Device fault diagnosis method based on associated hidden feature map

Abstract

The application relates to a device fault diagnosis method based on an associated hidden feature map and belongs to the technical field of device fault diagnosis. The method solves the problems that label data is difficult to obtain in the prior art, there are many monitoring variables, and the monitoring variables have an associated relationship, so that the device fault diagnosis precision is low. The fault diagnosis method comprises the following steps: obtaining to-be-diagnosed fault data of a to-be-diagnosed device; constructing a to-be-diagnosed nested self-encoder model according to the to-be-diagnosed fault data, and training the to-be-diagnosed nested self-encoder model until a model training stop condition is met, and taking parameters of the trained to-be-diagnosed nested self-encoder model as a parameter matrix of the to-be-diagnosed fault data; constructing a hidden feature map of the to-be-diagnosed fault data according to the parameter matrix of the to-be-diagnosed fault data, determining a fault node of the to-be-diagnosed device based on the hidden feature map of the to-be-diagnosed fault data, and realizing rapid and accurate fault diagnosis of the device.

Description

Technical Field

[0001] This invention relates to the field of equipment fault diagnosis technology, and in particular to a method for equipment fault diagnosis based on associated hidden feature maps. Background Technology

[0002] With the rapid advancement of science and technology, the widespread application of various types of equipment in today's society is becoming increasingly common, and the performance requirements for them are also rising. The safe and stable operation of these devices is directly related to the stable development of the national economy. Therefore, when equipment malfunctions, quickly and accurately locating the fault is of paramount importance.

[0003] In existing technologies, equipment fault diagnosis typically utilizes a limited number of variables related to the fault mode and does not consider the interrelationships between components. This results in an incomplete coverage of variables involved in the fault and an insufficient reflection of the interrelationships between these variables when diagnosing complex equipment, leading to less than ideal diagnostic results. Deep learning methods have been increasingly used in fault diagnosis in recent years; however, existing deep learning methods are often black-box models, lacking interpretability in fault diagnosis, and labeled data is often difficult to obtain.

[0004] Therefore, there is an urgent need for a technical solution for equipment fault diagnosis. Summary of the Invention

[0005] Based on the above analysis, the embodiments of the present invention aim to provide a device fault diagnosis method based on associated latent feature maps, in order to solve the problems of low accuracy in device fault diagnosis caused by insufficient label data, numerous monitoring variables, and the existence of correlations between monitoring variables in the prior art.

[0006] This invention provides a device fault diagnosis method based on associated hidden feature maps, the fault diagnosis method comprising:

[0007] Acquire the fault data to be diagnosed of the device to be diagnosed. The fault data to be diagnosed includes a first preset number of operation monitoring data groups to be diagnosed. The time interval between two adjacent operation monitoring data groups to be diagnosed is the same.

[0008] A nested autoencoder model is constructed based on the fault data to be diagnosed, and the constructed nested autoencoder model is trained based on the fault data to be diagnosed until the model training cutoff condition is met. The parameters of the trained nested autoencoder model are used as the parameter matrix of the fault data to be diagnosed.

[0009] A hidden feature map of the fault data to be diagnosed is constructed based on the parameter matrix of the fault data to be diagnosed, and the fault node of the equipment to be diagnosed is determined based on the hidden feature map of the fault data to be diagnosed.

[0010] A further improvement to the above method, the step of determining the fault node of the device to be diagnosed based on the hidden feature map of the fault data to be diagnosed, includes:

[0011] Based on the hidden feature map of the fault data to be diagnosed, the degree of each node is determined sequentially; the node with the smallest degree is taken as the fault node of the device to be diagnosed.

[0012] A further improvement to the above method, the step of determining the fault node of the device to be diagnosed based on the hidden feature map of the fault data to be diagnosed, further includes:

[0013] If there are two or more nodes with the smallest degree, then all nodes with the smallest degree are selected as nodes to be filtered.

[0014] Determine the clustering coefficient of each node to be screened, and select the node with the smallest clustering coefficient as the fault node of the device to be diagnosed.

[0015] A further improvement to the above method, the step of constructing the hidden feature map of the fault data to be diagnosed based on the parameter matrix of the fault data to be diagnosed, includes:

[0016] The nodes of the hidden feature map are determined based on the type of the component to be diagnosed, and there is a one-to-one correspondence between the type of component to be diagnosed and the node.

[0017] The parameter matrix of the fault data to be diagnosed is binarized according to the preset threshold to be diagnosed. The nodes are connected according to the result of the binarization to obtain the hidden feature map of the fault data to be diagnosed.

[0018] A further improvement to the above method, the step of constructing a nested autoencoder model to be diagnosed based on the fault data to be diagnosed, includes:

[0019] The monitoring variables to be diagnosed in the operational monitoring data group are divided according to the type of component to be diagnosed, and the number of component types to be diagnosed is determined.

[0020] A nested autoencoder model for diagnosis is constructed based on the number of component types to be diagnosed. The nested autoencoder model for diagnosis includes a system autoencoder and sub-autoencoders for the number of component types to be diagnosed. Both the sub-autoencoders and the system autoencoder include a feature extraction module, a hidden space, and a feature reduction module.

[0021] The parameters of the trained system autoencoder are used as the parameter matrix of the fault data to be diagnosed.

[0022] A further improvement to the above method, the step of training the constructed nested autoencoder model based on the fault data to be diagnosed, includes:

[0023] Preprocess the fault data to be diagnosed by dividing the first preset number of operational monitoring data groups to be diagnosed into a second preset number of training samples.

[0024] The training set and validation set are randomly selected from the second preset number of training samples, and the training samples in the training set and validation set are used to train the nested autoencoder model to be diagnosed.

