Power plant unit early warning method based on cloud platform

Abstract

The invention discloses a power plant unit early warning method based on a cloud platform, and relates to the technical field of power plant unit early warning, and the method comprises the steps: collecting high-frequency dynamic data and low-frequency steady-state data of a power plant unit; performing spectral analysis on the high-frequency dynamic data, extracting high-frequency sub-signal features, extracting low-frequency trend features from the low-frequency steady-state data, and performing dimensionality reduction through principal component analysis to form a multi-level feature set; the method comprises the following steps: extracting a multi-level feature set before normal operation and fault from a historical database, labeling according to a fault type, generating an association rule and constructing a rule base, dynamically updating a rule weight based on rule prediction accuracy, generating a mode class by adopting an ART neural network, and updating a class center or creating a new mode class after real-time feature clustering; the rules in the rule base are divided into three levels according to the confidence coefficient and the fault emergency degree, the matching degree and the early warning index of the real-time features and the rules are calculated, the hierarchical response is triggered according to the early warning index, and the accuracy, the real-time performance and the comprehensiveness of fault early warning are remarkably improved.

Description

Technical Field

[0001] This invention relates to the field of early warning technology for power plant units, specifically to an early warning method for power plant units based on a cloud platform. Background Technology

[0002] In traditional power plant operation and management, unit condition monitoring and fault early warning mainly rely on locally deployed distributed control systems (DCS) and human experience judgment. Existing technologies collect physical parameters such as temperature, pressure, and vibration in real time by installing sensors on key equipment of the unit, such as boilers, steam turbines, and generators, and perform anomaly detection based on threshold alarms or simple mathematical models, such as moving averages and standard deviation analysis. Some advanced systems have introduced machine learning algorithms, such as support vector machines and neural networks, to train models through historical data to achieve fault classification or remaining life prediction.

[0003] However, human experience-based judgment is highly subjective and has a slow response time, while threshold-based alarm mechanisms are difficult to capture complex fault modes, such as dynamic frequency band characteristics, which can easily lead to false alarms or missed alarms. High-frequency dynamic data (such as bearing vibration) is instantaneous and complex, and simple mathematical models cannot effectively extract its sensitive frequency band characteristics. Trend analysis of low-frequency steady-state data (such as temperature and pressure) lacks dynamic adaptability and is difficult to identify special patterns such as step-like sudden changes.

[0004] Furthermore, the processing efficiency of high-dimensional feature sets is low, and the model relies on a large amount of historical data for training, making it difficult to adapt to the needs of real-time early warning. Existing technologies have not fully integrated the correlation features between high-frequency dynamic data and low-frequency steady-state data, resulting in insufficient comprehensiveness and accuracy of fault early warning and the inability to realize a graded response mechanism. Summary of the Invention

[0005] (a) Technical problems to be solved To address the shortcomings of existing technologies, this invention provides a cloud-based early warning method for power plant units. By collecting and fusing high-frequency dynamic data and low-frequency steady-state data in real time, and combining spectrum analysis, adaptive feature extraction, principal component dimensionality reduction, dynamic rule base construction, and ART neural network clustering technology, it effectively solves the problems of strong subjectivity, delayed response, difficulty in capturing complex fault modes, low efficiency in high-dimensional data processing, poor model adaptability, and lack of hierarchical response in traditional methods, significantly improving the accuracy, real-time performance, and comprehensiveness of fault early warning.

[0006] (II) Technical Solution To achieve the above objectives, the present invention provides the following technical solution: a power plant unit early warning method based on a cloud platform, comprising: Collect high-frequency dynamic data and low-frequency steady-state data from power plant units, segment the high-frequency dynamic data, periodically summarize the low-frequency steady-state data, and establish a linear trend model for the high-frequency dynamic data and a polynomial trend model for the low-frequency steady-state data. Spectral analysis is performed on high-frequency dynamic data to identify sensitive frequency bands and divide them into sub-signals. High-frequency sub-signal features are extracted, and low-frequency trend features are extracted from low-frequency steady-state data. The high-frequency sub-signal features and low-frequency trend features are concatenated to form an initial feature set. Principal component analysis is used to reduce the dimensionality of the feature set to form a multi-level feature set. Extract multi-level feature sets from historical databases for normal operation and before failure, label them according to failure type, generate association rules and build a rule base, dynamically update rule weights based on rule prediction accuracy, use ART neural network to generate pattern classes, and update class centers or create new pattern classes after real-time feature clustering. The rules in the rule base are divided into three levels according to confidence level and fault urgency. The matching degree between real-time features and rules and the early warning index are calculated, and the graded response is triggered according to the early warning index.

