A method for energy system fault diagnosis based on deep learning and Gram angle field images

Abstract

The present invention relates to an energy system fault diagnosis method based on deep learning and Gram angle field images. The method converts the one-dimensional time series of multiple operating parameters into two-dimensional GAF images based on the historical operating data of the energy system under various working conditions, and establishes an image sample set; on this basis, the generative network in deep learning is used to expand the sample of the fault image sample and balance the sample set; the fault diagnosis model based on deep learning is trained offline based on the balanced image sample set, and the GAF images of the actual operating data are classified to realize the online fault diagnosis of the energy system. Compared with the prior art, the present invention has the advantages of ensuring the energy supply reliability and equipment safety of the energy system and having high diagnostic accuracy.

Description

Technical Field

[0001] The present invention relates to the technical field of energy system fault diagnosis, and in particular to an energy system fault diagnosis method based on deep learning and Gram angle field images. Background Art

[0002] Energy is indispensable in people's production and life, and it is particularly important to ensure the reliability and stability of energy system energy supply. However, the structure of the energy system is very large and complex, the various devices are closely related, and there are many uncertain factors in the system. Once a system failure occurs, it is difficult to quickly detect and diagnose the fault through traditional manual detection methods, which will not only cause energy waste and equipment damage, but also directly affect the energy consumption of the user end, causing unnecessary losses. Therefore, an intelligent energy system fault diagnosis technology is needed that can quickly and accurately detect and diagnose system faults and provide a reliable basis for subsequent system recovery control.

[0003] In addition, since the probability of energy system failure is low and the system is in normal operation most of the time, there will be an imbalance in the number of normal samples and faulty samples in the system data set, which will affect the diagnostic accuracy of the final diagnosis model and cause faulty samples to be misclassified as normal samples, resulting in serious consequences. Currently, there is no diagnostic method that can effectively solve this problem. Summary of the invention

[0004] The purpose of the present invention is to overcome the defects of the above-mentioned prior art and to provide an energy system fault diagnosis method based on deep learning and Gram angle field images.

[0005] The purpose of the present invention can be achieved by the following technical solutions:

[0006] A method for energy system fault diagnosis based on deep learning and Gram angle field image, the method comprising the following steps:

[0007] Step 1: Collect historical operation data of the energy system and record the types of faults that occurred and the time period in which they occurred.

[0008] In this step, the historical operating data of the energy system collected should include data under normal operating conditions and system operating data under the fault type conditions that need to be diagnosed, and the type of fault that occurred and the time when the fault occurred must be recorded so that labels can be assigned to the data samples later.

[0009] Step 2: Preprocess the collected system operation data. Specifically, it includes the following steps:

[0010] Step 2.1: Missing value filling. Search for missing values ​​in the collected data and fill them using the linear interpolation method; the linear interpolation method is shown in formula 1:

[0011]

[0012] In the formula, x0 and x1 represent the time points before and after the missing value, respectively; y0 and y1 represent the values ​​corresponding to the time points before and after the missing value, respectively; Represent the time point and fill value of missing value respectively.

[0013] Step 2.2: Eliminate data from the start-up and shutdown phases of the unit. Determine the time of the two non-steady-state operation phases of unit startup and shutdown, and directly eliminate the data from these two phases.

[0014] Step 2.3: Outlier elimination: According to the Raida criterion (3σ criterion), outliers in the data are searched and removed, and then filled in using the linear interpolation method.

[0015] The Raida criterion (3σ criterion) is shown in formula (2):

[0016]

[0017] In the formula, X represents the time series of the operating data of a certain parameter of the system; i represents the value of the parameter at the i-th time point, i = 1, 2, ..., n; n represents the length of the time series; μ is the expectation of X; σ is the standard deviation of X; x j is the outlier in X, located at the jth time point, and the standard deviation is usually calculated using the Bessel formula: represents x1,x2,…,x n The average value of . Remove outliers x j Then, linear interpolation is performed using formula (1).

[0018] Step 2.4: Data normalization. Normalize the data to [0, 1] to avoid the impact caused by different orders of magnitude between data.

[0019] The normalization formula is as follows:

[0020]

[0021] Where X={x1,x2,…,x n} represents the time series of the operating data of a certain parameter of the system; x irepresents the value of the parameter at the i-th time point; n represents the length of the time series; min(X) represents the minimum value in the time series of the parameter operation data; max(X) represents the maximum value in the time series of the parameter operation data.

[0022] Step 3: Use the maximum relevant minimum redundancy (mRMR) technique to screen the system operating parameters and select ten operating parameters as input features.

