An intelligent operation monitoring and diagnosis platform for booster station fan

Abstract

The application relates to the technical field of computer data processing and artificial intelligence application, and discloses an intelligent operation monitoring and diagnosis platform for a booster station fan, which comprises a data preprocessing module; a dynamic health baseline construction module connected with the data preprocessing module and used for receiving a characteristic vector and a working condition data vector, and internally comprising a gating network and a group of conditional variational autoencoder (CVAE) expert models. In the application, a multi-dimensional health baseline which can adapt to working conditions is constructed, the logic of fault monitoring is innovated from judging whether a single parameter is out of limit to judging whether a multi-dimensional operation mode deviates from a proper health model, early degradation signs of equipment represented by weakly coordinated changes of multiple parameters can be sensitively captured, after the abnormality is captured, the platform can further automatically trace the root cause of destroying the health mode by introducing a causal reasoning mechanism, and the transformation from passive response alarm to active predictive diagnosis is realized.

Description

Technical Field

[0001] This invention relates to the fields of computer data processing and artificial intelligence application technology, and in particular to an intelligent operation monitoring and diagnostic platform for booster station fans. Background Technology

[0002] Substations are critical hubs in power systems. Their core equipment, the main transformer, generates a large amount of heat during operation, requiring large cooling fans and other auxiliary equipment for effective heat dissipation to ensure its safe and efficient operation. Therefore, the health and operational reliability of the cooling fans directly affect the service life of the main transformer and the stability of the entire power grid. Given the critical nature of this equipment, establishing an advanced and reliable operation monitoring and diagnostic system to accurately control the health status of the fans and perform predictive maintenance is of paramount importance for ensuring power system safety and reducing operation and maintenance costs.

[0003] Currently, the common technical solution for monitoring substation fans is to set fixed, empirical alarm thresholds for key components of the fan (such as bearings and motor windings) for physical quantities such as temperature and vibration. This method has certain drawbacks: First, its monitoring model is rigid and cannot adapt to the dynamic changes in normal operating parameters of the fan under different seasons, loads, and wear stages, leading to false alarms or missed alarms due to excessively high threshold settings in practical applications. Second, this method views the data of each measuring point in isolation, ignoring the fact that early faults often manifest as weak anomalies in the inherent correlation patterns between multiple parameters, making it difficult to capture the initial change signals of faults. Moreover, when an alarm occurs, the troubleshooting of the fault heavily relies on the personal experience of maintenance personnel, lacking objective and standardized diagnostic logic. Summary of the Invention

[0004] The technical problem this invention aims to solve is that existing monitoring technologies for booster station fans use fixed empirical alarm thresholds, which have drawbacks such as rigid models that are not adapted to dynamic changes and are prone to false alarms and missed alarms, difficulty in capturing early faults by viewing measurement point data in isolation, and reliance on personal experience for fault diagnosis and lack of standardized diagnostic logic. To address these shortcomings, we propose an intelligent operation monitoring and diagnosis platform for booster station fans.

[0005] To achieve the above objectives, this application adopts the following technical solution: an intelligent operation monitoring and diagnosis platform for booster station fans, comprising: a data preprocessing module, used to clean, normalize, and extract features from real-time acquired raw multidimensional sensor data and raw operating condition data, outputting high-dimensional feature vectors and operating condition data vectors; and a dynamic health baseline construction module, connected to the data preprocessing module, used to receive the feature vectors and operating condition data vectors, which internally includes a gating network and a set of conditional variational autoencoder (CVAE) expert models. The gating network calculates the weights of each expert model based on the operating condition data vectors, and each expert model reconstructs data based on the feature vectors and operating condition data vectors. The dynamic health baseline construction module performs weighted summation of the reconstruction errors of each expert model. The system generates a health index characterizing the health status of the wind turbine; an operational anomaly detection module, connected to the dynamic health baseline construction module, receives the time series of the health index and monitors the time series using statistical process control methods, triggering an early warning signal when a statistically significant anomaly is detected; a fault root cause diagnosis module, connected to the operational anomaly detection module, is activated after receiving an early warning signal, calculates the contribution of each input feature to the reconstruction error, and performs inference based on a preset fault cause-effect graph to determine and output the root cause of the fault; and a fault knowledge self-learning module, after the platform receives confirmed fault labels from external input, constructs a sample set and trains the diagnostic sub-network using a loss function to optimize the platform's ability to distinguish different fault categories.

