A method, system, device and medium for determining the starting mode of a hydroelectric generating unit by fusing multi-modal features

Abstract

The present application relates to the technical field of hydroelectric unit control optimization, and discloses a method, system, device and medium for determining the starting mode of a hydroelectric unit by fusing multi-modal features, which comprises the following steps: collecting original vibration signals under different starting modes; preprocessing the original vibration signals to obtain two types of time-series vibration signals during the starting process; extracting frequency domain features of the first type of signals by Markov transition field, i.e. structural reparameterization large core network; extracting time domain features of the second type of signals by bidirectional gate recurrent unit and global attention mechanism; obtaining vibration features under different starting modes and vibration feature reference vectors under the original starting state by using global average pooling layer and weighted fusion mode; and obtaining the optimal starting mode based on the distance between the vibration feature vectors and the reference vectors. The present application extracts richer and more essential features from multiple dimensions, calculates the dynamic characteristics of the starting process, and makes the quantitative determination of the starting mode more accurate, so as to quickly respond to the demand.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of hydroelectric unit control optimization, and in particular to a hydroelectric unit starting mode determination method, system, device and medium fusing multi-modal features. BACKGROUND

[0002] Under the background of accelerating the construction of new power systems and large-scale grid connection of renewable energy, the operation efficiency and response capability of hydroelectric units as important flexible adjustment resources of power systems are directly related to the stability, economy and safety of power grids. Starting mode as the core link of hydroelectric unit operation control not only affects the rapidity and reliability of unit start-stop, but also determines the comprehensive performance of the unit in multiple scenarios such as peak load regulation, frequency regulation and emergency standby. A reasonable starting strategy can significantly reduce mechanical loss of the unit, prolong the service life of the equipment, and optimize the efficiency of power load distribution; while an improper starting mode may lead to response lag, increased energy loss, and even threaten the safe operation of the power grid. Especially under the challenges of frequent extreme weather and high proportion of new energy access, the optimization of hydroelectric unit starting mode has important value and practical significance for improving the resilience of the power system and supporting the transformation of energy structure.

[0003] For the transition process of hydroelectric units from static to no-load rated speed, studies have shown that this stage is a high-risk interval for unit operation, and its working condition conversion involves complex dynamic response. Scientific selection of the starting strategy plays a decisive role in ensuring the safe and smooth operation of the unit. Specifically, the starting process of the water turbine refers to the process of adjusting the guide vane from the full-closed state to the corresponding water head no-load opening interval by the speed regulation system, driving the runner to complete the hydraulic transition process from static to rated speed. Although the duration of the hydraulic transition process is much shorter than the normal operation period of the unit, the excitation effect of unstable water flow load during this process will cause the vibration amplitude of the unit to increase significantly. If the vibration is too large, it is easy to cause severe vibration of key components such as the spiral case, runner and draft tube, which seriously threatens the safe and stable operation of the hydroelectric unit. At present, the determination of the starting mode of the hydroelectric unit is mostly based on a single basis, and the description of multi-source vibration during the starting process is relatively vague, lacking a system for judging based on the vibration of the unit. SUMMARY

[0004] In view of the above existing problems, the present application is proposed.

[0005] Therefore, the present application provides a hydroelectric unit starting mode determination method, system, device and medium fusing multi-modal features to solve the problem of lack of objective, accurate and multi-dimensional quantitative judgment standard for the current hydroelectric unit starting mode.

[0006] To solve the above technical problems, the present application provides the following technical solutions:

[0007] In a first aspect, the present application provides a fusion multi-modal feature hydroelectric generating set starting mode determination method, comprising: collecting original vibration signals under different starting modes;

[0008] The original vibration signals are preprocessed to obtain two types of time sequence vibration signals in the starting process, including a first type of time sequence signal representing local features and a second type of time sequence signal representing long-term changes;

[0009] The first type of signal is converted into a time-frequency image through a Markov transition field, and a structural reparameterization large kernel network is used to extract frequency domain features at each time point;

[0010] The time sequence features of the second type of signal are extracted through a bidirectional gated recurrent unit, and the time domain features at each time point are extracted using a global attention mechanism;

[0011] The frequency domain features and the time domain features at each time point are multiplied, the global information is integrated using a global average pooling layer, and the adaptive fusion of the time-frequency domain features is performed through a weighted fusion method to obtain vibration features under different starting modes;

[0012] Through the preprocessing, time-frequency domain feature extraction and weighted fusion method, a vibration feature reference vector under the original starting state of the generating set is obtained;

[0013] The distance between the vibration feature vectors under different starting modes and the reference vector is calculated, and the optimal starting mode of the hydroelectric generating set is obtained based on the distance value.