[0023] The feature selection using the maximum relevance minimum redundancy (mRMR) technique specifically includes the following steps:

[0024] Step 3.1: Use mutual information to measure the correlation between system parameters and system faults, and arrange the parameters in descending order; specifically:

[0025] First, determine the correlation between the feature and the fault label. The nonlinear correlation between the feature and the fault can be measured by mutual information, as shown in formula (4):

[0026]

[0027] In the formula, f i Represents a feature vector in the data set; f i j represents the jth element in the feature vector; c and c i Represents the label vector and the i-th element in the label vector respectively; I(f i ,c) represents the mutual information measure between a feature and the fault label.

[0028] Step 3.2: Determine the initial feature subset, select new features to add to the subset, calculate the correlation between the new feature subset and the fault and the subset redundancy, and re-sort the features according to the ratio of correlation to redundancy;

[0029] Among them, regarding the measurement of the correlation between the feature subset and the fault label: the feature subset contains multiple feature vectors, and the correlation between the feature subset and the fault label is quantitatively scored, as shown in formula (5):

[0030]

[0031] In the formula, S represents a feature subset in the data set; |S| represents the number of features contained in the feature subset; D(S,c) represents the correlation measure between the feature subset and the fault label.

[0032] Redundancy measure for feature subsets: If two features f in the feature subset i With f jIf the mutual information value between them is high, it means that there is redundancy between the two features. The redundancy score of feature subset S is shown in formula (6):

[0033]

[0034] Furthermore, for the feature selection index: all features are sorted in descending order according to the mutual information value between the feature and the fault label. The initial feature subset is determined, and new features are selected to be added to the feature subset. The selection criteria for the new features are shown in formula (7):

[0035]

[0036] Step 3.3: Select the top ten best features as model input features.

[0037] After calculating mRMR for all features, a new feature ranking is obtained. The top-ranked features are excellent features that are highly correlated with the fault label and have low redundancy between features. The top ten features are selected as input features for subsequent work.

[0038] Step 4: Set the time window and sliding step size of appropriate size, segment the time series of the ten input features, and form multiple data samples, each of which corresponds to a system operating condition.

[0039] Step 5: Connect the time series of each feature in the data sample end to end to form a one-dimensional sequence, and use the Gramian Angular Field conversion matrix to convert it into a two-dimensional GAF (Gramian Angular Field) image.

[0040] Step 6: According to the operating conditions corresponding to each data sample, assign labels to the converted GAF ​​images and build an image sample set.

[0041] Furthermore, the process of establishing the GAF image sample set is as follows:

[0042] First, set the appropriate time window size and sliding step size to segment the one-dimensional time series of input features to form multiple data samples, each of which corresponds to a system operating condition;

[0043] Secondly, the time series of each input feature in the data sample are spliced ​​into a one-dimensional sequence in the form of end-to-end connection in, Represents the data at the jth time point of the ith input feature; i = 1, 2, ..., 10, a total of ten input features; j = 1, 2, ..., t; t represents the set time window size.

[0044] Then, the one-dimensional sequence Xseq Perform polar coordinate conversion, convert the timestamp in the sequence into the radius value in the polar coordinate system, and convert the numerical value of the sequence into the angle value in the polar coordinate system, as shown in formula (8):

[0045]

[0046] In the formula, Represents X seq The i-th data in the value range is [0,1], and its corresponding timestamp is t i ; N represents the total number of timestamps; θ i Represents X seq The polar angle of the polar coordinate signal corresponding to the i-th data; r represents the polar radius of the polar coordinate signal corresponding to the i-th data.

[0047] According to the Gram-like angular field conversion matrix, the one-dimensional sequence mapped to the polar coordinate system is converted into a two-dimensional GAF image. The Gram-like angular field conversion matrix is ​​shown in formula (9):

[0048]

[0049] Where Gram matrix is ​​the transformation matrix, and N is the number of signals in the one-dimensional sequence.

[0050] According to the system operating conditions corresponding to each data sample, the GAF image is labeled accordingly. To facilitate subsequent supervised learning, the operating condition labels should be converted into digital labels. Finally, a GAF image sample set is formed.

[0051] Step 7: Build a CWGAN-GP generative network and train the network using all fault samples in the image sample set.

[0052] Step 8: Input a certain amount of “noise” and fault label conditions that obey the normal distribution into the trained CWGAN-GP to generate corresponding fault samples, expand the fault samples, and balance the sample set.

