Equipment fault early warning method based on neural network, medium and equipment

Abstract

The invention provides an equipment fault early warning method based on a neural network, a medium and equipment, and the method comprises the steps: collecting equipment state information at regular time, recognizing abnormal equipment, obtaining operation parameters and structure parameters of the abnormal equipment and a layout schematic diagram of a system where the abnormal equipment is located, constructing a dynamic coupling topology link, and extracting associated parameters; verifying and generating a first error weight in combination with a physical kinematics equation, generating a second error weight by matching historical fault data, and comprehensively generating fault state information; a detection mode is switched based on risk level evaluation, a system dynamic coupling topology link, associated parameters, abnormal equipment information and fault state information are input into a neural network model, fault link information is output, and finally first early warning information and second early warning information are generated according to a dynamic detection mode to realize two-stage early warning. According to the method, the fault conduction relation between equipment is modeled through the graph neural network, and fault path prediction and systematic risk early warning are realized.

Description

Technical Field

[0001] The present invention relates to the field of computer technology, and particularly to a method, medium, and device for equipment fault early warning based on a neural network. Background Art

[0002] Current Internet of Things device fault early warning methods are mostly limited to the status monitoring of single devices and lack the ability to model the system-level coupled fault conduction path. Traditional solutions usually adopt threshold alarms or machine learning models based on isolated historical data. Although they can identify local anomalies, they cannot capture fault propagation, and it is difficult to distinguish between occasional anomalies and potential systemic faults. The prediction of cascading faults lags behind, especially when anomalies are conducted through mechanical coupling links, and the risk evolution of associated devices cannot be pre-warned. Summary of the Invention

[0003] In view of the above problems, the present invention provides a method, medium, and device for equipment fault early warning based on a neural network, which solves the problem of late warning caused by the inability to predict cascading faults.

[0004] To achieve the above object, in the first aspect, the present application provides a method for equipment fault early warning based on a neural network, including: Regularly obtain equipment status information, and determine whether the equipment status information is abnormal. If so, record it as an abnormal device and trigger an abnormal evaluation step, including: Obtain abnormal device information, where the abnormal device information includes operating parameters and structural parameters; Obtain the layout schematic diagram of the system where the abnormal device is located, construct a system dynamic coupling topology link according to the layout schematic diagram, and generate associated parameters in combination with the abnormal device information. The associated parameters are the location information of the abnormal device in the system and the device information of the other devices connected by the abnormal device; Verify the equipment status information according to the physical kinematic equation based on the abnormal device information and the associated parameters, and generate a first error weight; Match the historical working condition fault information of the abnormal device in the historical database according to the abnormal device information, calculate the fitness between the equipment status information and the historical working condition fault information, and generate a second error weight; Generate the fault status information of the equipment according to the first error weight and the second error weight; Evaluate the risk level of the fault status information, and generate a detection mode according to the evaluation result. The detection mode includes a high-frequency detection mode and a low-frequency detection mode; Input the system dynamic coupling topology link, the associated parameters, the abnormal device information, and the fault status information into the neural network model to obtain fault link information, where the fault link information includes forward traceability prediction information for causing the fault of the abnormal device, backward fault prediction information of the abnormal device, and the time limit section; Generate a first warning message based on the faulty link information, the fault status information, and the abnormal device information. When the detection mode is switched to the high-frequency detection mode, trigger the low-frequency detection mode of the remaining devices in the current faulty link information and generate a second warning message.

[0005] Furthermore, constructing the system dynamic coupling topology link according to the layout schematic diagram includes: Obtain the mechanical connection relationship in the layout schematic diagram, and construct a directed graph structure with individual devices as nodes and mechanical connections as connecting lines, and perform weight initialization on the directed graph structure, including: The node weight of each node is generated by the inherent failure rate of the individual device, and the connecting line weight of each connecting line is obtained by calculating the material strength and design load of the nodes at both ends of the connecting line; Monitor the vibration signals between the devices of each connecting line, generate the actual energy attenuation rate according to the vibration signals, and generate a first weight adjustment parameter according to the actual energy attenuation rate; Monitor the cooperation parameters of the devices of each connecting line, calculate the motion cooperation coefficient between two devices according to the cooperation parameters, and generate a second weight adjustment parameter according to the motion cooperation coefficient; Adjust the connecting line weight according to the first weight adjustment parameter and the second weight adjustment parameter; And, obtain the used duration and the rated duration of each device, and generate a device failure rate adjustment parameter according to the used duration and the rated duration; Adjust the node weight according to the device failure rate adjustment parameter.

[0006] Furthermore, construct the system dynamic coupling topology link according to the layout schematic diagram, and generate the association parameter in combination with the abnormal device information, including: Mark the node of the abnormal device as an abnormal node, and conduct a conduction search along the direction of the connecting line of the abnormal node until the number of connecting lines in the first conduction path formed by the conduction search exceeds the preset tolerance value, and the preset tolerance value is generated by the product of the connecting line weights in the first conduction path and the conduction threshold; Mark the devices associated with the nodes covered by the first conduction path as affected devices; Perform the following steps for each affected device: Obtain the connecting line weight on the second conduction path from the abnormal node to the affected device; Calculate the geometric mean of the connecting line weights on the second conduction path as the path conduction efficiency; Obtain the node weight of the abnormal node, denoted as the abnormal node weight, and obtain the node weight of the affected device, denoted as the original node weight; Generate a weight influence coefficient for the affected device according to the abnormal node weight, the path conduction efficiency, and the original node weight; Update the original node weights according to the weight influence coefficient, denoted as the influence node weights; Repeat the above steps until the original node weights of all affected devices are updated; Generate the location information of the abnormal device in the system based on the first conduction path and the directed graph structure, and generate the device information of the remaining devices connected by the abnormal device based on the influence node weights and the affected devices, which is the correlation parameter.

[0007] Furthermore, verify the device status information using the physical kinematic equation based on the abnormal device information and the correlation parameter, and generate the first error weight, including: Generate the actual installation parameters and the theoretical installation parameters based on the abnormal device information and the correlation parameter; Calculate the deviation information between the actual installation parameters and the theoretical installation parameters, denoted as the installation deviation information; Perform physical deviation calculation based on the installation deviation information to obtain the calculation result, which is represented by formula (1). Formula (1) is as follows: ; In formula (1), is the position deviation, is the angle deviation, is the fit deviation, is the position sensitivity coefficient, which is related to the sensor type, is the actual installation position vector, is the theoretical installation position vector, is the state information space gradient, is the installation angle deviation value, is the higher-order error term related to the angle, is the th harmonic amplitude of the vibration signal, is the fundamental wave amplitude; Convert the calculation result into the first error weight. The first error weight is represented by formula (2). Formula (2) is as follows: ; In formula (2), is the first error weight, is the original data of the uncorrected device status information, is the compensated status information, is the maximum historical amplitude of the status information under normal device operation, is the sampling time, is the number of sampling points within the sliding window, is the tightness influence factor, is the tightness index, is the position deviation amount, is the position deviation threshold, is the enhancement coefficient.

[0008] Furthermore, the device status information includes abnormal vibration signals and abnormal spectrum characteristics generated by abnormal device operation, and the historical working condition fault information includes historical vibration signals, historical spectrum characteristics, and historical operation parameters; Match the historical working condition fault information of the abnormal device in the historical database according to the abnormal device information, calculate the fitness between the current device status information and the historical working condition fault information, and generate a second error weight including: Calculate the waveform similarity between the abnormal vibration signal and the historical vibration signal, and calculate the energy distribution difference value between the abnormal spectrum characteristic and the historical spectrum characteristic in the abnormal frequency band; Correct the waveform similarity and the energy distribution difference value according to the difference information between the operation parameters in the abnormal device information and the historical operation parameters; Fuse the corrected waveform similarity and the corrected energy distribution difference value to generate a second error weight.

[0009] Furthermore, evaluate the risk level of the fault status information, and generate a detection mode according to the evaluation result, including: Divide the abnormal monitoring sequence in the fault status information into multiple monitoring points according to the unit time, and each monitoring point has a monitoring value; Construct an abnormal monitoring sequence curve graph, where the ordinate of the abnormal monitoring sequence curve graph is the monitoring value, and the abscissa of the abnormal monitoring sequence curve graph is the monitoring time; Divide the high-risk monitoring interval, medium-risk monitoring interval, and low-risk monitoring interval in the abnormal monitoring sequence curve graph, and divide the multiple monitoring values to form a high-risk set, a medium-risk set, and a low-risk set; Calculate the proportion of the monitoring values in the high-risk set, medium-risk set, and low-risk set in the current abnormal monitoring sequence to generate an evaluation result of the abnormal monitoring sequence; If the evaluation result is high risk, monitor the abnormal device according to the high-frequency detection mode; If the evaluation result is medium risk, monitor the abnormal device according to the low-frequency detection mode; If the evaluation result is low risk, maintain the initial detection mode.

[0010] Furthermore, the neural network model includes a physical decoupling model. Input the system dynamic coupling topology link, associated parameters, abnormal device information, and fault status information into the neural network model to obtain the fault link information including: Extract the forward conduction path of the abnormal device from the system dynamic coupling topology link, record the devices in the forward conduction path as fault source devices, generate topology decoupling features, and extract the physical connection features between the fault source devices and the abnormal devices from the associated parameters; Input the topology decoupling features and physical connection features into the trained physical decoupling model to obtain the forward traceability prediction information for the fault of the abnormal device, including: Extract the fault contribution probabilities of multiple fault source devices according to the topology decoupling features and physical connection features; Adjust the fault contribution probability of each fault source device according to the physical constraint features, where the physical constraint features include energy conservation constraint and time sequence propagation constraint, eliminate the fault source devices that do not meet the physical constraint features, and generate the corrected contribution probabilities of the remaining fault source devices; Generate the forward traceability prediction information for the fault of the abnormal device according to the corrected contribution probabilities and their corresponding fault source devices.