[0014] As a preferred scheme of the fusion multi-modal feature hydroelectric generating set starting mode determination method, the first type of signal is converted into a time-frequency image through a Markov transition field, comprising:

[0015] The first type of signal is divided into segments;

[0016] For each segment of the first type of signal, the signal amplitude is divided according to a normal distribution to obtain hydroelectric generating set Markov regions , wherein the region is mapped with the local vibration signal ;

[0017] The vibration signals at different time points are discretized into state regions , the probability of the signal transition from region to region is calculated, and the probability is obtained to construct a Markov state transition matrix ;

[0018] Based on Markov state transition matrix A dynamic Markov transition field is adopted to generate a local vibration signal state transition matrix under multiple scales

[0019] The state transition matrix is mapped into an image, and an average is generated through a blur kernel Vibration signal MTF map.

[0020] As a preferred scheme of the fusion multi-modal feature hydroelectric generating set starting mode determination method, wherein: a structural re-parameterization large kernel network is used to extract frequency domain features at each time point, including:

[0021] An image frequency domain feature extraction architecture is built, the extraction architecture includes four stages, each stage contains RepLkBlock module, wherein, Each stage is sampled through cross-step convolution;

[0022] Each vibration signal MTF map is decomposed into vibration element maps using a multi-branch structure to construct the RepLkBlock module;

[0023] All RepLkBlock modules share the same set of large convolution kernel initial weights, and each module adjusts the kernel weight through a learnable scaling factor to train the iterative RepLkBlock module to obtain the optimal vibration element map;

[0024] Haar features, LBP features and HOG features of the vibration element map are calculated, and weighted fusion is performed to obtain the frequency domain features at each time point.

[0025] As a preferred scheme of the fusion multi-modal feature hydroelectric generating set starting mode determination method, wherein: the time sequence features of the second type of signal are extracted through a bidirectional gate recurrent unit, and the time domain features at each time point are extracted using a global attention mechanism, including:

[0026] The second type of signal is input into the bidirectional gate recurrent unit;

[0027] The reference vibration time domain signal feature parameters and the update parameters of the vibration time domain signal are calculated, and the time sequence features corresponding to each time point are output;

[0028] The correlation between the second type of signal and the time domain feature state of all initial states is calculated;

[0029] Based on the correlation, a feature screening module is constructed using an attention mechanism;

[0030] ​​The time sequence feature is input into the feature screening module to obtain a time domain feature at each time point.

[0031] As a preferred scheme of the fusion multi-modal feature hydroelectric generating set start mode determination method, the frequency domain feature at each time point is added, a global average pooling layer is used to integrate global information, and adaptive fusion of the time-frequency domain features is performed through a weighted fusion method, including:

[0032] A frequency domain feature matrix sequence is obtained based on the frequency domain feature at each time point.

[0033] The frequency domain feature matrix sequence and the time domain feature sequence are multiplied after normalization to obtain a time-frequency domain initial vector.

[0034] The time-frequency domain initial vectors obtained for each segmented signal segment are averaged.

[0035] A weight vector is generated by acquiring a relationship between the time-frequency domain fusion vector averages through a one-dimensional convolutional neural network.

[0036] The weight vector is multiplied by a time-frequency domain initial vector modulus to obtain a time-frequency domain fusion feature value of the vibration signal under different start modes.

[0037] As a preferred scheme of the fusion multi-modal feature hydroelectric generating set start mode determination method, the weight vector is generated by acquiring a relationship between the time-frequency domain fusion vector averages through a one-dimensional convolutional neural network, including:

[0038] The one-dimensional convolutional neural network uses two convolutional layers W1 and W2 structures.

[0039] The weight vector is calculated through a Sigmoid function.

[0040] As a preferred scheme of the fusion multi-modal feature hydroelectric generating set start mode determination method, the maximum distance between the vibration feature vectors under different start modes and the reference vector is the optimal start mode.

[0041] In a second aspect, the application provides a fusion multi-modal feature hydroelectric generating set start mode determination system, including:

[0042] The acquisition module is used to acquire original vibration signals under different start modes.

[0043] The preprocessing module is used to preprocess the original vibration signals to obtain two types of time sequence vibration signals in the start process, including a first type of time sequence signal representing local features and a second type of time sequence signal representing long-term changes.

[0044] The frequency domain feature extraction module is configured to convert the first type of signal into a time-frequency image through Markov transition field and extract frequency domain features at each time point by using a structure re-parameterization large kernel network.

[0045] The time domain feature extraction module is configured to extract time sequence features of the second type of signal through a bidirectional gated recurrent unit and extract time domain features at each time point by using a global attention mechanism.

[0046] The feature fusion module is configured to multiply the frequency domain features and the time domain features at each time point, integrate global information by using a global average pooling layer, and adaptively fuse the time-frequency domain features by using a weighted fusion manner to obtain vibration features under different starting modes.

[0047] The reference vector acquisition module is configured to obtain vibration feature reference vectors under the original starting mode of the unit by using the preprocessing, the time-frequency domain feature extraction and the weighted fusion manner.

