A flywheel energy storage device operating state online monitoring method and system

Abstract

The application provides a flywheel energy storage device operation state online monitoring method and system, including dividing signals into multiple sub-frequency bands based on frequency domain information entropy of each sensor signal, determining a sub-frequency band noise reduction mode according to the characteristics of the sub-frequency band, and obtaining the noise-reduced operation parameters; obtaining the operation stage and health index of the flywheel energy storage device corresponding to the training sample, and obtaining the cost matrix based on the operation stage and the health index; obtaining the ratio of the sample quantity of each operation state in the training sample subset where the current training sample is located, determining the adjusted cost matrix based on the cost matrix and the ratio, training the deep forest model using the adjusted cost matrix; inputting the noise-reduced operation parameters into the deep forest model, inputting the feature vector output by each forest layer in the deep forest model into the self-organizing mapping network, and obtaining the operation state of the flywheel energy storage device according to the output of the self-organizing mapping network corresponding to the forest layer and the identification result of the forest layer.

Description

Technical Field

[0001] This application belongs to the field of online monitoring, and in particular relates to a method and system for online monitoring of the operating status of flywheel energy storage equipment. Background Technology

[0002] Flywheel energy storage systems, as a highly efficient mechanical energy storage technology, offer advantages such as fast charging speed, long cycle life, and zero pollution, showing promising application prospects in areas such as power system frequency regulation, uninterruptible power supplies (UPS), rail transit energy recovery, microgrids, and new energy grid integration. Flywheel energy storage devices typically operate under high-speed rotation conditions, and their core components, such as high-speed motors, bearing systems, vacuum systems, and power electronic converters, are subjected to complex electromagnetic, mechanical, and thermal stresses over long periods, inevitably leading to performance degradation and even sudden failures. Accurate and real-time online monitoring of the operating status of flywheel energy storage devices, timely detection of early fault signs, and diagnosis of fault type and severity are of great significance for ensuring the safe and stable operation of the system, improving equipment reliability, and extending service life. However, the multi-source heterogeneous sensor signals generated during operation, such as vibration, temperature, current, and voltage, are often submerged in strong background noise and complex operating condition disturbances, especially the early, weak fault characteristic signals, which have extremely low signal-to-noise ratios and are difficult to extract effectively. Moreover, flywheel energy storage devices operate in various modes, such as acceleration, constant speed, deceleration, and standby. The state characteristics of the device vary under different modes, and the boundaries between states may be blurred. This places high demands on the adaptability and identification accuracy of the monitoring model. Summary of the Invention

[0003] To address the current problems of poor noise reduction and low monitoring accuracy due to the model's poor adaptability to dynamic operating conditions in the monitoring of flywheel energy storage devices, this application proposes an online monitoring method for the operating status of flywheel energy storage devices, including:

[0004] The operating parameters of the flywheel energy storage device are collected by sensors. Based on the frequency domain information entropy of each sensor signal, the signal is divided into multiple sub-frequency bands. The noise reduction method of the sub-frequency band is determined according to the characteristics of the sub-frequency band. The noise reduction data of each sub-frequency band of the same sensor is performed separately and then reconstructed to obtain the noise-reduced operating parameters.

[0005] The operating phase and health index of the flywheel energy storage device corresponding to the training sample are obtained, and a cost matrix is ​​obtained based on the operating phase and the health index; the ratio of the number of samples in each operating state in the training sample subset where the current training sample is located is obtained, and an adjusted cost matrix is ​​determined based on the cost matrix and the ratio; the deep forest model is trained using the adjusted cost matrix.

[0006] The noise-reduced operating parameters are input into the deep forest model, and the feature vectors output by each forest layer in the deep forest model are input into the self-organizing map network. The operating status of the flywheel energy storage device is obtained based on the output of the self-organizing map network corresponding to the forest layer and the identification result of the forest layer.

[0007] Optionally, the frequency domain information entropy based on the signals from each sensor divides the signal into multiple sub-frequency bands, including:

[0008] The amplitude spectrum of the signal from each sensor is calculated using Fast Fourier Transform, and the amplitude spectrum is divided into M equal-bandwidth frequency bands according to frequency.

[0009] Calculate the proportion of energy at each frequency point within the frequency band to the total energy of the frequency band, and determine the frequency domain information entropy of each frequency band based on the proportion;

[0010] Obtain the number N of target sub-bands, where N is less than M and M and N are both positive integers; with the optimization objective of maximizing the sum of information entropy within all target sub-bands, the process is carried out by iteratively merging adjacent frequency bands until N target sub-bands are obtained.

[0011] Optionally, determining the sub-band noise reduction method based on the characteristics of the sub-band includes:

[0012] Extract features from each sub-band, including energy center frequency, energy amplitude, spectral kurtosis value, and power spectral entropy value;

[0013] The extracted features of each sub-frequency band are matched with the feature combinations in the noise reduction rule base to determine the noise reduction method and corresponding parameter set for each sub-frequency band.