[0011] Further, the neural network model includes a physical coupling model. Input the system dynamic coupling topology link, associated parameters, abnormal device information, and fault status information into the neural network model. The obtained fault link information also includes: Extract the backward conduction path of the abnormal device from the system dynamic coupling topology link, mark the devices in the backward conduction path as risk devices, generate topology coupling features, and extract the physical coupling features between the risk devices and the abnormal devices from the associated parameters; Input the topology coupling features and physical coupling features into the trained physical coupling model to generate the backward fault prediction information for the abnormal device, including: Calculate the inherent conduction risk value of the risk device according to the topology coupling features; Calculate the risk correction coefficient related to the working condition according to the physical coupling features; Fuse the risk correction coefficient and the inherent conduction risk value to obtain the conduction risk probability of each risk device; Verify the multiple conduction risk probabilities using the mass conservation constraint, eliminate the risk devices that violate the mass conservation constraint, correct the conduction risk probabilities, obtain the corrected risk probabilities, and generate the time period corresponding to the current corrected risk probabilities; Generate the backward fault prediction information and time period for the abnormal device according to the corrected risk probabilities and their corresponding risk devices.

[0012] In a second aspect, the present invention also provides a computer-readable storage medium, on which computer program instructions are stored. When the computer program instructions are executed by a processor, the method described in the first aspect is implemented.

[0013] In a third aspect, the present invention further provides an electronic device, including a memory and a processor, where the memory is used to store one or more computer program instructions, and wherein the one or more computer program instructions are executed by the processor to implement the method described in the first aspect.

[0014] Different from the prior art, the above technical solution provides a method, medium and device for early warning of equipment failures based on a neural network. After the method periodically collects equipment status information to identify abnormal equipment, it obtains its operating parameters, structural parameters, and the layout schematic diagram of the system where it is located, constructs a dynamic coupling topological link, and extracts associated parameters; combines physical kinematic equations to verify and generate a first error weight, and at the same time matches historical fault data to generate a second error weight, and comprehensively generates fault status information; based on risk level assessment, switches the detection mode, inputs the system dynamic coupling topological link, associated parameters, abnormal equipment information, and fault status information into a neural network model, outputs fault link information, and finally generates a first warning information and a second warning information according to the dynamic detection mode to achieve dual-level warning. This method models the fault conduction relationship between equipment through a graph neural network to realize fault path prediction and systematic risk warning.

[0015] The above relevant records of the invention content are only an overview of the technical solution of this application. In order to enable those of ordinary skill in the art to more clearly understand the technical solution of this application, and thus can be implemented according to the content recorded in the description and the drawings, and in order to make the above objects, other objects, features and advantages of this application more easily understood, the following is described in conjunction with the specific implementation manners and drawings of this application. BRIEF DESCRIPTION OF THE DRAWINGS

[0016] The drawings are only used to illustrate the principles, implementation manners, applications, features, and effects of the specific implementation manners of the present invention and other related contents, and should not be considered as a limitation to this application.

[0017] In the drawings of the specification: Figure 1 It is a method step diagram of steps S101 to S109 of the early warning method described in the specific implementation manner; Figure 2 It is a method step diagram of steps S201 to S206 of the early warning method described in the specific implementation manner; Figure 3 It is a method step diagram of steps S301 to S309 of the early warning method described in the specific implementation manner; Figure 4 It is a method step diagram of steps S401 to S403 of the early warning method described in the specific implementation manner; Figure 5 It is a structural schematic diagram of the electronic device described in the specific implementation manner.

[0018] The reference numerals involved in the above-mentioned drawings are explained as follows: 1. Electronic device; 11. Memory; 12. Processor. Detailed implementation manners

[0019] To describe in detail the possible application scenarios, technical principles, specific implementable solutions, achievable purposes and effects of this application, etc., the following will be described in detail with reference to the specific examples listed and in conjunction with the drawings. The examples described in this article are only used to more clearly illustrate the technical solutions of this application, so they are only examples and cannot be used to limit the protection scope of this application.

[0020] Referring to "embodiments" in this article means that specific features, structures or characteristics described in connection with the embodiments may be included in at least one embodiment of this application. The term "embodiment" that appears in various positions in the specification does not necessarily refer to the same embodiment, nor does it particularly limit its independence or relevance to other embodiments. In principle, in this application, as long as there is no technical contradiction or conflict, the technical features mentioned in each embodiment can be combined in any way to form corresponding implementable technical solutions.

[0021] Unless otherwise defined, the meanings of the technical terms used in this article are the same as those generally understood by those skilled in the technical field to which this application belongs; the use of relevant terms in this article is only to describe specific embodiments and is not intended to limit this application.

[0022] In the description of this application, the phrase "and / or" is an expression used to describe the logical relationship between objects, indicating that three relationships can exist. For example, A and / or B means: there is A, there is B, and there is both A and B at the same time. In addition, the character " / " in this article generally represents an "or" logical relationship between the associated objects before and after.

[0023] In this application, terms such as "first" and "second" are only used to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply any actual quantity, primary or secondary, or order relationship between these entities or operations.

[0024] Without further limitations, in this application, the open-ended expressions such as "including", "comprising", "having" or other similar expressions used in a statement are intended to cover non-exclusive inclusion. These expressions do not exclude the possibility that there may be additional elements in the process, method or product including the said elements. Thus, in a process, method or product including a series of elements, it can include not only those defined elements, but also other elements not explicitly listed, or elements inherent to such process, method or product.

[0025] Similar to the understanding in the "Examination Guidelines", in this application, expressions such as "greater than", "less than", "exceeding" are understood as not including the recited number; expressions such as "above", "below", "within" are understood as including the recited number. In addition, in the description of the embodiments of this application, the meaning of "multiple" is two or more (including two). Similar expressions related to "many", such as "multiple groups", "multiple times", etc., are understood in this way unless otherwise specifically defined.

[0026] In the description of the embodiments of this application, the spatially related expressions used, such as "center", "longitudinal", "transverse", "length", "width", "thickness", "upper", "lower", "front", "rear", "left", "right", "vertical", "horizontal", "perpendicular", "top", "bottom", "inner", "outer", "clockwise", "counterclockwise", "axial", "radial", "circumferential", etc., the indicated orientation or positional relationship is based on the orientation or positional relationship shown in the specific embodiment or the attached drawings. It is only for the convenience of describing the specific embodiments of this application or for the reader's understanding, rather than indicating or implying that the device or component referred to must have a specific position, a specific orientation, or be constructed or operated in a specific orientation. Therefore, it should not be construed as a limitation to the embodiments of this application.

[0027] Please refer to Figure 1 , in the first aspect, this embodiment provides a device fault warning method based on a neural network, including: S101. Regularly obtain device status information, and determine whether the device status information is abnormal. If so, record it as an abnormal device and trigger an abnormal evaluation step, including: S102. Obtain abnormal device information, where the abnormal device information includes operating parameters and structural parameters; S103. Obtain the layout schematic diagram of the system where the abnormal device is located, construct a system dynamic coupling topological link according to the layout schematic diagram, and generate associated parameters in combination with the abnormal device information. The associated parameters are the position information of the abnormal device in the system and the device information of the other devices connected by the abnormal device; S104. Verify the device status information according to the abnormal device information and the associated parameters using a physical kinematic equation to generate a first error weight; S105. Match the historical working condition fault information of the abnormal device in the historical database according to the abnormal device information, calculate the fitness between the device status information and the historical working condition fault information, and generate a second error weight; S106. Generate the fault status information of the device according to the first error weight and the second error weight; S107. Conduct a risk level assessment on the fault status information, and generate a detection mode according to the assessment result. The detection mode includes a high-frequency detection mode and a low-frequency detection mode; S108. Input the system dynamic coupling topology link, associated parameters, abnormal device information, and fault status information into a neural network model to obtain fault link information. The fault link information includes forward traceability prediction information for causing the abnormal device fault, backward fault prediction information for the abnormal device, and a time-effective section; S109. Generate a first warning information according to the fault link information, the fault status information, and the abnormal device information. When the detection mode is switched to the high-frequency detection mode, trigger the low-frequency detection mode for the remaining devices in the current fault link information to generate a second warning information.

[0028] In step S101, the device status information is obtained from multi-source data such as vibration signals, temperature readings, and pressure values collected in real time by sensors. The abnormality judgment of the device status information is realized through compliance verification by a physical rule engine. The physical rule engine is a rule library built based on device physical laws (such as dynamic equations and thermodynamic principles) and is used to verify in real time whether the sensor data conforms to the theoretical model. The compliance verification is performed by envelope comparison of the real-time data with the theoretical operating parameters (such as the safe speed range of the bearing and the motor temperature threshold). If the threshold is exceeded for a certain continuous sampling period, it is determined that the device status information is abnormal.

[0029] In step S102, the operating parameters refer to dynamic variables such as the rotational speed, load rate, and power consumption of the device during real-time operation, which characterize the current operating intensity of the device; the structural parameters include inherent attributes such as the geometric dimensions, material properties, and assembly clearances of the device and are used to define the physical constraint conditions of the device. The operating parameters and the structural parameters jointly constitute the complete mathematical model input of the device operating state. The structural parameters are digitally extracted from the device technical drawings, while the operating parameters are derived from the real-time monitoring data stream.

