Cross-dimension impeller machinery dynamic risk prediction and visualization system

Abstract

The invention relates to the technical field of turbomachinery monitoring, and discloses a cross-dimension turbomachinery dynamic risk prediction and visualization system. A dynamic data acquisition module of the system acquires multi-source heterogeneous sensing data such as a vibration spectrum sequence, an aerodynamic load distribution matrix, a rotor axial displacement time sequence and a surface temperature field thermodynamic diagram in real time; the service degradation evaluation module extracts a vibration characteristic spectrum based on a historical degradation mode library, calculates an instability risk level in combination with an aerodynamic load deviation value, and fuses a displacement sudden change gradient and a temperature difference abnormal area to generate a service degradation track matrix; the multi-dimensional risk prediction module inputs the degradation track matrix into a space-time coupling prediction network, and outputs a dynamic risk assessment vector including time-varying fault probability distribution, a residual life prediction value and a risk propagation path set; and the risk visualization engine module constructs a three-dimensional risk thermal field model and generates a dynamic thermal grid layer superposed on the digital twinborn body.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of impeller machinery monitoring, in particular to a cross-dimension impeller machinery dynamic risk prediction and visualization system. BACKGROUND

[0002] As the core equipment in the fields of energy, power, aerospace, etc., the running state of impeller machinery is directly related to the continuity and safety of industrial production. Such equipment is usually served for a long time under complex working conditions of high temperature, high pressure and high speed rotation, and is prone to failure due to problems such as vibration anomaly, aerodynamic instability, rotor deviation or temperature anomaly, and even causes shutdown accidents, causing huge economic losses. Therefore, real-time monitoring and risk prediction of the running state of impeller machinery have become a key link to ensure the stable operation of industrial systems. In the existing impeller machinery monitoring technology, data acquisition mainly relies on single type sensors, such as obtaining vibration signals only through vibration sensors, or monitoring equipment temperature only by relying on temperature sensors, which is difficult to fully capture the complex running state of the equipment. The heterogeneous characteristics of multi-source data, i.e. the time sequence of vibration spectrum, the spatial distribution of aerodynamic load, the dynamic change of displacement data and the image features of temperature field, make the fusion analysis of different types of data face challenges. Traditional methods often process each type of data in isolation, and cannot establish the correlation between multiple parameters, resulting in one-sidedness in the evaluation of the degradation state of the equipment. In terms of service degradation evaluation, the existing technology is mostly based on threshold judgment of a single parameter, such as determining whether the equipment is abnormal only by whether the vibration amplitude exceeds the preset value, ignoring the coupling effect between parameters. For example, abnormal distribution of aerodynamic load may be related to sudden change of rotor axial displacement, and local abnormality of temperature field may also exacerbate the deterioration of vibration characteristics, so single parameter evaluation cannot reflect the overall degradation trend of the equipment. At the same time, the application of historical degradation patterns is insufficient, which makes it impossible to combine the long-term operation law of the equipment for trend prediction, and the dynamic adaptability of the evaluation result is poor. In the risk prediction link, the existing model mostly adopts static analysis method, which fails to fully consider the time-varying characteristics and spatial correlation characteristics of the equipment running state. The identification of risk propagation path lacks systematicness, and it is difficult to accurately capture the diffusion law of faults among different components, resulting in insufficient timeliness and accuracy of the prediction result. In addition, the risk visualization means is relatively simple, mostly in the form of numerical table or two-dimensional curve, which makes it difficult for the operator to intuitively understand the spatial distribution and dynamic change of the risk, affecting the timeliness and effectiveness of risk disposal. With the advancement of industrial digitization, digital twin technology is gradually increasing in the application of equipment monitoring, but the existing system has not realized the deep fusion of multi-source heterogeneous data and digital twin, and there is still a technical gap in the dynamic superposition and intuitive display of risk information on the digital twin, which cannot provide an immersive risk perception experience for the operator. SUMMARY

[0003] The present application aims to provide a cross-dimension turbomachinery dynamic risk prediction and visualization system to solve the problems raised in the background.

[0004] To achieve the above-mentioned purpose, the present application provides a cross-dimension turbomachinery dynamic risk prediction and visualization system, which comprises: A dynamic data acquisition module is configured to acquire real-time multi-source heterogeneous sensing data of the turbomachinery, wherein the multi-source heterogeneous sensing data comprises a vibration frequency spectrum sequence, an aerodynamic load distribution matrix, a rotor axial displacement time sequence, and a surface temperature field thermal map. A service degradation evaluation module is configured to extract a current vibration feature spectrum based on a historical degradation mode library of the vibration frequency spectrum sequence, calculate an aerodynamic instability risk level according to a deviation amount of the aerodynamic load distribution matrix from a preset working condition threshold, and generate a service degradation trajectory matrix by fusing a mutation gradient of the rotor axial displacement time sequence and an area of a temperature difference abnormal region of the surface temperature field thermal map. A multi-dimensional risk prediction module is configured to input the service degradation trajectory matrix into a space-time coupling prediction network, and output a dynamic risk evaluation vector comprising a time-varying failure probability distribution, a residual life prediction value of a key component, and a risk propagation path set. A risk visualization engine module is configured to construct a three-dimensional space risk thermal field model according to the dynamic risk evaluation vector, and generate a dynamic thermal grid layer superimposed on a turbomachinery digital twin.

[0005] Preferably, the dynamic data acquisition module comprises: A vibration feature analysis unit is configured to perform wavelet packet energy spectrum decomposition on the vibration frequency spectrum sequence, extract energy entropy values in a preset frequency band interval, and mark abnormal frequency band positions with energy entropy values exceeding an entropy threshold value. A load distribution verification unit is configured to compare the aerodynamic load distribution matrix with a reference load distribution template under a standard working condition, and calculate an area ratio and a maximum overrun amplitude of a local load overrun region. A displacement cooperative analysis unit is configured to associate the rotor axial displacement time sequence with a rotational speed pulse signal, and identify a phase synchronization error between a displacement mutation point and a rotational speed step point. A temperature field fusion unit is configured to map the surface temperature field thermal map to a three-dimensional grid model of the turbomachinery, and locate a spatial coordinate set of a temperature difference abnormal region.