[0006] Preferably, the health index is defined by the following formula: ,in Represents a health index. For feature vectors, For operating condition data vectors, The total number of expert models, The weights of the expert model For the first The total number of encoders in an expert model. For the first Decoder function for an expert model, This is the L2 norm squaring operation.

[0007] Preferably, the original operating condition data includes at least one or a combination of ambient temperature, ambient humidity, atmospheric pressure, motor active power, or the operating frequency of the variable frequency fan.

[0008] Preferably, the abnormal operation detection module uses an exponentially weighted moving average (EWMA) control chart for monitoring, and its EWMA statistic... The calculation method is defined by the following formula: ,in The current health index. This is the EWMA statistic from the previous time step. For the range of values ​​within The smoothing constant between them.

[0009] Preferably, in the fault root cause diagnosis module, the fault cause-effect graph is a directed acyclic graph constructed a priori based on the physical structure of the wind turbine equipment, the interaction relationship between components, and the fault propagation mechanism. Its nodes include physical components, sensor measurement points, and root cause fault types.

[0010] Preferably, the root cause diagnosis module uses a graph neural network to perform causal reasoning on the fault cause-effect graph to calculate the posterior probability of each root cause fault node.

[0011] Preferably, the sample set constructed by the fault knowledge self-learning module is a triplet sample, including anchor samples, positive samples, and negative samples. The loss function used is the triplet loss function, which is defined by the following formula:

[0012] ,in For the feature mapping function of the diagnostic subnetwork, For distance measurement function, These are the feature vectors of the anchor point, positive sample, and negative sample, respectively. These are the pre-defined positive definite constant boundaries.

[0013] Preferably, the diagnostic subnetwork is the encoder part of the CVAE expert model or a separate convolutional neural network.

[0014] Preferably, it also includes a remaining life prediction module, which is connected to the root cause diagnosis module. After the cause of the failure is confirmed, the module extracts the subsequent time series of the health index and uses a time series prediction model to predict the future degradation trend of the series in order to calculate the remaining life of the wind turbine components.

[0015] Preferably, it also includes a visualization and interaction module, which is used to display to users real-time curves of health indexes, EWMA control charts, diagnosed root causes of failures and their confidence levels, and predicted remaining service life in the form of a graphical interface.

[0016] The technical effects and advantages of this invention are as follows: By constructing a multi-dimensional health baseline that adapts to different operating conditions, the logic of fault monitoring is revolutionized from judging whether a single parameter exceeds its limit to judging whether a multi-dimensional operating mode deviates from its expected health paradigm. This allows for the sensitive capture of early signs of equipment degradation characterized by weak coordinated changes in multiple parameters, signs that would be completely ignored in traditional threshold systems. Once an anomaly is detected, the platform can further introduce a causal reasoning mechanism to automatically trace the root cause of the health disruption, achieving a shift from passive response alarms to proactive predictive diagnosis. Ultimately, through remaining life prediction and the self-learning capability of fault knowledge, this platform can help enterprises scientifically arrange maintenance before a failure occurs, greatly improving asset reliability and economic efficiency, and enabling the system to continuously optimize. Attached Figure Description

[0017] The disclosure of this invention is illustrated with reference to the accompanying drawings. It should be understood that the drawings are for illustrative purposes only and are not intended to limit the scope of protection of this invention. In the drawings, the same reference numerals are used to refer to the same parts:

[0018] Figure 1 This is a schematic diagram of the module connection structure of the present invention. Detailed Implementation

[0019] It is readily understood that, based on the technical solution of this invention, those skilled in the art can propose various interchangeable structural methods and implementations without altering the essential spirit of the invention. Therefore, the following detailed embodiments and accompanying drawings are merely illustrative examples of the technical solution of this invention and should not be considered as the entirety of the invention or as limitations or restrictions on the technical solution of this invention.

