Fan running state monitoring method based on artificial intelligence

Abstract

The invention discloses a fan running state monitoring method based on artificial intelligence. The method comprises the steps of data acquisition, data preprocessing, construction of a stacked sample convolutional interactive network model, construction of a running state monitoring model, design of a state abnormal threshold interval and fan running state monitoring. The invention relates to the technical field of fan operation state monitoring, in particular to a fan operation state monitoring method based on artificial intelligence, and the method comprises the steps: obtaining fan operation original data through data collection; a data preprocessing method of data cleaning, data denoising, feature selection, data normalization and data set segmentation is adopted; running state monitoring is performed by adopting a stacked sample convolutional interactive network model, so that the prediction precision of the model is enhanced; a federated learning framework is adopted to train a running state monitoring model, and data privacy is effectively guaranteed; a state anomaly threshold interval calculation method based on Monte Carlo discarding and kernel density estimation is adopted, and the state anomaly threshold interval can be dynamically adjusted.

Description

Technical Field

[0001] The present invention relates to the technical field of fan operation status monitoring, and in particular to a fan operation status monitoring method based on artificial intelligence. Background Art

[0002] Wind turbine operation status monitoring refers to the process of real-time monitoring and analysis of key parameters and equipment status of wind turbines during operation. Through sensors and monitoring systems, the vibration, temperature, speed and other data of wind turbines can be continuously tracked, and the data can be used to detect anomalies and potential faults. It can not only improve the operating efficiency of the equipment, but also extend the service life of the equipment, thereby maximizing the economic benefits and reliability of the wind power generation system.

[0003] However, traditional wind turbine operating status monitoring methods have the technical problem that they only focus on feature relationships in the low-dimensional spatial domain and extract features based on a single time resolution, making it difficult to effectively utilize features of different scales; traditional wind turbine operating status monitoring methods have the technical problem that a single wind turbine cannot collect sufficient and high-quality training data in a short period of time, and there is data privacy, resulting in low model accuracy; traditional wind turbine operating status monitoring methods have the technical problem that the state abnormality threshold is set as a static threshold, and inevitably introduces subjective uncertainty, resulting in poor monitoring performance and a high false alarm rate. Summary of the invention

[0004] In view of the above situation, in order to overcome the defects of the prior art, the present invention provides a wind turbine operating status monitoring method based on artificial intelligence. The traditional wind turbine operating status monitoring method only focuses on the feature relationship in the low-dimensional spatial domain and extracts features based on a single time resolution, and it is difficult to effectively utilize features of different scales. The present invention creatively adopts a stacked sample convolutional interactive network model for operating status monitoring, which can fully extract and fuse multi-dimensional and multi-resolution spatiotemporal features, thereby enhancing the prediction accuracy of the model; the traditional wind turbine operating status monitoring method has the problem that a single wind turbine cannot collect sufficient and high-quality training data in a short period of time, and there is data privacy, resulting in To address the technical problem of low model accuracy, this solution creatively adopts a federated learning framework to train the operation status monitoring model, which can effectively ensure data privacy while making full use of data collected by different wind turbines, and can obtain sufficient and high-quality training data more quickly; the traditional wind turbine operation status monitoring method has the problem that the state abnormality threshold is set as a static threshold, and inevitably introduces subjective uncertainty, resulting in poor monitoring performance and high false alarm rate. This solution creatively adopts a state abnormality threshold interval calculation method based on Monte Carlo discarding and kernel density estimation, which effectively quantifies the uncertainty in the prediction process, can dynamically adjust the state abnormality threshold interval, and improves the accuracy of the model.

[0005] The technical solution adopted by the present invention is as follows: The method for predicting enterprise customer churn based on big data provided by the present invention comprises the following steps:

[0006] Step S1: data collection;

[0007] Step S2: data preprocessing;

[0008] Step S3: constructing a stacked sample convolutional interaction network model;

[0009] Step S4: constructing an operation status monitoring model;

[0010] Step S5: designing the threshold interval of abnormal state;

[0011] Step S6: monitoring the operating status of the fan.

[0012] Further, in step S1, the data acquisition is used to acquire the original data required for monitoring the operation status of the wind turbine, specifically, to acquire multiple wind turbine operation original data sets from the monitoring control and data acquisition systems of multiple wind turbines in the same wind farm through data acquisition;

[0013] The multiple wind turbine operation original data sets, specifically, each wind turbine corresponds to a wind turbine operation original data set, and the wind turbine operation original data set specifically includes operation status data, environmental data and operation cycle data. The operation status data specifically includes speed, torque, power generation, lubricating oil level and vibration data; the environmental data specifically includes wind speed, temperature, air pressure and humidity; the operation cycle data specifically includes cumulative operation time and start and stop times.

