Management method of intelligent power plant based on digital twinning and AI technologies

Abstract

The invention discloses a management method of a smart power plant based on digital twinning and AI technologies, and relates to the technical field of power plant management, and the method comprises the steps: constructing an initial signal data set composed of equipment operation parameters; performing hierarchical analysis on the initial signal data set, mapping equipment operation parameters to a feature vector space, and generating a unified feature data set; comparing the equipment operation parameters in real time by using the operation state classifier, calculating an abnormal probability distribution matrix of the equipment operation parameters, and determining an abnormal risk level of current equipment operation; extracting abnormal fragment data from the unified feature data set according to the state deterioration early warning signal, performing segmentation processing to obtain a segmented abnormal data set, performing future trajectory deduction on the segmented abnormal data set, and determining a future abnormal development trajectory of the potential fault; the abnormal point of the equipment is determined based on the future abnormal development track, and the abnormal development track is associated and bound with the equipment operation parameter of the current abnormal point, so that the accuracy and timeliness of monitoring the abnormal state of the equipment operation are improved.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of power plant management, in particular to a management method of a smart power plant based on digital twinning and AI technology. BACKGROUND

[0002] In the operation monitoring of micro-devices in remote areas, there are challenges such as weak signals, scattered data, and difficulty in capturing abnormalities. To solve these problems, a unified signal acquisition architecture needs to be built to preprocess and extract features from multiple weak signals, and generate a standardized feature dataset. However, due to the complexity of the device operating environment and frequent signal interference, it is difficult to extract abnormal features and accurately map them to a unified feature space. At the same time, the device state monitoring system has a high false alarm rate when processing fragmented data, making it difficult to accurately determine the abnormal risk level. In addition, due to the short duration of the abnormality, traditional trend prediction algorithms are difficult to capture the abnormal development trajectory, resulting in delayed fault warning. These problems are interrelated and collectively affect the effective monitoring and fault warning of micro-devices, and an integrated solution is needed to comprehensively address the problems of weak signals, difficulty in capturing abnormalities, and large prediction deviations, in order to improve the accuracy and real-time performance of remote monitoring. SUMMARY

[0003] The present application aims to at least partially solve one of the technical problems in the related art. To this end, one object of the present application is to propose a management method of a smart power plant based on digital twinning and AI technology, which improves the accuracy and timeliness of device operation abnormal state monitoring.

[0004] One aspect of the present application provides a management method of a smart power plant based on digital twinning and AI technology, comprising:

[0005] Step S100: Construct a sensor signal acquisition architecture to obtain raw sensor signals, denoise and enhance the raw sensor signals, and obtain an initial signal dataset composed of device operating parameters;

[0006] Step S200: Hierarchical analysis of the initial signal dataset, mapping the device operating parameters to a feature vector space, and generating a unified feature dataset;

[0007] Step S300: Real-time comparison of the device operating parameters using a running state classifier, and calculating the abnormal probability distribution matrix to determine the abnormal risk level of the current device operation;

[0008] Step S400: Extracting abnormal segment data from the unified feature dataset according to the state deterioration warning signal and performing segmented processing to obtain a segmented abnormal dataset, and performing future trajectory inference on the segmented abnormal dataset to determine the future abnormal development trajectory of potential faults;

[0009] Step S500: determining an abnormal point of the device based on the future abnormal development trajectory, and binding the abnormal development trajectory with the device operation parameter of the current abnormal point;

[0010] The specific method for constructing the sensor signal collection architecture, obtaining the original sensor signal, and performing denoising and enhancement on the original sensor signal to obtain an initial signal data set composed of device operation parameters is as follows:

[0011] Step S110: constructing a unified sensor signal collection architecture, obtaining original sensor signals of various sensors, establishing a sensor classification mapping table according to the sensor types and signal strengths, and obtaining a first data stream with a classification identifier;

[0012] Step S120: performing signal strength evaluation on the first data stream, marking as a weak signal if the signal strength is less than a preset threshold, and marking as a normal signal if the signal strength is greater than or equal to the preset threshold;

[0013] Step S130: performing denoising processing on the weak signal through a filter bank, adjusting the filter cutoff frequency according to a denoising level parameter, and obtaining a second data stream after denoising;

[0014] Step S140: performing enhancement processing on the second data stream by using a signal amplification technology, calculating an amplification coefficient according to an enhancement multiple parameter and the signal strength of the original sensor signal, and obtaining a third data stream after enhancement;

[0015] Step S150: obtaining the amplitude range and frequency distribution characteristics of each sensor signal in the third data stream, mapping different sensor signals to a unified numerical interval through normalization processing, and determining a standardized fourth data stream;

[0016] Step S160: evaluating the fourth data stream by using a data quality evaluation method, retaining data records if the data quality index of the fourth data stream meets a preset standard, and eliminating abnormal records if the data quality index does not meet the preset standard, and generating a standardized initial signal data set; the initial signal data set contains device operation parameters processed based on sensor signals;

[0017] The method for obtaining the second data stream after denoising is as follows:

[0018] Step S131: setting a denoising level parameter, and dynamically adjusting the cutoff frequency according to the denoising level parameter;

[0019] Step S132: based on the cutoff frequency and the first data stream, applying a filter to perform denoising processing on the weak signal, and obtaining the second data stream after denoising; the second data stream includes the weak signal and the normal signal after denoising;

[0020] The method for obtaining the enhanced third data stream comprises the following steps: obtaining a signal strength of an original sensor signal, setting an enhancement multiple parameter, calculating an amplification coefficient based on the signal strength and the enhancement multiple parameter, and performing enhancement processing on the second data stream according to the amplification coefficient through a signal amplification technology to obtain the enhanced third data stream.

[0021] The specific method for performing hierarchical analysis on the initial signal data set, mapping the device operation parameters to a feature vector space, and generating a unified feature data set comprises the following steps:

[0022] Step S210: Defining the initial signal data set as a basic data stream, and performing structural decomposition on the basic data stream by using a hierarchical analysis technology to obtain signal hierarchical results.

[0023] Step S220: According to the signal hierarchical results after preliminary analysis, for the signal interference phenomenon, applying an abnormal feature extraction logic to process each layer of device operation parameters to determine the abnormal feature distribution.

