Intelligent identification method and system for bearing temperature vibration signal fault evolution path

Abstract

The application provides a bearing temperature and vibration signal fault evolution path intelligent identification method and system, relates to the technical field of fault identification, and comprises the following steps: collecting and time-aligning bearing temperature, vibration signals and working condition parameters to construct a multi-source heterogeneous data matrix; extracting multi-dimensional features after adaptive noise reduction processing; constructing a fault evolution mechanism causal correlation network by using a causal inference model; performing time series clustering to divide health state clusters based on the network and establishing a state transition model by using a hidden Markov model; and finally applying a Viterbi algorithm to decoding to obtain an optimal fault evolution path. The method can effectively identify the bearing fault evolution law and improve the prediction accuracy.

Description

Technical Field

[0001] This invention relates to the field of fault identification technology, and in particular to an intelligent identification method and system for the fault evolution path of bearing temperature and vibration signals. Background Technology

[0002] As a critical component in rotating machinery, the health of bearings directly impacts the safe operation and production efficiency of equipment. With the development of industrial automation and intelligence, the need for early diagnosis of bearing faults and identification of fault evolution paths is becoming increasingly urgent. Traditional bearing condition monitoring mainly relies on vibration signal analysis, using frequency domain feature extraction to determine the fault type. However, bearings experience various physical phenomena during actual operation, such as temperature increases and intensified vibrations. These phenomena are coupled and evolve together, forming a complex fault development process. Therefore, comprehensively utilizing the two most important monitoring signals—temperature and vibration—to construct an intelligent identification method for bearing fault evolution paths is of great significance for improving equipment operational reliability and predictive maintenance.

[0003] Existing bearing fault evolution path identification technologies suffer from the following shortcomings: First, most methods focus only on a single signal source (e.g., analyzing only vibration signals), neglecting the coupling relationships between multiple signal sources such as temperature and vibration. This makes it difficult to comprehensively reflect the physical process of fault development, resulting in incomplete fault feature representation and affecting identification accuracy. Second, traditional methods typically cluster or classify signal features directly, lacking in-depth exploration of the causal relationships between feature variables. This fails to reveal the intrinsic mechanism and propagation path of fault development, leading to a lack of interpretability in fault diagnosis results. Summary of the Invention

[0004] This invention provides a method and system for intelligent identification of bearing temperature vibration signal fault evolution paths, which can solve the problems in the prior art.

[0005] A first aspect of the present invention provides a method for intelligent identification of bearing temperature vibration signal fault evolution paths, comprising:

[0006] Temperature signals, vibration signals, and operating parameters of the bearing are collected during continuous operation. The temperature signals, vibration signals, and operating parameters are then time-aligned according to the sampling time to construct a multi-source heterogeneous data matrix.

[0007] After adaptive noise reduction processing of the multi-source heterogeneous data matrix, temperature change rate features, vibration and shock energy features, temperature-vibration cross-entropy features, and operating load features are extracted to form a multi-dimensional feature set;

[0008] A causal inference model is used to mine causal relationships among the feature variables in the multidimensional feature set. By calculating the conditional independence and transit entropy between feature variables, a causal association network reflecting the fault evolution mechanism is constructed. The nodes of the causal association network represent feature variables, the directed edges represent the direction of causal influence, and the edge weights represent the causal strength coefficients.

[0009] Based on the aforementioned causal relationship network, a temporal clustering algorithm is used to divide the historical operating data of the bearing into multiple health state clusters. A hidden Markov model is then used to model the transition patterns between these state clusters, resulting in a state transition matrix and an observation probability matrix. The temporal clustering algorithm determines the cluster boundaries by minimizing the differences in the causal relationship network structure within each cluster.

[0010] Based on the state transition matrix and the observation probability matrix, the Viterbi algorithm is used to decode the optimal state sequence from the initial healthy state to the target fault state, and the optimal state sequence is the fault evolution path.

[0011] The step involves using a causal inference model to mine causal relationships among the feature variables in the multidimensional feature set. By calculating the conditional independence and transit entropy among the feature variables, a causal association network reflecting the fault evolution mechanism is constructed, including:

[0012] Construct a joint probability distribution model of each feature variable in the multidimensional feature set, and calculate the conditional independence test statistic of any two feature variables under the condition of a given third feature variable by performing conditional probability decomposition on the joint probability distribution model.

[0013] The multidimensional feature set is segmented based on a sliding time window. Within each time window, the information transfer amount from the source feature variable to the target feature variable is calculated. The information transfer amount is measured by the reduction of the conditional entropy of the historical state of the source feature variable to the current state of the target feature variable, thus obtaining the transfer entropy matrix between each pair of feature variables.

[0014] The conditional independence test statistic and the transit entropy matrix are fused together for discrimination. When the conditional independence test statistic of the feature variable pair is lower than the independence threshold and the transit entropy is higher than the causal threshold, it is determined that there is a causal relationship between the feature variable pair, and a directed edge is established in the causal relationship network.

[0015] The directed edges in the causal network are assigned edge weights, which are calculated by weighting the normalized value of the transit entropy of the corresponding feature variable pair with the inverse of the conditional independence test statistic. The numerical value of the edge weight represents the magnitude of the causal strength coefficient.

[0016] The segmentation of the multidimensional feature set based on a sliding time window, and the calculation of the information transfer amount from the source feature variable to the target feature variable within each time window, includes:

[0017] The window length and sliding step size of the sliding time window are set. The window length is determined according to the bearing failure evolution time scale, and the sliding step size is determined according to the data sampling frequency.

[0018] Within each sliding time window, extract the historical state sequence of the source feature variable before the current time and the state value of the target feature variable at the current time, and construct the joint probability distribution of the historical state sequence of the source feature variable and the current state of the target feature variable;

[0019] Calculate the conditional entropy of the target feature variable under the given historical state sequence of the source feature variable, and calculate the unconditional entropy of the target feature variable. Use the difference between the unconditional entropy and the conditional entropy as the information transfer amount from the source feature variable to the target feature variable.

[0020] Traverse all feature variable pairs in the multidimensional feature set, repeat the above information transfer calculation process for each feature variable pair, and organize the information transfer of all feature variable pairs into a transfer entropy matrix. The row index of the transfer entropy matrix represents the source feature variable, and the column index represents the target feature variable.

[0021] Based on the causal relationship network, a temporal clustering algorithm is used to divide the historical operating data of the bearing into multiple health state clusters, and a hidden Markov model is used to model the transition patterns between state clusters, including:

[0022] The historical operating data of the bearing is divided into multiple data segments in chronological order. Each data segment corresponds to a time interval. For each data segment, a corresponding causal relationship network instance is constructed.

[0023] A causal network structure similarity measurement function is defined, which is obtained by comprehensively calculating the node correspondence, edge connection pattern consistency, and edge weight Euclidean distance between two causal network instances.

[0024] Hierarchical clustering is used to cluster the causal association network instances corresponding to all data segments. The hierarchical clustering method takes minimizing the variance of the causal association network structure similarity measurement function within the cluster as the objective function and determines the cluster boundary through a bottom-up merging strategy to obtain multiple healthy state clusters.

[0025] A hidden Markov model is established for the multiple health state clusters, and the multiple health state clusters are used as the set of hidden states of the hidden Markov model. The multidimensional feature set is used as the set of observation variables. The state transition matrix is ​​calculated by statistically analyzing the transition frequency of adjacent data segments to which the health state clusters belong in the historical operation data. The observation probability matrix is ​​calculated by statistically analyzing the feature distribution of data segments within each health state cluster.

