A method for discovering causal relationships in electromechanical equipment monitoring data based on hierarchical knowledge embedding

Abstract

This invention relates to the field of causal relationship discovery technology for electromechanical equipment monitoring data, specifically to a method for causal relationship discovery based on hierarchical knowledge embedding in electromechanical equipment monitoring data. The method includes: collecting electromechanical equipment monitoring data; establishing a hierarchical knowledge constraint matrix based on the knowledge information of the electromechanical equipment; establishing a causal discovery model; weighting and adjusting the causal strength calculated by the causal discovery model based on the hierarchical knowledge constraint matrix, or masking and modulating the attention matrix of the causal discovery model based on the hierarchical knowledge constraint matrix; processing the electromechanical equipment monitoring data based on the knowledge-embedded causal discovery model to obtain a causal network adjacency matrix; performing conflict edge detection based on the hierarchical knowledge constraint matrix, updating the causal network adjacency matrix, and using the updated causal network adjacency matrix as the causal relationship analysis result. This invention can improve the performance of causal relationship identification in monitoring data.

Description

Technical Field

[0001] This invention relates to the field of causal relationship discovery technology for electromechanical equipment monitoring data, and specifically to a method for causal relationship discovery based on hierarchical knowledge embedding in electromechanical equipment monitoring data. Background Technology

[0002] With the development of science and technology, various complex electromechanical equipment are widely used in modern society. As the required functions become increasingly diverse, systems are gradually becoming more sophisticated and complex, containing dozens or even hundreds of interconnected monitoring variables. Accurately identifying the causal relationships between these variables is crucial for understanding system operating mechanisms, diagnosing fault propagation paths, and achieving health status assessments.

[0003] In existing technologies, causal relationship discovery methods typically focus on purely data-driven statistical analysis, including Granger causality tests, neural network transfer entropy, and attention-enhanced causal modeling. Due to the high-dimensional nonlinearity of complex systems and the limited sample size, no purely data-driven method can completely avoid identifying spurious causality. Therefore, in practice, there are often numerous causal connections that violate physical laws. Furthermore, the search space of these methods grows exponentially with the number of variables, making it extremely difficult to identify causal relationships in complex systems using traditional purely data-driven methods. Even if identification is possible, the results are often unsatisfactory.

[0004] Therefore, there is an urgent need for a technical solution for constructing causal networks that integrates domain knowledge constraints. Summary of the Invention

[0005] In view of the above problems, the present invention provides a method for discovering causal relationships in electromechanical equipment monitoring data based on hierarchical knowledge embedding, which solves the technical problems of existing pure data-driven methods, such as large search space, many false causal relationships, and lack of domain knowledge constraints, which makes it difficult to identify causal relationships.

[0006] This invention provides a method for discovering causal relationships in electromechanical equipment monitoring data based on hierarchical knowledge embedding, comprising the following steps: Step S1: Collect monitoring data of electromechanical equipment. The monitoring data of electromechanical equipment includes multiple time point data arranged in sequence, and each time point data includes multiple monitoring variables. Step S2: Establish a hierarchical knowledge constraint matrix based on the knowledge information of the electromechanical equipment; the knowledge information of the electromechanical equipment includes the degree of mutual influence between two monitoring variables among the multiple monitoring variables; Step S3: Establish a causal discovery model; perform knowledge embedding, including: weighting and adjusting the causal strength calculated by the causal discovery model based on the hierarchical knowledge constraint matrix, or masking and modulating the attention matrix of the causal discovery model based on the hierarchical knowledge constraint matrix. Step S4: Based on the causal discovery model after knowledge embedding, process the monitoring data of electromechanical equipment to obtain the causal network adjacency matrix. The elements in the causal network adjacency matrix represent the causal strength between monitoring variables. Step S5: Detect conflict edges based on the hierarchical knowledge constraint matrix, update the causal network adjacency matrix, and use the updated causal network adjacency matrix as the result of causal relationship analysis.

[0007] Preferably, in step S1, the electromechanical equipment is a spacecraft attitude control system; the multiple monitoring variables specifically include: frame angular velocity command, frame angle, rotor speed pulse count, roll angle estimation, pitch angle estimation, yaw angle estimation, roll angular velocity estimation, pitch angular velocity estimation, yaw angular velocity estimation, yaw maneuver target angle, left solar panel rotation angle, right solar panel rotation angle, rotor motor current, fixed end bearing temperature, sliding end bearing temperature, and frame motor current.

[0008] Preferably, step S2 specifically includes: Step S2-1: Establish knowledge information for electromechanical equipment, including strong constraint knowledge, weak constraint knowledge, and exploratory constraint knowledge; the strong constraint knowledge, weak constraint knowledge, and exploratory constraint knowledge record the constraint types of any two monitored variables. Step S2-2: Determine the constraint type of any two monitoring variables based on the knowledge information of the electromechanical equipment, and determine the weight coefficient of the two monitoring variables based on the constraint type; use the weight coefficient of any two monitoring variables as elements of the hierarchical knowledge constraint matrix, and finally establish the hierarchical knowledge constraint matrix.