[0024] Step S230: If the abnormal feature distribution exceeds a preset abnormal feature distribution threshold, performing filtering processing on the device operation parameters of the corresponding signal layer to obtain a filtered signal data stream.

[0025] Step S240: Through the filtered signal data stream, performing dimension reduction processing on the signal data stream by using a principal component analysis algorithm, mapping to a unified feature vector space, and judging the integration effect of the feature vectors.

[0026] Step S250: According to the integration effect of the feature vectors, for the feature vectors that do not meet the integration standard, applying a standardization adjustment logic for correction to obtain an adjusted feature vector set.

[0027] Step S260: Obtaining the adjusted feature vector set, performing feature matching and fusion processing on the signal data streams of different sensor types to determine a final unified feature data set.

[0028] The specific method for applying the abnormal feature extraction logic to process each layer of device operation parameters to determine the abnormal feature distribution comprises the following steps: performing feature extraction on each layer of device operation parameters by using a sliding window method, presetting a sliding window size and a sliding step, calculating the features of the device operation parameters in each window, and the features include mean, standard deviation, maximum fluctuation amplitude and gradient change rate; calculating an abnormal detection index, comparing the abnormal detection index with an index threshold, if the absolute value of the abnormal detection index is greater than the index threshold, it is determined that the device operation parameter is abnormal, the abnormal feature of the abnormal device operation parameter is marked as 1, otherwise it is marked as 0, and the abnormal feature distribution is obtained.

[0029] The method for determining whether the abnormal feature distribution exceeds the preset abnormal feature distribution threshold is: summing the total of the abnormal feature distribution, if the total is greater than the preset abnormal feature distribution threshold, it means exceeding, if the total is less than or equal to the preset abnormal feature distribution threshold, it means not exceeding.

[0030] The method for determining the integration effect of the feature vector is: first, the principal component analysis algorithm is used to reduce the dimension of the signal data stream, the feature vector of the reduced signal data stream is mapped to a unified feature vector space, and the reserved information proportion of the feature vector of the reduced signal data stream is calculated as the integration effect of the feature vector.

[0031] The specific method for determining the abnormal risk level of the current equipment operation by using the running state classifier to compare the equipment operation parameters in real time and calculate the abnormal probability distribution matrix is:

[0032] Step S310: standardizing the equipment operation parameters through data preprocessing to obtain standardized feature data;

[0033] Step S320: using the running state classifier based on the support vector machine algorithm to perform state classification processing on the standardized feature data, if the standardized feature data matches the normal operation mode, output the normal state identifier, if the standardized feature data deviates from the normal operation mode, output the abnormal state identifier and the deviation degree value;

[0034] Step S330: calculating the equipment state abnormal probability according to the abnormal state identifier and the deviation degree value, if the deviation degree value exceeds the deviation threshold, calculating the abnormal probability density function, and constructing the abnormal probability distribution matrix based on the abnormal probability density function;

[0035] Step S340: performing risk level quantization processing through the abnormal probability distribution matrix, if the abnormal probability distribution is concentrated in the high-risk interval, determining the risk level as the high-risk state, if the abnormal probability distribution is concentrated in the medium-risk interval, determining the risk level as the early warning state;

[0036] Step S350: using the sliding window mechanism to perform time series analysis on the risk level to obtain risk level change trend data, if the risk level shows an upward trend, triggering the state deterioration early warning signal;

[0037] The input of the running state classifier is the standardized feature data, and the output is the running state identifier of the standardized feature data; the running state identifier includes the normal state identifier and the abnormal state identifier;

[0038] The method for calculating the abnormal probability distribution matrix is: using a Bayesian probability device to calculate an abnormal probability for input standardized feature data x and a deviation degree value d, using a normal distribution to calculate an abnormal probability density function of the deviation degree value if the deviation degree value exceeds a preset deviation threshold, and constructing an abnormal probability distribution matrix according to the abnormal probability density function;

[0039] The method for triggering the state deterioration early warning signal is: defining a window length and a sliding step, calculating a first-order difference sequence of a risk value sequence in each window to determine whether the risk level shows an upward trend, determining that the risk level shows an upward trend if the first-order difference sequence is greater than 0 for three times in succession, and triggering the state deterioration early warning signal;

[0040] The specific method for extracting abnormal segment data from the unified feature data set according to the state deterioration early warning signal and performing segmentation processing to obtain a segmented abnormal data set, and performing future trajectory deduction on the segmented abnormal data set to determine a future abnormal development trajectory of a potential fault is:

[0041] Step S410: Extracting corresponding abnormal segment data from the unified feature data set for the state deterioration early warning signal;

[0042] Step S420: Applying a data screening rule to perform secondary filtering according to the abnormal segment data, removing segments that do not meet the duration condition through timestamp comparison, and determining selected abnormal segment data;

[0043] Step S430: Cutting the selected abnormal segment data according to a preset time interval to obtain segmented abnormal data segments;

[0044] Step S440: Extracting feature data of each segmented abnormal data segment, the feature data being a statistical indicator of each segmented abnormal data segment, the statistical indicator including a mean value and a standard deviation, and obtaining feature abnormal data;

[0045] Step S450: Classifying the feature abnormal data by using a support vector machine algorithm, and marking a segmented abnormal data segment as a high-risk abnormal segment if the feature data of the segmented abnormal data segment matches a preset abnormal pattern;

[0046] Step S460: Distinguishing the high-risk abnormal segment from a normal abnormal segment to generate a segmented abnormal data set;

[0047] Step S470: Using an LSTM algorithm to perform future trajectory deduction on the segmented abnormal data set to predict a future abnormal development trajectory;

[0048] The specific method for predicting the future abnormal development trajectory by using the LSTM algorithm on the segmented abnormal data set is: presetting a window length Q and a prediction step, using a sliding window method to obtain training samples from the segmented abnormal data set, the input of each training sample being the segmented abnormal data at the previous Q time points, the output data being the segmented abnormal data at the Q+1 time point, using mean square error as the loss function for training the LSTM model, and minimizing the value of the loss function as the training target; predicting the segmented abnormal data at each future time point based on the trained LSTM model, and the segmented abnormal data at each future time point forming the future abnormal development trajectory.

