Dynamic evaluation method based on defect trend change

Abstract

The present application relates to the technical field of defect trend analysis, and discloses a dynamic evaluation method based on defect trend change, comprising the following steps: S101, collecting sensor data, equipment state data, environment data and historical maintenance data in a converter station system, and representing the data through first feature data; S102, pre-processing the first feature data to obtain second feature data; S103, constructing a Bayesian network topology graph according to the second feature data; S104, constructing a Bayesian network model based on a PC algorithm according to the Bayesian network topology graph, and outputting probability values of various defects of the equipment; and S105, generating treatment measures according to the probability values of various defects of the equipment output by the Bayesian network model and in combination with a preset defect grade classification standard. The present application dynamically adjusts the Bayesian network structure through the PC algorithm, combines multi-source data collection and closed-loop management, and ensures the accuracy, adaptability and management integrity of defect prediction.

Description

Technical Field

[0001] This invention belongs to the field of defect trend analysis technology, specifically relating to a dynamic evaluation method based on defect trend changes. Background Technology

[0002] In the operation and management of intelligent converter stations, timely and accurate identification and prediction of equipment defects are crucial for ensuring the safe and stable operation of the system. Equipment defects not only affect the normal operation of the system but may also trigger wider power system failures. Therefore, rapid and effective defect identification and prediction can significantly improve maintenance foresight, reduce downtime and maintenance costs caused by failures, and ensure the stability and reliability of the power grid.

[0003] Currently, defect identification and prediction mainly rely on the integrated platform of the converter station. This platform uses multi-source data acquisition and deep learning algorithms to construct a "black box model" for identifying potential defects and predicting trends, achieving automated monitoring and early warning. However, this method faces many problems in practical applications. First, deep learning models are complex and difficult to solve, and the prediction results are uncontrollable and uninterpretable, increasing the uncertainty of safety hazards and fault management. Second, the model's generalization ability is weak, relying on the assumption that training and testing data are identically distributed. However, in the complex and ever-changing environment of the converter station, the prediction results are difficult to maintain stability, limiting its application effectiveness in diverse scenarios. Furthermore, this method lacks further decision support for identified defects, lacks hierarchical management, hazard investigation, and time-varying trend analysis, and fails to comprehensively reflect the equipment's operating status.

[0004] Therefore, there is an urgent need for a dynamic evaluation method based on the changes in defect trends to solve the above problems. Summary of the Invention

[0005] This invention provides a dynamic evaluation method based on defect trend changes, which solves the technical problems of related technologies that lack interpretability, have weak generalization ability and lack comprehensive decision support, making it difficult to adapt to complex actual operating environments.

[0006] This invention provides a dynamic evaluation method based on defect trend changes, comprising the following steps:

[0007] Step S101: Collect sensor data, equipment status data, environmental data, and historical maintenance data from the converter station system, and represent them using the first feature data;

[0008] The sensor data includes: equipment temperature, equipment current, equipment pressure, and equipment vibration; the equipment status data includes: equipment operating status and equipment load data; the environmental data includes: ambient temperature and ambient humidity; and the historical maintenance data includes: maintenance time interval and maintenance frequency.

[0009] Step S102: Preprocess the first feature data to obtain the second feature data;

[0010] Step S103: Construct a Bayesian network topology graph based on the second feature data. The Bayesian network topology graph includes: nodes, directed edges, and conditional probability tables of nodes.

[0011] Step S104: Construct a Bayesian network model based on the PC algorithm according to the Bayesian network topology diagram. The input of the Bayesian network model is the second feature data, and the output is the probability value of various defects in the equipment.

[0012] The types of defects that occur in equipment include: over-temperature defects, abnormal current defects, abnormal vibration defects, and abnormal pressure defects.

[0013] Step S105: Based on the probability values ​​of various defects in the device output by the Bayesian network model and the preset defect level classification standards, generate handling measures.

[0014] Furthermore, the device operating status includes: normal, standby, and overload. The device operating status is represented by integer encoding, and the device load data is represented by real number encoding.