[0009] Preferably, in step S3, the causal discovery model is an information-based causal discovery model or a neural network-based causal discovery model. When the causal discovery model is based on information theory, candidate causal pair reconstruction and hierarchical enhancement factor adjustment are performed on the information theory-based causal discovery model. When the causal discovery model is based on a neural network, the neural network causal discovery model is subjected to soft mask modulation of the attention mechanism and adaptive fusion processing of the dual-branch architecture.

[0010] Preferably, in step S3, the steps of reconstructing candidate causal pairs and adjusting hierarchical enhancement factors in the information-theory-based causal discovery model specifically include: The candidate causal pair set of the information theory-based causal discovery model is screened based on the hierarchical knowledge constraint matrix, and causal strength is calculated only for variable pairs whose weight coefficients are greater than a preset threshold. A hierarchical enhancement factor is established to weight and adjust the causal strength calculation results. The expression is as follows:

[0011] in, For the original causal strength values ​​between the monitoring variables at positions i and j in the hierarchical knowledge constraint matrix, The enhanced causal strength values ​​between the monitoring variables at positions i and j in the hierarchical knowledge constraint matrix; The hierarchical enhancement factor between the monitoring variables at positions i and j in the hierarchical knowledge constraint matrix;

[0012] in, This represents the value at position i, j in the hierarchical knowledge constraint matrix. , , , To enhance the coefficient, satisfy > > > .

[0013] Preferably, in step S3, the step of performing soft masking modulation of the attention mechanism and adaptive fusion processing of the dual-branch architecture on the causal discovery model of the neural network specifically includes: The original attention matrix of the attention mechanism is subjected to soft mask modulation, expressed as:

[0014] Where M represents the hierarchical knowledge constraint matrix, This is the attention matrix after soft masking. This is the original attention matrix; The attention matrix after soft masking is renormalized, as shown in the expression:

[0015] Where A is the modified attention matrix, and ε is a small constant; In a neural network model, feature fusion between constrained guided branches and free exploration branches is expressed as follows:

[0016]

[0017] in, This represents the fusion feature of the i-th monitored variable. Let i be the fusion weight of the i-th monitoring variable. and These are the output characteristics of the constrained guided branch and the free exploration branch, respectively. This indicates the total number of monitored variables.

[0018] Preferably, step S4 specifically includes: Step S4-1: Preprocess the monitoring data of electromechanical equipment, including standardization, detrending and time alignment. Step S4-2: Input the preprocessed monitoring data into the causal discovery model after knowledge embedding, and calculate the causal strength value between variable pairs; Step S4-3: Binarize the causal strength value according to the preset significance threshold to obtain the causal network adjacency matrix.

[0019] Preferably, step S5 specifically includes: Step S5-1: Compare the causal network adjacency matrix with the hierarchical knowledge constraint matrix, and mark the edges where the values ​​in the causal network adjacency matrix and the values ​​in the hierarchical knowledge constraint matrix conflict as conflict edges. Step S5-2: Classify the conflict edges according to their physical rationality. The physical rationality classification specifically includes: time sequence violation, knowledge base incompleteness, conditional dependency causality, and unverified. Step S5-3: Remove conflicting edges of the temporal violation class; for conflicting edges of the knowledge base incomplete class, update the values ​​in the causal network adjacency matrix with weak constraint weights.

[0020] Preferably, step S5-1 specifically includes: If the values ​​at positions i and j in the adjacency matrix of the causal network are equal to 1 and the values ​​at positions i and j in the corresponding hierarchical knowledge constraint matrix are less than a preset threshold, then the edge formed by the i-th and j-th monitoring variables is marked as a conflict edge.

[0021] Preferably, step S5 further includes: The network topology features are calculated based on the adjacency matrix of the causal network. These features include: network density, network entropy, inter-layer connection density ratio, node degree distribution, PageRank centrality, betweenness centrality, and clustering coefficient.

[0022] Compared with the prior art, the present invention has at least the following beneficial effects: (1) By systematically integrating multi-level domain knowledge into the causal discovery process through a hierarchical knowledge constraint matrix, the differences in the certainty levels of knowledge are fully considered, so that the constructed causal network has both physical interpretability and data adaptability. (2) By designing soft weight functions and differentiating embedding strategies, the data-driven exploration capability is preserved while the generation of false causality is effectively suppressed, which significantly improves the accuracy of causal relationship identification and the coverage of strong constraints. (3) By using candidate domain reconstruction and hierarchical enhancement factor mechanism, the causal search space is greatly reduced, the computational efficiency is significantly improved, and the accuracy of identifying true causal relationships is guaranteed. (4) Through the physical rationality classification mechanism of conflict edges, dynamic updates of the domain knowledge base and data-driven knowledge discovery are realized, and a virtuous cycle feedback mechanism of knowledge and data is constructed. (5) By using the Bayesian knowledge-data fusion framework, prior knowledge and data likelihood are unified into the probability inference system, providing a theoretical basis for the quantification of uncertainty in causal networks. Attached Figure Description

[0023] The accompanying drawings are for illustrative purposes only and are not intended to limit the scope of the invention.

[0024] Figure 1 The flowchart shows the causal relationship discovery method for electromechanical equipment monitoring data based on hierarchical knowledge embedding provided by the present invention.