[0049] The specific method for determining the abnormal point of the equipment based on the future abnormal development trajectory and associating the abnormal development trajectory with the device operating parameters of the current abnormal point is:

[0050] Step S510: determining the corresponding sensor signal based on the abnormal development trajectory to obtain the abnormal point of the equipment, and associating the device operating parameters at the abnormal point with the abnormal development trajectory;

[0051] Step S520: making a corresponding decision on the equipment at the current abnormal point based on the abnormal segment risk level of the abnormal development trajectory, if the future abnormal development trajectory is a high-risk abnormal segment, sending a warning to the maintenance personnel, and if the future abnormal development trajectory is an ordinary abnormal segment, taking the abnormal point as a key monitoring object for continuous monitoring.

[0052] One aspect of the present application provides a management system of a smart power plant based on digital twinning and AI technology, comprising:

[0053] An initial signal acquisition module is configured to construct a sensor signal acquisition architecture, acquire original sensor signals, denoise and enhance the original sensor signals, and obtain an initial signal data set composed of device operating parameters;

[0054] A feature vector extraction module is configured to hierarchically analyze the initial signal data set, map the device operating parameters to a feature vector space, and generate a unified feature data set;

[0055] An abnormal risk assessment module is configured to use a running state classifier to compare the device operating parameters in real time and calculate an abnormal probability distribution matrix to determine the abnormal risk level of the current device operation;

[0056] A future trajectory prediction module is configured to extract abnormal segment data from the unified feature data set according to a state deterioration warning signal and perform segmented processing to obtain a segmented abnormal data set, perform future trajectory extrapolation on the segmented abnormal data set, and determine the future abnormal development trajectory of the potential fault;

[0057] The trajectory positioning processing module is configured to determine an abnormal point of the device based on a future abnormal development trajectory, and bind the abnormal development trajectory with a device operation parameter of a current abnormal point.

[0058] The management method of the smart power plant based on the digital twinning and AI technology provided in the application has the following advantages over the prior art:

[0059] The application solves the problem of format difference of multi-source device signals by constructing a unified collection architecture, supports weak signal enhancement and denoising, retains important abnormal information under low signal-to-noise ratio, and improves data availability.

[0060] The application introduces a hierarchical structured representation of device parameters, analyzes the feature differences of different types of signals at different levels, extracts multiple abnormal indicators based on a sliding window, realizes fine-grained recognition, enhances the precision and context awareness of abnormal recognition, improves the accuracy of abnormal recognition, and significantly reduces the misjudgment rate.

[0061] The application uses an SVM model to determine whether the device is in an abnormal state in real time, realizes intelligent determination and labeling management of the running state, improves the accuracy of abnormal state determination of the device operation by quantifying the abnormal probability and risk level, and uses a sliding window mechanism to perform time series analysis on the risk changes, thereby improving the recognition ability of abnormal trends.

[0062] The application tracks the future trend of the anomaly by segmentation and trajectory prediction, drives the trajectory prediction by the abnormal determination result, forms a dynamic early warning closed loop, and improves the advance and controllability of abnormal response.

[0063] The application backtracks the feature data corresponding to the risk warning signal, realizes the time slicing extraction of abnormal events, constructs a predictable abnormal behavior segment through fine segmentation and feature statistics, uses an LSTM network to model the future trajectory of the segmented data, predicts the possible development trend of the fault, and improves the prediction accuracy and advance amount. BRIEF DESCRIPTION OF DRAWINGS

[0064] Figure 1 The method flowchart of the management method of the smart power plant based on the digital twinning and AI technology provided in the application is provided;

[0065] Figure 2 The acquisition method flowchart of the unified feature data set provided in the application is provided;

[0066] Figure 3 The discrimination method flowchart of the abnormal risk level provided in the application is provided;

[0067] Figure 4 The functional module diagram of the management system of the smart power plant based on the digital twinning and AI technology provided in the application is provided. DETAILED DESCRIPTION

[0068] For a better understanding of the present application, various aspects of the present application will be described in more detail below with reference to the accompanying drawings. It is to be noted that these detailed description is merely descriptive of exemplary embodiments of the present application and is not intended in any way to limit the scope of the present application. Throughout the specification, like drawing reference numerals will refer to like elements. The expression "and / or" includes any and all combinations of one or more of the associated listed items.

[0069] In the drawings, the size, dimensions, and shapes of elements have been slightly adjusted for ease of illustration. The drawings are merely exemplary and not strictly drawn to scale. As used in this document, the terms "substantially", "approximately", and similar terms are used as terms of approximation and not as terms of degree, and are intended to account for the inherent deviations in a measuring or computing process that would be recognized by those of ordinary skill in the art. Additionally, in the present application, the order of the steps of the process described does not necessarily indicate the order in which the processes occur in actual operation, unless explicitly limited or otherwise derivable from the context.

[0070] It should also be understood that expressions such as "include", "including", "have", "has", "contain", and / or "containing", and the like, are open-ended terms that are intended to mean that the listed item is present, but not excluding the presence of one or more other features, elements, components, and / or combinations thereof. In addition, when expressions such as "at least one of" appear after a list of two or more items, it is meant that any of the listed items can be present, individually or in combination with one or more of the other listed items. Furthermore, when describing embodiments of the present application, the use of "may" means "one or more embodiments of the present application". Also, the use of the term "exemplary" is intended to refer to an example or illustration.

[0071] Unless otherwise defined, all terms used in this document, including engineering and scientific terms, have the same meaning as commonly understood by one of ordinary skill in the art to which this application belongs. It should also be understood that the words defined in commonly used dictionaries should be interpreted as having a meaning that is consistent with their meaning in the context of the relevant art and not in an idealized or overly formal sense unless expressly so defined herein.

[0072] It should be noted that the embodiments and features in the embodiments of the present application can be combined with each other without conflict. The present application will be described in detail below with reference to the accompanying drawings and in conjunction with embodiments.

[0073] Example 1

[0074] As shown in the following, the management method of the smart power plant based on digital twin and AI technology provided by the present application comprises: Figure 1 a first step of acquiring a plurality of data of the power plant;

[0075] Step S100: constructing a sensor signal acquisition architecture, acquiring original sensor signals, denoising and enhancing the original sensor signals, and obtaining an initial signal dataset composed of device operation parameters;

[0076] The specific method for constructing the sensor signal acquisition architecture, acquiring the original sensor signals, denoising and enhancing the original sensor signals, and obtaining the initial signal dataset composed of device operation parameters is as follows:

[0077] Step S110: constructing a unified sensor signal acquisition architecture, acquiring original sensor signals of various sensors, establishing a sensor classification mapping table according to sensor types and signal strengths, and obtaining a first data stream with classification identifiers;

[0078] The first data stream is composed of sensor signals of different signal strengths collected by various sensors, and each sensor signal has a classification identifier of sensor type and signal strength.