[0030] In step S103, the layout schematic diagram includes system-level information such as the transmission chain connection method and the power transmission path. Preferably, the construction of the dynamic coupling topology link is based on the analysis of the vibration transfer characteristics and energy conduction paths between nodes, and the edge weight (i.e., the weight of the connection path between two nodes) is dynamically corrected by measuring the vibration signal transmission time delay between adjacent devices. The generation process of the associated parameters is completed by calculating the fault sensitivity of adjacent devices in the topology network and is used to quantify the fault conduction influence of the abnormal device in the system.

[0031] In step S104, physical kinematic equations are used to calculate theoretical fault characteristic parameters (such as characteristic frequencies, vibration modes). The specific content will be described in detail later. The first error weight is calculated based on the deviation between the measured characteristics and the theoretical values, representing the abnormal confidence level at the physical verification level. For example, in the case of a gear tooth breakage fault, the phase difference between the actual meshing frequency and the theoretically calculated value will directly affect this weight value.

[0032] In step S105, the matching of historical operating condition fault information is achieved using the dynamic time warping algorithm. The time series characteristics of the device state information are extracted through a sliding time window, and pattern similarity calculations are performed with the historical fault database. The second error weight is generated by converting the historical matching coefficient, reflecting the degree of coincidence between the current anomaly and the typical fault mode. Preferably, the calculation of the second error weight takes into account the time-varying effects of operating condition parameters (such as ambient temperature, lubrication state) and is corrected using the weighted Euclidean distance.

[0033] In steps S106 and S107, the fault state information can be obtained by fusing the first error weight and the second error weight through a Bayesian probability framework, which is used to quantify the probability estimate of the device being in a fault state. When the first error weight and the second error weight are positively correlated, the system will trigger a confidence acceleration accumulation mechanism. For example, when physical feature deviations and historical pattern matches occur simultaneously, the fault probability grows exponentially. Preferably, different monitoring strategies are activated according to the fault confidence score, and a fuzzy logic decision tree is used to implement risk level assessment. When the fault confidence score crosses a preset threshold, the monitoring strategy switches from baseline sampling to high-frequency acquisition. The preset threshold is determined by the statistical distribution of historical fault data and is dynamically corrected in combination with the device fault sensitivity level and operating environment parameters.

[0034] In step S108, preferably, the neural network model includes a physical decoupling model and a physical coupling model. The generation of fault link information depends on the spatio-temporal propagation simulation of the physical coupling model. The forward traceability prediction information locates the root cause device of the fault by inversely solving the fault characteristic transfer equation, and the backward fault prediction information can use the Monte Carlo method to simulate the conduction path of the fault along the topological network, which is used to predict the fault spread sequence. The time-sensitive section characterizes the time sensitivity of fault development, and the time window of each conduction stage is calculated through the device remaining life prediction model.

[0035] The step process for realizing equipment fault warning in this embodiment can be understood as follows: Regularly collect multi-source data such as equipment vibration, temperature, and pressure, and identify abnormal equipment through comparing the theoretical parameter envelope and continuous over-threshold judgment. Obtain the operating parameters and structural parameters of the abnormal equipment, construct a dynamic coupling topological link by combining with the system layout schematic diagram, and calculate the fault sensitivity of adjacent equipment to generate correlation parameters. Subsequently, execute the physical kinematic equation verification to generate the first error weight, and match the historical fault mode through the dynamic time warping algorithm to generate the second error weight. Based on Bayesian probability, fuse the first error weight and the second error weight to obtain the fault state probability, and use the fuzzy logic decision tree to evaluate the risk level. When the confidence level crosses the dynamic threshold, switch the detection mode. Input the topological link, correlation parameters, and fault information into the neural network model, solve the root cause equipment of the fault through the reverse equation, predict the potential fault sequence by Monte Carlo simulation of the conduction path, and generate the aging section in combination with the remaining life model. Finally, output the first warning information, activate the low-frequency monitoring of the associated equipment in the high-frequency detection mode, and synchronously generate the second warning information including the maintenance window period.

[0036] In this embodiment, by integrating the physical rule engine and the neural network model, multi-dimensional verification of the abnormal state of the equipment and accurate prediction of the fault conduction link are realized, effectively improving the warning reliability. Based on the construction of the dynamic coupling topological link and the calculation of the fault sensitivity of adjacent equipment, the quantitative analysis ability of the system-level fault conduction influence is strengthened. The first error weight generated by the physical kinematic equation verification and the second error weight of the historical pattern matching significantly improve the confidence level of the fault state evaluation. Through the detection mode switching strategy triggered by the dynamic threshold and the spatio-temporal propagation simulation of the neural network model, forward fault root cause location and backward conduction path prediction are synchronously realized. The generation of the aging section and the activation mechanism of the low-frequency monitoring of the associated equipment provide an accurate maintenance window period and a fault spread blocking strategy for equipment maintenance, reducing the risk of equipment cascading failures.

[0037] Please refer to Figure 2 , further, constructing the system dynamic coupling topological link according to the layout schematic diagram includes: S201. Obtain the mechanical connection relationship in the layout schematic diagram, construct a directed graph structure with individual equipment as nodes and mechanical connections as connection lines, and perform weight initialization on the directed graph structure, including: S202. The node weight of each node is generated by the inherent failure rate of a single piece of equipment, and the connection line weight of each connection line is obtained by calculating the material strength and design load of the nodes at both ends of the connection line; S203. Monitor the vibration signals between the equipment of each connection line, generate the actual energy attenuation rate according to the vibration signals, and generate the first weight adjustment parameter according to the actual energy attenuation rate; S204. Monitor the cooperation parameters of the devices on each connection line, calculate the motion cooperation coefficient between the two devices according to the cooperation parameters, and generate a second weight adjustment parameter according to the motion cooperation coefficient; S205. Adjust the connection line weight according to the first weight adjustment parameter and the second weight adjustment parameter; In addition, obtain the used duration and the rated duration of each device, and generate a device failure rate adjustment parameter according to the used duration and the rated duration; S206. Adjust the node weight according to the device failure rate adjustment parameter.

[0038] In step S201, the mechanical connection relationship is obtained by analyzing the assembly constraint information such as the transmission shaft system cooperation and the coupling interface position in the layout schematic diagram, and specifically comes from the geometric topology structure of the device CAD drawing or the physical contact relationship definition in the assembly process file. The construction of the directed graph structure follows the power transmission direction. In the weight initialization stage, basic parameters are assigned to the nodes and connection lines to provide an initial calculation benchmark for subsequent dynamic correction.

[0039] In step S202, the inherent failure rate of the node represents the basic failure probability of the device under standard working conditions, and is calculated by combining the design life parameter with the real-time health score (based on the cumulative operation duration and working conditions), reflecting the material fatigue characteristics. The calculation of the connection line weight depends on the ratio of the material strength (such as the yield limit of bearing steel) and the design load (such as the rated torque of the gear) of the two end nodes, and specifically shows the overload resistance coefficient, which is used to quantify the force transmission stability of the mechanical connection under theoretical working conditions. For example, when the ratio of the material strength to the design load at both ends of the connection line reaches 1.2, the initial weight is set to 0.8 to represent a high-reliability force transmission path.

[0040] In step S203, preferably, the vibration signal is collected by acceleration sensors installed on the devices at both ends of the connection line, and the actual energy attenuation rate is generated by calculating the frequency domain integral difference of the vibration power spectral density at the input end and the output end, reflecting the loss characteristics of the vibration energy during the transmission process. The first weight adjustment parameter is generated according to the deviation degree between the actual energy attenuation rate and the theoretical transmission efficiency (calculated based on the material damping coefficient). For example, when the measured attenuation rate exceeds the theoretical value by 15%, the adjustment parameter reduces the connection line weight by 0.3 to mark abnormal force transmission.

[0041] In step S204, preferably, the cooperation parameters include real-time operation indexes such as the rotational speed synchronization error and load fluctuation correlation between devices, which are monitored and obtained through a high-precision encoder and a torque sensor. The motion cooperation coefficient is calculated by the mutual information entropy algorithm within a sliding time window to quantify the dynamic matching degree of the operation states between devices. The second weight adjustment parameter is generated according to the amplitude of the deviation of the cooperation coefficient from the ideal value (such as 1.0 when completely synchronized). When the rotational speed deviation between devices continuously exceeds 2%, the coefficient decreases, triggering the attenuation of the weight value, thereby weakening the conduction influence of this connection line in the topological link.

[0042] In step S205, the device failure rate adjustment parameter is calculated by combining the ratio of the used duration to the rated duration (i.e., the life consumption rate) with the working condition severity coefficient (such as the high-temperature working condition acceleration aging coefficient). For example, when the life consumption rate reaches 80%, the node weight is amplified by a factor of 1.5 to increase the failure probability estimation. The adjustment of the node weight synchronously considers the conduction influence brought by the change of the connection line weight. For example, the weight increase of the adjacent nodes of the high-attenuation connection line is limited to avoid distortion of the failure rate assessment.

[0043] In step S206, the determination of the high-risk conduction link is based on whether the cumulative weight of the path (the superposition value of the weights of each connection line) exceeds the system safety threshold, and at the same time, it is detected whether the path contains nodes with a health degree lower than the critical value (such as the life consumption rate ). The system safety threshold can be calculated and generated through a dynamic correction model based on the weight statistical distribution of historical fault conduction paths, combined with the device fault sensitivity level and operation environment parameters. For example, if the cumulative weight of a certain conduction path reaches 4.2 (threshold 3.5) and contains two nodes with a health degree of 0.85, it is marked as a high-risk link and a warning is triggered. The final value of the node weight is obtained by correcting the inherent failure rate through the life consumption rate and is used for the iterative calculation of the fault probability of the conduction path.