[0015] Furthermore, the preprocessing steps for the first feature data include:

[0016] Step S201: Use the 3σ criterion to detect outliers in sensor data and environmental data, and replace the outliers with the median;

[0017] Step S202: The sensor data and environmental data that have undergone outlier replacement processing are smoothed using the sliding time window method;

[0018] Step S203: The time-domain waveform of the equipment vibration is converted into a frequency-domain waveform using a Fast Fourier Transform. In the frequency-domain waveform, the horizontal axis represents frequency and the vertical axis represents amplitude. The frequency component with the highest amplitude is extracted from the frequency-domain waveform as the dominant frequency. The dominant frequency energy is obtained based on the dominant frequency, a dominant frequency energy feature is created, and it is appended to the sensor data. The formula for the dominant frequency energy is: E=|A(f)| 2 , where A(f) represents the amplitude corresponding to the main frequency f;

[0019] Step S204: Normalize the first feature data processed in the above steps using the z-score normalization method to obtain the second feature data.

[0020] Furthermore, in the Bayesian network topology graph, the types of nodes include: sensor nodes, equipment status nodes, defect nodes, and external factor nodes. Sensor nodes are used to monitor the operating parameters of the equipment in real time, including temperature nodes, current nodes, pressure nodes, and vibration nodes. Equipment status nodes reflect the current operating status and historical information of the equipment, including equipment operating status nodes, equipment load nodes, and historical maintenance record nodes. Defect nodes are used to indicate the defects or fault states of the equipment, including over-temperature defect nodes, current abnormality defect nodes, pressure abnormality defect nodes, and vibration abnormality defect nodes. External factor nodes reflect the impact of the external environment on the operation of the equipment, including ambient temperature nodes and ambient humidity nodes.

[0021] The directed edges of the Bayesian network topology graph are based on real physical and causal relationships. Each directed edge represents a causal relationship between two nodes. The direction of the edge points from the causal variable to the outcome variable to reflect the causal relationship. The two nodes connected by the directed edge are represented by the parent node and child node, respectively, according to the direction. In addition, the weight of the directed edge is determined according to the degree of influence between the nodes. The weight value represents the strength of the influence of the previous node on the next node.

[0022] Furthermore, the steps for constructing a Bayesian network model based on the PC algorithm from the Bayesian network topology graph include:

[0023] Step S301: Define an initial conditional probability table for each node in the Bayesian network topology graph, and combine the Bayesian network topology graph to obtain the basic Bayesian network, where the conditional probability table represents the probability of a node under different conditions.

[0024] Step S302: Use the PC algorithm to dynamically adjust the edges in the Bayesian network. The dynamic adjustment includes: adding edges, removing edges, and adjusting the direction of edges.

[0025] Step S303: Recalculate the conditional probability table of each node in the Bayesian network after processing in step S302, and combine it with the dynamically adjusted Bayesian network to obtain the updated Bayesian network.

[0026] Step S304: Perform probability calculations based on the updated Bayesian network to obtain the probability values ​​of various defects occurring in the device.

[0027] Furthermore, the conditional probability table of the node in step S301 is divided into a conditional probability table for nodes without parent nodes and a conditional probability table for nodes with parent nodes. Nodes without parent nodes only have their own probability distribution and do not depend on other nodes. Nodes with parent nodes set probability values ​​according to the different states of the parent nodes and combine them into a conditional probability table.

[0028] Furthermore, the core of the PC algorithm in step S302 is to dynamically adjust the edge structure based on the conditional independence test. The dynamic adjustment steps include:

[0029] Step S3021: Perform independence tests on a pair of nodes that are not directly connected under various conditions. If the conditions of the pair of nodes are related, add an edge between them.

[0030] Step S3022: Perform independence checks on a pair of directly connected nodes under various conditions. If the pair of nodes are independent under the conditions, remove the edge between them.

[0031] Step S3023: Determine the direction of the edge according to the PC algorithm and preset rules. The preset rules include: v-structure rule and direction propagation rule. The v-structure rule means that if nodes A and B are conditionally independent, they become conditionally related after being given node C. Then the three nodes form a v-structure, which is represented as: A→C←B. The direction propagation rule means that the propagation starts from a known direction and is passed to the adjacent nodes to avoid forming a loop.

[0032] Furthermore, the specific steps of step S304 include:

[0033] Step S3041: Define target nodes and evidence nodes, and assign values ​​to evidence nodes according to the currently collected data. The target node refers to the defect node, and the evidence node refers to the node of the currently observed data, that is, the node other than the target node.