[0025] Figure 2 A detailed flowchart of the causal relationship discovery method for electromechanical equipment monitoring data based on hierarchical knowledge embedding provided by the present invention.

[0026] Figure 3 The HKE framework three-layer constraint architecture diagram provided by this invention.

[0027] Figure 4 This is a structural diagram of the AINN model provided by the present invention.

[0028] Figure 5 A schematic diagram of the Bayesian knowledge-data fusion framework provided by this invention.

[0029] Figure 6 A comparison chart of knowledge coverage and conflict rate for the various methods provided in this invention.

[0030] Figure 7 A comparison chart of F1 scores for the various methods provided in this invention.

[0031] Figure 8 The physical rationality classification and distribution diagram of conflict edges provided by this invention. Detailed Implementation

[0032] To better understand the above-described objectives, features, and advantages of the present invention, the invention will be further described in detail below with reference to the accompanying drawings and specific embodiments. It should be noted that, unless otherwise specified, the embodiments of the present invention and the features thereof can be combined with each other. Furthermore, the present invention can be implemented in other ways different from those described herein; therefore, the scope of protection of the present invention is not limited to the specific embodiments disclosed below.

[0033] To illustrate the effectiveness of the method proposed in this invention, the following detailed description of the above technical solution is provided through a specific embodiment, such as... Figure 1 As shown, a method for discovering causal relationships in electromechanical equipment monitoring data based on hierarchical knowledge embedding is disclosed. The specific implementation steps are as follows: Step S1: Collect monitoring data of electromechanical equipment. The monitoring data of electromechanical equipment includes multiple time point data arranged in sequence, and each time point data includes multiple monitoring variables. In this step, the present invention first collects monitoring data from electromechanical equipment. The electromechanical equipment in this invention can be a spacecraft attitude control system, an aircraft engine, a nuclear power plant, a high-speed train, etc. The monitoring data includes multiple time-point data arranged sequentially, and each time-point data includes multiple monitoring variables. These monitoring variables reflect the state parameters of the electromechanical equipment during operation.

[0034] In one specific embodiment, the present invention uses the spacecraft attitude control system as the system to be evaluated. The operational monitoring data of the spacecraft attitude control system can be collected through feedback signals from various sensors and actuators. For the spacecraft attitude control system, the specific monitoring variables are shown in Table 1.

[0035] Table 1

[0036] The above-mentioned monitored variables cover multiple aspects such as command signals, status feedback, actuator response, and thermal characteristics of the spacecraft attitude control system.

[0037] In another specific embodiment, the present invention uses an aero-engine as the system to be evaluated, and the specific monitoring variables are shown in Table 2.

[0038] Table 2

[0039] The monitoring data of the aero-engine at all collected time points are used as the monitoring data of the aero-engine, as shown in Table 3: Table 3

[0040] As shown in Table 3, the time points are sequentially arranged from time point 1 to time point 20631, totaling 20631 sets of monitoring data, which serve as the monitoring data for this aero-engine. The initial quantity can be reasonably set according to actual conditions; here, the initial quantity is 18. Each row of data represents the complete operating state of the engine at a certain moment, and the temporal arrangement of multiple rows of data reflects the evolution of the engine's operating state over time.

[0041] Step S2: Establish a hierarchical knowledge constraint matrix based on the knowledge information of the electromechanical equipment; the knowledge information of the electromechanical equipment includes the degree of mutual influence between two monitoring variables among the multiple monitoring variables; In this step, the present invention establishes a hierarchical knowledge constraint matrix based on the knowledge information of electromechanical equipment. The present invention first establishes the knowledge information of electromechanical equipment, which includes strongly constrained knowledge, weakly constrained knowledge, and exploratory constrained knowledge, including the following steps: Obtain the system's physical model document, control block diagram, and historical fault records; The relationships between monitoring variables determined by physical laws such as energy conservation, mass conservation, and momentum conservation in the physical model document are used as strongly constrained knowledge. The relationships between monitoring variables determined by the deterministic link relationships of the control loops in the control block diagram are treated as strongly constrained knowledge. The relationship between the fault propagation path in historical fault records and the monitoring variables determined by empirical rules is used as weakly constrained knowledge. The relationships between monitoring variables, determined by the connections between modules in the system topology, are used as exploratory constraint knowledge.

[0042] The following specific example describes the process of acquiring knowledge information for an aero-engine system.

[0043] (1) Relationship between monitoring variables related to physical laws For aero-engine systems, for example: according to the law of conservation of energy, the fuel flow rate Wf inevitably affects the high-pressure turbine outlet temperature T48; according to the law of conservation of mass, the high-pressure compressor outlet pressure Ps30 inevitably affects the combustion chamber outlet pressure P40; according to the law of conservation of momentum, the core rotational speed Nc inevitably affects the high-pressure turbine outlet temperature T48. These examples represent strongly constrained knowledge.

[0044] (2) Relationship of monitoring variables related to the control block diagram For example: the throttle valve resolution angle TRA inevitably affects the fuel flow rate Wf through the control system; the fuel flow rate Wf inevitably affects the high-pressure turbine outlet temperature T48 through the combustion process; the core speed Nc inevitably affects the throttle valve resolution angle TRA through the feedback loop. The above examples represent strongly constrained knowledge.