[0044] This embodiment realizes the multi-dimensional verification of the abnormal state of the device and the accurate prediction of the fault conduction link by integrating the physical rule engine and the neural network model. Based on the construction of the dynamic coupling topological link and the weight dynamic adjustment mechanism, the mechanical connection relationship, the actual energy attenuation rate, and the device cooperation parameters are integrated in real time to improve the visualization analysis and quantitative evaluation ability of the fault conduction path. Through the double-parameter iterative correction of the first weight adjustment parameter of the node weight and the second weight adjustment parameter of the connection line weight, the change of the inherent failure rate of the device and the abnormality of the mechanical force transmission state are synchronously reflected, enhancing the spatio-temporal correlation of the system-level fault probability calculation. Combining the conduction path risk determination strategy triggered by the dynamic threshold, the key links containing nodes with a high life consumption rate and an excessive cumulative weight are effectively identified, providing a decision basis for equipment maintenance with both fault root cause location and conduction trend prediction, optimizing the selection of preventive maintenance timing and suppressing the risk of cascading fault spread.

[0045] Please refer to Figure 3 , further, construct a system dynamic coupling topology link according to the layout schematic diagram, and generate associated parameters in combination with abnormal device information, including: S301. Denote the node of the abnormal device as an abnormal node. Taking the abnormal node as the center, conduct a conduction search along the connection line direction of the abnormal node until the number of connection lines in the first conduction path formed by the conduction search exceeds the preset tolerance value. The preset tolerance value is generated by the product of the connection line weights in the first conduction path and the conduction threshold; S302. Denote the devices associated with the nodes covered by the first conduction path as affected devices; For each affected device, perform the following steps: S303. Obtain the connection line weights on the second conduction path from the abnormal node to the affected device; S304. Calculate the geometric mean of the connection line weights on the second conduction path as the path conduction efficiency; S305. Obtain the node weight of the abnormal node, denoted as the abnormal node weight, and obtain the node weight of the affected device, denoted as the original node weight; S306. Generate a weight influence coefficient for the affected device according to the abnormal node weight, path conduction efficiency, and original node weight; S307. Update the original node weight according to the weight influence coefficient, denoted as the influence node weight; S308. Repeat the foregoing steps until the original node weights of all affected devices are updated; S309. Generate the location information of the abnormal device in the system according to the first conduction path and the directed graph structure, and generate the device information of the other devices connected by the abnormal device according to the influence node weight and the affected device, which are the associated parameters.

[0046] In step S301, when forming the first conduction path by conduction search, the generation logic of the preset tolerance value can be understood as comparing the continuous product of the connection line weights in the first conduction path with the preset conduction threshold. When the product result is lower than the threshold, the search stops. This conduction threshold usually comes from the minimum conduction efficiency benchmark value allowed by the system, which is used to prevent low-reliability paths from being included in the analysis scope. The connection line number dynamic restriction mechanism of the first conduction path characterizes the active constraint ability of the system on the abnormal propagation range.

[0047] In step S302, the definition of the affected device depends on the topological coverage range of the first conduction path, and the mapping of the device association relationship is realized by traversing all directly or indirectly connected nodes on the first conduction path.

[0048] In step S303, the connection line weights on the second conduction path are obtained by backtracking the shortest conduction link from the abnormal node to the affected device, and the weight sequence of all connection lines on this path is extracted.

[0049] In step S304, taking the geometric mean of the connection line weights on the second conduction path as the path conduction efficiency can better reflect the actual attenuation characteristics of vibration energy among series devices. When any connection line weight on the second conduction path approaches zero, the geometric mean will decrease significantly, causing the overall conduction efficiency to decay rapidly, thereby automatically reflecting the substantial blocking state of the mechanical force transmission path.

[0050] In step S305, the abnormal node weight is obtained by calling the real-time health assessment result of the device, which is generated by correcting the inherent failure rate of the device with the life consumption rate and reflects the severity of the abnormality at the current moment. The original node weight is the baseline failure probability estimate of the affected device before being affected by conduction.

[0051] In step S306, when generating the weight influence coefficient, the dynamic balance relationship between the external abnormal input and the device's own anti-interference ability is quantified by multiplying the abnormal node weight by the path conduction efficiency and then dividing by the original node weight. The larger the coefficient value, the more it indicates that the abnormal energy conduction is not effectively absorbed by the local node and the more severely the device is affected by the abnormality.

[0052] In step S307, when updating the original node weight, a linear superposition or non-linear correction strategy of the influence coefficient and the original value can be adopted, which is specifically manifested as converting the abnormal influence on the conduction path into an incremental adjustment of the node failure probability. The acquisition of the influence node weight is essentially a mathematical modeling of fault risk conduction, and the quantitative expression of system-level abnormal diffusion is realized through weight iterative update.

[0053] In step S308, the repeated update mechanism ensures that the propagation process of abnormal energy in the topological network is completely mapped by traversing the second conduction paths of all affected devices, forming a globally linked weight correction result.

[0054] In step S309, the location information of the abnormal device in the system is the three-dimensional topological coordinates of the abnormal device, including the hierarchical coordinate and the intensity coordinate. The hierarchical coordinate is determined by the maximum hop count of the conduction path, representing the longitudinal propagation depth of the abnormality in the system topology; the intensity coordinate selects the highest influence coefficient among the affected devices, reflecting the lateral radiation intensity of the abnormal effect. The correlation parameters provide double support of topological structure and quantitative parameters for the early warning of cascading failures by establishing a dynamic coupling relationship model between the abnormal device and the rest of the system.

[0055] In this embodiment, by constructing a dynamic coupling topology link and associated parameter generation mechanism, accurate modeling of the abnormal conduction path of the device and system-level risk quantification assessment are realized. Based on the layout schematic diagram, a weighted directed graph structure of nodes and connection lines is established. The coverage range of the first conduction path of the abnormal node is determined through conduction search, the affected devices are identified, and their second conduction paths are traced back. Subsequently, the path conduction efficiency is calculated using the geometric mean, and the weight influence coefficient is generated by combining the abnormal node weight and the original node weight, and the failure probability estimates of each device are iteratively updated. Finally, the propagation depth, radiation intensity, and associated device status of the abnormal device are characterized by the location information of the abnormal device in the system and the device information of the other devices connected to the abnormal device.

[0056] This method effectively improves the visualization recognition ability and risk quantification accuracy of the cascading failure conduction path through conduction threshold to constrain the abnormal propagation range, geometric mean to reflect the path attenuation characteristics, and weight iteration correction. It provides a decision-making basis with both topological structure and dynamic parameters for abnormal root cause location and preventive maintenance strategy formulation, and enhances the active prevention and control ability of complex system fault conduction blocking.

[0057] Furthermore, according to the abnormal device information and associated parameters, the physical kinematic equation of the device status information is verified to generate the first error weight, including: Generate the actual installation parameters and theoretical installation parameters according to the abnormal device information and associated parameters; Calculate the deviation information between the actual installation parameters and the theoretical installation parameters, denoted as the installation deviation information; Perform physical deviation calculation according to the installation deviation information to obtain the calculation result, which is represented by formula (1). Formula (1) is as follows: ; In formula (1), is the position deviation, is the angle deviation, is the fit deviation, is the position sensitivity coefficient, related to the sensor type, is the actual installation position vector, is the theoretical installation position vector, is the state information space gradient, is the installation angle deviation value, is the high-order error term related to the angle, is the th harmonic amplitude of the vibration signal, is the fundamental wave amplitude; Convert the calculation result into the first error weight. The first error weight is represented by formula (2). Formula (2) is as follows: ; In formula (2), is the first error weight, is the original data of the uncalibrated device status information, is the compensated status information, is the maximum historical amplitude of the status information under the normal operating state of the device, is the sampling time, is the number of sampling points within the sliding window, is the tightness influence factor, is the tightness index, is the position deviation amount, is the position deviation threshold, is the enhancement coefficient.

[0058] In this embodiment, the installation error analysis of the device is realized by fusing the installation parameter deviation analysis and the dynamic error weight calculation through the physical kinematic equation.

[0059] According to the abnormal device information and associated parameters, by parsing the sensor type identifier, historical operation data and device association topology, the actual installation parameters and theoretical installation parameters are generated, characterizing the difference between the actual installation conditions and the design of the device in terms of spatial layout and mechanical connection.

[0060] According to the installation position deviation amount, calculate the reference value of the sensor measurement error caused thereby; combine the installation angle deviation value to correct the distortion of the status information caused by non-ideal installation; by analyzing the harmonic components of the vibration signal, infer the actual tightening state of the mechanical connection.

[0061] In formula (1), can be the calibration coefficient of the accelerometer or strain gauge, reflects the signal attenuation caused by the installation tilt, is derived from the three-dimensional positioning measurement data in the device coordinate system, is generated based on the preset coordinates of the layout schematic diagram.

[0062] In formula (2), is the calculation formula of the first error weight under the normal state (i.e., when ), and for other states except this, it is the abnormal state, and at this time the calculation formula of the first error weight is , where the tightness influence factor quantifies the amplification coefficient of the error weight caused by the loosening of the mechanical connection (the value range is 0.1 - 1.0), reflecting the non-linear law of the error amplification caused by different loosening degrees, and is calibrated through the bolt vibration table test; the tightness index reflects the actual tightening state of the mechanical connection (0 = completely loose, 1 = ideally tightened); the first error weight For subsequent fault state judgment; position deviation amount is the Euclidean distance between the actual value and the theoretical value of the installation position obtained from formula (1); position deviation threshold is the critical deviation value for triggering non - linear enhancement (usually taking the upper limit of the design tolerance); enhancement coefficient is the exponential term coefficient of the control weight with respect to the growth rate of the position deviation (the value range is 0.5 - 2.0), optimized through finite element simulation.

