Cloud CMS diagnosis platform based on edge calculation and AI model

Abstract

The invention relates to the technical field of wind power state monitoring, and discloses a cloud CMS diagnosis platform based on edge calculation and an AI model. The system comprises an edge computing system deployed at a wind power plant side and a centralized diagnosis and analysis system at a cloud end. The edge computing system is responsible for collecting and processing data in real time, operating various lightweight early warning models and generating a preliminary early warning event; the cloud system comprises an early warning aggregation module, a diagnosis credibility evaluation module and a cloud intelligent diagnosis module; the early warning aggregation module aggregates the early warning events of the same equipment into an evidence set; the diagnosis credibility evaluation module generates a comprehensive diagnosis confidence coefficient by quantitatively analyzing evidence consistency, model credibility and data quality; the cloud intelligent diagnosis module performs deep analysis based on the evidence set and outputs a final diagnosis report with the confidence coefficient; according to the method, traditional discrete and shallow alarm is upgraded to diagnosis with quantitative credibility evaluation, and the intelligent level and decision-making efficiency of wind power operation and maintenance are improved.

Description

Technical Field

[0001] This invention relates to the field of wind power condition monitoring technology, and in particular to a cloud-based CMS diagnostic platform based on edge computing and AI models. Background Technology

[0002] As an important component of clean energy, the health status of wind power equipment directly affects power generation efficiency and operating costs. Condition monitoring systems (CMS) are a key means to ensure the safe and stable operation of wind turbine units. Currently, the CMS platforms widely deployed in the industry are mainly based on sensor data such as vibration and temperature to achieve real-time monitoring and alarm of equipment status.

[0003] However, most existing CMS platforms are limited to basic alarm functions; their core alarm logic usually relies on preset, relatively simple rule models, mainly including: alarms based on static or dynamic thresholds, alarms based on short-term trend increases, and alarms based on long-term trend increases. These methods are essentially anomaly detection for single or limited-dimensional data indicators, and their output results are only whether an alarm is triggered and a simple alarm level.

[0004] This basic alarm mode has significant limitations: it can only indicate that an anomaly may have occurred somewhere, but it cannot determine the core diagnostic issues such as what the fault is, where the root cause is, how serious it is, and what maintenance measures should be taken. After receiving a large number of such discrete and superficial alarm messages, maintenance personnel still need to rely on their personal experience to manually screen, correlate and diagnose the fault. This not only results in slow response speed and high maintenance costs, but also makes it difficult to guarantee the accuracy and consistency of the diagnostic results, which can easily lead to misjudgment, missed diagnosis, or over-maintenance.

[0005] Therefore, a key technical problem that urgently needs to be solved in the current wind power operation and maintenance field is: how to upgrade the existing CMS system, which can only provide fragmented primary alarm information, into an intelligent diagnostic decision platform that can automatically perform in-depth fault analysis, locate the root cause, and output quantitative and reliable assessments. Summary of the Invention

[0006] The technical problem to be solved by this invention is that the existing status monitoring system only has the simple threshold alarm function. To address this, we propose a cloud-based CMS diagnostic platform based on edge computing and AI models.

[0007] To achieve the above objectives, this application adopts the following technical solution: a cloud-based CMS diagnostic platform based on edge computing and AI models, including an edge computing system deployed at the wind farm and a centralized diagnostic and analysis system deployed in the cloud; the edge computing system is used to collect monitoring data of wind turbine units in real time, perform data processing and feature extraction, and generate preliminary early warning event streams based on multiple early warning models; the cloud-based centralized diagnostic and analysis system is used to aggregate and deeply analyze the preliminary early warning event streams from various wind farms, specifically including: an early warning aggregation module, used to aggregate data received within a preset sliding time window for each wind turbine unit as an aggregation unit, targeting that wind turbine unit... All preliminary warning events of the unit are aggregated into a warning evidence set with a unified context identifier; the diagnostic confidence assessment module is used to quantitatively evaluate the warning evidence set; this module is configured to: for the warning evidence set originating from the same wind turbine unit and aggregated within a continuous time window, generate a comprehensive diagnostic confidence coefficient by comprehensively quantifying the consistency of warning events within the set, the historical reliability of the warning models corresponding to each warning event, and the real-time quality of the data on which each warning event is based; the cloud-based intelligent diagnostic module performs fault diagnosis analysis based on the warning evidence set and outputs a final diagnostic report with the comprehensive diagnostic confidence coefficient.