[0045] (3) Relationship of monitoring variables related to historical fault records For example, according to historical maintenance records, an abnormal high-pressure turbine outlet temperature T48 is often accompanied by a change in the low-pressure turbine outlet temperature T50; an abnormal high-pressure compressor outlet static pressure Ps30 may affect the low-pressure compressor outlet pressure P24. These examples illustrate weak constraint knowledge.

[0046] (4) Relationship of monitoring variables related to system topology The connection relationships between modules are determined based on the system topology. As shown in Table 4, the 18 monitored variables are divided according to functional modules.

[0047] Table 4

[0048] For the functional modules mentioned above, for example, variables in the fan module may affect variables in the low-pressure compressor module, and variables in the low-pressure compressor module may affect variables in the high-pressure compressor module. The above examples illustrate the exploration of constraint knowledge.

[0049] Subsequently, this invention establishes a hierarchical knowledge constraint matrix based on the knowledge information of electromechanical equipment, specifically including: Set both rows and columns of the hierarchical knowledge constraint matrix to the total number of monitored variables; Based on the knowledge information of electromechanical equipment, determine the constraint type of any two monitoring variables, and determine the weight coefficient of the two monitoring variables based on the constraint type; use the weight coefficient of any two monitoring variables as the elements of the hierarchical knowledge constraint matrix, and finally establish the hierarchical knowledge constraint matrix.

[0050] The step of determining the constraint type of any two monitored variables based on the knowledge information of electromechanical equipment includes: determining which of the following types of constraint knowledge—strong constraint knowledge, weak constraint knowledge, and exploratory constraint knowledge—the two monitored variables conform to, and thus obtaining the corresponding constraint type. That is, if they conform to strong constraint knowledge, the constraint type is determined to be strong constraint; if they conform to weak constraint knowledge, the constraint type is determined to be weak constraint; and if they conform to exploratory constraint knowledge, the constraint type is determined to be exploratory constraint.

[0051] This invention sets corresponding weight coefficients for each type of constraint, with the weight coefficients decreasing sequentially for strong constraints, weak constraints, and exploratory constraints.

[0052] The following specific example illustrates the process of establishing a hierarchical knowledge constraint matrix for a spacecraft attitude control system. First, the constraint type for any two monitored variables is determined based on the knowledge information of the spacecraft attitude control system. Then, the weight coefficients of the two monitored variables are determined by the constraint type, as illustrated in Table 5, which shows some examples.

[0053] Table 5

[0054] Table 5 illustrates the weight coefficients of two monitored variables. These weight coefficients are used as elements of the hierarchical knowledge constraint matrix to ultimately establish the hierarchical knowledge constraint matrix of the spacecraft attitude control system, as shown in Table 6.

[0055] Table 6

[0056] Through the hierarchical setting of the weighting coefficients described above, the hierarchical knowledge constraint matrix of this invention fully quantifies the degree of prior support of domain knowledge for the causal relationships of each variable. This matrix will play a core constraint and guiding role in the subsequent knowledge embedding process of the causal discovery model, enabling the causal discovery process to fully utilize domain knowledge to reduce the search space and suppress false causality, while retaining the data-driven exploration capability to discover potential causal relationships not explicitly included in the knowledge base.

[0057] Figure 2 and Figure 3 It illustrates the hierarchical relationship between the strong constraint layer (physical laws + control block diagram), the weak constraint layer (empirical formulas + fault paths), and the exploratory constraint layer (topology connections).

[0058] Step S3: Establish a causal discovery model; perform knowledge embedding, including: weighting and adjusting the causal strength calculated by the causal discovery model based on the hierarchical knowledge constraint matrix, or masking and modulating the attention matrix of the causal discovery model based on the hierarchical knowledge constraint matrix. In this step, the present invention establishes a causal discovery model and embeds corresponding knowledge into different causal discovery models, as described in detail below.

[0059] The causal discovery model of this invention can be an information-theoretic causal discovery model or a neural network-based causal discovery model.

[0060] When the causal discovery model is an information theory-based causal discovery model, this invention performs candidate causal pair reconstruction and hierarchical enhancement factor adjustment processing on the information theory-based causal discovery model.

[0061] The causal discovery model based on information theory can adopt existing causal discovery models in this field, such as the transfer entropy method, mutual information method, conditional entropy method, and relative entropy method. This invention does not limit these basic information theory-based causal discovery models.

[0062] This invention adjusts the candidate causal pairs input to the information theory-based causal discovery model and the calculated causal strength output, as detailed below.

[0063] (1) Reconstruction of candidate causal pairs In this step, the present invention filters the candidate causal pair set of the information theory-based causal discovery model according to the hierarchical knowledge constraint matrix, and only calculates the causal strength of variable pairs with weight coefficients greater than a preset threshold.

[0064] In a specific embodiment, for an aero-engine system, the number of complete candidate causal pairs is 18 × 17 = 306. After candidate domain reconstruction, the number of candidate causal pairs can be reduced to below 100, and the search space is significantly reduced.

