A new energy power station centralized monitoring method based on multi-source data fusion

Abstract

The application relates to the technical field of new energy power station monitoring, and particularly discloses a new energy power station centralized monitoring method based on multi-source data fusion, which collects vibration sequences, gear box rotating speeds, generator powers and oil temperature data; the vibration sequences are divided into waveform frames, input into a lightweight time sequence convolution network composed of a depth separable convolution and an expansion convolution, and early pitting probability and main frequency offset are output; stable fault confidence is generated through buffer pool weighted sliding voting, and fault characteristic frequencies are obtained through median filtering; the rotating speeds, the powers, the oil temperatures and the fault characteristic frequencies are combined into a fusion feature vector, input into a decision rule tree, and an alarm level is output; when early warning is triggered, the data is uploaded for network iteration updating; the application retains full-band information under the condition that edge computing power is limited, improves early fault detection sensitivity, and improves model adaptability through incremental updating.

Description

Technical Field

[0001] This invention relates to the field of new energy power plant monitoring technology, specifically to a centralized monitoring method for new energy power plants based on multi-source data fusion. Background Technology

[0002] Centralized monitoring systems for new energy power plants (such as wind farms) typically rely on multi-source data fusion technology to aggregate vibration monitoring data, SCADA operating parameters, and meteorological data to a regional control center for analysis. In existing solutions, edge computing nodes are deployed at the power plant level to extract features from high-frequency signals collected by vibration sensors through Fourier transforms and envelope spectrum analysis, and then fuse these features with data such as rotational speed and power output to achieve real-time detection of early gearbox faults. New energy power plants, primarily wind farms and photovoltaic power plants, are usually distributed across different geographical areas. The control center needs to conduct unified remote monitoring, operation scheduling, and fault early warning for the power generation equipment, substations, and transmission lines of these power plants. Among these, rotating components such as gearboxes and generators in wind turbine units are high-risk areas for faults, and the efficient processing and accurate identification of their vibration monitoring data at the edge is a crucial step for the control center to achieve early warning.

[0003] Although the central control center can aggregate various types of data such as SCADA, meteorology, and electrical energy, existing edge computing solutions still face computing power bottlenecks in processing high-frequency vibration signals. In existing technologies, edge computing nodes, due to limited computing power, are forced to downsampling or truncating high-frequency components to meet real-time requirements when processing high-frequency vibration data, leading to the loss of key features of early pitting corrosion in gearboxes. Simultaneously, offline-trained monitoring models struggle to adapt to dynamic changes in actual operating conditions and lack sufficient penalty for missed faults, resulting in low fault identification sensitivity and poor generalization ability across power plants. Summary of the Invention

[0004] The purpose of this invention is to provide a centralized monitoring method for new energy power plants based on multi-source data fusion, so as to solve the problems mentioned above.

[0005] The objective of this invention can be achieved through the following technical solutions:

[0006] A centralized monitoring method for new energy power plants based on multi-source data fusion includes the following steps:

[0007] S1: Collect the raw time-domain vibration sequence output by the wind turbine vibration monitoring system, as well as gearbox speed, generator power and oil temperature data;

[0008] S2: The original temporal vibration sequence is divided into continuous original waveform frames according to the sliding time window. Each original waveform frame is input into a lightweight temporal convolutional network. The lightweight temporal convolutional network is composed of alternating stacks of depthwise separable convolution and dilated convolution. The output is the probability of early pitting and the corresponding main frequency shift estimate of this frame.

[0009] S3: The probability of early pitting and the estimated main frequency offset of multiple consecutive frames are fed into the confidence accumulation buffer. A stable fault confidence is generated by weighted sliding voting. The main frequency offset is then filtered by median filtering to output the fault characteristic frequency.

[0010] S4: Combine the stable fault confidence, fault characteristic frequency, and gearbox speed, generator power, and oil temperature data collected in the same time period into a fusion feature vector. Input the fusion feature vector into a pre-trained decision rule tree. Each node of the decision rule tree is composed of feature thresholds mined from historical samples, and outputs the corresponding alarm level.

[0011] S5: Trigger local early warning instructions based on the alarm level, and simultaneously package and upload the original waveform frames, fused feature vectors and alarm levels before and after the triggering of the early warning to the central control center for iterative updates of the lightweight temporal convolutional network.

[0012] As a further aspect of the present invention: the step of inputting each original waveform frame into a lightweight temporal convolutional network specifically includes:

[0013] Perform depthwise separable convolution operations on the original waveform frames to extract temporal local impulse features;

[0014] The obtained feature maps are sequentially passed through three layers of dilated convolution with increasing dilation rates to progressively enlarge the receptive field in order to capture fault impulse responses at different scales.

[0015] The output feature maps of each layer are fused element by element according to preset weights, and the eight channels with the highest response intensity are pruned and retained.

[0016] Global average pooling is performed on the retained channels, and the probability of early pitting is output through linear mapping. At the same time, the center frequency of the dilated convolution kernel corresponding to the channel with the largest response is taken as the main frequency offset estimate.

[0017] As a further aspect of the present invention: the obtained feature map is sequentially passed through three layers of dilated convolution with increasing dilation rates, specifically including:

[0018] The feature map is input into the first dilated convolution layer, with a dilation rate of 2, and the first-level receptive field feature is output.