[0014] The determined noise reduction algorithm and parameter set are applied to the corresponding sub-frequency band data to perform noise reduction processing.

[0015] Optionally, obtaining the cost matrix based on the operating phase and the health index includes:

[0016] Determine the running stage index and health index level index of the current training sample, and retrieve the cost matrix of the current sample from the cost lookup table based on the running stage index and health index level index.

[0017] Optionally, training the deep forest model using the adjusted cost matrix includes:

[0018] Within any decision tree of the deep forest, when selecting a split attribute for a node that aggregates a set of training samples, the adjusted cost matrix corresponding to each training sample in the set is obtained.

[0019] When the training sample set of the node is divided into at least one subset using a certain attribute, for each subset, each training sample in the subset is traversed, and the predicted category is determined according to the ratio between the true category of the training sample and the predicted category. The corresponding misclassification cost is found in the adjusted cost matrix of the training sample, and the misclassification costs of all training samples in the current subset are summed and divided by the number of samples in the subset to obtain the cost.

[0020] The attribute that minimizes the sum of costs of all the subsample sets formed after splitting is taken as the splitting attribute of the node.

[0021] Optionally, obtaining the operating status of the flywheel energy storage device based on the output of the self-organizing mapping network corresponding to the forest layer and the identification result of the forest layer includes:

[0022] The weight of each neuron in the self-organizing map network is determined based on the normalized frequency of activation of each neuron during training.

[0023] The feature vector output by each forest layer in the deep forest model is input into the self-organizing map network corresponding to that forest layer.

[0024] The state and state probability of the flywheel energy storage device are determined based on the output feature vector of each forest layer of the deep forest model, and the weights of the activated neurons of the self-organizing map network corresponding to the forest layer are obtained. The output of each forest layer is obtained by weighting the state probability with the weights.

[0025] The operating status of the flywheel energy storage device is obtained by summing and weighting the outputs of all forest layers.

[0026] This application also proposes an online monitoring system for the operating status of a flywheel energy storage device, including:

[0027] The preprocessing unit is used to collect the operating parameters of the flywheel energy storage device through sensors, divide the signal into multiple sub-frequency bands based on the frequency domain information entropy of each sensor signal, determine the sub-frequency band noise reduction method according to the characteristics of the sub-frequency band, and reconstruct the noise-reduced operating parameters after performing noise reduction on the data of each sub-frequency band of the same sensor.

[0028] The training unit is used to obtain the operating stage and health index of the flywheel energy storage device corresponding to the training sample, and obtain the cost matrix based on the operating stage and the health index; obtain the ratio of the number of samples in each operating state in the training sample subset where the current training sample is located, determine the adjusted cost matrix based on the cost matrix and the ratio, and use the adjusted cost matrix to train the deep forest model.

[0029] The identification unit is used to input the noise-reduced operating parameters into the deep forest model, input the feature vector output by each forest layer in the deep forest model into the self-organizing map network, and obtain the operating status of the flywheel energy storage device based on the output of the self-organizing map network corresponding to the forest layer and the identification result of the forest layer.

[0030] Optionally, the frequency domain information entropy based on the signals from each sensor divides the signal into multiple sub-frequency bands, including:

[0031] The amplitude spectrum of the signal from each sensor is calculated using Fast Fourier Transform, and the amplitude spectrum is divided into M equal-bandwidth frequency bands according to frequency.

[0032] Calculate the proportion of energy at each frequency point within the frequency band to the total energy of the frequency band, and determine the frequency domain information entropy of each frequency band based on the proportion;

[0033] Obtain the number N of target sub-bands, where N is less than M and M and N are both positive integers; with the optimization objective of maximizing the sum of information entropy within all target sub-bands, the process is carried out by iteratively merging adjacent frequency bands until N target sub-bands are obtained.

[0034] Optionally, determining the sub-band noise reduction method based on the characteristics of the sub-band includes:

[0035] Extract features from each sub-band, including energy center frequency, energy amplitude, spectral kurtosis value, and power spectral entropy value;

[0036] The extracted features of each sub-frequency band are matched with the feature combinations in the noise reduction rule base to determine the noise reduction method and corresponding parameter set for each sub-frequency band.

[0037] The determined noise reduction algorithm and parameter set are applied to the corresponding sub-frequency band data to perform noise reduction processing.

[0038] Optionally, obtaining the cost matrix based on the operating phase and the health index includes:

[0039] Determine the running stage index and health index level index of the current training sample, and retrieve the cost matrix of the current sample from the cost lookup table based on the running stage index and health index level index.

[0040] Optionally, training the deep forest model using the adjusted cost matrix includes:

[0041] Within any decision tree of the deep forest, when selecting a split attribute for a node that aggregates a set of training samples, the adjusted cost matrix corresponding to each training sample in the set is obtained.