[0063] The process of generating the first error weight can be understood as: calculating the difference degree between the original state information and the state information after installation error compensation; weighting and amplifying the difference degree according to the tightness index; fusing the installation position deviation amount and the angle deviation value into the final weight according to a preset ratio; when there are installation parameters beyond the allowable range, performing non - linear enhancement on the weight value.

[0064] Compare the first error weight with the dynamic threshold, and mark it as a physical verification anomaly when it exceeds the threshold. Preferably, the dynamic threshold is adaptively adjusted based on the statistical distribution of the error weights under the normal operating conditions of the device's history, combined with real - time operating environment parameters, for distinguishing physical verification anomalies from random fluctuations within the allowable range.

[0065] The verification process of this embodiment realizes the multi - parameter coupling analysis of the mechanical connection state, spatial positioning accuracy, and signal acquisition quality through physical modeling of installation parameter deviation and dynamic weight calculation, providing a quantitative criterion for equipment health state assessment that simultaneously considers static installation errors and dynamic operating characteristics. The generation logic of the error weight focuses on reflecting the conduction and amplification effect of installation defects on state information, and enhances the detection sensitivity to concealed installation faults through the non - linear superposition mechanism of the tightness index and the position deviation.

[0066] Please refer to Figure 4 , further, the device state information includes abnormal vibration signals and abnormal spectrum characteristics generated by abnormal device operation, and the historical condition fault information includes historical vibration signals, historical spectrum characteristics, and historical operating parameters; Match the historical condition fault information of the abnormal device in the historical database according to the abnormal device information, calculate the fitness degree between the current device state information and the historical condition fault information, and generate the second error weight including: S401. Calculate the waveform similarity between the abnormal vibration signal and the historical vibration signal, and calculate the energy distribution difference value between the abnormal spectrum characteristic and the historical spectrum characteristic within the abnormal frequency band; S402. Correct the waveform similarity and the energy distribution difference value according to the difference information between the operating parameters in the abnormal device information and the historical operating parameters; S403. Fuse the corrected waveform similarity and the corrected energy distribution difference value to generate the second error weight.

[0067] In this embodiment, through time-frequency domain feature matching and working condition offset compensation, a historical experience quantitative evaluation of the abnormal state of the equipment is realized.

[0068] In step S401, the waveform similarity between the abnormal vibration signal and the historical vibration signal is calculated through the dynamic time warping algorithm for time-domain feature comparison, eliminating the difference in cycle length caused by rotational speed fluctuations. The similarity score is obtained by calculating the minimum cumulative path distance, characterizing the morphological consistency of the time-domain waveform after elastic alignment on the time axis; preferably, the root mean square deviation method of the frequency band energy ratio is used to calculate the energy distribution difference value between the abnormal spectrum feature and the historical spectrum feature in the abnormal frequency band for frequency-domain feature comparison, extracting the amplitude spectrum data in the preset abnormal frequency band, and respectively calculating the Euclidean distance of the energy distribution of the current and historical spectra at each frequency point, which is used to quantify the offset degree of high-frequency resonance or specific harmonic components.

[0069] In step S402, the waveform similarity and the energy distribution difference value are corrected by constructing a mapping function of the parameter difference rate and the feature confidence degree to achieve working condition offset compensation. Specifically, by calculating the percentage difference between the operating parameters such as the load rate and rotational speed value of the current equipment and the historical cases, a parameter offset coefficient is generated; subsequently, the S-shaped function is used to perform attenuation compensation on the waveform similarity and proportional amplification on the energy distribution difference value.

[0070] The process of using the S-shaped function to perform attenuation compensation on the waveform similarity can be understood as: inputting the operating parameter difference rate into the S-shaped function to generate a [0,1] attenuation coefficient. The larger the parameter difference, the smaller the output value, so that the waveform similarity decays inversely with the degree of working condition offset. The process of using the S-shaped function to perform proportional amplification on the energy distribution difference value can be understood as: by adjusting the slope and offset of the S-shaped function, the parameter difference rate is mapped to interval, where is the preset maximum amplification factor, which is determined by the system according to engineering experience or experimental calibration. When the operating parameter difference rate reaches the maximum threshold, the output value of the S-shaped function approaches , ensuring that there is a controllable upper limit for the energy difference amplification (such as represents the highest amplification of 3 times), avoiding excessive correction under extreme working conditions. The higher the difference rate, the larger the output value, and the energy distribution difference value is multiplied by this coefficient to achieve amplification.

[0071] This step is used to eliminate the interference of working condition fluctuations on feature matching. For example, the systematic increase in the amplitude of the vibration signal under high-load working conditions should not affect the determination of the similarity of the waveform morphology.

[0072] In step S403, preferably, the process of obtaining the second error weight includes a hierarchical feature fitness determination mechanism: when both the corrected waveform similarity (time-domain feature) and the energy distribution difference value (frequency-domain feature) exceed the first threshold, it indicates that the current working condition highly matches the historical optimal case. At this time, directly use the historical minimum error weight as the second error weight to inherit the verified diagnostic conclusion and avoid repeated calculations; when either the waveform similarity or the energy distribution difference value is between the first threshold and the second threshold, use the sliding weighted average algorithm to dynamically adjust the weight ratio according to the deviation degrees of the waveform similarity and the energy distribution difference value, and correct the second error weight to ensure that partially matching cases can still contribute to the correction reference; when both the corrected waveform similarity and the corrected energy distribution difference value are lower than the second threshold, it is identified as a new abnormal mode, and a preset maximum error weight is given to trigger the independent analysis module.

[0073] Preferably, the first threshold is set according to the normal working condition parameter fluctuation range, and the second threshold is determined by the historical extreme value or the system tolerance upper limit. The two are combined through statistical analysis and experimental calibration to divide the difference rate interval to balance sensitivity and stability.

[0074] This embodiment calculates the waveform similarity between the abnormal vibration signal and the historical vibration signal based on the dynamic time warping algorithm to eliminate the influence of rotational speed fluctuations on the time-domain waveform shape comparison. At the same time, the root mean square deviation method is used to quantify the energy distribution difference value within the abnormal frequency band to capture high-frequency resonance or harmonic offset characteristics; a non-linear mapping between the operating parameter difference rate and the feature confidence level is constructed, and the waveform similarity is attenuated and corrected and the energy distribution difference value is proportionally amplified through the working condition offset compensation to eliminate the feature matching interference caused by working condition fluctuations such as load and rotational speed; based on the dual-threshold hierarchical determination mechanism of the first threshold and the second threshold (the first threshold comes from the normal fluctuation range, and the second threshold is determined by the historical extreme value), the corrected waveform similarity and the corrected energy distribution difference value are fused, and the second error weight is dynamically generated according to the full match, partial match, and new abnormal scenarios.

[0075] This method effectively improves the robustness of abnormal diagnosis through time-frequency joint analysis and working condition adaptive correction: it can inherit the historical optimal conclusion to avoid redundant calculations when highly matching, and can also retain the effective correction reference through the sliding weighted algorithm in partially matching scenarios, while ensuring the triggering of independent analysis for new abnormal modes, achieving a balance between historical experience reuse and unknown fault identification, and taking into account diagnostic sensitivity and system stability.

[0076] Furthermore, a risk level assessment is performed on the fault status information, and the detection modes generated according to the assessment results include: Divide the abnormal monitoring sequence in the fault status information into multiple monitoring points according to the unit time, and each monitoring point has a monitoring value; Construct an abnormal monitoring sequence curve graph, where the vertical axis of the abnormal monitoring sequence curve graph is the monitoring value, and the horizontal axis of the abnormal line monitoring sequence curve graph is the monitoring time; Divide the high-risk monitoring interval, medium-risk monitoring interval, and low-risk monitoring interval in the abnormal monitoring sequence curve graph, and divide multiple monitoring values to form a high-risk set, a medium-risk set, and a low-risk set; Calculate the proportion of the monitoring values in the high-risk set, medium-risk set, and low-risk set in the current abnormal monitoring sequence, and generate an evaluation result of the abnormal monitoring sequence; If the evaluation result is high risk, monitor the abnormal device according to the high-frequency detection mode; If the evaluation result is medium risk, monitor the abnormal device according to the low-frequency detection mode; If the evaluation result is low risk, maintain the initial detection mode.

[0077] In this embodiment, through the time-series feature analysis of the abnormal monitoring sequence and the dynamic division of the risk interval, the hierarchical evaluation of the device fault state and the adaptive adjustment of the detection mode are realized.

[0078] Dividing the abnormal monitoring sequence into multiple monitoring points per unit time can be understood as segmenting the continuously collected abnormal monitoring data at a fixed time interval (such as every minute or every hour). Each segment corresponds to a monitoring point, and its monitoring value is the statistic of the sensor data within that time period (such as the mean, peak value, or variance), which is used to quantify the abnormal degree of the device during that time period.

[0079] When constructing the abnormal monitoring sequence curve graph, the monitoring value on the vertical axis reflects the fluctuation intensity of the device abnormal state, and the monitoring time on the horizontal axis marks the time-series trajectory of the abnormal evolution. The abnormal mutation points or trend accumulation effects can be visually identified through the curve shape.

[0080] Dividing the high-risk monitoring interval, medium-risk monitoring interval, and low-risk monitoring interval can be understood as setting vertical numerical boundaries in the curve graph based on preset thresholds or historical statistical distributions, and dividing the vertical axis into numerical ranges corresponding to different risk levels. For example, the interval exceeding the upper threshold is defined as the high-risk monitoring interval, the interval between the upper and lower limits is the medium-risk monitoring interval, and the interval below the lower limit is the low-risk monitoring interval, so as to classify the monitoring values of each monitoring point into the corresponding risk sets.