[0019] The first-level receptive field features are input into the second-level dilated convolution, with the dilation rate of the second layer set to 4, and the second-level receptive field features are output.

[0020] The second-level receptive field features are input into the third-level dilated convolution, with the dilation rate of the third layer set to 8, and the third-level receptive field features are output.

[0021] As a further aspect of the present invention: the global average pooling of the retained channels, followed by linear mapping to output the probability of early pitting, specifically includes:

[0022] Extract the maximum response value, minimum response value, and arithmetic mean from the feature map of each retained channel to form a ternary statistical set for the corresponding retained channel;

[0023] Based on a pre-calibrated fuzzy measure, fuzzy integral operations are performed on the ternary statistical group to generate channel fuzzy features;

[0024] Arrange all channel fuzzy features into a feature sequence according to channel order;

[0025] The feature sequence is input into a nonlinear mapping network with a single intermediate node, and the probability of early pitting is output.

[0026] As a further aspect of the present invention: the step of generating a stable fault confidence level through weighted sliding voting and performing median filtering on the main frequency offset estimate specifically includes:

[0027] An exponentially decaying weight is assigned to the early pitting erosion of five consecutive frames in the buffer pool, with the weight value being larger the closer the frame time.

[0028] The probability of each frame is multiplied by its corresponding weight, then summed and divided by the total weight to obtain the stable fault confidence.

[0029] Meanwhile, the estimated main frequency offset values ​​of the five frames are arranged in ascending order of value, and the estimated value at the middle position is taken as the output fault characteristic frequency.

[0030] When the probability of early pitting in the latest frame is less than 0.3, immediately clear all historical probability and offset estimates in the cache pool.

[0031] As a further aspect of the present invention: the process of constructing the decision rule tree is as follows:

[0032] Collect a historical sample set containing stable fault confidence, fault characteristic frequency, gearbox speed, generator power and oil temperature and corresponding actual alarm levels;

[0033] Weighted Gini impurity is used as the evaluation index for node splitting, where the penalty weight for missed samples is higher than that for false positive samples.

[0034] Iterate through the value of each sample on each feature, calculate the decrease in weighted Gini impurity of the two subsets after splitting when the value is used as the splitting threshold, and select the feature with the largest decrease and its threshold as the basis for splitting the current node.

[0035] The complete decision rule tree structure is generated until the alarm level of the sample within the node is unique or the number of samples is less than five.

[0036] As a further aspect of the present invention: the use of weighted Gini impurity as an evaluation index for node splitting specifically includes:

[0037] Each sample is assigned a baseline weight of 1.

[0038] When the true alarm level of a sample is higher than the warning level and the output level of the sample in the current node split is lower than the true level, it is judged as a missed sample. Based on the frequency band to which the fault characteristic frequency of the missed sample belongs, a penalty multiplier of three to five times is added on the basis of the benchmark weight.

[0039] When the actual alarm level of a sample is lower than the warning level but the output level is higher than the actual level, it is judged as a false alarm sample and a penalty multiplier of one is added.

[0040] When calculating the weighted Gini impurity, the final penalty weight of each sample is used as its contribution coefficient in the split evaluation.

[0041] As a further aspect of the present invention: the calculation of weighted Gini impurity specifically includes:

[0042] Based on the final penalty weight of each sample, the number of weighted samples corresponding to each alarm level within the current node is counted.

[0043] Add up the weighted sample counts for all alarm levels to get the total weighted sample count for the node;

[0044] Divide the weighted sample size for each alarm level by the total weighted sample size for each node to obtain the weighted probability for each level.

[0045] Calculate the sum of squares of each weighted probability, and subtract the sum of squares from the value 1 to obtain the basic impurity;

[0046] Obtain the ratio of the interquartile range to the median of the fault characteristic frequencies of all samples within a node. When the ratio exceeds 0.4, multiply the basic impurity by the ratio, and use the product as the final weighted Gini impurity.

[0047] As a further aspect of the present invention: S5 specifically includes:

[0048] When the alarm level reaches the warning threshold, the original waveform frames of the current frame and the two frames before and after it are extracted to form a short time segment.

[0049] Short time segments, corresponding fused feature vectors and alarm levels are compressed using run-length encoding, and then uploaded to the central control center after adding edge node identifiers and collection timestamps.

[0050] After decompression, the central control center marks the data packets as positive samples and stores them in the incremental training pool at a ratio of 1:3 with historical negative samples.

[0051] Every 100 positive samples are accumulated, a fine-tuning training of the lightweight temporal convolutional network is initiated, and the updated network parameters are distributed to the edge nodes in the form of differences.

[0052] The beneficial effects of this invention are:

[0053] (1) Under the condition of limited computing power of edge computing nodes, by adopting a lightweight temporal convolutional network composed of alternating stacks of depth separable convolution and dilated convolution, the full-band information of the original vibration signal is preserved, avoiding the problem of loss of early pitting features due to downsampling or truncation of high frequency in traditional methods. Thus, the detection sensitivity of early faults of high-speed shaft of gearbox is improved while meeting the real-time processing delay requirements.