[0042] When the training sample set of the node is divided into at least one subset using a certain attribute, for each subset, each training sample in the subset is traversed, and the predicted category is determined according to the ratio between the true category of the training sample and the predicted category. The corresponding misclassification cost is found in the adjusted cost matrix of the training sample, and the misclassification costs of all training samples in the current subset are summed and divided by the number of samples in the subset to obtain the cost.

[0043] The attribute that minimizes the sum of costs of all the subsample sets formed after splitting is taken as the splitting attribute of the node.

[0044] Optionally, obtaining the operating status of the flywheel energy storage device based on the output of the self-organizing mapping network corresponding to the forest layer and the identification result of the forest layer includes:

[0045] The weight of each neuron in the self-organizing map network is determined based on the normalized frequency of activation of each neuron during training.

[0046] The feature vector output by each forest layer in the deep forest model is input into the self-organizing map network corresponding to that forest layer.

[0047] The state and state probability of the flywheel energy storage device are determined based on the output feature vector of each forest layer of the deep forest model, and the weights of the activated neurons of the self-organizing map network corresponding to the forest layer are obtained. The output of each forest layer is obtained by weighting the state probability with the weights.

[0048] The operating status of the flywheel energy storage device is obtained by summing and weighting the outputs of all forest layers.

[0049] This invention employs an adaptive sub-band partitioning and feature matching denoising method based on the frequency domain information entropy of sensor signals. This method can more precisely suppress noise in different frequency bands and improve the signal-to-noise ratio of the operating parameters after denoising. Furthermore, by constructing a cost matrix dynamically adjusted based on the operating stage and health index, and applying it to the training process of a deep forest model, the model can improve its sensitivity to identifying critical fault states based on the differences in misjudgment risk under different operating conditions and equipment health states, as well as the ratio of the number of samples of different categories in the training samples. In addition, the feature vectors output from each forest layer of the deep forest model are further input into a self-organizing map network for weight optimization and state fusion. Utilizing the topology preservation and clustering capabilities of the self-organizing map network for features, the accuracy and robustness of operating state identification are further improved. Attached Figure Description

[0050] Figure 1 This is a flowchart of Example 1;

[0051] Figure 2This is a schematic diagram showing the frequency bands divided into 10 equal bandwidths.

[0052] Figure 3 This is a schematic diagram after merging equal bandwidth frequency bands;

[0053] Figure 4 This is a diagram illustrating the use of 60 as the attribute for the current node.

[0054] Figure 5 This is a diagram illustrating the use of 65 as the attribute for the current node.

[0055] Figure 6 This is a schematic diagram of Deep Forest and SOM. Detailed Implementation

[0056] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of this application, and not all of them. All other embodiments obtained by those skilled in the art based on the embodiments of this application without creative effort are within the scope of protection of this application.

[0057] Specific embodiments, such as Figure 1 As shown, this application proposes an online monitoring method for the operating status of a flywheel energy storage device, comprising:

[0058] S1. The operating parameters of the flywheel energy storage device are collected by sensors. Based on the frequency domain information entropy of each sensor signal, the signal is divided into multiple sub-frequency bands. The noise reduction method of the sub-frequency band is determined according to the characteristics of the sub-frequency band. The noise reduction of each sub-frequency band data of the same sensor is performed separately and then reconstructed to obtain the noise-reduced operating parameters.

[0059] Data from various types of sensors, including vibration, temperature, current, voltage, rotational speed, and so on, is acquired in real time. For each sensor's signal set, a short-time Fourier transform or wavelet packet decomposition is performed to obtain the frequency components and energy distribution within different time windows. The frequency domain information entropy of the signal within each sub-band is calculated under different candidate frequency band partitioning strategies. In one embodiment, the frequency information entropy uses Shannon information entropy; frequency bands with higher information entropy contain richer signal information. A frequency band partitioning method that optimizes the sum or weighted sum of the information entropies of all sub-bands is selected, thereby decomposing the original broadband signal into several narrowband sub-signals. The characteristics of each partitioned sub-band are analyzed, such as the energy concentration, kurtosis, spectral kurtosis, and power spectral density shape of the sub-band signal, and these characteristics are compared with a noise reduction knowledge base. The knowledge base pre-stores the typical characteristics of different noise types, such as Gaussian white noise, impulse noise, periodic narrowband interference, and power frequency interference, at different frequency bands, as well as the most effective noise reduction algorithm categories for these characteristics, such as empirical mode decomposition, variational mode decomposition, spectral subtraction, adaptive line spectrum enhancement, and specific type filters. In a more specific embodiment, it also includes recommended parameters. By matching the features with the knowledge base, the most suitable noise reduction method and parameters are determined for each sub-frequency band, and the corresponding noise reduction processing is performed independently on the data of each sub-frequency band. All the noise-reduced sub-frequency band signals are fused and reconstructed through inverse transformation or direct superposition to obtain the sensor's noise-reduced operating parameter sequence.