[0034] Step S3042: Calculate the probability of the target node and evidence node appearing together step by step according to the conditional probability table. The calculation formula is as follows: Where P(target|evidence) represents the probability of the target node and the evidence node being together, P(evidence|target) represents the conditional probability of observing evidence when the defect exists, P(target) represents the pre-defined probability of the defect, and P(evidence) represents the probability of the evidence node.

[0035] Step S3043: The observed evidence is propagated layer by layer to the defect node through the conditional probability table in the Bayesian network to obtain the probability values ​​of various defects occurring in the equipment.

[0036] The beneficial effects of this invention are as follows: This invention dynamically adjusts the edge structure of a Bayesian network based on the PC algorithm, and combines conditional independence testing and directional propagation rules to ensure the accuracy of network causal relationships, making the defect prediction results more interpretable and reliable.

[0037] This invention employs multi-source heterogeneous data acquisition and multi-dimensional feature preprocessing, which can adapt to the complex and ever-changing environmental conditions of converter stations, avoid the dependence of traditional models on the independent and identically distributed assumption, and improve the stability of the model in actual operation.

[0038] This invention not only identifies defects, but also generates handling measures based on defect level standards, and combines them with the power grid management platform to achieve closed-loop management, monitor the progress of defect handling in real time, and ensure the integrity of the management process. Attached Figure Description

[0039] Figure 1 This is a flowchart of the dynamic evaluation method based on defect trend changes of the present invention;

[0040] Figure 2 This is a flowchart of the preprocessing of the first feature data according to the present invention;

[0041] Figure 3 This is the Bayesian network topology diagram of the present invention;

[0042] Figure 4 This is a flowchart of the process for constructing a Bayesian network model according to the present invention;

[0043] Figure 5 This is a flowchart of the dynamic adjustment edge structure of the present invention;

[0044] Figure 6 This is a flowchart of the probability calculation based on the updated Bayesian network according to the present invention. Detailed Implementation

[0045] The subject matter described herein will now be discussed with reference to exemplary embodiments. It should be understood that these embodiments are discussed only to enable those skilled in the art to better understand and implement the subject matter described herein, and changes may be made to the function and arrangement of the elements discussed without departing from the scope of this specification. Various processes or components may be omitted, substituted, or added as needed in the examples. Furthermore, features described in some examples may be combined in other examples.

[0046] It should be noted that, unless otherwise defined, the technical or scientific terms used in one or more embodiments of the present invention should have the ordinary meaning understood by one of ordinary skill in the art to which this invention pertains. The terms "first," "second," and similar terms used in one or more embodiments of the present invention do not indicate any order, quantity, or importance, but are merely used to distinguish different components. Terms such as "comprising" or "including" mean that the element or object preceding the word encompasses the elements or objects listed after the word and their equivalents, without excluding other elements or objects. Terms such as "connected" or "linked" are not limited to physical or mechanical connections, but can include electrical connections, whether direct or indirect. Terms such as "upper," "lower," "left," and "right" are used only to indicate relative positional relationships; when the absolute position of the described object changes, the relative positional relationship may also change accordingly.

[0047] like Figures 1-6 As shown, the dynamic evaluation method based on defect trend changes includes the following steps:

[0048] Step S101: Collect sensor data, equipment status data, environmental data, and historical maintenance data from the converter station system, and represent them using the first feature data;

[0049] The sensor data includes: equipment temperature, equipment current, equipment pressure, and equipment vibration; the equipment status data includes: equipment operating status and equipment load data; the environmental data includes: ambient temperature and ambient humidity; and the historical maintenance data includes: maintenance time interval and maintenance frequency.

[0050] Step S102: Preprocess the first feature data to obtain the second feature data;

[0051] Step S103: Construct a Bayesian network topology graph based on the second feature data. The Bayesian network topology graph includes: nodes, directed edges, and conditional probability tables of nodes.

[0052] Step S104: Construct a Bayesian network model based on the PC algorithm according to the Bayesian network topology diagram. The input of the Bayesian network model is the second feature data, and the output is the probability value of various defects in the equipment.

[0053] The types of defects that occur in equipment include: over-temperature defects, abnormal current defects, abnormal vibration defects, and abnormal pressure defects.