[0007] Furthermore, high-frequency dynamic data includes turbine bearing vibration and generator stator current, while low-frequency steady-state data includes boiler main steam temperature and main steam pressure. Data is collected through a distributed sensor network and real-time operating status tags of the unit are collected synchronously through a distributed control system. The operating status tags include four types of tags: startup, full load, reduced load, and shutdown. When segmenting high-frequency dynamic data: during the startup and load reduction phases, each data point is divided into a data segment of 512 data points, and during the full load phase, each data point is divided into a data segment of 1024 data points. A non-overlapping sliding window is used to achieve continuous segmentation. When periodically summarizing low-frequency steady-state data: every 30 seconds during the start-up and load reduction phases, and every 60 seconds during the full-load phase.

[0008] Furthermore, during the unit startup phase, for each data point, take one adjacent point before and after it to form a three-point sequence, and replace the original data point with the median value; During the stable operation phase of the unit, wavelet threshold filtering technology is adopted, and the db4 wavelet basis function is selected for three-level decomposition. The high-frequency coefficients obtained by decomposition are processed with soft threshold to suppress high-frequency noise. When establishing a linear trend model for high-frequency dynamic data, the window size is set to 10 consecutive data points corresponding to the sampling frequency. When establishing a polynomial trend model for low-frequency steady-state data, data from five consecutive summary periods with an order of 2 are used, and the data trend is fitted using the least squares method.

[0009] Furthermore, the normal fluctuation range is set as ±3σ interval calculated based on historical data of similar operating conditions, where σ is the standard deviation of historical data; When identifying isolated outliers, the deviation between the data point and the trend model prediction must exceed the normal fluctuation range, and the two adjacent data points before and after it must be within the normal fluctuation range. For isolated outliers in high-frequency dynamic data, interpolation repair is performed using the weighted average of the three valid data points before and after the outlier, with the weights decreasing linearly according to their distance from the outlier. For isolated outliers in low-frequency steady-state data, linear interpolation of data from two adjacent summary periods is used to calculate the repair value. The repaired data must meet the requirement that the rate of change of adjacent data points does not exceed 1.2 times the historical maximum rate of change under the same operating conditions.

[0010] Furthermore, the high-frequency dynamic data is subjected to spectrum analysis by fast Fourier transform to determine the sensitive frequency bands corresponding to different fault types. Based on the sensitive frequency band, an adaptive frequency band partitioning algorithm is used to divide the sub-signal: Extract the peak values ​​in the spectrum that exceed the baseline value and the valley values ​​between adjacent peak values. The baseline value is 1.5 times the average amplitude of the spectrum under normal operating conditions. Calculate the frequency band energy corresponding to each peak, and mark the frequency band corresponding to the peak whose energy percentage is ≥10% of the total energy as the main frequency band; The main frequency bands are subdivided by valley values. If the distance between adjacent valley values ​​is ≤5Hz, they are merged into one frequency band. The bandwidth of each final frequency band is 10-50Hz. For each frequency band, a Butterworth bandpass filter is used to decompose the high-frequency dynamic data into sub-signals.

[0011] Furthermore, the characteristic parameters of each sub-signal are calculated, including energy proportion, peak frequency, and waveform distortion. The feature parameters are standardized using z-score. For low-frequency steady-state data, the rate of change is calculated using a sliding window, with the window size dynamically adjusted according to the data type: the temperature data window is set to 5 minutes, and the pressure data window is set to 10 minutes. The change pattern of the data is determined based on the rate of change, including continuous rise, periodic fluctuation and step-like abrupt change; The high-frequency sub-signal features and low-frequency trend features are concatenated to form an initial feature set. The dimensionality is reduced by principal component analysis algorithm, and principal components with a cumulative variance contribution rate of ≥95% are retained to form a multi-level feature set.