[0025] Based on a further improvement to the above method, the preprocessing of the fault data to be diagnosed, dividing the first preset number of operational monitoring data groups to be diagnosed into a second preset number of training samples, includes:

[0026] The first preset number of operational monitoring data sets to be diagnosed are divided according to the operating conditions to be diagnosed, and the mean and standard deviation of each monitoring variable to be diagnosed under each operating condition are determined; each operational monitoring data set to be diagnosed includes data of a third preset number of monitoring variables to be diagnosed.

[0027] Based on the mean and standard deviation of all monitoring variables under all operating conditions to be diagnosed, the first preset number of monitoring data sets to be diagnosed are updated by standard deviation standardization and normalization to obtain the updated first preset number of monitoring data sets to be diagnosed.

[0028] Based on a preset sliding time window, the updated first preset number of operational monitoring data sets to be diagnosed are segmented to obtain a second preset number of training samples.

[0029] Based on a further improvement of the above method, the standard deviation of the first preset number of operational monitoring data sets to be diagnosed is standardized using the following formula:

[0030]

[0031] Where, x n，p This represents the p-th monitoring variable data in the n-th monitoring data group to be diagnosed. s represents the average value of the p-th monitoring variable to be diagnosed, corresponding to the operating condition to which the n-th monitoring data group to be diagnosed belongs. p This represents the standard deviation of the p-th monitoring variable to be diagnosed, corresponding to the operating condition to which the n-th monitoring data group to be diagnosed belongs. This represents the p-th monitoring variable data in the n-th monitoring data group to be diagnosed after standardization of standard deviation, and T represents the first preset quantity.

[0032] Based on further improvements to the above method, the first preset number of operational monitoring data sets to be diagnosed, after standardization of the standard deviation, are normalized using the following formula:

[0033]

[0034] Among them, u n，p This represents the p-th monitoring variable data in the n-th monitoring data group to be diagnosed after the update. This represents the minimum value of the p-th monitoring variable under the operating condition to which the n-th monitoring data group to be diagnosed belongs. The maximum value of the p-th monitoring variable under the operating condition to which the n-th monitoring data group to be diagnosed belongs.

[0035] A further improvement to the above method, the step of training the nested autoencoder model to be diagnosed using training samples from the training and validation sets, includes:

[0036] The diagnostic feature extraction variables of the same type of component to be diagnosed in the training samples are input into the feature extraction module of the sub-autoencoder corresponding to the nested autoencoder model to be diagnosed. The feature extraction module is used to encode the diagnostic feature extraction variables to obtain the feature values ​​under the type of component to be diagnosed and save them to the hidden space. The feature reduction module is used to receive the feature values ​​under the type of component to be diagnosed and decode them to obtain the predicted values ​​under the type of component to be diagnosed. The diagnostic comparison and verification variables of the same type of component to be diagnosed in the training samples are input into the feature reduction module of the sub-autoencoder corresponding to the nested autoencoder model to be diagnosed. The feature reduction module is also used to compare the predicted values ​​under the type of component to be diagnosed and the diagnostic comparison and verification variables, and determine whether to perform backpropagation based on the comparison results.

[0037] In the system autoencoder of the nested autoencoder model to be diagnosed, the feature extraction module of the system autoencoder receives the feature values ​​of all the component types to be diagnosed from all the sub-autoencoders, merges the feature values ​​of all the component types to be diagnosed to obtain the feature values ​​of the training sample, and saves them to the hidden space; the feature reduction module decodes the feature values ​​of the training sample to obtain the predicted value of the training sample.

[0038] In the system autoencoder of the nested autoencoder model to be diagnosed, the feature reduction module of the system autoencoder receives the diagnostic feature extraction variables of all diagnostic component types of the training sample as the true values ​​of the training sample; the feature reduction module receives the predicted values ​​of the training sample in the hidden space and compares them with the true values ​​of the training sample, and determines whether to perform backpropagation based on the comparison results.

[0039] Compared with the prior art, the present invention can achieve at least one of the following beneficial effects:

[0040] 1. Construct a first set of operational monitoring data for the device under diagnosis using multivariate data of the fault data of the device under diagnosis. Then, calculate the corresponding hidden feature map based on the fault data of the device under diagnosis, associate the multivariate data, and fully consider the feature data of the associated variables in the device under diagnosis so that the hidden feature map can fully express the device under diagnosis and quickly and accurately diagnose the fault location of the device under diagnosis.

[0041] 2. By using the nested autoencoder model to be diagnosed, and combining multiple sub-autoencoders and the system autoencoder, the parameter matrix of the fault data to be diagnosed is determined. This ensures that the parameter matrix of the fault data to be diagnosed fully considers the potential correlation constraints of each monitoring variable to be diagnosed in the device, thereby improving the accuracy of fault location diagnosis.

[0042] 3. By preprocessing the fault data to be diagnosed, standardizing and normalizing it, the impact of different operating conditions on the fault diagnosis of the equipment under the diagnosis can be reduced, thereby further improving the accuracy of fault location diagnosis.

[0043] In this invention, the above-described technical solutions can be combined with each other to achieve more preferred combinations. Other features and advantages of this invention will be set forth in the following description, and some advantages may become apparent from the description or be learned by practicing the invention. The objects and other advantages of this invention can be realized and obtained from what is particularly pointed out in the description and drawings. Attached Figure Description

[0044] The accompanying drawings are for illustrative purposes only and are not intended to limit the invention. Throughout the drawings, the same reference numerals denote the same parts.

[0045] Figure 1 This is a flowchart illustrating the device fault diagnosis method based on associated hidden feature maps provided in an embodiment of the present invention.

