram of the cost-sensitive optimization and alarm strategy of the present invention. Detailed Implementation

[0020] This invention proposes a reliability analysis method for helicopter propeller bearings based on deep attention and belief fusion. Its core lies in constructing a data processing flow that integrates deep feature self-learning and expert belief rule reasoning. For example... Figures 1 to 3 As shown, the steps of the present invention are as follows: Step 1: Construct a helicopter propeller bearing condition monitoring dataset and perform preprocessing.

[0021] Step 1.1, Data Acquisition and Sample Definition.

[0022] This step aims to construct a bearing lifecycle dataset encompassing multiple operating conditions. First, assuming the helicopter rotor bearing's safety condition is known, multi-sensor time series data are collected using sensor arrays installed on the helicopter rotor shaft and transmission components. This includes vibration acceleration signals, stator current signals, and lubricating oil temperature signals. Second, the length of the sliding window is set. and moving step size The multi-sensor time series is divided into multiple consecutive segments, with each segment serving as a sample. Finally, the health status of the helicopter propeller bearing within the corresponding sliding window is used as the label for the sample. The health status can be, for example, 1-normal state, 2-early degradation, 3-serious malfunction, etc.; thus, a health status is obtained from... A state monitoring dataset consists of 3 labeled samples; each sample contains a set of samples of length 1. Multi-sensor time series.

[0023] Step 1.2: Preprocess the samples in the state monitoring dataset to generate the input forms required for the deep feature extraction branch and belief rule reasoning branch in the deep belief fusion analysis model.

[0024] (1) For each sample The multi-sensor time series data were subjected to max-min normalization to obtain the normalized samples. As input to the deep feature extraction branch.

[0025] (2) For each sample Develop key statistical feature vectors Extraction; in this embodiment Vibration kurtosis was chosen to characterize the impact features. The root mean square (RMS) value of vibration was chosen to characterize the energy features. Lubricating oil temperature was selected to characterize environmental thermal stress; then, max-min normalization was performed on each key statistical feature of all samples to obtain a normalized key statistical feature vector. As input to the deep belief fusion analysis model.

[0026] Step 2: Construct and train a deep belief fusion analysis model, which includes a parallel deep feature extraction branch, a belief rule reasoning branch, and an evidence fusion module.

[0027] Step 2.1: Construct a deep feature extraction branch; the input to this branch is the normalized sample. It aims to automatically capture long-range temporal degradation features; the output is the probability distribution of the sample belonging to each health state. The deep feature extraction branch includes an embedding layer, a positional encoding layer, a self-attention layer, and an output layer, where: The embedding layer is used to perform linear mapping on the normalized samples to obtain high-dimensional feature representations. In this scheme, the embedding layer specifically uses a fully connected layer (Linear Projection) to map low-dimensional sensor signals to a latent space of the model dimension, thereby converting physical signals into continuous vector representations.

[0028] The positional encoding layer is used to represent the high-dimensional features of the embedding layer output. Perform time-series labeling to obtain a feature matrix with time location information. In this scheme, the position encoding layer uses sine and cosine functions to generate absolute position codes, which are then superimposed on the high-dimensional feature representation. The above is used to obtain the characteristic matrix. .

[0029] A multi-head self-attention layer is used to process the feature matrix output by the positional encoding layer. Perform multi-head scaled dot product attention computation to obtain contextual features that include degenerate dependencies. In this scheme, the self-attention layer adopts a multi-head scaling dot product attention mechanism, which first processes the feature matrix... Mapped to query matrix Key matrix Sum matrix And calculate the attention weights according to the following formula: ; in The dimension of the key vector, superscript This indicates transposition; this multi-head structure can monitor both short-term impacts and long-term wear trends in bearing signals in parallel.

[0030] A Dropout layer follows the self-attention layer. During the training phase, it is used to randomly discard some neuron outputs to prevent overfitting.