[0020] like Figure 1 As shown, an intelligent operation monitoring and diagnosis platform for booster station fans logically consists of a data preprocessing module, a dynamic health baseline construction module, an abnormal operation status detection module, a fault root cause diagnosis module, a remaining life prediction module, a fault knowledge self-learning module, and a visualization interaction module, which work sequentially and collaboratively.

[0021] The system architecture of this invention platform adopts a layered and progressive design concept, forming a complete technical closed loop from data acquisition to intelligent diagnosis and then to visualization. The various modules of the system are connected through standardized data interfaces, forming a collaborative organic whole.

[0022] The system architecture, from top to bottom, includes: a data acquisition layer responsible for real-time acquisition of raw data from field sensors and control systems; a data processing layer containing a data preprocessing module for cleaning, standardizing, and feature extraction of raw data; a core analysis layer comprising four core modules: dynamic health baseline construction, operational anomaly detection, fault root cause diagnosis, and fault knowledge self-learning; and an application service layer providing remaining life prediction and interactive visualization functions. This architecture design ensures the system is modular, scalable, and self-evolving, providing comprehensive intelligent monitoring and diagnostic capabilities for substation fans.

[0023] First, the data preprocessing module, as the entry point for the platform's data flow, processes the raw data collected in real time from various sensors and control systems on-site. The raw data includes two main categories: one is multi-dimensional sensor data directly reflecting the equipment's status, such as vibration, temperature, and noise; the other is raw operating condition data characterizing the equipment's operating conditions. This raw operating condition data specifically includes, but is not limited to, key parameters affecting the fan's operating status, such as ambient temperature, ambient humidity, atmospheric pressure, motor active power, and the operating frequency of the variable frequency fan. The processing steps of this module include: utilizing... The module uses criteria or isolated forest algorithms to filter out transient outliers and spikes in the data stream. Next, to eliminate differences in dimensions and numerical ranges between different physical quantities, Z-Score normalization is employed to convert all data into a standard normal distribution with a uniform scale. Finally, the module performs feature engineering, such as performing Fast Fourier Transform on vibration and acoustic time-series waveform data to extract characteristic frequency amplitudes related to specific faults in the frequency domain. It also calculates the mean, variance, kurtosis, and other time-domain statistical features of each signal. The output of this module is a high-dimensional feature vector containing rich information about the wind turbine's operating status. and the operating condition data vector formed after cleaning and standardization. These will be sent together to subsequent modules for in-depth analysis.

[0024] The benefit of implementing this module is that it transforms raw, noisy sensor signals and operating parameters into structured, more information-dense, and scale-uniform mathematical vectors, laying a solid data foundation for accurate pattern recognition by subsequent advanced models, and significantly improving the analytical performance and robustness of the entire platform.

[0025] Next, the dynamic health baseline construction module receives the preprocessed high-dimensional feature vector from the previous module. and operating condition data vector This module employs a hybrid expert architecture, consisting of a gating network and a set of parallel conditional variational autoencoder (CVAE) expert models. During operation, the working condition data vector... The input is given to a gating network, which dynamically assigns weights to each CVAE expert model in real time based on the characteristics of the current operating conditions. These weights reflect the applicability of each expert model to describing the health status under the current specific operating conditions. Meanwhile, the feature vectors... and operating condition data vector The encoder is fed into all CVAE expert models, each containing an encoder. and a decoder Its task is based on input. and Reconstruct a model that is as close as possible to The output vector Since the model is based on operating conditions, its reconstruction process is closely coupled with those conditions. Finally, the module uses the following formula to weight and sum the reconstruction errors of all expert models according to their respective weights, resulting in a single, comprehensive health index. : , in the formula The health index is a scalar value that quantifies the degree of deviation between the current device status and the health baseline. The larger the value, the worse the health status. The input is a multidimensional sensor feature vector with dimension . ,in The number of features; The input is a vector of real-time operating condition data with dimension . ,in The number of operating parameters; The total number of expert models is preset based on the complexity of the working conditions; For the index of expert models, ; The gating network determines the operating conditions. Calculated The weights of each expert model satisfy... ; For the first The encoder function of a CVAE expert model maps the input data to the latent space: ,in These are latent variables; For the first A decoder function for a CVAE expert model reconstructs data from the latent space: ,in The reconstructed feature vector; The square of the L2 norm (Euclidean distance) is used to calculate the difference between the original vector and the reconstructed vector; The value precisely quantifies the degree of deviation between the current actual operating mode of the wind turbine and its expected healthy mode under the current operating conditions.