[0006] Preferably, the service degradation evaluation module comprises: A degradation mode matching unit is configured to retrieve a degradation mode matched with the abnormal frequency band position from the historical degradation mode library, and output a first risk level. The instability risk quantification unit is configured to calculate the aerodynamic instability risk grade according to the exceeding amplitude gradient when the area proportion of the local load exceeding area exceeds an area threshold and the maximum exceeding amplitude exceeds an amplitude threshold. The displacement risk correlation unit is configured to generate an axial displacement risk coefficient according to the phase synchronization error and the amplitude increment of the displacement mutation point. The fusion calculation unit is configured to generate the service degradation trajectory matrix by weightedly fusing the first risk grade, the aerodynamic instability risk grade, and the axial displacement risk coefficient.

[0007] Preferably, the multi-dimensional risk prediction module comprises: The space-time feature extraction unit is configured to perform a space-time convolution operation on the service degradation trajectory matrix to extract a cross-scale risk feature vector. The failure probability prediction unit is configured to output the time-varying failure probability distribution in a future time window by processing the cross-scale risk feature vector through a gated recurrent network. The life attenuation calculation unit is configured to calculate the remaining life prediction value of the key component according to a mapping relationship between the cross-scale risk feature vector and a component material fatigue curve. The path deduction unit is configured to generate the risk propagation path set by simulating the diffusion process of risk in the topological structure of the turbomachinery based on a graph neural network.

[0008] Preferably, the risk visualization engine module comprises: The space mapping unit is configured to map the time-varying failure probability distribution to corresponding grid cells of a three-dimensional grid model of the turbomachinery. The thermal field generation unit is configured to calculate color gradient values of the grid cells according to the remaining life prediction value of the key component, and generate the three-dimensional space risk thermal field model by combining the spatial coordinates of the risk propagation path set. The dynamic rendering unit is configured to synchronously update the three-dimensional space risk thermal field model with the operating state of the turbomachinery, and output the dynamic thermal grid layer.

[0009] Preferably, the system further comprises: The anomaly tracing module is configured to extract all paths in the risk propagation path set that pass through a specific grid cell when the probability value of the specific grid cell in the time-varying failure probability distribution exceeds a probability threshold, and trace back to a source grid cell. The anomaly tracing module comprises a path screening unit configured to calculate the propagation intensity weight of each path, and screen paths with weights exceeding a weight threshold to form a main propagation chain.

[0010] Preferably, the system further comprises an adaptive monitoring module configured to adjust a collection frequency of corresponding sensing data according to a spatial position of the source grid cell. The adaptive monitoring module comprises a frequency decision unit configured to calculate an increment coefficient of the collection frequency based on a path length and a propagation intensity weight of the main propagation chain.

[0011] Preferably, the system further comprises a heat grid optimization module configured to perform a grid density dynamic adjustment on the dynamic heat grid layer. The heat grid optimization module comprises: a risk gradient analysis unit configured to calculate a risk probability gradient value between adjacent grid cells; a density controller unit configured to insert a subdivided grid layer in a corresponding region when the risk probability gradient value exceeds a gradient threshold value; a transition layer generation unit configured to configure a gradient transition grid between the subdivided grid layer and the original grid layer.

[0012] Preferably, the system further comprises a multi-dimensional data cleaning module configured to calculate a historical data storage redundancy according to an update frequency of the dynamic risk assessment vector. The multi-dimensional data cleaning module comprises: a redundancy evaluation unit configured to count a number of historical versions of the service degradation trajectory matrix; a cleaning decision unit configured to delete an earliest version data in a descending order of storage time when the number of historical versions exceeds a version threshold value.

[0013] Preferably, the system further comprises a cloud collaboration module configured to encrypt and upload the dynamic risk assessment vector to a cloud knowledge base. The cloud collaboration module comprises: an encryption retrieval unit configured to initiate a similar case retrieval request to the cloud to obtain a matched historical risk assessment case set; a local correction unit configured to compare a deviation amount between the historical risk assessment case set and the current dynamic risk assessment vector, and correct a weight parameter of the space-time coupled prediction network.

[0014] Compared with the prior art, the present application has the following advantages: The cross-dimension turbomachinery dynamic risk prediction and visualization system realizes comprehensive perception, accurate prediction and intuitive presentation of turbomachinery operation risk through multi-module cooperation. The dynamic data acquisition module breaks through the limitations of traditional single data acquisition, and includes multi-source heterogeneous data such as vibration spectrum sequence, aerodynamic load distribution matrix, rotor axial displacement time sequence and surface temperature field thermal map in the monitoring range, fully capturing the operation characteristics of the equipment in different dimensions, and providing rich basic data support for subsequent risk assessment. The cooperative collection of such multi-source data can more completely reflect the real operation state of the equipment, avoiding risk misjudgment caused by one-sided data. The service degradation assessment module constructs a more comprehensive degradation assessment system by fusing historical degradation patterns and real-time operation parameters. The historical degradation pattern library based on vibration spectrum sequence extracts the current vibration characteristic spectrum, which can identify potential abnormalities combined with the long-term operation rules of the equipment; the aerodynamic instability risk level is calculated by the deviation of the aerodynamic load distribution matrix from the preset working condition threshold, realizing dynamic monitoring of aerodynamic performance; and the service degradation trajectory matrix is generated by fusing the mutation gradient of the rotor axial displacement time sequence and the temperature difference abnormal area of the surface temperature field thermal map, which comprehensively considers the coupling effect of mechanical displacement and thermal characteristics, making the degradation assessment results more consistent with the actual degradation process of the equipment, and more accurately depicting the evolution trajectory of the equipment from normal operation to potential failure. The multi-dimensional risk prediction module introduces a time-space coupling prediction network, fully considering the time dynamic change and spatial correlation characteristics of the equipment operation state. After inputting the service degradation trajectory matrix into the network, the output dynamic risk assessment vector contains time-varying fault probability distribution, key component residual life prediction value and risk propagation path set, which not only can update the risk probability in real time, but also can clearly show the diffusion path of risk among the components of the equipment, making the risk prediction no longer limited to static assessment at a single time point, but forming a dynamic evolving risk panorama. This time-space coupled prediction method can more accurately grasp the risk development trend and provide more targeted information for risk warning. The risk visualization engine module realizes the intuitive presentation of risk information by constructing a three-dimensional space risk thermal field model and superimposing it on the turbomachinery digital twin to generate a dynamic thermal grid layer. The operator can intuitively observe the spatial distribution and dynamic change of risk in different parts of the equipment through the digital twin, without relying on complex numerical analysis to quickly understand the risk state. This visualization method converts abstract risk data into concrete spatial thermal distribution, reducing the difficulty of risk information interpretation, and helping operators more timely and accurately grasp the operation risk of the equipment and improve the efficiency of risk disposal. The cooperation of each module forms a complete closed loop from data acquisition, degradation evaluation, risk prediction to visualization presentation, realizes the whole process dynamic management of the risk of turbomachinery, adapts to the actual demand of equipment monitoring under complex working conditions, and provides all-round technical support for the safe and stable operation of turbomachinery. BRIEF DESCRIPTION OF DRAWINGS