[0048] The calculation and judgment module is configured to calculate distances between the vibration feature vectors under different starting modes and the reference vectors, and obtain the optimal starting mode of the hydroelectric generating unit based on the distance values.

[0049] In a third aspect, the present application provides a computer device, comprising:

[0050] a memory and a processor;

[0051] The memory is configured to store computer executable instructions, and the processor is configured to execute the computer executable instructions, so as to realize the steps of the method for determining the starting mode of the hydroelectric generating unit by fusing multi-modal features.

[0052] In a fourth aspect, the present application provides a computer readable storage medium, which stores computer executable instructions, and the computer executable instructions are executed by a processor to realize the steps of the method for determining the starting mode of the hydroelectric generating unit by fusing multi-modal features.

[0053] Compared with the prior art, the present application has the following beneficial effects: the present application converts one-dimensional signals into MTF images, captures spatial patterns and dependencies of vibration frequencies in a global range by using a super large receptive field of RepLkNet, and ensures high efficiency by re-parameterization; in combination with a bidirectional gated recurrent unit and a global attention mechanism, more abundant and essential features can be extracted from complex vibration signals; the dynamic characteristics of the starting process are calculated from multiple dimensions by fusing time domain features, frequency domain features and time-frequency domain features, the judgment result is more accurate, the evaluation method is quantized, the demand can be quickly responded, the starting strategy with the smallest damage to the unit is selected, and a basis is provided for intelligent operation and maintenance of the hydroelectric generating unit and safe and economic dispatching of the power grid. BRIEF DESCRIPTION OF DRAWINGS

[0054] In order to more clearly illustrate the technical solutions of the embodiments of the present application, the drawings needed to be used in the embodiments will be briefly introduced as follows. Obviously, the drawings in the following description are only some embodiments of the present application, and all other drawings obtained by those skilled in the art without creative effort based on these drawings should also fall within the protection scope of the present application.

[0055] Figure 1 The overall flowchart of the fusion multi-modal feature hydroelectric generating unit start-up mode determination method according to an embodiment of the present application.

[0056] Figure 2 The guide vane opening degree-time curve diagram in the fusion multi-modal feature hydroelectric generating unit start-up mode determination method according to an embodiment of the present application.

[0057] Figure 3 The unit speed-time curve diagram in the fusion multi-modal feature hydroelectric generating unit start-up mode determination method according to an embodiment of the present application.

[0058] Figure 4 The vibration trend signal diagram in the fusion multi-modal feature hydroelectric generating unit start-up mode determination method according to an embodiment of the present application.

[0059] Figure 5 The vibration feature signal diagram in the fusion multi-modal feature hydroelectric generating unit start-up mode determination method according to an embodiment of the present application.

[0060] Figure 6 The MFT-RepLkNet feature map at a certain time point in the fusion multi-modal feature hydroelectric generating unit start-up mode determination method according to an embodiment of the present application.

[0061] Figure 7 The fusion multi-modal feature results of different start-up modes in the fusion multi-modal feature hydroelectric generating unit start-up mode determination method according to an embodiment of the present application. DETAILED DESCRIPTION

[0062] In order to make the above purposes, features and advantages of the present application more obvious and easy to understand, the specific embodiments of the present application will be described in detail below with reference to the drawings in the specification. Obviously, the described embodiments are only some embodiments of the present application, and not all embodiments. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative effort should fall within the protection scope of the present application.

[0063] Embodiment 1, refer to Figure 1For an embodiment of the present application, a fusion multi-modal feature hydroelectric generating set starting mode determination method is provided, comprising:

[0064] S100: Collecting original vibration signals under different starting modes;

[0065] S200: Preprocessing the original vibration signals to obtain two types of time series vibration signals in the starting process, including a first type of time series signal representing local features and a second type of time series signal representing long-term changes;

[0066] S300: Converting the first type of signal into a time-frequency image through a Markov transition field, and extracting frequency domain features at each time point using a structure reparameterization large kernel network;

[0067] S400: Extracting time series features of the second type of signal through a bidirectional gated recurrent unit, and extracting time domain features at each time point using a global attention mechanism;

[0068] S500: Multiplying the frequency domain features and the time domain features at each time point, integrating global information using a global average pooling layer, and adaptively fusing the time-frequency domain features through a weighted fusion method to obtain vibration features under different starting modes;

[0069] S600: Obtaining vibration feature reference vectors under the original starting state of the generating set through the preprocessing, time-frequency domain feature extraction and weighted fusion methods;

[0070] S700: Calculating the distance between the vibration feature vectors under different starting modes and the reference vectors, and obtaining the optimal starting mode of the hydroelectric generating set based on the distance value.

