Improved deep learning model-based refrigeration unit fault detection method

Abstract

Disclosed in the present invention is an improved deep learning model-based refrigeration unit fault detection method. The method uses an LOF algorithm to remove outliers from a fault dataset, and then uses ADASYN technology to solve the problem of data imbalance. In addition, in respect of the problems that existing refrigeration unit fault diagnosis deep learning models are prone to network degradation, and refrigeration unit fault diagnosis models generally lack weighting critical features, the present invention first alleviate, on the basis of ResNet, the problem of network performance degradation which is prone to occur in deep neural network training processes, and then integrates a CBAM for capturing critical features in fault data, so as to improve the feature extraction capability of a network. Experimental results show that the LOF-ADASYN-ResNet-CBAM method provided by the present invention effectively diagnoses refrigeration unit faults.

Description

Refrigeration unit fault detection method based on improved deep learning model TECHNICAL FIELD

[0001] The application relates to the technical field of refrigeration unit fault diagnosis, and particularly relates to a refrigeration unit fault detection method based on an improved deep learning model. BACKGROUND

[0002] Building energy consumption is a hot issue in today's social development, and the refrigeration unit serves as an important component of air conditioning energy consumption. Relevant research shows that the operation energy consumption of the refrigeration unit accounts for about 30-40% of the building energy consumption, and when the unit is in a fault operation state, the equipment energy consumption will increase by about 15-30%. Timely troubleshooting of the refrigeration unit is of great significance to improving energy utilization.

[0003] Machine learning has been successfully applied in early fault diagnosis of refrigeration units. In recent years, methods based on deep learning have rapidly developed in the field of refrigeration unit fault diagnosis. On the one hand, deep learning can learn the complex nonlinear features of time series data to achieve high fault diagnosis accuracy. For example, Han uses a deep neural network (DNN) to learn the nonlinear features of fault data, and the model achieves high efficiency and stability, thereby improving the fault diagnosis accuracy. On the other hand, deep learning has the ability to extract deep features through the combination of multiple layers of neural networks. Long short-term memory networks (LSTM) have advantages in processing high-dimensional, strongly coupled and high-time-varying data. He et al. successfully applied a combination of multiple LSTMs in the field of refrigeration unit fault diagnosis to extract deep features of fault data.

[0004] However, in the actual training of deep learning models, with the increase of network layers, problems such as gradient disappearance and network degradation are prone to occur, which leads to difficulties in network training, reduced performance and other negative effects. In view of the above problems, Glorot X et al. proposed an improved strategy of Xavier initialization parameters to improve the performance and convergence speed of the network in deep structures. Later, Ioffe S et al. proposed a batch normalization technology (Batch Normalization, BN) which helps to suppress internal covariate shift and alleviate the network degradation problem.

[0005] And the emergence of residual neural network (ResNet) has made great progress in solving the above problems. ResNet model effectively solves the network degradation problem through residual connection, and compared with Xavier initialization parameters and BN technology, it focuses more on improving gradient flow and avoiding information loss in deep network, and has the ability to deeply mine data features. For example, Liu Feitian et al. used ResNet_21 for fault diagnosis of centrifugal refrigeration unit, which improved the CNN in refrigerant leakage fault. However, the ResNet_21 model has a deep layer, including 21 convolutional layers, which leads to poor training efficiency and overfitting problem, and still has more misjudgment and omission in the refrigerant excess fault state. The reason is that the model focuses on deepening the network layer and ignores the influence of key features in fault diagnosis results.

[0006] In the field of fault diagnosis, fault diagnosis should pay attention to the importance of key features for fault recognition. Deep learning without attention mechanism usually uses fixed weights to process the features of each element in the sequence. Studies have shown that introducing attention mechanism can strengthen the attention degree to key features. The mechanism is in the stage of continuous development. Convolutional block attention module (CBAM) captures key features more comprehensively through channel attention and spatial attention mechanism, and improves the fault diagnosis performance of the model on the time series data of refrigeration unit.