[0031] The output layer is used to process the contextual features output by the self-attention layer. After performing nonlinear transformation and normalization, the probability distribution of the deep feature extraction branch output is obtained. The output layer specifically includes a feedforward network (FFN) and a Softmax normalization function; the probability distribution The sample contains the first The probability of a certain health state is calculated; additionally, a Dropout layer is applied after the output of the feedforward network to further enhance the regularization effect. Simultaneously, the deep feature vector processed by the feedforward network (FFN) activation and Dropout layer is... It is extracted and used as input for the deep collaboration mechanism part in the reasoning branch of the belief rule base.

[0032] Step 2.2: Construct the reasoning branch of the belief rule base; the input of this branch is the normalized key statistical feature vector of the sample. The output is the distribution of belief levels for each health state. .

[0033] The belief rule base reasoning branch and the aforementioned deep feature extraction branch are parallel and collaborative. The deep feature extraction branch focuses on uncovering implicit temporal correlations, while the belief rule base reasoning branch focuses on logical reasoning based on interpretable features. The two are connected through a subsequent evidence fusion module. The specific construction of the belief rule base reasoning branch consists of two sub-steps: initial rule base construction and parameter adaptive optimization. (1) Construction of the initial belief rule base based on expert knowledge.

[0034] This step aims to describe the nonlinear mapping relationship between a specific set of bearing condition characteristics and health levels.

[0035] First, define the prerequisite attributes for the rule: select the normalized key statistical feature vectors of the samples. The three key statistical features included are used as attributes for rule input, namely... The kurtosis of the vibration signal characterizes the impact features. The root mean square (RMS) value is used to characterize the vibrational energy characteristics, and is selected as follows: The temperature of the lubricating oil is used to characterize the environmental thermal stress.

[0036] Secondly, define attribute reference values: based on the helicopter transmission system design manual and expert experience, set semantic reference levels for each attribute; for example, for vibration energy characteristics, set three reference points based on the root mean square (RMS) value: {low (L), medium (M), high (H)}, and... Set as anchor point.

[0037] Redefining the rule consequent: Setting the health status assessment result set for the bearing These correspond to the three health states: {normal state, early degradation, and severe failure}.

[0038] Finally, a rule table is generated: combining the reference levels of all attributes forms a series of "IF-THEN" rules, thus forming a belief rule base; where the first... Rules The form is expressed as follows: ; in," " is the logical "AND" operator. Indicates the first In rule number 1, Reference level for each attribute; These correspond to impact characteristics, vibration energy characteristics, and environmental thermal stress characteristics, respectively. For the first The attribute weights of each attribute are initially given by experts (expert experience); For the first The rule determines the state to be healthy. The confidence level of the belief These correspond to normal state, early degradation, and serious failure, respectively. The initial value is given by experts; Indicates the first The rule weight of a rule represents its relative importance or reliability within the entire belief rule base; a higher weight indicates that the rule carries greater weight in the final fusion reasoning. ; The initial value is set by experts based on experience; it is a trainable parameter that is automatically corrected during training.

[0039] (2) Confidence transformation of samples based on Gaussian membership.

[0040] This step transforms the input numerical samples into rule-processable belief levels.

[0041] For the normalized key statistical feature vector of the sample It includes three key statistical features (i.e., attributes). The Gaussian membership function is used to calculate its relationship with each reference center in the rule. Matching degree : ; in, For the first set by experts Rules The Middle The center value of the reference level for each attribute; It is a natural exponential function; The width parameter is set by experts to control the shape of the Gaussian membership function.

[0042] (3) Rule activation and parameter adaptive optimization.

[0043] To address the issue of static parameters, this approach utilizes a state monitoring dataset to train the belief rule base, enabling it to dynamically adapt to the actual degradation patterns of propeller bearings.

[0044] First, perform rule activation calculation, calculating the... Activation weight of the rule This weight represents the degree to which the current sample triggers the rule, and the formula is: ; in, , Indicates the first , No. The rule weight of each rule For the first The attribute weight of each attribute; The number of rules in the belief rule base. The total number of attributes, in this embodiment ; , For the current sample The attribute for the first Rule No. The degree of matching of the rules.