[0053] In order to solve the sample imbalance problem in the model training process, the present invention will expand the fault samples to balance the sample set, which includes the following steps:

[0054] First, build a CWGAN-GP generative network. The network structure includes a generative network and a discriminative network. The generative network is used to generate images, and the discriminative network is used to judge the true or false images. Secondly, the "noise" and label conditions that obey the normal distribution are input to the generator to generate a certain number of samples, and two labels are attached to the samples, one is the pseudo sample label and the other is the fault label. Furthermore, the true sample label is attached to the fault sample in the image sample set, and combined with the above-mentioned generated samples to form a training sample for the discriminative network. In the present invention, "noise" is a random vector that obeys the normal distribution.

[0055] During training, the discriminant network continuously improves its ability to distinguish the authenticity of training samples and the category to which the samples belong, and feeds the discrimination results back to the generative network. The generative network is trained based on the results fed back by the discriminant network, which can improve the quality of the generated images and make the images closer to the real images of the category to which they belong. The discriminant network and the generative network reach Nash equilibrium after continuous game learning. Finally, a certain amount of "noise" and label conditions can be input into the generative network in the trained CWGAN-GP to generate fault samples corresponding to the label conditions and achieve sample set balance.

[0056] As a further preferred method, in order to make the GAN model training more stable and avoid problems such as gradient disappearance, the Wasserstein distance is used to characterize the gap between the distribution of real fault samples and the distribution of generated fault samples, and is used as the loss function of the generation network and the discriminant network. In addition, a gradient penalty term (GP) is added to the loss function of the discriminant network, and finally the WGAN-GP model is formed. The gradient penalty term is shown in formula (10):

[0057]

[0058] In the formula, x r ,x g Represent samples of the true distribution and the generated distribution respectively; They represent interpolation distribution samples and interpolation distribution respectively, ∈ is a random number in the range [0,1]; gp is the gradient penalty term, λ is the weight of the gradient penalty term, usually 10; It represents the evaluation of the authenticity of the interpolation distribution samples by the discriminant network.

[0059] On the basis of WGAN-GP, image labels are added for training, so that the generative network generates corresponding fault samples according to the input label values, and finally forms the CWGAN-GP model for fault sample generation. The loss function of the generative network and the discriminative network in CWGAN-GP is shown in formula (11):

[0060]

[0061] In the formula, y represents the additional condition value, that is, the image label; L(D) and L(G) are the loss functions of the discriminant network and the generative network respectively; x r ,x g Represent the samples of the real distribution and the generated distribution respectively; p r ,p g Represent the real sample distribution and the generated sample distribution respectively.

[0062] Step 9: Build a 2D-CNN image classification network and train the network using the balanced image sample set; perform fault diagnosis by classifying GAF images.

[0063] The establishment of the fault diagnosis model in this step includes the following steps:

[0064] Step 9.1: Build a two-dimensional convolutional network (2D-CNN) image classification model. The network includes input layer, convolution layer, pooling layer, fully connected layer and output layer. The image features in the two-dimensional image are extracted through the convolution layer and pooling layer, the nonlinear relationship between the image features and the fault type is fitted through the fully connected layer, and the probability of the image belonging to each category is output through the output layer. The one with the highest probability is the diagnosis result.

[0065] Step 9.2: To avoid overfitting of the model, add batch normalization layer and Dropout layer to the network.

[0066] Step 9.3: Divide the balanced image sample set into a training set and a test set in a ratio of 8:2. Input the training set into the diagnostic model in batches for training to improve the diagnostic accuracy of the model. Repeat multiple trainings until the model converges.

[0067] Step 9.4: Use the test set to test the diagnostic model and comprehensively evaluate the diagnostic model through multiple accuracy evaluation indicators, such as confusion matrix, accuracy, recall rate and F1 score.

[0068] Step 9.5: Input the GAF image converted from the actual operating data into the trained fault diagnosis model, and the operating condition to which the image belongs can be quickly output to achieve online fault diagnosis.

[0069] The energy system fault diagnosis method based on deep learning and Gram angle field image provided by the present invention has at least the following beneficial effects compared with the prior art:

[0070] 1) The present invention first uses the maximum relevant minimum redundancy (mRMR) technology to screen the operating parameters of the energy system and selects ten parameters as input features of the diagnostic model; the feature screening work can not only reduce the number of input features but also ensure that the model has a high diagnostic accuracy; and because the number of input features is reduced, the model training speed will be accelerated and the computing power requirements of the computer will also be reduced. According to the feature screening results, some sensors in the actual energy system can be appropriately reduced to reduce the initial investment cost of the system.

[0071] 2) The present invention uses the Gram angular field matrix to convert the one-dimensional time series of multiple input features into a two-dimensional GAF image. Since the polar coordinate system can realize the unique mapping relationship between the time series and the polar coordinates, the complete original data information is retained, and the time characteristics in the data are also retained; in addition, by splicing the time series of the input features into a one-dimensional sequence and then performing GAF image conversion, the coupling relationship between the feature data can be condensed, and the coupling features can be extracted through subsequent 2D-CNN.