[0026] The defined causal network structure similarity measurement function is obtained by comprehensively calculating the node correspondence, edge connection pattern consistency, and edge weight Euclidean distance between two causal network instances, including:

[0027] Node alignment is performed on two causal network instances to be compared. Since the nodes of the two causal network instances both represent feature variables in the multidimensional feature set, a one-to-one correspondence between nodes is established through the names of the feature variables.

[0028] Construct adjacency matrices for two causal network instances. The element values ​​of the adjacency matrices indicate whether there is a directed edge between corresponding node pairs. Calculate the consistency coefficient of the edge connection pattern by comparing the two adjacency matrices element by element. The consistency coefficient of the edge connection pattern is the proportion of the number of elements with the same value at the same position in the two adjacency matrices to the total number of elements.

[0029] Extract the common directed edges from two causal network instances, calculate the square of the difference in edge weights for each common directed edge, and sum and take the square root of the square of the sum of the squares of the differences in edge weights of all common directed edges to obtain the Euclidean distance of the edge weights.

[0030] The consistency coefficient of the edge connection pattern and the Euclidean distance of the edge weight are weighted and combined. In the weighted combination, the consistency coefficient of the edge connection pattern is given a positive weight, and the Euclidean distance of the edge weight is given a negative weight. After normalization, the output value of the causal association network structure similarity measurement function is obtained.

[0031] The step of decoding the optimal state sequence from the initial healthy state to the target fault state using the Viterbi algorithm based on the state transition matrix and the observation probability matrix includes:

[0032] Obtain the bearing operation observation sequence of the evolution path to be identified, wherein the bearing operation observation sequence consists of a multi-dimensional feature set of continuous time moments;

[0033] Initialize the path probability matrix and path backtracking matrix of the Viterbi algorithm. The rows of the path probability matrix correspond to the set of hidden states of the hidden Markov model, and the columns correspond to the time index of the bearing operation observation sequence. The path backtracking matrix is ​​used to record the optimal predecessor state of each hidden state at each time.

[0034] Starting from the beginning of the bearing operation observation sequence, the local optimal path probability of each hidden state at each time moment is calculated sequentially. The local optimal path probability is determined by the maximum value of the product of the path probability of all possible predecessor states at the previous time moment, the corresponding state transition probability in the state transition matrix, and the probability of the current hidden state generating the current observation value in the observation probability matrix. The predecessor state that makes the local optimal path probability reach the maximum value is recorded in the path backtracking matrix.

[0035] At the termination time of the bearing operation observation sequence, the hidden state with the highest probability value in the path probability matrix is ​​selected as the optimal termination state. Starting from the optimal termination state, the reverse backtracking is performed based on the predecessor state information recorded in the path backtracking matrix to obtain the optimal state sequence from the initial healthy state to the target fault state.

[0036] A second aspect of the present invention provides an intelligent identification system for the fault evolution path of bearing temperature vibration signals, comprising:

[0037] The first unit is used to collect temperature signals, vibration signals, and operating parameters of the bearing during continuous operation. The temperature signals, vibration signals, and operating parameters are time-sequentially aligned according to the sampling time to construct a multi-source heterogeneous data matrix.

[0038] The second unit is used to perform adaptive noise reduction processing on the multi-source heterogeneous data matrix, and then extract temperature change rate features, vibration and shock energy features, temperature and vibration cross-entropy features, and working condition load features to form a multi-dimensional feature set.

[0039] The third unit is used to mine causal relationships among the feature variables in the multidimensional feature set using a causal inference model. By calculating the conditional independence and transit entropy between feature variables, a causal association network reflecting the fault evolution mechanism is constructed. The nodes of the causal association network represent feature variables, the directed edges represent the direction of causal influence, and the edge weights represent the causal strength coefficients.

[0040] The fourth unit is used to divide the historical operating data of the bearing into multiple health state clusters based on the causal association network using a temporal clustering algorithm, and to model the transition rules between state clusters using a hidden Markov model to obtain the state transition matrix and the observation probability matrix. The temporal clustering algorithm determines the cluster boundary by minimizing the differences in the causal association network structure within the cluster.

[0041] The fifth unit is used to decode the optimal state sequence from the initial healthy state to the target fault state using the Viterbi algorithm based on the state transition matrix and the observation probability matrix. The optimal state sequence is the fault evolution path.

[0042] A third aspect of the present invention,

[0043] An electronic device is provided, comprising:

[0044] processor;

[0045] Memory used to store processor-executable instructions;

[0046] The processor is configured to invoke instructions stored in the memory to execute the aforementioned method.

[0047] Fourth aspect of the embodiments of the present invention,

[0048] A computer-readable storage medium is provided, having stored thereon computer program instructions that, when executed by a processor, implement the aforementioned method.

[0049] The beneficial effects of this application are as follows:

[0050] By constructing a multi-source heterogeneous data matrix, the time-series alignment of temperature signals, vibration signals, and operating parameters was achieved, solving the problem of multi-source heterogeneous data fusion and providing a complete and consistent data foundation for subsequent analysis.

[0051] By employing a causal inference model to mine causal relationships among feature variables, a causal association network reflecting the failure evolution mechanism was constructed, breaking through the limitations of traditional correlation analysis and revealing the intrinsic mechanism and propagation path of bearing failure evolution.

[0052] By combining Hidden Markov Models and Viterbi Algorithms, the state transition laws are modeled and the optimal state sequence is decoded, enabling intelligent identification of the evolution path from the initial healthy state to the target fault state, thus providing a scientific basis for bearing fault early warning and maintenance decisions. Attached Figure Description

[0053] Figure 1 This is a flowchart illustrating the intelligent identification method for bearing temperature vibration signal fault evolution path according to an embodiment of the present invention. Detailed Implementation

[0054] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, 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.

[0055] The technical solution of the present invention will be described in detail below with reference to specific embodiments. These specific embodiments can be combined with each other, and the same or similar concepts or processes may not be described again in some embodiments.

[0056] Figure 1 This is a flowchart illustrating the intelligent identification method for bearing temperature vibration signal fault evolution path according to an embodiment of the present invention. Figure 1 As shown, the method includes:

[0057] Temperature signals, vibration signals, and operating parameters of the bearing are collected during continuous operation. The temperature signals, vibration signals, and operating parameters are then time-aligned according to the sampling time to construct a multi-source heterogeneous data matrix.

[0058] After adaptive noise reduction processing of the multi-source heterogeneous data matrix, temperature change rate features, vibration and shock energy features, temperature-vibration cross-entropy features, and operating load features are extracted to form a multi-dimensional feature set;

[0059] A causal inference model is used to mine causal relationships among the feature variables in the multidimensional feature set. By calculating the conditional independence and transit entropy between feature variables, a causal association network reflecting the fault evolution mechanism is constructed. The nodes of the causal association network represent feature variables, the directed edges represent the direction of causal influence, and the edge weights represent the causal strength coefficients.

[0060] Based on the aforementioned causal relationship network, a temporal clustering algorithm is used to divide the historical operating data of the bearing into multiple health state clusters. A hidden Markov model is then used to model the transition patterns between these state clusters, resulting in a state transition matrix and an observation probability matrix. The temporal clustering algorithm determines the cluster boundaries by minimizing the differences in the causal relationship network structure within each cluster.