[0012] Furthermore, multi-level feature sets of normal operation and pre-failure periods are extracted from the historical database and labeled according to the failure type; The MapReduce framework is used to generate a global frequent itemset, and items with support of less than 30% are deleted; The F_List table is generated by sorting the support in descending order. The F_List is divided into Q groups by a load balancing grouping strategy. Each group corresponds to a Reduce node. Each Reduce node builds an FP-Tree in parallel to mine local frequent itemsets. Merge local frequent itemsets to generate global frequent itemsets, extract feature combinations with confidence ≥80% → construct a rule base using fault type rules; Set initial values ​​for rule weights, update weights based on rule prediction accuracy, and use an ART neural network to cluster the feature set. Set an initial value for the warning parameter ρ, and dynamically adjust the ρ value based on the clustering results; The initial pattern class is generated using the feature set of normally operating samples as training data; When the Euclidean distance d ≤ ρ between the real-time feature and the class center vector, the feature is assigned to the class and the class center is updated; when d > ρ, a new pattern class is created.

[0013] Furthermore, a pattern class ID-rule grouping mapping table is established, and rules are matched in priority order; Calculate the warning index for the matching rules. The warning index is the sum of the products of the rule weight, confidence and matching degree of all matching rules. When the warning index is ≥0.8, a Level 1 warning response is triggered; When 0.6 ≤ warning index < 0.8, a Level II warning response is triggered; When the warning index is less than 0.6, a Level III warning response is triggered.

[0014] An electronic device includes a memory, a processor, and a computer program stored in the memory and running on the processor, wherein the processor executes the computer program to implement a cloud-based power plant unit early warning method.

[0015] A computer-readable storage medium storing a computer program, which, when executed, implements a power plant unit early warning method based on a cloud platform.

[0016] (III) Beneficial Effects This invention provides a cloud-based early warning method for power plant units, which has the following advantages: (1) High-frequency dynamic data and low-frequency steady-state data of power plant units are efficiently collected through a distributed sensor network. Combined with segmented processing, median filtering and wavelet threshold filtering techniques, noise interference is effectively suppressed, and linear and polynomial trend models are established to accurately capture the data change patterns. Isolated outliers are accurately identified by setting a normal fluctuation range, and weighted interpolation or linear interpolation is used for repair, ensuring the continuity and stability of the data, significantly improving data quality, and providing a reliable foundation for subsequent fault early warning.

[0017] (2) Accurately identify fault-sensitive frequency bands through spectrum analysis, adaptively divide frequency bands and extract sub-signal features such as energy ratio and peak frequency. At the same time, use sliding window analysis to analyze the change pattern of low-frequency steady-state data, combine one-hot coding to quantize trend features, and use principal component analysis (PCA) to reduce dimensionality and retain more than 95% of key information, significantly compress data dimensions and eliminate redundancy, forming a high-quality multi-level feature set, deeply integrating high / low frequency features, greatly improving the comprehensiveness and efficiency of fault feature extraction, and providing an accurate and efficient data foundation for subsequent rule mining and real-time early warning.

[0018] (3) By distributing and mining historical operation data, a fault rule base is constructed. High reliability rules are selected based on confidence level ≥ 80%, and the rule weights are dynamically optimized according to the prediction accuracy, which significantly improves the applicability of rules and the reliability of early warning. At the same time, ART neural network is used to adaptively cluster real-time features, automatically update the class center or create new pattern classes, enhance the system's ability to identify unknown fault modes, realize the continuous optimization of the rule base and the flexible expansion of fault modes, and greatly improve the adaptability and accuracy of the early warning model.

[0019] (4) By dividing the priority of the three-level rules into three levels based on confidence level and fault urgency, and combining them with the dynamic early warning index, we can achieve accurate graded response, significantly improve the accuracy and timeliness of response, minimize accident losses, and optimize the allocation of operation and maintenance resources through differentiated strategies to reduce unnecessary intervention. Attached Figure Description

[0020] Figure 1 This is a schematic diagram of the steps of the power plant unit early warning method based on a cloud platform according to the present invention; Figure 2 This is a schematic diagram of the feature extraction and rule generation process of the present invention. Detailed Implementation