[0046] Figure 2 This is one of the structural schematic diagrams of the nested autoencoder model to be diagnosed provided in an embodiment of the present invention;

[0047] Figure 3 This is a schematic diagram of a data segmentation structure based on a preset sliding time window provided in an embodiment of the present invention;

[0048] Figure 4 This is a schematic diagram of the structure for constructing a latent feature map based on a parameter matrix, provided in an embodiment of the present invention.

[0049] Figure 5 This is the second schematic diagram of the structure of the nested autoencoder model to be diagnosed provided in an embodiment of the present invention. Detailed Implementation

[0050] Preferred embodiments of the present invention will now be described in detail with reference to the accompanying drawings, which form part of this application and are used together with the embodiments of the present invention to illustrate the principles of the present invention, but are not intended to limit the scope of the present invention.

[0051] A specific embodiment of the present invention discloses a device fault diagnosis method based on associated hidden feature maps, such as... Figure 1 As shown, the fault diagnosis method includes:

[0052] Step S1: Obtain the fault data to be diagnosed of the device to be diagnosed. The fault data to be diagnosed includes a first preset number of operation monitoring data groups to be diagnosed. The time interval between two adjacent operation monitoring data groups to be diagnosed is the same.

[0053] Step S2: Construct a nested autoencoder model to be diagnosed based on the fault data to be diagnosed, and train the constructed nested autoencoder model to be diagnosed based on the fault data to be diagnosed until the model training cutoff condition is met. Use the parameters of the trained nested autoencoder model to be diagnosed as the parameter matrix of the fault data to be diagnosed.

[0054] Step S3: Construct a hidden feature map of the fault data to be diagnosed based on the parameter matrix of the fault data to be diagnosed, and determine the fault node of the equipment to be diagnosed based on the hidden feature map of the fault data to be diagnosed.

[0055] Specifically, during the operation of the equipment, a fixed time interval is designated as a data collection point. At each data collection point, various sensors collect the equipment's operational data, which is then used as the operational monitoring data set for each data collection point. The equipment can be used for engines, space stations, nuclear power plants, satellites, high-speed trains, and so on.

[0056] Specifically, when the device is an engine, as shown in Table 1 below, it can collect engine operation monitoring data through 15 sensors.

[0057] Table 1. Names and physical meanings of 15 sensors

[0058] 1 Wf fuel flow pps 2 Nf physical rotational speed of fan blades Rpm 3 Nc Spindle physical speed Rpm 4 T24 low-pressure compressor outlet temperature °R 5 T30 High pressure compressor outlet temperature °R 6 T48 High-pressure steam turbine outlet temperature °R 7 T50 Low-pressure steam turbine outlet temperature °R 8 P15 Total pressure in bypass pipe Psia 9 P2 Fan inlet total pressure Psia 10 P21 Total pressure at the fan outlet Psia 11 P24 Total pressure at the outlet of the low-pressure compressor Psia 12 PS30 Static pressure at the outlet of the high-pressure compressor Psia 13 P40 Total pressure at burner outlet Psia 14 P50 Total pressure at the outlet of the low-pressure steam turbine Psia 15 P52 Total pressure at the outlet of the high-pressure steam turbine Psia

[0059] Specifically, the engine's operational monitoring data set from time point 1 to 11764 is shown in Table 2 below.

[0060] Table 2 Engine Operation Monitoring Data Set

[0061] 1 1 502.42 600.15 3.8553 10005 2 2 502.33 600.06 3.8523 10013 3 0 502.42 600.21 3.8633 10017 … 11764 2 473.75 604.08 4.0405 24994

[0062] Specifically, as shown in Table 2, the data collection time points are sequentially arranged from time point 1 to time point 11764, totaling 11764 sets of operational monitoring data. These include operating condition types 0, 1, and 2. It can be understood that the preset operating conditions can include three types, with different operating conditions indicating that the engine is operating in different states, such as acceleration, deceleration, or stationary states.

[0063] Specifically, such as Figure 1 As shown, in step S1, when the device malfunctions, the diagnostic fault data of the device to be diagnosed is acquired. The first preset number can be reasonably set according to the time interval between two adjacent acquisition time points. The larger the time interval between two adjacent acquisition time points, the smaller the first preset number; the smaller the time interval between two adjacent acquisition time points, the larger the first preset number. The first preset number can be set to 300-1000.

[0064] For example, referring to Table 2, if the engine fails at the 10,000th data collection time, and the first preset number is 500:

[0065] You can select 500 sets of operational monitoring data from 9500 to 9999 as the 500 sets of operational monitoring data to be diagnosed for the engine. At this time, the time interval between two adjacent sets of operational monitoring data to be diagnosed is 0 collection time points.

[0066] Alternatively, the time interval between two adjacent sets of operational monitoring data to be diagnosed can be selected as one collection time point. That is, 500 sets of operational monitoring data at collection time points of 9999, 9997, 9995...8999 can be selected as 500 sets of operational monitoring data to be diagnosed for the engine.

[0067] Specifically, after obtaining the fault data of the device to be diagnosed, in step S2, a nested autoencoder model to be diagnosed is constructed based on the fault data.

[0068] Preferably, such as Figure 2 As shown, the step of constructing a nested autoencoder model based on the fault data to be diagnosed includes:

[0069] The monitoring variables to be diagnosed in the operational monitoring data group are divided according to the type of component to be diagnosed, and the number of component types to be diagnosed is determined.

[0070] A nested autoencoder model for diagnosis is constructed based on the number of component types to be diagnosed. The nested autoencoder model for diagnosis includes a system autoencoder and sub-autoencoders for the number of component types to be diagnosed. Both the sub-autoencoders and the system autoencoder include a feature extraction module, a hidden space, and a feature reduction module.

[0071] The parameters of the trained system autoencoder are used as the parameter matrix of the fault data to be diagnosed.