[0054] (2) By constructing a decision rule tree with weighted Gini impurity as the splitting index and adding a penalty weight higher than that for false alarm samples to the missed samples, the trained rule tree is more sensitive to the missed situation, reducing the risk of faults being missed in actual operation and maintenance; at the same time, combined with the incremental collection of positive samples and differential update of network parameters by the central control center, the monitoring model is continuously optimized in the actual operating environment, improving the adaptability of different wind farms and different operating conditions. Attached Figure Description

[0055] The invention will now be further described with reference to the accompanying drawings.

[0056] Figure 1 This is a flowchart of the method of the present invention;

[0057] Figure 2 This is a flowchart of the process of inputting each original waveform frame into a lightweight temporal convolutional network in this invention.

[0058] Figure 3 This is a flowchart of the median filtering process for the main frequency offset estimation in this invention. Detailed Implementation

[0059] 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.

[0060] The embodiments of the present invention are described in detail using wind turbine generators as an example. However, those skilled in the art will understand that for other new energy power generation equipment containing rotating machinery (such as the drive device of a photovoltaic tracking system or the cooling fan of an energy storage system), under similar conditions of vibration, speed, power and temperature monitoring, the ideas of the present invention can also be adapted and implemented.

[0061] Please see Figure 1 As shown, this invention is a centralized monitoring method for new energy power plants based on multi-source data fusion, comprising the following steps:

[0062] S1: Collect the raw time-domain vibration sequence output by the wind turbine vibration monitoring system, as well as gearbox speed, generator power and oil temperature data;

[0063] S2: The original temporal vibration sequence is divided into continuous original waveform frames according to the sliding time window. Each original waveform frame is input into a lightweight temporal convolutional network. The lightweight temporal convolutional network is composed of alternating stacks of depthwise separable convolution and dilated convolution. The output is the probability of early pitting and the corresponding main frequency shift estimate of this frame.

[0064] S3: The probability of early pitting and the estimated main frequency offset of multiple consecutive frames are fed into the confidence accumulation buffer. A stable fault confidence is generated by weighted sliding voting. The main frequency offset is then filtered by median filtering to output the fault characteristic frequency.

[0065] S4: Combine the stable fault confidence, fault characteristic frequency, and gearbox speed, generator power, and oil temperature data collected in the same time period into a fusion feature vector. Input the fusion feature vector into a pre-trained decision rule tree. Each node of the decision rule tree is composed of feature thresholds mined from historical samples, and outputs the corresponding alarm level.

[0066] S5: Trigger local early warning instructions based on the alarm level, and simultaneously package and upload the original waveform frames, fused feature vectors and alarm levels before and after the triggering of the early warning to the central control center for iterative updates of the lightweight temporal convolutional network.

[0067] In S1, the raw time-domain vibration sequence output by the wind turbine vibration monitoring system, as well as gearbox speed, generator power, and oil temperature data, are collected, specifically including:

[0068] A piezoelectric accelerometer is installed at the high-speed shaft bearing housing of the gearbox of each wind turbine. The output signal of the accelerometer is conditioned by a charge amplifier and then continuously sampled at a sampling frequency of 12.8 kHz to obtain the original time-domain vibration sequence. The unit of the value of each sampling point is gravitational acceleration.

[0069] Simultaneously, a magnetoelectric speed sensor is installed at the input shaft end of the gearbox. This sensor outputs a pulse frequency signal, which is converted into a speed per minute (SPM) value via a frequency-to-digital converter and recorded as the gearbox speed. The data update frequency is 10 Hz. A power transmitter is installed on the generator output side. This transmitter converts the three-phase active power into a 4-20 mA DC current signal, which is then converted from analog to digital to obtain the generator power value. The update frequency is 1 Hz. A platinum resistance temperature sensor is embedded in the gearbox lubricating oil sump. The resistance value is measured by a bridge circuit and converted into a Celsius temperature value, recorded as the oil temperature data. The update frequency is 0.5 Hz.

[0070] All analog or digital signals output by the aforementioned sensors are connected to the programmable logic controller (PLC) in the main control cabinet of the wind turbine. The PLC scans and reads all data at a 1-second interval, and aligns the read gearbox speed, generator power, and oil temperature data with the original time-domain vibration sequence output by the vibration monitoring system according to the millisecond-level timestamps provided by the global positioning system. Then, the packaged data is transmitted to the memory cache of the edge computing node on the power station side via industrial Ethernet, awaiting subsequent processing steps.

[0071] Please see Figure 2 As shown in S2, the original temporal vibration sequence is divided into continuous original waveform frames according to a sliding time window. Each original waveform frame is input into a lightweight temporal convolutional network, which is composed of alternating stacks of depthwise separable convolutions and dilated convolutions. The output includes the probability of early pitting and the corresponding main frequency shift estimate for this frame, specifically including:

[0072] When the original temporal vibration sequence is divided into continuous original waveform frames by a sliding time window, the width of the sliding time window is set to 0.1 seconds, corresponding to 1280 sampling points. The overlap rate between adjacent frames is 50%, that is, a frame is cut out every 0.05 seconds. Before each original waveform frame is fed into the lightweight temporal convolutional network, it is first normalized: the arithmetic mean and standard deviation of all sampling points in the frame are calculated, and then the arithmetic mean and standard deviation of each sampling point are subtracted from the average value and divided by the standard deviation, so that the mean of the entire frame data is 0 and the standard deviation is 1.