[0021] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0022] Please see Figures 1-2 This invention provides a cloud-based early warning method for power plant units, comprising the following steps: Step 1: Collect high-frequency dynamic data and low-frequency steady-state data of the power plant units, segment the high-frequency dynamic data, periodically summarize the low-frequency steady-state data, and establish a linear trend model for the high-frequency dynamic data and a polynomial trend model for the low-frequency steady-state data. Step one includes the following: Step 101: The cloud platform collects high-frequency dynamic data and low-frequency steady-state data of the power plant units through a distributed sensor network. The high-frequency dynamic data includes turbine bearing vibration and generator stator current, while the low-frequency steady-state data includes boiler main steam temperature and main steam pressure. The real-time operating status tags of the power plant units are collected synchronously through the distributed control system (DCS), including four types of tags: start-up, full load, load reduction, and shutdown. Step 102: The collected high-frequency dynamic data is segmented. During the startup and load reduction phases, each data point is divided into a data segment of 512 data points. During the full load phase, each data segment is divided into a data segment of 1024 data points. Continuous segmentation is achieved through a non-overlapping sliding window. For low-frequency steady-state data, the data is summarized every 30 seconds during the startup and load reduction phases and every 60 seconds during the full load phase. Step 103: During the unit startup phase, for each data point, take one adjacent point before and after it to form a 3-point sequence, and replace the original data point with the median. During the stable operation phase of the unit, wavelet threshold filtering is used, and the db4 wavelet basis function is selected for 3-level decomposition. The high-frequency coefficients obtained by decomposition are processed with soft threshold to suppress high-frequency noise generated by the inherent vibration of the equipment. Step 104: For high-frequency dynamic data, a linear trend model of adjacent data points is established using the sliding window method. The window size is set to 10 consecutive data points corresponding to the sampling frequency of the data. For example, the window for 1000Hz high-frequency data is 10 data points within 10ms. For low-frequency steady-state data, a polynomial trend model with 5 consecutive summarization periods and an order of 2 is used. The data change trend is fitted by the least squares method. Step 105: Set the normal fluctuation range to 3 times the standard deviation interval calculated based on historical data of similar working conditions, i.e. ±3σ, where σ is the standard deviation of historical data. When any data point deviates from the predicted value of the trend model beyond the normal fluctuation range, and the two adjacent data points before and after it are within the normal fluctuation range, it is determined to be an isolated outlier. Step 106: For isolated outliers in high-frequency dynamic data, interpolation repair is performed using the weighted average of the three valid data points before and after the outlier. The weights decrease linearly with distance from the outlier, and the sum of the weights is 1. For example, the weight of the nearest point is 0.3, the second nearest point is 0.15, and the farthest point is 0.5. For isolated outliers in low-frequency steady-state data, the repair value is calculated by linear interpolation of data from two adjacent summary periods. That is, the repair value = the value of the previous period + (the proportion of the period in which the outlier is located × (the value of the next period - the value of the previous period)). The repaired data must meet the continuity check, that is, the rate of change of adjacent data points does not exceed 1.2 times the historical maximum rate of change under the same operating conditions. Otherwise, the repair is repeated until the continuity check is met.

[0023] When using this method, please refer to steps 101 to 106: By efficiently acquiring high-frequency dynamic data and low-frequency steady-state data from power plant units through a distributed sensor network, and effectively suppressing noise interference through segmented processing, median filtering, and wavelet threshold filtering techniques, linear and polynomial trend models are established to accurately capture data change patterns. Isolated outliers are accurately identified by setting a normal fluctuation range, and weighted interpolation or linear interpolation is used for repair, ensuring data continuity and stability, significantly improving data quality, and providing a reliable foundation for subsequent fault early warning.

[0024] Step 2: Perform spectrum analysis on the high-frequency dynamic data to determine the sensitive frequency bands and divide them into sub-signals. Extract the features of the high-frequency sub-signals and extract the low-frequency trend features from the low-frequency steady-state data. Combine the high-frequency sub-signal features and the low-frequency trend features to form an initial feature set. Reduce the dimensionality through principal component analysis to form a multi-level feature set. Step two includes the following: Step 201: For high-frequency dynamic data, utilize historical fault data stored on the cloud platform, which includes at least 3 years of typical unit fault records and covers more than 20 fault types such as bearing wear and rotor imbalance. Perform spectrum analysis through Fast Fourier Transform (FFT) to obtain a spectrum diagram and determine the sensitive frequency bands corresponding to different fault types. For example, bearing faults correspond to the 100-300Hz frequency band, where the energy proportion of fault characteristic signals is ≥40%, and rotor imbalances correspond to the 50-80Hz frequency band, where the signal-to-noise ratio of characteristic signals is ≥15dB. Step 202: Based on sensitive frequency bands, an adaptive frequency band allocation algorithm is used to divide the frequency bands. The specific steps include: Step 2021: Use a spectrum analysis tool to extract all peak values ​​in the spectrum that exceed the baseline value and the valley values ​​between adjacent peak values. The baseline value is 1.5 times the average amplitude of the spectrum under normal operating conditions. Step 2022: Calculate the frequency band energy corresponding to each peak. The frequency band energy is the total energy within the 10Hz bandwidth where the peak is located. Mark the frequency band corresponding to the peak whose energy percentage is ≥10% of the total energy as the main frequency band. Step 2023: Subdivide the main frequency bands by valley values. If the distance between adjacent valley values ​​is ≤5Hz, merge them into one frequency band. Ensure that each final divided frequency band contains at least one peak signal with fault-sensitive characteristics and the bandwidth range of the frequency band is 10-50Hz. Use a Butterworth bandpass filter (order 8, passband ripple ≤1dB) for each frequency band to decompose the high-frequency dynamic data into multiple sub-signals.