[0060] S2, obtain the operating stage and health index of the flywheel energy storage device corresponding to the training sample, and obtain the cost matrix based on the operating stage and the health index; obtain the ratio of the number of samples in each operating state in the training sample subset where the current training sample is located, determine the adjusted cost matrix based on the cost matrix and the ratio, and train the deep forest model using the adjusted cost matrix;

[0061] Specifically, during the model training phase, for each labeled sample used for training, in addition to its corresponding real operating state label, two additional pieces of information are required: the operating stage of the flywheel device when the sample was collected, which includes the start-up acceleration stage, stable energy storage stage, rapid discharge stage, and low-power standby stage; and the health index of the overall health status of the device at that time. In one embodiment, the health index is obtained through information such as the device's cumulative operating time, historical maintenance records, and degradation of key components.

[0062] Based on the operational phase and the health index, a cost matrix is ​​obtained. In one embodiment, the cost matrix is ​​a table, obtained by looking up the table using the operational phase and health index. The cost matrix represents the penalty weights to be assigned when a true state is misclassified as another state under specific conditions, i.e., operational phase and health status. During the training of the deep forest model, especially when processing each batch or subset of training samples, the number of samples contained in each known operational state category in the current batch or subset is counted, and the ratio between them is calculated. If the number of samples in some states is much less than that in other states, i.e., there is a class imbalance, the cost matrix of the current samples is adjusted according to these sample ratios to obtain an adjusted cost matrix. In one embodiment, the adjustment method is to increase the cost of misclassifying minority class samples and decrease the cost of misclassifying majority class samples; in another embodiment, a nonlinear mapping is performed based on the proportion of each class and the basic cost. During the growth of each decision tree in the deep forest, when selecting the optimal attribute of the split node, this adjusted cost matrix is ​​used to evaluate the expected misclassification cost caused by different split choices, and the attribute that minimizes the total expected cost is taken as the split attribute.

[0063] S3, input the noise-reduced operating parameters into the deep forest model, input the feature vector output by each forest layer in the deep forest model into the self-organizing map network, and obtain the operating status of the flywheel energy storage device based on the output of the self-organizing map network corresponding to the forest layer and the identification result of the forest layer.

[0064] When judging the current operating parameters, noise reduction is performed first in the same way as in S1. The noise-reduced data is then input into the pre-trained deep forest model. The deep forest model consists of multiple cascaded forest layers, each containing multiple different types of random forests. In one embodiment, after the input data passes through the first forest layer, each forest within that layer provides a class probability vector for the input. These vectors are concatenated as the input features for the next forest layer. In another embodiment, the features output by each forest layer pass through a small network to obtain a class probability distribution vector, while the output of each forest layer is a feature vector.

[0065] The feature vector or class probability vector output by each forest layer is input into a pre-trained Self-Organizing Map (SOM) network corresponding to that forest layer. A SOM is an unsupervised learning neural network that maps high-dimensional input data onto a typically discrete network of neurons while preserving the topological structure of the input data. During training, the SOM network learns different feature patterns represented by different regions based on the distribution of the input feature vectors, and each neuron forms a weight or strength based on its activation frequency during training or the density of data points in its neighborhood. When a new feature vector is input into the SOM network, it finds an optimal matching unit—the neuron most similar to its input pattern. A fusion decision is made by combining the preliminary identification results of the current forest layer's operating state (e.g., the highest probability class) with the weights of the neurons activated by that forest layer's feature vector in the corresponding SOM network or the typical state information represented by that neuron's region. In one embodiment, the weights of the SOM neurons are used to adjust the state probabilities given by the forest layer, or the topological location information output by the SOM is fed back as additional features to subsequent decision modules. By combining the output results of all forest layers after optimization or calibration of their corresponding SOM networks, the current operating status of the flywheel energy storage device is obtained through weighted averaging or voting mechanisms.

[0066] In one embodiment, the frequency domain information entropy based on each sensor signal divides the signal into multiple sub-frequency bands, including:

[0067] The amplitude spectrum of the signal from each sensor is calculated using Fast Fourier Transform, and the amplitude spectrum is divided into M equal-bandwidth frequency bands according to frequency.

[0068] Calculate the proportion of energy at each frequency point within the frequency band to the total energy of the frequency band, and determine the frequency domain information entropy of each frequency band based on the proportion;

[0069] Obtain the number N of target sub-bands, where N is less than M and M and N are both positive integers; with the optimization objective of maximizing the sum of information entropy within all target sub-bands, the process is carried out by iteratively merging adjacent frequency bands until N target sub-bands are obtained.

[0070] The following example uses a vibration sensor on a flywheel energy storage device. Assume this vibration sensor has a sampling rate of 1024 kHz.