[0014] Furthermore, in step S2, the data preprocessing is used to preprocess the collected raw data, and specifically includes the following steps:

[0015] Step S21: data cleaning, for cleaning the original data, specifically processing the missing values ​​and duplicate values ​​in each of the wind turbine operation original data sets to obtain a plurality of roughly processed data sets;

[0016] Step S22: data denoising, which is used to denoise the coarsely processed data, specifically using a Kalman filter algorithm to denoise each of the coarsely processed data sets to obtain multiple denoised data sets;

[0017] Step S23: feature selection, for selecting features, specifically, performing feature selection on each of the denoised data sets by calculating the Pearson correlation coefficient to obtain a plurality of fan operation feature sets;

[0018] Step S24: data normalization, which is used to normalize the wind turbine operation characteristics, specifically using the minimum-maximum method to normalize the data of each wind turbine operation characteristic set to obtain multiple normalized data sets;

[0019] Step S25: data set segmentation, for segmenting the data set, specifically segmenting each of the normalized data sets to obtain a plurality of wind turbine state training sets and a plurality of wind turbine state test sets;

[0020] Step S26: preprocessing, specifically preprocessing the plurality of wind turbine operation original data sets through the data cleaning, the data denoising, the feature selection, the data normalization and the data set segmentation to obtain a plurality of wind turbine status training sets and a plurality of wind turbine status test sets.

[0021] Further, in step S3, the stacked sample convolutional interaction network model is constructed, which is used to construct the stacked sample convolutional interaction network model, and specifically includes the following steps:

[0022] Step S31: Design a coordinate attention mechanism, the steps include:

[0023] Step S311: Coordinate attention pooling, specifically using a one-dimensional average pooling operation with a pooling kernel of (1, W) and a one-dimensional average pooling operation with a pooling kernel of (H, 1) to pool the input feature map of the coordinate attention mechanism Processing, in which represents the input feature map of the coordinate attention mechanism, C represents the number of channels of the input feature map of the coordinate attention mechanism, H represents the height of the input feature map of the coordinate attention mechanism, W represents the width of the input feature map of the coordinate attention mechanism, and the input feature map of the coordinate attention mechanism At the output of the pooling operation in each channel, the calculation formula is as follows:

[0024] ;

[0025] In the formula, represents the output of the pooling operation with height h on the cth channel, represents the output of the pooling operation with width w on the cth channel, H represents the height of the input feature map of the coordinate attention mechanism, W represents the width of the input feature map of the coordinate attention mechanism, Represents the value of the coordinate (h,i) position of the input feature map of the coordinate attention mechanism on the cth channel, Represents the value of the coordinate (j, w) position on the cth channel of the input feature map of the coordinate attention mechanism;

[0026] Step S312: coordinate attention feature fusion, the formula used is as follows:

[0027] ;

[0028] In the formula, represents the coordinate attention fusion feature of the c-th channel, represents the sigmoid function, represents the convolution operation function with a convolution kernel size of 1×1, Represents a splicing operation;

[0029] Step S313: Calculate the weight matrix, specifically divide the coordinate attention fusion feature into two independent tensors along the spatial dimension, and calculate the weight matrix. The formula used for calculating the weight matrix is ​​as follows:

[0030] ;

[0031] In the formula, represents the coordinate attention weight matrix, represents the tensor of the coordinate attention fusion feature divided along the height, A tensor representing the coordinate attention fusion feature divided along the width;

[0032] Step S314: Calculate the coordinate attention mechanism output, the formula used is as follows:

[0033] ;

[0034] In the formula, represents the coordinate attention mechanism output, represents element-wise multiplication;

[0035] Step S32: constructing a sample convolution interaction submodule, the steps include:

[0036] Step S321: performing even-odd extraction, specifically, downsampling the input time series of the submodule into an odd subsequence and an even subsequence by using even-odd extraction;

[0037] Step S322: interactive learning, the formula used is as follows:

[0038] ;

[0039] In the formula, represents an odd subsequence, represents an even sequence, represents the hyperbolic tangent function, Indicates that the convolution kernel size is The one-dimensional convolution operation function, Indicates that the convolution kernel size is The one-dimensional convolution operation function, represents the Mish activation function, represents the odd subsequence after interactive learning, represents the even-child sequence after interactive learning;

[0040] Step S323: Calculate the submodule output using the following formula:

[0041] ;

[0042] In the formula, represents the odd subsequence output of the submodule, represents the even subsequence output of the submodule, Indicates that the convolution kernel size is The one-dimensional convolution operation function, Indicates that the convolution kernel size is One-dimensional convolution operation function;

[0043] Step S33: constructing a coordinate attention sample convolution interaction module, specifically, constructing a coordinate attention sample convolution interaction module through the design of the coordinate attention mechanism and the construction of the sample convolution interaction submodule to obtain a coordinate attention sample convolution interaction module;

[0044] Step S34: constructing a stacked sample convolutional interaction network model, specifically by constructing an encoder structure and a decoder structure to construct the stacked sample convolutional interaction network model to obtain a running status monitoring model, the encoder structure specifically includes a multi-layer coordinate attention sample convolutional interaction module and a connection module, the decoder structure is specifically a fully connected layer with a residual link, the multi-layer coordinate attention sample convolutional interaction module is specifically arranged according to a binary tree structure, and the connection module is used to connect the outputs of the multi-layer coordinate attention sample convolutional interaction module;

[0045] Step S35: Calculate the model loss using the following formula:

[0046] ;

[0047] In the formula, Loss represents the model loss, T represents the length of the time series, and L represents the total number of layers of the multi-layer coordinate attention sample convolution interaction module. represents the t-th predicted value of the l-th layer, Represents the tth actual value.