[0079] Step S120: performing signal strength evaluation on the first data stream, marking as weak signals if the signal strength is less than a preset threshold, and marking as normal signals if the signal strength is greater than or equal to the preset threshold.

[0080] The value of the preset threshold is set by a person skilled in the art according to experience.

[0081] Step S130: denoising the weak signals by a filter bank, adjusting the filter cutoff frequency according to a denoising level parameter, and obtaining a second data stream after denoising.

[0082] The method for obtaining the second data stream after denoising is as follows:

[0083] Step S131: setting a denoising level parameter and dynamically adjusting the cutoff frequency according to the denoising level parameter.

[0084] Step S132: based on the cutoff frequency and the first data stream, applying a filter to denoise the weak signals, and obtaining the second data stream after denoising; the second data stream includes the weak signals and the normal signals after denoising.

[0085] The calculation formula of the cutoff frequency is: f c =f base -α×L; wherein, L is the denoising level parameter, α is an adjustment factor, and f base is an initial denoising level parameter; preferably, the adjustment factor α is equal to 0.1.

[0086] Step S140: performing enhancement processing on the second data stream using a signal amplification technology, calculating an amplification factor according to the enhancement multiple parameter and the signal strength of the original sensor signal, and obtaining an enhanced third data stream;

[0087] The enhanced third data stream is obtained by: obtaining the signal strength of the original sensor signal, setting an enhancement factor parameter, calculating an amplification factor based on the signal strength and the enhancement factor parameter, and performing enhancement processing on the second data stream according to the amplification factor using a signal amplification technology to obtain the enhanced third data stream;

[0088] The amplification factor is the ratio of the enhancement factor parameter to the signal intensity;

[0089] Step S150: Acquire the amplitude range and frequency distribution characteristics of each sensor signal in the third data stream, map different sensor signals to a unified numerical range through normalization processing, and determine a standardized fourth data stream;

[0090] Step S160: Evaluate the fourth data stream using a data quality assessment method. If the data quality index of the fourth data stream meets the preset standard, retain the data record. If the data quality index does not meet the preset standard, remove the abnormal record to generate a standardized initial signal data set; the initial signal data set includes the device operating parameters processed based on the sensor signal;

[0091] The data quality index includes an integrity index and a consistency index. The calculation formula of the integrity index is: Wherein, Mq is the number of missing sensor signal points, and Mz is the total number of sensor signal points; the consistency index includes the variance and volatility of the sensor signal, and the volatility is the absolute difference between the sensor signals at two adjacent sampling time points;

[0092] The preset standard is set by those skilled in the art based on experience, wherein the preset standard for the integrity index is that the integrity is less than 5%, and the preset standard for the consistency index is that the consistency index is greater than the minimum value of the historical values ​​of each consistency index.

[0093] Furthermore, the initial signal data set is time-sequenced and indexed according to the acquisition frequency and data flow characteristics, and a data storage structure that supports fast retrieval is established;

[0094] For example, by constructing a unified sensor signal acquisition architecture, in order to solve the problem of scattered data sources and weak signals of micro devices in remote scenes, first at the data acquisition level, assuming that 100 micro sensors are deployed on devices in remote areas, each sensor collects running data once every second, the data types include temperature, vibration and current, the collection range is-40 to 85 degrees Celsius, 0 to 10 Hz and 0 to 5 amperes, the original data stream is transmitted to the edge gateway through the Bluetooth protocol, the transmission rate is controlled at 1Mbps to ensure signal stability, and the gateway preliminarily caches the data with a cache capacity of 10MB to prevent data loss. Then, in the signal preprocessing stage, in order to solve the problem of weak collected signals and noise interference, a wavelet transform denoising algorithm is used, the decomposition level is set to 5 layers, the high-frequency noise is filtered out by the soft threshold method, the noise ratio in the original signal is 20%, and the signal-to-noise ratio is improved to 15dB after denoising. At the same time, the signal is enhanced, the Hilbert transform is used to extract the signal envelope, the signal characteristics are enhanced, and the amplitude is increased by about 30%. The standardized initial signal dataset is obtained, the data sampling rate is unified to 100Hz, and the storage format is 16-bit floating point number, which ensures the accuracy of subsequent analysis. In order to verify the effect, a group of vibration signal data is selected, the original signal peak value is 0.5V, the peak value after denoising is 0.48V, and the peak value after enhancement is 0.62V, which indicates that the signal quality is significantly improved. The whole process is realized through automatic script, and the data is seamlessly connected from acquisition to preprocessing. The edge gateway and the cloud synchronize data through the MQTT protocol, the delay is controlled within 50ms to ensure real-time performance, and combined with the device running state analysis, if the signal fluctuation exceeds the threshold value of 10%, the alarm mechanism is automatically triggered, forming a business closed loop and improving the system reliability.

[0095] Step S200: Hierarchical analysis is performed on the initial signal dataset, device running parameters are mapped to a feature vector space, and a unified feature dataset is generated;

[0096] As shown in Figure 2 The specific method for the hierarchical analysis of the initial signal dataset, mapping the device running parameters to the feature vector space, and generating the unified feature dataset is as follows:

[0097] Step S210: Defining the initial signal dataset as a basic data stream, using hierarchical analysis technology to structurally decompose the basic data stream to obtain signal hierarchical results;

[0098] Step S220: According to the signal hierarchical results after preliminary analysis, for signal interference phenomenon, applying abnormal feature extraction logic to process each layer of device running parameters to determine the abnormal feature distribution;

[0099] The application anomaly feature extraction logic processes each layer of device operation parameters, and the specific method for determining the anomaly feature distribution is as follows: a sliding window method is used to extract features of each layer of device operation parameters, a sliding window size and a sliding step are preset, features of device operation parameters in each window are calculated, the features include mean, standard deviation, maximum fluctuation amplitude and gradient change rate; an anomaly detection index is calculated, the anomaly detection index is compared with an index threshold value, if an absolute value of the anomaly detection index is greater than the index threshold value, it is determined that the device operation parameters are abnormal, the anomaly feature of the abnormal device operation parameters is marked as 1, otherwise it is marked as 0, and the anomaly feature distribution is obtained;

[0100] Step S230: If the anomaly feature distribution exceeds the preset anomaly feature distribution threshold value, the device operation parameters of the corresponding signal layer are filtered to obtain a filtered signal data stream.