[0073] A depthwise separable convolution operation is performed on the normalized original waveform frame. This operation consists of two steps:

[0074] The first step is to use three one-dimensional convolution kernels to perform channel-by-channel convolution on the frame data. Each convolution kernel has a length of 5 and a stride of 1, outputting intermediate feature maps for three channels. The convolution calculation method for each channel is as follows: multiply the five weight coefficients of the convolution kernel by the corresponding five consecutive sampling points in the input sequence, sum them, and then add a bias value.

[0075] The second step involves performing pointwise convolution on the three intermediate feature maps using a single convolutional kernel of length 1. This weighted sum of the response values ​​at the same location across the three channels is then activated by a linear rectified function, outputting a final depthwise separable convolutional feature map. This feature map has one channel and a length of 1280 points, the same as the input frame. In this process, the initial weights of the convolutional kernel are obtained through pre-training on a cloud platform using a large number of historical vibration samples, and the initial bias value is set to 0.

[0076] The feature map output from the depthwise separable convolution is sequentially passed through three layers of dilated convolutions with increasing dilation rates. The dilation rate of the first layer is set to 2, and the kernel length is 5. The calculation method is as follows: on the input feature map, five positions are selected at each time, including the current index position and two points to the left and right. The values ​​at these positions are multiplied by the corresponding five weights of the convolution kernel, summed, and then activated by a linear rectified function after adding a bias. The output is the first-level receptive field feature map, with a length of 1280 points and 8 channels.

[0077] The second layer of dilated convolution receives the first-level receptive field feature map, with a dilation rate of 4 and a kernel length of 5. During calculation, it selects 5 positions, including the current index and 4 points to the left and right, and outputs the second-level receptive field feature map, which still has a length of 1280 points and expands the number of channels to 16.

[0078] The third dilated convolutional layer receives the second-level receptive field feature map, with a dilation rate of 8 and a kernel length of 5, and outputs the third-level receptive field feature map, with a length of 1280 points and 32 channels. The first, second, and third-level receptive field feature maps are concatenated according to the channel dimension to obtain a feature map with a total of 8 + 16 + 32 = 56 channels.

[0079] The 56 channels are compressed back to 8 channels using a pointwise convolution kernel of length 1. The pointwise convolution is calculated as follows: for each position, the values ​​of the 56 channels are multiplied by the corresponding weight coefficients and then summed, plus a bias value. After activation by a linear rectified function, a multi-scale fused impulse response map is obtained with 8 channels and a length of 1280 points.

[0080] The multi-scale fused impulse response map was pruned, retaining the 8 channels with the highest response intensities. Response intensity was defined as the sum of the squares of the values ​​at all positions within each channel. Data from these 8 channels was extracted, while data from the remaining channels were discarded. Subsequently, global average pooling was performed on each retained channel, i.e., the arithmetic mean of 1280 positions within that channel was calculated, resulting in 8 averages.

[0081] The above eight average values ​​are further processed to output the probability of early pitting corrosion. The specific steps are as follows:

[0082] The first step is to extract the maximum response value, minimum response value, and arithmetic mean of the feature map of each channel for each channel, and form a ternary statistical set.

[0083] The second step involves performing fuzzy integral operations on the ternary statistical set based on the pre-calibrated fuzzy measure. The fuzzy measure is pre-calibrated as follows: 500 known sample frames containing early pitting faults and 500 normal sample frames are taken. For each sample frame, three statistical values ​​are calculated, and the weighting coefficients of each of the three statistical values ​​are fitted using the least squares method, denoted as... , , And satisfy Fuzzy integral value The calculation formula is:

[0084] ;

[0085] in, For the maximum response value, For the minimum response value, The arithmetic mean is given, and all three values ​​are pre-normalized to the [0,1] interval. , , The corresponding fuzzy measure value is obtained by fitting historical samples. In this formula, the smaller of each statistical value and its weight is taken first, and then the maximum value among these three smaller values ​​is taken as the channel fuzzy feature.

[0086] The third step is to arrange the fuzzy features of the 8 channels into an 8-dimensional feature sequence according to the channel number.

[0087] The fourth step involves inputting the feature sequence into a nonlinear mapping network with a single intermediate node, which consists of two layers:

[0088] The first layer is fully connected, which maps the 8-dimensional input to a 4-dimensional intermediate vector. Each output value is equal to the dot product of the input vector and the corresponding weight vector, plus the bias, and then activated by the hyperbolic tangent function.

[0089] The second layer is a single-node fully connected layer that maps the weighted sum of the 4-dimensional intermediate vectors to a range of 0 to 1 using a sigmoid function, thus obtaining the probability of early pitting erosion. The weights and biases of this nonlinear mapping network are also determined by pre-training on the cloud platform.

[0090] Simultaneously, after completing the three-layer dilated convolution, the channel with the strongest response intensity is identified from the eight channels of the multi-scale fused impulse response map, and the center frequency of the dilated convolution kernel corresponding to that channel is recorded. Since the convolution kernels of the three-layer dilated convolution are given physical meaning during training, the convolution kernel of each channel has a fixed center frequency (ranging from 0 to 6.4 kHz), and this center frequency is taken as the estimated value of the main frequency offset for output.