[0072] 3) The present invention uses the CWGAN-GP generative network to expand the fault samples and generate a large number of fault images, thereby solving the problem of sample imbalance; using the balanced fault samples to train the diagnosis model can improve the diagnostic accuracy of the model and reduce the occurrence of "false positive" diagnosis results; compared with ordinary generative networks, CWGAN-GP is more stable in training, the generated samples are of higher quality, and the corresponding fault samples can be generated according to the input label conditions.

[0073] 4) The present invention uses CNN as a diagnostic model to classify image samples and realize energy system fault detection and diagnosis. CNN has the advantages of local perception, sparse connection and weight sharing. It can well extract deep features in the image and fit the nonlinear relationship between features and faults. It has high diagnostic accuracy and does not require the addition of expert knowledge, and can achieve high intelligence. BRIEF DESCRIPTION OF THE DRAWINGS

[0074] Figure 1 A schematic diagram of a specific implementation process of an energy system fault diagnosis method based on deep learning and Gram angle field images provided in an embodiment of the present invention;

[0075] Figure 2 A schematic diagram of a specific implementation flow of step S1 data preprocessing provided in an embodiment of the present invention;

[0076] Figure 3 A specific implementation process and principle diagram of the feature selection work in step S2 provided in an embodiment of the present invention;

[0077] Figure 4A schematic diagram of the specific implementation process and principle of converting the multi-feature time series into a two-dimensional GAF image in step S3 provided in an embodiment of the present invention;

[0078] Figure 5 A schematic diagram of the CWGAN-GP model training principle provided in an embodiment of the present invention;

[0079] Figure 6 A schematic diagram of generating a network structure and parameter settings in the CWGAN-GP model provided in an embodiment of the present invention;

[0080] Figure 7 A schematic diagram of discriminating network structure and parameter settings in the CWGAN-GP model provided in an embodiment of the present invention;

[0081] Figure 8 A schematic diagram of the principle of a 2D-CNN provided in an embodiment of the present invention;

[0082] Fig. 9 A schematic diagram of the 2D-CNN diagnostic model structure and parameter settings provided by an embodiment of the present invention;

[0083] Fig.10 A confusion matrix of a diagnostic model test set obtained by training the unbalanced sample set provided by an embodiment of the present invention;

[0084] Fig.11 A confusion matrix of a diagnostic model test set obtained by training the balanced sample set provided by an embodiment of the present invention;

[0085] Fig.12 A schematic diagram of the diagnostic accuracy and overall diagnostic accuracy of the energy system fault diagnosis method based on deep learning and Gram angular field images provided in an embodiment of the present invention for various types of system conditions of centrifugal chillers. DETAILED DESCRIPTION

[0086] The present invention is described in detail below in conjunction with the accompanying drawings and specific embodiments. Obviously, the described embodiments are part of the embodiments of the present invention, rather than all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative work should belong to the scope of protection of the present invention.

[0087] Example

[0088] The present invention relates to an energy system fault diagnosis method based on deep learning and Gram angle field images, aiming to solve the problems of slow response speed, low diagnostic accuracy, insensitivity to early faults, excessive reliance on expert experience, etc. in the prior art, and realize rapid and accurate detection and diagnosis of energy system faults.

[0089] With the continuous maturity of artificial intelligence technology, fault diagnosis technology with machine learning as the core has been widely used. Fault diagnosis technology based on machine learning relies on algorithm models to extract potential features in system data, and through multiple training of the model, it can obtain the nonlinear relationship between data features and system faults. Although fault diagnosis technology based on machine learning has achieved good results, machine learning algorithms are usually shallow in structure and have fewer adjustable parameters, which limits the model's ability to extract deep features of historical operating data of energy systems, and further limits the model's ability to learn non-stationary relationships.

[0090] As computer computing power continues to increase, deep learning methods were proposed in the early 20th century. Deep learning methods are based on machine learning methods, and further deepen the algorithm structure. They can extract deep features of data, and improve the structure to achieve different functions through different improvements. Deep learning methods have been widely used in various fields, such as image recognition, speech processing, and text processing, and can achieve good results. In the field of fault diagnosis, deep learning methods can extract deep features from a large amount of operating data, learn the complex nonlinear relationship between data and faults, and then accurately detect and diagnose faults without the addition of expert knowledge, achieving high intelligence. As a classic deep learning algorithm, convolutional neural network (CNN) has the advantages of local perception, weight sharing, and sparse connection. It is very suitable for processing image data, so it has attracted widespread attention in the field of image recognition. There is a coupling relationship between the various parameters of the energy system. The occurrence of a fault will cause multiple parameters to change to varying degrees. In order to achieve higher fault diagnosis accuracy, the coupling relationship between input features should be fully considered. For this reason, the one-dimensional time series of multiple input features are converted into a two-dimensional GAF (Gramian Angular Field) image, which not only retains the temporal characteristics of the data, but also extracts the coupling characteristics between the data. The data samples are converted into image samples, and then CNN is used to identify the image to achieve fault diagnosis.

