An energy-saving fan fault prediction method and system based on multi-sensor fusion

Abstract

The application discloses an energy-saving fan fault prediction method and system based on multi-sensor fusion, and relates to the technical field of intelligent operation and maintenance, comprising: applying short-time rotating speed disturbance in the stable operation section of the energy-saving fan, and synchronously collecting inlet pressure, outlet pressure, flow, motor current, motor power, bearing vibration, sealing cavity acoustic emission, sealing seat temperature and shell thermal image to obtain a disturbance period synchronous time sequence section; reorganizing the disturbance period synchronous time sequence section according to the periods before, during and after disturbance, and combining the continuous enhancement of the sealing cavity acoustic emission, the slow recovery of the outlet pressure, the lag of the motor power backfall and the expansion of the sealing circumferential thermal spot to determine the sealing contact change, so as to obtain a contact state sequence and a degradation characteristic sequence containing a sealing stiffness degradation index. The application can identify the sealing degradation risk in advance by judging the health state of the degradation state sequence and matching the maintenance strategy, so that the fault prediction and health management are changed from post-analysis to pre-judgment.

Description

Technical Field

[0001] This invention relates to the field of intelligent operation and maintenance technology, and in particular to an energy-saving wind turbine fault prediction method and system based on multi-sensor fusion. Background Technology

[0002] With the widespread application of industrial energy-saving fans in chemical, metallurgical, and wastewater treatment industries, related technologies for equipment operation status monitoring and fault prediction are constantly developing. Existing technologies typically collect key parameters of the energy-saving fan's operation online by deploying pressure sensors, flow sensors, and vibration sensors, and then combine these with threshold judgment, frequency domain analysis, or machine learning-based modeling methods to assess the equipment's operating status. Simultaneously, some methods are beginning to introduce multi-source information fusion, comprehensively analyzing multiple signals such as vibration, acoustic emission, and temperature to improve the ability to identify equipment status under complex operating conditions. In terms of intelligent algorithms, deep learning methods such as recurrent neural networks and graph neural networks are gradually being applied to temporal modeling and structural relationship modeling, providing new technical pathways for the dynamic analysis of equipment operating status.

[0003] However, existing technologies still have certain limitations in predicting seal-related faults. On the one hand, traditional methods often rely on single or limited sensor signals, lacking a systematic characterization of the coupling relationship between seal contact changes, flow channel response, and thermal distribution evolution, resulting in insufficient ability to identify latent leaks and early degradation characteristics. On the other hand, existing methods are mostly based on natural operating data for modeling, lacking active excitation mechanisms, making it difficult to effectively stimulate the observable response of the sealing end face under stable operating conditions, thus introducing uncertainty into the degradation trend modeling. Summary of the Invention

[0004] In view of the aforementioned existing problems, the present invention is proposed.

[0005] Therefore, this invention provides an energy-saving wind turbine fault prediction method based on multi-sensor fusion to solve the problems of insufficient coupling of multi-source sensor information and difficulty in continuously modeling the evolution of seal degradation.

[0006] To solve the above-mentioned technical problems, the present invention provides the following technical solution:

[0007] In a first aspect, the present invention provides an energy-saving wind turbine fault prediction method based on multi-sensor fusion, which includes,

[0008] A short-term speed disturbance is applied within the stable operation segment of the energy-saving fan, and inlet pressure, outlet pressure, flow rate, motor current, motor power, bearing vibration, acoustic emission of the sealing cavity, temperature of the sealing seat, and thermal image of the shell are collected simultaneously to obtain the disturbance period synchronization time sequence segment.

[0009] The disturbance cycle synchronization time sequence segments are recombined into the periods before disturbance, during disturbance, and disturbance recovery. Combined with the continuous enhancement of acoustic emission from the sealed cavity, the slow recovery of outlet pressure, the lag in motor power decline, and the expansion of circumferential hot spots in the seal, the changes in sealing contact are determined, resulting in a contact state sequence and a degradation characteristic sequence including the sealing stiffness degradation index.

[0010] Based on the contact state sequence and the degradation feature sequence containing the sealing stiffness degradation index, the pre-disturbance adjacency relationship, the mid-disturbance adjacency relationship, the disturbance recovery adjacency relationship, and the stage state transition relationship are constructed. Combined with the disturbance period synchronization time sequence segment, physical edge weights are assigned to each adjacency relationship according to the fluid-structure interaction constraint of the sealing end face to form a physical constraint stage adaptive joint graph structure. The graph convolutional neural network is used to perform joint encoding of position association and temporal association to obtain the fused state sequence.

[0011] Bidirectional long short-term memory prediction is performed on the fusion state sequence to obtain the degenerate state sequence;

[0012] The system performs health status assessment and maintenance strategy matching on the degradation state sequence, and outputs energy-saving fan failure prediction results and seal maintenance suggestions.

[0013] As a preferred embodiment of the energy-saving wind turbine fault prediction method based on multi-sensor fusion described in this invention, wherein obtaining the disturbance period synchronization time sequence segment specifically includes:

[0014] During the continuous operation of the energy-saving fan, the continuous changes in speed, flow rate and outlet pressure are monitored, and the operating range with slow speed changes, continuous flow rate fluctuations and stable outlet pressure is selected to obtain the stable operating segment.

[0015] During the stable operation segment, a speed disturbance command is sent to the frequency converter drive to obtain the disturbance period;

[0016] Data from each sensor is recorded synchronously according to the disturbance period and aligned with a unified time stamp to obtain a disturbance period synchronization time sequence segment.

[0017] As a preferred embodiment of the energy-saving wind turbine fault prediction method based on multi-sensor fusion described in this invention, the step of recombining the disturbance period synchronization time sequence segments into pre-disturbance, during-disturbance, and disturbance recovery periods specifically includes:

[0018] Extract the rotational speed change trajectory and pressure response trajectory from the synchronous time segment of the disturbance period to determine the start time of the rotational speed disturbance, the duration of the disturbance, and the time when the rotational speed falls back to its original value, thus obtaining the pre-disturbance period, the mid-disturbance period, and the disturbance recovery period.

[0019] The data corresponding to each time period are rearranged in a unified time sequence to obtain the reconstructed time sequence segment.

[0020] As a preferred embodiment of the energy-saving wind turbine fault prediction method based on multi-sensor fusion described in this invention, the step of obtaining the contact state sequence and degradation feature sequence specifically includes:

[0021] The acoustic emission enhancement section, pressure recovery delay section, power fall-off delay section, and sealing circumferential hot spot expansion section are extracted from the reconstructed time sequence segment to obtain the stage response segment;

[0022] By correlating the stage response segments, the contact stages of the seal transition from frictional contact to separation and then to re-stabilization are divided, resulting in a contact state sequence.

[0023] The contact state sequence is mapped to the changes in flow rate and pressure difference, and a sealing stiffness degradation index is constructed based on the speed disturbance amplitude, pressure difference recovery response, flow response, power drop response, acoustic emission response, and hot spot expansion response to obtain the degradation characteristic sequence.

[0024] As a preferred embodiment of the energy-saving wind turbine fault prediction method based on multi-sensor fusion described in this invention, the step of mapping the contact state sequence to flow rate changes and pressure difference changes specifically includes:

[0025] The outlet pressure recovery amplitude and recovery duration are extracted at each stage corresponding to the contact state sequence to obtain the differential pressure maintenance characteristics.

[0026] The flow rate change and the corresponding motor power consumption are extracted at each stage of the contact state sequence to obtain the power consumption characteristics per unit flow rate.

[0027] The duration of continuous temperature rise of the sealing seat and the duration of continuous expansion of the circumferential hot spot of the sealing seat are extracted at each stage corresponding to the contact state sequence to obtain the temperature rise extension characteristics.

[0028] The degree of sustained acoustic emission enhancement in the sealed cavity is extracted to obtain acoustic emission enhancement characteristics;

[0029] The response slope between the differential pressure maintenance characteristics and the flow rate change characterizes the equivalent stiffness response term of the seal. The motor power drop hysteresis, the degree of acoustic emission enhancement, and the duration of continuous expansion of the circumferential hot spot of the seal characterize the seal contact degradation correction term. The equivalent stiffness response term of the seal and the seal contact degradation correction term are normalized and fused to obtain the seal stiffness degradation index.

[0030] By arranging the differential pressure maintenance characteristics, unit flow power consumption characteristics, temperature rise extension characteristics, acoustic emission enhancement characteristics, and sealing stiffness degradation index in a stage-wise order, a degradation characteristic sequence is obtained.

[0031] As a preferred embodiment of the energy-saving wind turbine fault prediction method based on multi-sensor fusion described in this invention, the step of using a graph convolutional neural network for joint encoding of location association and temporal association specifically includes:

[0032] Based on the three-stage contact changes corresponding to the contact state sequence, the inlet pressure, outlet pressure, flow rate, motor current, motor power, bearing vibration, acoustic emission of the sealing cavity, temperature of the sealing seat, and thermal image of the shell are mapped as graph nodes to obtain the stage node set;

[0033] Based on the sensor installation position relationship, a pre-disturbance adjacency relationship is constructed; based on the coupling relationship between the sealed end face and the flow channel response, a mid-disturbance adjacency relationship is constructed; based on the stage response continuity relationship during the speed drop process, a disturbance recovery adjacency relationship is constructed, resulting in a stage correlation diagram.