[0007] Therefore, the present application comprehensively considers the above problems, and aims to use a shallow ResNet model as the research basis, each residual block containing two convolutional layers. ResNet deeply mines data features while being able to alleviate the degradation problem that often occurs in deep neural network training. Combined with the perception ability of CBAM to key features, the advantages of the two are combined, which has a high fault recognition rate while reducing the complexity of the existing model, and establishes a ResNet-CBAM refrigeration unit fault detection model.

[0008] However, in actual engineering cases, there are some abnormal points (outliers) in the operation data of the refrigeration unit, and their local distribution density is low, which causes misjudgment of the diagnosis results of the classifier. The present application uses LOF (Local Outlier Factor) algorithm to detect and eliminate abnormal points in the data set, ensuring the quality of the data set.

[0009] And in actual engineering cases, the refrigeration unit is usually running in a healthy state, thus leading to the relative scarcity of fault data, and the data imbalance may cause the machine learning model to be biased towards the majority class when diagnosing faults, resulting in low fault detection rate, reducing the generalization performance of the model, and affecting the fault diagnosis result. In view of the imbalance of the refrigeration unit operation data, there are many data enhancement methods. ADASYN has more intuitive interpretation and stability than GAN and VAE, and is more targeted and adaptive than other oversampling methods such as SMOTE when dealing with class imbalance data.

[0010] After removing the abnormal points (outliers) in the data set by the LOF algorithm, the data set is expanded by ADASYN, and finally the ResNet-CBAM detection model is used for fault diagnosis.

[0011] SUMMARY

[0012] In order to remove some abnormal points (outliers) in the refrigeration unit operation data in actual engineering and cope with the problem of scarcity of fault data, the LOF-ADASYN algorithm is used to process the data set. In order to improve the accuracy of the refrigeration unit fault diagnosis and solve the problem that the existing refrigeration unit fault diagnosis technology is prone to network degradation and the detection model generally lacks weight weighting of key features, the present application provides a refrigeration unit deep learning fault diagnosis method based on LOF-ADASYN combined with ResNet-CBAM model, which realizes accurate diagnosis of refrigeration unit faults.

[0013] In the first aspect, the application provides a refrigeration unit fault detection method based on an improved deep learning model, which comprises the following steps:

[0014] Step one, obtain the operating parameters of the refrigeration unit under normal operating state and different fault types as a data set.

[0015] Step two, detect the local outliers in the data set by using the local outlier factor algorithm and remove them, and then expand the data set by using the adaptive synthetic sampling algorithm.

[0016] Step three, normalize the data set.

[0017] Step four, build a fault detection model. The fault detection model includes a ResNet module and a CBAM module. The ResNet module includes three residual blocks connected in sequence; each residual block introduces a Dropout layer. The Dropout layer randomly shuts down part of the neurons of the 1DCNN layer connected to it at a dropout rate p. The input of the residual block is added to the output of the Dropout layer by identity mapping, as the output of the residual block.

[0018] The output of the ResNet module is taken as the input of the CBAM module; the CBAM module comprises a channel attention module and a spatial attention module connected in sequence. In the spatial attention module, the channel attention weight output by the channel attention module is subjected to maximum pooling and average pooling in the channel of each feature point, to obtain feature maps respectively. The channel-based splicing is performed on the feature maps

[0019] The output feature of the CBAM module is input to a full connection layer. The fault detection model is trained by using the data set obtained in step two.

[0020] Step five, input the parameters of the refrigeration unit to be measured into the fault detection model, to determine whether the refrigeration unit to be measured has a fault and the type of the fault.

[0021] As a preferred, in the ResNet module, the input of the first residual block is the operation data of the refrigeration unit, and the output of the last residual block is the output of the ResNet module; the output of the ResNet module is input to the intermediate hidden layer; the hidden layer output feature h = [h1, h2,..., hn] obtained by the ResNet module extracting relevant features is sequentially input to the channel attention module and the spatial attention module. N

[0022] As a preferred, the expression of the output feature Y d of the Dropout layer is as follows:

[0023] Wherein, x is the input of the Dropout layer, and p is the dropout rate.