[0025] Step 203: Calculate the characteristic parameters of each sub-signal, including energy proportion, peak frequency, and waveform distortion. Wherein, energy proportion = root mean square energy of sub-signal / total root mean square energy of original high-frequency signal, peak frequency = frequency value corresponding to the largest amplitude value in the sub-signal spectrum, waveform distortion = (sub-signal peak value - sub-signal effective value) / sub-signal effective value, the effective value is the root mean square value, and the characteristic parameters are processed using the z-score standardization formula, the standardized value = (original characteristic value - historical normal data characteristic mean) / historical normal data characteristic standard deviation; Step 204: For low-frequency steady-state data, calculate the rate of change of the low-frequency steady-state data using a sliding window. The window size is dynamically adjusted according to the data type and historical fluctuation cycle: the temperature data window is set to 5 minutes, based on the fact that the short-term fluctuation cycle of historical temperature data is concentrated in 5-10 minutes; the pressure data window is set to 10 minutes, based on the fact that the pressure parameter has a large inertia and the fluctuation cycle is usually 10-15 minutes. The formula for calculating the rate of change is: Rate of change = (last data value in the window - first data value in the window) / window duration; Step 205: When the rate of change of three consecutive sliding windows is >0, and the rate of the next window is ≥10% higher than the rate of the previous window, the change pattern of the data is determined to be a continuous increase; through Fourier transform analysis, if there is a significant fundamental frequency periodically and the energy proportion of this frequency component is ≥30%, the change pattern of the data is determined to be a periodic fluctuation; if the rate of change of a single window is >3 times the average rate of change of historical data of the same type, and the rate fluctuation of subsequent windows is ≤5%, the change pattern of the data is determined to be a step-like mutation; each pattern is converted into a three-dimensional feature vector through one-hot encoding, such as continuous increase corresponding to [1, 0, 0], and the standard deviation of the data within the window is added as a stability feature; Step 206: The characteristic parameters of the sub-signal are used as high-frequency sub-signal features, and the change rate, change mode and stability features are used as low-frequency trend features. The high-frequency sub-signal features and low-frequency trend features are combined into an initial feature set by feature splicing. The initial feature set is then dimensionality-reduced by principal component analysis (PCA) algorithm, and principal components with a cumulative variance contribution rate of ≥95% are retained to finally form a multi-level feature set.

[0026] When using this method, refer to steps 201 to 206: By accurately identifying fault-sensitive frequency bands through spectrum analysis, adaptively dividing frequency bands and extracting sub-signal features such as energy proportion and peak frequency, while using sliding window analysis to analyze the variation patterns of low-frequency steady-state data and combining one-hot encoding to quantify trend features, principal component analysis (PCA) is used to reduce dimensionality and retain more than 95% of key information, significantly compressing data dimensionality and eliminating redundancy, forming a high-quality multi-level feature set, deeply integrating high / low frequency features, greatly improving the comprehensiveness and efficiency of fault feature extraction, and providing an accurate and efficient data foundation for subsequent rule mining and real-time early warning.