[0034] Based on the fluid-structure interaction model of the sealing end face, the physical transmission direction between outlet pressure, flow rate, acoustic emission of the sealing cavity, temperature of the sealing seat and thermal image of the shell is determined, and the initial edge weights corresponding to each physical transmission direction are determined according to the sealing gap sensitivity coefficient, pressure difference response sensitivity coefficient, flow leakage sensitivity coefficient, friction heat transfer coefficient and circumferential heat diffusion distance.

[0035] Based on the frictional contact stage, separation transition stage and restabilization stage corresponding to the contact state sequence, the initial edge weights are modulated in stages to obtain the physical edge weights before disturbance, the physical edge weights during disturbance and the physical edge weights after disturbance recovery.

[0036] Connect the corresponding nodes in the stage association diagram according to the change order of the contact state sequence to obtain the stage transition diagram;

[0037] Map the stage node data and degradation feature sequence corresponding to the stage node set in the perturbation period synchronization time segment to the stage association graph and stage transition graph to obtain the physical constraint stage adaptive joint graph structure.

[0038] By using a graph convolutional neural network to perform node propagation, stage aggregation, and sequence encoding along a physically constrained stage adaptive joint graph structure, a fused state sequence is obtained.

[0039] As a preferred embodiment of the energy-saving wind turbine fault prediction method based on multi-sensor fusion described in this invention, the step of using a graph convolutional neural network to perform node propagation, stage aggregation, and sequence encoding along a physically constrained stage adaptive joint graph structure specifically includes:

[0040] The positional association features of the stable operation phase are extracted along the adjacency relationship before the disturbance to obtain the state features before the disturbance;

[0041] The stage coupling features between the sealing end face and the flow channel are extracted along the adjacency relationship in the disturbance to obtain the state features in the disturbance;

[0042] During the node propagation process of pre-disturbance state characteristics, during-disturbance state characteristics, and during-disturbance recovery state characteristics, the intensity of node information transmission is limited by the physical edge weight of the corresponding stage, and the stage representation of the acoustic emission node of the sealing cavity, the temperature node of the sealing seat, and the thermal imaging node of the shell is corrected by the sealing stiffness degradation index.

[0043] The pre-disturbance state features, mid-disturbance state features, and post-disturbance state features are concatenated along the stage state transition relationship and fused with the corresponding degradation feature sequence to obtain the fused state sequence.

[0044] As a preferred embodiment of the energy-saving wind turbine fault prediction method based on multi-sensor fusion described in this invention, the step of performing bidirectional long short-term memory prediction on the fused state sequence specifically includes:

[0045] Extract the fusion state sequence corresponding to the current disturbance cycle, and retrieve the fusion state sequences corresponding to several historical disturbance cycles that are consecutive to the current disturbance cycle. Then, concatenate them in chronological order to obtain a continuous fusion sequence.

[0046] By using a bidirectional long short-term memory network to perform forward and backward modeling on the continuous fusion sequence, a degenerate state sequence is obtained.

[0047] As a preferred embodiment of the energy-saving wind turbine fault prediction method based on multi-sensor fusion described in this invention, the step of determining the health status and matching maintenance strategies for the degradation state sequence specifically includes:

[0048] Extract the stage change results, continuous change results, evolution direction results, and energy-saving loss changes from the degradation state sequence to obtain the discrimination results;

[0049] The growth trend of the sealing stiffness degradation index is extracted from the degradation state sequence to obtain the sealing stiffness degradation trend results;

[0050] The judgment results and sealing stiffness degradation trend results are correlated with the preset health status categories to obtain the current health status conclusion;

[0051] Based on the current health status conclusions, evolution direction results, changes in energy-saving losses, and sealing stiffness degradation trends, the failure prediction results for energy-saving fans are determined.

[0052] The current health status conclusion and the energy-saving fan failure prediction result are matched with the processing rules corresponding to the sealing assembly to obtain the maintenance strategy result;

[0053] When maintenance strategy results indicate that abnormal contact continues to prolong and energy loss continues to increase, output a shutdown inspection of the sealing end face, sealing gap and sealing seat fit condition for sealing maintenance recommendations.

[0054] If the maintenance strategy results indicate that the contact condition has recovered and the energy-saving loss still has an expanding trend, output the sealing maintenance recommendations for the next maintenance window and review the sealing surface.

[0055] If the maintenance strategy results indicate that the degradation state remains stable, output a seal maintenance recommendation to continue operation and maintain online monitoring.

[0056] Secondly, the present invention provides an energy-saving wind turbine fault prediction system based on multi-sensor fusion, comprising,

[0057] The disturbance acquisition module applies a short-term speed disturbance during the stable operation segment of the energy-saving fan and simultaneously acquires inlet pressure, outlet pressure, flow rate, motor current, motor power, bearing vibration, acoustic emission of the sealing cavity, temperature of the sealing seat, and thermal image of the shell to obtain a disturbance cycle synchronous time sequence segment.

[0058] The time-period recombination module recombines the synchronous time sequence segments of the disturbance cycle into the periods before, during, and after the disturbance. It also combines the continuous enhancement of acoustic emission from the sealed cavity, the slow recovery of the outlet pressure, the lag in the decline of motor power, and the expansion of the circumferential hot spot of the seal to determine the changes in the sealing contact, thereby obtaining the contact state sequence and the degradation characteristic sequence containing the sealing stiffness degradation index.

[0059] The graph encoding module constructs pre-disturbance adjacency relationships, mid-disturbance adjacency relationships, disturbance recovery adjacency relationships, and stage state transition relationships based on the contact state sequence and the degradation feature sequence containing the sealing stiffness degradation index. It also combines the disturbance period synchronization time segment and assigns physical edge weights to each adjacency relationship according to the fluid-structure interaction constraint of the sealing end face to form a physical constraint stage adaptive joint graph structure. The graph convolutional neural network is used to perform joint encoding of position association and temporal association to obtain the fused state sequence.

[0060] The time-series prediction module performs bidirectional long short-term memory prediction on the fused state sequence to obtain the degenerate state sequence;

[0061] The predictive management module performs health status assessment and maintenance strategy matching on the degradation state sequence, and outputs energy-saving fan failure prediction results and seal maintenance suggestions.

[0062] The beneficial effects of this invention are as follows: By applying short-term speed perturbations within stable operating segments, the sealing end face generates a repeatable response process under controlled conditions, thereby avoiding interference from natural operating condition fluctuations on data consistency and improving the comparability and reliability of data sources. By simultaneously acquiring pressure, flow rate, motor power, acoustic emission, and thermal imaging information, a multi-dimensional characterization of sealing contact, fluid transport, and thermal evolution is achieved, effectively enhancing the perception of latent leaks and minute contact changes. By dividing the perturbation cycle into a frictional contact stage, a separation transition stage, and a restabilization stage, and combining differential pressure maintenance characteristics, unit flow power consumption characteristics, temperature rise extension characteristics, and acoustic emission enhancement characteristics, a physically meaningful degradation feature sequence is constructed, expanding the degradation characterization from a single indicator to a multi-dimensional coupled indicator. A graph convolutional neural network is used to jointly encode the positional correlation between sensors and the temporal correlation between stages, enabling the fusion expression of multi-source information at the structural level. A bidirectional long short-term memory network is used to model the continuous fusion sequence of multiple perturbation cycles, allowing the degradation state to not only reflect the current state but also characterize the historical evolution path and future trends. By performing health status assessment and maintenance strategy matching on the degradation state sequence, a closed-loop processing from data acquisition to decision output is achieved, transforming fault prediction and health management from post-analysis to pre-judgment. This not only allows for the early identification of seal degradation risks but also provides targeted maintenance recommendations based on changes in energy-saving losses, thereby improving overall energy efficiency while ensuring equipment operational reliability. Attached Figure Description

[0063] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the following description of the embodiments will be briefly introduced. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0064] Figure 1 This is a flowchart of an energy-saving wind turbine fault prediction method based on multi-sensor fusion.

[0065] Figure 2 Flowchart for generating synchronous time-series segments of disturbance cycles and spatial characterization of shell thermal images.

[0066] Figure 3 A flowchart is constructed for the contact state sequence, sealing stiffness degradation index, and degradation characteristic sequence.

[0067] Figure 4 This is a flowchart of the adaptive joint graph structure construction and graph volume encoding process in the physical constraint stage.

[0068] Figure 5 Flowchart for health status assessment and maintenance decision-making. Detailed Implementation

[0069] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings.

[0070] Many specific details are set forth in the following description in order to provide a full understanding of the invention. However, the invention may also be practiced in other ways different from those described herein, and those skilled in the art can make similar extensions without departing from the spirit of the invention. Therefore, the invention is not limited to the specific embodiments disclosed below.

[0071] Secondly, the term "one embodiment" or "embodiment" as used herein refers to a specific feature, structure, or characteristic that may be included in at least one implementation of the present invention. The phrase "in one embodiment" appearing in different places in this specification does not necessarily refer to the same embodiment, nor is it a single or selective embodiment that is mutually exclusive with other embodiments.

[0072] Reference Figures 1-4 This is one embodiment of the present invention, which provides a method for predicting the faults of energy-saving wind turbines based on multi-sensor fusion, including the following steps:

[0073] S1. Apply a short-term speed disturbance during the stable operation segment of the energy-saving fan, and simultaneously collect inlet pressure, outlet pressure, flow rate, motor current, motor power, bearing vibration, acoustic emission of the sealing cavity, temperature of the sealing seat, and thermal image of the shell to obtain the disturbance period synchronization time segment.