[0091] Please see Figure 3 As shown, in S3, the probability of early pitting and the estimated main frequency offset of multiple consecutive frames are fed into the confidence accumulation buffer. A stable fault confidence is generated through weighted sliding voting, and the main frequency offset estimate is subjected to median filtering to output the fault characteristic frequency, specifically including:

[0092] A confidence accumulation buffer is established in the memory of the edge computing node. This buffer stores the early pitting probability and the estimated main frequency offset corresponding to five consecutive raw waveform frames. After processing a new frame, the early pitting probability and the estimated main frequency offset of that frame are pushed into the buffer, while the oldest frame data in the buffer is discarded, so that the buffer always contains only the information of the most recent five frames.

[0093] When assigning exponential decay weights to the early pitting of five consecutive frames in the buffer pool, the weight values ​​are set sequentially from the closest to the furthest in time.

[0094] Specifically, the weight of the latest frame is set to 1, the weight of the previous frame is set to 1 / 2, the weight of the frame before that is set to 1 / 4, the weight of the frame before that is set to 1 / 8, and the weight of the oldest frame is set to 1 / 16. These five weight values ​​form a weight sequence that decreases exponentially.

[0095] When calculating the confidence of stable faults, the probability of early pitting in each frame is multiplied by its corresponding weight to obtain five weighted probability values.

[0096] The five weighted probability values ​​are summed together to obtain a total weighted probability. Simultaneously, the five weight values ​​themselves are also summed together to obtain a total weight.

[0097] The quotient obtained by dividing the sum of weighted probabilities by the sum of weights is the confidence level for a stable fault. For example, if the probability of the latest frame is 0.8, the previous frame is 0.7, the frame before that is 0.6, the frame before that is 0.5, and the oldest frame is 0.4, then the sum of weighted probabilities is 0.8 x 1 + 0.7 x 1 / 2 + 0.6 x 1 / 4 + 0.5 x 1 / 8 + 0.4 x 1 / 16, and the sum of weights is 1 + 1 / 2 + 1 / 4 + 1 / 8 + 1 / 16. Dividing the two gives the confidence level for a stable fault.

[0098] When performing median filtering on the estimated main frequency offset, the estimated main frequency offset values ​​of the five frames in the buffer pool are rearranged in ascending order to form an ascending sequence. Then, the value at the third position in this ascending sequence, i.e., the estimated value in the middle order, is taken as the output fault characteristic frequency. For example, if the five estimated values ​​are 120 Hz, 115 Hz, 125 Hz, 118 Hz, and 122 Hz, and after sorting they become 115, 118, 120, 122, and 125, then 120 Hz is taken as the output fault characteristic frequency.

[0099] When the probability of early pitting in the latest frame is less than 0.3, it is considered that the current vibration signal does not contain any fault characteristics, and all historical probability and historical offset estimates stored in the confidence accumulation buffer are immediately cleared.

[0100] After the clearing operation is completed, data from subsequent new frames will start accumulating again from the empty pool, avoiding interference from outdated data in subsequent judgments. The clearing threshold of 0.3 was determined based on the statistical results of a large number of fault-free samples. Under normal operating conditions, the typical fluctuation range of the probability of early pitting is between 0 and 0.25. Using 0.3 as the clearing threshold ensures that the buffer pool is only reset when the fault-free state is confirmed.

[0101] In S4, the stable fault confidence, fault characteristic frequency, and gearbox speed, generator power, and oil temperature data collected during the same time period are combined into a fused feature vector. This fused feature vector is then input into a pre-trained decision rule tree. Each node of the decision rule tree consists of feature thresholds mined from historical samples, and the corresponding alarm level is output, specifically including:

[0102] Historical sample sets are pre-collected in the cloud platform. Each sample includes stable fault confidence, fault characteristic frequency, gearbox speed, generator power, oil temperature, and the corresponding actual alarm level. The actual alarm level is divided into four levels: level 0 indicates no alarm, level 1 indicates attention, level 2 indicates warning, and level 3 indicates alarm. During sample collection, the above five feature values ​​are extracted from data packets uploaded from the edge nodes of each wind farm, and the operation and maintenance personnel label the actual alarm level according to the actual fault occurrence. At least 5000 samples are collected, including normal and fault samples under various operating conditions.

[0103] When constructing the decision rule tree, the weighted Gini impurity is used as the evaluation index for node splitting.

[0104] Assign a baseline weight to each historical sample, with an initial value of 1.

[0105] The weights are adjusted based on the actual situation of the sample: when the actual alarm level of the sample is higher than level 2 (i.e., the actual alarm is a warning or alarm) and the predicted output level of the sample during the splitting process of the current node is lower than its actual level, the sample is judged as a missed sample.

[0106] For missed detection samples, a penalty multiplier is applied based on the frequency band of their fault characteristic frequency, specifically divided into three frequency bands: a 3x penalty multiplier is applied when the fault characteristic frequency is below 100 Hz, a 4x penalty multiplier is applied when the fault characteristic frequency is between 100 Hz and 400 Hz, and a 5x penalty multiplier is applied when the fault characteristic frequency is above 400 Hz. The final penalty weight is equal to the baseline weight multiplied by the added penalty multiplier. When the true alarm level of a sample is lower than level 2 (i.e., the true alarm level is no alarm or notice) and the predicted output level of the sample is higher than its true level, the sample is judged as a false alarm sample, and a 1x penalty multiplier is applied, meaning the final penalty weight is twice the baseline weight. For samples that are neither missed nor false alarms, the final penalty weight remains at 1.