[0026] The advantage of implementing this dynamic health baseline construction module is that it creates a health status reference system that can accurately sense and adapt to various operating scenarios. This enables the platform to effectively distinguish between parameter fluctuations caused by changes in normal operating conditions and real anomalies caused by early equipment degradation, thereby fundamentally solving the problems of high false alarm rate and low sensitivity of traditional fixed threshold methods under varying operating conditions.

[0027] Subsequently, the operational anomaly detection module continuously outputs data to the dynamic health baseline construction module. Time series monitoring is used to capture... For a small but persistent upward trend in the sequence, this module uses the Exponentially Weighted Moving Average (EWMA) control chart technique. The module calculates the EWMA statistic in real time using the following formula: , in the formula For at a certain point in time The EWMA statistic is used to monitor trends in health indices. For at a certain point in time Real-time health index value; For the previous time point The EWMA statistic; Smoothing constant This determines the degree of importance attached to current and historical data; smaller... The value assigns more weight to historical data and is more sensitive to slow trends.

[0028] Here It is an adjustable smoothing constant, and its value ranges from 1 to 10. Between these points, the sensitivity to historical data is adjusted, and the system will adjust based on the initial health phase. The sequence automatically calculates the center line and upper control limit of the control chart. During continuous monitoring, once... When the value exceeds the upper control limit, the module determines that the device status has deteriorated significantly in a statistical sense and immediately triggers an anomaly warning, transmitting the warning signal and related data to subsequent modules.

[0029] The advantage of implementing this operational anomaly detection module is that by capturing the dynamic trends of health index sequences rather than instantaneous values, the platform can issue early warnings at the very early stages of fault development, greatly extending the time window from when a fault can be detected to when it becomes a functional failure, and providing ample time for implementing planned predictive maintenance.

[0030] Once an early warning signal is received, the root cause diagnosis module is activated. Its task is to accurately deduce the underlying physical cause of the problem from complex anomalies. This module first traces back the data points that led to the early warning and then calculates... To quantify each input feature The contribution of the total reconstruction error to the formula is shown in the figure. For the first The contribution score of each input feature. For the first The original true values ​​of each feature, that is, the actual data collected from the sensors, preprocessed, and input into the dynamic health baseline construction module. For the first The reconstructed values ​​of each feature, To calculate the squared difference between the original value and the healthy template value, features with the highest contribution ranking are considered the most significant evidence of the anomaly. This evidence is then input into a pre-constructed causal graph of wind turbine failure, which incorporates expert knowledge. This causal graph is a directed acyclic graph constructed based on the physical structure of the wind turbine equipment, the interaction relationships between components, and the prior mechanism of fault propagation. The nodes in the graph include physical components, sensor measurement points, and root cause fault types. The built-in graph neural network algorithm performs inference on this graph, aggregating neighbor node information by learning the message passing function between nodes to simulate the propagation process of the fault impact in the causal graph, thereby updating the representation of each node. After inference, the graph neural network outputs the posterior probability of each potential root cause fault node in the graph. The platform outputs the node with the highest posterior probability as the final diagnostic conclusion.