[0074] S1.1. During the continuous operation of the energy-saving fan, the real-time operating flow of speed, flow rate and outlet pressure is continuously read, and the speed, flow rate and outlet pressure are continuously monitored in chronological order; when the speed change curve remains flat, the flow rate change curve does not change abruptly and the outlet pressure change curve is within a stable fluctuation range, the corresponding operating interval is extracted from the real-time operating flow as a stable operating segment.

[0075] The stable fluctuation range refers to the interval within a statistical time window where the outlet pressure fluctuates slightly randomly around its operating average without showing a trend deviation. This stable fluctuation range is obtained through statistical analysis of historical stable operating data. Specifically, it involves selecting outlet pressure data from energy-saving fans under long-term normal operation without regulatory intervention, calculating the pressure average and fluctuation amplitude within each window using a time window sliding method, and characterizing the fluctuation intensity using the pressure standard deviation. When the deviation of the outlet pressure from the corresponding average within the current time window does not exceed three times the standard deviation under historical stable operating conditions, and no monotonically increasing or decreasing trend appears within multiple consecutive time windows, the corresponding interval is determined to be the stable fluctuation range. Pressure signals under normal operating conditions typically follow the random fluctuation patterns observed in engineering measurements. A range of three times the standard deviation can cover the vast majority of normal fluctuation intervals, while avoiding misjudging abnormal fluctuations caused by valve position adjustments, sudden load changes, or pipeline disturbances as stable states. This ensures that the captured stable operating segments only contain sealing response information under a single operating background, improving the consistency and reliability of sealing status identification in subsequent fault prediction and health management.

[0076] Under the conditions that the speed change curve remains flat, the flow rate change curve has no sudden changes, and the outlet pressure change curve is within the above-mentioned stable fluctuation range, the corresponding operating interval is extracted from the real-time operating flow as a stable operating segment. The stable operating segment corresponds to the energy-saving fan being in normal transmission state without sudden changes in operating conditions, start-stop switching, or large load adjustments. The purpose of selecting the stable operating segment is to avoid the mixed influence of upstream pipeline fluctuations, downstream valve position changes, and human operation interference on the sealing status identification, thereby ensuring that fault prediction and health management face the sealing response information under the same operating background.

[0077] S1.2. During the stable operation segment, a short-term speed disturbance command is sent to the frequency converter drive. The short-term speed disturbance command corresponds to a single speed change process with a fixed amplitude and a fixed duration. The short-term speed disturbance command causes the energy-saving fan to go through three consecutive stages in sequence: disturbance start, disturbance process, and disturbance recovery.

[0078] In this embodiment, the short-term speed disturbance adopts a small speed change mode that does not deviate from the normal operating range. The duration of the short-term speed disturbance is limited to the point that a complete pressure response, power response and sealing thermal response can be formed. After the short-term speed disturbance ends, the speed returns to the original speed corresponding to the stable operating segment, thereby forming a complete disturbance cycle.

[0079] The short-term speed disturbance adopts a single short-cycle method, which aims to make the sealing end face form an observable response under short-term operating conditions, while avoiding long-term operating condition deviation from causing a continuous impact on the normal operation and energy-saving status of the energy-saving fan.

[0080] S1.3. After the disturbance period is formed, the inlet pressure, outlet pressure, flow rate, motor current, motor power, bearing vibration, acoustic emission from the sealing cavity, sealing seat temperature, and shell thermal image are recorded synchronously for the continuous time periods before the disturbance begins, during the disturbance, and after the disturbance recovers. Inlet pressure, outlet pressure, and flow rate are used to reflect changes in the energy-saving fan's delivery status; motor current and motor power are used to reflect changes in the drive load under short-term speed disturbances; bearing vibration is used to reflect the accompanying response of rotating components during the disturbance period; acoustic emission from the sealing cavity is used to reflect the high-frequency friction and leakage scouring characteristics during the sealing contact change process; sealing seat temperature is used to reflect the thermal state evolution of the sealing area; and shell thermal image is used to reflect changes in the circumferential thermal distribution of the seal. The synchronous acquisition of inlet pressure, outlet pressure, flow rate, motor current, motor power, bearing vibration, acoustic emission from the sealing cavity, sealing seat temperature, and shell thermal image within the same disturbance period aims to ensure comparability of pressure response, power response, acoustic emission response, and thermal response under a unified time background, providing a consistent data source for subsequent quantitative processing of latent leakage in the seal.

[0081] S1.4. Add a unified time mark to the inlet pressure, outlet pressure, flow rate, motor current, motor power, bearing vibration, acoustic emission of the sealing cavity, temperature of the sealing seat, and thermal image of the shell, and align the time marks according to the unified time mark.

[0082] When there are differences in the sampling frequency of inlet pressure, outlet pressure, flow rate, motor current, motor power, bearing vibration, acoustic emission of the sealing cavity, and sealing seat temperature, resampling and interpolation alignment are performed with a unified time axis to ensure that there is a unique set of inlet pressure, outlet pressure, flow rate, motor current, motor power, bearing vibration, acoustic emission of the sealing cavity, and sealing seat temperature at the same time.

[0083] Since the thermal images of the shell are two-dimensional spatial image data, the original two-dimensional thermal image matrix is ​​not directly used as one-dimensional time-series data for unified time-stamp alignment. Instead, spatial feature processing is performed on each frame of the shell thermal image.

[0084] Specifically, the region of interest in the sealing area is determined based on the installation position of the sealing seat in the thermal image of the housing. The region of interest in the sealing area is then expanded into a polar coordinate thermal image region along the sealing center and the sealing circumferential direction. Using the stable thermal image before the disturbance as the temperature reference, temperature difference is performed on the thermal image frames during the disturbance and the disturbance recovery stage to obtain the circumferential temperature rise distribution map of the seal.

[0085] In the circumferential temperature rise distribution map of the seal, hot spot areas are extracted based on the temperature rise threshold, which is determined by the average temperature and temperature standard deviation of the stable thermal image area before disturbance; when the temperature rise value of a certain pixel exceeds the temperature rise threshold, the pixel is identified as a hot spot pixel.

[0086] Connectivity analysis is performed on the hot spot pixels to extract the hot spot area, the highest temperature of the hot spot, the average temperature of the hot spot, the circumferential span of the hot spot, the radial expansion distance of the hot spot, the centroid angle of the hot spot, and the duration of the hot spot, so as to obtain the thermal image spatial feature vector corresponding to each frame of shell thermal image.

[0087] The thermal image spatial feature vectors corresponding to each frame of the thermal image of the shell are arranged in chronological order to form a thermal image spatial feature sequence of the shell. When the frame rate of the thermal image of the shell is lower than the sampling frequency of other one-dimensional sensors, the thermal image spatial feature sequence of the shell is aligned to a unified time axis by timestamp nearest neighbor mapping or linear interpolation, so that each time step corresponds to a set of thermal image spatial feature vectors of the shell.

[0088] S1.5. After the unified time identification is completed, the inlet pressure, outlet pressure, flow rate, motor current, motor power, bearing vibration, acoustic emission of the sealing cavity, temperature of the sealing seat, and spatial characteristic sequence of the thermal image of the shell are extracted into data segments of the same length according to the start and end time of the disturbance cycle, and then arranged in a unified time order to form a strictly aligned disturbance cycle synchronous time sequence segment within the same disturbance cycle.

[0089] The disturbance cycle synchronization time segment fully preserves the pressure changes, flow rate changes, power changes, vibration changes, acoustic emission changes, temperature changes, and thermal image spatial characteristic changes caused by short-term speed disturbances. The disturbance cycle synchronization time segment can characterize both the instantaneous operating conditions of the energy-saving fan in the stable operation segment and the contact change process of the sealing end face under short-term operating condition changes.

[0090] S2. The disturbance cycle synchronization time sequence segments are recombined into the periods before disturbance, during disturbance, and during disturbance recovery. Combined with the continuous enhancement of acoustic emission from the sealed cavity, the slow recovery of outlet pressure, the lag in motor power decline, and the expansion of circumferential hot spots in the seal, the changes in sealing contact are determined, resulting in a contact state sequence and a degradation characteristic sequence containing the sealing stiffness degradation index.

[0091] S2.1. Read the synchronous timing segment of the disturbance cycle, locate the starting point and the end point of recovery corresponding to the short-term speed disturbance in the speed change trajectory, and locate the moment when the pressure response begins to change and the moment when the pressure recovers and tends to stabilize in the outlet pressure change trajectory; when the speed is at the original level before the disturbance and the outlet pressure remains at the level corresponding to the stable operating segment, determine the period before the disturbance; when the speed enters the change range corresponding to the short-term speed disturbance, determine the period during the disturbance; when the speed recovers to the level corresponding to the stable operating segment and the outlet pressure enters the recovery process, determine the disturbance recovery period.

[0092] The inlet pressure, outlet pressure, flow rate, motor current, motor power, bearing vibration, acoustic emission of the sealing cavity, temperature of the sealing seat, and thermal image spatial feature sequences of the pre-disturbance period, the mid-disturbance period, and the disturbance recovery period are rearranged in a unified time sequence while maintaining a unified time identifier. After connecting the pre-disturbance period, the mid-disturbance period, and the disturbance recovery period end to end, a recombined time sequence segment is formed that simultaneously retains the stage boundaries and the original time sequence relationships.

[0093] S2.2. Extract the continuous enhancement section of acoustic emission from the sealed cavity, the delay section of outlet pressure recovery, the delay section of motor power drop-off, and the circumferential hot spot expansion section of the seal from the reconstructed time sequence segment.