[0081] By counting the number of monitoring points falling into the high-risk set, medium-risk set, and low-risk set in the current anomaly monitoring sequence, and calculating their respective proportions of the total number of monitoring points. For example, if the proportion of the high-risk set reaches 60%, it is evaluated as high risk, which is used to characterize the persistence and severity of the abnormal state. The high-frequency detection mode means shortening the data acquisition cycle (such as increasing from once every 10 minutes to once every minute) to enhance the monitoring density, and is applicable to quickly capturing the details of abnormal evolution in a high-risk state; the low-frequency detection mode extends the acquisition interval (such as adjusting from once every 10 minutes to once every 30 minutes), and balances resource consumption and monitoring requirements in a medium-risk state; the initial detection mode maintains the original acquisition frequency and is applicable to normal monitoring in a low-risk state.

[0082] Among them, the threshold definition of the risk interval can be dynamically adjusted in combination with the device type, historical failure data, and industry standards. For example, the numerical interval corresponding to the typical abnormal mode is divided through cluster analysis, or the threshold boundary is updated according to the recent data distribution using the sliding window statistical method. The evaluation rule for the proportion of monitoring values can adopt the weighted cumulative method. For example, a higher weight is assigned to the high-risk set, and when the weighted score exceeds the set threshold, the corresponding risk level is determined. The detection mode switching mechanism needs to embed lag filtering processing to avoid frequent mode jumps caused by instantaneous fluctuations. For example, when three consecutive evaluation cycles are all high risk, it switches to the high-frequency mode.

[0083] In this embodiment, through the time-series feature analysis of the anomaly monitoring sequence and the dynamic division of the risk interval, the hierarchical evaluation of the device fault state and the adaptive adjustment of the detection mode are realized. Its method process includes: dividing the continuous anomaly monitoring sequence into discrete monitoring points with monitoring values according to the unit time, constructing an anomaly monitoring sequence curve graph with the vertical axis being the monitoring value and the horizontal axis being the monitoring time, dividing the high-risk monitoring interval, medium-risk monitoring interval, and low-risk monitoring interval based on the preset or dynamically adjusted vertical threshold, counting the proportion of monitoring points in each risk set to generate the evaluation result, and finally automatically switching to the high-frequency, low-frequency, or initial detection mode according to the determination of high risk, medium risk, or low risk.

[0084] This method is based on the dynamic division of the vertical threshold of the monitoring curve graph, optimizes the risk interval boundary by combining the device type and historical data, and can accurately capture the intensity and evolution trend of abnormal fluctuations; through the statistical analysis of the proportion of monitoring points and the weighted evaluation rule, it takes into account the characterization of the persistence of the abnormal state and the response to sudden events, and avoids misjudgment by a single threshold; the detection mode switching mechanism embeds lag filtering processing to suppress the interference of instantaneous fluctuations and ensure the stability of mode adjustment; the dynamically adjusted high-frequency detection mode and low-frequency detection mode optimize the resource utilization rate in medium- and low-risk scenarios while ensuring the monitoring density in high-risk states, forming a closed-loop adaptive control system, and achieving the balance between the sensitivity of fault warning and the economic operation of the system.

[0085] Furthermore, the neural network model includes a physical decoupling model. The system dynamic coupling topological link, associated parameters, abnormal device information, and fault status information are input into the neural network model, and the obtained fault link information includes: Extract the forward conduction path of the abnormal device from the system dynamic coupling topological link, mark the devices in the forward conduction path as fault source devices, generate topological decoupling features, and extract the physical connection features between the fault source devices and the abnormal devices from the associated parameters; Input the topological decoupling features and physical connection features into the trained physical decoupling model to obtain the forward traceability prediction information for the fault of the abnormal device, including: Extract the fault contribution probabilities of multiple fault source devices according to the topological decoupling features and physical connection features; Adjust the fault contribution probability of each fault source device according to the physical constraint features. The physical constraint features include energy conservation constraint and time sequence propagation constraint. Eliminate the fault source devices that do not meet the physical constraint features and generate the corrected contribution probabilities of the remaining fault source devices; Generate the forward traceability prediction information for the fault of the abnormal device according to the corrected contribution probabilities and their corresponding fault source devices.

[0086] In this embodiment, extracting the forward conduction path of the abnormal device from the system dynamic coupling topological link can be understood as identifying the device link that may conduct faults upstream of the abnormal device based on the real-time connection relationship and interaction intensity between system nodes, and using a weighted path search algorithm (such as Dijkstra's algorithm). Mark the devices with direct or indirect dynamic coupling relationships with the abnormal device in the path as fault source devices for locating potential fault propagation sources.

[0087] Extract structural attributes such as node connectivity, path weight distribution, and conduction directionality from the topological link. For example, quantify the hub role of each node in fault propagation by calculating the betweenness centrality of each node, or characterize the activity degree of the conduction path by the cumulative value of the path weights as topological decoupling features for describing the structural features of fault propagation.

[0088] Parse physical interaction parameters such as material properties, mechanical connection stiffness, or energy transfer coefficients between the fault source device and the abnormal device from the system parameter library. For example, extract the torque transfer coefficient of the coupling or the fluid resistance coefficient of the pipeline as physical connection features for quantifying the fault conduction ability at the physical level.

[0089] The training of the physical decoupling model uses the verified propagation chains in historical fault cases as positive samples, generates negative samples by simulating situations that violate physical laws such as abnormal energy transfer or time sequence inversion. During the training process, the sensitivity of the model to physical constraints is strengthened through a contrastive loss function, and at the same time, an adversarial training mechanism is combined to improve the robustness to noise interference, enabling the model to embed physical prior knowledge while retaining the advantages of data-driven, enhancing the interpretability and engineering credibility of the prediction results.

[0090] The physical decoupling model inputs topological features (such as node connection degree, conduction path weight) and physical features (such as material conduction coefficient, mechanical impedance) into a multi-layer perceptron network, establishes a topological-physical coupling relationship through a feature cross-fusion layer, and uses an attention mechanism to calculate the state correlation strength of each potential fault source device to the abnormal device, and outputs a normalized weight value as the fault contribution probability, representing the relative influence degree of the device in the conduction path.

[0091] The energy conservation constraint means verifying whether the energy transfer efficiency from the fault source device to the abnormal device conforms to the laws of thermodynamics or the mechanical efficiency threshold, for example, checking whether the attenuation rate of vibration energy on the conduction path is within a reasonable range; the time sequence propagation constraint means verifying whether the arrival time sequence of the fault feature signal in the topological link conforms to the relationship between the wave propagation speed and the path length, for example, judging the causal rationality by comparing the path distance and the signal propagation time delay. For devices that meet the constraints, the original probability is retained, and for devices that violate the constraints, the weight is reduced or set to zero according to the attenuation coefficient, and finally the modified contribution probability is generated by normalization to ensure that the result conforms to the actual physical laws and propagation path logic.

[0092] Arrange the modified contribution probabilities in descending order, and output the list of fault source devices that meet the physical constraints and their associated conduction paths, which is the forward traceability prediction information, providing a priority guidance for fault troubleshooting.

[0093] In this embodiment, the forward conduction path of the abnormal device is extracted based on the real-time topological link and the topological decoupling features are generated, and the conduction ability is quantified by combining the physical connection parameters; the initial contribution probability of the fault source device is extracted through the physical decoupling model, and the topological-physical two-dimensional feature cross-fusion and the attention mechanism are used to strengthen the recognition ability of the key nodes of the conduction path; then through the physical constraint feature verification mechanism, the abnormal prediction results that violate the laws of thermodynamics or signal propagation logic are eliminated, the false correlation interference is filtered, and the modified contribution probability is generated to ensure that the prediction result is consistent with the actual physical laws of the project; the list of fault sources that conform to the physical laws and are sorted by the degree of influence is output, which is the forward traceability prediction information. This method enables the physical decoupling model to still maintain a robust discrimination ability for the real fault conduction link in a complex noise environment.

[0094] Further, the neural network model includes a physical coupling model. Inputting the system dynamic coupling topological link, associated parameters, abnormal device information, and fault status information into the neural network model, the obtained fault link information further includes: Extract the backward conduction path of the abnormal device in the system dynamic coupling topological link, mark the devices in the backward conduction path as risk devices, generate topological coupling features, and extract the physical coupling features between the risk devices and the abnormal devices from the associated parameters; Input the topological coupling features and physical coupling features into the trained physical coupling model to generate the backward fault prediction information of the abnormal device, including: Calculate the inherent conduction risk value of the risk device according to the topological coupling features; Calculate the risk correction coefficient related to the working condition according to the physical coupling features; Fuse the risk correction coefficient and the inherent conduction risk value to obtain the conduction risk probability of each risk device; Verify the multiple conduction risk probabilities by applying the mass conservation constraint, eliminate the risk devices that violate the mass conservation constraint, correct the conduction risk probabilities, obtain the corrected risk probabilities, and generate the time period corresponding to the current corrected risk probability; Generate the backward fault prediction information and time period of the abnormal device according to the corrected risk probabilities and their corresponding risk devices.

[0095] In this embodiment, extracting the backward conduction path of the abnormal device in the system dynamic coupling topological link can be understood as being implemented by a reverse path search algorithm (such as improved breadth-first search), which represents the potential path network where the abnormal device fault may spread to downstream devices, used to identify the secondary risk device group that may be affected by the current abnormal state. Its search direction is opposite to the actual physical flow direction of the fault conduction, but the dynamic coupling strength weights between nodes are retained to reflect the conduction probability.