[0061] Based on the state transition matrix and the observation probability matrix, the Viterbi algorithm is used to decode the optimal state sequence from the initial healthy state to the target fault state, and the optimal state sequence is the fault evolution path.

[0062] For example, temperature signals, vibration signals, and operating parameters of the bearing are collected during continuous operation. Temperature signals can be acquired using thermocouples or infrared temperature sensors, recording the bearing surface temperature at a frequency of no less than ten times per second; vibration signals are acquired using accelerometers, recording the bearing vibration acceleration value at a sampling frequency of no less than 10 kHz; operating parameters, including speed, load, and ambient temperature, are acquired through corresponding sensors.

[0063] To ensure the temporal consistency of multi-source data, temperature signals, vibration signals, and operating parameters need to be time-aligned according to their sampling times. Specifically, this involves downsampling high-frequency signals based on the lowest sampling frequency, ensuring a one-to-one correspondence between different signal types on the time axis. For example, when the vibration signal sampling frequency is 20kHz, the temperature signal sampling frequency is 10Hz, and the operating parameter sampling frequency is 1Hz, the vibration signal is averaged every 2000 points, and the temperature signal is averaged every 10 points, ultimately forming an aligned data matrix based on 1Hz.

[0064] The aligned multi-source heterogeneous data matrix may contain noise interference, requiring adaptive denoising. Temperature signals are denoised using wavelet thresholding with a db4 wavelet basis and a three-layer decomposition method, employing soft thresholding to remove noise. Vibration signals are denoised using adaptive filtering, with the filter order automatically adjusted based on signal characteristics. Operating parameters are denoised using median filtering to remove outliers. This denoising process ensures the accuracy of subsequent feature extraction.

[0065] Based on the denoised data matrix, a multidimensional feature set is extracted. Temperature change rate features are obtained by calculating the temperature difference between adjacent time points divided by the time interval, and its mean, standard deviation, kurtosis, and skewness are statistically analyzed. Vibration and shock energy features are extracted through envelope demodulation analysis, including peak power coefficient, margin coefficient, and impulse coefficient. Temperature-vibration cross-entropy features reflect the coupling relationship between temperature and vibration signals by calculating their mutual information entropy. Operating load features include average rotational speed and load fluctuation. All features are normalized to form a feature matrix with uniform dimensions.

[0066] This paper employs a causal inference model to uncover causal relationships among feature variables. For example, the causal inference model used in this method is a hybrid model combining a structure learning algorithm based on a Bayesian network framework and transfer entropy analysis. First, an undirected skeleton graph of the feature variables is constructed using the PC algorithm, and the dependencies between variables are determined through conditional independence tests. Then, the direction of the edges is determined based on transfer entropy calculation; a larger transfer entropy value indicates a stronger causal relationship. For instance, if the transfer entropy value of vibration feature to temperature feature is significantly higher than that of temperature feature to vibration feature, then the causal direction from vibration feature to temperature feature is determined. Finally, a causal association network is constructed, where nodes represent feature variables, directed edges represent the direction of causal influence, and edge weights represent the causal strength coefficient.

[0067] Based on the constructed causal relationship network, a temporal clustering algorithm is used to divide the historical operating data of the bearings into multiple health state clusters. During the clustering process, the similarity of the causal relationship network structure within different time windows is calculated as a distance metric. A dynamic time warping algorithm is used to align the temporal features, and then the density clustering method DBSCAN is used to determine the number and boundaries of the state clusters. The clustering results may include multiple categories such as normal state, early degradation state, intermediate degradation state, and fault state.

[0068] For the identified state clusters, a Hidden Markov Model (HMM) is used to model the state transition patterns. Each state cluster is treated as a hidden state of the HMM, and the observation sequence is the feature vector at the corresponding time step. The Baum-Welch algorithm is used to estimate the model parameters, yielding the state transition matrix and the observation probability matrix. The state transition matrix describes the probability of transitioning from one state to another, while the observation probability matrix describes the probability of observing a specific feature vector in a given state.

[0069] Based on the state transition matrix and the observation probability matrix, the Viterbi algorithm is used to decode the optimal state sequence. Given the current observation feature sequence of the bearing, the Viterbi algorithm can find the most likely hidden state sequence that generated this observation sequence, that is, the optimal state sequence from the initial healthy state to the target failure state. This state sequence is the evolution path of the bearing failure and can be used for failure prediction and health management.

[0070] In practical applications, the remaining service life of a bearing can be estimated based on the fault evolution path obtained through decoding and the residence time distribution of each state. For example, for a bearing in an early degradation state, the expected failure time of the bearing can be estimated based on the average transition time from this state to the failure state in historical data, thereby allowing for advance planning of maintenance and avoiding equipment downtime losses due to sudden failures. The fault evolution mechanism is mainly manifested as follows: fatigue of the bearing material leads to the generation and propagation of microcracks, causing changes in local stress distribution, which in turn leads to changes in vibration characteristics; at the same time, increased friction generates additional heat, causing the temperature to rise, which in turn accelerates material degradation, forming a positive feedback loop. This evolution process is reflected in the causal links and loops between characteristic variables in the causal network.

[0071] The advantage of this method lies in fully integrating the complementary information of temperature and vibration signals, revealing the intrinsic mechanism of fault evolution through causal relationship analysis. It can not only identify the fault evolution path, but also explain the influence relationship between different features, providing a new technical means for bearing health monitoring and predictive maintenance.

[0072] In one optional implementation, the step of using a causal inference model to mine causal relationships among the feature variables in the multidimensional feature set, and constructing a causal association network reflecting the fault evolution mechanism by calculating the conditional independence and transit entropy between feature variables, includes:

[0073] Construct a joint probability distribution model of each feature variable in the multidimensional feature set, and calculate the conditional independence test statistic of any two feature variables under the condition of a given third feature variable by performing conditional probability decomposition on the joint probability distribution model.

[0074] The multidimensional feature set is segmented based on a sliding time window. Within each time window, the information transfer amount from the source feature variable to the target feature variable is calculated. The information transfer amount is measured by the reduction of the conditional entropy of the historical state of the source feature variable to the current state of the target feature variable, thus obtaining the transfer entropy matrix between each pair of feature variables.

[0075] The conditional independence test statistic and the transit entropy matrix are fused together for discrimination. When the conditional independence test statistic of the feature variable pair is lower than the independence threshold and the transit entropy is higher than the causal threshold, it is determined that there is a causal relationship between the feature variable pair, and a directed edge is established in the causal relationship network.

[0076] The directed edges in the causal network are assigned edge weights, which are calculated by weighting the normalized value of the transit entropy of the corresponding feature variable pair with the inverse of the conditional independence test statistic. The numerical value of the edge weight represents the magnitude of the causal strength coefficient.

[0077] In the fault diagnosis system, a causal inference model is used to process the multidimensional feature set and construct a causal association network reflecting the fault evolution mechanism, providing a basis for equipment health status assessment. The specific implementation process is as follows:

[0078] Acquire multidimensional feature set data, which includes various sensor signals or extracted feature variables collected during equipment operation. These feature variables may include physical quantities such as temperature, vibration, and pressure, as well as their statistical characteristics.