[0071] It should be noted that the traditional selection of starting strategies largely depends on the historical experience and subjective judgment of engineers, and in most cases only individual parameters such as total starting time or maximum speed are considered, often ignoring the mechanical vibration state. The starting of the generating set is a fast, non-stationary and nonlinear transition process, and the vibration signal processing method is difficult to effectively capture its instantaneous and complex dynamic change characteristics and long-distance dependence in the entire process.

[0072] Therefore, in view of the above main problems, steps S100-S700 are set to extract vibration signals, adopt different time-frequency domain signal extraction and fusion methods, fully reflect the dynamic characteristics of the starting process from multiple dimensions, and make the determination method quantitative, so as to quickly respond to the demand.

[0073] Embodiment 2, refer to Figure 1 For an embodiment of the present application, based on the above embodiment, a fusion multi-modal feature hydroelectric generating set starting mode determination method is provided.

[0074] In the embodiment of the present application, the original vibration signals under different starting modes are collected in step S100. Specifically, the vibration acceleration sensor and the strain gauge are installed on the bearing seat, the unit top cover, the guide vane connecting rod, and the generator frame to collect the original vibration signals.

[0075] In the embodiment of the present application, the original vibration signals are preprocessed in step S200 to obtain two types of time series vibration signals in the starting process, including the first type of time series signal representing local features and the second type of time series signal representing long-term changes.

[0076] Specifically, the original vibration signals at multiple time points in the starting process are intercepted, and the intercepted signals are decomposed by using the variational mode decomposition (VMD) method to obtain the two types of time series vibration signals in the starting process.

[0077] In the embodiment of the present application, the first type of signal is converted into a time-frequency image by using the Markov transition field (MTF) in step S300, including the following steps A1-A5.

[0078] A1: The first type of signal is divided into segments;

[0079] Specifically, the first type of signal, i.e., the local vibration signal ; wherein is the number of local vibration signals;

[0080] A2: For each segment of the first type of signal, the signal amplitude is divided according to the normal distribution to obtain Markov regions of the hydroelectric generating unit , wherein the region is mapped with the local vibration signal .

[0081] Specifically, the normal distribution division strategy means that the statistical quantity contained in the divided region obeys the normal distribution.

[0082] A3: The vibration signals at different time points are discretized into state regions , the probability of the signal transition from the region to the region is calculated to obtain the probability is , so as to construct the Markov state transition matrix .

[0083] A4: Based on the Markov state transition matrix, a dynamic Markov transition field is used to generate the local vibration signal state transition matrix under multiple scales ;

[0084] It should be noted that the transformation probability of the matrix control only depends on the two time points before and after, so the dynamic MTF method is preferred to further generate the local vibration signal state transition matrix under multiple time scales .

[0085] Exemplarily, the MTF method generates a transition matrix which can be represented as:

[0086] (1);

[0087] wherein, in formula (1) represents the transition probability of the local vibration signal to ; represents the index of the Markov region (state), ranging from 1 to ; is the number of local vibration signals.

[0088] A5: Map the state transition matrix to an image, and generate vibration signal MTF images through fuzzy kernel averaging.

[0089] It should be noted that the first type of signal is divided into segments, because subsequent MTF directly generates images for multiple signals will result in too large size, which will cause too large storage required for subsequent calculation, so cutting is performed to avoid the problem of too large image size and calculation amount caused by directly generating MTF images for the entire long sequence.

[0090] In the embodiments of the present application, the frequency domain features at each time point are extracted by using a structure re-parameterized large kernel network (RepLkNet) in step S300, including the following steps B1-B4:

[0091] B1: Build an image frequency domain feature extraction architecture, which includes four stages, each stage containing a RepLkBlock, wherein , image sampling is performed between each stage through cross-step convolution;

[0092] Specifically, the sampling formula can be represented as:

[0093] (2);

[0094] wherein, denotes the convolution sampling with step 2; denotes the down-sampled vibration signal MTF feature map; i.e. the vibration signal MTF feature map of the last stage.

[0095] B2: using a multi-branch structure to decompose each vibration signal MTF map into vibration element maps to construct the RepLkBlock module;

[0096] Specifically, the decomposition process is represented as:

[0097] (3);

[0098] wherein, denotes the RepLkBlock module to be trained; denotes the large kernel convolution decomposition weight; denotes the large kernel convolution decomposition; denotes the point convolution decomposition; denotes the depth convolution decomposition; , denotes the decomposition residual parameter; denotes a local vibration signal.

[0099] Each vibration element map is shown in formula (4):

[0100] (4);

[0101] wherein, denotes the feature map set; , denotes the picture size height and width; denotes the real space; denotes the channel number of the feature map in the stage;

[0102] B3: let all RepLkBlock modules share the same set of large convolution kernel initial weights, and each module adjusts the kernel weight through a learnable scaling factor to train the iterative RepLkBlock module to obtain the optimal vibration element map;

[0103] Specifically, the adjusted weight can be represented as:

[0104] (5);

[0105] wherein, denotes the updated large kernel convolution weight, The shared large convolution kernel initial weight is not updated.