[0048] Further, in step S4, the operation status monitoring model is constructed to construct a model required for monitoring the operation status of the wind turbine, specifically, the operation status monitoring model is obtained by constructing a federated learning framework;

[0049] The operation status monitoring model is constructed, specifically comprising the following steps:

[0050] Step S41: constructing a main model and a sub-model, specifically constructing a single main model and multiple sub-models, the main model adopts the stacked sample convolutional interaction network model, and the sub-model adopts the stacked sample convolutional interaction network model;

[0051] Step S42: sub-model training, specifically, each sub-model is trained on a wind turbine status training set corresponding to a wind turbine, and the wind turbine status test set is used to evaluate the model performance, and the trained sub-model parameters are sent to the main model;

[0052] Step S43: updating the main model parameters, specifically updating the main model parameters based on the sub-model parameters, and sending the updated main model parameters to the sub-models for a new round of training. The formula used for updating the main model parameters is as follows:

[0053] ;

[0054] In the formula, represents the main model parameters of the d+1th training round, represents the sub-model parameters of the n-th sub-model in the d-th training round, represents the data volume of the wind turbine status training set of the nth sub-model, M represents the total data volume of all wind turbine status training sets, N represents the total number of sub-models, and d represents the current training round;

[0055] Step S44: constructing an operation status monitoring model, specifically by constructing a main model and a sub-model, training the sub-model and updating the main model parameters, building a federated learning framework and performing training, and using the trained main model as the operation status monitoring model.

[0056] Further, in step S5, the state abnormality threshold interval design is used to design the state abnormality threshold interval, which specifically includes the following steps:

[0057] Step S51: Monte Carlo discard design, specifically, performing multiple forward propagations on each input, in which the weight of the operating status monitoring model is regarded as a random variable and obeys the Bernoulli distribution in each forward propagation. The formula used in the Monte Carlo discard design is as follows;

[0058] ;

[0059] In the formula, represents the weight of the running state monitoring model during the qth forward propagation, p represents the probability that the weight of the running state monitoring model is retained, Represents the weight of the operating status monitoring model during the qth forward propagation The probability that the weights of the operating status monitoring model are retained is p.

[0060] Step S52: Calculate the model prediction value, the formula used is as follows:

[0061] ;

[0062] Where x represents the input of the operation status monitoring model, represents the predicted value of the running state monitoring model input x after the introduction of Monte Carlo discard, Q represents the total number of forward propagation, Indicates that the weight of the operating status monitoring model is the weight of the operating status monitoring model during the qth forward propagation The predicted value of the running state monitoring model input x when the running state monitoring model is running;

[0063] Step S53: Calculate the probability density function, specifically by Considered as the predicted value about the time step, the probability density function is calculated based on the kernel density estimation method, and the formula used is as follows:

[0064] ;

[0065] In the formula, represents the probability density function, b represents the bandwidth coefficient of the Gaussian kernel function, represents the Gaussian kernel function, Indicates that the weight of the running status monitoring model is the weight of the running status monitoring model during the first forward propagation The predicted value of the running state monitoring model input x when the running state monitoring model is Indicates that the weight of the operating status monitoring model is the weight of the operating status monitoring model during the Qth forward propagation The predicted value of the running state monitoring model input x when the running state monitoring model is running;

[0066] Step S54: Calculate the state abnormality threshold interval, the formula used is as follows:

[0067] ;

[0068] In the formula, V represents the abnormal state threshold interval, Indicates the lower limit of the abnormal status threshold. Indicates the upper limit of the abnormal status threshold. Represents the probability density function Integrate to get the cumulative distribution function , represents the cumulative distribution function The inverse function of Indicates the confidence level.

[0069] Further, in step S6, the wind turbine operating status monitoring is specifically to use the operating status monitoring model to monitor the wind turbine operating status. If the model prediction value of the operating status monitoring model does not continuously exceed the abnormal state threshold interval, the wind turbine operates normally. If the model prediction value of the operating status monitoring model continuously exceeds the abnormal state threshold interval, the wind turbine operates abnormally.

[0070] The beneficial effects achieved by the present invention using the above scheme are as follows:

[0071] (1) In view of the technical problems that traditional wind turbine operating status monitoring methods only focus on feature relationships in the low-dimensional spatial domain and extract features based on a single time resolution, making it difficult to effectively utilize features of different scales, this scheme creatively adopts a stacked sample convolutional interaction network model for operating status monitoring, which can fully extract and fuse multi-dimensional and multi-resolution spatiotemporal features, thereby enhancing the prediction accuracy of the model.

[0072] (2) In view of the technical problem that a single wind turbine cannot collect sufficient and high-quality training data in a short period of time, and there is data privacy, resulting in low model accuracy in traditional wind turbine operating status monitoring methods, this solution creatively adopts a federated learning framework to train the operating status monitoring model. While effectively ensuring data privacy, it can make full use of the data collected by different wind turbines and obtain sufficient and high-quality training data more quickly.