[0015] Figure 1 A timing diagram of the cross-dimension turbomachinery dynamic risk prediction and visualization system described in the application; Figure 2 A workflow diagram of the service degradation evaluation module; Figure 3 A workflow diagram of the multi-dimensional risk prediction module; Figure 4 A workflow diagram of the abnormality tracing module. DETAILED DESCRIPTION

[0016] The technical solutions in the embodiments of the application will be clearly and completely described below with reference to the drawings in the embodiments of the application. Obviously, the described embodiments are only part of the embodiments of the application, rather than all the embodiments. Based on the embodiments in the application, all other embodiments obtained by those skilled in the art without creative labor fall within the scope of protection of the application.

[0017] Please refer to Figure 1 The application provides a cross-dimension turbomachinery dynamic risk prediction and visualization system, which comprises: The dynamic data acquisition module continuously obtains real-time data streams from various sensors installed on the turbomachinery. These data include vibration spectrum sequences representing the dynamic characteristics of the machinery, aerodynamic load distribution matrices reflecting the distribution state of aerodynamic load, rotor axial displacement time series indicating the axial position of the rotor, and surface temperature field thermograms depicting the temperature distribution of the machinery surface. These data constitute a multi-source heterogeneous sensor data set.

[0018] The service degradation evaluation module receives the above-mentioned multi-source heterogeneous sensor data and performs comprehensive evaluation. The module uses the stored historical degradation mode library to perform feature extraction and pattern matching on the currently obtained vibration spectrum sequences. At the same time, the deviation between the aerodynamic load distribution matrix and the preset standard working condition threshold is calculated to evaluate the aerodynamic stability. In addition, the mutation characteristics in the rotor axial displacement time series and their synchronicity with the speed change are analyzed, and the area information of the abnormal temperature difference area in the surface temperature field thermogram is combined. By fusing the above-mentioned evaluation results in multiple aspects, a service degradation trajectory matrix is generated, which comprehensively represents the current health state and development trend of the equipment.

[0019] The multi-dimensional risk prediction module receives the service degradation trajectory matrix as the core input. The module utilizes a spatio-temporal coupled prediction network to process the input matrix. The network analyzes the spatio-temporal evolution patterns contained in the trajectory matrix and outputs a dynamic risk assessment vector. The vector contains three key prediction information: the probability distribution of different component or region failures in the future specific time period, the expected remaining working time of key components before failure, and the potential diffusion direction and path of potential failures among the internal structure of the equipment.

[0020] The risk visualization engine module receives the dynamic risk assessment vector and converts it into intuitive visualization information. Based on the three-dimensional digital model of the turbomachinery, the module spatially maps the predicted risk information. Specifically, it constructs a three-dimensional spatial risk thermal field model that expresses the risk probability, remaining life, and risk propagation direction at different positions on the three-dimensional grid in the form of color gradient and transparency changes. Finally, this dynamically changing risk thermal field model is superimposed on the turbomachinery digital twin in the form of a dynamic thermal grid layer, realizing real-time, three-dimensional, and intuitive display of risks.

[0021] Embodiment 1: refer to Figure 2 The dynamic data acquisition module serves as the data entry of the system, and its core function is to capture and preliminarily process multi-source heterogeneous sensor data from turbomachinery in real time. The module contains four functional units, each processing a specific type of data stream. The vibration feature analysis unit is responsible for processing continuous vibration spectrum sequence data. The unit first performs wavelet packet transform decomposition on the input vibration spectrum sequence. This decomposition process decomposes the original spectrum signal into a series of pre-set, mutually orthogonal frequency band intervals. After decomposition, the unit calculates the energy entropy value in each specified frequency band interval. The energy entropy value is a quantitative indicator that measures the complexity or disorder degree of signal energy distribution in a specific frequency band. After calculation, the unit compares the energy entropy value of each frequency band with the pre-set entropy threshold value. When the energy entropy value of a certain frequency band continuously or significantly exceeds its corresponding entropy threshold value, the frequency band is marked as an abnormal frequency band, and its specific location information is recorded. These marked abnormal frequency band positions indicate possible abnormal frequency components in the vibration signal.

[0022] The load distribution verification unit focuses on processing aerodynamic load distribution matrix data. The unit internally stores or has access to a pre-established reference load distribution template under standard operating conditions. This template represents the ideal or typical aerodynamic load spatial distribution of the equipment in a normal, stable operating state. The real-time acquired aerodynamic load distribution matrix is compared point by point with the reference template. During the comparison process, the difference between the real-time load value at each spatial position and the threshold value at the corresponding position of the reference template is calculated. All spatial points whose real-time load values exceed the threshold values at their corresponding positions are identified, and these points form the local load overrun area. The unit further calculates the proportion of the area of these overrun areas to the total area of the entire analysis area. At the same time, among all the overrun points, the maximum overrun amplitude value whose load value exceeds the threshold value is found. The area proportion and the maximum overrun amplitude are key indicators for quantifying the degree of deviation of the current aerodynamic load distribution from the normal state.

[0023] The displacement coordination analysis unit processes two associated signals: the rotor axial displacement time series signal and the rotational speed pulse signal. The unit first analyzes the rotor axial displacement time series curve and detects points where the displacement value changes suddenly and significantly, i.e., displacement mutation points. These mutation points usually correspond to abnormal changes in the rotor operating state. At the same time, the unit analyzes the rotational speed pulse signal and identifies points where the rotational speed changes in steps, i.e., rotational speed step points. Rotational speed step points usually reflect changes in device control instructions or load state. To evaluate the coordination between rotor axial displacement changes and rotational speed changes, the unit calculates the time difference between each detected displacement mutation point and the nearest rotational speed step point. This time difference is the phase synchronization error. A larger phase synchronization error indicates that the displacement change has not followed the rotational speed change in time, which may indicate mechanical response delay or potential problems. In addition, the unit also records the amplitude increment of the displacement mutation point, i.e., the magnitude of the displacement mutation.