[0079] Based on the acquired multidimensional feature set, a joint probability distribution model among the feature variables is constructed. Assuming the feature variable set is X = {X1, X2, ..., Xn}, the joint probability distribution P(X1, X2, ..., Xn) of these variables is constructed, where n is the number of features. Using probabilistic graphical models such as Bayesian networks or Gaussian graphical models, the joint probability distribution is conditionally decomposed to obtain P(X1, X2, ..., Xn) = ∏ᵢP(Xi|Pai), where Pai is the set of parent nodes of Xi.

[0080] For any two feature variables Xi and Xj, given a third set of feature variables Z, calculate the conditional independence test statistic. This can be done using the mutual information method based on information theory or the partial correlation coefficient method based on regression. For example, the conditional mutual information I(Xi;Xj|Z) can be used to assess the dependence of Xi and Xj given Z; when this value is close to zero, it indicates that the two variables tend to be independent under the given conditions.

[0081] Next, a sliding time window technique is used to perform time-series segmentation on the multidimensional feature set. The window length *w* and sliding step size *s* are set to segment the time series of each feature variable. In practical applications, the window length can be determined based on the dynamic characteristics of the equipment; for example, several rotor cycles might be chosen as the window length for rotating machinery.

[0082] Within each time window, the amount of information transferred from the source feature variable to the target feature variable, i.e., the transfer entropy, is calculated. For the source variable X and the target variable Y, the transfer entropy measures the contribution of the historical state of X to the prediction of the current state of Y. The transfer entropy calculation is based on the reduction in conditional entropy; specifically, it is the difference between the conditional entropy of the target variable considering only its own historical state and the conditional entropy considering the historical state of the source variable. By comparing the improvement in predictive ability of the target variable based on its own history with that after incorporating the history of the source variable, the flow of information from the source variable to the target variable is quantified. The transfer entropy is calculated for all feature variable pairs, forming an n×n transfer entropy matrix TE, where TE... ij This represents the transfer entropy value from variable i to variable j.

[0083] The conditional independence test statistic and the transfer entropy matrix are fused to determine the causal relationship between feature variables. A conditional independence threshold θ1 and a transfer entropy threshold θ2 are set. For a feature variable pair (Xi, Xj), if the conditional independence test statistic is lower than θ1 and the transfer entropy TE is lower than θ2, then the causal relationship between the feature variables is determined. ij If the values ​​are higher than θ2, then Xi is determined to have a causal influence on Xj, and a directed edge from Xi to Xj is established in the causal association network. The thresholds θ1 and θ2 can be determined through cross-validation or domain expert knowledge to balance the accuracy and completeness of the causal association judgment.

[0084] Weights are assigned to directed edges in a causal network to reflect the strength of the causal relationship. The edge weight W can be calculated as a weighted combination of the normalized value of the transitive entropy and the reciprocal of the conditional independence test statistic: W ij = α·(TE ij / TE max ) + (1-α)·(1 / CI ij ), where TE max To maximize the propagation entropy, CI ijThis is the conditional independence statistic, and α is the tradeoff parameter (0 ≤ α ≤ 1). The larger the edge weight, the stronger the causal relationship.

[0085] The constructed causal network not only reveals the causal structure among feature variables but also quantifies the causal strength through edge weights, providing a powerful tool for root cause analysis of failures. By tracing the causal paths in the network, the initial source of failure can be identified and the failure propagation mechanism can be understood, thereby enabling the formulation of precise maintenance strategies and improving equipment operational reliability.

[0086] In one optional implementation, the step of segmenting the multidimensional feature set based on a sliding time window and calculating the information transfer amount from the source feature variable to the target feature variable within each time window includes:

[0087] The window length and sliding step size of the sliding time window are set. The window length is determined according to the bearing failure evolution time scale, and the sliding step size is determined according to the data sampling frequency.

[0088] Within each sliding time window, extract the historical state sequence of the source feature variable before the current time and the state value of the target feature variable at the current time, and construct the joint probability distribution of the historical state sequence of the source feature variable and the current state of the target feature variable;

[0089] Calculate the conditional entropy of the target feature variable under the given historical state sequence of the source feature variable, and calculate the unconditional entropy of the target feature variable. Use the difference between the unconditional entropy and the conditional entropy as the information transfer amount from the source feature variable to the target feature variable.

[0090] Traverse all feature variable pairs in the multidimensional feature set, repeat the above information transfer calculation process for each feature variable pair, and organize the information transfer of all feature variable pairs into a transfer entropy matrix. The row index of the transfer entropy matrix represents the source feature variable, and the column index represents the target feature variable.

[0091] During the bearing fault evolution process, the collected multidimensional feature set needs to be segmented to capture the dynamic characteristics of fault development. Setting the sliding time window is a crucial step, as the window length directly affects the accuracy of extracting the relationships between feature variables.

[0092] When setting the sliding time window, the window length is typically determined based on the bearing's fault evolution timescale. For example, for low-speed bearings, whose fault evolution cycle is longer, a larger window length can be set, such as 10,000 sampling points; while for high-speed bearings, whose fault evolution is faster, a smaller window length can be set, such as 2,000 sampling points. The sliding step size is determined based on the data sampling frequency. A larger step size, such as 100 sampling points, can be selected when the sampling frequency is high; a smaller step size, such as 20 sampling points, can be selected when the sampling frequency is low. This ensures smooth window movement and an appropriate amount of information.

[0093] Within each sliding time window, a relationship model between the source feature variable and the target feature variable needs to be constructed. First, the historical state sequence of the source feature variable X is extracted, namely X(tk), X(t-k+1), ..., X(t-1), where k is the length of the historical state, typically set according to the system response characteristics, and can be selected as 3-5 time steps. Simultaneously, the state value Y(t) of the target feature variable Y at the current time t is extracted. Then, the joint probability distribution P(X^k_t, Y_t) of the historical state sequence of the source feature variable and the current state of the target feature variable is constructed, where X^k_t represents the sequence of k historical states before time t.

[0094] Probability distributions can be constructed using histograms. This involves dividing the range of feature variables into several equally wide intervals, counting the number of sample points within each interval, and then normalizing the count. For example, dividing the feature range [min, max] into eight equal intervals and counting the frequency of samples falling into each interval yields a discrete probability distribution. In practical applications, the number of intervals can be dynamically adjusted based on the amount of data and distribution characteristics. When the data volume is large, the number of intervals can be increased to improve accuracy.

[0095] When calculating the amount of information transferred, first calculate the unconditional entropy H(Y) of the target feature variable Y, representing the uncertainty of Y; then calculate the conditional entropy H(Y|X^k) given the historical state sequence of the source feature variable X, representing the uncertainty of Y after knowing the historical information of X. The difference between the two is the transfer entropy TE(X→Y) = H(Y) - H(Y|X^k), representing the amount of information transferred from X to Y.

[0096] To comprehensively analyze the information flow between feature variables, it is necessary to traverse all feature variable pairs in the multidimensional feature set. Assuming the feature set contains n feature variables, the transfer entropy of n×(n-1) feature pairs needs to be calculated. For each feature variable pair (i,j), the transfer entropy TE(i→j) from feature i to feature j is calculated, forming an n×n transfer entropy matrix. The element in the i-th row and j-th column of the matrix represents the amount of information transferred from feature i to feature j. Diagonal elements are usually set to 0 because information transfer from itself to itself is meaningless.