[0045] Subsequently, data-driven parameter adjustments are performed; although the initial values ​​of the aforementioned parameters are given by experts, these parameters (rule weights) Attribute weight Belief confidence level All of these are trainable parameters. In the model training phase of step 3, the model is trained using samples with health status labels in the state monitoring dataset through backpropagation to minimize the prediction error loss function, thereby automatically fine-tuning these parameters. The physical meaning is that if a rule frequently causes misjudgments in the samples of the state monitoring dataset, the model will automatically reduce the rule weight of that rule. This corrects the biases in expert knowledge.

[0046] (4) Deep Collaboration Mechanism: In order to enhance the dynamic adaptability of BRB, the deep feature vector output by the output layer of the deep feature extraction branch is used to... A correction factor is generated by passing through a fully connected layer. The belief confidence is added to the reasoning branch of the belief rule base (weighted fusion, such as linear interpolation). Above, thus affecting the confidence level of beliefs. Calibration is performed; this allows the belief rule base to not only rely on pre-defined logic, but also to self-calibrate using implicit features discovered by deep networks.

[0047] (5) Output section; Finally, using the evidence reasoning algorithm, the calibrated belief confidence level corresponding to each rule is calculated. According to its corresponding activation weight The data is aggregated to generate a distribution of belief levels for the three health states within the health status assessment result set. .

[0048] Step 2.3 Construct the evidence fusion module.

[0049] The Evidence Fusion Layer is located at the end of the deep belief fusion analysis model. It is responsible for receiving real-time calculation results from the deep feature extraction branch and the belief rule reasoning branch, and performing decision-level consistency fusion.

[0050] (1) Definition and source of fusion parameters.

[0051] The input to the evidence fusion module is the output of the sample after passing through the deep feature extraction branch and the belief rule reasoning branch.

[0052] (2) Evidence fusion calculation process.

[0053] This approach utilizes the Dempster-Shafer evidence theory to fuse the two independent sources of evidence mentioned above; in the Dempster-Shafer evidence theory, and These represent propositional variables for different health states derived from the deep feature extraction branch and the belief rule inference branch; they can be the health state assessment result set. Any subset of; The deep feature extraction branch supports the proposition. The probability distribution, Indicate the supporting proposition of the reasoning branch of the belief rule. The distribution of belief levels; when the proposition is a single state of health, there are , .

[0054] Among them, propositions and It can be a single proposition or a compound proposition. A single proposition refers to... Compound propositions, for example, are , ;at this time , ; Similarly.

[0055] First, calculate the conflict coefficient between the two types of propositions. This is used to measure the degree of inconsistency between the deep belief fusion analysis model and expert rule judgments, and its calculation formula is as follows: ; Subsequently, the consistency evidence was orthogonally normalized and fused based on the conflict coefficient to obtain the final distribution of health status beliefs. The calculation formula is as follows: ; (3) Calculation of comprehensive health score.

[0056] To transform the probability distribution into an intuitive single-value indicator, a pre-defined level utility value is set for each health state. Finally, combining the distribution of beliefs about health status... Compared with the preset level utility value Calculate the final comprehensive health score. The formula is: ; Among them, the level utility value A preset constant is used to determine the impact of health status on the operational safety and economy of the rotor system; for example, it can be set within the range of [0,1]. In this embodiment, it can be set as follows: In practical applications, specific values ​​can be adjusted based on the knowledge of domain experts.

[0057] Step 3, Model Training and Parameter Optimization The state monitoring dataset is input into the deep belief fusion analysis model, and end-to-end training is performed using a cost-sensitive joint loss function; firstly, the loss function is constructed. This function includes the MSE prediction error, the KL divergence (used to constrain the consistency of the two branch distributions), and a cost term. Its calculation formula is as follows: ; in, Indicates overall health score Mean squared error between the true value and the actual value; the actual value refers to the quantitative score of the health status {normal status, early degradation, severe failure} in the label. For example, the actual value for normal status is set to 1, early degradation is set to 0.5, and severe failure is set to 0, etc. These are hyperparameters used to balance the weights of various losses. They can be set empirically at the beginning, for example, to 1. They can be adjusted during the training phase to optimize the model's performance on the preset validation set. express Divergence; This represents the cost term, which imposes different penalties for different types of misclassification. It is The preset matrix, its first... Line number Column elements This represents the true state of health. Mistakenly diagnosed as healthy The penalty coefficient; this matrix is ​​set based on expert knowledge, for example, imposing a higher penalty on missed faults than on false faults.