[0024] As a preferred, the value of the dropout rate p is 0.2.

[0025] As a preferred, the residual block comprises a first 1DCNN layer, an activation function layer, a BN layer, a second 1DCNN layer, a Dropout layer and an identity mapping layer connected in sequence; the input of the first 1DCNN layer and the output of the Dropout layer are jointly input to the identity mapping layer for superposition.

[0026] As a preferred, in the channel attention module, the hidden layer output feature output by the ResNet module is subjected to global maximum pooling and global average pooling operations, to obtain feature maps respectively. ​The two inputted multi-layer perceptrons generate two feature maps 1 and 2 respectively. In the multi-layer perceptron, the channel dimension of the feature map is reduced by channel reduction. The reduction ratio r of the channel reduction is a factor of the number of running parameters, and r>1. The feature map 1 and the feature map 2 are spliced to obtain a new feature map; the new feature map is activated by a Sigmoid function, and each element in the new feature map is assigned a new weight; the weight is the channel attention weight M C (F).

[0027] As preferred, the feature vector Y obtained after the weight weighting of the CBAM module is input to a fully connected layer. The weight parameters in the fully connected layer combine the input features to generate high-dimensional global features Z. The weight parameters will be learned and updated through the back propagation algorithm during the training process. The global features Z are input to the output layer of the fully connected layer. The output layer uses a Softmax function. The output layer includes n neurons. n is the number of refrigeration unit states; the n neurons are respectively used to judge n-1 fault states and the normal state of the refrigeration unit.

[0028] As preferred, the specific process of step two is as follows:

[0029] 1) Construct the reachable distance reach_d k (b,o) of any point b in the data set k (b,o)=max{d k (o),d(b,o)} (1)

[0030] Wherein, point o is any one of the neighboring points of point b; d k (o) is the distance between the kth neighboring point and point o; d(b,o) represents the Euclidean distance between point b and point o.

[0031] 2) Calculate the local reachable density Ird k (b) as follows:

[0032] Wherein, N k (b) is the set of K-neighbor data points of point b:

[0033] 3) Calculate the local outlier factor LOF k (b) of point b:

[0034] 4) After removing the local outlier factor LOF k (b), determine the number of samples G to be generated: G=(m l -m s )*β (4)

[0035] wherein m l represents the number of existing majority class samples; m s represents the number of existing minority class samples, and β∈[0,1].

[0036] 5) Calculate the proportion r of majority class samples in the K nearest neighbors of each minority class sample i : r i = Δ i / K (5)

[0037] wherein Δ i is the number of samples belonging to the majority class in the K nearest neighbors of the i-th minority class sample;

[0038] 6) Calculate the normalized proportion as follows:

[0039] 7) Calculate the number g of new samples that each minority class sample needs to generate i as follows:

[0040] 8) For each minority class sample x i , generate g i new samples: s i =x i +(x zi -x i )*γ (8)

[0041] wherein s i is the newly generated sample; x zi is a randomly selected minority class sample in the K nearest neighbors of the minority class sample x i ; and γ∈[0,1] is a random number.

[0042] As a preference, the operating parameters include one or more temperature parameters, one or more valve position parameters, one or more pressure parameters, one or more flow parameters, one or more power parameters, and one or more current parameters.

[0043] As a preference, the identified chiller unit is a water chiller unit; the fault types of the water chiller unit include one or more of excessive lubricating oil, condenser fouling, refrigerant leakage, excessive refrigerant, non-condensable gas, insufficient cooling water, and insufficient chilled water.

[0044] In a second aspect, the present application provides a computer device, comprising a memory, a processor and a computer program stored in the memory and executable on the processor, wherein the memory stores the computer program; and the processor executes the improved deep learning model-based refrigerating unit fault detection method.