[0101] The method for determining whether the anomaly feature distribution exceeds the preset anomaly feature distribution threshold value is as follows: the sum of the anomaly feature distribution is counted, if the sum is greater than the preset anomaly feature distribution threshold value, it is indicated that the anomaly feature distribution exceeds the preset anomaly feature distribution threshold value, and if the sum is less than or equal to the preset anomaly feature distribution threshold value, it is indicated that the anomaly feature distribution does not exceed the preset anomaly feature distribution threshold value.

[0102] Step S240: The principal component analysis algorithm is used to reduce the dimension of the signal data stream through the filtered signal data stream, the signal data stream is mapped to a unified feature vector space, and the integration effect of the feature vector is determined.

[0103] The method for determining the integration effect of the feature vector is as follows: the principal component analysis algorithm is used to reduce the dimension of the signal data stream, the feature vector of the reduced signal data stream is mapped to a unified feature vector space, and the proportion of the reserved information of the feature vector of the reduced signal data stream is calculated as the integration effect of the feature vector.

[0104] Step S250: According to the integration effect of the feature vector, the standardization adjustment logic is applied to correct the feature vector that does not meet the integration standard to obtain an adjusted feature vector set.

[0105] The adjustment method of the feature vector set is as follows: the feature vector that does not meet the integration standard is normalized, the normalized feature vector is Z-score standardized to obtain an adjusted feature vector set.

[0106] Step S260: The adjusted feature vector set is obtained, the feature matching and fusion processing are performed on the signal data streams of different sensor types, and a final unified feature data set is determined.

[0107] Further, a cluster analysis method is used to perform group verification on the unified feature data set to obtain a verified feature grouping result; the cluster analysis method is used to perform group verification on the unified feature data set to determine whether the final unified feature data set can semantically distinguish the category and state of the sensor;

[0108] Specifically, based on the processing of the initial signal dataset, the real-time parameter acquisition module first performs a hierarchical analysis of the multi-source signals. Assume that the initial dataset contains signal streams from 10 micro-devices, each generating 1000 data points per second. Signal types include voltage, current, and frequency, with data ranges of 0-5V, 0-10A, and 50-60Hz, respectively. The system uses a hierarchical analysis algorithm to separate the signals by type and extracts the frequency domain characteristics of each signal through Fourier transform. For example, the dominant frequency component of the frequency signal is calculated, resulting in a dominant frequency value of 55.3Hz. Its amplitude and phase offset of 0.2rad are recorded for subsequent analysis. To address the frequent signal interference, the system extracts anomaly features from the data streams of different micro-devices. For example, if device 1's data stream contains interference noise with a standard deviation exceeding a threshold of 2.5, the system decomposes the signal using a wavelet transform algorithm, extracting anomaly feature vectors. For example, if the noise component accounts for 15.2% of the energy, the feature vectors are mapped to a unified feature vector space. Principal component analysis is then used to reduce the multidimensional features of the 10 devices to a three-dimensional space, retaining 95% of the variance. This generates a feature vector such as [0.45, 0.33, 0.22]. Finally, the integrated unified feature dataset is normalized to scale all eigenvalues ​​to the range of 0 to 1. For example, the original vector [0.45, 0.33, 0.22] is normalized to [0.9, 0.66, 0.44]. This data is then stored in a standard CSV format file containing the device ID, timestamp, and feature values ​​for subsequent machine learning model training. This process is automated using automated scripts. The real-time data stream processing module updates the feature dataset every minute to ensure the timeliness of the feature dataset.

[0109] Step S300: using the operating status classifier to compare the equipment operating parameters in real time, and calculate its abnormal probability distribution matrix to determine the abnormal risk level of the current equipment operation;

[0110] like Figure 3 The figure shows a flow chart of the abnormal risk level determination method provided by this application. The specific method of using the operating state classifier to compare the equipment operating parameters in real time and calculate its abnormal probability distribution matrix to determine the abnormal risk level of the current equipment operation is as follows:

[0111] Step S310: Standardizing the equipment operating parameters through data preprocessing to obtain standardized characteristic data;

[0112] Step S320: using a running state classifier based on a support vector machine algorithm to perform state classification processing on the standardized feature data, if the standardized feature data matches a normal running mode, outputting a normal state identifier, if the standardized feature data deviates from the normal running mode, outputting an abnormal state identifier and a deviation degree value;

[0113] The input of the running state classifier is the standardized feature data, and the output is a running state identifier of the standardized feature data; the running state identifier includes a normal state identifier and an abnormal state identifier;

[0114] The model expression of the running state classifier is: Wherein, x is the input standardized feature data, x i is a support vector, y i is the output running state identifier, K(x i , x) is a kernel function, b is a bias term, i represents the current i-th support vector being traversed, and N represents the total number of support vectors;

[0115] The deviation degree value represents the distance of the standardized feature data from the hyperplane;

[0116] Step S330: calculating a device state abnormal probability through a Bayesian probability calculation device according to the abnormal state identifier and the deviation degree value, if the deviation degree value exceeds a deviation threshold, calculating an abnormal probability density function, and constructing an abnormal probability distribution matrix based on the abnormal probability density function;

[0117] The calculation method of the abnormal probability distribution matrix is: for the input standardized feature data x and its deviation degree value d, using a Bayesian probability calculation device to calculate an abnormal probability, if the deviation degree value exceeds a preset deviation threshold, using a normal distribution to calculate an abnormal probability density function of the deviation degree value, and constructing an abnormal probability distribution matrix according to the abnormal probability density function;

[0118] The calculation formula of the abnormal probability density function is: Wherein, sigma represents the standard deviation of the normal distribution, and mu represents the mean value of the normal distribution;

[0119] Step S340: performing risk level quantification processing through the abnormal probability distribution matrix, if the abnormal probability distribution is concentrated in a high-risk interval, determining the risk level as a high-risk state, if the abnormal probability distribution is concentrated in a medium-risk interval, determining the risk level as a pre-warning state;