[0094] The continuous enhancement section of acoustic emission from the sealed cavity corresponds to the interval in which the acoustic emission from the sealed cavity continuously increases over time; the delayed recovery section of the outlet pressure corresponds to the continuous interval in which the outlet pressure has not yet recovered to the level corresponding to the stable operating segment during the disturbance recovery period; the delayed decline section of motor power corresponds to the continuous interval in which the speed has entered the recovery process but the motor power still remains at a high level.

[0095] The circumferential hot spot extension section of the seal is determined by the spatial feature sequence of the shell thermal image, rather than directly by the original two-dimensional thermal image frame. Specifically, the circumferential hot spot extension section of the seal is jointly determined by the hot spot area, the highest hot spot temperature, the average hot spot temperature, the circumferential span of the hot spot, the radial extension distance of the hot spot, the centroid angle of the hot spot, and the duration of the hot spot in the spatial feature sequence of the shell thermal image.

[0096] When the area of ​​the hot spot increases within a continuous time step, or the circumferential span of the hot spot increases within a continuous time step, or the radial expansion distance of the hot spot continues to increase, it is determined that the circumferential hot spot of the seal is in an expanding state; when the above-mentioned expanding state continues to exist, the corresponding time interval is determined as the expanding section of the circumferential hot spot of the seal.

[0097] When the highest and average temperatures of the hot spot increase, but the area, circumferential span, and radial expansion distance of the hot spot do not continue to increase, it is only considered as a localized increase in temperature, and not as a separate section of the sealed circumferential hot spot expansion.

[0098] The continuous enhancement section of acoustic emission from the sealed cavity, the delay section of outlet pressure recovery, the delay section of motor power drop-off, and the circumferential hot spot expansion section of the sealed cavity determined by the spatial feature sequence of the shell thermal image are mapped onto a unified time axis to form a stage response segment.

[0099] S2.3. Determine the sealing contact changes based on the sequential and overlapping relationships of the stage response segments on a unified time axis; when the continuous enhancement segment of acoustic emission from the sealing cavity and the delayed segment of motor power decline appear synchronously, and the delayed segment of outlet pressure recovery has been formed and the segment of circumferential hot spot expansion of the seal begins to appear, the seal is determined to be in the friction contact stage; when the continuous enhancement segment of acoustic emission from the sealing cavity begins to weaken, the outlet pressure enters the recovery process, the motor power declines from a high level to the level corresponding to the stable operation segment, and the circumferential hot spot of the seal continues to expand but the expansion range tends to slow down, the seal is determined to be in the separation transition stage; when the outlet pressure recovers to the level corresponding to the stable operation segment, the motor power decline is completed, the acoustic emission from the sealing cavity remains stable, and the circumferential hot spot of the seal no longer continues to expand, the seal is determined to be in the restabilization stage; the friction contact stage, separation transition stage, and restabilization stage are arranged in chronological order to form a contact state sequence.

[0100] S2.4. Correspond the contact state sequence with the changes in flow rate and differential pressure, and extract the degradation feature sequence by combining the changes in sealing seat temperature, changes in the circumferential hot spot of the seal, and changes in acoustic emission of the sealing cavity; statistically analyze the outlet pressure recovery amplitude and recovery time in the friction contact stage, separation transition stage, and re-stabilization stage to form the differential pressure maintenance feature; extract the flow rate change and corresponding motor power consumption in the friction contact stage, separation transition stage, and re-stabilization stage to form the power consumption per unit flow rate; extract the duration of continuous temperature rise of the sealing seat and the duration of continuous expansion of the circumferential hot spot of the seal in the friction contact stage, separation transition stage, and re-stabilization stage to form the temperature rise extension feature; extract the degree of continuous enhancement of acoustic emission of the sealing cavity in the friction contact stage, separation transition stage, and re-stabilization stage to form the acoustic emission enhancement feature.

[0101] After the differential pressure maintenance characteristics, unit flow power consumption characteristics, temperature rise extension characteristics, and acoustic emission enhancement characteristics are formed, a sealing stiffness degradation index is constructed based on the speed disturbance amplitude, differential pressure recovery response, flow response, power fall-off response, acoustic emission response, and hot spot expansion response. The sealing stiffness degradation index is used to characterize the degree of decrease in the equivalent support capacity of the sealing end face to short-term speed disturbance.

[0102] Specifically, in each stage corresponding to the contact state sequence, the response slope between the outlet pressure recovery amplitude and the flow rate change is used as the sealing equivalent stiffness response term. The sealing equivalent stiffness response term reflects the ability of the sealing area to maintain pressure difference under unit flow rate disturbance. When the outlet pressure recovery amplitude decreases, the outlet pressure recovery time increases, and the flow rate change under the same disturbance increases, the sealing equivalent stiffness response term decreases, indicating that the sealing gap constraint ability decreases.

[0103] After the equivalent stiffness response term of the seal is formed, the degree of motor power fallback hysteresis, the degree of continuous enhancement of acoustic emission of the sealing cavity, and the duration of continuous expansion of the circumferential hot spot of the seal are used as the seal contact degradation correction terms. Among them, the degree of motor power fallback hysteresis characterizes the continuity of frictional load on the sealing end face, the degree of continuous enhancement of acoustic emission of the sealing cavity characterizes the enhancement of micro-friction or leakage scouring on the sealing end face, and the duration of continuous expansion of the circumferential hot spot of the seal characterizes the continuity of frictional heat generation and circumferential heat transfer.

[0104] The sealing equivalent stiffness response term and the sealing contact degradation correction term are normalized and then fused according to the time sequence of the three stages before disturbance, during disturbance, and after disturbance recovery to obtain the sealing stiffness degradation index. When the sealing equivalent stiffness response term continuously decreases and the sealing contact degradation correction term continuously increases, the sealing stiffness degradation index increases, indicating that the sealing end face support constraint capacity decreases, contact abnormalities increase, and leakage risk increases.

[0105] The differential pressure maintenance characteristics, unit flow power consumption characteristics, temperature rise extension characteristics, acoustic emission enhancement characteristics, and sealing stiffness degradation index are arranged in the order of the stages corresponding to the contact state sequence to form a degradation characteristic sequence.

[0106] S3. Based on the contact state sequence and the degradation feature sequence containing the sealing stiffness degradation index, construct the pre-disturbance adjacency relationship, the adjacency relationship during the disturbance, the adjacency relationship after the disturbance recovery, and the stage state transition relationship. Combined with the synchronous time sequence segment of the disturbance period, assign physical edge weights to each adjacency relationship according to the fluid-structure interaction constraint of the sealing end face to form a physical constraint stage adaptive joint graph structure. Use graph convolutional neural network to perform joint encoding of position association and temporal association to obtain the fused state sequence.

[0107] In this embodiment, the adjacency relationships before disturbance, during disturbance, and after disturbance recovery are not directly determined by the correlation of node data. Instead, the physical propagation direction and initial edge weights are determined under the constraints of the fluid-structure interaction model of the sealed end face. Then, stage modulation is performed according to the contact state sequence to form a physical constraint stage adaptive joint graph structure. This physical constraint stage adaptive joint graph structure serves as the input of the graph convolutional neural network, making the node propagation direction and propagation intensity of the graph convolutional neural network correspond to the sealing end face pressure difference maintenance, leakage flow change, frictional heat generation, and circumferential heat diffusion mechanism.

[0108] S3.1. The inlet pressure, outlet pressure, flow rate, motor current, motor power, bearing vibration, acoustic emission of the sealing cavity, temperature of the sealing seat, and thermal image of the shell are classified into stages according to the friction contact stage, separation transition stage, and restabilization stage.

[0109] Among them, the inlet pressure, outlet pressure, flow rate, motor current, motor power, bearing vibration, acoustic emission from the sealing cavity, and sealing seat temperature are all one-dimensional time-series data. The sampling value segments of the corresponding stages are directly extracted according to a unified time axis to obtain the stage time-series characteristics of each one-dimensional sensor.

[0110] The thermal image of the casing is a two-dimensional spatial image data. Before mapping it into graph nodes, each frame of the casing thermal image is first subjected to spatial feature processing. Specifically, the region of interest in the sealing area is determined according to the installation position of the sealing seat in the casing thermal image. The region of interest in the sealing area is unfolded into a polar coordinate thermal image region along the sealing center and the sealing circumferential direction. Using the stable thermal image before the disturbance as the temperature reference, temperature difference is performed on the thermal image frames during the disturbance and the disturbance recovery stage to obtain the circumferential temperature rise distribution map of the seal.

[0111] In the circumferential temperature rise distribution map of the seal, hot spot regions are extracted based on a temperature rise threshold, which is determined by the average temperature and temperature standard deviation of the stable thermal image region before disturbance. When the temperature rise value of a certain pixel exceeds the temperature rise threshold, that pixel is identified as a hot spot pixel. Connectivity analysis is performed on the hot spot pixels to extract the hot spot area, maximum hot spot temperature, average hot spot temperature, circumferential span of the hot spot, radial expansion distance of the hot spot, centroid angle of the hot spot, and duration of the hot spot, thus obtaining the thermal image spatial feature vector corresponding to each frame of the shell thermal image.

[0112] The thermal image spatial feature vectors corresponding to each frame of the shell thermal image are arranged according to a unified time axis to form a shell thermal image spatial feature sequence. When the shell thermal image frame rate is lower than the sampling frequency of other one-dimensional sensors, the shell thermal image spatial feature sequence is aligned to a unified time axis by timestamp nearest neighbor mapping or linear interpolation, so that each time step corresponds to a set of shell thermal image spatial feature vectors.