[0106] It should be noted that only the scaling factor and the gradient residual are updated during the back propagation of B3 to reduce the training time. At the same time, the calculation speed is accelerated by replacing the standard convolution with the depth separable convolution, and the performance of the large kernel is maintained by the algebraically equivalent transformation. Finally, the vibration parallel graph is repeatedly iterated to obtain the vibration parallel graph to minimize the residual parameters, so as to obtain the optimal vibration tuple graph.

[0107] B4: Calculate the image pixel intensity difference (Haar) feature, local binary (LBP) feature and direction gradient histogram (HOG) feature of the vibration tuple graph, and obtain the frequency domain feature at each time point by weighted fusion.

[0108] Specifically, the average value of the Haar feature, the LBP feature and the HOG feature is used as the weighted value, and the pixel matrix of the vibration parallel graph is weighted and added to obtain the frequency domain feature at each time point.

[0109] It should be noted that in step S300, the one-dimensional signal is converted into an image by MTF-RepLkNet, the super large receptive field of RepLkNet is used to capture the spatial mode and dependence of the vibration frequency in the global range, and the high efficiency is ensured by reparameterization.

[0110] In the embodiment of the application, the time sequence feature of the second type signal is extracted by the bidirectional gated recurrent unit (BiGRU) in step S400, and the time domain feature at each time point is extracted by using the global attention mechanism (GA), including steps C1-C5:

[0111] C1: inputting the second type signal into the bidirectional gated recurrent unit;

[0112] C2: calculating the reference vibration time domain signal feature parameter and the updated parameter of the vibration time domain signal, and outputting the time sequence feature corresponding to each time point;

[0113] Specifically, the reference vibration time domain signal feature parameter can be represented as:

[0114] (6) ;

[0115] wherein: represents the unit trend vibration signal at the moment after decomposition; is a normalization coefficient; , , respectively represent the initial parameters set under the current vibration signal and the initial parameters set under the previous vibration signal; respectively represent the initial parameters set under the current vibration signal and the initial parameters set under the previous vibration signal; respectively represent the initial parameters set under the current vibration signal and the initial parameters set under the previous vibration signal;

[0116] Specifically, the vibration time domain signal update parameter Under the action of the Rule activation function, the update parameter is determined by the output of the previous time unit and the current time input .

[0117] Vibration time domain signal update parameter The specific formula is as follows:

[0118] (7);

[0119] Wherein: , respectively represent the constantly updated weight parameters; represent the updated residual value.

[0120] Specifically, the output time sequence feature at the final time is represented by formula (8):

[0121] (8);

[0122] Further, the feature extracted by the BiGRU from the unit trend vibration signal is only related to the data before and after a single signal point. In order to make the obtained time domain feature more accurate, the global attention mechanism is used to update the feature vector , and the specific process is represented as follows:

[0123] C3: calculate the correlation of the second type of signal with the time domain feature state under all initial states;

[0124] The correlation between the unit trend vibration signal and the time domain feature state under all initial states is calculated , which is represented as:

[0125] (9);

[0126] Wherein: and respectively represent the self-learning parameters and the convolution learning parameters; represent the global learning parameters; represent the bias term.

[0127] C4: Based on the correlation, a feature screening module is constructed using an attention mechanism;

[0128] Specifically, it can be expressed as:

[0129] (10);

[0130] Among them: represents the capture information (Query), represents the key information (Key), represents the value information (Value); The distance between two vibration signal points is represented by the function. The data can be normalized by using the function. By calculating the difference between and, the corresponding value weight is obtained. These weights are substituted into the value information to calculate the final global attention value GA.

[0131] C5: The time sequence features are input into the feature screening module to obtain the time domain features at each time point.

[0132] It should be noted that in step S400, the BiGRU captures the bidirectional short-term dependence of the sequence, and the GlobalAttention mechanism can automatically focus on the most critical and representative time points in the entire starting process, ignore noise, and extract the most core time domain evolution rule.

[0133] In the embodiments of the present application, the frequency domain features and the time domain features at each time point are added in step S500, the global information is integrated by using a global average pooling layer (1D-CNN), and the adaptive fusion of the time-frequency domain features is performed by a weighted fusion method, including the following steps D1-D5:

[0134] D1: Based on the frequency domain features at each time point, a frequency domain feature matrix sequence is obtained;

[0135] D2: The frequency domain feature matrix sequence and the time domain feature sequence are multiplied after being normalized to obtain an initial time-frequency domain feature vector;

[0136] D3: The average value of the initial time-frequency domain vector obtained for each segmented signal segment is taken;

[0137] D4: The relationship between the mean values of the time-frequency domain fusion vectors is obtained by a one-dimensional convolutional neural network to generate a weight vector.