[0097] In practical applications, such as a bearing vibration monitoring system that collects multiple characteristic variables including radial vibration, axial vibration, and temperature, the transfer entropy matrix calculated using the above method can be observed to significantly increase the amount of information transferred from temperature characteristics to vibration characteristics in the early stages of a fault. This indicates that temperature changes foreshadow changes in vibration characteristics, providing an important basis for early fault warning.

[0098] By calculating the transfer entropy matrix piecewise using a sliding window, the temporal evolution of information transfer relationships between feature variables can be tracked, thereby more accurately capturing the dynamic changes in the fault development process and providing reliable technical support for bearing health status assessment and fault prediction.

[0099] After the transfer entropy matrix is ​​constructed, the causal relationship network between feature variables can be further analyzed to identify key feature variables and their influence paths, providing input for subsequent fault diagnosis and prediction models, and realizing accurate characterization and prediction of bearing fault evolution process.

[0100] In one optional implementation, based on the causal correlation network, a temporal clustering algorithm is used to divide the historical operating data of the bearing into multiple health state clusters, and a hidden Markov model is used to model the transition patterns between state clusters, including:

[0101] The historical operating data of the bearing is divided into multiple data segments in chronological order. Each data segment corresponds to a time interval. For each data segment, a corresponding causal relationship network instance is constructed.

[0102] A causal network structure similarity measurement function is defined, which is obtained by comprehensively calculating the node correspondence, edge connection pattern consistency, and edge weight Euclidean distance between two causal network instances.

[0103] Hierarchical clustering is used to cluster the causal association network instances corresponding to all data segments. The hierarchical clustering method takes minimizing the variance of the causal association network structure similarity measurement function within the cluster as the objective function and determines the cluster boundary through a bottom-up merging strategy to obtain multiple healthy state clusters.

[0104] A hidden Markov model is established for the multiple health state clusters, and the multiple health state clusters are used as the set of hidden states of the hidden Markov model. The multidimensional feature set is used as the set of observation variables. The state transition matrix is ​​calculated by statistically analyzing the transition frequency of adjacent data segments to which the health state clusters belong in the historical operation data. The observation probability matrix is ​​calculated by statistically analyzing the feature distribution of data segments within each health state cluster.

[0105] The bearing health status clustering and transition pattern modeling method based on causal association network first divides the historical operating data of the bearing according to time sequence, then constructs a causal association network for each data segment, and identifies multiple health states of the bearing through similarity measurement function and hierarchical clustering method. Finally, a hidden Markov model is used to model the transition patterns between these states.

[0106] The historical operating data of the bearing is divided into multiple data segments according to time sequence. Assuming there is a bearing operating data segment of length T, it can be divided into M data segments of fixed length L. For the i-th data segment, its corresponding time interval is [(i-1)×L+1, i×L]. For example, for a bearing operating data segment of 10,000 points, if each data segment is set to a length of 500 points, it can be divided into 20 data segments.

[0107] For each data segment, a corresponding causal association network instance is constructed. Each feature in the bearing's multidimensional feature set is used as a node in the network. The causal relationships between nodes are determined using methods such as Granger causality tests or transitive entropy, and directed edges are established. The weights of the edges can be calculated using causal strength, for example, by using the F-statistic or the transformed p-value of the Granger causality test. Assuming that the bearing's multidimensional feature set contains 10 features, including the RMS value, peak value, kurtosis, and frequency domain energy of the vibration signal, each causal association network instance contains 10 nodes, and the edges between the nodes are established based on the causal test results.

[0108] Next, we define the similarity measurement function for causal network structures. Assume there are two causal network instances G1 and G2, with node sets V1 and V2, edge sets E1 and E2, and edge weight sets W1 and W2, respectively. The similarity measurement function S(G1,G2) can be calculated comprehensively from the following three aspects:

[0109] Node correspondence similarity Snode: Since nodes in different network instances represent the same features, the node correspondence is a one-to-one mapping, therefore Snode=1.

[0110] Edge connection pattern consistency (Sedge): Calculates the proportion of edges that exist in both networks to the total number of edges, i.e., Sedge = (|E1∩E2|) / (|E1∪E2|).

[0111] Edge weight Euclidean distance Sweight: For edges that exist in both networks, calculate the Euclidean distance between their weights.

[0112] The final similarity metric function can be expressed as:

[0113] S(G1,G2)=α×Snode+β×Sedge+γ×Sweight, where α, β, and γ are weight coefficients, and α+β+γ=1.

[0114] Hierarchical clustering is used to cluster causal network instances corresponding to all data segments. The hierarchical clustering algorithm starts by treating each network instance as an independent cluster, then progressively merges the two most similar clusters until a termination condition is met. During the merging process, the inter-cluster distance can be calculated using the average link method, which is the average similarity between all pairs of network instances in two clusters. The objective function of clustering is to minimize the variance of the intra-cluster similarity metric. The clustering termination condition can be set to the number of clusters reaching a preset value K, or the inter-cluster distance being greater than a threshold θ. In this way, M data segments are divided into K healthy clusters C1, C2, ..., CK.

[0115] For example, suppose there are 20 data fragments corresponding to causal association network instances. After hierarchical clustering, they may be divided into 3 health state clusters: C1 represents the normal state, containing data fragments 1-7; C2 represents the slightly deteriorated state, containing data fragments 8-15; and C3 represents the severely deteriorated state, containing data fragments 16-20.

[0116] A Hidden Markov Model (HMM) is established for the multiple health state clusters obtained from the partitioning. The K health state clusters are used as the set of hidden states S={S1, S2, ..., SK} of the HMM, and the multidimensional feature set of the bearing is used as the set of observed variables O={O1, O2, ..., ON}.

[0117] The state transition matrix A is calculated based on the transition frequency of adjacent data segments belonging to the health state clusters in historical operating data. For any two states Si and Sj, the transition probability aij represents the probability of transitioning from state Si to state Sj, and is calculated using the formula aij = nij / ni, where nij represents the number of times state Si transitions to state Sj, and ni represents the total number of times state Si occurs.

[0118] The observation probability matrix B is calculated based on the feature distribution of data segments within each health state cluster. For a state Si and an observation Ok, the observation probability bik represents the probability of observing Ok in state Si. It can be estimated by statistically analyzing the distribution of features within each state cluster, for example, by fitting the feature distribution using a Gaussian mixture model.

[0119] Finally, the parameters of the Hidden Markov Model are trained using a forward-backward algorithm to obtain the complete model parameters λ=(A, B, π), where π is the initial state probability distribution. The Viterbi algorithm can then be used to infer the most probable state sequence based on the observation sequence, thereby enabling the identification and prediction of the bearing's health status.

[0120] The above method enables the clustering and modeling of bearing health status based on causal association networks, providing an effective technical means for bearing health status monitoring and fault early warning.

[0121] In one optional implementation, the definition of a causal network structure similarity metric function is obtained by comprehensively calculating the node correspondence, edge connection pattern consistency, and Euclidean distance of edge weights between two causal network instances, including:

[0122] Node alignment is performed on two causal network instances to be compared. Since the nodes of the two causal network instances both represent feature variables in the multidimensional feature set, a one-to-one correspondence between nodes is established through the names of the feature variables.

[0123] Construct adjacency matrices for two causal network instances. The element values ​​of the adjacency matrices indicate whether there is a directed edge between corresponding node pairs. Calculate the consistency coefficient of the edge connection pattern by comparing the two adjacency matrices element by element. The consistency coefficient of the edge connection pattern is the proportion of the number of elements with the same value at the same position in the two adjacency matrices to the total number of elements.