[0045] In a third aspect, the present application provides a readable storage medium storing a computer program; when the computer program is executed by a processor, the improved deep learning model-based refrigerating unit fault detection method is implemented.

[0046] The present application has the following advantages:

[0047] 1. The present application introduces a Dropout layer in the ResNet module of the ResNet-CBAM model. The Dropout layer randomly turns off part of the neurons of the 1D CNN layer connected thereto at a dropout rate p to reduce the risk of overfitting and enhance the generalization ability of the model; and the present application introduces a multi-scale convolution in the spatial attention module to capture different receptive field features in the feature map; based on this, the present application alleviates the network degradation problem and focuses on the extraction of key features, overcoming the problems of "easy network degradation" and "lack of weight weighting of key features of fault data" in existing deep learning models for refrigerating unit fault diagnosis.

[0048] 2. The present application successfully removes the outliers in the data set using the LOF algorithm, ensures the quality of the data set, and then successfully increases the samples of fault data using ADASYN (adaptive synthetic sampling), realizes data enhancement, and solves the problem of data imbalance in actual engineering.

[0049] 3. The detection model of the present application overcomes the characteristics of "EO, RO and RL as refrigerating unit system faults have strong coupling and are difficult to identify", and realizes high-accuracy identification of the three types of faults. BRIEF DESCRIPTION OF DRAWINGS

[0050] Fig. 1 is a refrigerating unit fault diagnosis flowchart of the present application;

[0051] Fig. 2 is a residual block structure diagram in the present application;

[0052] Fig. 3 is a model structure diagram of CBAM in the present application;

[0053] Fig. 4 is a LOF-ADASYN-ResNet-CBAM model structure diagram used in the present application;

[0054] Fig. 5 is a model parameter tuning diagram of ResNet-CBAM in the present application;

[0055] FIG. 6 is a loss function curve of the training sample set and the validation sample set of the ResNet-CBAM model in the application under the balanced data set;

[0056] FIG. 7 is an accuracy iterative curve of each detection model of the application under an oversampling rate of 200%;

[0057] FIG. 8 is a comparison diagram of the diagnostic results of the model used in the application and a plurality of existing models on the test sample set under an oversampling rate of 200%. DETAILED DESCRIPTION

[0058] In order to more clearly illustrate the purpose, technical scheme and advantages of the refrigeration unit, the application will be described in more detail below in conjunction with specific embodiments and with reference to the accompanying drawings. The embodiments only represent some embodiments of the application.

[0059] Embodiment 1

[0060] As shown in FIGS. 1-4, a refrigeration unit fault detection method based on an improved deep learning model is based on an end-to-end fault detection model. The fault detection model is based on the Python3.7.0 platform, the Pytorch1.13.1 deep learning framework is used to build the fault detection model, and the training is performed on the hardware device of the NVIDIAGEForce RTX 3060 Ti GPU. In this embodiment, the refrigeration unit is a water chiller; in some other embodiments, the refrigeration unit can also be a air-cooled heat pump unit or other types of refrigeration unit.

[0061] The refrigeration unit fault detection method comprises the following steps:

[0062] Step one, obtain the operating parameters of the water chiller under normal operating conditions and different fault types, and form a sample set with labels. In this embodiment, the ASHRAE RP-1043 data set is used, which is derived from the description of the water chiller fault simulation experiment in the ASHRAE RP-1043 standard. The experiment involves a centrifugal water chiller with a rated power of 90 cold tons (about 316kW), using R134a as the refrigerant. ASHRAE simulated 27 operating conditions under normal conditions and 7 fault conditions by adjusting three control variables: chilled water temperature, condenser water inlet temperature and condenser cooling load. Each fault condition has been tested for four different levels of fault grades. The simulation experiment runs for about 14 hours for each state, and the data sampling interval is 10 seconds. Here, Nor represents the normal state, EO, RL, and RO represent system faults, and CF, NC, FWC, and FWE represent local faults. The specific fault type names are shown in Table 1.