[0027] Step 3: Extract multi-level feature sets from the historical database for normal operation and before failure, label them according to failure type, generate association rules and build a rule base, dynamically update rule weights based on rule prediction accuracy, use ART neural network to generate pattern classes, update class centers or create new pattern classes after real-time feature clustering; Step three includes the following: Step 301: Extract a large number of multi-level feature sets from the historical database of the cloud platform for normal operation and before failure, classify and label them according to the failure type, and store the labeling results as a "feature set ID-failure type-occurrence time" association table; Step 302: Use the MapReduce framework to scan the feature set to generate a global frequent 1-itemset, delete items with support <30%, sort by support in descending order to generate an F_List table, divide the F_List into Q groups using a load balancing grouping strategy, with each group corresponding to a Reduce node, and each Reduce node builds an FP-Tree in parallel, recursively mining local frequent itemsets. For example, bearing fault-related rules are processed on node 1, and steam temperature abnormality rules are processed on node 2. Merge local frequent itemsets to generate a global frequent itemset, improve I / O efficiency through a distributed caching mechanism, and finally extract the "feature combination → fault type" rule to build a rule base. The rule requires a confidence level ≥80% (i.e., the probability that the consequent is true when the antecedent of the rule occurs is ≥80%). Support refers to the frequency of a certain feature combination (i.e., a set of multiple features related to the fault) in a historical sample set. For example, if the feature combination "bearing vibration 100-300Hz energy ratio > 0.4 and temperature change rate > 2℃ / min" appears 350 times in 1000 historical samples, then its support is 35% (350 / 1000). Step 303: Set the initial value of the rule weight (e.g., 1), and update the rule weight according to the rule prediction accuracy at a fixed period (e.g., every seven days): If the accuracy is greater than 85%, the rule weight = rule weight × (1 + α), α = (accuracy - 85%) / 10%, with an upper limit of 2.0; If the accuracy is less than 70%, the rule weight = rule weight × (1-α), α = (70% - accuracy) / 10%, with a lower limit of 0.5; Step 304: Adopt the ART neural network mode, set the input layer dimension to be consistent with the feature set dimension, set the initial value of the warning parameter ρ (e.g., 0.75), and dynamically adjust it according to the clustering results: if the number of clusters is less than 10, decrease ρ by 0.05 to expand the clustering range; if the number of clusters is more than 20, increase ρ by 0.05 to narrow the clustering range. At the same time, select the feature set of 1000 normal operating samples as training data, generate the initial pattern class through competitive learning, and the class center vector is the mean of all samples in the class. Calculate the Euclidean distance d between the real-time feature and the class center vector. If d ≤ ρ, then it is assigned to the class where the class center feature is located, and the class center vector is updated. If d > ρ, then a new pattern class is created, with the real-time feature as the initial center. When using this method, refer to steps 301 to 304: By constructing a fault rule base through distributed mining of historical operational data, high-reliability rules are selected based on a confidence level of ≥80%, and rule weights are dynamically optimized according to prediction accuracy, significantly improving rule applicability and early warning reliability. At the same time, ART neural networks are used to adaptively cluster real-time features, automatically updating class centers or creating new pattern classes, enhancing the system's ability to identify unknown fault modes. This achieves continuous optimization of the rule base and flexible expansion of fault modes, greatly improving the adaptability and accuracy of the early warning model.

[0028] Step 4: Divide the rules in the rule base into three levels according to confidence level and fault urgency, calculate the matching degree between real-time features and rules and the early warning index, and trigger a graded response based on the early warning index.

[0029] Step four includes the following: Step 401: Divide the rules in the rule base into three priority levels: Level 1 rules correspond to a confidence level ≥ 90% and are associated with urgent faults (such as bearing jamming), with the highest priority; Level 2 rules correspond to a confidence level of 80%-90% and are associated with potential faults (such as minor leaks); Level 3 rules correspond to a confidence level of 70%-80% and are associated with parameters exceeding limits. Establish a "Pattern Class ID - Rule Grouping" mapping table (i.e., pattern classes and rules with the same characteristics). During retrieval, rules are matched according to priority order. Step 402: For the matching rules, calculate the warning index, which is the sum of the product of the rule weight, confidence and matching degree of all matching rules, where the matching degree = 1-D / ρ, and D is the Euclidean distance between the real-time feature and the rule antecedent; Step 403: When the warning index is ≥0.8, a Level 1 warning response is triggered. The cloud platform sends a load reduction command to the unit's DCS system through an industrial control protocol (such as ModbusTCP). The load reduction range is 10% × (warning index - 0.8) / 0.2 (maximum reduction of 30%). At the same time, the operation and maintenance personnel are notified via SMS and system pop-up (if no confirmation is received within 5 minutes, the duty phone number will be automatically dialed), and real-time data is uploaded to the emergency command platform. When 0.6 ≤ warning index < 0.8, a level 2 warning response is triggered, a diagnostic report is generated, including abnormal features, correlation rules, and historical similar cases, and is pushed to the responsible person through the operation and maintenance system, prompting key inspection areas, such as checking the boiler superheater temperature, and starting special data collection, increasing the sampling frequency of the corresponding parameters by 10 times; When the warning index is greater than 0.8, a level 3 warning response is triggered. The warning time, characteristic value, and correlation rules are recorded in the system log. The monitoring period is set to the original period / 2, such as shortening it from 5 minutes to 2.5 minutes. The parameter changes are continuously tracked. If the warning is not upgraded within 24 hours, it will be automatically lifted.