[0072] Specifically, as shown in Table 1, each set of operational monitoring data to be diagnosed includes 15 sensors, and each sensor corresponds to one monitoring variable to be diagnosed, meaning each set of monitoring data to be diagnosed includes 15 monitoring variables to be diagnosed. The monitoring variables to be diagnosed are categorized according to the type of component to be diagnosed, as shown in Table 3. The types of components to be diagnosed include combustion chamber, low-pressure compressor, high-pressure compressor, low-pressure turbine, high-pressure turbine, fan (speed), and fan (pressure), totaling 7 types. Therefore, the number of component types to be diagnosed is 7, and the number of sub-autoencoders in the nested autoencoder model to be diagnosed is set to 7.

[0073] Table 3 categorizes the monitoring variables for diagnosis based on the type of the component to be diagnosed.

[0074]

[0075] That is, based on the 15 diagnostic monitoring variables included in each diagnostic operation monitoring data group shown in Table 1, and categorized by the type of diagnostic component, the number of diagnostic component types is 7. For example... Figure 2 As shown, the constructed nested sub-encoder model for diagnosis includes one system autoencoder and seven sub-autoencoders. Both the sub-autoencoders and the system autoencoder include a feature extraction module, a hidden space, and a feature reduction module. After the nested sub-encoder model for diagnosis is trained, the parameters of the system autoencoder are used as the parameter matrix of the fault data to be diagnosed.

[0076] Specifically, such as Figure 1 As shown, in step S2, the constructed nested autoencoder model to be diagnosed is trained based on the fault data to be diagnosed until the model training cutoff condition is met. The model training cutoff condition can be the number of training iterations or the loss reaching a threshold, which is not specifically limited here. The parameters of the trained nested autoencoder model to be diagnosed are used as the parameter matrix of the fault data to be diagnosed.

[0077] Preferably, training the constructed nested autoencoder model for diagnosis based on the fault data to be diagnosed includes:

[0078] Preprocess the fault data to be diagnosed by dividing the first preset number of operational monitoring data groups to be diagnosed into a second preset number of training samples.

[0079] The training set and validation set are randomly selected from the second preset number of training samples, and the training samples in the training set and validation set are used to train the nested autoencoder model to be diagnosed.

[0080] Specifically, the preprocessing of the fault data to be diagnosed, which divides the first preset number of operational monitoring data groups to be diagnosed into a second preset number of training samples, includes:

[0081] The first preset number of operational monitoring data sets to be diagnosed are divided according to the operating conditions to be diagnosed, and the mean and standard deviation of each monitoring variable to be diagnosed under each operating condition are determined; each operational monitoring data set to be diagnosed includes data of the third number of monitoring variables to be diagnosed.

[0082] Based on the mean and standard deviation of all monitoring variables under all operating conditions to be diagnosed, the first preset number of monitoring data sets to be diagnosed are updated by standard deviation standardization and normalization to obtain the updated first preset number of monitoring data sets to be diagnosed.

[0083] Based on a preset sliding time window, the updated first preset number of operational monitoring data sets to be diagnosed are segmented to obtain a second preset number of training samples.

[0084] Specifically, the preset operating conditions can be divided into three categories: acceleration, deceleration, and stationary states. The first preset number of operational monitoring data sets to be diagnosed are divided into three groups according to these operating conditions. For example, if 500 operational monitoring data sets to be diagnosed are grouped according to operating conditions, with 200 sets for acceleration, 150 sets for deceleration, and 150 sets for stationary states, then the mean and standard deviation of each monitored variable in each operating condition group can be calculated. Specifically, the mean and standard deviation of 15 monitored variables in the 200 sets for acceleration, 150 sets for deceleration, and 150 sets for stationary states can be calculated.

[0085] As shown in Table 1, each set of operational monitoring data to be diagnosed includes data for 15 monitoring variables to be diagnosed, meaning that the sensors and monitoring variables to be diagnosed correspond to each other, and the third preset number is the number of sensors.

[0086] After determining the mean and standard deviation of all monitoring variables to be diagnosed under all operating conditions for the fault data to be diagnosed, the standard deviation is standardized and normalized for 500 sets of operational monitoring data to be diagnosed for the fault data to be diagnosed, and the updated first preset number of sets of operational monitoring data to be diagnosed is obtained.

[0087] Preferably, the standard deviation of the first preset number of operational monitoring data sets to be diagnosed is standardized using the following formula:

[0088]

[0089] Where, x n，p This represents the p-th monitoring variable data in the n-th monitoring data group to be diagnosed. s represents the average value of the p-th monitoring variable to be diagnosed, corresponding to the operating condition to which the n-th monitoring data group to be diagnosed belongs. p This represents the standard deviation of the p-th monitoring variable to be diagnosed, corresponding to the operating condition to which the n-th monitoring data group to be diagnosed belongs. This represents the p-th monitoring variable data in the n-th monitoring data group to be diagnosed after standardization of standard deviation, and T represents the first preset quantity.

[0090] Specifically, refer to Table 1, x n，p This represents the p-th monitoring variable data in the n-th monitoring data group to be diagnosed (i.e., the p-th monitoring variable out of 15 monitoring variables to be diagnosed). The p-th monitoring variable data in the n-th monitoring data set to be diagnosed after standardization of standard deviation is shown in Table 1, where p ranges from 1 to 15.

[0091] Preferably, the first preset number of operational monitoring data sets to be diagnosed, after standardization of the standard deviation, are normalized using the following formula:

[0092]

[0093] Among them, u n，p This represents the p-th monitoring variable data in the n-th monitoring data group to be diagnosed after the update. This represents the minimum value of the p-th monitoring variable under the operating condition to which the n-th monitoring data group to be diagnosed belongs. The maximum value of the p-th monitoring variable under the operating condition to which the n-th monitoring data group to be diagnosed belongs.