[0120] The high-risk interval is [0.6, +∞), the medium-risk interval is [0.3, 0.6), and the low-risk interval is (-∞, 0.3];

[0121] Step S350: Time series analysis of the risk level is performed using a sliding window mechanism to obtain risk level trend data, and if the risk level shows an upward trend, a state deterioration early warning signal is triggered;

[0122] The triggering method of the state deterioration early warning signal is: defining a window length and a sliding step, for the risk value sequence in each window, calculating its first-order difference sequence to determine whether the risk level shows an upward trend, if the first-order difference sequence is greater than 0 for three consecutive times, it is determined that the risk level shows an upward trend, and a state deterioration early warning signal is triggered;

[0123] Specifically, for the technical content of constructing a micro device state monitoring system based on a unified feature data set, the specific implementation method can be integrated into a complete process. First, the unified feature data set includes temperature, vibration frequency and current, the data sampling frequency is once per minute, the data set has been standardized, the temperature range is 20 to 80 degrees Celsius, the vibration frequency is 10 to 50 Hz, and the current is 5 to 20 amperes. The system first constructs a running state classifier through historical data, uses a support vector machine algorithm, and divides the device state into three categories: normal, slight abnormality and serious abnormality. The training data of the classifier is based on 10,000 records in the past month, of which the normal state accounts for 70%, the slight abnormality accounts for 20%, and the serious abnormality accounts for 10%. Through cross-validation, the accuracy rate of the classifier reaches 92.5%. Then, the system compares the real-time collected device running parameters, for example, the current collected data is temperature 45.3 degrees Celsius, vibration frequency 28.7 Hz, and current 12.4 amperes, which is input into the classifier, and the output current state is normal, but further analysis of the abnormal probability is required. Immediately, the system calls the probability distribution calculation engine, adopts a Gaussian distribution model, calculates the abnormal probability of each feature, assumes that the temperature mean is 40 degrees Celsius, the standard deviation is 5, the vibration frequency mean is 25 Hz, the standard deviation is 3, and the current mean is 10 amperes, the standard deviation is 2, the temperature abnormal probability is 0.21, the vibration frequency is 0.18, and the current is 0.27. The weighted average is obtained. The comprehensive abnormal probability is 0.22. Finally, the system determines that the current state is a low risk level according to the preset threshold, stores the result in the database, and triggers the log record, if the probability exceeds 0.3, an early warning information is automatically generated and pushed to the monitoring platform, ensuring the closed-loop logic of real-time monitoring and feedback.

[0124] Step S400: Extracting abnormal segment data from the unified feature data set according to the state deterioration early warning signal and performing segmentation processing to obtain a segmented abnormal data set, and performing future trajectory deduction on the segmented abnormal data set to determine the future abnormal development trajectory of the potential fault;

[0125] The specific method for extracting abnormal segment data from the unified feature data set according to the state deterioration early warning signal and performing segmentation processing to obtain a segmented abnormal data set, and performing future trajectory deduction on the segmented abnormal data set to determine the future abnormal development trajectory of the potential fault is as follows:

[0126] Step S410: Extracting corresponding abnormal segment data from the unified feature data set for the state deterioration early warning signal;

[0127] Step S420: According to the abnormal segment data, applying a data screening rule for secondary filtering, removing segments that do not meet the duration condition through timestamp comparison, and determining selected abnormal segment data;

[0128] The segment that does not meet the duration condition is removed through timestamp comparison in order to remove abnormal segment data that does not meet the duration condition according to the characteristic of short abnormal duration, so as to retain selected abnormal segments.

[0129] Step S430: Cutting the selected abnormal segment data according to a preset time interval to obtain segmented abnormal data segments;

[0130] The selected abnormal segment data is cut according to the preset time interval in order to ensure that each segment has an independent time identifier;

[0131] Step S440: Extracting feature data of each segmented abnormal data segment, the feature data being a statistical indicator of each segmented abnormal data segment, the statistical indicator including mean and standard deviation, and obtaining feature abnormal data;

[0132] Step S450: Classifying the feature abnormal data using a support vector machine algorithm, and if the feature data of a certain segmented abnormal data segment matches a preset abnormal pattern, marking it as a high-risk abnormal segment;

[0133] Step S460: Distinguishing the high-risk abnormal segment from the ordinary abnormal segment to generate a segmented abnormal data set;

[0134] The segmented abnormal data set is an abnormal segment sequence containing historical abnormal changes;

[0135] Step S470: Using an LSTM algorithm to perform future trajectory deduction on the segmented abnormal data set to predict the future abnormal development trajectory;

[0136] The specific method for predicting the future abnormal development trajectory by using the LSTM algorithm on the segmented abnormal data set is: presetting a window length Q and a prediction step, using a sliding window method to obtain training samples from the segmented abnormal data set, the input of each training sample being the segmented abnormal data at the previous Q time points, the output data being the segmented abnormal data at the Q+1 time point, using mean square error as the loss function for training the LSTM model, and minimizing the value of the loss function as the training target; predicting the segmented abnormal data at each future time point based on the trained LSTM model, and the segmented abnormal data at each future time point forming the future abnormal development trajectory.

[0137] Specifically, in the process of processing abnormal data screening and segmentation, first, abnormal segment data is extracted from the unified feature data set. The sensor data set contains 10,000 time series data, and each data records the vibration frequency of the device running, with a sampling frequency of 10 times per second. The abnormal confirmation signal is set as the case where the vibration frequency exceeds 50Hz. By using a threshold-based screening algorithm, the data set is traversed, and the data points with vibration frequency greater than 50Hz are marked as abnormal, and the corresponding time segment data is extracted, for example, in the 2000th to 2050th data interval, it is found that the vibration frequency continuously exceeds 50Hz for 5 seconds. Then, in order to solve the problem of short abnormal duration and difficult to capture, a time window division strategy is used to segment the abnormal segment data. The time window is set to 1 second, and the sliding step is set to 0.5 second. By using a loop algorithm, the 5-second abnormal segment is divided into 8 overlapping 1-second window segments, each window containing 10 data points. The mean and standard deviation of the vibration frequency in each window are calculated. It is found that the mean of the third window is 52.3Hz and the standard deviation is 1.2, which is significantly higher than that of other windows, indicating that the abnormal characteristics in this window are more significant. Then, these segmented data are integrated into a segmented abnormal data set and stored as a structured table, containing fields such as window number, start time, end time, mean and standard deviation, so as to facilitate subsequent analysis. Through the above method, from data extraction to segmentation processing to correlation analysis, a complete technical chain is formed, ensuring accurate capture and effective analysis of abnormal data.