[0091] Based on the above content, the design idea of ​​the present invention is to use a generative network to expand the fault samples, balance the sample set, and then train the fault diagnosis model. The present invention builds a CWGAN-GP generative network. The network structure includes a generative network and a discriminative network. The image is generated by the generative network, and the discriminative network is used to judge the true or false of the image. Secondly, "noise" ("noise" refers to a random vector that obeys a normal distribution) and label conditions are input into the generator to generate a certain number of samples, and two labels are attached to the samples, one is a pseudo-sample label, and the other is a fault label. Furthermore, a true sample label is attached to the fault sample in the image sample set, and combined with the above-mentioned generated samples into a training sample for the discriminative network.

[0092] During training, the discriminant network continuously improves its ability to distinguish the authenticity of training samples and the category to which the samples belong, and feeds the discrimination results back to the generative network. The generative network is trained based on the results fed back by the discriminant network, which can improve the quality of the generated images and make the images closer to the real images of the category to which they belong. The discriminant network and the generative network reach Nash equilibrium after continuous game learning. Finally, a certain amount of "noise" and label conditions can be input into the generative network in the trained CWGAN-GP to generate fault samples corresponding to the label conditions and achieve sample set balance.

[0093] The specific steps are as follows:

[0094] S1: Collect historical operating data of the energy system under various working conditions and pre-process the data.

[0095] S2: The maximum relevance minimum redundancy (mRMR) technique is used for feature selection to select ten system parameters as input features.

[0096] S3: Convert the one-dimensional time series of input features into a two-dimensional GAF image, and attach corresponding labels to the images according to the system operating conditions of the time series to establish a GAF image sample set.

[0097] S4: Build a generative network in deep learning and use the fault samples in the image sample set established in step S3 to perform model training.

[0098] S5: Input "noise" and additional conditions into the generative network to generate a certain number of fault samples of various types to achieve sample expansion and balance the sample set.

[0099] S6: Build a fault diagnosis model based on deep learning, and use the balanced image sample set in step S5 to train the model, so as to accurately obtain the fault type to which the image belongs.

[0100] S7: Perform a comprehensive accuracy evaluation on the fault diagnosis model of step S6.

[0101] The experimental data used in the following examples are the centrifugal chiller fault experimental data (ASHRAE-1043-RP) open sourced by the American Society of Heating, Refrigerating and Air-Conditioning Engineers (ASHRAE). The energy system fault diagnosis method based on deep learning and Gram angular field images proposed in the present invention is verified based on the experimental data in the examples. The specific implementation process is as follows: Figure 1 shown.

[0102] The data used in the embodiment is a 90-ton centrifugal chiller fault test data, which includes operating data of seven fault types and operating data under normal conditions. The seven fault types are: condenser scaling, excessive compressor lubricating oil, non-condensable gas mixed in the refrigerant, reduced condenser water flow, reduced evaporator water flow, refrigerant leakage and refrigerant overcharge. Each fault includes four different fault levels. The time granularity for recording the unit operation data is 10 seconds. Since the open source data set is directly used in the embodiment, the data set has made detailed records and arrangements of the operating data of the unit under various working conditions, so there is no need to perform data collection steps.

[0103] However, in the process of collecting centrifugal chiller operation data, there may be problems such as sensor failure or signal interruption, so there will be abnormal values ​​and missing values ​​in the data set, and the operation data of the chiller start-up and shutdown stages is in a non-steady state and has no reference value. Therefore, it is necessary to preprocess the data set used. The specific implementation steps are as follows: Figure 2 shown.

[0104] Firstly, the missing values ​​in the data are filled by linear interpolation. Secondly, the non-steady-state operation data of the chiller during the start-up and shutdown stages are eliminated. Furthermore, the abnormal values ​​in the parameters of the chiller operation data are detected according to the Raida criterion (3σ criterion), and then the abnormal values ​​are filled by linear interpolation after being eliminated. Finally, in order to avoid the adverse effects caused by the order of magnitude differences between the parameter data, the data are normalized according to the maximum and minimum values ​​of each parameter, and the data are unified into the range of 0 to 1.