[0071] Performing a Fast Fourier Transform (FFT) on a vibration signal yields its amplitude spectrum within the frequency range of 0 to 5120 Hz. The amplitude spectrum displays the distribution of signal energy at different frequencies. The complete spectrum is initially divided into M equal-bandwidth frequency bands; for example, if M = 256, then each initial band is 20 Hz wide. Figure 2The diagram illustrates dividing 1000 Hz into 10 frequency bands. For any one of these 256 initial frequency bands, such as the band from 1000 Hz to 1020 Hz, the proportion of energy at each frequency point to the total energy of that 1000-1020 Hz band is calculated. Based on these energy proportions, the specific frequency domain information entropy for this 20 Hz wide initial frequency band is calculated using an information entropy calculation formula, such as the Shannon entropy formula. The more complex the energy distribution, the higher the entropy value is usually. This results in 256 corresponding frequency domain information entropy values. Furthermore, if N=4, that is, by merging adjacent initial frequency bands to reduce the number of frequency bands from 256 to 4, the most suitable merging method is to maximize the sum of the frequency domain information entropies of the final 4 target sub-bands. Specifically, different merging strategies are tried, such as merging the 1st to 10th initial frequency bands into the first target sub-band, merging the 11th to 25th initial frequency bands into the second target sub-band, and so on, until four target sub-bands are formed. Figure 3 As shown. For each possible merging scheme resulting in 16 target sub-bands, the combined frequency domain information entropy is recalculated and summed. This process continues until a partitioning method is found that maximizes the sum of the information entropies of these 16 target sub-bands among all possible merging methods. In one embodiment, other methods can be used to determine the conditions for stopping the attempts, such as the number of iterations or a preset multiple of the average value.

[0072] In one embodiment, determining the sub-band noise reduction method based on the characteristics of the sub-band includes:

[0073] Extract features from each sub-band, including energy center frequency, energy amplitude, spectral kurtosis value, and power spectral entropy value;

[0074] The extracted features of each sub-frequency band are matched with the feature combinations in the noise reduction rule base to determine the noise reduction method and corresponding parameter set for each sub-frequency band.

[0075] The determined noise reduction algorithm and parameter set are applied to the corresponding sub-frequency band data to perform noise reduction processing.

[0076] Taking a vibration sensor as an example, after dividing the frequency into sub-bands, the characteristics of each independent sub-band are calculated. Assume that three sub-bands are obtained after processing: sub-band A covers 0 to 100 Hz, sub-band B covers 100 to 500 Hz, and sub-band C covers 500 to 2000 Hz. The energy center frequency of sub-band A is calculated to be 30 Hz, with an energy amplitude of 5 and a spectral kurtosis of 2.5, indicating a relatively stable energy distribution in this band. The power spectral entropy is 0.4, which together constitute the characteristics of the sub-band.

[0077] The extracted feature values ​​are combined and matched with a denoising rule base. This rule base stores a large amount of expert knowledge and experience data, linking different feature combination patterns with the optimal denoising algorithm and its core parameters. For example, one rule in the rule base states: if the spectral kurtosis of a sub-band is greater than 6 and the power spectral entropy is between 0.5 and 0.7, it indicates that the band may be subject to random impulse noise interference. In this case, a wavelet transform-based threshold denoising method should be used, and a wavelet basis such as sym4 is recommended, with a decomposition level of 4 and a heuristic hard threshold. Another rule states: if the energy center frequency of a sub-band is close to an integer multiple of the power frequency, such as 48 to 52 Hz, and the energy amplitude is significantly higher than that of neighboring frequency bands, power frequency interference may exist. In this case, an adaptive notch filter should be used, and the notch point should be precisely set according to the actual measured center frequency. For the feature combination of sub-band A, a rule recommending the use of empirical mode decomposition combined with a specific mode function selection strategy will be matched. Sub-band C uses a wavelet threshold denoising rule. After determining the specific noise reduction method and parameters, the noise reduction method is applied to the corresponding sub-frequency band data to perform noise reduction processing.

[0078] In one specific embodiment, obtaining the cost matrix based on the operating phase and the health index includes:

[0079] Determine the running stage index and health index level index of the current training sample, and retrieve the cost matrix of the current sample from the cost lookup table based on the running stage index and health index level index.

[0080] The operation of a flywheel energy storage device can be divided into multiple stages. For example, the process of the device accelerating from a standstill and storing energy is defined as operation stage index 0; the state where the device maintains stable energy after reaching its rated speed is defined as operation stage index 1; the process of the device decelerating and releasing energy is defined as operation stage index 2; and the standby / dormant state where the device waits for instructions at low power consumption is defined as operation stage index 3. The overall health status of the device is assessed based on sensor data and historical operating records, and this health status is quantified into a health index, for example, from 0 to 100. For example, a health score above 85 corresponds to health level index 0, representing excellent condition; 70 to 84 corresponds to level index 1, representing good condition; 50 to 69 corresponds to level index 2, representing average condition; and below 50 corresponds to level index 3, representing poor condition. When a training sample is collected, both the operation stage index (e.g., 1) and the health index level index (e.g., 3) are obtained simultaneously.