[0124] Extract the common directed edges from two causal network instances, calculate the square of the difference in edge weights for each common directed edge, and sum and take the square root of the square of the sum of the squares of the differences in edge weights of all common directed edges to obtain the Euclidean distance of the edge weights.

[0125] The consistency coefficient of the edge connection pattern and the Euclidean distance of the edge weight are weighted and combined. In the weighted combination, the consistency coefficient of the edge connection pattern is given a positive weight, and the Euclidean distance of the edge weight is given a negative weight. After normalization, the output value of the causal association network structure similarity measurement function is obtained.

[0126] In this embodiment, a structural similarity metric function for causal association networks is used to quantitatively evaluate the degree of structural similarity between two instances of causal association networks. This metric function comprehensively considers multi-dimensional features such as node correspondence, consistency of edge connection patterns, and Euclidean distance of edge weights, providing a foundation for subsequent clustering analysis of causal association networks.

[0127] Node alignment is performed on two causal network instances to be compared. Since the nodes in the two causal network instances represent feature variables in a multidimensional feature set, a one-to-one correspondence between nodes can be established through the names of the feature variables.

[0128] Suppose two causal network instances are G1 and G2, with node sets V1 and V2, respectively. A mapping relationship M: V1 → V2 can be established using feature variable names, such that for each node v1i in V1, there is a corresponding node v2j = M(v1i) in V2. This mapping relationship ensures that the structural information at corresponding positions in the two network instances can be correctly compared in subsequent steps.

[0129] Construct adjacency matrices A1 and A2 for two causal network instances. The element values ​​of the adjacency matrices represent whether there is a directed edge between corresponding node pairs. If there is a directed edge between node i and node j, the element A[i][j] in the adjacency matrix is ​​set to 1; otherwise, it is set to 0. Calculate the consistency coefficient of the edge connection pattern by comparing the two adjacency matrices element by element.

[0130] Specifically, the consistency coefficient C of the edge connection pattern is calculated as follows: Traverse each corresponding element in the two adjacency matrices A1 and A2, count the number of elements with the same value, and then divide by the total number of elements. Assuming each network has n nodes, the consistency coefficient can be expressed as: the total number of identical elements divided by the total number of elements in the adjacency matrices (n×n). This coefficient reflects the degree of similarity in the topology of the two causal networks.

[0131] For example, if two networks each have 3 nodes, their adjacency matrices are as follows:

[0132] A1 = [[0,1,0], [0,0,1], [0,0,0]];

[0133] A2 = [[0,1,0], [0,0,0], [1,0,0]];

[0134] By comparing each element, it was found that there are 7 positions with the same element value, and the total number of elements is 9. Therefore, the consistency coefficient is 7 / 9.

[0135] Subsequently, common directed edges are extracted from the two causal network instances. Common directed edges are those connecting the same pairs of nodes in both networks. For each common directed edge, the square of the difference in its edge weight is calculated. Edge weights typically represent the strength of the causal relationship and are important information in causal networks.

[0136] Suppose two networks share m edges, and the weight of the i-th common edge is w1i in G1 and w2i in G2. Then the Euclidean distance D of the edge weights can be calculated as the square root of the sum of the squares of the differences in the weights of all common edges. This distance reflects the degree of difference in the edge weight distribution between the two networks.

[0137] Finally, the consistency coefficients of the edge connection patterns and the Euclidean distance of the edge weights are weighted and combined to obtain the final similarity metric for the causal network structures. In the weighted combination, the consistency coefficients of the edge connection patterns are given positive weights, indicating that the higher the consistency, the more similar the networks are; the Euclidean distance of the edge weights is given negative weights, indicating that the larger the distance, the less similar the networks are.

[0138] Specifically, weighting factors α and β can be set such that the similarity S = α×C - β×D, where α+β=1, and both α and β are positive values. Then, the S value is normalized to ensure that the final similarity measure falls within the [0,1] interval, which facilitates subsequent comparative analysis.

[0139] This similarity measurement method, which comprehensively considers network topology and edge weight information, can more fully capture the structural differences between causal network instances, providing a reliable quantitative basis for subsequent network classification, clustering, and evolutionary analysis.

[0140] In one optional implementation, the step of decoding the optimal state sequence from the initial healthy state to the target fault state using the Viterbi algorithm based on the state transition matrix and the observation probability matrix includes:

[0141] Obtain the bearing operation observation sequence of the evolution path to be identified, wherein the bearing operation observation sequence consists of a multi-dimensional feature set of continuous time moments;

[0142] Initialize the path probability matrix and path backtracking matrix of the Viterbi algorithm. The rows of the path probability matrix correspond to the set of hidden states of the hidden Markov model, and the columns correspond to the time index of the bearing operation observation sequence. The path backtracking matrix is ​​used to record the optimal predecessor state of each hidden state at each time.

[0143] Starting from the beginning of the bearing operation observation sequence, the local optimal path probability of each hidden state at each time moment is calculated sequentially. The local optimal path probability is determined by the maximum value of the product of the path probability of all possible predecessor states at the previous time moment, the corresponding state transition probability in the state transition matrix, and the probability of the current hidden state generating the current observation value in the observation probability matrix. The predecessor state that makes the local optimal path probability reach the maximum value is recorded in the path backtracking matrix.

[0144] At the termination time of the bearing operation observation sequence, the hidden state with the highest probability value in the path probability matrix is ​​selected as the optimal termination state. Starting from the optimal termination state, the reverse backtracking is performed based on the predecessor state information recorded in the path backtracking matrix to obtain the optimal state sequence from the initial healthy state to the target fault state.

[0145] The process involves acquiring a bearing operation observation sequence with a known evolutionary path. This sequence consists of a multi-dimensional feature set spanning consecutive time points. In practical applications, vibration signals can be collected using sensors mounted on the bearing. Feature parameters, such as root mean square (RMS), peak value, kurtosis, and skewness, are extracted using time-domain analysis, frequency-domain analysis, or time-frequency analysis to form feature vectors. These include time-domain features such as RMS, peak value, kurtosis, and skewness, as well as frequency-domain features such as energy distribution and dominant frequency components across different frequency bands. For example, in a motor bearing monitoring system, vibration data is collected every 10 minutes, and an 8-dimensional feature vector is extracted. After 24 hours of continuous monitoring, an observation sequence of 144 time points is generated.

[0146] Initialize the path probability matrix and path backtracking matrix required for the Viterbi algorithm. The rows of the path probability matrix correspond to the set of hidden states in the Hidden Markov Model, and the columns correspond to the time indices of the bearing operation observation sequence. Assuming the bearing state is divided into four types—healthy, early wear, moderate wear, and severe failure—then the path probability matrix has 4 rows and the number of columns equal to the length of the observation sequence. Each element in the path probability matrix represents the maximum probability of being in the corresponding hidden state at that time. For the first time step, only the probability of the healthy state is 1 during initialization, while the probabilities of the other states are 0. For other time steps, the initial probabilities of all states are set to 0. The path backtracking matrix has the same dimension as the path probability matrix and is used to record the optimal predecessor state for each hidden state at each time step. Its initial value can be set to -1 to indicate that the predecessor state has not yet been determined.