[0054] Step S105: Based on the probability values ​​of various defects in the device output by the Bayesian network model and the preset defect level classification standards, generate handling measures.

[0055] In one embodiment of the present invention, data acquisition in the converter station system primarily relies on various sensors and equipment monitoring systems to comprehensively monitor equipment status and environmental conditions by collecting different types of data. This data is recorded in real time and transmitted to the system via an integrated platform. Specifically, temperature, current, pressure, and vibration frequencies of the equipment are collected by installing temperature sensors, current sensors, pressure sensors, and vibration sensors on the equipment; ambient temperature and humidity are collected by installing external temperature and humidity sensors; equipment operating status and load data are collected through the equipment management system; and historical maintenance data are automatically extracted through the integrated platform and the power grid management system.

[0056] In one embodiment of the present invention, the device vibration is represented by a time-domain waveform, where the horizontal axis represents a time point and the vertical axis represents the intensity of the vibration, expressed as acceleration. The device operating states include: normal, standby, and overload. The device operating states are represented by integer codes, for example, normal, standby, and overload are represented by 1, 2, and 3 respectively, then the code for the device operating state of standby is 2. The device load data is represented by real number codes, for example, when the device load reaches 70%, it is represented by the real number 0.7.

[0057] In one embodiment of the present invention, the step of preprocessing the first feature data includes:

[0058] Step S201: Use the 3σ criterion to detect outliers in sensor data and environmental data, and replace the outliers with the median. Specifically, calculate the mean and standard deviation of sensor data and environmental data, determine the normal data range as μ ± 3σ, detect outliers outside the range, and replace the detected outliers with the median to ensure the stability and accuracy of the data, where μ is the mean and σ is the standard deviation.

[0059] Step S202: The sliding time window method is used to smooth the sensor data and environmental data that have undergone outlier replacement processing. Specifically, a preset time window is selected, the average value of the data within the window is calculated as the smoothing value, and the window is gradually slid forward to smooth the entire dataset, eliminating short-term fluctuations and maintaining the long-term trend of the data.

[0060] Step S203: The time-domain waveform of the equipment vibration is converted into a frequency-domain waveform using a Fast Fourier Transform. In the frequency-domain waveform, the horizontal axis represents frequency and the vertical axis represents amplitude. The frequency component with the highest amplitude is extracted from the frequency-domain waveform as the dominant frequency. The dominant frequency energy is obtained based on the dominant frequency, a dominant frequency energy feature is created, and it is appended to the sensor data. The formula for the dominant frequency energy is: E=|A(f)| 2 , where A(f) represents the amplitude corresponding to the main frequency f;

[0061] Step S204: Normalize the first feature data processed in the above steps using the z-score normalization method to obtain the second feature data.

[0062] In one embodiment of the present invention, the types of nodes in the Bayesian network topology graph include: sensor nodes, device status nodes, defect nodes, and external factor nodes. Sensor nodes are used to monitor the operating parameters of the device in real time, including temperature nodes, current nodes, pressure nodes, and vibration nodes. Device status nodes reflect the current operating status and historical information of the device, including device operating status nodes, device load nodes, and historical maintenance record nodes. Defect nodes are used to indicate the defects or fault states of the device, including over-temperature defect nodes, current abnormality defect nodes, pressure abnormality defect nodes, and vibration abnormality defect nodes. External factor nodes reflect the influence of the external environment on the operation of the device, including ambient temperature nodes and ambient humidity nodes.

[0063] The directed edges of the Bayesian network topology are based on real physical and causal relationships. Each directed edge represents a causal association between two nodes, with the direction of the edge pointing from the causal variable to the outcome variable to reflect the causal relationship. The two nodes connected by the directed edge are represented by the parent node and child node, respectively, according to the direction. In addition, the weight of the directed edge is determined according to the degree of influence between the nodes. The weight value represents the strength of the influence of the previous node on the next node. For example, the temperature of the equipment directly affects the probability of an overheating defect. Therefore, there is a directed edge between the temperature node and the overheating defect node, pointing from the temperature node to the overheating defect node. The weight of this directed edge represents the probability of an overheating defect occurring at the current equipment temperature. Similarly, the ambient temperature directly affects the equipment temperature. Therefore, there is a directed edge between the ambient temperature node and the temperature node, pointing from the ambient temperature node to the temperature node. The weight of this directed edge represents the degree of influence of the current ambient temperature on the equipment temperature.