[0138] In the embodiments of the present application, the relationship between the mean values of the time-frequency domain fusion vectors is obtained by a one-dimensional convolutional neural network to generate a weight vector in D4, including:

[0139] ​​​D4.1: One-dimensional convolutional neural network uses two convolutional layer W1 and W2 structure;

[0140] D4.2: The weight vector is calculated by Sigmoid function .

[0141] Specifically, it is represented as:

[0142] (11);

[0143] Wherein, Residual error between input value and output value; W2 represents the weight matrix of the second convolutional layer in 1D-CNN; W1 represents the weight matrix of the first convolutional layer in 1D-CNN; is the average post-input time-frequency domain initial vector.

[0144] D5: The weight vector is multiplied by the time-frequency domain initial vector modulus to obtain the time-frequency domain fusion feature value of the vibration signal under different starting modes.

[0145] Specifically, the time-frequency domain fusion feature value can be represented as:

[0146] (12);

[0147] Wherein, Time-frequency domain fusion feature value; Time-frequency domain initial vector modulus.

[0148] In the embodiment of the application, the vibration feature reference vector under the original starting mode of the unit is obtained by the above-mentioned pretreatment, time-frequency domain feature extraction and weighted fusion method, i.e. steps S200-S500, in step S600;

[0149] In the embodiment of the application, the maximum distance between the vibration feature vector under different starting modes and the reference vector is the optimal starting mode in step S700.

[0150] In summary, the pros and cons of the starting process are quantified as a calculable distance between the vibration feature vector and the reference vector, thereby providing an objective and unified mathematical scoring standard for different starting strategies, rather than the traditional experience standard; the energy of the vibration frequency component is captured by MTF-RepLkNet; the long-term trend is captured by BiGRU-GlobalAttention, and the time domain features are fused by 1D-CNN, which comprehensively and deeply describes the dynamic characteristics of the starting process from multiple dimensions, and the judgment result is more comprehensive and accurate.

[0151] Example 3, refer to Figures 2-7Based on the previous embodiment, the present embodiment provides a practical case effect verification of the method for determining the starting mode of the hydroelectric generating set by fusing multi-modal features, so as to embody the beneficial effects of the present application.

[0152] For a test on the No. 6 unit of a certain hydropower station, the unit capacity is 1000 MW, and the rated speed is 111.1 r / min. The strain signals of the guide vane connecting rod position under different starting modes are collected on site for analysis, and the data sampling frequency is 2400 Hz. The unit running speed and guide vane opening parameters are as shown in Figure 2 and Figure 3 It can be seen from the figure that the final opening of the five modes reaches the stable no-load opening, and the total time experienced from small to large is mode 1, mode 2, mode 3.1, mode 3.2, and mode 4. Among them, the guide vane opening change is the largest in the first stage of mode 2 and mode 4; the speed of mode 3.1 and 3.2 increases rapidly between 30s~40s; mode 4 stays at the position of about 8% guide vane opening for the longest time. The speed of mode 1 and mode 2 rises the fastest, the speed of mode 2 increases slowly in the first 30s, and then increases rapidly after the second increase of the guide vane opening, the speed of mode 3.1 and 3.2 increases rapidly after 60s, and the speed of mode 4 changes relatively slowly.

[0153] Among them, mode 3.1 and mode 3.2 are the same adjustment mode, but there are differences in the details of the guide vane opening, mode 3.1 continues to increase at a faster speed after reaching the set value, and mode 3.2 increases slowly, and is unified as mode 3 in the subsequent Figure 6 and Figure 7

[0154] First, the strain signals are preprocessed and decomposed into two types of time series vibration signals, and the decomposition results at a certain time point are as shown in Figure 4 and Figure 5 The MFT-RepLkNet method is used to extract the frequency domain signal features of the vibration characteristic signals, and the results are as shown in Figure 6 The time domain features of the vibration trend signals are extracted by using the bidirectional gate recurrent unit, and the adaptive fusion of the time-frequency domain features is realized by using the weighted fusion method, and the vibration information under different starting modes is obtained, and the results are as shown in Figure 7 It can be seen that, compared with mode 1, the feature values of mode 2 and mode 4 deviate greatly, and the vibration changes more obviously compared with mode 1. The feature value trend of mode 3 is basically the same as the original mode, but there is a large deviation in the beginning.

[0155] ​Embodiment 4, the above is a schematic scheme of a fusion multi-modal feature hydroelectric generating set start-up mode determination method. It should be noted that the technical scheme of the fusion multi-modal feature hydroelectric generating set start-up mode determination system belongs to the same concept as the technical scheme of the fusion multi-modal feature hydroelectric generating set start-up mode determination method described above. The technical scheme of the fusion multi-modal feature hydroelectric generating set start-up mode determination system in this embodiment is not described in detail. Please refer to the description of the technical scheme of the fusion multi-modal feature hydroelectric generating set start-up mode determination method described above.