[0008] Preferably, the edge computing system generates preliminary warning events in the following manner: Vibration and temperature waveforms, as well as speed and power operating condition signals, are synchronously acquired from the wind turbine's condition monitoring sensors via standard industrial protocols; the raw waveforms undergo preprocessing including filtering, noise reduction, and operating condition slicing, and preset time-domain statistical features and frequency-domain spectral features are extracted; the extracted feature data is input into a lightweight warning model library deployed at the edge, which runs multiple independent AI warning models in parallel for real-time analysis; the outputs of each warning model are integrated to generate a structured preliminary warning event, which includes at least a warning model type identifier, a normalized anomaly intensity value, a precise microsecond-level trigger timestamp, and a measurement point code identifying the specific sensor location.

[0009] Preferably, the lightweight early warning model running in the edge computing system includes at least three types with different analysis dimensions: a threshold early warning model for determining whether vibration characteristic values ​​exceed preset static or dynamic thresholds; an early warning model for determining whether the growth rate of vibration characteristic values ​​in adjacent analysis periods exceeds the short-term trend growth rate; and an early warning model for determining whether the abnormal acceleration trend of the growth rate of vibration characteristic values ​​in long-term periods exceeds the long-term trend growth rate. The normalized intensity value in the preliminary early warning event is the result of mapping the original output intensity values ​​with different dimensions from different types of early warning models to a unified, dimensionless numerical range through a preset normalization function.

[0010] Preferably, the early warning aggregation module is specifically configured as follows: taking a single wind turbine as the basic aggregation unit, a sliding time window mechanism is used to listen to and cache the continuously uploaded preliminary early warning event stream in real time; for the cached events, correlation analysis is performed based on their turbine identifier and timestamp, and all events falling within the same sliding window are dynamically aggregated into a logically unified early warning evidence set; each generated early warning evidence set is assigned a globally unique tracking identifier, along with the window start and end times, the list of included events, and the original data index, forming a complete diagnostic context.

[0011] Preferably, the diagnostic confidence assessment module includes: a consistency quantification unit, used to calculate the internal synergy of the early warning evidence set, which is achieved by fusing the intensity value dispersion and trigger time dispersion of all early warning events within the set; a multi-source information fusion unit, used to acquire and integrate the historical accuracy reputation score of the early warning model associated with each early warning event in the early warning evidence set, as well as the real-time quality score of the original data on which the event was based; and a confidence synthesis unit, used to calculate and generate a comprehensive diagnostic confidence coefficient based on the consistency quantification result, the multi-source information fusion result, and the quantity information of early warning events using a predefined algorithm.

[0012] Preferably, the historical accuracy reputation score of the early warning model is dynamically maintained and obtained through the following methods: The platform maintains a cloud-based model performance knowledge base, continuously recording the historical trigger records of each early warning model and its corresponding subsequent on-site verification results; when the diagnostic report output by the platform is verified by on-site operation and maintenance, the relevant early warning events and their triggering models are traced back according to the verification results, and the historical accuracy reputation score of the corresponding model in the model performance knowledge base is dynamically adjusted according to the preset reputation update algorithm; when needed, the diagnostic credibility assessment module queries the model performance knowledge base in real time to obtain the latest reputation score of the corresponding model.

[0013] Preferably, the consistency quantification unit calculates the internal synergy of the early warning evidence set through the following steps: S1. Intensity dispersion calculation: calculate the standard deviation of the normalized intensity value of each early warning event in the set relative to its arithmetic mean; S2. Time dispersion calculation: calculate the standard deviation of the trigger timestamps of each early warning event in the set, and divide it by the average of the differences between each pair of these timestamps to obtain the time dispersion factor; S3. Synergy synthesis: combine the intensity dispersion and the time dispersion factor by weighting or multiplying them to form a diagnostic intention dispersion index. The higher the index value, the worse the consistency of events within the set.