[0094] Specifically, during normalization, each monitoring variable data in all monitoring data groups of the fault to be diagnosed is mapped to the interval [0,1].

[0095] Specifically, the step of segmenting the updated first preset number of operational monitoring data sets to be diagnosed based on a preset sliding time window to obtain a second preset number of training samples includes:

[0096]

[0097] Where T represents the first preset quantity, ws represents the size of the preset sliding time window, ss represents the step size of the preset sliding time window, and Floor[] represents rounding down;

[0098] For each training sample, the monitoring data set to be diagnosed is divided according to the type of component to be diagnosed, and the feature extraction variables and the comparison and verification variables to be diagnosed are determined in each type of component to be diagnosed.

[0099] like Figure 3 As shown, the horizontal axis represents the fault data to be diagnosed arranged in chronological order, from the start time point to the end time point of the data collection. The vertical axis represents the data from the 15 sensors corresponding to each data collection time point. A preset sliding time window slides from the start time point to the end time point, dividing the first preset number of fault monitoring data groups into a second preset number of training samples.

[0100] It is worth noting that each training sample includes ws sets of monitoring data to be diagnosed, and the number of training samples can be controlled by adjusting the step size of the sliding time window.

[0101] For each training sample, the monitoring data set to be diagnosed is divided according to the type of component to be diagnosed, and the feature extraction variables and the comparison and verification variables to be diagnosed are determined in each type of component to be diagnosed.

[0102] Referring to Table 3, after dividing each set of operational monitoring data to be diagnosed according to the type of component to be diagnosed, it is also necessary to further divide it into the feature extraction variable u to be diagnosed. in Comparison and validation of the variable u to be diagnosed ou That is, within each type of component to be diagnosed, variables are divided into features to be extracted and variables to be compared and validated. For example, in the combustion chamber, the combustion flow rate is used as the feature to be extracted variable u. in The total pressure at the burner outlet is used as the variable to be compared and verified in the diagnosis. ou .

[0103] It is worth noting that the data to be diagnosed, i.e., the first preset number of operational monitoring data sets to be diagnosed, is divided into a second preset number of training samples. Each training sample includes a set of operational monitoring data sets to be diagnosed with a sliding time window. This set of operational monitoring data sets with a sliding time window is divided into two parts: one part is the feature extraction variable to be diagnosed, and the other part is the comparison and verification variable to be diagnosed. As shown in Table 4, the data to be diagnosed includes a total of 500 collection time points. The sliding time window size is 10, and the step size of the sliding time window is 5. This results in {Floor[(500-10) / 5]+1} training samples, or 99 training samples. Thus, the first training sample consists of the first to the tenth collection time points, the second training sample consists of the sixth to the fifteenth collection time points, and so on, resulting in 99 training samples. Each training sample includes the diagnostic data set with 10 collection time points. The operation monitoring data group is divided according to the component types in Table 3. The fuel flow rate, low-pressure compressor outlet temperature, high-pressure compressor outlet temperature, low-pressure turbine outlet temperature, fan blade physical speed, fan inlet total pressure, and high-pressure turbine outlet temperature at the 10 collection points are used as feature extraction variables. The total pressure at the burner outlet, the total pressure at the low-pressure compressor outlet, the static pressure at the high-pressure compressor outlet, the total pressure at the low-pressure turbine outlet, the main shaft physical speed, the total pressure at the fan outlet, and the total pressure in the bypass pipe at the 10 collection points are used as comparison and verification variables.

[0104] Table 4. Training sample division of the fault data to be diagnosed

[0105] 1 1 502.42 600.15 3.8553 10005 2 2 502.33 600.06 3.8523 10013 3 0 502.42 600.21 3.8633 10017 … 500 2 473.75 604.08 4.0405 24994

[0106] Specifically, such as Figure 2 As shown, the nested autoencoder model to be diagnosed includes one system autoencoder and seven sub-autoencoders. Both the sub-autoencoders and the system autoencoder include a feature extraction module, a hidden space, and a feature reduction module. The feature extraction module is used for encoding, the hidden space is used to store the encoded feature values, and then the feature values ​​are transmitted to the feature reduction module for decoding. It is worth noting that the feature extraction module of the system autoencoder obtains the feature values ​​of each sub-autoencoder from the sub-autoencoders.

[0107] Preferably, training the nested autoencoder model to be diagnosed using training samples from the training and validation sets includes:

[0108] The diagnostic feature extraction variables of the same type of component to be diagnosed in the training samples are input into the feature extraction module of the sub-autoencoder corresponding to the nested autoencoder model to be diagnosed. The feature extraction module is used to encode the diagnostic feature extraction variables to obtain the feature values ​​under the type of component to be diagnosed and save them to the hidden space. The feature reduction module is used to receive the feature values ​​under the type of component to be diagnosed and decode them to obtain the predicted values ​​under the type of component to be diagnosed. The diagnostic comparison and verification variables of the same type of component to be diagnosed in the training samples are input into the feature reduction module of the sub-autoencoder corresponding to the nested autoencoder model to be diagnosed. The feature reduction module is also used to compare the predicted values ​​under the type of component to be diagnosed and the diagnostic comparison and verification variables, and determine whether to perform backpropagation based on the comparison results.

[0109] In the system autoencoder of the nested autoencoder model to be diagnosed, the feature extraction module of the system autoencoder receives the feature values ​​of all the component types to be diagnosed from all the sub-autoencoders, merges the feature values ​​of all the component types to be diagnosed to obtain the feature values ​​of the training sample, and saves them to the hidden space; the feature reduction module decodes the feature values ​​of the training sample to obtain the predicted value of the training sample.