[0073] (3) The traditional wind turbine operating status monitoring method has the problem that the state abnormality threshold is set as a static threshold, which inevitably introduces subjective uncertainty, resulting in poor monitoring performance and high false alarm rate. This scheme creatively adopts a state abnormality threshold interval calculation method based on Monte Carlo discarding and kernel density estimation, which effectively quantifies the uncertainty in the prediction process, can dynamically adjust the state abnormality threshold interval, and improves the accuracy of the model. BRIEF DESCRIPTION OF THE DRAWINGS

[0074] Figure 1 A schematic diagram of the flow of the fan operating status monitoring method based on artificial intelligence provided by the present invention;

[0075] Figure 2 This is a schematic diagram of the process of data preprocessing in step S2;

[0076] Figure 3 Schematic diagram of the process of building a convolutional interaction network model of stacked samples in step S3;

[0077] Figure 4 A schematic diagram of the process of constructing the operating status monitoring model in step S4;

[0078] Figure 5Schematic diagram of the process designed for the abnormal state threshold interval of step S5.

[0079] The accompanying drawings are used to provide further understanding of the present invention and constitute a part of the specification. They are used to explain the present invention together with the embodiments of the present invention and do not constitute a limitation of the present invention. DETAILED DESCRIPTION

[0080] The technical solutions in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, rather than all the embodiments; based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative work are within the scope of protection of the present invention.

[0081] In the description of the present invention, it should be understood that terms such as “upper”, “lower”, “front”, “back”, “left”, “right”, “top”, “bottom”, “inside” and “outside” indicating directions or positional relationships are based on the directions or positional relationships shown in the accompanying drawings, and are only for the convenience of describing the present invention and simplifying the description, rather than indicating or implying that the device or element referred to must have a specific direction, be constructed and operated in a specific direction, and therefore should not be understood as limiting the present invention.

[0082] Example 1, see Figure 1 The technical solution adopted by the present invention is as follows: The enterprise customer churn prediction method based on big data provided by the present invention comprises the following steps:

[0083] Step S1: data collection;

[0084] Step S2: data preprocessing;

[0085] Step S3: constructing a stacked sample convolutional interaction network model;

[0086] Step S4: constructing an operation status monitoring model;

[0087] Step S5: designing the threshold interval of abnormal state;

[0088] Step S6: monitoring the operating status of the fan.

[0089] Example 2, see Figure 1 In step S1, the data acquisition is used to acquire the original data required for monitoring the operation status of the wind turbine, specifically, to acquire multiple wind turbine operation original data sets from the monitoring control and data acquisition systems of multiple wind turbines in the same wind farm through data acquisition;

[0090] The multiple wind turbine operation original data sets, specifically, each wind turbine corresponds to a wind turbine operation original data set, and the wind turbine operation original data set specifically includes operation status data, environmental data and operation cycle data. The operation status data specifically includes speed, torque, power generation, lubricating oil level and vibration data; the environmental data specifically includes wind speed, temperature, air pressure and humidity; the operation cycle data specifically includes cumulative operation time and start and stop times.

[0091] Example 3, see Figure 1 and Figure 2 This embodiment is based on the above embodiment. In step S2, the data preprocessing is used to preprocess the collected raw data, and specifically includes the following steps:

[0092] Step S21: data cleaning, for cleaning the original data, specifically processing the missing values ​​and duplicate values ​​in each of the wind turbine operation original data sets to obtain a plurality of roughly processed data sets;

[0093] Step S22: data denoising, which is used to denoise the coarsely processed data, specifically using a Kalman filter algorithm to denoise each of the coarsely processed data sets to obtain multiple denoised data sets;

[0094] Step S23: feature selection, for selecting features, specifically, performing feature selection on each of the denoised data sets by calculating the Pearson correlation coefficient to obtain a plurality of fan operation feature sets;

[0095] Step S24: data normalization, which is used to normalize the wind turbine operation characteristics, specifically using the minimum-maximum method to normalize the data of each wind turbine operation characteristic set to obtain multiple normalized data sets;

[0096] Step S25: data set segmentation, for segmenting the data set, specifically segmenting each of the normalized data sets to obtain a plurality of wind turbine state training sets and a plurality of wind turbine state test sets;

[0097] Step S26: preprocessing, specifically preprocessing the plurality of wind turbine operation original data sets through the data cleaning, the data denoising, the feature selection, the data normalization and the data set segmentation to obtain a plurality of wind turbine status training sets and a plurality of wind turbine status test sets.

[0098] Example 4, see Figure 1 and Figure 3 This embodiment is based on the above embodiment. In step S3, the stacked sample convolutional interaction network model is constructed to construct a stacked sample convolutional interaction network model, which specifically includes the following steps:

[0099] Step S31: Design a coordinate attention mechanism, the steps include:

[0100] Step S311: Coordinate attention pooling, specifically using a one-dimensional average pooling operation with a pooling kernel of (1, W) and a one-dimensional average pooling operation with a pooling kernel of (H, 1) to pool the input feature map of the coordinate attention mechanism Processing, in which represents the input feature map of the coordinate attention mechanism, C represents the number of channels of the input feature map of the coordinate attention mechanism, H represents the height of the input feature map of the coordinate attention mechanism, W represents the width of the input feature map of the coordinate attention mechanism, and the input feature map of the coordinate attention mechanism At the output of the pooling operation in each channel, the calculation formula is as follows:

[0101] ;

[0102] In the formula, represents the output of the pooling operation with height h on the cth channel, represents the output of the pooling operation with width w on the cth channel, H represents the height of the input feature map of the coordinate attention mechanism, W represents the width of the input feature map of the coordinate attention mechanism, Represents the value of the coordinate (h,i) position of the input feature map of the coordinate attention mechanism on the cth channel, Represents the value of the coordinate (j, w) position on the cth channel of the input feature map of the coordinate attention mechanism;

[0103] Step S312: coordinate attention feature fusion, the formula used is as follows:

[0104] ;

[0105] In the formula, represents the coordinate attention fusion feature of the c-th channel, represents the sigmoid function, represents the convolution operation function with a convolution kernel size of 1×1, Represents a splicing operation;

[0106] Step S313: Calculate the weight matrix, specifically divide the coordinate attention fusion feature into two independent tensors along the spatial dimension, and calculate the weight matrix. The formula used for calculating the weight matrix is ​​as follows:

[0107] ;

[0108] In the formula, represents the coordinate attention weight matrix, represents the tensor of the coordinate attention fusion feature divided along the height, A tensor representing the coordinate attention fusion feature divided along the width;

[0109] Step S314: Calculate the coordinate attention mechanism output, the formula used is as follows:

[0110] ;

[0111] In the formula, represents the coordinate attention mechanism output, represents element-wise multiplication;

[0112] Step S32: constructing a sample convolution interaction submodule, the steps include:

[0113] Step S321: performing even-odd extraction, specifically, downsampling the input time series of the submodule into an odd subsequence and an even subsequence by using even-odd extraction;

[0114] Step S322: interactive learning, the formula used is as follows:

[0115] ;

[0116] In the formula, represents an odd subsequence, represents an even sequence, represents the hyperbolic tangent function, Indicates that the convolution kernel size is The one-dimensional convolution operation function, Indicates that the convolution kernel size is The one-dimensional convolution operation function, represents the Mish activation function, represents the odd subsequence after interactive learning, represents the even-child sequence after interactive learning;

[0117] Step S323: Calculate the submodule output using the following formula:

[0118] ;

[0119] In the formula, represents the odd subsequence output of the submodule, represents the even subsequence output of the submodule, Indicates that the convolution kernel size is The one-dimensional convolution operation function, Indicates that the convolution kernel size is One-dimensional convolution operation function;

[0120] Step S33: constructing a coordinate attention sample convolution interaction module, specifically, constructing a coordinate attention sample convolution interaction module by designing the coordinate attention mechanism and constructing the sample convolution interaction submodule to obtain a coordinate attention sample convolution interaction module;

[0121] Step S34: constructing a stacked sample convolutional interaction network model, specifically by constructing an encoder structure and a decoder structure to construct the stacked sample convolutional interaction network model to obtain a running status monitoring model, the encoder structure specifically includes a multi-layer coordinate attention sample convolutional interaction module and a connection module, the decoder structure is specifically a fully connected layer with a residual link, the multi-layer coordinate attention sample convolutional interaction module is specifically arranged according to a binary tree structure, and the connection module is used to connect the outputs of the multi-layer coordinate attention sample convolutional interaction module;

[0122] Step S35: Calculate the model loss using the following formula:

[0123] ;

[0124] In the formula, Loss represents the model loss, T represents the length of the time series, and L represents the total number of layers of the multi-layer coordinate attention sample convolution interaction module. represents the t-th predicted value of the l-th layer, Represents the tth actual value.

[0125] By performing the above operations, this solution creatively adopts a stacked sample convolutional interaction network model for operation status monitoring, which can fully extract and fuse multi-dimensional and multi-resolution spatiotemporal features and enhance the prediction accuracy of the model.

[0126] Example 5, see Figure 1 and Figure 4 , this embodiment is based on the above embodiment, in step S4, the operation status monitoring model is constructed to construct a model required for monitoring the operation status of the wind turbine, specifically, the operation status monitoring model is obtained by constructing a federated learning framework;

[0127] The operation status monitoring model is constructed, specifically comprising the following steps:

[0128] Step S41: constructing a main model and a sub-model, specifically constructing a single main model and multiple sub-models, the main model adopts the stacked sample convolutional interaction network model, and the sub-model adopts the stacked sample convolutional interaction network model;

[0129] Step S42: sub-model training, specifically, each sub-model is trained on a wind turbine status training set corresponding to a wind turbine, and the wind turbine status test set is used to evaluate the model performance, and the trained sub-model parameters are sent to the main model;

[0130] Step S43: updating the main model parameters, specifically updating the main model parameters based on the sub-model parameters, and sending the updated main model parameters to the sub-models for a new round of training. The formula used for updating the main model parameters is as follows:

[0131] ;

[0132] In the formula, represents the main model parameters of the d+1th training round, represents the sub-model parameters of the n-th sub-model in the d-th training round, represents the data volume of the wind turbine status training set of the nth sub-model, M represents the total data volume of all wind turbine status training sets, N represents the total number of sub-models, and d represents the current training round;

[0133] Step S44: constructing an operation status monitoring model, specifically by constructing a main model and a sub-model, training the sub-model and updating the main model parameters, building a federated learning framework and performing training, and using the trained main model as the operation status monitoring model.