[0063] Table 1 ASHRAE RP-1043 SL1 fault types

[0064] In this embodiment, the ASHRAE-1043-RP data collection interval is 10 seconds, and a total of 64 parameters are recorded. Among them, 48 parameters are directly collected by sensors, including 29 temperature parameters, 7 valve position parameters, 5 pressure parameters, 2 flow parameters, 1 power parameter and 1 current parameter, etc. The remaining 16 parameters are calculated in real time by VisSim software. Because the chiller experiences large changes in some variables during the short time of starting and stopping, the effectiveness of these data is lower during data collection. Therefore, when selecting the data set, the steady-state data is used.

[0065] Two different embodiments are set up, corresponding to balanced data set and unbalanced data set respectively.

[0066] The embodiment under the balanced data set uses seven types of fault data with the least severe level (SL1) (see Table 1 for fault types) and normal data as the sample set, 1200 for each state, to form a 9600x64-dimensional sample data set ω, which is randomly divided into training set, validation set and test set according to the ratio of 6:2:2;

[0067] The embodiment under the unbalanced data set uses 1200 normal data and 300 fault data for each type, and performs experiments under the unbalanced ratio of 4:1. The data set division method is consistent with that of the balanced data set embodiment;

[0068] Step two, use LOF (Local Outlier Factor) algorithm to detect local outliers in the data set and remove them, and then use ADASYN (Adaptive Synthetic Sampling) to expand the data set;

[0069] In the embodiment of expanding the fault data set using LOF-ADASYN, the data under each state in the training set and the validation set is 1200 under the balanced data set. For this purpose, 1200 data under the normal state are taken, and the rest of each fault state is first removed using the LOF algorithm to remove outliers in the fault data set, and then the ADASYN method is used to collect different amounts of data according to different oversampling rates (when the oversampling rate is 100%, the original fault data for each type is 600; when the oversampling rate is 200%, the original fault data for each type is 400), and finally the data is balanced, the total data of the training set and the validation set is about 9600, and the proportion of the training set and the validation set is divided as 5:1, and the test set is selected from the original ASHRAE RP-1043 data 3000, that is, 375 for each type, to verify the generalization ability of the deep learning model.

[0070] "Over-sampling rate" refers to the proportion of the number of synthetic samples generated by using over-sampling technology to the number of original minority class samples when dealing with data imbalance problems.

[0071] The steps of LOF-ADASYN to expand the fault dataset are as follows:

[0072] 1) Calculate the local density of each point:

[0073] Calculate the distance between point b and the Kth nearest neighbor, denoted as d k (b), d(b,o) represents the Euclidean distance between b and o: reach_d k (b,o) = max{d k (o), d(b,o)} (1)

[0074] 2) Calculate the local reachable density Ird k (b), N k (b) is the set of K nearest neighbor data points of b:

[0075] 3) Calculate the local outlier factor LOF k (b) of point b:

[0076] 4) After removing the local outlier factor determined in the above steps, determine the number of samples to be generated: G = (m l -m s )*β (4)

[0077] Where m l represents the number of existing majority class samples, and m s represents the number of existing minority class samples, and β ∈ [0,1]. When β = 1, the majority class and the synthetic minority class samples are finally equal, and G is the total number of samples finally generated;

[0078] 5) Calculate the proportion of majority class samples in the K nearest neighbors of each minority class sample: r i = Δ i / K (5)

[0079] Calculate the ratio r i , the numerator Δ i is the number of samples belonging to the majority class in the K nearest neighbors of the ith minority class sample;

[0080] 6) Standardize r i :

[0081] 7) Calculate the number of new samples to be generated for each minority class sample:

[0082] 8) For each minority sample x i , generate g i new samples: s i = x i + (x zi - x i )*γ (8)

[0083] In the above formula, s i is the newly generated sample; x zi is a randomly selected minority sample in the Kth nearest neighbor of x i , and γ∈[0,1] is a random number.