[0107] When calculating the weighted Gini impurity, the first step is to count the number of weighted samples for each alarm level (level 0, level 1, level 2, level 3) based on the final penalty weights of all samples within the current node. Specifically, for a given alarm level, the final penalty weights of all samples at that level are summed to obtain the weighted number of samples for that level.

[0108] Add the weighted sample counts for the four levels to obtain the total weighted sample count for the node. Then divide the weighted sample count for each alarm level by the total weighted sample count for the node to obtain the weighted probability of each level within the node.

[0109] For example, if the weighted sample size for level 0 is 200, level 1 is 100, level 2 is 50, level 3 is 10, and the total weighted sample size is 360, then the weighted probabilities for each level are approximately 0.556 for 200 divided by 360, approximately 0.278 for 100 divided by 360, approximately 0.139 for 50 divided by 360, and approximately 0.028 for 10 divided by 360.

[0110] To calculate the basic impurity: First, square each weighted probability to obtain four squared values. Sum these squared values ​​to get a sum of squares. Subtract this sum of squares from the value 1; the difference is the basic impurity. Next, obtain the fault characteristic frequency values ​​of all samples within the current node. Sort these frequency values ​​from smallest to largest, find the first quartile (the value at the 25th percentile after sorting) and the third quartile (the value at the 75th percentile). Calculate the interquartile range (IMR), which is equal to the third quartile minus the first quartile. Simultaneously, find the median (the value at the 50th percentile). Calculate the ratio of the IMR to the median. If this ratio exceeds 0.4, multiply the basic impurity by this ratio; the product is the final weighted Gini impurity. If the ratio is less than or equal to 0.4, the final weighted Gini impurity equals the basic impurity.

[0111] In constructing the decision rule tree, for the current node, each feature (stable fault confidence, fault feature frequency, gearbox speed, generator power, oil temperature) and the value of each sample on that feature are traversed, with each value serving as a candidate splitting threshold. For each candidate threshold, the samples within the current node are divided into a left subset and a right subset based on whether the feature value is less than the threshold. The weighted Gini impurity of the left and right subsets is calculated separately, and then the weighted sums are calculated according to the proportion of the total weight of each subset to the total weight of the parent node, yielding the weighted impurity after splitting. The weighted Gini impurity of the parent node is subtracted from the weighted impurity after splitting to obtain the impurity reduction corresponding to that threshold. After traversing all candidate thresholds, the feature with the largest impurity reduction and its corresponding threshold are selected as the splitting criterion for the current node.

[0112] Split the current node into two child nodes, left and right, and recursively execute the above process for each child node.

[0113] The recursive splitting terminates when: all samples within the current node have a unique alarm level (i.e., all are of the same level), or the number of samples within the current node is less than 5. When either condition is met, splitting stops, the node is marked as a leaf node, and the alarm level that appears most frequently within that node is recorded as the output level. The final generated decision rule tree is a binary tree structure, where each internal node stores a feature name and a splitting threshold, and each leaf node stores an alarm level. After the tree is built on the cloud platform, it is distributed to each edge node as a parameter file for real-time inference.

[0114] In S5, local early warning instructions are triggered based on the alarm level. Simultaneously, the original waveform frames before and after the alarm trigger, the fused feature vectors, and the alarm level are packaged and uploaded to the central control center for iterative updates of the lightweight temporal convolutional network. Specifically, this includes:

[0115] When the alarm level output by S4 reaches the warning threshold, the warning threshold is set to level 2 or level 3, that is, when the alarm level is warning or alarm, subsequent operations are triggered.

[0116] The current original waveform frame and the two frames before and after it are extracted from the memory cache of the edge computing node, for a total of five original waveform frames. Each frame contains 1280 sampling points, and the total number of sampling points is 6400. These five frames are arranged in chronological order to form a short time segment.

[0117] At the same time, the fused feature vectors corresponding to these five frames are extracted. Each fused feature vector contains five values: stable fault confidence, fault feature frequency, gearbox speed, generator power and oil temperature. There are a total of 25 values ​​for the five frames.

[0118] The aforementioned short time-series segments, 25 fusion feature values, and the alarm levels corresponding to these five frames are compressed using run-length encoding. The specific method of run-length encoding is as follows: the 6400 sample point values ​​in the short time-series segments are scanned in the order of appearance. Whenever consecutive identical values ​​are encountered, the value and its number of consecutive occurrences are recorded, forming a compressed sequence consisting of a value and a repetition count pair. For the fusion feature values ​​and alarm levels, due to significant numerical fluctuations, they are stored directly in their original numerical form.

[0119] After compression, an edge node identifier and a collection timestamp are appended to the header of the compressed data packet. The edge node identifier uses the unique number of the wind turbine, and the collection timestamp is the time point corresponding to the current frame, accurate to milliseconds.