[0024] The temperature field fusion unit processes surface temperature field thermograph data. The core task of this unit is to associate two-dimensional temperature distribution image information with the three-dimensional geometry of the turbomachinery. It first maps the acquired surface temperature field thermograph data to the three-dimensional grid model surface of the turbomachinery. The mapping process ensures that temperature data points and surface grid elements of the three-dimensional grid model establish a corresponding relationship. After mapping is complete, the unit judges the temperature value of each grid element according to the normal temperature range threshold set for the current operating condition of the equipment. Grid elements whose temperature values deviate significantly from their normal temperature range are identified as temperature difference abnormal points. The continuous area formed by the aggregation of these abnormal points is the temperature difference abnormal area. The unit finally outputs the spatial coordinate set of all grid elements covered by these temperature difference abnormal areas, accurately identifying the spatial location of the abnormal high temperature on the surface of the equipment.

[0025] The service degradation assessment module receives the processing results from the above four data acquisition units, conducts comprehensive state assessment and generates a service degradation trajectory matrix. The degradation mode matching unit utilizes the abnormal frequency band position information output by the vibration feature analysis unit. This unit accesses the historical degradation mode library, which is a knowledge base storing the characteristic abnormal frequency band patterns appearing in the vibration spectrum when various typical failure modes occurred in the device in history and the corresponding risk level information. The unit searches the historical library for the most matching or similar historical degradation mode to the currently detected abnormal frequency band position. The matching process may involve pattern similarity calculation or rule matching. After successful matching, the unit outputs the first risk level associated with the historical mode, which represents the degree of failure risk implied by the current vibration anomaly.

[0026] The instability risk quantification unit processes the area proportion and maximum overrun amplitude data provided by the load distribution verification unit. According to the preset logic, only when the area proportion of the local load overrun area exceeds a set area threshold and the maximum overrun amplitude also exceeds a set amplitude threshold, it is considered that the current aerodynamic state has a significant instability risk. If both conditions are met, the unit outputs a specific, quantified aerodynamic instability risk level according to the distribution characteristics of the overrun area and the change gradient of the overrun amplitude through the preset quantification rules or calculation models. This level reflects the possibility of system instability caused by abnormal aerodynamic load distribution.

[0027] The displacement risk correlation unit processes the phase synchronization error and displacement mutation point amplitude increment data provided by the displacement coordination analysis unit. According to the size of the phase synchronization error and the amplitude increment of the displacement mutation, combined with the preset correlation rules or functions, the unit calculates and generates an axial displacement risk coefficient. Generally, the larger the phase synchronization error, the more serious the problem of displacement and speed change out of synchronization, and the higher the risk; the larger the amplitude increment of the displacement mutation, the more severe the degree of displacement anomaly, and the higher the risk. This coefficient comprehensively reflects the risk level implied by the rotor axial displacement anomaly.

[0028] The fusion computing unit is the core output unit of the service degradation assessment module. It receives the first risk level from the degradation pattern matching unit, the aerodynamic instability risk level from the instability risk quantification unit, and the axial displacement risk coefficient from the displacement risk correlation unit. In addition, the spatial coordinate set information of the temperature difference abnormal area may also be input or used to assist in weight distribution. This unit performs weighted fusion calculation on the three main risk indicators according to the pre-set weight coefficient scheme. The weight coefficient can be determined based on expert experience, historical data analysis or equipment criticality analysis, for example, more attention may be paid to vibration risk or aerodynamic risk under certain working conditions. The result of the fusion calculation is a multi-dimensional matrix, i.e. a service degradation trajectory matrix. The structure of this matrix is designed to comprehensively represent the degradation state and trend of the equipment. The rows of the matrix may represent different evaluation dimensions or indicators, and the columns may represent different time points or different equipment sub-regions / components. Each element value in the matrix integrates the degradation degree information of the corresponding dimension, time point or region. This service degradation trajectory matrix provides the core, multi-source information fusion input data for the subsequent risk prediction module.

[0029] Embodiment 2: see Figure 3 The core function of the multi-dimensional risk prediction module is to use deep learning technology to deeply analyze the service degradation trajectory matrix and predict the future risk state of the equipment. This module relies on an architecture called a spatio-temporal coupling prediction network, which is composed of multiple functional units. The spatio-temporal feature extraction unit is the starting point for the network to process input data. This unit receives the service degradation trajectory matrix from the service degradation assessment module as input. The matrix is usually a multi-dimensional data structure, which contains spatial information, temporal information and multiple evaluation indicators. The core operation of the spatio-temporal feature extraction unit is to perform spatio-temporal convolution. This involves using a specially designed convolution kernel to slide across the spatial and temporal dimensions of the matrix. Convolution operations can capture both the correlation of spatially adjacent region states and the evolution patterns over time. For example, a three-dimensional convolution kernel can slide in both spatial and temporal dimensions to extract feature patterns within a local spatio-temporal neighborhood. Another implementation is to use separate convolutions: first use a two-dimensional spatial convolution kernel to extract spatial correlation features between different regions at the same time point, and then use a one-dimensional temporal convolution kernel to extract time series features of the state change over time at the same spatial position. The purpose of these convolution operations is to extract abstract cross-scale risk feature patterns from complex service degradation trajectories that can reflect the essence of risk evolution. After multiple layers of convolution processing, the unit outputs a highly condensed cross-scale risk feature vector that contains key information representing the risk evolution trend of the current and historical state of the equipment.

[0030] The failure probability prediction unit receives the cross-scale risk feature vector output by the spatio-temporal feature extraction unit. Since the risk feature vector usually contains time series information, this unit employs a gated recurrent network structure for processing. Gated recurrent networks, such as long short-term memory (LSTM) or gated recurrent unit (GRU), are chosen for their ability to effectively learn and memorize long-term dependencies in time series. The network structure of this unit learns the dynamic patterns of the risk feature vector over time. Based on the learned patterns, the network predicts the likelihood of failure of each component or spatial region of the device within a future set time window. The output result is a time-varying failure probability distribution. This distribution is a complex data structure, usually containing spatial and temporal dimensions. In the spatial dimension, it identifies each grid cell or component on the device model; in the temporal dimension, it gives the probability estimate of each spatial cell failing at a series of discrete time points in the future. For example, it can be represented as a three-dimensional array: dimension one represents the spatial location, dimension two represents the future time point, and dimension three stores the predicted failure probability value of that location at that time point. This distribution dynamically depicts the evolution of the risk probability over time on the spatial structure of the device.