[0031] The advantage of implementing this root cause diagnosis module is that it deeply integrates data-driven pattern recognition with expert knowledge based on physical mechanisms. This makes the diagnostic conclusions have clear physical meaning and interpretability, providing clear and accurate maintenance guidance for maintenance personnel, thereby avoiding blind maintenance and significantly improving maintenance efficiency and accuracy.

[0032] In parallel with the root cause diagnosis module, the remaining life prediction module is activated after the cause of the failure is identified. This module aims to predict the future health trend of the damaged component, and extracts the health index. The subsequent time series reflects the degradation process of this component. The module employs time series prediction models, such as autoregressive integral moving average models or long short-term memory networks, to predict the degradation process of this component. The system fits and extrapolates the future development trend of the sequence to predict it. Simultaneously, the system presets a failure threshold based on historical data or industry standards. The module calculates and predicts... The remaining service life of the wind turbine component is determined by the time required for the curve to reach the failure threshold.

[0033] The advantage of the remaining life prediction module is that it elevates the diagnostic conclusion from a static fault identification level to a dynamic life prediction level, providing direct and quantitative decision-making basis for spare parts procurement, maintenance resource allocation, and production planning. It is a key link in realizing predictive maintenance.

[0034] Finally, the fault knowledge self-learning module endows the platform with the function of continuous learning and self-improvement. When a fault is finally confirmed by on-site operation and maintenance, the confirmed fault tag is fed back into the platform. After receiving this tag, this module automatically uses the characteristic data at the time of the fault as anchor samples. And retrieve similar fault data from the historical database as positive samples. And all other types of fault or health data as negative samples This constructs one or more training triples. Subsequently, the fault knowledge self-learning module uses the triple loss function to fine-tune a dedicated diagnostic sub-network online. This diagnostic sub-network can be the encoder part of the CVAE expert model or an independent convolutional neural network. The goal of this training process is to aggregate all sample points of the same type of fault in the feature space defined by the network, while separating them from other types of sample points. The triple loss function is defined as follows: , in the formula The calculated triplet loss value is used to guide the update of network parameters, with the goal of minimizing this loss; The feature mapping function for the diagnostic subnetwork maps the input data to a high-dimensional feature space; The distance metric function in the feature space is Euclidean distance. , , These are the anchor points, positive and negative sample data after... The feature vector obtained after function mapping; The feature vector of the anchor point sample comes from the currently confirmed fault data; The feature vectors of positive samples are derived from historical fault data of the same type as the anchor points. The feature vectors of negative samples come from other fault or health status data of different classes than the anchor point; A pre-defined positive definite constant boundary is used to define the minimum distance that should exist between samples of different classes, ensuring that the distance between pairs of samples of different classes is at least greater than the distance between pairs of samples of the same class. .

[0035] The fault knowledge self-learning module establishes a feedback loop from practice to model optimization. Every manually confirmed fault diagnosis is transformed into a supervised learning opportunity, enabling the platform to continuously improve its accuracy in identifying known faults and gradually accumulate the ability to deal with newly emerging fault modes.

[0036] In addition, the platform includes a visualization and interaction module, which serves as the human-computer interface between the platform and operations personnel. This module retrieves key results and data from all the aforementioned modules, and its core function is to transform this data into intuitive graphs, charts, and diagnostic reports, integrating them into a unified graphical user interface for user display. Specific content displayed includes: health index. It displays real-time and historical change curves; EWMA control charts and anomaly warning markers; the most likely root causes of failures, affected components and their confidence levels; and the remaining service life prediction results of key components. It transforms the complex algorithm calculation results within the platform into visual information that is easy to understand and operate, significantly reducing the threshold for using the system and making the operation and maintenance decision-making process more efficient and reliable.

[0037] To clearly demonstrate the technical effects of the present invention, the following two specific embodiments are provided for illustration:

[0038] Example 1: The online monitoring system of the cooling fan of the main transformer in a certain substation showed that the equipment was operating normally, and the instantaneous readings of all key parameters (such as overall vibration value and bearing temperature) did not exceed the set fixed alarm threshold. However, the intelligent operation monitoring and diagnosis platform of this invention generated a continuously rising health index through its dynamic health baseline construction module. This triggered an internal system alert.