[0065] (2) Adjustment of stratified enhancement factors In this step, a hierarchical enhancement factor is established to weight and adjust the causal strength calculation results. The expression is as follows:

[0066] in, For the original causal strength values ​​between the monitoring variables at positions i and j in the hierarchical knowledge constraint matrix, The enhanced causal strength values ​​between the monitoring variables at positions i and j in the hierarchical knowledge constraint matrix; This represents the hierarchical enhancement factor between the monitoring variables at positions i and j in the hierarchical knowledge constraint matrix.

[0067]

[0068] in, This represents the value at position i, j in the hierarchical knowledge constraint matrix. , , , To enhance the coefficient, satisfy > > > In a specific example, =1.2, =1.0, =0.8, =0.6.

[0069] Specifically, such as Figure 5 The diagram illustrates a causal discovery model based on information theory. This framework uses domain knowledge as a prior distribution and data observations as a likelihood function, obtaining the posterior distribution of causal relationships through Bayesian inference. This framework is applicable to any causal discovery model, providing a universal theoretical foundation for knowledge and data fusion.

[0070] When the causal discovery model is a neural network-based causal discovery model, this invention performs soft mask modulation of the attention mechanism and adaptive fusion processing of the dual-branch architecture on the neural network causal discovery model.

[0071] The causal discovery model based on neural networks can adopt existing causal discovery models in the field, such as attention mechanism methods, Transformer methods, and graph neural network methods. This invention does not limit these basic neural network-based causal discovery models.

[0072] This invention adjusts the attention matrix and fusion weights of a neural network-based causal discovery model, as detailed below.

[0073] (1) Soft mask modulation of attention mechanism In this step, the present invention introduces a soft masking mechanism to achieve knowledge embedding in the attention calculation process of the neural network-based causal discovery model.

[0074] The formula for calculating the attention mechanism is:

[0075] Where Q is the query matrix and K is the key matrix. The dimension of the key vector. This represents the softmax activation function. This is the original attention matrix.

[0076] Using the hierarchical knowledge constraint matrix as a soft mask matrix, the original attention matrix is ​​modulated, as expressed by:

[0077] Where M represents the hierarchical knowledge constraint matrix, This is the attention matrix after soft masking.

[0078] The attention matrix after soft masking is renormalized, as shown in the expression:

[0079] Where A is the final attention matrix, and ε is a small constant, usually set to 1e-8.

[0080] Through the above steps, this invention performs element-wise product modulation on the original attention matrix calculated by the attention mechanism, thereby integrating domain knowledge into the feature extraction process of the neural network.

[0081] (2) Adaptive fusion of dual-branch architecture In this step, the invention constructs a dual-branch architecture, building two parallel feature extraction paths within the neural network model. The first branch is a constraint-guided branch, using the constrained attention matrix for feature extraction; the second branch is a free exploration branch, using the original attention matrix for feature extraction. The constraint-guided branch fully utilizes prior information from domain knowledge, focusing on causal relationships with high certainty in the knowledge base; the free exploration branch relies entirely on a data-driven learning process, capable of discovering potential causal relationships not included in the knowledge base. The outputs of the two branches are adaptively fused using variable-level fusion weights, which are determined based on the weight distribution of each variable in the hierarchical knowledge constraint matrix.

[0082] The outputs of the two branches are adaptively fused using variable-level fusion weights, expressed as follows:

[0083]

[0084] in, This represents the fusion feature of the i-th monitored variable. Let i be the fusion weight of the i-th monitoring variable. and These are the output characteristics of the constrained guided branch and the free exploration branch, respectively. This indicates the total number of monitored variables.

[0085] Specifically, such as Figure 4 The diagram illustrates the architecture of a neural network-based causal discovery model, comprising an input layer, a feature extraction layer, an attention mechanism layer, hidden layers, and an output layer. This architecture can be generalized to other attention-based neural network models.

[0086] Step S4: Based on the causal discovery model after knowledge embedding, process the monitoring data of electromechanical equipment to obtain the causal network adjacency matrix. The elements in the causal network adjacency matrix represent the causal strength between monitoring variables. Step S4-1: Preprocess the monitoring data of electromechanical equipment, including standardization, detrending and time alignment. In this step, the present invention first preprocesses the monitoring data of electromechanical equipment, including standardization, detrending and time alignment.

[0087] Standardization can employ methods such as, but is not limited to, Z-score standardization, maximum-minimum normalization, and robust standardization. In a preferred embodiment, Z-score standardization is used:

[0088] Where X[i,t] is the original value of the i-th variable at time t, and μ[i] and σ[i] are the mean and standard deviation of the variable, respectively.

[0089] Specifically, detrending processing can employ, but is not limited to, first-order differencing, polynomial fitting, and filter methods. In a preferred embodiment, first-order differencing is used:

[0090] Specifically, time alignment ensures that the sampling times of all variables are consistent. For missing values, methods such as linear interpolation, spline interpolation, forward padding, and backward padding can be used to fill in the missing values.

[0091] Step S4-2: Input the preprocessed monitoring data into the causal discovery model after knowledge embedding, and calculate the causal strength value between variable pairs; In this step, the preprocessed monitoring data is input into the causal discovery model after knowledge embedding. For information theory-based models, each variable pair ( → The enhanced causal strength value CS[i,j] is obtained; for neural network-based models, the causal strength matrix is ​​calculated through model forward propagation.