[0084] Step three, normalize the sample data set ω, adopt Min-Max normalization technique to ensure that the result value is mapped between [0-1], as follows:

[0085] The sample set processed in step two is divided into training sample set w train , test sample set w test and validation sample w val in proportion; and the sample data set w is normalized pretreated, adopting Min-Max normalization technique to ensure that the output value is mapped between [0-1]. The formula of this technique is as follows:

[0086] Where x and x' are the data before and after normalization respectively; min and max are the maximum and minimum values in the data respectively.

[0087] Step four, build and train the ResNet-CBAM model.

[0088] The parameter selection in the detection model has an important influence on the detection performance of the model. Therefore, the ResNet-CBAM model is parameter-optimized in this embodiment to achieve the best performance of the model. However, the setting of hyperparameters has randomness and largely depends on experimental experience.

[0089] Since the input dimension of the chiller fault data is 64, the channel reduction ratio in the multi-layer perception (MLP) in CBAM is generally the input dimension factor of the chiller fault data. After trying various reduction ratios (such as 8, 16, 32), it is determined through experiments that r = 16 can provide the best diagnostic performance.

[0090] The functions of the two fully connected layers in the multi-layer perceptron (MLP) in the CBAM are to compress and expand the features, cooperate with the nonlinear activation function, calculate the channel attention weight, and thus enhance the focusing ability of the model on the key features. The MLP in the channel attention module includes two fully connected layers. The first fully connected layer reduces the channel number to 4 through a reduction ratio, performs feature screening in a low-dimensional space, and the dimension reduction process reduces the number of parameters and the computational complexity. By reducing the dimension and the amount of information, the key feature information in the input MLP is preserved. The second fully connected layer restores the compressed channel number 4 to 64 dimensions, maps the low-dimensional features back to the high-dimensional space, and reconstructs richer feature representations, i.e., strengthens the features. The weight weighting of the key features is realized through the activation function (Sigmoid) in the second fully connected layer. In summary, the expansion process strengthens and weights the key features screened by the above compression process.

[0091] The initial hyperparameters of the detection model used are shown in Table 2. In this embodiment, the influence of the key hyperparameters on the model diagnosis performance is compared through experiments. In each experiment, one hyperparameter is adjusted, while the other hyperparameters remain unchanged. Each experiment is repeated 5 times, and the average value of the diagnosis index is recorded. The experimental results are shown in FIG. 5. Finally, it is determined that the number of residual blocks of the ResNet-CBAM model is 3, the number of convolutional layers of each residual block is 2, a total of 6 convolutional layers, denoted as ResNet_6. ResNet_6 reduces 15 convolutional layers compared with ResNet_21. In ResNet_6, the convolution kernel size is 3, the learning rate is 0.001, and the fault diagnosis effect is the best. The remaining hyperparameters are consistent with Table 2.

[0092] Table 2 Initial hyperparameter table of ResNet-CBAM network structure

[0093] The ResNet-CBAM model includes a ResNet module and a CBAM module.

[0094] The ResNet module includes three residual blocks connected in sequence; each residual block includes a first 1D CNN (one-dimensional convolutional neural network) layer, an activation function layer, a BN (batch normalization) layer, a second 1D CNN layer, a Dropout (random inactivation) layer, and an identity mapping layer connected in sequence, each layer performing relevant feature extraction; the input of the first 1D CNN layer and the output of the Dropout (random inactivation) layer are jointly input to the identity mapping layer for superposition.

[0095] Based on the above structure, the internal structure of the residual block contains two main branches, one branch performs identity mapping x, that is, the input feature of the first residual block, and the other branch is the residual information F(x) learned by each neural network layer, and the two are added to form the output H(x) of the residual block;

[0096] The formula for constructing the residual block is: H(x) = F(x) + x (10)

[0097] The CBAM module includes a channel attention module and a spatial attention module connected in sequence, and a multi-scale convolution is introduced in the spatial attention module to capture different receptive field features in the feature map.