[0030] When using this method, refer to the content of steps 401 to 403: By prioritizing rules into three levels based on confidence level and fault urgency, and combining this with a dynamic early warning index, precise tiered responses are achieved, significantly improving response accuracy and timeliness, minimizing accident losses, and optimizing the allocation of operational resources through differentiated strategies to reduce unnecessary interventions.

[0031] An electronic device includes a memory, a processor, and a computer program stored in the memory and running on the processor, wherein the processor executes the computer program to implement a cloud-based early warning method for power plant units provided by the methods described above.

[0032] A computer-readable storage medium includes a memory, a processor, and a computer program stored on the memory and running on the processor, wherein the processor executes the computer program to implement a cloud-based power plant unit early warning method provided by the methods described above.

[0033] In the application, the various formulas mentioned are all calculated by removing dimensions and taking their numerical values. The formulas are derived from software simulations based on a large amount of collected data to obtain the most recent real-world results. The coefficients in the formulas are set by those skilled in the art according to the actual situation.

[0034] The above embodiments can be implemented, in whole or in part, by software, hardware, firmware, or any other combination thereof. When implemented in software, the above embodiments can be implemented, in whole or in part, as a computer program product. Those skilled in the art will recognize that the units and algorithm steps of the various examples described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, computer software, and combinations thereof. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution.

[0035] The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; they may be located in one place or distributed across multiple network units. Some or all of the units can be selected to achieve the purpose of this embodiment, depending on actual needs.

[0036] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any changes or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application.

Claims

1. A power plant unit early warning method based on a cloud platform, characterized in that: include: Collect high-frequency dynamic data and low-frequency steady-state data from power plant units, segment the high-frequency dynamic data, periodically summarize the low-frequency steady-state data, and establish a linear trend model for the high-frequency dynamic data and a polynomial trend model for the low-frequency steady-state data. Spectral analysis is performed on high-frequency dynamic data to identify sensitive frequency bands and divide them into sub-signals. High-frequency sub-signal features are extracted, and low-frequency trend features are extracted from low-frequency steady-state data. The high-frequency sub-signal features and low-frequency trend features are concatenated to form an initial feature set. Principal component analysis is used to reduce the dimensionality of the feature set to form a multi-level feature set. Extract multi-level feature sets from historical databases for normal operation and before failure, label them according to failure type, generate association rules and build a rule base, dynamically update rule weights based on rule prediction accuracy, use ART neural network to generate pattern classes, and update class centers or create new pattern classes after real-time feature clustering. The rules in the rule base are divided into three levels according to confidence level and fault urgency. The matching degree between real-time features and rules and the early warning index are calculated, and the graded response is triggered according to the early warning index.

2. The power plant unit early warning method based on a cloud platform according to claim 1, characterized in that: High-frequency dynamic data includes turbine bearing vibration and generator stator current, while low-frequency steady-state data includes boiler main steam temperature and main steam pressure. Data is collected through a distributed sensor network and real-time operating condition tags of the unit are collected synchronously through a distributed control system. The operating condition tags include four types: startup, full load, reduced load, and shutdown. When segmenting high-frequency dynamic data: during the startup and load reduction phases, each data point is divided into a data segment of 512 data points, and during the full load phase, each data point is divided into a data segment of 1024 data points. A non-overlapping sliding window is used to achieve continuous segmentation. When periodically summarizing low-frequency steady-state data: during the startup and load reduction phases, the data is summarized every 30 seconds, and during the full load phase, the data is summarized every 60 seconds.

3. The power plant unit early warning method based on a cloud platform according to claim 2, characterized in that: During the unit startup phase, a three-point sequence is formed by taking one adjacent point before and after each data point, and the median value is used to replace the original data point. During the stable operation phase, wavelet threshold filtering technology is used, and the db4 wavelet basis function is selected for three-level decomposition. The high-frequency coefficients obtained by decomposition are processed with soft thresholding to suppress high-frequency noise. When establishing a linear trend model for high-frequency dynamic data, the window size is set to 10 consecutive data points corresponding to the sampling frequency. When establishing a polynomial trend model for low-frequency steady-state data, data from 5 consecutive summary periods are used, with an order of 2, and the data change trend is fitted by the least squares method.