[0110] In the system autoencoder of the nested autoencoder model to be diagnosed, the feature reduction module of the system autoencoder receives the diagnostic feature extraction variables of all diagnostic component types of the training sample as the true values ​​of the training sample; the feature reduction module receives the predicted values ​​of the training sample in the hidden space and compares them with the true values ​​of the training sample, and determines whether to perform backpropagation based on the comparison results.

[0111] Specifically, such as Figure 2 and Figure 5 As shown, the diagnostic feature extraction variables of the fault data to be diagnosed are input into the constructed nested autoencoder model. Specifically, the diagnostic feature extraction variables belonging to the same type of component to be diagnosed are input into the feature extraction module of the same sub-autoencoder, and the diagnostic feature verification variables belonging to the same type of component to be diagnosed are input into the feature reduction module of the same sub-autoencoder. At the same time, the diagnostic feature extraction variables of the training sample are input into the feature reduction module of the system autoencoder.

[0112] like Figure 3 As shown, the feature extraction variables of the same type of component to be diagnosed in the training samples are input into the feature extraction module of the sub-autoencoder corresponding to the nested autoencoder model to be diagnosed. The feature extraction module is used to encode the feature extraction variables to be diagnosed, obtain the feature values ​​under the type of component to be diagnosed, and save them to the hidden space. The feature reduction module is used to receive the feature values ​​under the type of component to be diagnosed and decode them to obtain the predicted values ​​under the type of component to be diagnosed.

[0113] The diagnostic comparison and verification variables of the same type of component to be diagnosed in the training samples are input into the feature reduction module of the sub-autoencoder corresponding to the nested autoencoder model to be diagnosed. The feature reduction module is also used to compare the predicted value and the diagnostic comparison and verification variables under the type of component to be diagnosed, and to determine whether to perform backpropagation based on the comparison result.

[0114] In the system autoencoder of the nested autoencoder model to be diagnosed, the feature extraction module of the system autoencoder receives the feature values ​​of all the component types to be diagnosed from all the sub-autoencoders, merges the feature values ​​of all the component types to be diagnosed to obtain the feature values ​​of the training sample, and saves them to the hidden space; the feature reduction module decodes the feature values ​​of the training sample to obtain the predicted value of the training sample; in the system autoencoder of the nested autoencoder model, the feature reduction module of the system autoencoder receives the feature extraction variables of all the component types to be diagnosed from the training sample as the true value of the training sample; the feature reduction module receives the predicted value of the training sample from the hidden space and compares it with the true value of the training sample, and determines whether to perform backpropagation based on the comparison result.

[0115] The nested autoencoder model for diagnosis is trained using 99 training samples of the fault data to be diagnosed until the model training cutoff condition is met. The model training cutoff condition can be the number of training iterations or the loss reaching a threshold, which will not be specifically limited here.

[0116] When the training of the nested autoencoder model to be diagnosed is completed, the parameters in the nested autoencoder model to be diagnosed are used as the parameter matrix of the running cycle. The parameters in the nested autoencoder model to be diagnosed are the parameter matrix of the hidden space of the system autoencoder. This parameter matrix is ​​a 7×7 matrix and is used as the parameter matrix of the fault data to be diagnosed.

[0117] For example, such as Figure 5 As shown, the nested autoencoder model to be diagnosed includes 7 sub-autoencoders and 1 system autoencoder. The 7 diagnostic feature extraction variables from a training sample of the fault data to be diagnosed are input from the left side. The inputs are fed into each sub-autoencoder. Within the feature extraction module of the first sub-autoencoder, the input features to be diagnosed are sequentially encoded using an LSTM (Long Short-Term Memory) network with an attention mechanism and an FCNN (Fully Connected Neural Network), resulting in... The hidden space stores the encoded feature value h1. Specifically, the feature value h1 in the hidden space is passed through an FCNN network to achieve feature reduction to obtain mid1, and then passed through another FCNN network to obtain mid_s1.

[0118] Meanwhile, the seven feature variables of a training sample of the fault data to be diagnosed are input from the right side, i.e. In the special reduction module of the first sub-autoencoder, compare mid_s1 and Perform backpropagation.

[0119] Similarly, the remaining sub-autoencoder processes are the same, each used to handle different monitored variables to be diagnosed.

[0120] Furthermore, within the system's autoencoder, the feature extraction module receives the feature values ​​from each sub-autoencoder and merges them to form midc. This midc is then processed by an FCNN parameter network to obtain mid, and mid is further processed by another FCNN network to obtain u. in‘ Simultaneously, the feature reduction module of the system's autoencoder receives the feature extraction variable u to be diagnosed. in u in for The sum of and is compared within the feature reduction module of the system's autoencoder. in and u in‘ Backpropagation is then performed based on the comparison results.

[0121] It is worth noting that when the model training cutoff condition is met, the parameter matrix of the FCNN parameter network of the system autoencoder is taken out as the parameter matrix of the fault data to be diagnosed.

[0122] Specifically, such as Figure 1 As shown, in step S3, a hidden feature map of the fault data to be diagnosed is constructed using the parameter matrix of the fault data to be diagnosed.

[0123] Preferably, the step of constructing a latent feature map of the fault data to be diagnosed based on the parameter matrix of the fault data to be diagnosed includes:

[0124] The nodes of the hidden feature map are determined based on the type of the component to be diagnosed, and there is a one-to-one correspondence between the type of component to be diagnosed and the node.

[0125] The parameter matrix of the fault data to be diagnosed is binarized according to the preset threshold to be diagnosed. The nodes are connected according to the result of the binarization to obtain the hidden feature map of the fault data to be diagnosed.