[0014] Preferably, the real-time data quality score used in the multi-source information fusion unit is calculated synchronously by the edge computing system when the early warning event is generated and encapsulated in the event data packet; the calculation basis of the real-time quality score includes: the signal-to-noise ratio of the original vibration signal used to generate the early warning in the effective frequency band, the sampling integrity index of the waveform data, and the operating condition stability score of the wind turbine power and speed during the signal acquisition period.

[0015] Preferably, the diagnostic intention dispersion index in step S3 is calculated using the following formula: ;in, Indicator of diagnostic intention dispersion; The total number of warning events in the warning evidence set; For the first Normalized intensity values ​​of each warning event; This is the average of the normalized intensity values ​​for all warning events. The standard deviation of the timestamps for all warning events; It is the average of the absolute values ​​of the differences between all pairwise trigger timestamps. and These are extremely small constants, used to ensure that the strength term is never zero and to avoid the denominator being zero, respectively.

[0016] Preferably, the confidence level synthesis unit calculates the comprehensive diagnostic confidence coefficient using the following formula: ;in, This represents the overall diagnostic confidence coefficient; Indicates the first The historical accuracy and reputation score of the early warning model corresponding to each early warning event; Indicates the generation of the first The real-time quality score of the data used in each early warning event; This is a preset baseline constant used to adjust the gain strength of the confidence level based on the amount of evidence. This represents the natural logarithm function.

[0017] The technical effects and advantages of this invention are as follows: This invention constructs an edge-cloud diagnostic architecture, completely changing the limitation of traditional CMS systems that can only provide basic alarms. At the edge, the system runs multiple lightweight early warning models in parallel, generating multi-dimensional preliminary evidence. In the cloud, a unified set of early warning evidence is formed through an early warning aggregation module, and a diagnostic credibility assessment module is introduced. This module generates a comprehensive diagnostic confidence coefficient by quantitatively analyzing evidence consistency, model historical reputation, and data quality. Finally, the cloud-based intelligent diagnostic module outputs an enhanced diagnostic report with this quantitative coefficient. Thus, the platform achieves a leap from fragmented alarms to deep and reliable diagnosis, upgrading the operation and maintenance mode from a passive response relying on human experience to proactive decision-making based on quantitative confidence assessment, significantly improving the accuracy, objectivity, and operational efficiency of fault diagnosis. Attached Figure Description

[0018] 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:

[0019] Figure 1 This is a data flow diagram of the present invention. Detailed Implementation

[0020] 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.

[0021] Reference Figure 1 A cloud-based CMS diagnostic platform based on edge computing and AI models, which includes an edge computing system deployed on the wind farm side and a centralized diagnostic and analysis system deployed in the cloud.

[0022] The edge computing system is deployed on the wind farm side, close to the wind turbine. Its main function is to realize local real-time data processing and preliminary early warning. Specifically, the edge computing system synchronously collects vibration, temperature raw waveforms, and operating condition signals such as speed and power from the condition monitoring sensors of the wind turbine through standard industrial protocols. These sensors typically include accelerometers, temperature sensors, etc., and are used to monitor the operating status of key components of the wind turbine.

[0023] The acquired raw waveform data is first preprocessed, including but not limited to: filtering to remove high-frequency noise or power frequency interference; noise reduction, such as wavelet noise reduction and empirical mode decomposition noise reduction; and operating condition slicing. Operating condition slicing refers to dividing the raw waveform data into data segments under stable operating conditions based on the operating condition signals such as the wind turbine's speed and power, in order to eliminate the impact of operating condition changes on feature extraction.

[0024] The preprocessed data is used to extract preset time-domain statistical features, such as root mean square value, peak value, kurtosis, margin, and waveform factor, and frequency-domain spectral features, such as envelope demodulation spectrum, power spectral density, and cepstrum. These features are key indicators reflecting the operating status of the equipment. The extracted feature data is input into a lightweight early warning model library deployed at the edge. This model library runs multiple independent AI early warning models in parallel for real-time analysis. These lightweight early warning models include at least three types with different analytical dimensions:

[0025] Threshold warning model: Used to determine whether vibration characteristic values ​​exceed preset static or dynamic thresholds. Static thresholds are fixed values ​​set based on equipment design parameters or historical experience; dynamic thresholds can be adaptively adjusted based on equipment operating conditions, ambient temperature, and other factors. Short-term trend increase warning model: Used to determine whether the growth rate of vibration characteristic values ​​exceeds a reasonable range within adjacent analysis periods; for example, monitoring the growth rate of a characteristic value over the past 24 hours, triggering a warning if it exceeds a certain percentage. Long-term trend increase warning model: Used to determine whether the growth rate of vibration characteristic values ​​shows an abnormal acceleration trend over a longer period, such as several weeks or months; this is usually determined through regression analysis or curve fitting of the historical trend of the characteristic value.