[0064] In one embodiment of the present invention, the step of constructing a Bayesian network model based on the PC algorithm according to the Bayesian network topology graph includes:

[0065] Step S301: Define an initial conditional probability table for each node in the Bayesian network topology graph, and combine the Bayesian network topology graph to obtain the basic Bayesian network, where the conditional probability table represents the probability of a node under different conditions.

[0066] Step S302: Use the PC algorithm to dynamically adjust the edges in the Bayesian network. The dynamic adjustment includes: adding edges, removing edges, and adjusting the direction of edges.

[0067] Step S303: Recalculate the conditional probability table of each node in the Bayesian network after processing in step S302, and combine it with the dynamically adjusted Bayesian network to obtain the updated Bayesian network.

[0068] Step S304: Perform probability calculations based on the updated Bayesian network to obtain the probability values ​​of various defects occurring in the device.

[0069] In one embodiment of the present invention, the conditional probability table of the nodes in step S301 is divided into a conditional probability table for nodes without parent nodes and a conditional probability table for nodes with parent nodes. Nodes without parent nodes include: ambient temperature nodes, ambient humidity nodes, and historical maintenance record nodes. Nodes without parent nodes only have their own probability distribution and do not depend on other nodes. The ambient temperature and ambient humidity nodes are set to a uniform distribution, and the historical maintenance record nodes are given an initial probability based on the frequency distribution of each state in the historical records. Nodes with parent nodes are those other than those without parent nodes. Their probability values ​​are set according to different states of the parent node. For example, the equipment temperature changes due to the influence of ambient temperature, and its probability expression is: P(device temperature|ambient temperature) represents the conditional probability table of the influence of ambient temperature on the equipment temperature. P(high equipment temperature|high ambient temperature) represents the probability that a high ambient temperature will cause the equipment temperature to be high, and P(low equipment temperature|low ambient temperature) represents the probability that a low ambient temperature will cause the equipment temperature to be low.

[0070] In one embodiment of the present invention, the core of the PC algorithm in step S302 is to dynamically adjust the edge structure based on the conditional independence test, wherein the conditional independence test is used to determine whether two variables are independent under a given third conditional variable, and the dynamic adjustment step includes:

[0071] Step S3021: Perform independence tests on a pair of nodes that are not directly connected under various conditions. If the conditions of the pair of nodes are related, add an edge between them.

[0072] Step S3022: Perform independence checks on a pair of directly connected nodes under various conditions. If the pair of nodes are independent under the conditions, remove the edge between them.

[0073] Step S3023: Determine the edge direction based on the PC algorithm and preset rules. The preset rules include: v-structure rules and direction propagation rules. The v-structure rule means that if conditionally independent nodes A and B become conditionally related after being given node C, then the three nodes form a v-structure, represented as A→C←B. For example, equipment load and ambient temperature are conditionally independent, but after being given equipment temperature, equipment load and equipment temperature become related, forming a v-structure. This structure clarifies the edge direction, ensuring the causal effect of A and B on C. Without this conditional independence relationship, the edge direction cannot be uniquely determined. The direction propagation rule means: starting from a known direction, propagation is passed to adjacent nodes to avoid forming loops. For example, if it is known that equipment load A affects equipment temperature B, and there is an edge BC, where C represents an over-temperature defect, according to the direction propagation rule, BC can be directed to B→C, forming A→B→C. The direction propagation rule helps avoid closed loops in the network structure, thereby maintaining the acyclicity of the Bayesian network.

[0074] By adjusting the edge structure as described above, the PC algorithm achieves accurate causal relationship expression within the network structure, and ensures that the edge directions are reasonable and acyclic through the v-structure and directional propagation rules. This design not only improves the model's accuracy and computational efficiency but also provides reliable causal chain support for equipment defect prediction.

[0075] In one embodiment of the present invention, step S304 specifically includes the following steps:

[0076] Step S3041: Define target nodes and evidence nodes, and assign values ​​to evidence nodes according to the currently collected data. The target node refers to the defect node, and the evidence node refers to the node of the currently observed data, that is, the node other than the target node.