[0058] During the optimization process, the Adam optimizer is used to synchronously update all trainable parameters in the deep belief fusion analysis model, thereby achieving closed-loop optimization of data-driven knowledge optimization and knowledge-constrained data features.

[0059] Step 4: Online inference and reliability alerts. In the practical application stage after the deep belief fusion analysis model has been trained, in order to solve the problem that deep learning models can usually only provide point estimates but cannot evaluate their own prediction credibility, this invention utilizes Monte Carlo Dropout (MC-Dropout) technology. The specific implementation process for the random sampling is as follows: (1) Monte Carlo random sampling mechanism.

[0060] This mechanism is applied to the deep feature extraction branch. The specific execution logic is as follows: During the inference phase, the Dropout layer in the deep feature extraction branch is forcibly kept enabled; the multi-sensor time series to be tested is acquired, and after sliding window partitioning and preprocessing, the sequence within one of the sliding windows is... (Its format is consistent with the sample, and its length is...) This allows it to continuously undergo deep belief fusion analysis through the model. The forward propagation is repeated; due to the randomness of Dropout, the deep belief fusion analysis model structure is slightly different for each forward propagation, thus yielding... Slightly different comprehensive health score sequences among groups ;in For the first The overall health score is obtained from the forward propagation.

[0061] (2) Calculation of distribution statistics parameters.

[0062] Based on the above Based on the sampling results, calculate the statistical properties of the predicted distribution; first, calculate the predicted mean. The final output health status assessment value (point estimate) is calculated using the following formula: ; in, For the first The overall health score is obtained from the forward propagation.

[0063] Next, calculate the prediction variance. Used to characterize the deep belief fusion analysis model for Uncertainty in the judgment; the larger the variance, the greater the discrepancy in the model's predictions under different Dropout modes, i.e., the lower the confidence level. Its calculation formula is: ; in This is the model accuracy term, representing the inherent noise of the data; it can be set to a small constant, such as 0.01 or 0.001, based on expert experience; or it can be used as a trainable parameter of the model during training.

[0064] (3) Confidence interval and dual alarm strategy.

[0065] First, based on the preset confidence level Calculate the confidence interval of the prediction results. The formula is: ; in, The upper side of the standard normal distribution quantiles; As the lower confidence limit, Upper confidence level; The significance level is (e.g., often taken as 0.05 or 0.01). An alarm will be issued only if both of the following conditions are met: The first is the threshold condition, namely the predicted mean. Below the preset health threshold (A potential malfunction may be present); secondly, the confidence condition, i.e., the lower confidence limit. It is also below the preset risk threshold. .

[0066] This strategy confirms that the model has a high degree of confidence in the fault judgment and eliminates misjudgments caused by model instability or data noise.

[0067] Among them, health threshold For example, by statistically analyzing the comprehensive health scores of samples in normal condition and early deterioration in the condition monitoring dataset, a quantile that allows the vast majority of samples to be correctly identified can be set; risk threshold. Then, the lower quantile is set according to the maximum acceptable risk level.

[0068] Therefore, in practical applications, by continuously acquiring multi-sensor time series data and using a sliding window for partitioning and preprocessing, the health status assessment value predicted for each sliding window can be obtained by following the above process. Simultaneously, an alarm is triggered when the above conditions are met, enabling online monitoring of the rotor bearing's reliability.

[0069] This invention can be widely applied to reliability assessment, predictive maintenance, and flight safety assurance of aero-engines and rotor systems, and has engineering value that combines high precision, interpretability, and reliability.