[0026] The outputs of each early warning model are integrated to generate structured preliminary early warning events. Each preliminary early warning event includes at least an early warning model type identifier, such as "threshold model - RMS exceeding limit" or "short-term trend - kurtosis anomalous growth", a normalized anomalous intensity value, a precise microsecond-level trigger timestamp, and a measurement point code identifying the specific sensor location. The normalized intensity value is the result of mapping the original output intensity values ​​from different types of early warning models, which have different dimensions, to a unified, dimensionless numerical range through a preset normalization function, so as to facilitate subsequent fusion and comparison.

[0027] The cloud-based centralized diagnostic and analysis system is deployed in the cloud and is responsible for receiving preliminary early warning event streams uploaded from the edge computing system, and performing deeper aggregation, diagnosis, and credibility assessment. The system includes an early warning aggregation module, a diagnostic credibility assessment module, and a cloud-based intelligent diagnostic module.

[0028] The early warning aggregation module is used to aggregate multiple preliminary early warning events from the same wind turbine within a preset time window into a single early warning evidence set. Specifically, this module uses a single wind turbine as the basic aggregation unit and employs a sliding time window mechanism to monitor and cache the continuously uploaded preliminary early warning event stream in real time. For example, a 1-hour sliding window can be set, sliding every 5 minutes. For cached events, correlation analysis is performed based on their turbine identifier and timestamp, dynamically aggregating all events falling within the same sliding window into a logically unified early warning evidence set. For example, if multiple measuring points of a wind turbine's gearbox simultaneously trigger vibration threshold early warning and short-term trend early warning within 1 hour, these events will be aggregated into the same evidence set. Each generated early warning evidence set is assigned a globally unique tracking identifier and includes the window start and end times, a list of included events, and an original data index, forming a complete diagnostic context for subsequent modules to process and trace.

[0029] The diagnostic credibility assessment module is used to quantitatively evaluate the early warning evidence set and generate a comprehensive diagnostic confidence coefficient. This module achieves this by integrating the consistency of early warning events within the set, the historical performance reputation of the early warning models corresponding to each early warning event, and the real-time quality of the data on which each early warning event is based. Specifically, the diagnostic credibility assessment module includes: a consistency quantification unit, used to calculate the internal coherence of the early warning evidence set. This is achieved by integrating the intensity value dispersion and trigger time dispersion of all early warning events within the set; intensity dispersion calculation: calculating the standard deviation of the normalized intensity value of each early warning event within the set relative to its arithmetic mean; the larger the standard deviation, the more dispersed the event intensity values ​​and the worse the consistency; time dispersion calculation: calculating the standard deviation of the trigger timestamps of each early warning event within the set, and dividing it by the average of the differences between each pair of these timestamps to obtain the time dispersion factor. A larger time dispersion factor indicates more dispersed event trigger times and poorer consistency. Coordination synthesis: The intensity dispersion and the time dispersion factor are combined through weighted summation or multiplication to form a diagnostic intention dispersion index. A higher index value indicates poorer consistency among events within the set. Specifically, the diagnostic intention dispersion index is calculated using the following formula: ;in, Indicator of diagnostic intention dispersion; The total number of warning events in the warning evidence set; For the first Normalized intensity values ​​of each warning event; This is the average of the normalized intensity values ​​for all warning events. The standard deviation of the timestamps for all warning events; It is the average of the absolute values ​​of the differences between all pairs of trigger timestamps; and These are extremely small constants, used to ensure that the strength term is never zero and to avoid the denominator being zero, respectively.

[0030] Multi-source information fusion unit: used to acquire and integrate the historical accuracy and reputation score of the early warning model associated with each early warning event in the early warning evidence set, as well as the real-time quality score of the original data on which the event was based.