[0105] The centrifugal chiller fault experimental data set collected a total of 64 operating parameters, including 43 measurement parameters and 21 calculation parameters. In order to improve the training efficiency and diagnostic performance of the fault diagnosis model, it is necessary to perform feature selection on the operating parameters. The maximum relevance minimum redundancy (mRMR) technology is used to select the ten parameters with the highest correlation with the system operating conditions from the 64 operating parameters as input features. The flow chart and schematic diagram of the feature selection work are shown in the figure. Figure 3 shown.

[0106] First, the mutual information coefficient between each parameter and the fault label is calculated to characterize the correlation between the parameter and the fault label, and the parameters are arranged in descending order according to the mutual information coefficient.

[0107] Secondly, an initial feature subset is determined, new features are selected to be added to the subset, the correlation between the new feature subset and the fault and the subset redundancy are calculated, and the features are re-sorted by the ratio of correlation to redundancy.

[0108] Finally, the top ten best features are selected as the input features of the fault diagnosis model, which include the following features: condenser water inlet temperature, evaporator water outlet temperature, compressor efficiency, condenser water outlet temperature, oil supply pressure, evaporator water inlet temperature, refrigerant pressure in the condenser, evaporator cooling rate, condenser progressive temperature and subcooling temperature of the refrigerant in the condenser.

[0109] The time series of the ten input features selected in this embodiment are converted into two-dimensional image samples according to the GAF image conversion method of the present invention. The specific process of GAF image conversion and the schematic diagram of each step are as shown in FIG. Figure 4 shown.

[0110] First, set the time window to 2 minutes and the sliding step to 1 minute, split the ten input feature time series, and obtain multiple time series samples. Each sample includes ten time series with input feature lengths of 2 minutes. Secondly, concatenate the ten feature time series in each sample to form a one-dimensional sequence, and transform the time domain of the sequence. Furthermore, perform polar coordinate conversion on the signal values ​​in the one-dimensional sequence. The conversion formula is as follows: the sequence signal value corresponds to the polar angle in the polar coordinates, and the timestamp corresponds to the radius in the polar coordinates.

[0111]

[0112] In the formula, Represents X seq The i-th data in the value range is [0,1], and its corresponding timestamp is t i ; N represents the total number of timestamps; θ i Represents X seq The polar angle of the polar coordinate signal corresponding to the i-th data; r represents the polar radius of the polar coordinate signal corresponding to the i-th data.

[0113] Finally, the polarized one-dimensional sequence is substituted into the Gram-like angular field matrix to obtain a two-dimensional GAF image.

[0114] Each time series corresponds to an operating condition of the chiller, so each GAF image also has a corresponding operating condition. A digital label is attached to each image to represent the different operating conditions of the unit. Finally, 6928 normal operating condition image samples and 433 images of the other seven types of fault images are obtained, with a total of 9959 image samples.

[0115] In this embodiment, the ratio of the number of normal samples to the number of various fault samples is 16:1, and there is a sample imbalance problem. The CWGAN-GP in the present invention is used to generate fault samples and balance the sample set. The schematic diagram of the CWGAN-GP model training principle is as follows: Figure 5As shown. By adding the fault label condition, the generation of samples of the specified fault type is realized. Through the game between the generation network and the discriminant network, the Nash equilibrium is finally achieved to optimize the performance of the diagnostic network. After training the CWGAN-GP model with all the fault samples in the sample set, the corresponding fault samples can be output by inputting the "noise" vector and fault label that obey the normal distribution into the generation network. Finally, the CWGAN-GP model is used to generate 6495 image samples of each type of fault, so that the number of each type of fault sample is consistent with the number of normal samples, both of which are 6928 images, and the total number of image sample sets is 55424 images.

[0116] The structure and hyperparameter settings of the generated network in the CWGAN-GP model, such as Figure 6 As shown in Figure 2, the structure and hyperparameter settings of the discriminant network are as follows: Figure 7 The generator network consists of three deconvolution layers and three convolution layers. The deconvolution layer is used to expand the image with a smaller size, while the convolution layer is used to extract the effective features in the image. The discriminator network consists of four convolution layers, which are used to extract the potential features of the input image to judge the image. The parameter settings of each layer of the network are shown in Figure 6 and Figure 7 There are instructions in.

[0117] The balanced GAF ​​image sample set is divided into a training set and a test set at a ratio of nearly 8:2, where the proportion of each type of image is consistent with the total sample set. Therefore, the training set contains 50,424 images, and each type of image contains 6,303 images; the test set contains 5,000 images, and each type of image contains 625 images.