[0126] Specifically, a latent feature map of the fault data to be diagnosed is constructed based on the parameter matrix of the fault data to be diagnosed. Specifically, each component type to be diagnosed is used as a node in the latent feature map, such as... Figure 4 As shown in the right-hand hidden feature map, there are a total of 7 nodes: v1, v2, v3, v4, v5, v6, and v7, representing 7 component types to be diagnosed. Figure 4The left image shows the parameter matrix after binarization. The preset threshold can be set reasonably according to the actual situation. The preset threshold is less than 1, and the preset threshold is preferably 0.2.

[0127] like Figure 4 As shown, by connecting the 7 nodes in the right figure according to the parameter matrix in the left figure, the hidden feature map shown in the right figure is obtained. The hidden feature map includes 7 nodes v1, v2, v3, v4, v5, v6 and v7, and also includes 11 connections e1, e2, e3, e4, e5, e6, e7, e8, e9, e10 and e11.

[0128] Specifically, such as Figure 1 As shown, in step S1, the fault node of the device to be diagnosed is determined based on the hidden feature map of the fault data to be diagnosed.

[0129] Preferably, determining the fault node of the device to be diagnosed based on the hidden feature map of the fault data includes:

[0130] Based on the hidden feature map of the fault data to be diagnosed, the degree of each node is determined sequentially; the node with the smallest degree is taken as the fault node of the device to be diagnosed.

[0131] Specifically, such as Figure 4 As shown, the hidden feature graph includes 7 nodes: v1, v2, v3, v4, v5, v6, and v7, and 11 connections: e1, e2, e3, e4, e5, e6, e7, e8, e9, e10, and e11.

[0132] Specifically, the degree of a node represents the number of connections to that node. For example, the degree of node v1 is 3 (e1, e4, and e6), the degree of node v2 is 3 (e1, e2, and e1), the degree of node v3 is 3 (e2, e3, and e9), the degree of node v4 is 3 (e4, e5, and e7), the degree of node v5 is 4 (e3, e5, e6, and e8), the degree of node v6 is 4 (e8, e7, e9, and e10), and the degree of node v7 is 2 (e10 and e11).

[0133] from Figure 4 As can be seen from the graph, node v7 has the smallest degree, so node v7 is the fault node of the device to be diagnosed.

[0134] Preferably, the step of determining the fault node of the device to be diagnosed based on the hidden feature map of the fault data to be diagnosed further includes:

[0135] If there are two or more nodes with the smallest degree, then all nodes with the smallest degree are selected as nodes to be filtered.

[0136] Determine the clustering coefficient of each node to be screened, and select the node with the smallest clustering coefficient as the fault node of the device to be diagnosed.

[0137] It is understandable that the clustering coefficient of each node is the number of nodes connected to that node.

[0138] Specifically, in some cases, there are two or more nodes with the smallest degree. In this case, all nodes with the smallest degree are selected as nodes to be screened, the clustering coefficient of each node to be screened is calculated, and the node with the smallest clustering coefficient is selected as the fault node of the device to be diagnosed.

[0139] As can be understood, as shown in Table 3, by matching the type of the component to be diagnosed to the fault node, the location of the fault in the device to be diagnosed can be determined.

[0140] Compared with existing technologies, the device fault diagnosis method based on associated latent feature maps provided in this invention constructs a first number of operational monitoring data groups to be diagnosed based on the multivariate data of the fault data of the device to be diagnosed. Then, it calculates the corresponding latent feature maps based on the fault data to be diagnosed, associates the multivariates, and fully considers the associated variable feature data in the device to be diagnosed, so that the latent feature maps can fully express the device to be diagnosed and quickly and accurately diagnose the fault location of the device to be diagnosed. At the same time, through the nested autoencoder model to be diagnosed, the parameter matrix of the fault data to be diagnosed is determined by combining multiple sub-autoencoders and the system autoencoder. This ensures that the parameter matrix of the fault data to be diagnosed fully considers the potential association constraints of each monitoring variable to be diagnosed in the device to be diagnosed, thereby improving the accuracy of fault location diagnosis. Furthermore, by preprocessing the fault data to be diagnosed, standardizing and normalizing it, the influence of the device to be diagnosed under different operating conditions on fault diagnosis is reduced, further improving the accuracy of fault location diagnosis.

[0141] Those skilled in the art will understand that all or part of the processes of the methods described in the above embodiments can be implemented by a computer program instructing related hardware, and the program can be stored in a computer-readable storage medium. The computer-readable storage medium may be a disk, optical disk, read-only memory, or random access memory, etc.

[0142] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any changes or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in the present invention should be included within the scope of protection of the present invention.

Claims

1. A method for diagnosing equipment faults based on associated latent feature maps, characterized in that, The fault diagnosis method includes: Acquire the fault data to be diagnosed of the device to be diagnosed. The fault data to be diagnosed includes a first preset number of operation monitoring data groups to be diagnosed. The time interval between two adjacent operation monitoring data groups to be diagnosed is the same. A nested autoencoder model is constructed based on the fault data to be diagnosed, and the constructed nested autoencoder model is trained based on the fault data to be diagnosed until the model training cutoff condition is met. The parameters of the trained nested autoencoder model are used as the parameter matrix of the fault data to be diagnosed. Construct a hidden feature map of the fault data to be diagnosed based on the parameter matrix of the fault data to be diagnosed, and determine the fault node of the equipment to be diagnosed based on the hidden feature map of the fault data to be diagnosed. The set of operational monitoring data to be diagnosed includes fuel flow rate, fan blade physical speed, low-pressure compressor outlet temperature, and fan inlet total pressure. The method of determining the fault node of the device to be diagnosed based on the hidden feature map of the fault data includes: Based on the hidden feature map of the fault data to be diagnosed, the degree of each node is determined in turn; the node with the smallest degree is taken as the fault node of the device to be diagnosed. The step of constructing a latent feature map of the fault data to be diagnosed based on the parameter matrix of the fault data to be diagnosed includes: The nodes of the hidden feature map are determined based on the type of the component to be diagnosed, and there is a one-to-one correspondence between the type of component to be diagnosed and the node. The parameter matrix of the fault data to be diagnosed is binarized according to the preset threshold to be diagnosed. The nodes are connected according to the result of the binarization to obtain the hidden feature map of the fault data to be diagnosed. The step of constructing a nested autoencoder model for diagnosis based on the fault data to be diagnosed includes: The monitoring variables to be diagnosed in the operational monitoring data group are divided according to the type of component to be diagnosed, and the number of component types to be diagnosed is determined. A nested autoencoder model for diagnosis is constructed based on the number of component types to be diagnosed. The nested autoencoder model for diagnosis includes a system autoencoder and sub-autoencoders for the number of component types to be diagnosed. Both the sub-autoencoders and the system autoencoder include a feature extraction module, a hidden space, and a feature reduction module. The parameters of the trained system autoencoder are used as the parameter matrix of the fault data to be diagnosed.