[0031] The historical accuracy reputation score of the early warning model is dynamically maintained and obtained through the following methods: The platform maintains a cloud-based model performance knowledge base, continuously recording the historical trigger records of each early warning model and its corresponding subsequent on-site verification results; when the diagnostic report output by the platform is verified by on-site operation and maintenance, based on the confirmation or rejection of the verification results, it traces back to the relevant early warning events and their triggering models, and dynamically adjusts the historical accuracy reputation score of the corresponding model in the model performance knowledge base according to the preset reputation update algorithm, such as Bayesian update or moving average algorithm; when needed, the diagnostic credibility assessment module queries the model performance knowledge base in real time to obtain the latest reputation score of the corresponding model.

[0032] The real-time data quality score is calculated synchronously by the edge computing system when an early warning event is generated and encapsulated in the event data packet. The calculation basis of the real-time quality score includes: the signal-to-noise ratio of the original vibration signal used to generate the early warning in the effective frequency band, the sampling integrity index of the waveform data, such as whether there is packet loss or data missing; and the operating condition stability score of the wind turbine power and speed during the signal acquisition period, such as whether the operating condition fluctuates drastically. These indicators comprehensively reflect the reliability of the original data.

[0033] Confidence Synthesis Unit: Used to calculate and generate a comprehensive diagnostic confidence coefficient based on consistency quantification results, multi-source information fusion results such as model reputation score and data quality score, and the quantity information of warning events, using a predefined algorithm; specifically, the comprehensive diagnostic confidence coefficient is calculated using the following formula: ;in, This represents the overall diagnostic confidence coefficient; Indicates the first The historical accuracy and reputation score of the early warning model corresponding to each early warning event; Indicates the generation of the first The real-time quality score of the data used in each early warning event; This is a preset baseline constant used to adjust the gain strength of the confidence level based on the amount of evidence. The formula represents the natural logarithm function; it comprehensively considers the consistency within the warning event, the model reputation and data quality of each event, and the amount of evidence, so that the CDC value can fully reflect the credibility of the diagnostic conclusion.

[0034] The cloud-based intelligent diagnostic module is used to perform fault diagnosis analysis based on the early warning evidence set and output a final diagnostic report with a comprehensive diagnostic confidence coefficient. This module can integrate more complex AI diagnostic models to perform in-depth analysis on the aggregated evidence set and identify the specific fault type, fault location, and fault severity. The final diagnostic report not only includes fault information but also includes a comprehensive diagnostic confidence coefficient generated by the diagnostic credibility assessment module, providing quantitative decision-making basis for operation and maintenance personnel.

[0035] The overall workflow of the platform of this invention is as follows: Data acquisition and preprocessing (edge ​​side): The edge computing system acquires wind turbine sensor data in real time and performs preprocessing such as filtering, noise reduction, and operating condition slicing.

[0036] Feature extraction and preliminary early warning (edge ​​side): Extract time-domain and frequency-domain features from preprocessed data and input them into a lightweight early warning model library; Multiple parallel AI early warning models, such as threshold early warning, short-term trend increase early warning, and long-term trend growth early warning, generate structured preliminary early warning events based on feature values ​​and calculate real-time data quality scores.

[0037] Event Upload and Aggregation (Cloud): The initial warning event stream is uploaded to the cloud-based centralized diagnosis and analysis system; the warning aggregation module uses wind turbine units as units and adopts a sliding time window mechanism to aggregate multiple initial warning events of the same unit within a preset time window into a warning evidence set and assigns a unique tracking identifier.

[0038] Diagnostic Confidence Assessment (Cloud): The diagnostic confidence assessment module receives the set of early warning evidence, calculates the internal synergy degree (DID) through the consistency quantification unit, and obtains the historical reputation score and real-time data quality score of each early warning model through the multi-source information fusion unit; the confidence synthesis unit integrates this information to calculate the comprehensive diagnostic confidence coefficient (CDC).

[0039] Intelligent Diagnosis and Report Generation (Cloud): The cloud-based intelligent diagnosis module performs in-depth fault diagnosis analysis based on the early warning evidence set and generates the final diagnostic report in conjunction with CDC.