[0031] The life degradation calculation unit also receives the cross-scale risk feature vector as input. The focus of this unit is to predict the remaining useful life of the key components before failure. Its working principle is based on the correlation between physical mechanisms such as material fatigue, wear, or aging and the operating state of the device. The unit stores or accesses material fatigue curves or empirical degradation models of key components internally. The cross-scale risk feature vector contains information reflecting the current stress state, cumulative damage degree, or degradation rate of the components. The life degradation calculation unit establishes a mapping relationship between these risk features and the material fatigue curves or degradation models. This mapping can be achieved in several ways: using a trained regression model to directly predict the remaining life from the input features; or based on a physics-based damage accumulation model, using the information in the feature vector to calculate the cumulative damage degree, and then calculating the remaining life. The unit outputs the predicted value of the remaining life of the key components, which is usually a time quantity representing the length of time the component is expected to safely operate under the predicted degradation path.

[0032] The path inference unit is responsible for simulating and predicting the propagation paths of potential faults in the internal structure of the equipment. This unit builds a model based on graph neural networks. First, a topological structure graph of the turbomachinery needs to be constructed. This graph abstracts the physical structure of the equipment: the nodes in the graph represent key components, functional units, or partitioned spatial regions of the equipment; the edges in the graph represent potential fault propagation paths or influence relationships between nodes. The establishment of edges is based on physical connections, functional dependencies, or known fault propagation mechanisms. Each edge can be assigned a weight or attribute, representing the degree of difficulty or strength of propagation. The path inference unit inputs the cross-scale risk feature vector as the initial state or feature of the nodes into the graph neural network. The core operation of the graph neural network is message passing: each node updates its state according to its current state and the states of neighboring nodes and the attributes of the connecting edges. Through multiple rounds of message passing iterations, the network simulates how risk features spread and propagate from the source node to other nodes along the edges in the graph structure. After simulation and inference, the unit outputs a set of risk propagation paths. This set contains multiple predicted path sequences that start from the potential risk source node, propagate through the edge connections in the graph, and reach other nodes. Each path records an ordered sequence of nodes, representing the possible direction chain of risk diffusion. This set reveals how a local problem may evolve into a global problem.

[0033] The task of the risk visualization engine module is to transform the abstract prediction results output by the multi-dimensional risk prediction module into intuitive and understandable three-dimensional visual graphics, and superimpose them on the digital twin model of the turbomachinery. The spatial mapping unit processes the time-varying fault probability distribution. The core of this unit is to map the predicted fault probability values to the corresponding spatial positions of the three-dimensional grid model of the turbomachinery. Specifically, for each grid cell in the three-dimensional grid model, according to its position in the equipment structure, find the predicted probability value at a specific future time point in the time-varying fault probability distribution corresponding to the position. Then, according to the size of the probability value, assign a visual attribute value to each grid cell. The most commonly used visual channels are color and transparency. For example, using a color mapping scheme from cool to warm colors, the higher the probability value, the more the color tends to be red and possibly less transparent. In this way, the spatial distribution of probability values is intuitively presented through the color and transparency of grid cells.

[0034] The thermal field generation unit is responsible for integrating various risk information to construct the final three-dimensional thermal field model. This unit processes the residual life prediction values of critical components and the set of risk propagation paths. First, for critical components, the corresponding color gradient values are calculated based on their residual life prediction values. Here, a different color mapping scheme is usually adopted from the failure probability to avoid confusion. For example, long residual life can be represented by green, medium life by yellow, and short life by orange or red. The color gradient values directly reflect the length of the residual life. Next, the set of risk propagation paths is processed. This unit extracts the spatial coordinates corresponding to the sequence of nodes passed by each path in the set. In the three-dimensional scene, along these sequences of spatial coordinate points, visual path indicators are generated. Common indicators include: connecting the nodes on the path with lines of different colors; adding special markers or highlighting effects on the grid cells passed by the path; or adding icons at the starting point and critical nodes of the path. Finally, the thermal field generation unit superimposes the failure probability distribution processed by the spatial mapping unit, the residual life color gradient values of critical components, and the spatial indication indicators of risk propagation paths on the same three-dimensional grid model. When merging, visual conflicts need to be handled, for example, path indication indicators need to be eye-catching enough to distinguish from the probability thermal map. The final output is a comprehensive three-dimensional spatial risk thermal field model that contains rich risk information. This model simultaneously displays in three-dimensional space: the probability of failure at different locations, the life status of critical components, and the potential direction and path of risk diffusion.

[0035] The dynamic rendering unit is the output link of the visualization engine, ensuring that the visualization results are synchronized with the real-time or predicted state of the device. This unit receives continuous updates from the multi-dimensional risk prediction module or directly receives sensor data streams to drive the state update of the digital twin. When a new dynamic risk assessment vector arrives, the dynamic rendering unit triggers the update of the entire visualization process: the spatial mapping unit updates the grid color / transparency based on the new time-varying failure probability distribution; the thermal field generation unit updates the color of critical components and the path indication based on the new residual life prediction values and the set of risk propagation paths; the thermal field generation unit reconstructs or updates the three-dimensional spatial risk thermal field model. At the same time, this unit is responsible for synchronizing the updated three-dimensional spatial risk thermal field model with the current state of the turbomachinery digital twin. The rendering engine calculates effects such as light, perspective, model transformation, etc. in real time. Finally, the dynamic rendering unit outputs the dynamically generated thermal grid layer to the graphical display device. This layer is dynamically changing, as the prediction results update or the device state changes, the risk thermal map, life indication color block, and risk propagation path line on the layer will be updated in real time, providing users with an immediate and intuitive perception of the current and future risk state of the device.

[0036] Example 3: see Figure 4When the failure probability value corresponding to a certain grid cell in the time-varying failure probability distribution generated by the multi-dimensional risk prediction module exceeds the preset critical threshold, the system triggers an abnormality tracing mechanism. This mechanism analyzes the set of risk propagation paths and reversely traces the potential origin of the high-risk location. The abnormality tracing module first filters out all propagation paths containing the high-risk grid cell from the set of risk propagation paths. These paths record the possible trajectory of risk spreading from the source to the current high-risk location. Each path consists of a series of ordered grid cells, representing the intermediate node sequence of risk propagation.

[0037] The path filtering unit assesses the importance of the extracted propagation paths. This unit calculates the propagation intensity weight of each path, which reflects the activity level and risk transmission ability of the path. The calculation of the propagation intensity weight uses the following formula: ; where: Wp represents the propagation intensity weight of path P; u represents the grid cell on path P; a u is the probability value of grid cell u in the current time-varying failure probability distribution; b u is the activity coefficient of grid cell u in the historical propagation record; l P represents the hop count (path length) of path P; is a small constant introduced to avoid a zero denominator. The activity coefficient b u is obtained by counting the frequency of this grid cell as an intermediate node in historical risk propagation. The higher the frequency, the larger the coefficient.