[0077] Step S3042: Calculate the probability of the target node and evidence node appearing together step by step according to the conditional probability table. The calculation formula is as follows: Where P(target|evidence) represents the probability of the target node and the evidence node being together, P(evidence|target) represents the conditional probability of observing evidence when the defect exists, P(target) represents the pre-defined probability of the defect, and P(evidence) represents the probability of the evidence node.

[0078] Step S3043: The observed evidence is propagated layer by layer to the defect node through the conditional probability table in the Bayesian network to obtain the probability values ​​of various defects occurring in the equipment.

[0079] In one embodiment of the present invention, a defect level standard is preset based on expert experience and the actual operating requirements of the equipment. The levels are divided as follows: a probability between 0 and 0.3 is a low defect level, a probability between 0.3 and 0.6 is a medium defect level, a probability between 0.6 and 0.9 is a high defect level, and a probability between 0.9 and 1 is an extremely high defect level. The probability values ​​of various defects of the equipment output by the Bayesian network model are compared with the preset defect level standard to determine the level of each defect. Corresponding handling measures are generated according to the defect level. For example, the handling measures for a low defect level are: no intervention is required, but regular monitoring is necessary to record the changing trend of the defect probability; the handling measures for a medium defect level are: increase the monitoring frequency and arrange regular inspections; the handling measures for a high defect level are: arrange preventive maintenance and, when convenient, briefly stop the machine for further inspection; the handling measures for an extremely high defect level are: immediately stop the machine and arrange a comprehensive overhaul to prevent serious equipment failure.

[0080] In one embodiment of the present invention, the invention can also obtain ledger information and maintenance records from the integrated platform and power grid management platform based on the unique identifier of the defective object, continuously monitor the progress of defect handling, and achieve comprehensive closed-loop management. This process covers all stages from defect discovery to completion of handling, ensuring data integrity and traceability, and further improving the efficiency of power grid operation and maintenance management and the safety and stability of equipment.

[0081] The embodiments of this example have been described above. However, this example is not limited to the specific implementation methods described above. The specific implementation methods described above are merely illustrative and not restrictive. Those skilled in the art can make many other forms based on the guidance of this example, and all of them are within the protection scope of this example.

Claims

1. A dynamic evaluation method based on a change in defect trend, characterized by, The method comprises the following steps: In step S101, sensor data, device state data, environment data and historical maintenance data in the converter station system are collected and represented by first feature data; The sensor data comprises device temperature, device current, device pressure and device vibration, the device state data comprises device operating state and device load data, the environment data comprises environment temperature and environment humidity, and the historical maintenance data comprises maintenance time interval and maintenance times; In step S102, the first feature data is preprocessed to obtain second feature data; In step S103, a Bayesian network topology graph is constructed according to the second feature data, the Bayesian network topology graph comprising nodes, directed edges and conditional probability tables of the nodes; In step S104, a Bayesian network model based on a PC algorithm is constructed according to the Bayesian network topology graph, the input of the Bayesian network model being the second feature data and the output being probability values of various defects of the device; The defect types of the various defects of the device comprise over-temperature defects, current abnormal defects, vibration abnormal defects and pressure abnormal defects; The step of constructing the Bayesian network model based on the PC algorithm according to the Bayesian network topology graph comprises: In step S301, an initial conditional probability table is defined for each node in the Bayesian network topology graph, and a basic Bayesian network is obtained in combination with the Bayesian network topology graph, wherein the conditional probability table represents the probability of the node under different conditions; In step S302, the PC algorithm is used to dynamically adjust the edges in the Bayesian network, the dynamic adjustment comprising adding edges, removing edges and adjusting the direction of the edges; In step S303, the conditional probability table of each node is recalculated for the Bayesian network processed in step S302, and an updated Bayesian network is obtained in combination with the dynamically adjusted Bayesian network; In step S304, probability calculation is performed according to the updated Bayesian network to obtain the probability values of the various defects of the device; In step S105, the processing measures are generated according to the probability values of the various defects of the device output by the Bayesian network model in combination with a preset defect level division standard.

2. The dynamic evaluation method based on defect trend variation according to claim 1, characterized in that, The device operating state comprises normal, standby and overload, and the device operating state is represented by integer coding, and the device load data is represented by real number coding.