[0092] Step S4-3: Binarize the causal strength value according to the preset significance threshold to obtain the causal network adjacency matrix.

[0093] Binarization is performed based on a preset significance threshold. This threshold can be determined using statistical significance tests, such as a permutation test to generate a null hypothesis distribution, with the upper bound of the confidence interval at a preset confidence level (e.g., 95% or 99%) as the threshold. If the causal strength value CS[i,j] is greater than the threshold τ, then A[i,j] = 1; otherwise, A[i,j] = 0. The resulting binarized matrix A is the adjacency matrix of the causal network.

[0094] Step S5: Detect conflict edges based on the hierarchical knowledge constraint matrix, update the causal network adjacency matrix, and use the updated causal network adjacency matrix as the result of causal relationship analysis.

[0095] In this step, the present invention compares and analyzes the causal network adjacency matrix and the hierarchical knowledge constraint matrix obtained in the preceding steps, identifies causal connections that are discovered through data-driven discovery but conflict with domain knowledge, i.e. conflict edges, classifies conflict edges according to their physical rationality, and removes conflict edges that violate the time sequence. The details are described below.

[0096] Step S5-1: Compare the causal network adjacency matrix with the hierarchical knowledge constraint matrix, and mark the edges where the values ​​in the causal network adjacency matrix and the values ​​in the hierarchical knowledge constraint matrix conflict as conflict edges. In this step, the rule for identifying conflict edges in this invention is as follows: if the value of the causal network adjacency matrix A[i,j]=1 and the value of the hierarchical knowledge constraint matrix M[i,j]<a preset threshold, then (Xi→Xj) is marked as a conflict edge. In a preferred embodiment, the preset threshold is set to 0.5, that is, when the prior weight corresponding to the causal connection discovered by the data is low, a conflict is considered to exist.

[0097] Step S5-2: Classify the conflict edges according to their physical rationality. The classification specifically includes: time sequence violation, incomplete knowledge base, conditional dependency causality, and unverified. The following describes the classification of the physical rationality of conflict edges through a specific embodiment. In one specific embodiment, the classification rules for the physical rationality of conflict edges are shown in Table 7.

[0098] Table 7

[0099] Specifically, timing violation-type conflict edges refer to situations where their causal direction clearly violates the known physical timing relationships of the system. Judgment conditions include, but are not limited to: When the direction of causal connection is inconsistent with the objective sequence of material transfer in the system, it is determined that the airflow direction is reversed. When the direction of causal connection is inconsistent with the basic process of energy input-conversion-output in the system, it is determined that the energy flow is reversed; When the direction of the causal connection is inconsistent with the basic control logic in the control system where the command signal points from the control unit to the execution unit, it is determined to be a reverse control signal.

[0100] For any causal connection that satisfies any of the above conditions, it is determined to violate basic physical or control temporal relations and is classified as a temporal violation conflict edge.

[0101] Specifically, incomplete knowledge base conflict edges refer to situations where the causal connection does not violate the system's basic physical temporal relationships, but the current domain knowledge base does not explicitly contain this causal relationship. The criteria for determining this include, but are not limited to: The source and target variables of the causal connection have a reasonable module relationship in the system's functional structure; The causal connection conforms to the general path characteristics of the transfer of matter, energy or information in the system in terms of direction; The causal connection does not violate fundamental physical laws such as energy conservation and mass conservation, and has a reasonable explanation from an engineering perspective.

[0102] Causal connections that meet the above conditions are identified as conflict edges in the knowledge base incompleteness class and are retained as potential valid causal relationships.

[0103] Specifically, conditionally dependent causal conflict edges refer to situations where the causal connection does not hold in all operating states, but only exhibits a significant causal relationship under specific conditions. The criteria for determining this include, but are not limited to: This causal connection only becomes significant under specific operating conditions; This causal connection is only valid within a specific time interval during system operation; The significance of this causal connection is related to operating conditions, control status, or load level.

[0104] For causal connections that meet the above conditions, they are identified as conditional dependent causal classes, and the corresponding triggering condition information is recorded.

[0105] Step S5-3: Remove conflicting edges of the temporal violation class; for conflicting edges of the knowledge base incomplete class, update the values ​​in the causal network adjacency matrix with weak constraint weights.

[0106] The following is a description of step S5-3 of a specific embodiment.

[0107] Different approaches are taken for different types of conflict edges. For type A conflict edges in Table 7, they are directly deleted from the causal network adjacency matrix, i.e., A[i,j]=0. For type B conflict edges, the causal connection is added to the domain knowledge base and assigned an appropriate weight coefficient, usually set to the weight of the weak constraint layer, such as between 0.6 and 0.7. For type C conflict edges, the triggering conditions are marked in the domain knowledge base, such as "valid under high-altitude cruise conditions" or "valid under high load operation". For type D conflict edges, they are retained in the causal network but marked as "to be verified", awaiting further engineering analysis or experimental verification.

[0108] Finally, the updated causal network adjacency matrix is ​​used as the result of the causal relationship analysis.