[0156] The embodiment also provides another fusion multi-modal feature hydroelectric generating set start-up mode determination system, which comprises:

[0157] The acquisition module is configured to acquire original vibration signals under different start-up modes.

[0158] The preprocessing module is configured to preprocess the original vibration signals to obtain two types of time series vibration signals in the start-up process, including a first type of time series signal representing local features and a second type of time series signal representing long-term changes.

[0159] The frequency domain feature extraction module is configured to convert the first type of signal into a time-frequency image through Markov transition field, and extract frequency domain features at each time point using a structure re-parameterization large kernel network.

[0160] The time domain feature extraction module is configured to extract time series features of the second type of signal through a bidirectional gated recurrent unit, and extract time domain features at each time point using a global attention mechanism.

[0161] The feature fusion module is configured to multiply the frequency domain features and the time domain features at each time point, integrate global information using a global average pooling layer, and perform adaptive fusion of the time-frequency domain features through a weighted fusion method to obtain vibration features under different start-up modes.

[0162] The reference vector acquisition module is configured to obtain vibration feature reference vectors under the original start-up state of the generating set through the preprocessing, time-frequency domain feature extraction and weighted fusion method.

[0163] The calculation and judgment module is configured to calculate the distance between the vibration feature vectors under different start-up modes and the reference vectors, and obtain the optimal start-up mode of the hydroelectric generating set based on the distance value.

[0164] The embodiment also provides a computer device, which comprises a memory and a processor. The memory is configured to store computer executable instructions, and the processor is configured to execute the computer executable instructions to implement the fusion multi-modal feature hydroelectric generating set start-up mode determination method proposed in the above embodiment.

[0165] The embodiment also provides a storage medium, which stores a computer program, and the computer program is executed by a processor to implement the method for determining the starting mode of the hydroelectric generating set by fusing multi-modal features as proposed in the above embodiment.

[0166] The storage medium proposed in the embodiment belongs to the same inventive concept as the method for determining the starting mode of the hydroelectric generating set by fusing multi-modal features proposed in the above embodiment, and the technical details not described in the embodiment can be referred to the above embodiment, and the embodiment has the same beneficial effects as the above embodiment.

[0167] Through the above description of the embodiments, those skilled in the art can clearly understand that the present application can be realized by means of software and necessary universal hardware, and of course can also be realized by hardware. Based on such understanding, the technical solutions of the present application or the part that contributes to the prior art can be embodied in the form of a software product, and the computer software product can be stored in a computer readable storage medium, such as a floppy disk, a read-only memory (ROM), a random access memory (RAM), a FLASH, a hard disk, or an optical disc, etc., and includes a number of instructions to make a computer device (which can be a personal computer, a server, or a network device, etc.) execute the methods of various embodiments of the present application.

[0168] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present application but not limit the present application, and although the present application has been described in detail with reference to the preferred embodiments, those skilled in the art should understand that the technical solutions of the present application can be modified or replaced by equivalents without departing from the spirit and scope of the present application, and all of them should be covered in the scope of the claims of the present application.

Claims

1. A method for determining the start-up mode of a hydropower unit by integrating multimodal characteristics, characterized in that, include: Collect raw vibration signals under different power-on modes; The original vibration signal is preprocessed to obtain two types of time-series vibration signals during the power-on process, including the first type of time-series signal that characterizes local features and the second type of time-series signal that characterizes long-term changes. The first type of signal is converted into a time-frequency image using a Markov transformed field, and the frequency domain features at each time point are extracted using a structure-reparameterized large kernel network, including: An image frequency domain feature extraction architecture is constructed, which includes four stages, each stage comprising... There are one RepLkBlock module, among which... Image sampling is performed between each stage using strided convolution; The MTF plot of each vibration signal is decomposed using a multi-branch structure. A vibration tuple diagram is used to construct the RepLkBlock module; All RepLkBlock modules share the same set of initial weights for large convolutional kernels. Each module adjusts the kernel weights using a learnable scaling factor to train and iterate the RepLkBlock modules to obtain the optimal vibration tuple diagram. The Haar, LBP, and HOG features of the vibration tuple diagram are calculated and weighted to obtain the frequency domain features at each time point. The temporal features of the second type of signal are extracted by a bidirectional gated recurrent unit, and the temporal features at each time point are extracted by a global attention mechanism. The frequency domain features and time domain features at each time point are multiplied together, and the global average pooling layer is used to integrate the global information. The time and frequency domain features are adaptively fused through a weighted fusion method to obtain the vibration features under different power-on modes. The vibration characteristic reference vector of the unit under the original start-up state is obtained through the preprocessing, time-frequency domain feature extraction and weighted fusion method. Calculate the distance between the vibration characteristic vector and the reference vector under different start-up modes, and obtain the optimal start-up mode of the hydropower unit based on the distance value.