[0113] Nine graphical nodes were determined for each of the following stages: in the friction contact stage, inlet pressure, outlet pressure, flow rate, motor current, motor power, bearing vibration, acoustic emission from the sealing cavity, temperature of the sealing seat, and thermal image spatial characteristic sequence of the shell. Nine graphical nodes were also determined for each of the following stages: inlet pressure, outlet pressure, flow rate, motor current, motor power, bearing vibration, acoustic emission from the sealing cavity, temperature of the sealing seat, and thermal image spatial characteristic sequence of the shell. Nine graphical nodes were also determined for each of the following stages: inlet pressure, outlet pressure, flow rate, motor current, motor power, bearing vibration, acoustic emission from the sealing cavity, temperature of the sealing seat, and thermal image spatial characteristic sequence of the shell. A total of 27 graphical nodes were obtained for the three stages, which together constitute the stage node set.

[0114] The node data corresponding to each graph node in the stage node set is given by the sampled value segment of the corresponding stage in the disturbance period synchronization time sequence segment. The node data corresponding to the shell thermal image node is given by the shell thermal image spatial feature sequence of the corresponding stage, rather than directly given by the original two-dimensional thermal image matrix. The stage features corresponding to each graph node in the stage node set are given by the pressure difference maintenance feature, unit flow power consumption feature, temperature rise extension feature, acoustic emission enhancement feature, and sealing stiffness degradation index.

[0115] S3.2. After the stage node set is determined, the pre-disturbance adjacency relationship is constructed based on the sensor installation position relationship; the inlet pressure, outlet pressure and flow rate are connected along the energy-saving fan flow channel direction, the motor current and motor power are connected along the drive end, and the bearing vibration, sealing cavity acoustic emission, sealing seat temperature and shell thermal image are connected along the sealing end. The pre-disturbance adjacency relationship reflects the actual installation position association during the stable operation stage.

[0116] After the adjacency relationship before the disturbance is determined, the adjacency relationship during the disturbance is constructed based on the fluid-structure interaction model of the sealing end face. The fluid-structure interaction model of the sealing end face is used to describe the physical transfer relationship between pressure difference, sealing gap, leakage flow, frictional contact, acoustic emission response, frictional heat generation and circumferential heat diffusion under short-term speed disturbance.

[0117] In the fluid-structure interaction model of the sealing end face, the edge between the outlet pressure node and the flow node represents the influence of pressure difference change on leakage flow and delivery flow; the edge between the flow node and the motor power node represents the influence of flow disturbance on the driving load; the edge between the sealing cavity acoustic emission node and the sealing seat temperature node represents the influence of frictional contact or leakage scouring on local heat accumulation; the edge between the sealing seat temperature node and the shell thermal imaging node represents the influence of heat diffusion from the sealing seat to the shell circumferentially; and the outlet pressure node, flow node, and sealing cavity acoustic emission node together point to the stage characteristic representation corresponding to the sealing stiffness degradation index.

[0118] After determining the physical transmission direction between outlet pressure, flow rate, motor power, acoustic emission of the sealing cavity, temperature of the sealing seat, and thermal image of the shell based on the fluid-structure interaction model of the sealing end face, the initial edge weights corresponding to each physical transmission direction are determined according to the sealing gap sensitivity coefficient, pressure difference response sensitivity coefficient, flow leakage sensitivity coefficient, frictional heat transfer coefficient, and circumferential heat diffusion distance.

[0119] Among them, the sealing gap sensitivity coefficient is used to characterize the impact of sealing gap changes on flow leakage and differential pressure maintenance capability; the differential pressure response sensitivity coefficient is used to characterize the responsiveness of outlet pressure recovery amplitude and recovery time to changes in sealing state; the flow leakage sensitivity coefficient is used to characterize the sensitivity of flow change to seal degradation under the same speed disturbance; the frictional heat transfer coefficient is used to characterize the contribution of enhanced acoustic emission from the sealing cavity and motor power drop hysteresis to the temperature rise of the sealing seat; and the circumferential heat diffusion distance is used to characterize the spatial influence range of hot spots expanding along the circumferential direction of the seal in the thermal image of the shell.

[0120] In this embodiment, the initial edge weights are obtained by normalizing the sensitivity coefficients of the corresponding physical transmission directions. When the sensitivity coefficient of a certain physical transmission direction is large, the initial edge weights in that physical transmission direction are large, and the propagation intensity of node information in the graph convolutional neural network in that direction increases accordingly. When a certain physical transmission direction has no direct correspondence with the fluid-structure interaction mechanism of the sealing end face, no physical edge is established in that direction or a low edge weight is assigned, thereby suppressing the propagation of data correlation without physical meaning.

[0121] After the initial edge weights are formed, stage modulation is performed according to the friction contact stage, separation transition stage, and restabilization stage corresponding to the contact state sequence. In the friction contact stage, the edge weights between the acoustic emission node of the sealing cavity, the motor power node, the sealing seat temperature node, and the shell thermal imaging node are increased to highlight the transfer of frictional load and frictional heat generation. In the separation transition stage, the edge weights between the outlet pressure node, the flow node, and the motor power node are increased to highlight the pressure difference recovery and flow leakage changes. In the restabilization stage, the recovery edge weights between the outlet pressure node, the flow node, and the shell thermal imaging node are increased to highlight the continuity between pressure recovery, flow recovery, and thermal imaging recovery.

[0122] After the adjacency relationships during the disturbance are determined, the disturbance recovery adjacency relationships are constructed based on the stage response continuity during the speed reduction process. The graph nodes corresponding to speed recovery, outlet pressure recovery, motor power reduction, sealing seat temperature reduction, and shell thermal image stabilization are connected in the order of recovery. The disturbance recovery adjacency relationships reflect the continuity response during the disturbance recovery stage. The edge weights in the disturbance recovery adjacency relationships are modulated by the pressure difference response sensitivity coefficient, power reduction lag, temperature rise reduction time, and circumferential thermal diffusion distance during the disturbance recovery stage, ensuring that the node propagation intensity during the recovery stage corresponds to the sealing state recovery process.

[0123] S3.3. According to the order of the friction contact stage, separation transition stage and re-stabilization stage corresponding to the contact state sequence, the corresponding graph nodes of the same sensor in adjacent stages are connected in sequence to obtain the stage state transition relationship; the graph nodes corresponding to the inlet pressure, outlet pressure, flow rate, motor current, motor power, bearing vibration, acoustic emission of the sealing cavity, temperature of the sealing seat and thermal image of the shell are connected in sequence along the friction contact stage, separation transition stage and re-stabilization stage respectively, and the stage state transition relationship completely preserves the stage switching order in the contact state sequence.

[0124] The node data in the stage node set and the stage features in the degradation feature sequence are jointly mapped to the adjacency relationship before disturbance, the adjacency relationship during disturbance, the adjacency relationship after disturbance recovery, and the stage state transition relationship to obtain the physical constraint stage adaptive joint graph structure. Each graph node in the physical constraint stage adaptive joint graph structure contains the sample value fragment of the corresponding stage, as well as the pressure difference maintenance feature, unit flow power consumption feature, temperature rise extension feature, acoustic emission enhancement feature, and sealing stiffness degradation index of the corresponding stage. Each graph edge contains the sensor position relationship, the stage state transition relationship, and the physical edge weight determined by the fluid-structure interaction model of the sealing end face.

[0125] S3.4. A graph convolutional neural network is used to perform node propagation, stage aggregation, and sequence encoding along a physically constrained stage adaptive joint graph structure. Node propagation in the graph convolutional neural network is obtained using a neighborhood aggregation method with physical edge weights, expressed as:

[0126] ;

[0127] in, Indicates the first The node representation of a graph node after one physical constraint neighborhood aggregation. Indicates the first The original node representation of a graph node. Indicates the relationship with the first Each graph node is a set of neighboring graph nodes that are directly connected to it. Representing neighboring nodes The original node representation, Indicates the first Each stage of nodes To the node Physical edge weights This represents a nonlinear mapping function.

[0128] The node propagation constrains the pressure difference transmission, leakage flow transmission, frictional heat generation transmission, and circumferential heat diffusion transmission in the fluid-structure interaction model of the sealed end face to the information propagation process of the graph convolutional neural network through physical edge weights. This makes the information fusion between different sensing nodes no longer rely solely on statistical correlation, but is limited by the sealing mechanism.

[0129] After node propagation is completed, positional correlation features of the stable operation stage are extracted along the pre-disturbance adjacency relationship to obtain the pre-disturbance state features; stage coupling features between the sealing end face and the flow channel are extracted along the mid-disturbance adjacency relationship to obtain the mid-disturbance state features; stage correlation features between pressure recovery, power drop, and thermal image stabilization are extracted along the disturbance recovery adjacency relationship to obtain the disturbance recovery state features; during the node propagation process of the pre-disturbance state features, mid-disturbance state features, and disturbance recovery state features, the node information transmission intensity is limited by the physical edge weight of the corresponding stage, and the stage representation of the sealing cavity acoustic emission node, sealing seat temperature node, and shell thermal image node is corrected by the sealing stiffness degradation index, so that the sealing contact anomaly, frictional heat generation, and hot spot expansion have continuous and interpretable physical meaning in the fused state sequence.