[0081] A pre-established cost lookup table is obtained. Preferably, the cost lookup table is a two-dimensional table, with one dimension corresponding to the runtime index and the other dimension corresponding to the health index level index. Each cell of the table stores a specific cost matrix. When the runtime index of a training sample is 1 and the health index level index is 3, the pre-defined cost matrix corresponding to row 1 and column 3 is located from the cost lookup table. The cost matrix represents the relative cost or risk of misjudging a true state as another state. For example, under the above conditions, the cost of misjudging a state that is actually a serious fault as a normal state may be set to a very high value, such as 100, while the cost of misjudging a slight anomaly as normal is 30, and the cost of a correct judgment is 0. In the case of runtime index 0 and health index 0, i.e., acceleration and excellent state, another cost matrix is ​​obtained from the table, thus adapting to the risk assessment needs under different scenarios.

[0082] In one embodiment, training the deep forest model using the adjusted cost matrix includes:

[0083] Within any decision tree of the deep forest, when selecting a split attribute for a node that aggregates a set of training samples, the adjusted cost matrix corresponding to each training sample in the set is obtained.

[0084] When the training sample set of the node is divided into at least one subset using a certain attribute, for each subset, each training sample in the subset is traversed, and the predicted category is determined according to the ratio between the true category of the training sample and the predicted category. The corresponding misclassification cost is found in the adjusted cost matrix of the training sample, and the misclassification costs of all training samples in the current subset are summed and divided by the number of samples in the subset to obtain the cost.

[0085] The attribute that minimizes the sum of costs of all the subsample sets formed after splitting is taken as the splitting attribute of the node.

[0086] In the process of training a deep forest model, specifically in the construction of each decision tree, when a node of a tree gathers a batch of training samples and needs to select an optimal operating parameter, such as vibration amplitude or motor temperature, to distribute these samples, special attention needs to be paid to the personalized misclassification cost of each sample. Suppose the current node has 10 training samples, of which 3 samples are severe fault states collected when the flywheel is discharging at high speed and its health is low, and the other 7 are normal state samples under normal operation. For these 3 severe fault samples, their respective adjusted cost matrices will give extremely high penalty values ​​to those cases that misclassify severe faults as normal or minor faults, for example, a misclassification cost of 100. However, for the 7 normal samples, the cost matrix may only have a cost of 5 for misclassifying them as minor faults.

[0087] When evaluating the splitting of 10 samples at the current node using parameter A at a certain threshold, such as a temperature above 60 degrees Celsius, let's assume this split produces two subsets: a high-temperature subset containing 2 severely faulty samples and 2 normal samples, and a low-temperature subset containing 1 severely faulty sample and 5 normal samples. For the high-temperature subset, a provisional predicted category is determined based on the class ratio of these 4 samples within the subset, for example, 50% severely faulty and 50% normal, let's say severely faulty. Then, the cost is calculated for each of these 4 samples: the 2 samples that are truly severely faulty are correctly predicted, so the cost is 0; the 2 samples that are truly normal are predicted as severely faulty, and the misclassification cost is found from their respective adjusted cost matrices, for example, 20 for each, totaling 40. Therefore, the average cost of the high-temperature subset is 10. Similarly, the cost is calculated for the low-temperature subset, and the average cost of all subsets is summed. During this process, various combinations of parameters and thresholds are tried for splitting, such as... Figure 4 , Figure 5 As shown, the splitting method that minimizes the sum of costs of all the resulting subsample sets is ultimately selected, thereby constructing a decision tree that can effectively identify and avoid high-risk misjudgments.

[0088] In one embodiment, obtaining the operating status of the flywheel energy storage device based on the output of the self-organizing mapping network corresponding to the forest layer and the identification result of the forest layer includes:

[0089] The weight of each neuron in the self-organizing map network is determined based on the normalized frequency of activation of each neuron during training.

[0090] The feature vector output by each forest layer in the deep forest model is input into the self-organizing map network corresponding to that forest layer.

[0091] The state and state probability of the flywheel energy storage device are determined based on the output feature vector of each forest layer of the deep forest model, and the weights of the activated neurons of the self-organizing map network corresponding to the forest layer are obtained. The output of each forest layer is obtained by weighting the state probability with the weights.

[0092] The operating status of the flywheel energy storage device is obtained by summing and weighting the outputs of all forest layers.