[0134] By performing the above operations, this solution creatively uses a federated learning framework to train the operation status monitoring model, which can effectively ensure data privacy while making full use of the data collected by different wind turbines, and can obtain sufficient and high-quality training data more quickly.

[0135] Example 6, see Figure 1 and Figure 5 This embodiment is based on the above embodiment. In step S5, the state abnormality threshold interval design is used to design the state abnormality threshold interval, and specifically includes the following steps:

[0136] Step S51: Monte Carlo discard design, specifically, performing multiple forward propagations on each input, in which the weight of the operating status monitoring model is regarded as a random variable and obeys the Bernoulli distribution in each forward propagation. The formula used in the Monte Carlo discard design is as follows;

[0137] ;

[0138] In the formula, represents the weight of the running state monitoring model during the qth forward propagation, p represents the probability that the weight of the running state monitoring model is retained, Represents the weight of the operating status monitoring model during the qth forward propagation The probability that the weights of the operating status monitoring model are retained is p.

[0139] Step S52: Calculate the model prediction value, the formula used is as follows:

[0140] ;

[0141] Where x represents the input of the operation status monitoring model, represents the predicted value of the running state monitoring model input x after the introduction of Monte Carlo discard, Q represents the total number of forward propagation, Indicates that the weight of the operating status monitoring model is the weight of the operating status monitoring model during the qth forward propagation The predicted value of the running state monitoring model input x when the running state monitoring model is running;

[0142] Step S53: Calculate the probability density function, specifically by Considered as the predicted value about the time step, the probability density function is calculated based on the kernel density estimation method, and the formula used is as follows:

[0143] ;

[0144] In the formula, represents the probability density function, b represents the bandwidth coefficient of the Gaussian kernel function, represents the Gaussian kernel function, Indicates that the weight of the running status monitoring model is the weight of the running status monitoring model during the first forward propagation The predicted value of the running state monitoring model input x when the running state monitoring model is Indicates that the weight of the operating status monitoring model is the weight of the operating status monitoring model during the Qth forward propagation The predicted value of the running state monitoring model input x when the running state monitoring model is running;

[0145] Step S54: Calculate the state abnormality threshold interval, the formula used is as follows:

[0146] ;

[0147] In the formula, V represents the abnormal state threshold interval, Indicates the lower limit of the abnormal status threshold. Indicates the upper limit of the abnormal status threshold. Represents the probability density function Integrate to get the cumulative distribution function , represents the cumulative distribution function The inverse function of Indicates the confidence level.

[0148] By performing the above operations, the traditional wind turbine operation status monitoring method has the problem that the state abnormality threshold is set as a static threshold, and inevitably introduces subjective uncertainty, resulting in poor monitoring performance and high false alarm rate. This scheme creatively adopts a state abnormality threshold interval calculation method based on Monte Carlo discarding and kernel density estimation, which effectively quantifies the uncertainty in the prediction process, can dynamically adjust the state abnormality threshold interval, and improves the accuracy of the model.

[0149] Embodiment 7, see Figure 1 This embodiment is based on the above embodiment. In step S6, the wind turbine operating status monitoring is specifically to use the operating status monitoring model to monitor the wind turbine operating status. If the model prediction value of the operating status monitoring model does not continuously exceed the abnormal state threshold interval, the wind turbine operates normally. If the model prediction value of the operating status monitoring model continuously exceeds the abnormal state threshold interval, the wind turbine operates abnormally.

[0150] It should be noted that, in this article, relational terms such as first and second, etc. are only used to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Moreover, the terms "include", "comprise" or any other variants thereof are intended to cover non-exclusive inclusion, so that a process, method, article or device including a series of elements includes not only those elements, but also other elements not explicitly listed, or also includes elements inherent to such process, method, article or device.

[0151] While the embodiments of the present invention have been shown and described, it will be appreciated by those skilled in the art that many changes, modifications, substitutions and variations can be made to the embodiments without departing from the principles and spirit of the invention.

[0152] The present invention and its embodiments are described above, and such description is not restrictive. The drawings show only one embodiment of the present invention, and the actual structure is not limited thereto. In short, if ordinary technicians in the field are inspired by it, without departing from the purpose of the invention, they can design a structure and embodiment similar to the technical solution without creativity, which should belong to the protection scope of the present invention.

Claims

1. A fan operation status monitoring method based on artificial intelligence, characterized in that: The method comprises the following steps: Step S1: data collection; Step S2: data preprocessing, performing data cleaning, data denoising, feature selection, data normalization and data set segmentation on the collected original data of fan operation to obtain multiple fan state training sets and multiple fan state test sets; Step S3: constructing a stacked sample convolutional interaction network model, specifically, obtaining a stacked sample convolutional interaction network model by designing a coordinate attention mechanism and constructing a sample convolutional interaction submodule; Step S4: constructing an operation status monitoring model, specifically, obtaining an operation status monitoring model by constructing a federated learning framework; Step S5: designing the state abnormality threshold interval, specifically, obtaining the state abnormality threshold interval by Monte Carlo discarding design and calculating the probability density function based on the kernel density estimation method; Step S6: monitoring the operating status of the fan, specifically monitoring the operating status of the fan through the operating status monitoring model, and judging the operating status of the fan in combination with the abnormal status threshold interval.