[0109] like Figure 6 As shown, the comparison of knowledge coverage and conflict rate for each method is presented. Figure 5 The horizontal axis represents different causal discovery methods, and the vertical axis represents knowledge coverage (green bars) and conflict rate (red bars). It can be seen that the method using a knowledge embedding strategy has significant advantages over the purely data-driven method in terms of both knowledge coverage and conflict rate control.

[0110] like Figure 7 The figure shows a comparison of the F1 scores of various methods. The F1 score takes into account both precision and recall in causal identification; a higher score indicates better method performance. Figure 6The study compared F1 scores under two evaluation criteria: a strict criteria (requiring a perfect match) and a lenient criteria (allowing partial matches). The method using the knowledge embedding strategy performed well under both criteria.

[0111] like Figure 8 As shown, a distribution map illustrating the physical rationality classification of conflict edges is presented. Figure 7 The pie chart clearly shows the proportion of conflict edges in four categories: A (temporal violation), B (incomplete knowledge base), C (conditional dependency causality), and D (to be verified). This distribution chart helps to understand the relationship between data-driven discovery and domain knowledge, and guides the optimization and improvement of the knowledge base.

[0112] In some embodiments, after the causal network is constructed, network topology features can be calculated based on the causal network adjacency matrix for applications such as system health assessment, fault diagnosis, and anomaly detection. Network topology features include, but are not limited to, network density, network entropy, inter-layer connection density ratio, node degree distribution, PageRank centrality, betweenness centrality, and clustering coefficient.

[0113] Compared with existing technologies, the method for constructing causal networks for complex systems based on hierarchical knowledge embedding provided in this invention systematically integrates multi-level domain knowledge into the causal discovery process through a hierarchical knowledge constraint matrix. It fully considers the differences in the certainty levels of knowledge, enabling the constructed causal network to possess both physical interpretability and data adaptability. Furthermore, through soft weight function design and differentiated embedding strategies, it retains data-driven exploration capabilities while effectively suppressing the generation of false causality, significantly improving the accuracy of causal relationship identification and strong constraint coverage. Simultaneously, through candidate domain reconstruction and hierarchical enhancement factor mechanisms, it significantly reduces the causal search space, significantly improves computational efficiency, and ensures the accuracy of identifying true causal relationships. Finally, through the physical rationality classification mechanism of conflict edges, it realizes dynamic updates of the domain knowledge base and data-driven knowledge discovery, constructing a virtuous cycle feedback mechanism between knowledge and data.

[0114] While the specific embodiments of the present invention depict actions or steps in a particular order, this should be understood as requiring such actions or steps to be performed in the shown specific order or sequential order, or requiring all illustrated actions or steps to be performed to achieve the desired result. In certain environments, multitasking and parallel processing may be advantageous. Similarly, although several specific implementation details are included in the above discussion, these should not be construed as limiting the scope of this disclosure. Certain features described in the context of individual embodiments may also be implemented in combination in a single implementation. Conversely, various features described in the context of a single implementation may also be implemented individually or in any suitable sub-combination in multiple implementations. The above descriptions are merely preferred embodiments of the present invention, but the scope of protection of the present invention is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the technical scope disclosed in the present invention should be included within the scope of protection of the present invention.

[0115] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any changes or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in the present invention should be included within the scope of protection of the present invention.

Claims

1. A method for discovering causal relationships in electromechanical equipment monitoring data based on hierarchical knowledge embedding, characterized in that, Includes the following steps: Step S1: Collect monitoring data of electromechanical equipment. The monitoring data of electromechanical equipment includes multiple time point data arranged in sequence, and each time point data includes multiple monitoring variables. Step S2: Establish a hierarchical knowledge constraint matrix based on the knowledge information of the electromechanical equipment; the knowledge information of the electromechanical equipment includes the degree of mutual influence between two monitoring variables among the multiple monitoring variables; Step S3: Establish a causal discovery model; Knowledge embedding includes: weighting and adjusting the causal strength calculated by the causal discovery model based on the hierarchical knowledge constraint matrix, or masking and modulating the attention matrix of the causal discovery model based on the hierarchical knowledge constraint matrix; Step S4: Process the electromechanical equipment monitoring data based on the causal discovery model after knowledge embedding to obtain the causal network adjacency matrix. The elements in the causal network adjacency matrix represent the causal strength between monitoring variables. Step S5: Detect conflict edges based on the hierarchical knowledge constraint matrix, update the causal network adjacency matrix, and use the updated causal network adjacency matrix as the result of causal relationship analysis.

2. The method for discovering causal relationships in electromechanical equipment monitoring data based on hierarchical knowledge embedding as described in claim 1, characterized in that, In step S1, the electromechanical equipment is a spacecraft attitude control system; the multiple monitoring variables specifically include: frame angular velocity command, frame angle, rotor speed pulse count, roll angle estimation, pitch angle estimation, yaw angle estimation, roll angular velocity estimation, pitch angular velocity estimation, yaw angular velocity estimation, yaw maneuver target angle, left solar panel rotation angle, right solar panel rotation angle, rotor motor current, fixed end bearing temperature, sliding end bearing temperature, and frame motor current.