[0096] The generation of the topological coupling features includes calculating the cumulative value of the shortest path weights from each risk device to the abnormal device, which is used to quantify the activity of the conduction path at the topological level; at the same time, extract the nodes with high betweenness centrality in the path as key coupling points, which reflect their hub role in the fault diffusion process. For example, screen out the key nodes by counting the occurrence frequency of the nodes in the conduction path to enhance the ability to identify the conduction bottleneck.

[0097] Preferably, the physical coupling features include the dynamic interaction coefficient and the failure sensitivity. Extracting the physical coupling features of risk devices and abnormal devices from the correlation parameters can be understood as follows: the dynamic interaction coefficient quantifies the physical connection strength between risk devices and abnormal devices by analyzing parameters such as the mechanical transmission efficiency, fluid resistance coefficient, or conductivity between devices in the system parameter library; the failure sensitivity calculates the performance degradation rate of risk devices in the current operating state based on real-time operating condition data (such as vibration amplitude, temperature gradient) and device design thresholds, and is used to evaluate their vulnerability to external fault conduction.

[0098] The physical coupling model is trained using a curriculum learning strategy. Initially, only basic topological constraints are introduced, and complex physical rules such as mass conservation and power flow balance are gradually superimposed, enabling the model to master the mapping relationship from structural features to physical laws in stages; adversarial training forces the model to distinguish false conduction paths by generating adversarial samples with high topological coupling but violating physical constraints. For example, a contradictory feature combination with a high topological path weight but a zero dynamic interaction coefficient is constructed to enhance the model's robustness to abnormal operating conditions. Positive samples are sourced from the verified subsequent fault device chains in the historical fault logs, while negative samples are generated by randomly replacing the end device of the conduction path or inserting physically unreachable nodes, ensuring that the samples cover both real conduction and physical violation scenarios. The loss function design emphasizes the punishment for physical constraint violations. For example, an additional loss term is imposed on the prediction results that violate mass conservation, and at the same time, contrastive loss is used to strengthen the difference between positive and negative samples, enabling the physical coupling model to accurately separate conduction modes that conform to engineering reality in the feature space.

[0099] The calculation of the inherent conduction risk value fuses the path weight and the critical node density in the topological coupling features through a multi-layer perceptron network, and outputs the basic probability reflecting the threat of device failure diffusion at the structural level; the risk correction coefficient is weighted and calculated through the dynamic interaction coefficient and the failure sensitivity in the physical coupling features. For example, the coefficients of devices with high transmission efficiency and low failure sensitivity are adjusted downwards to suppress their risk contribution; the fusion of the conduction risk probability uses a gating mechanism to dynamically adjust the weight ratio of topological and physical features. For example, under high-pressure operating conditions, the decision weight of physical coupling features is strengthened, enabling the model to adapt to the changes in conduction characteristics under different operating environments.

[0100] The verification of the mass conservation constraint is achieved by comparing the input-output deviation of the material flow or energy flow on the fault conduction path. For example, in a hydraulic system, it is verified whether the flow rate loss caused by fault propagation is within the allowable range of pipeline leakage. If the deviation exceeds the threshold, it is determined that the path violates the conservation law, and the corresponding risk equipment is then eliminated. The time-effective section for correcting the risk probability is generated based on the time-delay distribution law of similar conduction paths in historical fault cases, and the time window for the propagation of the fault impact is estimated in combination with the current system operation rate (such as the rotational speed of a rotating machine). For example, the earliest and latest time intervals when abnormal conditions may occur in the risk equipment are estimated by the ratio of the path length to the propagation speed of the medium.

[0101] In this embodiment, a backward fault prediction mechanism is constructed by integrating the structural characteristics of the system's dynamic coupling topological link and the physical interaction parameters. Based on the reverse path search algorithm, the potential conduction paths of abnormal equipment are extracted and the risk equipment is marked to generate topological coupling features and physical coupling features. The inherent conduction risk value and the operating condition-related risk correction coefficient are calculated through a physical coupling model, and a gating mechanism is used to dynamically fuse the two to generate the conduction risk probability. The mass conservation constraint is applied to eliminate the illegal paths and correct the risk probability, and the time-effective section is estimated by combining the historical time-delay law and the operation rate. This method effectively improves the positioning accuracy of multi-source concurrent faults through the topological-physical two-dimensional feature cross-validation and dynamic weight adjustment mechanism; the curriculum learning strategy and adversarial training are combined to ensure the robustness of the model under complex operating conditions, and the mass conservation constraint excludes false conduction paths; the time-effective section prediction provides a clear time window for operation and maintenance decision-making, and the gating mechanism ensures that the model adaptively adjusts the contribution weights of the topological coupling features and the physical coupling features, enhancing the adaptability in different operating environments, and finally forming a fault conduction prediction system that conforms to engineering physical laws and has time-sequence guiding value.

[0102] In a second aspect, this embodiment also provides a computer-readable storage medium, on which computer program instructions are stored, and the computer program instructions, when executed by a processor, implement the method described in the first aspect.

[0103] The computer program involved in this embodiment can be stored in a computer-readable storage medium, which includes but is not limited to magnetic disks, magnetic tapes, magnetic cards, floppy disks, flash memories, optical discs, optical cards, read-only memories (ROMs), random access memories (RAMs), erasable programmable ROMs (EPROMs), and electrically erasable programmable ROMs (EEPROMs), etc. It also includes other biological, physical, or chemical structures that can achieve the same or equivalent functions as the above-listed storage media, such as units with information storage capabilities like DNA, RNA, proteins, etc. In a specific embodiment, the storage medium involved can be one of the above medium types or a combination of the above medium types. In different embodiments, the computer program involved in the embodiment can be centrally stored in a single medium or distributedly stored in multiple media. The memory containing the computer-readable storage medium can be a non-volatile memory or a random access memory. These computer-readable storage media can be built into the device or can be an external device or a part of an external device connected to the device involved in the embodiment. In some embodiments, the memory with the computer-readable storage medium is deployed locally; in other embodiments, a scheme of deploying the memory away from the processor can also be adopted, such as a network-attached memory accessed via an RF circuit or an external port and a communication network, where the communication network can be the Internet, one or more internal networks, a local area network (LAN), a wide area wireless network (WLAN), a storage area network (SAN), etc., or a suitable combination thereof, as long as it can enable the computer device to access the memory. In addition, the computer program involved in the embodiment can be stored in plaintext / ciphertext form or can be designed as training data and be integrally reorganized and implicitly stored in the parameter states of a deep neural network or other machine learning models through model training.

[0104] Please refer to Figure 5 , in a third aspect, this embodiment also provides an electronic device 1, including a memory 11 and a processor 12, where the memory 11 is used to store one or more computer program instructions, and wherein, the one or more computer program instructions are executed by the processor 12 to implement the method described in the first aspect.

[0105] The processor described in this embodiment can be implemented by hardware, firmware, software, or a combination thereof. It can use circuits, one or more application specific integrated circuits (ASICs), digital signal processors (DSPs), digital signal processing devices (DSPDs), programmable logic devices (PLDs), field programmable gate arrays (FPGAs), central processing units (CPUs), controllers, microcontrollers, microprocessors, or at least one of the like, and also includes other physical, biological, or chemical structures that can achieve functions similar to or equivalent to those of the above-listed processors, such as biological neurons, quantum computing units, DNA computing units, etc., so that the processor can execute some steps, all steps, or any combination of the steps mentioned in the computer programs or methods involved in the various embodiments of the present application.

[0106] Different from the prior art, the above technical solution has the following beneficial effects: The present invention realizes the system-level prediction and early warning of the device fault conduction path by constructing a neural network model with deep embedding of dynamic coupling topology links and physical rules. Specifically, the above technical solution abstracts the mechanical connection relationship and energy transfer characteristics between devices into a weighted topology network, generates multi-dimensional error weights by combining the verification of physical kinematic equations and historical working conditions matching, reversely locates the faulty source device through a physical decoupling model, and predicts the backward conduction path and time-effective section by using a physical coupling model. The topological coupling feature and the physical coupling feature are dynamically fused through a gating mechanism, and the conduction paths that violate physical laws are eliminated by combining the mass conservation constraint, forming a fault propagation prediction system that conforms to engineering practice. This method breaks through the limitation of traditional single-device monitoring, integrates complex physical rules in stages through a curriculum learning strategy, and enhances the model's recognition ability for topological-physical contradictory samples through adversarial training, simultaneously improving the multi-source fault location accuracy and the adaptability to abnormal working conditions; calculates the time-effective section based on the path length and the medium propagation speed, and predicts the potential spread sequence by combining the Monte Carlo path simulation, realizing the full-link spatio-temporal dynamic early warning from single-point anomalies to system-level risks, providing decision support for fault conduction blocking with both physical interpretability and dynamic adaptability for Internet of Things devices, and significantly improving the operation reliability of complex systems.

[0107] Finally, it should be noted that although the above embodiments have been described in the text and drawings of the specification of this application, the patent protection scope of this application cannot be limited thereby. Any technical solutions obtained by equivalent structure or equivalent process substitution or modification based on the essential concept of this application and using the content recorded in the text and drawings of the specification of this application, as well as the direct or indirect implementation of the technical solutions of the above embodiments in other related technical fields, etc., are all included in the patent protection scope of this application.