[0039] The diagnostic process and results of the platform of this invention: After confirming the abnormal trend of the health index, the abnormal operation status detection module activated the root cause diagnosis module. This module deeply analyzed the operation data and found that although the absolute value of the parameter did not exceed the limit, the synergistic relationship between the vibration energy of the bearing in a specific frequency band and the bearing temperature had significantly deviated from the healthy mode. Contribution analysis confirmed that these two features were the key evidence that contributed the most to this abnormality. Subsequently, the platform matched and reasoned with the built-in fault causal spectrum based on physical mechanism. Graph neural network calculation showed that the root cause node of bearing inner ring wear could better explain the joint phenomenon of specific vibration intensification and synchronous temperature rise. Finally, the platform output a clear diagnostic conclusion: there is inner ring wear in the drive end bearing, with a confidence level of 91%.

[0040] Based on this diagnostic conclusion, the maintenance personnel conducted a focused inspection of the bearing during the planned shutdown for maintenance. After disassembly, it was confirmed that the inner ring of the bearing did indeed have slight scratches and early pitting corrosion. This embodiment shows that the present invention achieves early warning and accurate diagnosis in the early stage of a fault, when traditional threshold alarm systems remain silent, by capturing weak anomalies in the synergistic relationship of multiple parameters. This fundamentally solves the problem of missed alarms caused by insufficient sensitivity and isolated judgment in traditional methods.

[0041] Example 2: The traditional vibration monitoring system of a certain wind turbine issued an alarm for excessive vibration, but the alarm information could not indicate the root cause of the fault, and the maintenance personnel could not immediately determine whether it was due to poor bearing lubrication, blade adhesion, or other mechanical failure.

[0042] The diagnostic process and results of the platform of this invention: When analyzing the wind turbine data, the dynamic health baseline construction module of this platform also generated a significantly increased health index. After the fault root cause diagnosis module was activated, it conducted in-depth analysis of various characteristics. The analysis found that the increase in vibration energy was mainly concentrated in the wind turbine frequency and its harmonics, while the high-frequency component energy characterizing bearing faults was normal. At the same time, the bearing temperature only increased slightly. The platform identified the increased frequency vibration as the main abnormal evidence and the slight increase in bearing temperature as secondary evidence.

[0043] Based on this evidence package, the platform performs intelligent reasoning in the fault cause-effect diagram. The reasoning process shows that if the problem is due to poor bearing lubrication, the temperature rise should be more significant and the vibration characteristics should match the bearing failure frequency. If the problem is due to imbalance caused by blade attachments, the core feature should be increased rotational frequency vibration. The current evidence has the highest degree of matching with the fault mechanism of foreign matter attached to the blades. Therefore, the platform outputs a clear diagnostic conclusion: the cooling fan blades are caused by dynamic imbalance, with a confidence level of 88%.

[0044] Based on this diagnosis, the maintenance personnel prioritized the inspection and cleaning of the wind turbine blades. The results showed that a large amount of dirt was attached to the surface of the blades. After cleaning, the vibration value of the wind turbine returned to normal. This embodiment shows that the present invention can effectively distinguish between fault types with similar symptoms but different root causes. Compared with traditional methods that can only provide vague alarms, the present invention can provide diagnostic results with clear direction, which greatly improves maintenance efficiency and avoids the time and economic costs caused by misjudgment.

[0045] The technical scope of this invention is not limited to the content described above. Those skilled in the art can make various modifications and variations to the above embodiments without departing from the technical concept of this invention, and all such modifications and variations should fall within the protection scope of this invention.