[0120] Compressed data packets are uploaded to the central control center server via the 4G communication module built into the edge node using the message queue telemetry transmission protocol.

[0121] After receiving the data packet, the central control center performs a decompression operation: according to the run-length encoding rules, the repetition pairs in the compressed sequence are restored to the original continuous numerical sequence, the complete short time-series segment is recovered, and the fused feature values ​​and alarm levels are read out at the same time.

[0122] After decompression, the data packet is marked as a positive sample because its corresponding alarm level reaches the warning threshold, indicating that there is indeed an anomaly or fault during that period. At the same time, the central control center has pre-stored a large number of historical negative samples. Negative samples are defined as data packets collected under normal operating conditions, whose alarm level is always level 0 or 1 and which have been confirmed by maintenance personnel to have no faults.

[0123] To balance the training data, the current positive sample is stored in the incremental training pool along with three times the number of negative samples randomly selected from the historical negative sample database, with a positive sample to negative sample ratio of 1:3. For example, if one positive sample is uploaded, three negative samples are selected to match it and stored together in the incremental training pool.

[0124] In the incremental training pool, a fine-tuning training of the lightweight temporal convolutional network is initiated every 100 positive samples accumulated. The specific process of fine-tuning training is as follows: using the network parameters currently being used by the edge node as the initial values, the samples in the incremental training pool (containing 100 positive samples and 300 negative samples) are used as training data, and the network parameters are iteratively updated using the backpropagation algorithm for a total of 20 iterations, with a batch size of 32 samples in each round.

[0125] After training, the differences between the updated network parameters and the original network parameters are calculated, i.e., the change in each weight coefficient. These changes are then compiled into a difference parameter matrix. This difference parameter matrix has the same size as the original network parameters, but its numerical range is smaller, with most changes concentrated near zero.

[0126] The difference parameter matrix is ​​compressed using run-length encoding, and the compressed difference data packets are sent to the corresponding edge nodes via a message queue telemetry transmission protocol. After receiving the difference data packets, the edge nodes first decompress and restore the difference parameter matrix, then add the difference value to each element of the currently used network parameters to obtain the updated network parameters, which replace the original parameters, thus completing the online iterative update of the lightweight temporal convolutional network.

[0127] The working principle of this invention is as follows: The original time-domain vibration sequence, gearbox speed, generator power, and oil temperature data of the wind turbine are collected. The vibration sequence is divided into original waveform frames according to a sliding time window. A lightweight temporal convolutional network, composed of alternating stacks of depthwise separable convolutions and dilated convolutions, outputs the probability of early pitting and the estimated main frequency shift. The probability and shift estimates of multiple consecutive frames are fed into a confidence accumulation buffer. A stable fault confidence level is generated through weighted sliding voting, and the shift estimate is filtered by median to obtain the fault characteristic frequency. The stable fault confidence level, fault characteristic frequency, speed, power, and oil temperature are combined into a fused feature vector, which is input into a pre-trained decision rule tree to output an alarm level. A local warning is triggered based on the alarm level, and the waveform frames before and after the triggering time, the fused feature vector, and the alarm level are packaged and uploaded to the central control center for iterative updates of the lightweight temporal convolutional network.

[0128] The foregoing has provided a detailed description of one embodiment of the present invention, but this description is merely a preferred embodiment and should not be construed as limiting the scope of the invention. All equivalent variations and modifications made within the scope of the claims of this invention should still fall within the patent coverage of this invention.

Claims

1. A centralized monitoring method for new energy power plants based on multi-source data fusion, characterized in that, Includes the following steps: S1: Collect the raw time-domain vibration sequence output by the wind turbine vibration monitoring system, as well as gearbox speed, generator power and oil temperature data; S2: The original temporal vibration sequence is divided into continuous original waveform frames according to the sliding time window. Each original waveform frame is input into a lightweight temporal convolutional network. The lightweight temporal convolutional network is composed of alternating stacks of depthwise separable convolution and dilated convolution. The output is the probability of early pitting and the corresponding main frequency shift estimate of this frame. S3: The probability of early pitting and the estimated main frequency offset of multiple consecutive frames are fed into the confidence accumulation buffer. A stable fault confidence is generated by weighted sliding voting. The main frequency offset is then filtered by median filtering to output the fault characteristic frequency. S4: Combine the stable fault confidence, fault characteristic frequency, and gearbox speed, generator power, and oil temperature data collected in the same time period into a fusion feature vector. Input the fusion feature vector into a pre-trained decision rule tree. Each node of the decision rule tree is composed of feature thresholds mined from historical samples, and outputs the corresponding alarm level. S5: Trigger local early warning instructions based on the alarm level, and simultaneously package and upload the original waveform frames, fused feature vectors and alarm levels before and after the triggering of the early warning to the central control center for iterative updates of the lightweight temporal convolutional network.

2. The centralized monitoring method for new energy power plants based on multi-source data fusion according to claim 1, characterized in that, The step of inputting each original waveform frame into a lightweight temporal convolutional network specifically includes: Perform depthwise separable convolution operations on the original waveform frames to extract temporal local impulse features; The obtained feature maps are sequentially passed through three layers of dilated convolution with increasing dilation rates to progressively enlarge the receptive field in order to capture fault impulse responses at different scales. The output feature maps of each layer are fused element by element according to preset weights, and the eight channels with the highest response intensity are pruned and retained. Global average pooling is performed on the retained channels, and the probability of early pitting is output through linear mapping. At the same time, the center frequency of the dilated convolution kernel corresponding to the channel with the largest response is taken as the main frequency offset estimate.