[0093] A self-organizing map (SOM) network paired with a specific forest layer receives a large number of feature vectors from the output of that forest layer during the training phase. Assume this network has 100 output neurons. If neuron A is activated a total of 2000 times when processing feature vectors representing the normal high-speed operation of a flywheel, and the total number of activations for all neurons is 50,000, then neuron A's normalized activation frequency (NEM) is 0.04. 0.04 is neuron A's weight, representing its relative importance in recognizing features representing the normal high-speed operation state. Similarly, another neuron B might be activated 1500 times due to frequent responses to features representing a slightly unbalanced state, and its weight would be 0.03. Each neuron receives a weight based on its response frequency to a specific data pattern during training. In one embodiment, the deep forest and SOM are trained layer by layer, i.e., the first forest layer and its corresponding SOM are trained first, then the first forest layer, the second forest layer, and their corresponding SOMs are trained, iterating until all forest layers and SOMs are included.

[0094] During online monitoring, a certain layer of the deep forest model outputs a feature vector. This vector represents a preliminary state probability distribution, either from the forest layer itself or through a small MLP network. For example, the probability of a normal state is 0.7, the probability of a slight imbalance is 0.2, and the probability of a bearing failure is 0.1. Simultaneously, this feature vector is also fed into its corresponding self-organizing map network, such as... Figure 6 As shown. Suppose this vector activates the aforementioned neuron A, and neuron A is assigned a weight of 0.04. The initial state probabilities given by this forest layer—0.7 for normal, 0.2 for slight imbalance, and 0.1 for bearing failure—are then weighted and adjusted using the weight 0.04 to form the output of this forest layer after correction by the self-organizing map network. This process can be performed on each layer or selected forest layers in the deep forest. For example, another forest layer might output probabilities of 0.15 for normal, 0.6 for slight imbalance, and 0.25 for bearing failure, activating a neuron with a weight of 0.035 in its self-organizing map network; its output will also be adjusted using this weight. All adjusted forest layer outputs are then aggregated, for example, by summing the weighted probabilities of each state or by majority voting, combining the insights of all layers to arrive at the most reliable current operating state of the entire flywheel energy storage device.

[0095] The above embodiments are only used to illustrate the technical solutions of the present invention, and are not intended to limit it. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention. In addition, various different implementations of the embodiments of the present invention can be arbitrarily combined, as long as they do not violate the spirit of the embodiments of the present invention, and they should also be regarded as the content disclosed in the embodiments of the present invention.

Claims

1. A method for online monitoring of the operating status of a flywheel energy storage device, characterized in that, include: The operating parameters of the flywheel energy storage device are collected by sensors. Based on the frequency domain information entropy of each sensor signal, the signal is divided into multiple sub-frequency bands. The noise reduction method of the sub-frequency band is determined according to the characteristics of the sub-frequency band. The noise reduction data of each sub-frequency band of the same sensor is performed separately and then reconstructed to obtain the noise-reduced operating parameters. The operating phase and health index of the flywheel energy storage device corresponding to the training sample are obtained, and a cost matrix is ​​obtained based on the operating phase and the health index; the ratio of the number of samples in each operating state in the training sample subset where the current training sample is located is obtained, and an adjusted cost matrix is ​​determined based on the cost matrix and the ratio; the deep forest model is trained using the adjusted cost matrix. The noise-reduced operating parameters are input into the deep forest model, and the feature vectors output by each forest layer in the deep forest model are input into the self-organizing map network. The operating status of the flywheel energy storage device is obtained based on the output of the self-organizing map network corresponding to the forest layer and the identification result of the forest layer. The step of obtaining the operating status of the flywheel energy storage device based on the output of the self-organizing mapping network corresponding to the forest layer and the identification result of the forest layer includes: The weight of each neuron in the self-organizing map network is determined based on the normalized frequency of activation of each neuron during training. The feature vector output by each forest layer in the deep forest model is input into the self-organizing map network corresponding to that forest layer. The state and state probability of the flywheel energy storage device are determined based on the output feature vector of each forest layer of the deep forest model, and the weights of the activated neurons of the self-organizing map network corresponding to the forest layer are obtained. The output of each forest layer is obtained by weighting the state probability with the weights. The operating status of the flywheel energy storage device is obtained by summing and weighting the outputs of all forest layers.

2. The method according to claim 1, characterized in that, The frequency domain information entropy based on the signals from each sensor divides the signal into multiple sub-frequency bands, including: The amplitude spectrum of the signal from each sensor is calculated using Fast Fourier Transform, and the amplitude spectrum is divided into M equal-bandwidth frequency bands according to frequency. Calculate the proportion of energy at each frequency point within the frequency band to the total energy of the frequency band, and determine the frequency domain information entropy of each frequency band based on the proportion; Obtain the number N of target sub-bands, where N is less than M and M and N are both positive integers; with the optimization objective of maximizing the sum of information entropy within all target sub-bands, the process is carried out by iteratively merging adjacent frequency bands until N target sub-bands are obtained.