[0147] Starting from the initial moment of the bearing operation observation sequence, the local optimal path probability of each hidden state at each moment is calculated sequentially. For state j at moment t, the local optimal path probability is calculated as follows: traverse all possible predecessor states i at moment t-1, calculate the joint probability of transitioning from state i to state j and observing the current observation value under state j, which is the product of the path probability of state i at moment t-1, the state transition probability of transitioning from state i to state j, and the observation probability of state j generating the current observation value. Then, select the maximum value among these joint probabilities as the local optimal path probability of state j at moment t. Simultaneously, record the index of the predecessor state that maximizes this local optimal path probability in the path backtracking matrix.

[0148] Specifically, let the observed value at time t be Ot, the state transition matrix be A, and the observation probability matrix be B. Then the local optimal path probability δ of state j at time t is... t The formula for calculating (j) is:

[0149] δ t (j) = max[δ t-1 [i] x A(i,j) x B(j,Ot)], where i traverses all possible predecessor states. Simultaneously, the state i that achieves this maximum value is recorded and stored in the corresponding position of the path backtracking matrix.

[0150] To illustrate with an example, suppose a bearing has four health states (healthy, early wear, moderate wear, and severe failure), and the eigenvector observed at time 20 is [2.5, 3.1, 1.8, 0.9]. The path probabilities for the four states at time 19 are known to be [0.6, 0.3, 0.1, 0], and the state transition matrix A and observation probability matrix B have been obtained through training. When calculating the local optimal path probability of the "moderate wear" state at time 20, it is necessary to calculate the probabilities of transitioning from the four states at the previous time. For example, the probability of transitioning from the "healthy" state is 0.6 × 0.05 × 0.02 (assuming A from healthy to moderate = 0.05, B generating the observation value in the moderate state = 0.02), the probability of transitioning from the "early wear" state is 0.3 × 0.2 × 0.02, and so on. The maximum value of 0.0012 is selected as the path probability of the "moderate wear" state at time 20, and its optimal preceding state is recorded as "early wear".

[0151] Finally, at the end of the bearing operation observation sequence, the hidden state with the highest probability value is selected as the optimal termination state from the path probability matrix. For example, at the last time T of the sequence, if the path probabilities of the four states are [0.1, 0.2, 0.4, 0.3], then the "moderate wear" state (probability 0.4) is selected as the optimal termination state. Then, starting from this optimal termination state, the process is reversed based on the preceding state information recorded in the path backtracking matrix, sequentially determining the optimal state at T-1, T-2, ... until the initial time, ultimately obtaining a complete optimal state sequence from the initial healthy state to the target fault state.

[0152] The state sequences identified in this way not only reflect the evolution of a bearing from healthy to faulty, but also accurately pinpoint the critical moments when the bearing's state changes, providing important data for predictive maintenance and fault early warning. For example, the state sequence identified in one experiment showed that the bearing transitioned from a healthy state to an early wear state after 50 hours of operation, and entered a moderate wear state after 75 hours, providing the equipment maintenance department with a clear maintenance time window.

[0153] The intelligent identification system for bearing temperature vibration signal fault evolution path according to an embodiment of the present invention includes:

[0154] The first unit is used to collect temperature signals, vibration signals, and operating parameters of the bearing during continuous operation. The temperature signals, vibration signals, and operating parameters are time-sequentially aligned according to the sampling time to construct a multi-source heterogeneous data matrix.

[0155] The second unit is used to perform adaptive noise reduction processing on the multi-source heterogeneous data matrix, and then extract temperature change rate features, vibration and shock energy features, temperature and vibration cross-entropy features, and working condition load features to form a multi-dimensional feature set.

[0156] The third unit is used to mine causal relationships among the feature variables in the multidimensional feature set using a causal inference model. By calculating the conditional independence and transit entropy between feature variables, a causal association network reflecting the fault evolution mechanism is constructed. The nodes of the causal association network represent feature variables, the directed edges represent the direction of causal influence, and the edge weights represent the causal strength coefficients.

[0157] The fourth unit is used to divide the historical operating data of the bearing into multiple health state clusters based on the causal association network using a temporal clustering algorithm, and to model the transition rules between state clusters using a hidden Markov model to obtain the state transition matrix and the observation probability matrix. The temporal clustering algorithm determines the cluster boundary by minimizing the differences in the causal association network structure within the cluster.

[0158] The fifth unit is used to decode the optimal state sequence from the initial healthy state to the target fault state using the Viterbi algorithm based on the state transition matrix and the observation probability matrix. The optimal state sequence is the fault evolution path.

[0159] A third aspect of the present invention,

[0160] An electronic device is provided, comprising:

[0161] processor;

[0162] Memory used to store processor-executable instructions;

[0163] The processor is configured to invoke instructions stored in the memory to execute the aforementioned method.

[0164] Fourth aspect of the embodiments of the present invention,

[0165] A computer-readable storage medium is provided, having stored thereon computer program instructions that, when executed by a processor, implement the aforementioned method.

[0166] This invention can be a method, apparatus, system, and / or computer program product. The computer program product may include a computer-readable storage medium having computer-readable program instructions loaded thereon for performing various aspects of the invention.

[0167] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some or all of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method for intelligent identification of bearing temperature vibration signal fault evolution path, characterized in that, include: Temperature signals, vibration signals, and operating parameters of the bearing are collected during continuous operation. The temperature signals, vibration signals, and operating parameters are then time-aligned according to the sampling time to construct a multi-source heterogeneous data matrix. After adaptive noise reduction processing of the multi-source heterogeneous data matrix, temperature change rate features, vibration and shock energy features, temperature-vibration cross-entropy features, and operating load features are extracted to form a multi-dimensional feature set; A causal inference model is used to mine causal relationships among the feature variables in the multidimensional feature set. By calculating the conditional independence and transit entropy between feature variables, a causal association network reflecting the fault evolution mechanism is constructed. The nodes of the causal association network represent feature variables, the directed edges represent the direction of causal influence, and the edge weights represent the causal strength coefficients. Based on the aforementioned causal relationship network, a temporal clustering algorithm is used to divide the historical operating data of the bearing into multiple health state clusters. A hidden Markov model is then used to model the transition patterns between these state clusters, resulting in a state transition matrix and an observation probability matrix. The temporal clustering algorithm determines the cluster boundaries by minimizing the differences in the causal relationship network structure within each cluster. Based on the state transition matrix and the observation probability matrix, the Viterbi algorithm is used to decode the optimal state sequence from the initial healthy state to the target fault state, and the optimal state sequence is the fault evolution path.

2. The method according to claim 1, characterized in that, The step involves using a causal inference model to mine causal relationships among the feature variables in the multidimensional feature set. By calculating the conditional independence and transit entropy among the feature variables, a causal association network reflecting the fault evolution mechanism is constructed, including: Construct a joint probability distribution model of each feature variable in the multidimensional feature set, and calculate the conditional independence test statistic of any two feature variables under the condition of a given third feature variable by performing conditional probability decomposition on the joint probability distribution model. The multidimensional feature set is segmented based on a sliding time window. Within each time window, the information transfer amount from the source feature variable to the target feature variable is calculated. The information transfer amount is measured by the reduction of the conditional entropy of the historical state of the source feature variable to the current state of the target feature variable, thus obtaining the transfer entropy matrix between each pair of feature variables. The conditional independence test statistic and the transit entropy matrix are fused together for discrimination. When the conditional independence test statistic of the feature variable pair is lower than the independence threshold and the transit entropy is higher than the causal threshold, it is determined that there is a causal relationship between the feature variable pair, and a directed edge is established in the causal relationship network. The directed edges in the causal network are assigned edge weights, which are calculated by weighting the normalized value of the transit entropy of the corresponding feature variable pair with the inverse of the conditional independence test statistic. The numerical value of the edge weight represents the magnitude of the causal strength coefficient.