[0138] Step S500: determining the abnormal point of the device based on the future abnormal development trajectory, and associating and binding the abnormal development trajectory with the device running parameter of the current abnormal point;

[0139] The specific method for determining the abnormal point of the device based on the future abnormal development trajectory and associating and binding the abnormal development trajectory with the device running parameter of the current abnormal point is:

[0140] Step S510: determining the corresponding sensor signal based on the abnormal development trajectory to obtain the abnormal point of the device, and associating the device running parameter at the abnormal point with the abnormal development trajectory.

[0141] Step S520: Make corresponding decisions on the equipment at the current abnormal point based on the abnormal segment risk level of the abnormal development trajectory. If the future abnormal development trajectory is a high-risk abnormal segment, send an early warning to the maintenance personnel. If the future abnormal development trajectory is an ordinary abnormal segment, the abnormal point will be taken as a key monitoring object and continuously monitored.

[0142] Example 2

[0143] like Figure 4 As shown in the figure, the management system of the smart power plant based on digital twin and AI technology provided by this application includes:

[0144] The initial signal acquisition module is used to build a sensor signal acquisition architecture, obtain the original sensor signal, denoise and enhance the original sensor signal, and obtain an initial signal data set consisting of equipment operating parameters;

[0145] The feature vector extraction module is used to perform hierarchical analysis on the initial signal data set, map the equipment operating parameters to the feature vector space, and generate a unified feature data set;

[0146] The abnormal risk assessment module is used to compare the equipment operating parameters in real time using the operating status classifier, calculate its abnormal probability distribution matrix, and determine the abnormal risk level of the current equipment operation;

[0147] The future trajectory prediction module is used to extract abnormal fragment data from the unified feature data set according to the state deterioration warning signal and perform segmented processing to obtain a segmented abnormal data set. The module then performs future trajectory prediction on the segmented abnormal data set to determine the future abnormal development trajectory of the potential fault.

[0148] The trajectory positioning processing module is used to determine the abnormal point of the equipment based on the future abnormal development trajectory, and associate and bind the abnormal development trajectory with the equipment operating parameters of the current abnormal point.

[0149] In addition, the parts of the above technical solutions provided in the embodiments of the present application that are consistent with the implementation principles of the corresponding technical solutions in the prior art are not described in detail to avoid excessive redundancy.

[0150] The above-described specific embodiments further illustrate the objectives, technical solutions, and beneficial effects of the present invention. It should be understood that the above description is merely a specific embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principles of the present invention shall be included within the scope of protection of the present invention.

Claims

1. A management method for smart power plants based on digital twins and AI technology, characterized by: include: Build a sensor signal acquisition architecture, obtain raw sensor signals, denoise and enhance the raw sensor signals, and obtain an initial signal dataset consisting of device operating parameters; Perform hierarchical analysis on the initial signal data set, map the equipment operating parameters to the feature vector space, and generate a unified feature data set; Use the operating status classifier to compare the equipment operating parameters in real time, calculate its abnormal probability distribution matrix, and determine the abnormal risk level of the current equipment operation; According to the state deterioration warning signal, abnormal fragment data is extracted from the unified feature data set and segmented to obtain a segmented abnormal data set. The future trajectory of the segmented abnormal data set is deduced to determine the future abnormal development trajectory of the potential fault; The abnormal point of the equipment is determined based on the future abnormal development trajectory, and the abnormal development trajectory is associated and bound with the equipment operating parameters of the current abnormal point.

2. The management method of a smart power plant based on digital twin and AI technology according to claim 1, characterized in that: The specific method of constructing a sensor signal acquisition architecture, acquiring raw sensor signals, denoising and enhancing the raw sensor signals, and obtaining an initial signal data set consisting of device operating parameters is as follows: Build a unified sensor signal acquisition architecture to obtain the original sensor signals of various sensors, establish a sensor classification mapping table based on sensor type and signal strength, and obtain the first data stream of classification identification; Evaluate the signal strength of the first data stream, and mark it as a weak signal if the signal strength is less than a preset threshold; and mark it as a normal signal if the signal strength is greater than or equal to the preset threshold; De-noising the weak signal through the filter bank, adjusting the filter cutoff frequency according to the de-noising level parameter, and obtaining a de-noised second data stream; The second data stream is enhanced by using a signal amplification technology, and an amplification factor is calculated according to the enhancement multiple parameter and the signal strength of the original sensor signal to obtain an enhanced third data stream; Obtaining the amplitude range and frequency distribution characteristics of each sensor signal in the third data stream, mapping the different sensor signals to a unified numerical range through normalization processing, and determining a standardized fourth data stream; The fourth data stream is evaluated using a data quality assessment method. If the data quality index of the fourth data stream meets the preset standard, the data record is retained. If the data quality index does not meet the preset standard, the abnormal record is eliminated to generate a standardized initial signal data set; the initial signal data set includes the equipment operating parameters processed based on the sensor signal.

3. The management method of a smart power plant based on digital twin and AI technology according to claim 2, characterized in that: The method for obtaining the enhanced third data stream is as follows: obtaining the signal strength of the original sensor signal, setting an enhancement multiple parameter, calculating an amplification factor based on the signal strength and the enhancement multiple parameter, and enhancing the second data stream according to the amplification factor using signal amplification technology to obtain the enhanced third data stream.

4. The management method of a smart power plant based on digital twin and AI technology according to claim 3 is characterized in that: The specific method of performing hierarchical analysis on the initial signal data set, mapping the device operating parameters to the feature vector space, and generating a unified feature data set is as follows: The initial signal data set is defined as the basic data stream, and the hierarchical analysis technology is used to perform structural decomposition on the basic data stream to obtain the signal hierarchical results; Based on the signal layering results after preliminary analysis, we apply abnormal feature extraction logic to process the operating parameters of each layer of equipment to determine the distribution of abnormal features in response to signal interference. If the abnormal feature distribution exceeds the preset abnormal feature distribution threshold, the device operating parameters of the corresponding signal layer are filtered to obtain the filtered signal data stream; The principal component analysis algorithm is used to reduce the dimension of the filtered signal data stream, map it to a unified eigenvector space, and judge the integration effect of the eigenvector; According to the integration effect of the feature vectors, the feature vectors that do not meet the integration standard are corrected by applying the standardized adjustment logic to obtain the adjusted feature vector set; Obtain the adjusted feature vector set, perform feature matching and fusion processing on the signal data streams of different sensor types, and determine the final unified feature data set.