3. The method of claim 1, wherein the defect trend change is determined by a change in a defect density of a defect type. The step of preprocessing the first feature data comprises: Step S201, using 3 criteria detects outliers in the sensor data and the environment data and replaces the outliers with median values; In step S202, the sensor data and the environment data processed by the outlier replacement are smoothed by using a sliding time window method; Step S203, the time-domain waveform of the device vibration is converted into a frequency-domain waveform representation by fast Fourier transform, in the frequency-domain waveform, the horizontal axis represents the frequency, and the vertical axis represents the amplitude, the frequency component with the highest amplitude is extracted from the frequency-domain waveform as the main frequency, the main frequency energy is obtained according to the main frequency, the main frequency energy feature is created, and it is spliced after the sensor data, and the formula of the main frequency energy is: Wherein represents the amplitude corresponding to the main frequency f. In step S204, the first feature data processed in the above steps is normalized by using a z-score normalization method to obtain the second feature data.

4. The method of claim 1, wherein the method is characterized by, The types of nodes in the Bayesian network topology include: sensor nodes, device state nodes, defect nodes, and external factor nodes, wherein the sensor nodes are used to monitor the operating parameters of the device in real time, including temperature nodes, current nodes, pressure nodes, and vibration nodes; the device state nodes reflect the current operating state and historical information of the device, including device operating state nodes, device load nodes, and historical maintenance record nodes; the defect nodes are used to represent the defects or fault states of the device, including over-temperature defect nodes, current anomaly defect nodes, pressure anomaly defect nodes, and vibration anomaly defect nodes; and the external factor nodes reflect the influence of the external environment on the operation of the device, including environment temperature nodes and environment humidity nodes. The directed edges of the Bayesian network topology are set based on real physical and causal relationships, each directed edge represents the causal association between two nodes, the direction of the edge points from the causal variable to the result variable to reflect the causal relationship, and the two nodes connected by the directed edge are represented by the parent node and the child node according to the direction; in addition, the weight of the directed edge is determined according to the influence degree between the nodes, and the weight value represents the influence strength of the previous node on the next node.

5. The method of claim 4, wherein the defect trend change is determined by a difference between a first defect trend and a second defect trend. The conditional probability table of the node in the step S301 is divided into the conditional probability table of the node without a parent node and the conditional probability table of the node with a parent node, wherein the node without a parent node only has its own probability distribution and does not depend on other nodes; and the node with a parent node sets the probability value according to the different states of the parent node and combines them into the conditional probability table.

6. The method of claim 4, wherein the defect trend change is determined by a change in a defect density of the defects. The core of the PC algorithm in the step S302 is to dynamically adjust the edge structure according to the conditional independence test, and the steps of the dynamic adjustment include: Step S3021, performing independence test under each condition for a pair of nodes not directly connected, and when the pair of nodes are conditionally dependent, an edge is added between them; Step S3022, performing independence test under each condition for a pair of directly connected nodes, and when the pair of nodes are conditionally independent, an edge is removed between them; In step S3023, the direction of the edge is determined according to the PC algorithm combined with preset rules, and the preset rules include a v structure rule and a direction propagation rule. The v structure rule means that nodes A and B that are conditionally independent become conditionally dependent after a given node C, and the three nodes form a v structure, which is denoted as: The direction propagation rule means that the direction is transmitted to adjacent nodes from a known direction to avoid forming a loop.

7. The method of claim 4, wherein the defect trend change is determined by a regression analysis of the defect data. The specific steps of the step S304 include: Step S3041, defining the target node and the evidence node, and assigning values to the evidence node according to the currently collected data, wherein the target node refers to the defect node, and the evidence node refers to the node of the current observation data, i.e., the node other than the target node; Step S3042, the joint probability of the target node and the evidence node is calculated according to the conditional probability table, and the calculation formula is: wherein P (X, E) represents the joint probability of the target node and the evidence node, P (E | D) represents the conditional probability of observing the evidence when the defect exists, P (D) represents the preset probability of the defect, P (E) represents the probability of the evidence node; Step S3043, propagating the observed evidence to the defect node through the conditional probability table in the Bayesian network layer by layer to obtain the probability values of various defects of the device.

Images