[0130] The pre-disturbance state characteristics, during-disturbance state characteristics, and disturbance recovery state characteristics are sequentially concatenated along the stage state transition relationship. These are then fused with the differential pressure maintenance characteristics, unit flow power consumption characteristics, temperature rise extension characteristics, acoustic emission enhancement characteristics, and sealing stiffness degradation index from the degradation characteristic sequence to obtain a fused state sequence. This fused state sequence fully expresses the sealing contact duration, the degree of insufficient differential pressure recovery, the energy consumption shift trend, the thermal anomaly propagation direction, and the sealing stiffness degradation trend.

[0131] S3.5. The graph convolutional neural network is trained using historical perturbation cycle synchronization time-series segments, historical contact state sequences, and historical degradation feature sequences. The historical perturbation cycle synchronization time-series segments are obtained according to S1.1 to S1.5, and the historical contact state sequences and historical degradation feature sequences are obtained according to S2.1 to S2.3. The friction contact phase, separation transition phase, and restabilization phase in the historical contact state sequences are used as supervision labels. The graph convolutional neural network uses two graph convolutional layers and one sequence output layer. The first graph convolutional layer has an output dimension of 64, and the second graph convolutional layer has an output dimension of 32. The training batch consists of 32 perturbation cycle samples, with 100 training epochs and a learning rate of 0.001.

[0132] During training, the historical disturbance cycle synchronization time sequence, historical contact state sequence, historical degradation feature sequence, and historical sealing stiffness degradation index are first transformed into a historical physical constraint stage adaptive joint graph structure according to S3.1 to S3.3. Then, node propagation and stage aggregation with physical edge weights are performed according to S3.4. Subsequently, the cross-entropy loss is calculated using the supervision labels corresponding to the historical contact state sequence, and the prediction deviation of the sealing stiffness degradation index is used as an auxiliary constraint term to participate in the parameter update.

[0133] The auxiliary constraint term is used to limit the fused state sequence output by the graph convolutional neural network from deviating from the physical degradation direction of the sealing end face. When the model output is inconsistent with the sealing stiffness degradation trend characterized by decreased pressure recovery capability, increased power consumption per unit flow, enhanced acoustic emission, and hot spot expansion, the auxiliary constraint term is increased, thereby prompting the graph convolutional neural network to learn the node propagation relationship consistent with the sealing mechanism.

[0134] Training stops when the cross-entropy loss of the validation samples does not decrease for 10 consecutive rounds. The trained graph convolutional neural network is used for node propagation, stage aggregation, and sequence encoding of the adaptive joint graph structure for the current physical constraint stage, ultimately obtaining the fused state sequence.

[0135] S4. Perform bidirectional long short-term memory prediction on the fusion state sequence to obtain the degenerate state sequence.

[0136] S4.1. Number and label the fusion state sequence according to the time order of the perturbation period, and determine the fusion state sequence corresponding to the current perturbation period as the nth. The sequence is extracted simultaneously with the sequence number. The sequence of fused states corresponding to the first few consecutive perturbation cycles is arranged in chronological order as follows: To the The fusion state sequence; the ... To the The fusion state sequence is continuous from beginning to end and has consistent intervals on the time axis, which satisfies the input requirement of bidirectional long short-term memory network for continuous time series.

[0137] In this embodiment, the number of consecutive disturbance periods is taken as The length of the fusion state sequence corresponding to each perturbation cycle remains consistent to ensure time step alignment.

[0138] S4.2. The first The fusion state sequence is spliced ​​together in chronological order to form a continuous fusion sequence. Each time step in the continuous fusion sequence contains the node encoding result of the corresponding time in the fusion state sequence. The continuous fusion sequence completely preserves the continuous evolution information of the sealing contact duration, the degree of insufficient pressure difference recovery, the energy consumption offset trend, the thermal anomaly expansion direction, and the change in the sealing stiffness degradation index in multiple disturbance cycles.

[0139] The continuous fusion sequence is input into a bidirectional long short-term memory network for forward and backward modeling. The bidirectional long short-term memory network consists of forward long short-term memory units and backward long short-term memory units. The forward long short-term memory units process the continuous fusion sequence in forward chronological order, while the backward long short-term memory units process the continuous fusion sequence in reverse chronological order. The long short-term memory network structure is derived from the gated recurrent unit extension structure in recurrent neural networks. Its core computation is based on the state control using input gates, forget gates, and output gates.

[0140] S4.3. By calculating the state update process of the Long Short-Term Memory (LSTM) unit, the hidden state at the current moment is obtained, expressed as:

[0141] ;

[0142] in, Indicates the first The hidden states at each time step are used to describe the comprehensive output of the time series features at the current time. This indicates the activation result of the output gate. Indicates cell state, This represents the hyperbolic tangent function, used for nonlinear mapping of cell states.

[0143] The update of cell state is determined by both input and historical information, expressed as:

[0144] ;

[0145] in, This represents the cell state at the previous time step, used to ensure the stability and continuity of time series information during propagation. This indicates the activation result of the forget gate, used to control the degree to which historical information is retained. This indicates the activation result of the input gate, used to control the degree to which the current input information is introduced. This represents the candidate state at the current time step.

[0146] In this embodiment, the change in the sealing stiffness degradation index is used as an explicit physical degradation input in the continuous fusion sequence. It enters the input gate and forget gate together with the node encoding result output by the graph convolutional neural network, so that the bidirectional long short-term memory network can learn the data-driven state evolution path and the physical evolution direction of sealing stiffness degradation at the same time during the time modeling process.

[0147] S4.4. The hidden state sequence output by the forward long short-term memory unit and the hidden state sequence output by the backward long short-term memory unit are concatenated at the same time step to obtain the joint output sequence of the bidirectional long short-term memory network. The joint output sequence contains both historical evolution information and future trend information, so that the sealing degradation change in the fused state sequence is bidirectionally constrained in the time dimension, thereby improving the stability of fault prediction and health management.

[0148] The joint output sequence of the bidirectional long short-term memory network is mapped in the time dimension, and the hidden state corresponding to each time step in the joint output sequence is mapped to the corresponding state value in the degradation state sequence. Each state value in the degradation state sequence is used to characterize the sealing contact change trend, pressure difference maintenance capability change trend, unit flow power consumption change trend, temperature rise expansion trend, and sealing stiffness degradation index change trend at the corresponding time step. After the degradation state sequence is continuously arranged on the time axis, the complete degradation state sequence is obtained.

[0149] The bidirectional long short-term memory network is trained using historical fusion state sequences and corresponding historical degradation state sequences. The historical fusion state sequences are obtained by splicing together the fusion state sequences corresponding to multiple historical disturbance cycles in chronological order. The historical degradation state sequences are generated by corresponding changes in sealing contact, pressure difference, energy consumption, temperature, and sealing stiffness degradation index recorded during actual operation.

[0150] During training, the historical fused state sequence is used as input, and the historical degraded state sequence is used as supervision labels. The mean squared error loss function is used to measure the difference between the prediction result and the supervision label, and the expression is:

[0151] ;

[0152] in, This represents the loss value, used to measure the deviation between the predicted result and the actual result. Indicates the total number of time steps. Indicates the first The actual degradation state value at each time step. Indicates the first The predicted degradation state value at each time step.

[0153] When the sealing stiffness degradation index in the historical degradation state sequence continues to increase, but the degradation state sequence output by the bidirectional long short-term memory network does not reflect the corresponding increasing trend, the mean square error loss increases, causing the network parameters to be updated in a direction that can express the continuous degradation of sealing stiffness.

[0154] The training batch consists of 32 consecutive fused sequences, with 100 training rounds and a learning rate of 0.001. Training is stopped when the loss value corresponding to the validation dataset no longer decreases in 10 consecutive training rounds. The bidirectional long short-term memory network after training is used for forward and backward modeling of the current consecutive fused sequences, and finally outputs a degenerate state sequence.

[0155] S5. Perform health status judgment and maintenance strategy matching on the degradation state sequence, and output the energy-saving fan failure prediction results and seal maintenance suggestions.

[0156] S5.1. Extract the stage change results, continuous change results, evolution direction results, and energy-saving loss changes from the degradation state sequence. The stage change results consist of the switching order and duration of the friction contact stage, separation transition stage, and re-stabilization stage in adjacent time steps. The continuous change results consist of the time intervals of continuous increase or continuous decrease in the degradation state sequence. The evolution direction results consist of the overall upward, overall downward, or stable trend of the degradation state sequence in adjacent disturbance cycles. The energy-saving loss changes consist of the synchronous changes of pressure difference maintenance characteristics and unit flow power consumption characteristics. Extract the growth trend of the sealing stiffness degradation index from the degradation state sequence to obtain the sealing stiffness degradation trend results. The stage change results, continuous change results, evolution direction results, energy-saving loss changes, and sealing stiffness degradation trend results together constitute the discrimination results used for fault prediction and health management.

[0157] The judgment results are mapped to health status categories, which are divided into healthy, slightly degraded, moderately degraded, severely degraded, and failure risk. When the contact state sequence corresponding to the stage change results is still mainly in the re-stabilization stage, and the energy-saving loss changes remain stable, and the sealing stiffness degradation index remains low or has no continuous upward trend, the current health status conclusion is determined to be healthy. When the stage change results show that the duration of the friction contact stage and the separation transition stage increases, and the energy-saving loss changes show a slow increase, and the sealing stiffness degradation index shows a slow upward trend, the current health status conclusion is determined to be slightly degraded. When the stage change results show an increase in the proportion of the separation transition stage, a continuous increase in energy-saving losses, and a continuous increase in the sealing stiffness degradation index during continuous disturbance cycles, the current health status is determined to be moderately degraded. When the stage change results show a continuous extension of the friction contact stage, a significant increase in energy-saving losses, and a faster growth rate in the sealing stiffness degradation index, the current health status is determined to be severely degraded. When the stage change results show that the re-stabilization stage is difficult to recover, a continuous deterioration in energy-saving losses, and a high and continuing-increasing sealing stiffness degradation index, the current health status is determined to be at risk of failure.