2. The method for determining the start-up mode of a hydropower unit by integrating multimodal characteristics as described in claim 1, characterized in that, Converting a Type I signal into a time-frequency image using a Markov transformed field includes: Classify the first type of signal into A segment; For each segment of the first type of signal, it is divided according to the signal amplitude using a normal distribution, resulting in... Markov region of hydroelectric power units , among which, region With local vibration signals Mutual mapping; Vibration signals at different times are discretized into state regions. Calculate the signal from the region Transfer mapping to region The probability between them, to obtain the probability for To construct the Markov state transition matrix; Based on the Markov state transition matrix, a dynamic Markov transformation field is used to generate the local vibration signal state transition matrix at multiple scales. ; The state transition matrix Mapped to an image and averaged using a blur kernel to generate... MTF graph of a vibration signal.

3. The method for determining the start-up mode of a hydropower unit by integrating multimodal characteristics as described in claim 2, characterized in that, The temporal features of the second type of signal are extracted using a bidirectional gated recurrent unit, and the temporal features at each time point are extracted using a global attention mechanism, including: The second type of signal is input to the bidirectional gated loop unit; Calculate the characteristic parameters of the reference vibration time-domain signal and the update parameters of the vibration time-domain signal, and output the time-series characteristics corresponding to each time point; Calculate the correlation between the second type of signal and the time-domain characteristic states under all initial states; Based on relevance, an attention mechanism is used to construct a feature selection module; Input the time-series features into the feature filtering module to obtain the time-domain features at each time point.

4. The method for determining the start-up mode of a hydropower unit by integrating multimodal characteristics as described in claim 3, characterized in that, The frequency domain features and time domain features at each time point are added together, and a global average pooling layer is used to integrate the global information. Adaptive fusion of time and frequency domain features is then performed through a weighted fusion method, including: A sequence of frequency domain feature matrices is obtained based on the frequency domain features at each time point; The frequency domain feature matrix sequence and the time domain feature sequence are normalized and then multiplied to obtain the initial vector of time-frequency domain features. Take the average value of the initial time-frequency domain vector obtained for each segmented signal segment; The relationship between the mean values ​​of the time-frequency domain fusion vector is obtained through a one-dimensional convolutional neural network, and a weight vector is generated. Multiplying the weight vector by the magnitude of the initial vector in the time-frequency domain yields the time-frequency domain fused feature values ​​of the vibration signal under different power-on modes.

5. The method for determining the start-up mode of a hydropower unit by integrating multimodal characteristics as described in claim 4, characterized in that, The step of obtaining the relationship between the mean values ​​of the time-frequency domain fusion vectors through a one-dimensional convolutional neural network and generating a weight vector includes: A one-dimensional convolutional neural network uses a structure with two convolutional layers, W1 and W2. The weight vector is calculated using the Sigmoid function.

6. The method for determining the start-up mode of a hydropower unit by integrating multimodal characteristics as described in claim 5, characterized in that, It also includes the optimal startup mode, which is the maximum distance between the vibration characteristic vector and the reference vector under different startup modes.

7. A system for determining the start-up mode of a hydropower unit by integrating multimodal characteristics, using the method for determining the start-up mode of a hydropower unit by integrating multimodal characteristics as described in any one of claims 1-6, characterized in that, include: The acquisition module is used to acquire raw vibration signals under different power-on modes; The preprocessing module is used to preprocess the original vibration signal to obtain two types of time-series vibration signals during the power-on process, including the first type of time-series signal characterizing local features and the second type of time-series signal characterizing long-term changes. The frequency domain feature extraction module is used to convert the first type of signal into a time-frequency image through the Markov transformation field, and to extract the frequency domain features at each time point using the structure reparameterized big kernel network. The temporal feature extraction module is used to extract the temporal features of the second type of signal through a bidirectional gated loop unit and to extract the temporal features at each time point using a global attention mechanism. The feature fusion module is used to multiply the frequency domain features and time domain features at each time point, integrate the global information using a global average pooling layer, and perform adaptive fusion of time and frequency domain features through a weighted fusion method to obtain vibration features under different power-on modes. The reference vector acquisition module is used to obtain the vibration feature reference vector of the unit under the original start-up state through the preprocessing, time-frequency domain feature extraction and weighted fusion method. The calculation and judgment module is used to calculate the distance between the vibration characteristic vector and the reference vector under different start-up modes, and to determine the optimal start-up mode of the hydropower unit based on the distance value.

8. A computer device, comprising: Memory and processor; The memory is used to store computer-executable instructions, and the processor is used to execute the computer-executable instructions. When the computer-executable instructions are executed by the processor, they implement the steps of the method for determining the start-up mode of a hydropower unit with integrated multimodal characteristics as described in any one of claims 1 to 6.

9. A computer-readable storage medium storing computer-executable instructions, which, when executed by a processor, implement the steps of the method for determining the start-up mode of a hydropower unit incorporating multimodal characteristics as described in any one of claims 1 to 6.

Images