[0038] The path filtering unit filters out secondary paths with a weight lower than the set threshold based on the calculated propagation intensity weight Wp. The remaining high-weight paths form the main propagation chain set, which are considered as the key propagation channels most likely to cause the current high-risk state. The abnormality tracing module reversely traces along the paths in the main propagation chain set. Starting from the identified high-risk grid cell, the direction of the path node sequence is reversed, and the state of each predecessor node is checked in turn. The tracing process records the risk probability gradient changes of each node on the path and finds the turning point of the probability value surge.

[0039] After the adaptive monitoring module locates the source grid cell, it analyzes the corresponding relationship between its spatial location and sensor deployment. The frequency decision unit calculates the adjustment amount of the collection frequency, mainly considering two characteristic parameters of the main propagation chain: path length L (hop count from the source to the high-risk location) and average propagation intensity weight . Shorter path length and higher average weight mean faster risk propagation speed and direct impact, requiring more intensive monitoring. The increment coefficient of the collection frequency has a non-linear relationship with these two parameters, prioritizing the monitoring density of high-risk source areas.

[0040] In the dynamic data acquisition module, sensor nodes associated with the source grid cell receive frequency adjustment instructions. The sampling rate of the vibration sensor is increased exponentially according to the increment coefficient, ensuring the capture of transient vibration characteristics. The reading interval of the temperature monitoring point is shortened accordingly to track local temperature rise trends. The data output frequency of the pneumatic load sensor is increased synchronously to monitor pressure fluctuation details. These adjustments enable the system to obtain a more complete state evolution process of the source area.

[0041] The anomaly source tracing module works in conjunction with the visualization engine module, and the tracing results are fed back to the three-dimensional space risk thermal field model in real time. The source grid cell is highlighted by special markers. The path on the main propagation chain is rendered in high-light color to enhance visual recognition. The dynamic rendering unit adds pulsation effects to the source area to attract the attention of operators. These visualization elements, together with other state indicators of the digital twin, form a complete risk tracing display.

[0042] When the risk probability of the source area continues to rise, the system will start secondary frequency adjustment. The new increment coefficient is calculated based on the probability change rate ΔP: ; Where: η is the adjusted increment coefficient; η0 is the initial increment coefficient; γ is the sensitivity adjustment parameter; ΔP represents the change in probability value of the source grid cell per unit time. This dynamic adjustment ensures that the monitoring intensity matches the risk evolution speed.

[0043] Each successful tracing instance updates the correlation strength parameters between grid cells, optimizing the accuracy of subsequent path deduction. The system gradually establishes a knowledge base of device-specific risk propagation patterns through continuous accumulation of tracing experience. These knowledge is used to calibrate the edge weight parameters in the graph neural network, improving the relevance of risk prediction.

[0044] Users can select any high-risk grid cell through the interface to trigger immediate tracing analysis. The system displays detailed state parameters of each node on the path while rendering the main propagation chain. The time axis control allows the user to trace back the history of risk propagation and observe the diffusion dynamics during a specific period. These interactive functions enhance the operator's understanding of complex risk evolution. When multiple high-risk sources need to be monitored simultaneously, the module allocates monitoring resources based on the criticality of each region. The source area of the core component is given a higher sampling rate, while the secondary area is appropriately reduced in frequency requirement. This resource allocation algorithm balances the monitoring accuracy and system load, maintaining stable operation performance. The identified source area features are supplemented to the degradation mode library, enriching the correspondence between vibration spectrum and fault types. Newly discovered risk propagation paths are added to the topology graph, expanding the training samples of the graph neural network. This closed-loop learning mechanism enables the system to continuously improve its risk identification capabilities.

[0045] In embodiment 4, the heat grid optimization module continuously analyzes the spatial variation characteristics of the risk distribution in the dynamic heat grid layer. The risk gradient analysis unit is responsible for calculating the risk probability difference between adjacent grid cells. This unit traverses the entire grid model, and for each grid cell, it calculates the absolute value of the difference between its risk probability value and that of all directly adjacent cells. The maximum of these differences is taken as the risk probability gradient value between this cell and its neighbors. This gradient value quantifies the degree of spatial variation of the risk distribution in the local area.

[0046] Table 1: Example of risk probability values and calculated gradient values for some grid cells in the rotor area of a turbomachinery at a certain time.

[0047]

[0048] When the maximum risk probability gradient value between a grid cell and its adjacent cells is detected to exceed this threshold (such as the gradient value of 0.62 between cells G-1027 and G-1101 in the table), it indicates that there is a significant risk boundary or high-risk concentration area in this region, and the current grid resolution may not be sufficient to accurately depict the details of the risk distribution. At this time, the density controller unit inserts a subdivided grid layer at the spatial location corresponding to the high-risk gradient area. The subdivided grid layer uses smaller grid cell sizes, for example, dividing the original size grid cells into multiple sub-cells. The number of inserted subdivision levels can be dynamically determined according to the degree to which the gradient value exceeds the threshold, and the greater the gradient value, the more subdivision levels that may be inserted. For example, for areas with gradient values much higher than the threshold, three levels of subdivided grids may be inserted to form a very fine local grid structure.

[0049] Inserting fine grids directly next to coarse grids can cause obvious jagged boundaries or discontinuities during rendering. The unit automatically generates one or more transition grid layers with intermediate-sized grid cells in the boundary area between the subdivided grid layer and the original grid layer. The cell sizes of these transition grid layers smoothly transition from the original grid size to the subdivided grid size. For example, between the original grid (size S) and the first-level subdivided grid (size S / 2), a transition grid layer with cell sizes of approximately S / 1.5 is inserted; if there is a second-level subdivision (size S / 4), another transition grid layer with cell sizes of approximately S / 3 is inserted between the second-level subdivision and the first-level subdivision. The existence of transition grid layers ensures that the entire grid model is geometrically continuous and visually smoothly transitioning, eliminating visual imperfections caused by sudden changes in grid density and improving the visualization quality and realism of the risk heat field model.