Claims

1. An intelligent operation monitoring and diagnosis platform for booster station fans, characterized in that, Comprise: A data preprocessing module for cleaning, normalizing and feature extraction processing of raw multi-dimensional sensor data and raw working condition data collected in real time, outputting high-dimensional feature vectors and working condition data vectors; a dynamic health baseline construction module connected with the data preprocessing module for receiving the feature vectors and working condition data vectors, which internally comprises a gating network and a set of conditional variational autoencoder (CVAE) expert models, the gating network calculates the weight of each expert model according to the working condition data vector, and each expert model reconstructs data according to the feature vector and the working condition data vector, the dynamic health baseline construction module generates a health index representing the health status of the fan by weighted summing the reconstruction errors of each expert model; An operating state anomaly detection module connected with the dynamic health baseline construction module for receiving the time series of the health index and monitoring the time series through statistical process control method, and triggering an early warning signal when a statistically significant anomaly is detected; A fault root cause diagnosis module connected with the operating state anomaly detection module, activated after receiving the early warning signal, for calculating the contribution of each input feature to the reconstruction error and reasoning with a pre-set fault causal graph to determine and output the root cause of the fault; A fault knowledge self-learning module for training the diagnosis sub-network by constructing a sample set and using a loss function to optimize the platform's ability to distinguish different fault categories after the platform receives an externally input confirmed fault label; the sample set constructed by the fault knowledge self-learning module is a three-tuple sample, including an anchor sample, a positive sample and a negative sample, and the loss function used is a three-tuple loss function, the implementation of the three-tuple loss function is defined by the following formula: wherein is a feature mapping function of the diagnosis sub-network, is a distance metric function, are feature vectors of anchor, positive, and negative samples, respectively, is a pre-set positive constant boundary. 2.The intelligent operation monitoring and diagnosis platform for booster station fans of claim 1, wherein: The implementation of the health index is defined by the following formula: wherein represents the health index, is a feature vector, is a working condition data vector, is the total number of expert models, is the weight of the expert model, is the total number of encoders of the th expert model, is the decoder function of the th expert model, is an L2 norm square operation. 3.The intelligent operation monitoring and diagnosis platform for booster station fans of claim 1, wherein: The raw working condition data includes at least one of environmental temperature, environmental humidity, atmospheric pressure, motor active power or variable frequency fan operating frequency, or a combination thereof.

4. The intelligent operation monitoring and diagnosis platform for booster station fans of claim 1, wherein: The operation state abnormality detection module monitors by using an exponential weighted moving average (EWMA) control chart, and an EWMA statistic quantity of the EWMA control chart is calculated by the following formula: The calculation manner of the EWMA statistic quantity is defined by the following formula: Wherein, is a health index at a current moment, is an EWMA statistic quantity at a previous moment, is a smoothing constant with a value range of between 0 and 1.

5. The intelligent operation monitoring and diagnosis platform for booster station fans of claim 1, wherein: In the fault root cause diagnosis module, the fault causal graph is a directed acyclic graph constructed a priori according to the physical structure of the fan equipment, the interaction relationship between components and the fault propagation mechanism, the nodes of which include physical components, sensor measurement points and root cause fault types.

6. The intelligent operation monitoring and diagnosis platform for booster station fans according to claim 5, characterized in that: The fault root cause diagnosis module uses a graph neural network to perform causal reasoning on the fault causal graph to calculate the posterior probability of each root cause fault node.

7. The intelligent operation monitoring and diagnosis platform for booster station fans of claim 1, wherein: The diagnosis sub-network is the encoder part of the CVAE expert model or an independent convolutional neural network.

8. The intelligent operation monitoring and diagnosis platform for booster station fans of claim 1, wherein: It also includes a remaining life prediction module connected with the fault root cause diagnosis module, which extracts the subsequent time series of the health index after the fault cause is confirmed, and uses a time series prediction model to predict the future degradation trend of the sequence to calculate the remaining useful life of the fan components.

9. The intelligent operation monitoring and diagnosis platform for booster station fans of claim 1, wherein: It also includes a visual interactive module for displaying the real-time curve of the health index, the EWMA control chart, the diagnosed fault root cause and its confidence, and the predicted remaining useful life to the user in the form of a graphical interface.

Images