4. The power plant unit early warning method based on a cloud platform according to claim 3, characterized in that: The normal fluctuation range is defined as ±3σ, calculated based on historical data of similar operating conditions, where σ is the standard deviation of historical data. When identifying isolated outliers, the deviation between the data point and the trend model prediction must exceed the normal fluctuation range, and the two adjacent data points before and after the outlier must be within the normal fluctuation range. For isolated outliers in high-frequency dynamic data, interpolation repair is performed using the weighted average of the three valid data points before and after the outlier, with the weights decreasing linearly according to their distance from the outlier. For isolated outliers in low-frequency steady-state data, the repair value is calculated by linear interpolation of data from two adjacent summary periods. The repaired data must satisfy the condition that the rate of change of adjacent data points does not exceed 1.2 times the historical maximum rate of change under similar operating conditions.

5. The power plant unit early warning method based on a cloud platform according to claim 1, characterized in that: Spectral analysis of high-frequency dynamic data using Fast Fourier Transform (FFT) was performed to determine sensitive frequency bands corresponding to different fault types. Based on these sensitive frequency bands, an adaptive frequency band partitioning algorithm was used to divide the data into sub-signals: peak values ​​exceeding the baseline value and valley values ​​between adjacent peak values ​​were extracted from the spectrum. The baseline value was 1.5 times the average amplitude of the spectrum under normal operating conditions. The frequency band energy corresponding to each peak was calculated, and the frequency bands corresponding to peaks with an energy percentage ≥ 10% of the total energy were marked as the main frequency bands. The main frequency bands were further subdivided by valley values. If the distance between adjacent valley values ​​was ≤ 5 Hz, they were merged into one frequency band, with each final frequency band having a bandwidth of 10-50 Hz. A Butterworth bandpass filter was used for each frequency band to decompose the high-frequency dynamic data into sub-signals.

6. The power plant unit early warning method based on a cloud platform according to claim 5, characterized in that: Calculate the characteristic parameters of each sub-signal, including energy percentage, peak frequency, and waveform distortion. The feature parameters are standardized using z-scores. The rate of change of low-frequency steady-state data is calculated using a sliding window, with the window size dynamically adjusted according to the data type: the window for temperature data is set to 5 minutes, and the window for pressure data is set to 10 minutes. The change pattern of the data is determined based on the rate of change, including continuous rise, periodic fluctuation, and step-like abrupt change. The high-frequency sub-signal features and low-frequency trend features are concatenated to form an initial feature set. The dimensionality is reduced by principal component analysis, and principal components with a cumulative variance contribution rate of ≥95% are retained to form a multi-level feature set.

7. The power plant unit early warning method based on a cloud platform according to claim 1, characterized in that: Multi-level feature sets are extracted from historical databases for normal operation and pre-failure periods, and labeled according to failure type. A global frequent itemset is generated using the MapReduce framework, and items with support <30% are deleted. An F_List table is generated by sorting items in descending order of support. The F_List is divided into Q groups using a load balancing grouping strategy, with each group corresponding to a Reduce node. Each Reduce node constructs an FP-Tree in parallel to mine local frequent itemsets. Local frequent itemsets are merged to generate a global frequent itemset, and feature combinations with confidence ≥80% are extracted to construct a rule base for failure type rules. Initial values ​​for rule weights are set, and the weights are updated based on the rule prediction accuracy. An ART neural network is used to cluster the feature set: an initial value for the warning parameter ρ is set, and the ρ value is dynamically adjusted based on the clustering results. The initial pattern class is generated using the feature set of normally operating samples as training data; When the Euclidean distance d ≤ ρ between the real-time feature and the class center vector, the feature is assigned to the class and the class center is updated; when d > ρ, a new pattern class is created.

8. The power plant unit early warning method based on a cloud platform according to claim 1, characterized in that: Establish a pattern class ID-rule grouping mapping table and match rules in priority order; calculate the warning index for the matched rules, which is the sum of the products of the rule weight, confidence and matching degree of all matched rules; When the warning index is ≥0.8, a Level I warning response is triggered; when 0.6≤the warning index<0.8, a Level II warning response is triggered; when the warning index<0.6, a Level III warning response is triggered.

9. An electronic device comprising a memory, a processor, and a computer program stored in the memory and running on the processor, characterized in that, When the processor executes the computer program, it implements the power plant unit early warning method based on a cloud platform as described in claims 1-8.

10. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a computer program, which, when executed, implements the power plant unit early warning method based on a cloud platform as described in claims 1-8.

Images