[0098] The working process of the ResNet-CBAM model is as follows:

[0099] Step 4-1. In the working process of the ResNet-CBAM model, N sequence data X = [X1, X2,..., X N ] samples of a time step are input into the ResNet module.

[0100] The first 1DCNN layer of the first residual block receives the chiller operation sequence data, and then each layer performs feature extraction in turn, and the output feature Y d = [Y d1 , Y d2 ,..., Y dn ] after the Dropout layer:

[0101] x is the input of Dropout, p is the dropout rate, is a scaling factor to ensure that the expected value of the neurons retained is times the original expected value, keeping the expected value of the Dropout layer output unchanged;

[0102] The input feature X = [X1, X2,..., X N ] is added to the output feature Y d = [Y d1 , Y d2 ,..., Y dn ] of the Dropout layer by identity mapping, and then input into the next residual block;

[0103] The input of the first residual block is the chiller operation data, and the output of the last residual block is the output of the ResNet module; the ResNet module output is input into the intermediate hidden layer;

[0104] The weight layer in ResNet is a 1D CNN layer. During training, w train is subjected to one-dimensional convolution and BN operation in the ResNet module. BN helps to reduce internal covariate shift. The dropout layer randomly turns off part of the neurons of the second 1D CNN layer with a probability of p = 0.2 to reduce the risk of overfitting and enhance the generalization ability of the model.

[0105] The formula of one-dimensional convolution is as follows:

[0106] where N represents the number of convolution kernels of the l-1 layer, represents the i-th feature map of the l layer, represents the corresponding convolution kernel weight matrix, represents the bias term, and f(·) represents a nonlinear activation function. In the ResNet module of the present application, a Relu activation function is used.

[0107] The main role of the ReLU (Rectified Linear Unit) activation function is to introduce a nonlinear mapping, so that the network can learn more complex feature representations. The formula of the ReLU activation function is as follows: x represents the input feature; F(x) = max(0, x) (13)

[0108] The formula of BN is as follows:

[0109] where, is the normalized feature value, and represents the input feature x i after batch normalization processing. μ and σ are the mean and standard deviation of the input feature x i on a batch. ε is a small normal number to ensure the stability of division, which is used to prevent the problem of division by zero when the variance is zero;

[0110] After the relevant features are extracted by the ResNet module, the hidden layer output features h = [h1, h2,..., h N ] are obtained.

[0111] Step 4-2. The hidden layer output h is first subjected to global maximum pooling (Maxpool) and global average pooling (Avgpool) operations in the CBAM module, and and are obtained. Subsequently, the and The feature map 1 and the feature map 2 in the illustration of the specification channel attention module are respectively generated, and the maximum pooling is used to capture the most significant features in each channel, and the average pooling is used to capture the average features in each channel. In the hidden layer of the MLP, the channel dimension is reduced by using channel reduction (R C / r*1*1 , where r is the reduction ratio). After the feature map 1 and the feature map 2 are subjected to element concatenation (Concat) and Sigmoid function activation, a new weight, i.e., a channel attention weight M C (F) = [M1, M2, …, M N ] is given to each element in the new feature map. The weight calculation process is as shown in formula 5:

[0112] where W1 and W0 are weights of the multi-layer perceptron, and the inputs of the two are shared. and represent the maximum pooling and the average pooling operations, respectively. σ(·) represents a Sigmoid function, and the Sigmoid function is as shown in formula 6:

[0113] Subsequently, the spatial attention module receives the channel attention weight feature as input, performs maximum pooling and average pooling on each channel of the feature point, and obtains and The two feature maps obtained are subjected to channel-based concatenation, and a multi-scale convolution operation is performed, which can simultaneously consider different sizes of receptive fields, capture information of different scales in space, and then perform feature fusion, so as to obtain more rich and comprehensive feature representation and improve the performance of the model. After Sigmoid function activation, the spatial attention weight M s (F) = [M s1 , M s2 , …, M sN ] is obtained, and a feature vector Y = [Y1, Y2, …, Y N ] weighted by the CBAM module weight is also obtained. The weight calculation process is as shown in formula 7:

[0114] In the formula, f n1 , f n2 , f n3 are the kernel sizes of the three convolution layers in the spatial attention module, and satisfy the constraint condition n1+2=n2=n3-2, and n1, n2, n3 are positive integers.