3. The method according to claim 1, characterized in that, The step of determining the sub-band noise reduction method based on the characteristics of the sub-band includes: Extract features from each sub-band, including energy center frequency, energy amplitude, spectral kurtosis value, and power spectral entropy value; The extracted features of each sub-frequency band are matched with the feature combinations in the noise reduction rule base to determine the noise reduction method and corresponding parameter set for each sub-frequency band. The determined noise reduction algorithm and parameter set are applied to the corresponding sub-frequency band data to perform noise reduction processing.

4. The method according to claim 1, characterized in that, The cost matrix obtained based on the operational phase and the health index includes: Determine the running stage index and health index level index of the current training sample, and retrieve the cost matrix of the current sample from the cost lookup table based on the running stage index and health index level index.

5. The method according to claim 1, characterized in that, The training of the deep forest model using the adjusted cost matrix includes: Within any decision tree of the deep forest, when selecting a split attribute for a node that aggregates a set of training samples, the adjusted cost matrix corresponding to each training sample in the set is obtained. When the training sample set of the node is divided into at least one subset using a certain attribute, for each subset, each training sample in the subset is traversed, and the predicted category is determined according to the ratio between the true category of the training sample and the predicted category. The corresponding misclassification cost is found in the adjusted cost matrix of the training sample, and the misclassification costs of all training samples in the current subset are summed and divided by the number of samples in the subset to obtain the cost. The attribute that minimizes the sum of costs of all the subsample sets formed after splitting is taken as the splitting attribute of the node.

6. An online monitoring system for the operating status of a flywheel energy storage device, characterized in that, include: The preprocessing unit is used to collect the operating parameters of the flywheel energy storage device through sensors, divide the signal into multiple sub-frequency bands based on the frequency domain information entropy of each sensor signal, determine the sub-frequency band noise reduction method according to the characteristics of the sub-frequency band, and reconstruct the noise-reduced operating parameters after performing noise reduction on the data of each sub-frequency band of the same sensor. The training unit is used to obtain the operating stage and health index of the flywheel energy storage device corresponding to the training sample, obtain the cost matrix based on the operating stage and the health index, obtain the ratio of the number of samples in each operating state in the training sample subset where the current training sample is located, determine the adjusted cost matrix based on the cost matrix and the ratio, and use the adjusted cost matrix to train the deep forest model. The identification unit is used to input the noise-reduced operating parameters into the deep forest model, input the feature vector output by each forest layer in the deep forest model into the self-organizing map network, and obtain the operating status of the flywheel energy storage device based on the output of the self-organizing map network corresponding to the forest layer and the identification result of the forest layer. The step of obtaining the operating status of the flywheel energy storage device based on the output of the self-organizing mapping network corresponding to the forest layer and the identification result of the forest layer includes: The weight of each neuron in the self-organizing map network is determined based on the normalized frequency of activation of each neuron during training. The feature vector output by each forest layer in the deep forest model is input into the self-organizing map network corresponding to that forest layer. The state and state probability of the flywheel energy storage device are determined based on the output feature vector of each forest layer of the deep forest model, and the weights of the activated neurons of the self-organizing map network corresponding to the forest layer are obtained. The output of each forest layer is obtained by weighting the state probability with the weights. The operating status of the flywheel energy storage device is obtained by summing and weighting the outputs of all forest layers.

7. The system according to claim 6, characterized in that, The frequency domain information entropy based on the signals from each sensor divides the signal into multiple sub-frequency bands, including: The amplitude spectrum of the signal from each sensor is calculated using Fast Fourier Transform, and the amplitude spectrum is divided into M equal-bandwidth frequency bands according to frequency. Calculate the proportion of energy at each frequency point within the frequency band to the total energy of the frequency band, and determine the frequency domain information entropy of each frequency band based on the proportion; Obtain the number N of target sub-bands, where N is less than M and M and N are both positive integers; with the optimization objective of maximizing the sum of information entropy within all target sub-bands, the process is carried out by iteratively merging adjacent frequency bands until N target sub-bands are obtained.

8. The system according to claim 6, characterized in that, The step of determining the sub-band noise reduction method based on the characteristics of the sub-band includes: Extract features from each sub-band, including energy center frequency, energy amplitude, spectral kurtosis value, and power spectral entropy value; The extracted features of each sub-frequency band are matched with the feature combinations in the noise reduction rule base to determine the noise reduction method and corresponding parameter set for each sub-frequency band. The determined noise reduction algorithm and parameter set are applied to the corresponding sub-frequency band data to perform noise reduction processing.

9. The system according to claim 6, characterized in that, The cost matrix obtained based on the operational phase and the health index includes: Determine the running stage index and health index level index of the current training sample, and retrieve the cost matrix of the current sample from the cost lookup table based on the running stage index and health index level index.

Images