Claims

1. A device failure early warning method based on a neural network, characterized in that: The method comprises: Obtain device status information regularly to determine whether the device status information is abnormal. If so, record it as an abnormal device and trigger an abnormal evaluation step, including: Acquiring abnormal device information, wherein the abnormal device information includes operating parameters and structural parameters; Obtaining a layout schematic diagram of the system where the abnormal device is located, constructing a system dynamic coupling topology link according to the layout schematic diagram, and generating association parameters in combination with the abnormal device information, wherein the association parameters are the location information of the abnormal device in the system and the device information of the remaining devices connected to the abnormal device; Performing physical kinematic equation verification on the device status information according to the abnormal device information and associated parameters to generate a first error weight; Matching the historical working condition fault information of the abnormal equipment in the historical database according to the abnormal equipment information, calculating the adaptability of the equipment state information and the historical working condition fault information, and generating a second error weight; Generate fault status information of the device according to the first error weight and the second error weight; Perform risk level assessment on fault status information and generate a detection mode based on the assessment result. The detection mode includes a high-frequency detection mode and a low-frequency detection mode. The system dynamic coupling topology link, associated parameters, abnormal equipment information and fault status information are input into the neural network model to obtain the fault link information, which includes the forward tracing prediction information of the abnormal equipment fault, the backward fault prediction information of the abnormal equipment and the time-effect section; The first warning information is generated according to the fault link information, the fault status information and the abnormal device information. When the detection mode is switched to the high-frequency detection mode, the low-frequency detection mode of the remaining devices in the current fault link information is triggered to generate the second warning information.

2. The neural network-based equipment failure early warning method according to claim 1 is characterized in that: Constructing a system dynamic coupling topology link according to the layout schematic diagram includes: Obtaining the mechanical connection relationship in the layout schematic diagram, and constructing a directed graph structure with a single device as a node and mechanical connections as connecting lines, and weight initializing the directed graph structure, including: The node weight of each node is generated by the inherent failure rate of a single device, and the connection line weight of each connection line is obtained by calculating the material strength and design load of the nodes at both ends of the connection line; Monitoring vibration signals between devices of each connection line, generating an actual energy decay rate according to the vibration signal, and generating a first weight adjustment parameter according to the actual energy decay rate; Monitoring the coordination parameters of the devices of each connection line, calculating the motion coordination coefficient between the two devices according to the coordination parameters, and generating a second weight adjustment parameter according to the motion coordination coefficient; Adjust the connection line weight according to the first weight adjustment parameter and the second weight adjustment parameter; And, obtaining the used time and rated time of each device, and generating a device failure rate adjustment parameter according to the used time and rated time; The node weight is adjusted according to the device failure rate adjustment parameter.

3. The neural network-based equipment failure early warning method according to claim 2 is characterized in that: According to the layout schematic diagram, the system dynamic coupling topology link is constructed, and the associated parameters generated by combining the abnormal device information include: The node of the abnormal device is recorded as an abnormal node, and a conduction search is performed along the connection line direction of the abnormal node with the abnormal node as the center, until the number of connection lines in the first conduction path formed by the conduction search exceeds a preset tolerance value, and the preset tolerance value is generated by multiplying the connection line weight product in the first conduction path by a conduction threshold; Recording the device associated with the node covered by the first conductive path as an affected device; Perform the following steps for each affected device: obtaining a connection line weight on a second conduction path from the abnormal node to the affected device; Calculating the geometric mean of the connection line weights on the second conduction path as the path conduction efficiency; Obtain the node weight of the abnormal node, recorded as the abnormal node weight, and obtain the node weight of the affected device, recorded as the original node weight; Generate weight influence coefficients of affected devices according to abnormal node weights, path conduction efficiency, and original node weights; Update the original node weight according to the weight influence coefficient, which is recorded as the influence node weight; Repeat the above steps until the original node weights of all affected devices are updated; The location information of the abnormal device in the system is generated according to the first conduction path and the directed graph structure, and the device information of the remaining devices connected to the abnormal device is generated according to the influencing node weight and the affected device, which is the association parameter.

4. The neural network-based equipment failure early warning method according to claim 1 is characterized in that: Performing a physical kinematics equation check on the device status information according to the abnormal device information and associated parameters to generate a first error weight includes: Generate actual installation parameters and theoretical installation parameters according to the abnormal device information and associated parameters; Calculating deviation information between the actual installation parameter and the theoretical installation parameter, and recording the deviation information as installation deviation information; The physical deviation is calculated according to the installation deviation information to obtain a calculation result, which is expressed by formula (1). The formula (1) is as follows: ； In formula (1), is the position deviation, is the angle deviation, To match the deviation, is the position sensitivity coefficient, which is related to the sensor type. is the actual installation position vector, is the theoretical installation position vector, is the state information space gradient, is the installation angle deviation value, is the angle-related high-order error term, The vibration signal Subharmonic amplitude, is the fundamental amplitude; The first error weight is converted according to the calculation result. The first error weight is expressed by formula (2). The formula (2) is as follows: ； In formula (2), is the first error weight, It is the uncorrected raw data of the equipment status information. is the status information after compensation, is the maximum historical amplitude of the status information under normal operation of the equipment, is the sampling time, is the number of sampling points in the sliding window, is the tightness influencing factor, is the tightness index, is the position deviation, is the position deviation threshold, is the enhancement factor.

5. The neural network-based equipment failure early warning method according to claim 1 is characterized in that: The equipment status information includes abnormal vibration signals and abnormal spectrum characteristics generated by abnormal equipment operation, and the historical operating condition fault information includes historical vibration signals, historical spectrum characteristics and historical operating parameters; Matching the historical working condition fault information of the abnormal equipment in the historical database according to the abnormal equipment information, calculating the adaptability of the current equipment state information and the historical working condition fault information, and generating a second error weight includes: Calculating the waveform similarity between the abnormal vibration signal and the historical vibration signal, and calculating the energy distribution difference value between the abnormal spectrum feature and the historical spectrum feature in the abnormal frequency band; Correct the waveform similarity and energy distribution difference value according to the difference information between the operating parameters in the abnormal device information and the historical operating parameters; The corrected waveform similarity and the corrected energy distribution difference value are fused to generate the second error weight.

6. The neural network-based equipment failure early warning method according to claim 1 is characterized in that: Performing a risk level assessment on the fault status information and generating a detection mode according to the assessment result includes: Dividing the abnormal monitoring sequence in the fault status information into a plurality of monitoring points according to unit time, each monitoring point having a monitoring value; Constructing an abnormal monitoring sequence curve graph, wherein the ordinate of the abnormal monitoring sequence curve graph is the monitoring value, and the abscissa of the abnormal monitoring sequence curve graph is the monitoring time; In the abnormal monitoring sequence curve diagram, a high-risk monitoring interval, a medium-risk monitoring interval and a low-risk monitoring interval are divided, and a plurality of monitoring values ​​are divided to form a high-risk set, a medium-risk set and a low-risk set; Calculate the proportion of the monitoring values ​​of the high-risk set, the medium-risk set, and the low-risk set in the current abnormal monitoring sequence, and generate an evaluation result of the abnormal monitoring sequence; If the assessment result is high risk, the abnormal equipment will be monitored according to the high-frequency detection mode; If the assessment result is medium risk, the abnormal equipment will be monitored according to the low-frequency detection mode; If the assessment result is low risk, the initial testing mode will be maintained.

7. The neural network-based equipment failure early warning method according to claim 1 is characterized in that: The neural network model includes a physical decoupling model, and the system dynamic coupling topology link, associated parameters, abnormal device information and fault status information are input into the neural network model to obtain the fault link information including: Extract the forward conduction path of the abnormal device in the system dynamic coupling topology link, record the device in the forward conduction path as the fault source device, generate the topological decoupling feature, and extract the physical connection feature between the fault source device and the abnormal device from the associated parameters; The topological decoupling features and physical connection features are input into the trained physical decoupling model to obtain forward tracing prediction information that causes abnormal equipment failures, including: Extract the fault contribution probability of multiple fault source devices based on topological decoupling characteristics and physical connection characteristics; The fault contribution probability of each of the fault source devices is adjusted according to the physical constraint characteristics, wherein the physical constraint characteristics include energy conservation constraints and timing propagation constraints, the fault source devices that do not meet the physical constraint characteristics are eliminated, and the modified contribution probability of the retained fault source devices is generated; According to the modified contribution probability and its corresponding fault source device, forward tracing prediction information causing the abnormal device failure is generated.

8. The neural network-based equipment failure early warning method according to claim 7 is characterized in that: The neural network model includes a physical coupling model, and the system dynamic coupling topology link, associated parameters, abnormal device information and fault status information are input into the neural network model to obtain the fault link information, which also includes: Extract the backward conduction path of abnormal devices in the system dynamic coupling topology link, mark the devices in the backward conduction path as risk devices, generate topological coupling features, and extract the physical coupling features of risk devices and abnormal devices from the associated parameters; The topological coupling features and physical coupling features are input into the trained physical coupling model to generate backward fault prediction information of abnormal equipment, including: Calculate the inherent conduction risk value of the risk equipment based on the topological coupling characteristics; Calculate the risk correction factor related to the working condition based on the physical coupling characteristics; The risk correction coefficient is combined with the inherent conduction risk value to obtain the conduction risk probability of each risk device; Applying mass conservation constraints to verify the multiple conduction risk probabilities, eliminating risk equipment that violates the mass conservation constraints, and correcting the conduction risk probabilities to obtain corrected risk probabilities, and generating a time limit section corresponding to the current corrected risk probabilities; According to the modified risk probability and the corresponding risk equipment, backward fault prediction information and time-effect section of the abnormal equipment are generated.

9. A computer-readable storage medium having computer program instructions stored thereon, characterized in that: The computer program instructions implement the method of any one of claims 1 to 8 when executed by a processor.

10. An electronic device comprising a memory and a processor, characterized in that: The memory is used to store one or more computer program instructions, wherein the one or more computer program instructions are executed by the processor to implement the method according to any one of claims 1 to 8.

Images