3. The method according to claim 2, characterized in that, The segmentation of the multidimensional feature set based on a sliding time window, and the calculation of the information transfer amount from the source feature variable to the target feature variable within each time window, includes: The window length and sliding step size of the sliding time window are set. The window length is determined according to the bearing failure evolution time scale, and the sliding step size is determined according to the data sampling frequency. Within each sliding time window, extract the historical state sequence of the source feature variable before the current time and the state value of the target feature variable at the current time, and construct the joint probability distribution of the historical state sequence of the source feature variable and the current state of the target feature variable; Calculate the conditional entropy of the target feature variable under the given historical state sequence of the source feature variable, and calculate the unconditional entropy of the target feature variable. Use the difference between the unconditional entropy and the conditional entropy as the information transfer amount from the source feature variable to the target feature variable. Traverse all feature variable pairs in the multidimensional feature set, repeat the above information transfer calculation process for each feature variable pair, and organize the information transfer of all feature variable pairs into a transfer entropy matrix. The row index of the transfer entropy matrix represents the source feature variable, and the column index represents the target feature variable.

4. The method according to claim 1, characterized in that, Based on the causal relationship network, a temporal clustering algorithm is used to divide the historical operating data of the bearing into multiple health state clusters, and a hidden Markov model is used to model the transition patterns between state clusters, including: The historical operating data of the bearing is divided into multiple data segments in chronological order. Each data segment corresponds to a time interval. For each data segment, a corresponding causal relationship network instance is constructed. A causal network structure similarity measurement function is defined, which is obtained by comprehensively calculating the node correspondence, edge connection pattern consistency, and edge weight Euclidean distance between two causal network instances. Hierarchical clustering is used to cluster the causal association network instances corresponding to all data segments. The hierarchical clustering method takes minimizing the variance of the causal association network structure similarity measurement function within the cluster as the objective function and determines the cluster boundary through a bottom-up merging strategy to obtain multiple healthy state clusters. A hidden Markov model is established for the multiple health state clusters, and the multiple health state clusters are used as the set of hidden states of the hidden Markov model. The multidimensional feature set is used as the set of observation variables. The state transition matrix is ​​calculated by statistically analyzing the transition frequency of adjacent data segments to which the health state clusters belong in the historical operation data. The observation probability matrix is ​​calculated by statistically analyzing the feature distribution of data segments within each health state cluster.

5. The method according to claim 4, characterized in that, The defined causal network structure similarity measurement function is obtained by comprehensively calculating the node correspondence, edge connection pattern consistency, and edge weight Euclidean distance between two causal network instances, including: Node alignment is performed on two causal network instances to be compared. Since the nodes of the two causal network instances both represent feature variables in the multidimensional feature set, a one-to-one correspondence between nodes is established through the names of the feature variables. Construct adjacency matrices for two causal network instances. The element values ​​of the adjacency matrices indicate whether there is a directed edge between corresponding node pairs. Calculate the consistency coefficient of the edge connection pattern by comparing the two adjacency matrices element by element. The consistency coefficient of the edge connection pattern is the proportion of the number of elements with the same value at the same position in the two adjacency matrices to the total number of elements. Extract the common directed edges from two causal network instances, calculate the square of the difference in edge weights for each common directed edge, and sum and take the square root of the square of the sum of the squares of the differences in edge weights of all common directed edges to obtain the Euclidean distance of the edge weights. The consistency coefficient of the edge connection pattern and the Euclidean distance of the edge weight are weighted and combined. In the weighted combination, the consistency coefficient of the edge connection pattern is given a positive weight, and the Euclidean distance of the edge weight is given a negative weight. After normalization, the output value of the causal association network structure similarity measurement function is obtained.

6. The method according to claim 1, characterized in that, The step of decoding the optimal state sequence from the initial healthy state to the target fault state using the Viterbi algorithm based on the state transition matrix and the observation probability matrix includes: Obtain the bearing operation observation sequence of the evolution path to be identified, wherein the bearing operation observation sequence consists of a multi-dimensional feature set of continuous time moments; Initialize the path probability matrix and path backtracking matrix of the Viterbi algorithm. The rows of the path probability matrix correspond to the set of hidden states of the hidden Markov model, and the columns correspond to the time index of the bearing operation observation sequence. The path backtracking matrix is ​​used to record the optimal predecessor state of each hidden state at each time. Starting from the beginning of the bearing operation observation sequence, the local optimal path probability of each hidden state at each time moment is calculated sequentially. The local optimal path probability is determined by the maximum value of the product of the path probability of all possible predecessor states at the previous time moment, the corresponding state transition probability in the state transition matrix, and the probability of the current hidden state generating the current observation value in the observation probability matrix. The predecessor state that makes the local optimal path probability reach the maximum value is recorded in the path backtracking matrix. At the termination time of the bearing operation observation sequence, the hidden state with the highest probability value in the path probability matrix is ​​selected as the optimal termination state. Starting from the optimal termination state, the reverse backtracking is performed based on the predecessor state information recorded in the path backtracking matrix to obtain the optimal state sequence from the initial healthy state to the target fault state.

7. A bearing temperature vibration signal fault evolution path intelligent identification system, used to implement the method as described in any one of claims 1-6, characterized in that, include: The first unit is used to collect temperature signals, vibration signals, and operating parameters of the bearing during continuous operation. The temperature signals, vibration signals, and operating parameters are time-sequentially aligned according to the sampling time to construct a multi-source heterogeneous data matrix. The second unit is used to perform adaptive noise reduction processing on the multi-source heterogeneous data matrix, and then extract temperature change rate features, vibration and shock energy features, temperature and vibration cross-entropy features, and working condition load features to form a multi-dimensional feature set. The third unit is used to mine causal relationships among the feature variables in the multidimensional feature set using a causal inference model. By calculating the conditional independence and transit entropy between feature variables, a causal association network reflecting the fault evolution mechanism is constructed. The nodes of the causal association network represent feature variables, the directed edges represent the direction of causal influence, and the edge weights represent the causal strength coefficients. The fourth unit is used to divide the historical operating data of the bearing into multiple health state clusters based on the causal association network using a temporal clustering algorithm, and to model the transition rules between state clusters using a hidden Markov model to obtain the state transition matrix and the observation probability matrix. The temporal clustering algorithm determines the cluster boundary by minimizing the differences in the causal association network structure within the cluster. The fifth unit is used to decode the optimal state sequence from the initial healthy state to the target fault state using the Viterbi algorithm based on the state transition matrix and the observation probability matrix. The optimal state sequence is the fault evolution path.

8. An electronic device, characterized in that, include: processor; Memory used to store processor-executable instructions; The processor is configured to invoke instructions stored in the memory to execute the method according to any one of claims 1 to 6.

9. A computer-readable storage medium having computer program instructions stored thereon, characterized in that, When the computer program instructions are executed by the processor, they implement the method described in any one of claims 1 to 6.

Images