5. The management method of a smart power plant based on digital twin and AI technology according to claim 4 is characterized in that: The specific method of applying abnormal feature extraction logic to process the operating parameters of each layer of equipment and determining the distribution of abnormal features is as follows: using a sliding window method to extract features of the operating parameters of each layer of equipment, presetting the sliding window size and sliding step size, and calculating the features of the equipment operating parameters in each window, wherein the features include the mean, standard deviation, maximum fluctuation amplitude and gradient change rate; calculating an anomaly detection index, comparing the anomaly detection index with the index threshold, and if the absolute value of the anomaly detection index is greater than the index threshold, then the equipment operating parameter is deemed abnormal, and the abnormal feature of the abnormal equipment operating parameter is marked as 1, otherwise it is marked as 0, thereby obtaining the distribution of abnormal features.

6. The management method of a smart power plant based on digital twin and AI technology according to claim 5, characterized in that: The specific method of using the operating status classifier to compare the equipment operating parameters in real time and calculate the abnormal probability distribution matrix thereof to determine the abnormal risk level of the current equipment operation is as follows: Standardize the equipment operating parameters through data preprocessing to obtain standardized characteristic data; An operating state classifier based on a support vector machine algorithm is used to perform state classification processing on the standardized feature data. If the standardized feature data matches the normal operating mode, a normal state identifier is output; if the standardized feature data deviates from the normal operating mode, an abnormal state identifier and a deviation degree value are output; Based on the abnormal state identification and deviation value, the abnormal probability of the equipment state is calculated through Bayesian probability. If the deviation value exceeds the deviation threshold, the abnormal probability density function is calculated and the abnormal probability distribution matrix is ​​constructed based on the abnormal probability density function. The risk level is quantified through the abnormal probability distribution matrix. If the abnormal probability distribution is concentrated in the high-risk range, the risk level is determined to be high-risk. If the abnormal probability distribution is concentrated in the medium-risk range, the risk level is determined to be a warning state. A sliding window mechanism is used to conduct time series analysis of risk levels to obtain risk level change trend data. If the risk level shows an upward trend, a status deterioration warning signal is triggered.

7. The management method of a smart power plant based on digital twin and AI technology according to claim 6, characterized in that: The triggering method of the state deterioration warning signal is as follows: a window length and a sliding step size are defined; for the risk value sequence within each window, the first-order difference sequence is calculated to determine whether the risk level is on an upward trend; if the first-order difference sequence is greater than 0 three times in a row, it is determined that the risk level is on an upward trend, and a state deterioration warning signal is triggered.

8. The management method of a smart power plant based on digital twin and AI technology according to claim 7, characterized in that: The specific method of extracting abnormal fragment data from the unified feature data set according to the state deterioration warning signal and performing segmentation processing to obtain a segmented abnormal data set, performing future trajectory deduction on the segmented abnormal data set, and determining the future abnormal development trajectory of the potential fault is as follows: For the early warning signal of state deterioration, the corresponding abnormal fragment data is extracted from the unified feature data set; Based on the abnormal fragment data, the data screening rules are applied to perform secondary filtering, and the fragments that do not meet the duration conditions are removed through timestamp comparison to determine the selected abnormal fragment data; Segment the selected abnormal fragment data according to the preset time interval to obtain segmented abnormal data fragments; Extracting characteristic data of each segmented abnormal data fragment, wherein the characteristic data is a statistical index of each segmented abnormal data fragment, wherein the statistical index includes a mean and a standard deviation, to obtain characterized abnormal data; The support vector machine algorithm is used to classify the characterized abnormal data. If the characteristic data of a segmented abnormal data segment matches the preset abnormal pattern, it is marked as a high-risk abnormal segment; Distinguish high-risk abnormal segments from common abnormal segments and generate segmented abnormal datasets; The LSTM algorithm is used to deduce the future trajectory of the segmented anomaly data set and predict the future anomaly development trajectory.

9. The management method of a smart power plant based on digital twin and AI technology according to claim 8, characterized in that: The specific method of using the LSTM algorithm to perform future trajectory deduction on the segmented anomaly data set and predict the future anomaly development trajectory is as follows: presetting a window length Q and a prediction step size, using a sliding window method to obtain training samples from the segmented anomaly data set, the input of each training sample is the segmented anomaly data of the previous Q moments, and the output data is the segmented anomaly data of the Q+1th moment, using the mean square error as the loss function for training the LSTM model, and minimizing the value of the loss function as the training goal; based on the trained LSTM model, predicting the segmented anomaly data at each future moment, the segmented anomaly data at each future moment constitutes the future anomaly development trajectory.

10. A management system for a smart power plant based on digital twins and AI technology, which is implemented based on the management method for a smart power plant based on digital twins and AI technology according to any one of claims 1 to 9, characterized in that: include: The initial signal acquisition module is used to build a sensor signal acquisition architecture, obtain the original sensor signal, denoise and enhance the original sensor signal, and obtain an initial signal data set consisting of equipment operating parameters; The feature vector extraction module is used to perform hierarchical analysis on the initial signal data set, map the equipment operating parameters to the feature vector space, and generate a unified feature data set; The abnormal risk assessment module is used to compare the equipment operating parameters in real time using the operating status classifier, calculate its abnormal probability distribution matrix, and determine the abnormal risk level of the current equipment operation; The future trajectory prediction module is used to extract abnormal fragment data from the unified feature data set according to the state deterioration warning signal and perform segmented processing to obtain a segmented abnormal data set. The module then performs future trajectory prediction on the segmented abnormal data set to determine the future abnormal development trajectory of the potential fault. The trajectory positioning processing module is used to determine the abnormal point of the equipment based on the future abnormal development trajectory, and associate the abnormal development trajectory with the equipment operating parameters of the current abnormal point.

Images