3. The centralized monitoring method for new energy power plants based on multi-source data fusion according to claim 2, characterized in that, The obtained feature map is sequentially passed through three layers of dilated convolution with increasing dilation rates, specifically including: The feature map is input into the first dilated convolution layer, with a dilation rate of 2, and the first-level receptive field feature is output. The first-level receptive field features are input into the second-level dilated convolution, with the dilation rate of the second layer set to 4, and the second-level receptive field features are output. The second-level receptive field features are input into the third-level dilated convolution, with the dilation rate of the third layer set to 8, and the third-level receptive field features are output.

4. The centralized monitoring method for new energy power plants based on multi-source data fusion according to claim 2, characterized in that, The process of performing global average pooling on the retained channels and outputting the probability of early pitting through linear mapping specifically includes: Extract the maximum response value, minimum response value, and arithmetic mean from the feature map of each retained channel to form a ternary statistical set for the corresponding retained channel; Based on a pre-calibrated fuzzy measure, fuzzy integral operations are performed on the ternary statistical group to generate channel fuzzy features; Arrange all channel fuzzy features into a feature sequence according to channel order; The feature sequence is input into a nonlinear mapping network with a single intermediate node, and the probability of early pitting is output.

5. A centralized monitoring method for new energy power plants based on multi-source data fusion according to claim 1, characterized in that, The process of generating stable fault confidence through weighted sliding voting and performing median filtering on the main frequency offset estimate specifically includes: An exponentially decaying weight is assigned to the early pitting erosion of five consecutive frames in the buffer pool, with the weight value being larger the closer the frame time. The probability of each frame is multiplied by its corresponding weight, then summed and divided by the total weight to obtain the stable fault confidence. Meanwhile, the estimated main frequency offset values ​​of the five frames are arranged in ascending order of value, and the estimated value at the middle position is taken as the output fault characteristic frequency. When the probability of early pitting in the latest frame is less than 0.3, immediately clear all historical probability and offset estimates in the cache pool.

6. The centralized monitoring method for new energy power plants based on multi-source data fusion according to claim 1, characterized in that, The process of constructing the decision rule tree is as follows: Collect a historical sample set containing stable fault confidence, fault characteristic frequency, gearbox speed, generator power and oil temperature and corresponding actual alarm levels; Weighted Gini impurity is used as the evaluation index for node splitting, where the penalty weight for missed samples is higher than that for false positive samples. Iterate through the value of each sample on each feature, calculate the weighted Gini impurity decrease of the two subsets after splitting when the value is used as the splitting threshold, and select the feature with the largest decrease and its threshold as the basis for splitting the current node. The complete decision rule tree structure is generated until the alarm level of the sample within the node is unique or the number of samples is less than five.

7. A centralized monitoring method for new energy power plants based on multi-source data fusion according to claim 6, characterized in that, The use of weighted Gini impurity as an evaluation index for node splitting specifically includes: Each sample is assigned a baseline weight of 1. When the true alarm level of a sample is higher than the warning level and the output level of the sample in the current node split is lower than the true level, it is judged as a missed sample. Based on the frequency band to which the fault characteristic frequency of the missed sample belongs, a penalty multiplier of three to five times is added on the basis of the benchmark weight. When the actual alarm level of a sample is lower than the warning level but the output level is higher than the actual level, it is judged as a false alarm sample and a penalty multiplier of one is added. When calculating the weighted Gini impurity, the final penalty weight of each sample is used as its contribution coefficient in the split evaluation.

8. A centralized monitoring method for new energy power plants based on multi-source data fusion according to claim 7, characterized in that, The calculation of the weighted Gini impurity specifically includes: Based on the final penalty weight of each sample, the number of weighted samples corresponding to each alarm level within the current node is counted. Add up the weighted sample counts for all alarm levels to get the total weighted sample count for the node; Divide the weighted sample size for each alarm level by the total weighted sample size for each node to obtain the weighted probability for each level. Calculate the sum of squares of each weighted probability, and subtract the sum of squares from the value 1 to obtain the basic impurity; Obtain the ratio of the interquartile range to the median of the fault characteristic frequencies of all samples within a node. When the ratio exceeds 0.4, multiply the basic impurity by the ratio, and use the product as the final weighted Gini impurity.

9. A centralized monitoring method for new energy power plants based on multi-source data fusion according to claim 1, characterized in that, S5 specifically includes: When the alarm level reaches the warning threshold, the original waveform frames of the current frame and the two frames before and after it are extracted to form a short time segment. Short time segments, corresponding fused feature vectors and alarm levels are compressed using run-length encoding, and then uploaded to the central control center after adding edge node identifiers and collection timestamps. After decompression, the central control center marks the data packets as positive samples and stores them in the incremental training pool at a ratio of 1:3 with historical negative samples. Every 100 positive samples are accumulated, a fine-tuning training of the lightweight temporal convolutional network is initiated, and the updated network parameters are distributed to the edge nodes in the form of differences.

Images