[0040] Feedback and Optimization (Cloud): Operation and maintenance personnel conduct on-site verification based on the diagnostic report; the verification results are fed back to the model performance knowledge base, which is used to dynamically adjust the historical accuracy and reputation score of relevant early warning models, thereby enabling continuous learning and optimization of the system.

[0041] This invention achieves the following advantages through the collaborative work of edge computing and centralized cloud-based diagnostics: Real-time performance and efficiency: Real-time early warnings at the edge significantly shorten fault detection time and reduce data transmission pressure. Accuracy and reliability: Multi-source information fusion and a reliable quantification evaluation mechanism effectively improve the accuracy and reliability of diagnostic results, reducing false alarms and missed alarms. Interpretability and decision support: The comprehensive diagnostic confidence coefficient provides maintenance personnel with quantitative decision-making basis, enabling them to take more confident actions. Adaptability and intelligence: The dynamic maintenance mechanism of model reputation allows the system to self-optimize based on actual operating results, continuously improving diagnostic performance. Scalability: The modular design makes the system easy to expand and facilitates the integration of new early warning models and diagnostic algorithms.

[0042] 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. A cloud-based CMS diagnostic platform based on edge computing and AI models, characterized in that: It includes an edge computing system deployed at the wind farm and a centralized diagnostic and analysis system deployed in the cloud; the edge computing system is used to collect monitoring data of wind turbine units in real time, perform data processing and feature extraction, and generate preliminary early warning event streams based on multiple early warning models; The cloud-based centralized diagnosis and analysis system is used to aggregate and deeply analyze the preliminary early warning event streams from various wind farms. Specifically, the system includes: an early warning aggregation module, which aggregates all preliminary early warning events received within a preset sliding time window for a single wind turbine into an early warning evidence set with a unified context identifier, using a single wind turbine as the aggregation unit; a diagnostic credibility assessment module, which quantitatively assesses the early warning evidence set; this module is configured to: for early warning evidence sets originating from the same wind turbine and aggregated within a continuous time window, generate a comprehensive diagnostic confidence coefficient by comprehensively quantifying the consistency of early warning events within the set, the historical reliability of the early warning models corresponding to each early warning event, and the real-time quality of the data on which each early warning event is based; and a cloud-based intelligent diagnosis module, which performs fault diagnosis analysis based on the early warning evidence set and outputs a final diagnostic report with the comprehensive diagnostic confidence coefficient.

2. The cloud-based CMS diagnostic platform based on edge computing and AI models according to claim 1, characterized in that, The edge computing system generates preliminary early warning events in the following way: it synchronously collects vibration, temperature raw waveforms, and speed and power operating condition signals from the condition monitoring sensors of the wind turbine through standard industrial protocols; it performs preprocessing on the raw waveforms, including filtering, noise reduction, and operating condition slicing, and extracts preset time-domain statistical features and frequency-domain spectral features from them; The extracted feature data is input into a lightweight early warning model library deployed at the edge. This model library runs multiple independent AI early warning models in parallel for real-time analysis. The outputs of each early warning model are integrated to generate a structured preliminary early warning event, which includes at least an early warning model type identifier, a normalized anomaly intensity value, a precise microsecond-level trigger timestamp, and a measurement point code that identifies the specific sensor location.

3. The cloud-based CMS diagnostic platform based on edge computing and AI models according to claim 2, characterized in that, The lightweight early warning model running in the edge computing system includes at least three types with different analysis dimensions: a threshold early warning model for determining whether vibration characteristic values ​​exceed preset static or dynamic thresholds; an early warning model for determining whether the growth rate of vibration characteristic values ​​in adjacent analysis periods exceeds the short-term trend increase; and an early warning model for determining whether the abnormal acceleration trend of the growth rate of vibration characteristic values ​​in a long period exceeds the long-term trend increase. The normalized intensity value in the preliminary early warning event is the result of mapping the original output intensity values ​​with different dimensions from different types of early warning models to a unified, dimensionless numerical range through a preset normalization function.