[0158] S5.2. Based on the current health status conclusion, evolution direction results, changes in energy-saving losses, and sealing stiffness degradation trend results, determine the energy-saving fan failure prediction results; when the current health status conclusion changes progressively from healthy to slightly degraded, moderately degraded, or severely degraded, and the evolution direction results continue to rise, the energy-saving fan failure prediction results indicate an increased risk of sealing degradation; when the current health status conclusion remains unchanged, the evolution direction results tend to stabilize, and the sealing stiffness degradation index does not show a continuous upward trend, the energy-saving fan failure prediction results indicate a stable sealing state; when the current health status conclusion further changes from severely degraded to failure risk, and the changes in energy-saving losses continue to rise, and the sealing stiffness degradation index maintains a high growth rate, the energy-saving fan failure prediction results indicate an increased risk of sealing failure.

[0159] The current health status conclusion and the energy-saving fan failure prediction results are matched with the processing rules corresponding to the sealing assembly to obtain the maintenance strategy result. The processing rules corresponding to the sealing assembly are established based on the statistical results of historical operating data. Specifically, a conditional discrimination relationship is constructed according to the health status category, evolution direction result, energy loss change, and sealing stiffness degradation trend result. Each processing rule consists of a discrimination condition composed of the health status conclusion, evolution direction result, energy loss change, and sealing stiffness degradation trend result, and a one-to-one correspondence is established with the corresponding maintenance strategy. The matching process adopts a conditional judgment method, that is, the current health status conclusion, evolution direction result, energy loss change, and sealing stiffness degradation trend result are compared with the discrimination conditions in the processing rule item by item. When each item simultaneously meets the discrimination condition in the same processing rule, the corresponding maintenance strategy result is determined.

[0160] The maintenance strategy results are a set of disposal conclusions that correspond one-to-one with the current operating status. Specifically, they include abnormal contact development results, contact recovery but abnormal energy consumption results, and stable operation results. Abnormal contact development results correspond to the combination of a continuous extension of the friction contact stage, a continuous increase in energy-saving losses, and a continuous increase in the sealing stiffness degradation index. Contact recovery but abnormal energy consumption results correspond to the combination of recovery of the re-stabilization stage, an increase in power consumption per unit flow rate, and a failure of the sealing stiffness degradation index to return to a stable range. Stable operation results correspond to the combination of stable stage changes, energy-saving losses, and the sealing stiffness degradation index.

[0161] When the maintenance strategy results indicate that abnormal contact continues to prolong and energy loss continues to increase, output a sealing maintenance recommendation to stop the machine and check the sealing end face, sealing gap, and sealing seat fit. When the maintenance strategy results indicate that abnormal contact continues to prolong, energy loss continues to increase, and the sealing stiffness degradation index continues to increase, output a sealing maintenance recommendation to stop the machine and check the wear of the sealing end face, the expansion of the sealing gap, and the loosening of the sealing seat fit. When the maintenance strategy results indicate that the contact condition has recovered but the energy loss still has an expanding trend, output a sealing maintenance recommendation to be included in the next maintenance window and to review the sealing surface. When the maintenance strategy results indicate that the contact condition has recovered but the sealing stiffness degradation index has not yet fallen back to a stable range, output a sealing maintenance recommendation to review the flatness of the sealing end face, the sealing gap, and the residual area of ​​the circumferential hot spot in the next maintenance window. When the maintenance strategy results indicate that the degradation condition remains stable, output a sealing maintenance recommendation to continue operation and maintain online monitoring.

[0162] This embodiment also provides an energy-saving fan fault prediction system based on multi-sensor fusion, including: a disturbance acquisition module, which applies a short-term speed disturbance during the stable operation segment of the energy-saving fan and simultaneously acquires inlet pressure, outlet pressure, flow rate, motor current, motor power, bearing vibration, acoustic emission of the sealing cavity, temperature of the sealing seat and thermal image of the shell to obtain a disturbance period synchronous time sequence segment;

[0163] The time-segment recombination module recombines the synchronous time segments of the disturbance cycle into the periods before, during, and after the disturbance. It also combines the continuous enhancement of acoustic emission from the sealed cavity, the slow recovery of the outlet pressure, the lag in the decline of motor power, and the expansion of the circumferential hot spot in the seal to determine the changes in the sealing contact, thus obtaining the contact state sequence and the degradation characteristic sequence.

[0164] The graph convolutional coding module constructs pre-disturbance adjacency relationships, mid-disturbance adjacency relationships, disturbance recovery adjacency relationships, and stage state transition relationships based on the contact state sequence. It also combines the disturbance periodic synchronization time sequence and the degradation feature sequence, and uses a graph convolutional neural network to perform joint encoding of position association and temporal association to obtain the fused state sequence.

[0165] The time-series prediction module performs bidirectional long short-term memory prediction on the fused state sequence to obtain the degenerate state sequence;

[0166] The predictive management module performs health status assessment and maintenance strategy matching on the degradation state sequence, and outputs energy-saving fan failure prediction results and seal maintenance suggestions.

[0167] In summary, this invention applies short-term speed perturbations within stable operating segments, causing the sealing end face to generate a repeatable response process under controlled conditions. This avoids interference from natural operating condition fluctuations on data consistency and improves the comparability and reliability of data sources. By simultaneously acquiring pressure, flow rate, motor power, acoustic emission, and thermal imaging information, a multi-dimensional characterization of sealing contact, fluid transport, and thermal evolution is achieved, effectively enhancing the perception of latent leaks and minute contact changes. By dividing the perturbation cycle into a frictional contact stage, a separation transition stage, and a restabilization stage, and combining differential pressure maintenance characteristics, unit flow power consumption characteristics, temperature rise extension characteristics, and acoustic emission enhancement characteristics, a physically meaningful degradation feature sequence is constructed, expanding the degradation characterization from a single indicator to a multi-dimensional coupled indicator. Graph convolutional neural networks are used to jointly encode the positional correlation between sensors and the temporal correlation between stages, enabling the fusion and expression of multi-source information at the structural level. A bidirectional long short-term memory network is used to model the continuous fusion sequence of multiple perturbation cycles, allowing the degradation state to not only reflect the current state but also characterize the historical evolution path and future trends. By performing health status assessment and maintenance strategy matching on the degradation state sequence, a closed-loop processing from data acquisition to decision output is achieved, transforming fault prediction and health management from post-analysis to pre-judgment. This not only allows for the early identification of seal degradation risks but also provides targeted maintenance recommendations based on changes in energy-saving losses, thereby improving overall energy efficiency while ensuring equipment operational reliability.

[0168] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention, and all such modifications or substitutions should be covered within the scope of the claims of the present invention.

Claims

1. A method for predicting faults in energy-saving wind turbines based on multi-sensor fusion, characterized in that: include, A short-term speed disturbance is applied within the stable operation segment of the energy-saving fan, and inlet pressure, outlet pressure, flow rate, motor current, motor power, bearing vibration, acoustic emission of the sealing cavity, temperature of the sealing seat, and thermal image of the shell are collected simultaneously to obtain the disturbance period synchronization time sequence segment. The disturbance cycle synchronization time sequence segments are recombined into the periods before disturbance, during disturbance, and disturbance recovery. Combined with the continuous enhancement of acoustic emission from the sealed cavity, the slow recovery of outlet pressure, the lag in motor power decline, and the expansion of circumferential hot spots in the seal, the changes in sealing contact are determined, resulting in a contact state sequence and a degradation characteristic sequence including the sealing stiffness degradation index. Based on the contact state sequence and the degradation feature sequence containing the sealing stiffness degradation index, the pre-disturbance adjacency relationship, the mid-disturbance adjacency relationship, the disturbance recovery adjacency relationship, and the stage state transition relationship are constructed. Combined with the disturbance period synchronization time sequence segment, physical edge weights are assigned to each adjacency relationship according to the fluid-structure interaction constraint of the sealing end face to form a physical constraint stage adaptive joint graph structure. The graph convolutional neural network is used to perform joint encoding of position association and temporal association to obtain the fused state sequence. Bidirectional long short-term memory prediction is performed on the fusion state sequence to obtain the degenerate state sequence; The system performs health status assessment and maintenance strategy matching on the degradation state sequence, and outputs energy-saving fan failure prediction results and seal maintenance suggestions.

2. The energy-saving wind turbine fault prediction method based on multi-sensor fusion as described in claim 1, characterized in that: The obtained perturbation period synchronization time segment specifically includes: During the continuous operation of the energy-saving fan, the continuous changes in speed, flow rate and outlet pressure are monitored, and the operating range with slow speed changes, continuous flow rate fluctuations and stable outlet pressure is selected to obtain the stable operating segment. During the stable operation segment, a speed disturbance command is sent to the frequency converter drive to obtain the disturbance period; Data from each sensor is recorded synchronously according to the disturbance period and aligned with a unified time stamp to obtain a disturbance period synchronization time sequence segment.