2. The fault diagnosis method according to claim 1, characterized in that, The method of determining the fault node of the device to be diagnosed based on the hidden feature map of the fault data to be diagnosed also includes: If there are two or more nodes with the smallest degree, then all nodes with the smallest degree are selected as nodes to be filtered. Determine the clustering coefficient of each node to be screened, and select the node with the smallest clustering coefficient as the fault node of the device to be diagnosed.

3. The fault diagnosis method according to claim 2, characterized in that, The step of training the constructed nested autoencoder model for diagnosis based on the fault data to be diagnosed includes: Preprocess the fault data to be diagnosed by dividing the first preset number of operational monitoring data groups to be diagnosed into a second preset number of training samples. The training set and validation set are randomly selected from the second preset number of training samples, and the training samples in the training set and validation set are used to train the nested autoencoder model to be diagnosed.

4. The fault diagnosis method according to claim 3, characterized in that, The preprocessing of the fault data to be diagnosed involves dividing a first preset number of operational monitoring data sets to be diagnosed into a second preset number of training samples, including: The first preset number of operational monitoring data sets to be diagnosed are divided according to the operating conditions to be diagnosed, and the mean and standard deviation of each monitoring variable to be diagnosed under each operating condition are determined; each operational monitoring data set to be diagnosed includes data of a third preset number of monitoring variables to be diagnosed. Based on the mean and standard deviation of all monitoring variables under all operating conditions to be diagnosed, the first preset number of monitoring data sets to be diagnosed are updated by standard deviation standardization and normalization to obtain the updated first preset number of monitoring data sets to be diagnosed. Based on a preset sliding time window, the updated first preset number of operational monitoring data sets to be diagnosed are segmented to obtain a second preset number of training samples.

5. The fault diagnosis method according to claim 4, characterized in that, The standard deviation of the first preset number of operational monitoring data sets to be diagnosed is standardized using the following formula: ； in, This represents the p-th monitoring variable data in the n-th monitoring data group to be diagnosed. This represents the average value of the p-th monitoring variable to be diagnosed, corresponding to the operating condition to which the n-th monitoring data group to be diagnosed belongs. This represents the standard deviation of the p-th monitoring variable to be diagnosed, corresponding to the operating condition to which the n-th monitoring data set to be diagnosed belongs. This represents the p-th monitoring variable data in the n-th monitoring data group to be diagnosed after standardization of standard deviation, and T represents the first preset quantity.

6. The fault diagnosis method according to claim 5, characterized in that, The first preset number of operational monitoring data sets to be diagnosed, after standardization of the standard deviation, are normalized using the following formula: ； in, This represents the p-th monitoring variable data in the n-th monitoring data group to be diagnosed after the update. This represents the minimum value of the p-th monitoring variable under the operating condition to which the n-th monitoring data group to be diagnosed belongs. The maximum value of the p-th monitoring variable under the operating condition to which the n-th monitoring data group to be diagnosed belongs.

7. The fault diagnosis method according to claim 6, characterized in that, The process of training the nested autoencoder model to be diagnosed using training samples from the training and validation sets includes: The diagnostic feature extraction variables of the same type of component to be diagnosed in the training samples are input into the feature extraction module of the sub-autoencoder corresponding to the nested autoencoder model to be diagnosed. The feature extraction module is used to encode the diagnostic feature extraction variables to obtain the feature values ​​under the type of component to be diagnosed and save them to the hidden space. The feature reduction module is used to receive the feature values ​​under the type of component to be diagnosed and decode them to obtain the predicted values ​​under the type of component to be diagnosed. The diagnostic comparison and verification variables of the same type of component to be diagnosed in the training samples are input into the feature reduction module of the sub-autoencoder corresponding to the nested autoencoder model to be diagnosed. The feature reduction module is also used to compare the predicted values ​​under the type of component to be diagnosed and the diagnostic comparison and verification variables, and determine whether to perform backpropagation based on the comparison results. In the system autoencoder of the nested autoencoder model to be diagnosed, the feature extraction module of the system autoencoder receives the feature values ​​of all the component types to be diagnosed from all the sub-autoencoders, merges the feature values ​​of all the component types to be diagnosed to obtain the feature values ​​of the training sample, and saves them to the hidden space; the feature reduction module decodes the feature values ​​of the training sample to obtain the predicted value of the training sample. In the system autoencoder of the nested autoencoder model to be diagnosed, the feature reduction module of the system autoencoder receives the diagnostic feature extraction variables of all diagnostic component types of the training sample as the true values ​​of the training sample; the feature reduction module receives the predicted values ​​of the training sample in the hidden space and compares them with the true values ​​of the training sample, and determines whether to perform backpropagation based on the comparison results.

Images