4. The cloud-based CMS diagnostic platform based on edge computing and AI models according to claim 1, characterized in that, The early warning aggregation module is specifically configured as follows: using a single wind turbine as the basic aggregation unit, a sliding time window mechanism is used to listen to and cache the continuously uploaded preliminary early warning event stream in real time; for the cached events, correlation analysis is performed based on their turbine identifier and timestamp, and all events falling within the same sliding window are dynamically aggregated into a logically unified early warning evidence set; each generated early warning evidence set is assigned a globally unique tracking identifier, along with the window start and end times, the list of included events, and the original data index, forming a complete diagnostic context.

5. The cloud-based CMS diagnostic platform based on edge computing and AI models according to claim 1, characterized in that, The diagnostic confidence assessment module includes: a consistency quantification unit, used to calculate the internal synergy of the early warning evidence set, which is achieved by fusing the intensity value dispersion and trigger time dispersion of all early warning events within the set; a multi-source information fusion unit, used to acquire and integrate the historical accuracy reputation score of the early warning model associated with each early warning event in the early warning evidence set, as well as the real-time quality score of the original data on which the event was based; and a confidence synthesis unit, used to calculate and generate the comprehensive diagnostic confidence coefficient based on the consistency quantification result, the multi-source information fusion result, and the quantity information of early warning events using a predefined algorithm.

6. The cloud-based CMS diagnostic platform based on edge computing and AI models according to claim 5, characterized in that, The historical accuracy reputation score of the early warning model is dynamically maintained and obtained through the following methods: The platform maintains a cloud-based model performance knowledge base, continuously recording the historical trigger records of each early warning model and its corresponding subsequent on-site verification results; When the diagnostic report output by the platform is verified by on-site operation and maintenance, the relevant early warning events and their triggering models are traced back according to the verification results, and the historical accuracy reputation score of the corresponding model in the model performance knowledge base is dynamically adjusted according to the preset reputation update algorithm. When needed, the diagnostic credibility assessment module retrieves the latest reputation score of the corresponding model from the model performance knowledge base in real time.

7. The cloud-based CMS diagnostic platform based on edge computing and AI models according to claim 5, characterized in that, The consistency quantification unit calculates the internal coherence of the early warning evidence set through the following steps: S1. Intensity dispersion calculation: calculate the standard deviation of the normalized intensity value of each early warning event in the set relative to its arithmetic mean; S2. Time dispersion calculation: calculate the standard deviation of the trigger timestamps of each early warning event in the set, and divide it by the average of the differences between each pair of these timestamps to obtain the time dispersion factor; S3. Coherence synthesis: combine the intensity dispersion and the time dispersion factor by weighting or multiplying to form a diagnostic intention dispersion index. The higher the index value, the worse the consistency of events within the set.

8. The cloud-based CMS diagnostic platform based on edge computing and AI models according to claim 5, characterized in that, The real-time data quality score used in the multi-source information fusion unit is calculated synchronously by the edge computing system when an early warning event is generated and encapsulated in the event data packet; The calculation basis for the real-time quality score includes: the signal-to-noise ratio of the original vibration signal used to generate the warning within the effective frequency band, the sampling integrity index of the waveform data, and the operating condition stability score of the wind turbine power and speed during the signal acquisition period.

9. The cloud-based CMS diagnostic platform based on edge computing and AI models according to claim 7, characterized in that, The diagnostic intention dispersion index in step S3 is calculated using the following formula: ;in, Indicator of diagnostic intention dispersion; The total number of warning events in the warning evidence set; For the first Normalized intensity values ​​of each warning event; This is the average of the normalized intensity values ​​for all warning events. The standard deviation of the timestamps for all warning events; It is the average of the absolute values ​​of the differences between all pairwise trigger timestamps. and These are extremely small constants, used to ensure that the strength term is never zero and to avoid the denominator being zero, respectively.

10. The cloud-based CMS diagnostic platform based on edge computing and AI models according to claim 9, characterized in that, The confidence level synthesis unit calculates the comprehensive diagnostic confidence coefficient using the following formula: ;in, This represents the overall diagnostic confidence coefficient; Indicates the first The historical accuracy and reputation score of the early warning model corresponding to each early warning event; Indicates the generation of the first The real-time quality score of the data used in each early warning event; This is a preset baseline constant used to adjust the gain strength of the confidence level based on the amount of evidence. This represents the natural logarithm function.

Images