[0070] The above embodiments are only used to illustrate the technical solutions of this application, and are not intended to limit them. Although this application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of this application, and should all be included within the protection scope of this application.

Claims

1. A rotor bearing reliability analysis method based on the fusion of deep attention and belief, characterized in that, include: A condition monitoring dataset is constructed by collecting multi-sensor time series data based on a sensor array installed on the helicopter rotor, and the samples in the condition monitoring dataset are preprocessed. A deep belief fusion analysis model is constructed and trained, comprising parallel deep feature extraction branches, belief rule inference branches, and evidence fusion modules. The deep feature extraction branch takes a normalized sample as input and outputs the probability distribution of the sample belonging to each health state. The rule base inference branch takes a normalized key statistical feature vector of the sample as input and outputs the belief degree distribution of the sample belonging to each health state. The evidence fusion module performs decision fusion on the outputs of the deep feature extraction branch and the belief rule inference branch to obtain the comprehensive health score corresponding to the sample. In the application phase, the time series data from multiple sensors are acquired in real time, and after sliding window partitioning and preprocessing, they are input into a trained deep belief fusion analysis model to obtain a health status assessment value. At the same time, an alarm is triggered when preset threshold conditions and confidence conditions are met, thereby realizing online monitoring of the reliability of the rotor bearing.

2. The rotor bearing reliability analysis method based on deep attention and belief fusion as described in claim 1, characterized in that, The multi-sensor time series is divided into multiple segments using a sliding window, with each segment serving as a sample. The health status of the helicopter rotor bearing corresponding to the segment is used as the label for the sample. Max-min normalization is performed on the multi-sensor time series of each sample to obtain the normalized sample; For each sample, key statistical features are extracted to construct a key statistical feature vector; The key statistical features include vibration kurtosis, vibration root mean square value, and lubricating oil temperature; For each key statistical feature of all samples, a max-min normalization process is performed to obtain a normalized key statistical feature vector.

3. The rotor bearing reliability analysis method based on deep attention and belief fusion according to claim 1, characterized in that, The deep feature extraction branch includes an embedding layer, a positional encoding layer, a self-attention layer, and an output layer, wherein: The embedding layer is used to perform linear mapping on the normalized samples to obtain high-dimensional feature representations; the embedding layer is a fully connected layer. The positional coding layer is used to perform temporal labeling on the high-dimensional feature representation to obtain a feature matrix with temporal position information; the positional coding layer uses sine and cosine functions to generate absolute position codes. The self-attention layer is used to perform multi-head scaling dot product attention calculation on the feature matrix to obtain context features containing degenerate dependencies; the self-attention layer is followed by a Dropout layer, which is used to randomly discard some neuron outputs during the training phase; The output layer is used to perform nonlinear transformation and normalization on the context features output by the self-attention layer to obtain the probability distribution of the samples output by the deep feature extraction branch belonging to each health state. The output layer specifically includes a feedforward network and a Softmax normalization function. The output of the feedforward network is followed by a Dropout layer. The deep feature vectors processed by the feedforward network activation and Dropout layer are extracted and used as the input of the deep collaboration mechanism part in the belief rule base inference branch.

4. The rotor bearing reliability analysis method based on deep attention and belief fusion according to claim 1, characterized in that, The process of constructing reasoning branches in the belief rule base includes: First, define the prerequisite attributes for the rules: select the key statistical features contained in the normalized key statistical feature vector of the samples as the attributes for rule input; Secondly, define attribute reference values: based on the helicopter transmission system design manual and expert experience, set semantic reference levels for each attribute; Redefine the rule consequent: Set the health status assessment result set of the bearing, corresponding to the three health statuses: {normal status, early degradation, and severe failure}. Finally, a rule table is generated: combining the reference levels of all attributes forms a series of "IF-THEN" rules, thus forming a belief rule base; where the first... Rules The form is expressed as follows: ； in," " is the logical AND operator, Indicates the first In rule number 1, Reference level for each attribute; These correspond to impact characteristics, vibration energy characteristics, and environmental thermal stress characteristics, respectively. For the first The attribute weight of each attribute; For the first The rule determines the state to be healthy. The confidence level of the belief These correspond to normal state, early degradation, and serious failure, respectively. Indicates the first The rule weight of each rule .