[0118] Build 2D-CNN to classify GAF images. The schematic diagram of 2D-CNN principle is as follows: Figure 8 The network model mainly includes an input layer, a convolutional layer, a maximum pooling layer, a fully connected layer, and an output layer. In this embodiment, the specific structure and hyperparameter settings of the constructed CNN are as follows: Fig. 9 As shown in the figure, the network consists of an input layer, three convolutional layers, three maximum pooling layers, three fully connected layers and an output layer. The specific parameters of each layer are set in Fig. 9 The convolution layer is used to extract image features, the pooling layer is used to reduce the parameter scale, and the fully connected layer is used to fit the complex nonlinear relationship between features and faults. Finally, the output layer uses the Softmax function to obtain the probability that the input image belongs to each category, and the category with the highest probability is the judgment result of the network.

[0119] The constructed 2D-CNN is trained using the GAF image training set. After multiple iterations, the network converges. The model performance is tested using the test set. The confusion matrix of the unbalanced sample set and the balanced sample set is as follows: Fig.10 , Fig.11 As shown, Fig.10 is the confusion matrix of the unbalanced sample set, Fig.11 is the confusion matrix of the balanced sample set (where label "0" represents the normal operation of the centrifugal chiller, label "1" represents the condenser scaling fault, label "2" represents the compressor lubricating oil excess fault, label "3" represents the refrigerant contains non-condensable gas fault, label "4" represents the condenser water flow reduction fault, label "5" represents the evaporator water flow reduction fault, label "6" represents the refrigerant leakage fault, label "7" represents the refrigerant overcharge fault), the diagonal of the confusion matrix represents the number of system operating conditions represented by the sample that the fault diagnosis model can correctly predict. It can be seen from the confusion matrix that the diagnostic model trained by the unbalanced sample set has achieved a high accuracy, but there will be "false positive" diagnostic results, while the accuracy of the model trained by the balanced sample set will be improved and the occurrence of "false positive" results will be reduced. Fig.12 Shown in the figure are the diagnostic accuracy of the fault diagnosis model for each type of working condition and the overall diagnostic accuracy. It can be seen that the overall diagnostic accuracy of the fault diagnosis model trained by the unbalanced sample set is 89.80%, and the overall diagnostic accuracy of the fault diagnosis model trained by the balanced sample set is 96.22%. The accuracy has been improved by 6.42%, reaching a relatively high level. In addition, the diagnostic accuracy of the diagnostic model after balancing the samples is also improved for each category of faults, and the diagnostic accuracy for each type of fault is above 95%. The above results show that the energy system fault diagnosis method based on deep learning and Gram angle field images provided by the present invention has superior effects.

[0120] The present invention converts the input feature time series of the energy system into a two-dimensional GAF image, which not only retains the time series characteristics of the original data, but also can extract the coupling characteristics between the features. On this basis, in view of the problem of sample imbalance, the generative network in deep learning, the CWGAN-GP model, is used to expand the fault image samples and balance the constructed image sample set. According to the balanced sample set, the proposed 2D-CNN image classification model is trained, and the key features of the GAF image are extracted by using the powerful feature extraction ability of 2D-CNN for the image, so as to realize image classification, thereby realizing the fault detection and diagnosis of the energy system. The above embodiments prove that the energy system fault diagnosis method based on deep learning and Gram angular field images proposed by the present invention can accurately diagnose various faults of centrifugal chillers, and its diagnostic accuracy is high. By comparing the diagnostic accuracy of the model trained by the balanced and unbalanced sample sets, the present invention improves the diagnostic accuracy of the fault diagnosis model to a certain extent by solving the problem of sample imbalance. Therefore, the energy system fault diagnosis method based on deep learning and Gram angular field images proposed by the present invention has extremely high practical value for fault detection and diagnosis of energy systems, and provides a feasible solution for intelligent fault diagnosis.

[0121] The above is only a specific embodiment of the present invention, but the protection scope of the present invention is not limited thereto. Any person familiar with the technical field can easily think of various equivalent modifications or substitutions within the technical scope disclosed by the present invention, and these modifications or substitutions should be included in the protection scope of the present invention. Therefore, the protection scope of the present invention shall be based on the protection scope of the claims.