[0050] The multi-dimensional data cleaning module runs independently and is responsible for managing the historical data versions stored by the system, optimizing the utilization of storage resources. During the continuous operation of the system, new service degradation trajectory matrices and dynamic risk assessment vectors are periodically generated, forming a historical data sequence. The redundancy assessment unit monitors the storage status of these core data. The core task of this unit is to count the number of historical versions of the service degradation trajectory matrix currently stored. For example, the system may generate a trajectory matrix version every hour, and the redundancy assessment unit records the total number of versions saved in the past period (such as 24 hours or a week).

[0051] The unit sets a maximum historical version number threshold. When the redundancy assessment unit reports that the number of stored service degradation trajectory matrix historical versions exceeds this threshold, the cleaning decision unit triggers the data deletion process. The deletion strategy strictly follows the principle of reverse order of storage time. The system first identifies all stored historical versions and sorts them according to their generation or storage timestamps, starting from the earliest stored version. The deletion operation continues until the number of remaining stored historical versions is equal to or lower than the set maximum threshold. For example, assuming the threshold is set to retain the last 50 versions, when the system detects that 60 versions are stored, the cleaning decision unit will delete the data of the 10 versions with the earliest generation time. This mechanism ensures that the system always retains the latest and most relevant historical data versions, providing necessary support for possible short-term rollback analysis or model parameter fine-tuning, while effectively controlling the unlimited growth of storage space. Similar data cleaning strategies also apply to other cumulative generated historical data objects such as dynamic risk assessment vectors.

[0052] Embodiment 5: The cloud collaboration module establishes a secure data channel between the local system and the remote cloud knowledge base, realizing prediction optimization based on group experience. When the module starts, the encryption retrieval unit first performs encryption conversion on the latest dynamic risk assessment vector generated locally. The encryption process uses a pre-set cryptography protocol to convert the numerical features in the vector into ciphertext form that cannot be directly interpreted. The converted encrypted data packet is uploaded to the designated receiving interface of the cloud knowledge base through a secure transmission protocol. At the same time, the unit constructs and sends an encrypted similar case retrieval request, which embeds the encrypted dynamic risk assessment vector as the query basis.

[0053] The cloud knowledge base receives the encrypted data packet and performs retrieval operations while maintaining data encryption or performing secure decryption according to agreed-upon strategies. The knowledge base internally stores a large number of historical cases, each containing the complete risk assessment record of a specific device within a specific time period and its context information. The retrieval algorithm analyzes the similarity of the query vector to the characteristics of the cases in the library, and the calculation process may involve encrypted domain similarity calculation or feature comparison after decryption. The algorithm outputs a number of historical cases with the highest matching degree to the query vector, forming an encrypted set of historical risk assessment cases. After encryption, the case set is returned to the local system through a secure channel.

[0054] In a secure environment, the unit performs decryption operations on the case set and the original dynamic risk assessment vector stored locally. After decryption, the unit performs multi-dimensional comparison analysis. The comparison content includes: the difference between the current predicted time-varying failure probability distribution and the actual failure probability distribution in the historical cases; the deviation between the current calculated key component residual life prediction value and the actual residual life record of the same component in the historical cases; the degree of agreement between the current generated risk propagation path set and the actual observed failure propagation path in the historical cases. These difference quantities are quantified as specific deviation parameters.

[0055] Based on the calculated set of deviation parameters, the local correction unit generates adjustment instructions for the weight parameters of the spatiotemporal coupling prediction network. The adjustment process uses an incremental learning strategy, modifying only the connection weights of specific layers in the network. For example, if the comparison finds that the actual occurrence probability of a certain type of bearing failure in historical cases is generally higher than the predicted value of the current model, the output weight of the corresponding bearing failure class in the network's failure probability prediction branch is increased. If the deviation analysis shows that the residual life prediction is systematically optimistic under high temperature conditions, the neuron weight that processes temperature-related features in the life decay calculation unit is adjusted. The weight correction amount is usually proportional to the size of the deviation parameter, determined by a pre-set mapping function.

[0056] The system continuously records the degree of conformity between the prediction performance of the corrected model and the subsequent actual device state. These records form feedback data for evaluating the actual effect of cloud-based collaborative optimization. When a new dynamic risk assessment vector is generated, the cloud-based collaborative module automatically starts a new round of encrypted upload, retrieval, and local correction process, forming a closed loop of continuous optimization.

[0057] The local system can choose to upload the current case's de-identified data to the cloud knowledge base after obtaining user authorization. The uploaded data is verified and integrated by the knowledge base, becoming a new historical case available for retrieval by other devices. This mechanism enables the cloud knowledge base to continuously accumulate diverse risk assessment experiences from different devices and different working conditions. After each weight correction of the spatiotemporal coupling prediction network, the system automatically saves a snapshot version of the network parameters. The snapshot is marked with the corresponding cloud retrieval timestamp and case identifier. The system retains the snapshot versions of the last several corrections. When the new prediction result shows abnormal fluctuations, the operator can roll back to the historical version of the network parameters to maintain system stability.

[0058] The operator can set the range limit of similar case retrieval, such as retrieving only cases of the same device type or limiting the similarity threshold of working condition parameters. The retrieval request can be attached with the feature encoding of the current operating environment of the device, and the cloud knowledge base prioritizes matching historical cases with similar environmental features. The retrieval result is returned with a matching score and a summary of the source device information.

[0059] The log records the trigger time of each correction, the details of the bias parameters involved, the adjusted weight position, and the correction value. These logs are used to analyze the long-term effects of cloud-based collaborative optimization and identify the relationship between specific bias patterns and optimal correction strategies. The analysis results are fed back to the correction algorithm to optimize the subsequent weight adjustment decision logic.

[0060] It should be noted that the relational terms herein such as first and second and the like are used solely to distinguish one entity or action from another, without necessarily requiring or implying any such actual relationship or order between such entities or actions. Moreover, the terms "comprises", "comprising", or any other variations thereof, are intended to cover a non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements does not include only those elements but can include other elements not expressly listed or inherent to such process, method, article, or apparatus.

[0061] Although embodiments of the present application have been shown and described, it is to be understood that various modifications, substitutions, replacements and changes can be made to these embodiments without departing from the principles and spirit of the present application, and the scope of the present application is defined by the appended claims and their equivalents.