2. The method for monitoring the operation status of a fan based on artificial intelligence according to claim 1, characterized in that: In step S3, the stacked sample convolutional interaction network model is constructed, which is used to construct the stacked sample convolutional interaction network model, and specifically includes the following steps: Step S31: Design a coordinate attention mechanism, the steps include: Step S311: Coordinate attention pooling, specifically using a one-dimensional average pooling operation with a pooling kernel of (1, W) and a one-dimensional average pooling operation with a pooling kernel of (H, 1) to pool the input feature map of the coordinate attention mechanism Processing, in which represents the input feature map of the coordinate attention mechanism, C represents the number of channels of the input feature map of the coordinate attention mechanism, H represents the height of the input feature map of the coordinate attention mechanism, W represents the width of the input feature map of the coordinate attention mechanism, and the input feature map of the coordinate attention mechanism At the output of the pooling operation in each channel, the calculation formula is as follows: ； In the formula, represents the output of the pooling operation with height h on the cth channel, represents the output of the pooling operation with width w on the cth channel, H represents the height of the input feature map of the coordinate attention mechanism, W represents the width of the input feature map of the coordinate attention mechanism, Represents the value of the coordinate (h,i) position of the input feature map of the coordinate attention mechanism on the cth channel, Represents the value of the coordinate (j, w) position on the cth channel of the input feature map of the coordinate attention mechanism; Step S312: coordinate attention feature fusion, the formula used is as follows: ； In the formula, represents the coordinate attention fusion feature of the c-th channel, represents the sigmoid function, represents the convolution operation function with a convolution kernel size of 1×1, Represents a splicing operation; Step S313: Calculate the weight matrix, specifically divide the coordinate attention fusion feature into two independent tensors along the spatial dimension, and calculate the weight matrix. The formula used for calculating the weight matrix is ​​as follows: ； In the formula, represents the coordinate attention weight matrix, represents the tensor of the coordinate attention fusion feature divided along the height, A tensor representing the coordinate attention fusion feature divided along the width; Step S314: Calculate the coordinate attention mechanism output, the formula used is as follows: ； In the formula, represents the coordinate attention mechanism output, represents element-wise multiplication; Step S32: constructing a sample convolution interaction submodule, the steps include: Step S321: performing even-odd extraction, specifically, downsampling the input time series of the submodule into an odd subsequence and an even subsequence by using even-odd extraction; Step S322: interactive learning, the formula used is as follows: ； In the formula, represents an odd subsequence, represents an even sequence, represents the hyperbolic tangent function, Indicates that the convolution kernel size is The one-dimensional convolution operation function, Indicates that the convolution kernel size is The one-dimensional convolution operation function, represents the Mish activation function, represents the odd subsequence after interactive learning, represents the even-child sequence after interactive learning; Step S323: Calculate the submodule output using the following formula: ； In the formula, represents the odd subsequence output of the submodule, represents the even subsequence output of the submodule, Indicates that the convolution kernel size is The one-dimensional convolution operation function, Indicates that the convolution kernel size is One-dimensional convolution operation function; Step S33: constructing a coordinate attention sample convolution interaction module, specifically, constructing a coordinate attention sample convolution interaction module through the design of the coordinate attention mechanism and the construction of the sample convolution interaction submodule to obtain a coordinate attention sample convolution interaction module; Step S34: constructing a stacked sample convolutional interaction network model, specifically by constructing an encoder structure and a decoder structure to construct the stacked sample convolutional interaction network model to obtain a running status monitoring model, the encoder structure specifically includes a multi-layer coordinate attention sample convolutional interaction module and a connection module, the decoder structure is specifically a fully connected layer with a residual link, the multi-layer coordinate attention sample convolutional interaction module is specifically arranged according to a binary tree structure, and the connection module is used to connect the outputs of the multi-layer coordinate attention sample convolutional interaction module; Step S35: Calculate the model loss using the following formula: ； In the formula, Loss represents the model loss, T represents the length of the time series, and L represents the total number of layers of the multi-layer coordinate attention sample convolution interaction module. represents the t-th predicted value of the l-th layer, Represents the tth actual value.

3. The method for monitoring the operation status of a fan based on artificial intelligence according to claim 1, characterized in that: In step S4, the operation status monitoring model is constructed to construct a model required for monitoring the operation status of the wind turbine, specifically, the operation status monitoring model is obtained by constructing a federated learning framework; The operation status monitoring model is constructed, specifically comprising the following steps: Step S41: constructing a main model and a sub-model, specifically constructing a single main model and multiple sub-models, the main model adopts the stacked sample convolutional interaction network model, and the sub-model adopts the stacked sample convolutional interaction network model; Step S42: sub-model training, specifically, each sub-model is trained on a wind turbine status training set corresponding to a wind turbine, and the wind turbine status test set is used to evaluate the model performance, and the trained sub-model parameters are sent to the main model; Step S43: updating the main model parameters, specifically updating the main model parameters based on the sub-model parameters, and sending the updated main model parameters to the sub-models for a new round of training. The formula used for updating the main model parameters is as follows: ； In the formula, represents the main model parameters of the d+1th training round, represents the sub-model parameters of the n-th sub-model in the d-th training round, represents the data volume of the wind turbine status training set of the nth sub-model, M represents the total data volume of all wind turbine status training sets, N represents the total number of sub-models, and d represents the current training round; Step S44: constructing an operation status monitoring model, specifically by constructing a main model and a sub-model, training the sub-model and updating the main model parameters, building a federated learning framework and performing training, and using the trained main model as the operation status monitoring model.