Claims

1. A method for energy system fault diagnosis based on deep learning and Gram angle field images, characterized in that: The following steps are involved: 1) Collect historical operating data of the energy system under various working conditions and pre-process the historical operating data; 2) Use the maximum correlation minimum redundancy technique to select features and select ten system parameters as input features; 3) Convert the one-dimensional time series of input features into a two-dimensional GAF image, and attach corresponding labels to the two-dimensional GAF image according to the system operating conditions of the time series to establish a GAF image sample set; 4) Build a generative network in deep learning and use the fault samples in the GAF image sample set established in step 3) to train the model; 5) Input noise and additional conditions into the generative network to generate various types of fault samples, achieve sample expansion, and balance the GAF graph sample set; 6) Build a fault diagnosis model based on deep learning, use the GAF image sample set balanced in step 5) to train the model, and obtain the fault type to which the image belongs; 7) Input the GAF image converted from the actual operating data into the trained fault diagnosis model, output the operating condition to which the image belongs, and realize online fault diagnosis; The specific steps of step 2) include: 21) Use mutual information metric to obtain nonlinear correlation between features and faults; 22) Obtaining a measure of the correlation between the feature subset and the fault label; 23) Determine the redundancy measure of the feature subset; 24) According to the mutual information value between the feature and the fault label, all features are sorted in descending order, the initial feature subset is determined, and new features are selected to be added to the feature subset; 25) After calculating mRMR for all features, a new feature ranking is obtained, and the top ten features are selected as input features for subsequent work; In step 3), the specific steps of establishing the GAF image sample set include: 31) Set the time window size and window sliding step size, intercept multiple time series of input features from the two-dimensional GAF image with corresponding labels, and obtain multiple data samples; 32) For a data sample, concatenate the time series of each input feature end to end into a one-dimensional sequence; 33) performing polar coordinate transformation on the one-dimensional sequence constructed in step 32); 34) The one-dimensional sequence mapped to the polar coordinate system is converted into a two-dimensional GAF image through a Gram-like angular field matrix; In step 4), the generative network is a generative adversarial network in deep learning. The generative adversarial network model in deep learning includes a generative network for generating images and a discriminative network for judging whether an image is true or false. In the training process of the generative adversarial network model in deep learning, the generative network and the discriminative network are made to learn from each other in a game to achieve Nash equilibrium; In the training process of the generative adversarial network model in deep learning, the Wasserstein distance is used to characterize the gap between the distribution of real fault samples and the distribution of generated fault samples, and is used as the loss function of the generative network and the discriminative network. A gradient penalty term is added to the loss function of the discriminative network to form the WGAN-GP model. In the training process of the generative adversarial network model in deep learning, image labels are added to the WGAN-GP model for training, so that the generative network generates corresponding fault samples according to the input label values, and finally forms a CWGAN-GP model for fault sample generation; the loss function of the generative network and the discriminative network in the CWGAN-GP model is: ， In the formula, is the additional conditional value, i.e., the image label; They are the loss functions of the discriminative network and the generative network respectively; are samples of the true distribution and the generated distribution respectively; are the real sample distribution and the generated sample distribution respectively.

2. The energy system fault diagnosis method based on deep learning and Gram angle field image according to claim 1 is characterized in that: In step 1), the specific steps of preprocessing the historical operation data include: 11) Retrieve the location of missing values ​​in historical operation data and fill in the missing values ​​using linear interpolation method; 12) Eliminate the data during the start-up and shutdown phases of the unit; 13) Eliminate outliers in system operation data according to the Raida criterion and use linear interpolation method to fill in data; 14) Normalize the data processed in step 13).

3. The energy system fault diagnosis method based on deep learning and Gram angle field image according to claim 1 is characterized in that: The specific content of achieving Nash equilibrium is: In the training process of the generative adversarial network model in deep learning, a small number of real fault sample images are used to train CWGAN-GP multiple times until the model converges; noise and fault image label conditions that obey the normal distribution are input to the generative network, and a corresponding number of fault sample images corresponding to the input labels are output, thereby expanding the fault samples so that the sample set is balanced.

4. The energy system fault diagnosis method based on deep learning and Gram angle field image according to claim 1 is characterized in that: In step 6), the fault diagnosis model based on deep learning adopts the constructed two-dimensional convolutional neural network model, and the constructed two-dimensional convolutional neural network model includes an input layer, three convolutional layers, three maximum pooling layers, three fully connected layers, an output layer, three batch normalization layers and three Dropout layers.

5. The energy system fault diagnosis method based on deep learning and Gram angle field image according to claim 4 is characterized in that: The training process of a 2D convolutional neural network consists of the following steps: 61) Randomly extract training samples and test samples in equal proportions from the balanced GAF ​​image sample set to train and test the diagnostic model; 62) Input the image sample training set into the built two-dimensional convolutional neural network model for model training; 63) According to the principle of gradient descent, the Adam optimizer is used to optimize the variable parameters in the model, and the model converges through multiple iterative training; 64) The test sample set is input into the diagnosis model. The output layer of the two-dimensional convolutional neural network model outputs the probability of the sample belonging to each category through the Softmax function. The category with the largest probability is the diagnosis result of the model. 65) Based on the test results, conduct a comprehensive accuracy evaluation of the diagnostic model.

Images