5. The rotor bearing reliability analysis method based on deep attention and belief fusion according to claim 4, characterized in that, The process of constructing reasoning branches in the belief rule base also includes: For the key statistical features contained in the normalized key statistical feature vector of the sample, the matching degree between them and the center of each reference value in the rule is calculated using the Gaussian membership function. Rule activation is calculated based on matching degree, and the first rule is calculated. Activation weight of the rule : ； in, , Indicates the first , No. The rule weight of each rule For the first The attribute weight of each attribute; The number of rules in the belief rule base. The total number of attributes; , For the current sample The attribute for the first Rule No. The degree of matching of the rules; Deep collaboration mechanism: The deep feature vector output by the deep feature extraction branch is passed through a fully connected layer to generate a correction factor, which is added to the belief confidence of the belief rule base inference branch to calibrate the belief confidence. Finally, the calibrated belief confidence scores for each rule are calculated according to their corresponding activation weights. The data is aggregated to generate a distribution of belief levels for each health status within the health status assessment set. .

6. The rotor bearing reliability analysis method based on deep attention and belief fusion according to claim 1, characterized in that, The input to the evidence fusion module is the output of the sample after passing through the deep feature extraction branch and the belief rule reasoning branch; The evidence fusion calculation process of the evidence fusion module is as follows: The Dempster-Shafer evidence theory is used to fuse two independent sources of evidence: the deep feature extraction branch and the belief rule reasoning branch. and The propositional variables representing different health states, derived from the deep feature extraction branch and the belief rule inference branch, are denoted as the health state assessment result set. Any subset of; These correspond to normal condition, early degradation, and severe failure, respectively. ; First, calculate the conflict coefficient between the two types of propositions. : ； in The deep feature extraction branch supports the proposition. The probability distribution, Indicate the supporting proposition of the reasoning branch of the belief rule. The distribution of belief levels; Subsequently, the consistency evidence was orthogonally normalized and fused based on the conflict coefficient to obtain the final distribution of health status beliefs. : ； Combining health status belief distribution Compared with the preset level utility value Calculate the final comprehensive health score. The formula is: 。 7. The rotor bearing reliability analysis method based on deep attention and belief fusion according to claim 1, characterized in that, In practical applications, the Dropout layer in the deep feature extraction branch is forced to remain enabled; the multi-sensor time series to be tested is taken, and after sliding window partitioning and preprocessing, the sequence within one of the sliding windows is selected. This allows it to continuously undergo deep belief fusion analysis model. Second forward propagation; obtained Slightly different comprehensive health score sequences among groups ;in For the first The overall health score obtained from the first forward propagation; Calculate the predicted mean The formula for the final output health status assessment value is: ； in, For the first The overall health score is obtained from the forward propagation.

8. The rotor bearing reliability analysis method based on deep attention and belief fusion according to claim 7, characterized in that, Calculate the prediction variance Used to characterize the deep belief fusion analysis model for Uncertainty in judgment: ； in This is the model accuracy term, representing the inherent noise of the data; Finally, based on the preset confidence level, and combined with the prediction variance and prediction mean, the confidence interval of the prediction result is determined; an alarm is issued only when both of the following conditions are met simultaneously: The first is the threshold condition, which is that the predicted mean is lower than the preset health threshold. Second, the confidence condition, that is, the lower confidence limit of the confidence interval is lower than the preset risk threshold.

9. A terminal device, comprising a processor, a memory, and a computer program stored in the memory; characterized in that, When the processor executes the computer program, it implements the rotor bearing reliability analysis method based on deep attention and belief fusion as described in any one of claims 1-8.

10. A computer-readable storage medium storing a computer program; characterized in that, When the computer program is executed by the processor, it implements the rotor bearing reliability analysis method based on deep attention and belief fusion as described in any one of claims 1-8.

Images