3. The energy-saving wind turbine fault prediction method based on multi-sensor fusion as described in claim 2, characterized in that: The reorganization of the disturbance period synchronization time sequence segments into pre-disturbance, during-disturbance, and disturbance recovery periods specifically includes: Extract the rotational speed change trajectory and pressure response trajectory from the synchronous time segment of the disturbance period to determine the start time of the rotational speed disturbance, the duration of the disturbance, and the time when the rotational speed falls back to its original value, thus obtaining the pre-disturbance period, the mid-disturbance period, and the disturbance recovery period. The data corresponding to each time period are rearranged in a unified time sequence to obtain the reconstructed time sequence segment.

4. The energy-saving wind turbine fault prediction method based on multi-sensor fusion as described in claim 3, characterized in that: The process of obtaining the contact state sequence and degradation feature sequence specifically includes: The acoustic emission enhancement section, pressure recovery delay section, power fall-off delay section, and sealing circumferential hot spot expansion section are extracted from the reconstructed time sequence segment to obtain the stage response segment; By correlating the stage response segments, the contact stages of the seal transition from frictional contact to separation and then to re-stabilization are divided, resulting in a contact state sequence. The contact state sequence is mapped to the changes in flow rate and pressure difference, and a sealing stiffness degradation index is constructed based on the speed disturbance amplitude, pressure difference recovery response, flow response, power drop response, acoustic emission response, and hot spot expansion response to obtain the degradation characteristic sequence.

5. The energy-saving wind turbine fault prediction method based on multi-sensor fusion as described in claim 4, characterized in that: The mapping of the contact state sequence to changes in flow rate and pressure difference specifically includes: The outlet pressure recovery amplitude and recovery duration are extracted at each stage corresponding to the contact state sequence to obtain the differential pressure maintenance characteristics. The flow rate change and the corresponding motor power consumption are extracted at each stage of the contact state sequence to obtain the power consumption characteristics per unit flow rate. The duration of continuous temperature rise of the sealing seat and the duration of continuous expansion of the circumferential hot spot of the sealing seat are extracted at each stage corresponding to the contact state sequence to obtain the temperature rise extension characteristics. The degree of sustained acoustic emission enhancement in the sealed cavity is extracted to obtain acoustic emission enhancement characteristics; The response slope between the differential pressure maintenance characteristics and the flow rate change characterizes the equivalent stiffness response term of the seal. The motor power drop hysteresis, the degree of acoustic emission enhancement, and the duration of continuous expansion of the circumferential hot spot of the seal characterize the seal contact degradation correction term. The equivalent stiffness response term of the seal and the seal contact degradation correction term are normalized and fused to obtain the seal stiffness degradation index. By arranging the differential pressure maintenance characteristics, unit flow power consumption characteristics, temperature rise extension characteristics, acoustic emission enhancement characteristics, and sealing stiffness degradation index in a stage-wise order, a degradation characteristic sequence is obtained.

6. The energy-saving wind turbine fault prediction method based on multi-sensor fusion as described in claim 5, characterized in that: The method of using graph convolutional neural networks for joint encoding of positional and temporal associations specifically includes: Based on the three-stage contact changes corresponding to the contact state sequence, the inlet pressure, outlet pressure, flow rate, motor current, motor power, bearing vibration, acoustic emission of the sealing cavity, temperature of the sealing seat, and thermal image of the shell are mapped as graph nodes to obtain the stage node set; Based on the sensor installation position relationship, a pre-disturbance adjacency relationship is constructed; based on the coupling relationship between the sealed end face and the flow channel response, a mid-disturbance adjacency relationship is constructed; based on the stage response continuity relationship during the speed drop process, a disturbance recovery adjacency relationship is constructed, resulting in a stage correlation diagram. Based on the fluid-structure interaction model of the sealing end face, the physical transmission direction between outlet pressure, flow rate, acoustic emission of the sealing cavity, temperature of the sealing seat and thermal image of the shell is determined, and the initial edge weights corresponding to each physical transmission direction are determined according to the sealing gap sensitivity coefficient, pressure difference response sensitivity coefficient, flow leakage sensitivity coefficient, friction heat transfer coefficient and circumferential heat diffusion distance. Based on the frictional contact stage, separation transition stage and restabilization stage corresponding to the contact state sequence, the initial edge weights are modulated in stages to obtain the physical edge weights before disturbance, the physical edge weights during disturbance and the physical edge weights after disturbance recovery. By connecting the corresponding nodes in the stage association diagram according to the change order of the contact state sequence, the stage transition diagram is obtained; Map the stage node data and degradation feature sequence corresponding to the stage node set in the perturbation period synchronization time segment to the stage association graph and stage transition graph to obtain the physical constraint stage adaptive joint graph structure. By using a graph convolutional neural network to perform node propagation, stage aggregation, and sequence encoding along a physically constrained stage adaptive joint graph structure, a fused state sequence is obtained.

7. The energy-saving wind turbine fault prediction method based on multi-sensor fusion as described in claim 6, characterized in that: The method of using a graph convolutional neural network to perform node propagation, stage aggregation, and sequence encoding along a physically constrained stage adaptive joint graph structure specifically includes: The positional association features of the stable operation phase are extracted along the adjacency relationship before the disturbance to obtain the state features before the disturbance; The stage coupling features between the sealing end face and the flow channel are extracted along the adjacency relationship in the disturbance to obtain the state features in the disturbance; During the node propagation process of pre-disturbance state characteristics, during-disturbance state characteristics, and during-disturbance recovery state characteristics, the intensity of node information transmission is limited by the physical edge weight of the corresponding stage, and the stage representation of the acoustic emission node of the sealing cavity, the temperature node of the sealing seat, and the thermal imaging node of the shell is corrected by the sealing stiffness degradation index. The pre-disturbance state features, mid-disturbance state features, and post-disturbance state features are concatenated along the stage state transition relationship and fused with the corresponding degradation feature sequence to obtain the fused state sequence.

8. The energy-saving wind turbine fault prediction method based on multi-sensor fusion as described in claim 7, characterized in that: The bidirectional long short-term memory prediction of the fused state sequence specifically includes: Extract the fusion state sequence corresponding to the current disturbance cycle, and retrieve the fusion state sequences corresponding to several historical disturbance cycles that are consecutive to the current disturbance cycle. Then, concatenate them in chronological order to obtain a continuous fusion sequence. By using a bidirectional long short-term memory network to perform forward and backward modeling on the continuous fusion sequence, a degenerate state sequence is obtained.

9. The energy-saving wind turbine fault prediction method based on multi-sensor fusion as described in claim 8, characterized in that: The process of determining the health status and matching maintenance strategies for the degraded state sequence specifically includes: Extract the stage change results, continuous change results, evolution direction results, and energy-saving loss changes from the degradation state sequence to obtain the discrimination results; The growth trend of the sealing stiffness degradation index is extracted from the degradation state sequence to obtain the sealing stiffness degradation trend results; The judgment results and sealing stiffness degradation trend results are correlated with the preset health status categories to obtain the current health status conclusion; Based on the current health status conclusions, evolution direction results, changes in energy-saving losses, and sealing stiffness degradation trends, the failure prediction results for energy-saving fans are determined. The current health status conclusion and the energy-saving fan failure prediction result are matched with the processing rules corresponding to the sealing assembly to obtain the maintenance strategy result; When maintenance strategy results indicate that abnormal contact continues to prolong and energy loss continues to increase, output a shutdown inspection of the sealing end face, sealing gap and sealing seat fit condition as a sealing maintenance recommendation. If the maintenance strategy results indicate that the contact condition has recovered and the energy loss still has an expanding trend, output the sealing maintenance recommendations for the next maintenance window and review the sealing surface. If the maintenance strategy results indicate that the degradation state remains stable, output a seal maintenance recommendation to continue operation and maintain online monitoring.

10. A fault prediction system for energy-saving wind turbines based on multi-sensor fusion, based on the fault prediction method for energy-saving wind turbines based on multi-sensor fusion as described in any one of claims 1 to 9, characterized in that: include, The disturbance acquisition module applies a short-term speed disturbance during the stable operation segment of the energy-saving fan and simultaneously acquires inlet pressure, outlet pressure, flow rate, motor current, motor power, bearing vibration, acoustic emission of the sealing cavity, temperature of the sealing seat, and thermal image of the shell to obtain a disturbance cycle synchronous time sequence segment. The time-period recombination module recombines the synchronous time sequence segments of the disturbance cycle into the periods before, during, and after the disturbance. It also combines the continuous enhancement of acoustic emission from the sealed cavity, the slow recovery of the outlet pressure, the lag in the decline of motor power, and the expansion of the circumferential hot spot of the seal to determine the changes in the sealing contact, thereby obtaining the contact state sequence and the degradation characteristic sequence containing the sealing stiffness degradation index. The graph encoding module constructs pre-disturbance adjacency relationships, mid-disturbance adjacency relationships, disturbance recovery adjacency relationships, and stage state transition relationships based on the contact state sequence and the degradation feature sequence containing the sealing stiffness degradation index. It also combines the disturbance period synchronization time segment and assigns physical edge weights to each adjacency relationship according to the fluid-structure interaction constraint of the sealing end face to form a physical constraint stage adaptive joint graph structure. The graph convolutional neural network is used to perform joint encoding of position association and temporal association to obtain the fused state sequence. The time-series prediction module performs bidirectional long short-term memory prediction on the fused state sequence to obtain the degenerate state sequence; The predictive management module performs health status assessment and maintenance strategy matching on the degradation state sequence, and outputs energy-saving fan failure prediction results and seal maintenance suggestions.

Images