4. The method for monitoring the operation status of a fan based on artificial intelligence according to claim 1, characterized in that: In step S5, the state abnormality threshold interval design is used to design the state abnormality threshold interval, which specifically includes the following steps: Step S51: Monte Carlo discard design, specifically, performing multiple forward propagations on each input, in which the weight of the operating status monitoring model is regarded as a random variable and obeys the Bernoulli distribution in each forward propagation. The formula used in the Monte Carlo discard design is as follows; ； In the formula, represents the weight of the running state monitoring model during the qth forward propagation, p represents the probability that the weight of the running state monitoring model is retained, Represents the weight of the operating status monitoring model during the qth forward propagation The probability that the weights of the operating status monitoring model are retained is p. Step S52: Calculate the model prediction value, the formula used is as follows: ； Where x represents the input of the operation status monitoring model, represents the predicted value of the running state monitoring model input x after the introduction of Monte Carlo discard, Q represents the total number of forward propagation, Indicates that the weight of the operating status monitoring model is the weight of the operating status monitoring model during the qth forward propagation The predicted value of the running state monitoring model input x when the running state monitoring model is running; Step S53: Calculate the probability density function, specifically by Considered as the predicted value about the time step, the probability density function is calculated based on the kernel density estimation method, and the formula used is as follows: ； In the formula, represents the probability density function, b represents the bandwidth coefficient of the Gaussian kernel function, represents the Gaussian kernel function, Indicates that the weight of the running status monitoring model is the weight of the running status monitoring model during the first forward propagation The predicted value of the running state monitoring model input x when the running state monitoring model is Indicates that the weight of the operating status monitoring model is the weight of the operating status monitoring model during the Qth forward propagation The predicted value of the running state monitoring model input x when the running state monitoring model is running; Step S54: Calculate the abnormal state threshold interval, the formula used is as follows: ； In the formula, V represents the abnormal state threshold interval, Indicates the lower limit of the abnormal status threshold. Indicates the upper limit of the abnormal status threshold. Represents the probability density function Integrate to get the cumulative distribution function , represents the cumulative distribution function The inverse function of Indicates the confidence level.

5. The method for monitoring the operation status of a fan based on artificial intelligence according to claim 1, characterized in that: In step S1, the data acquisition is used to acquire the original data required for monitoring the operation status of the wind turbine, specifically, to acquire multiple wind turbine operation original data sets from the monitoring control and data acquisition systems of multiple wind turbines in the same wind farm through data acquisition; The multiple wind turbine operation original data sets, specifically, each wind turbine corresponds to a wind turbine operation original data set, and the wind turbine operation original data set specifically includes operation status data, environmental data and operation cycle data. The operation status data specifically includes speed, torque, power generation, lubricating oil level and vibration data; the environmental data specifically includes wind speed, temperature, air pressure and humidity; the operation cycle data specifically includes cumulative operation time and start and stop times.

6. The method for monitoring the operation status of a fan based on artificial intelligence according to claim 1, characterized in that: In step S2, the data preprocessing is used to preprocess the collected raw data, and specifically includes the following steps: Step S21: data cleaning, for cleaning the original data, specifically processing the missing values ​​and duplicate values ​​in each of the wind turbine operation original data sets to obtain a plurality of roughly processed data sets; Step S22: data denoising, which is used to denoise the coarsely processed data, specifically using a Kalman filter algorithm to denoise each of the coarsely processed data sets to obtain multiple denoised data sets; Step S23: feature selection, for selecting features, specifically, performing feature selection on each of the denoised data sets by calculating the Pearson correlation coefficient to obtain a plurality of fan operation feature sets; Step S24: data normalization, which is used to normalize the wind turbine operation characteristics, specifically using the minimum-maximum method to normalize the data of each wind turbine operation characteristic set to obtain multiple normalized data sets; Step S25: data set segmentation, for segmenting the data set, specifically segmenting each of the normalized data sets to obtain a plurality of wind turbine state training sets and a plurality of wind turbine state test sets; Step S26: preprocessing, specifically preprocessing the plurality of wind turbine operation original data sets through the data cleaning, the data denoising, the feature selection, the data normalization and the data set segmentation to obtain a plurality of wind turbine status training sets and a plurality of wind turbine status test sets.

7. The method for monitoring the operation status of a fan based on artificial intelligence according to claim 1, characterized in that: In step S6, the wind turbine operating status monitoring is specifically to use the operating status monitoring model to monitor the wind turbine operating status. If the model prediction value of the operating status monitoring model does not continuously exceed the abnormal state threshold interval, the wind turbine operates normally. If the model prediction value of the operating status monitoring model continuously exceeds the abnormal state threshold interval, the wind turbine operates abnormally.

Images