Claims

1. A cross-dimensional turbomachinery dynamic risk prediction and visualization system, characterized by: include: A dynamic data acquisition module is used to acquire multi-source heterogeneous sensor data of the turbomachinery in real time, including vibration spectrum sequence, aerodynamic load distribution matrix, rotor axial displacement time series and surface temperature field thermal map; a service degradation assessment module, configured to extract a current vibration characteristic spectrum based on a historical degradation pattern library of the vibration spectrum sequence, calculate an aerodynamic instability risk level based on a deviation between the aerodynamic load distribution matrix and a preset operating condition threshold, and generate a service degradation trajectory matrix by integrating the sudden change gradient of the rotor axial displacement time series and the area of ​​the temperature difference abnormality region of the surface temperature field thermodynamic map; a multidimensional risk prediction module, configured to input the service degradation trajectory matrix into a spatiotemporal coupled prediction network and output a dynamic risk assessment vector comprising a time-varying failure probability distribution, a predicted value of the remaining life of key components, and a set of risk propagation paths; The risk visualization engine module is used to construct a three-dimensional risk thermal field model based on the dynamic risk assessment vector and generate a dynamic thermal grid layer superimposed on the digital twin of the impeller machinery.

2. The cross-dimensional turbomachinery dynamic risk prediction and visualization system according to claim 1, characterized in that: The dynamic data acquisition module includes: a vibration feature analysis unit, configured to perform wavelet packet energy spectrum decomposition on the vibration spectrum sequence, extract energy entropy values ​​within a preset frequency band interval, and mark abnormal frequency band positions where the energy entropy values ​​exceed an entropy value threshold; A load distribution verification unit is used to compare the aerodynamic load distribution matrix with a reference load distribution template under standard working conditions, and calculate the area ratio and maximum overload amplitude of the local load overload area; A displacement collaborative analysis unit, configured to correlate the rotor axial displacement timing with the speed pulse signal and identify the phase synchronization error between the displacement mutation point and the speed step point; The temperature field fusion unit is used to map the surface temperature field thermal map to the impeller machinery three-dimensional grid model and locate the spatial coordinate set of the temperature difference abnormal area.

3. The cross-dimensional turbomachinery dynamic risk prediction and visualization system according to claim 2, characterized in that: The service degradation assessment module includes: a degradation pattern matching unit, configured to retrieve a degradation pattern matching the abnormal frequency band position from the historical degradation pattern library and output a first risk level; an instability risk quantification unit, configured to calculate the aerodynamic instability risk level according to an excess amplitude gradient when the area ratio of the local load excess region exceeds an area threshold and the maximum excess amplitude exceeds an amplitude threshold; a displacement risk association unit, configured to generate an axial displacement risk coefficient according to the phase synchronization error and the amplitude increment of the displacement mutation point; A fusion calculation unit is used to weightedly fuse the first risk level, the aerodynamic instability risk level and the axial displacement risk coefficient to generate the service degradation trajectory matrix.

4. The cross-dimensional turbomachinery dynamic risk prediction and visualization system according to claim 3, characterized in that: The multidimensional risk prediction module includes: a spatiotemporal feature extraction unit, configured to perform a spatiotemporal convolution operation on the service degradation trajectory matrix to extract a cross-scale risk feature vector; a fault probability prediction unit, configured to process the cross-scale risk feature vector through a gated recurrent network and output the time-varying fault probability distribution within a future time window; A life decay calculation unit, configured to calculate a predicted value of the remaining life of the key component based on a mapping relationship between the cross-scale risk characteristic vector and a component material fatigue curve; The path deduction unit is used to simulate the risk diffusion process in the impeller machinery topology structure based on the graph neural network to generate the risk propagation path set.

5. The cross-dimensional turbomachinery dynamic risk prediction and visualization system according to claim 4, characterized in that: The risk visualization engine module includes: a spatial mapping unit, configured to map the time-varying fault probability distribution to corresponding grid cells of a three-dimensional grid model of the turbomachinery; a thermal field generation unit, configured to calculate a color gradient value of a grid cell according to the predicted remaining life of the key component, and generate the three-dimensional risk thermal field model in combination with the spatial coordinates of the risk propagation path set; A dynamic rendering unit is used to synchronously update the three-dimensional space risk thermal field model and the operating status of the impeller machinery and output the dynamic thermal grid layer.

6. The cross-dimensional turbomachinery dynamic risk prediction and visualization system according to claim 5, characterized in that: The system further comprises: An anomaly tracing module is used to extract all paths passing through the grid unit in the risk propagation path set and trace back to the source grid unit when the probability value of a specific grid unit in the time-varying fault probability distribution exceeds a probability threshold; The anomaly tracing module includes a path screening unit, which is used to calculate the propagation strength weight of each path and screen the paths whose weight exceeds the weight threshold to form the main propagation chain.

7. The cross-dimensional turbomachinery dynamic risk prediction and visualization system according to claim 6, characterized in that: The system further includes: an adaptive monitoring module for adjusting the collection frequency of corresponding sensor data according to the spatial position of the source grid unit; The adaptive monitoring module includes a frequency decision unit, which is used to calculate the incremental coefficient of the acquisition frequency based on the path length and propagation intensity weight of the main propagation chain.

8. The cross-dimensional turbomachinery dynamic risk prediction and visualization system according to claim 7, characterized in that: The system further comprises: a thermal grid optimization module for dynamically adjusting the grid density of the dynamic thermal grid layer; The thermal grid optimization module includes: Risk gradient analysis unit, used to calculate the risk probability gradient value between adjacent grid cells; a density controller unit, configured to insert a subdivided grid layer in a corresponding area when the risk probability gradient value exceeds a gradient threshold; The transition layer generation unit is used to configure a gradual transition mesh between the subdivided mesh layer and the original mesh layer.

9. The cross-dimensional turbomachinery dynamic risk prediction and visualization system according to claim 8, characterized in that: The system further comprises: a multidimensional data cleaning module for calculating the redundancy of historical data storage according to the update frequency of the dynamic risk assessment vector; The multidimensional data cleaning module includes: a redundancy evaluation unit, configured to count the number of historical versions of the service degradation trajectory matrix; The cleaning decision unit is used to delete the earliest version data in reverse order of storage time when the number of historical versions exceeds the version threshold.

10. The cross-dimensional turbomachinery dynamic risk prediction and visualization system according to claim 9, characterized in that: The system further includes: a cloud collaboration module for encrypting and uploading the dynamic risk assessment vector to a cloud knowledge base; The cloud collaboration module includes: The encrypted retrieval unit is used to initiate a similar case retrieval request to the cloud to obtain a matching historical risk assessment case set; The local correction unit is used to compare the deviation between the historical risk assessment case set and the current dynamic risk assessment vector, and correct the weight parameters of the spatiotemporal coupling prediction network.

Images