3. The method for discovering causal relationships in electromechanical equipment monitoring data based on hierarchical knowledge embedding as described in claim 2, characterized in that, Step S2 specifically includes: Step S2-1: Establish knowledge information for electromechanical equipment, including strong constraint knowledge, weak constraint knowledge, and exploratory constraint knowledge; the strong constraint knowledge, weak constraint knowledge, and exploratory constraint knowledge record the constraint types of any two monitored variables. Step S2-2: Determine the constraint type of any two monitoring variables based on the knowledge information of the electromechanical equipment, and determine the weight coefficient of the two monitoring variables based on the constraint type; use the weight coefficient of any two monitoring variables as elements of the hierarchical knowledge constraint matrix, and finally establish the hierarchical knowledge constraint matrix.

4. The method for discovering causal relationships in electromechanical equipment monitoring data based on hierarchical knowledge embedding as described in claim 3, characterized in that, In step S3, the causal discovery model is either an information-based causal discovery model or a neural network-based causal discovery model. When the causal discovery model is based on information theory, candidate causal pair reconstruction and hierarchical enhancement factor adjustment are performed on the information theory-based causal discovery model. When the causal discovery model is based on a neural network, the neural network causal discovery model is subjected to soft mask modulation of the attention mechanism and adaptive fusion processing of the dual-branch architecture.

5. The method for discovering causal relationships in electromechanical equipment monitoring data based on hierarchical knowledge embedding according to claim 4, characterized in that, Step S3, specifically including the steps of reconstructing candidate causal pairs and adjusting hierarchical enhancement factors in the information-theory-based causal discovery model, includes: The candidate causal pair set of the information theory-based causal discovery model is screened based on the hierarchical knowledge constraint matrix, and causal strength is calculated only for variable pairs whose weight coefficients are greater than a preset threshold. A hierarchical enhancement factor is established to weight and adjust the causal strength calculation results. The expression is as follows: in, For the original causal strength values ​​between the monitoring variables at positions i and j in the hierarchical knowledge constraint matrix, The enhanced causal strength values ​​between the monitoring variables at positions i and j in the hierarchical knowledge constraint matrix; The hierarchical enhancement factor between the monitoring variables at positions i and j in the hierarchical knowledge constraint matrix; in, This represents the value at position i, j in the hierarchical knowledge constraint matrix. , , , To enhance the coefficient, satisfy > > > .

6. The method for discovering causal relationships in electromechanical equipment monitoring data based on hierarchical knowledge embedding according to claim 4, characterized in that, In step S3, the step of performing soft masking modulation of the attention mechanism and adaptive fusion processing of the dual-branch architecture on the causal discovery model of the neural network specifically includes: The original attention matrix of the attention mechanism is subjected to soft mask modulation, expressed as: Where M represents the hierarchical knowledge constraint matrix, This is the attention matrix after soft masking. This is the original attention matrix; The attention matrix after soft masking is renormalized, as shown in the expression: Where A is the modified attention matrix. ε It is a small constant; In a neural network model, feature fusion between constrained guided branches and free exploration branches is expressed as follows: in, Indicates the first i The fusion characteristics of the monitored variables For the first i The fusion weights of the monitoring variables, and These are the output characteristics of the constrained guided branch and the free exploration branch, respectively. This indicates the total number of monitored variables.

7. The method for discovering causal relationships in electromechanical equipment monitoring data based on hierarchical knowledge embedding according to claim 6, characterized in that, Step S4 specifically includes: Step S4-1: Preprocess the monitoring data of electromechanical equipment, including standardization, detrending and time alignment. Step S4-2: Input the preprocessed monitoring data into the causal discovery model after knowledge embedding, and calculate the causal strength value between variable pairs; Step S4-3: Binarize the causal strength value according to the preset significance threshold to obtain the causal network adjacency matrix.

8. The method for discovering causal relationships in electromechanical equipment monitoring data based on hierarchical knowledge embedding according to claim 7, characterized in that, Step S5 specifically includes: Step S5-1: Compare the causal network adjacency matrix with the hierarchical knowledge constraint matrix, and mark the edges where the values ​​in the causal network adjacency matrix and the values ​​in the hierarchical knowledge constraint matrix conflict as conflict edges. Step S5-2: Classify the conflict edges according to their physical rationality. The physical rationality classification specifically includes: time sequence violation, knowledge base incompleteness, conditional dependency causality, and unverified. Step S5-3: Remove conflicting edges of the temporal violation class; for conflicting edges of the knowledge base incomplete class, update the values ​​in the causal network adjacency matrix with weak constraint weights.

9. The method for discovering causal relationships in electromechanical equipment monitoring data based on hierarchical knowledge embedding according to claim 8, characterized in that, Step S5-1 specifically includes: If the values ​​at positions i and j in the adjacency matrix of the causal network are equal to 1 and the values ​​at positions i and j in the corresponding hierarchical knowledge constraint matrix are less than a preset threshold, then the edge formed by the i-th and j-th monitoring variables is marked as a conflict edge.

10. The method for discovering causal relationships in electromechanical equipment monitoring data based on hierarchical knowledge embedding according to claim 9, characterized in that, Step S5 also includes: The network topology features are calculated based on the adjacency matrix of the causal network. These features include: network density, network entropy, inter-layer connection density ratio, node degree distribution, PageRank centrality, betweenness centrality, and clustering coefficient.

Images