[0115] Step 4-3. The feature vector Y = [Y1, Y2, …, Y N] into a fully connected layer. The weight parameters in the fully connected layer combine the features extracted by the ResNet-CBAM model to generate high-dimensional global features Z = [Z1, Z2,..., Z N' ](N' > N). The weight parameters are learned and updated during the training process through the backpropagation algorithm to minimize the loss function and optimize the performance of the network. Second, the global features Z = [Z1, Z2,..., Z N' ] are input into the output layer, which ensures that the output and input dimensions of the fully connected layer and the output layer are consistent. The output layer uses the Softmax function to convert the raw scores output by the model into a probability distribution for classification decision-making.

[0116] where Softmax(z) i represents the i-th element in the output vector of the Softmax function, e is the base of the natural logarithm, K is the length of the vector z, and the Softmax function converts each element z i in the vector to a real number between 0 and 1, so that the sum of all elements is 1.

[0117] The output layer output dimension contains n neurons, aiming to diagnose n states of the water chiller, i.e., n-1 fault states and 1 normal state. The optimization algorithm uses Yogi, which is a second-order gradient estimation-based optimization algorithm that combines the features of Adam and RMSProp and adjusts the learning rate based on the historical gradient information of the parameters. The loss function uses multi-classification cross-entropy (Categorical Cross-Entropy) to update the model parameters w and b to improve the probability of correct classification. After training the ResNet-CBAM model, save the best weights of the model;

[0118] Step five, train the ResNet-CBAM model on the training sample set w train and the validation sample set w val After training the ResNet-CBAM model on the training sample set w and the validation sample set w , save the best weights of the model. At the same time, generate the loss function curve. The loss function curve is a graph that measures the difference between the model's predicted values and the observed values during the model's iterative training process. As the training progresses, the loss curve decreases, indicating that the model's fitting effect on the training data gradually improves. The loss function curve of the embodiment under the balanced data set is shown in Figure 6. The optimization algorithm uses Yogi, which is a second-order gradient estimation-based optimization algorithm that combines the features of Adam and RMSProp and adjusts the learning rate based on the historical gradient information of the parameters. The loss function uses multi-classification cross-entropy (Categorical Cross-Entropy) to update the model parameters to improve the probability of correct classification;

[0119] Step six, the test sample set w test Load the model weight obtained in step 4, and obtain the diagnostic result of the final output layer as the basis for fault detection;

[0120] Step seven, the embodiment of expanding the fault data set by LOF-ADASYN, different oversampling rates are used to collect different amounts of fault data in the fault data expansion process, and after data balancing, the training process repeats the above steps.

[0121] Select the classic deep learning models 1DCNN, GRU, ResNet model and combined model for fault diagnosis performance comparison (SA is self-attention mechanism, MHSA is multi-head self-attention mechanism). The fault diagnosis results of the three embodiments are shown in Tables 3-7.

[0122] Table 3 Balanced embodiment chiller fault diagnosis model results

[0123] Table 4 Overall evaluation index of balanced embodiment chiller fault diagnosis model results

[0124] From Tables 3 and 4, it can be concluded that under the balanced data set, the recognition accuracy of 1DCNN and GRU models for EO, RO, and RL system faults is significantly lower than that of ResNet_6 model. ResNet_6 extracts more information from the data and integrates CBAM module, which pays more attention to the extraction of key features. For the four types of local faults, ResNet_6-CBAM model is slightly better than the other three detection models.

[0125] Table 5 Unbalanced embodiment chiller fault diagnosis model results

[0126] Table 6 Overall evaluation index of unbalanced embodiment chiller fault diagnosis model results

[0127] From Tables 5 and 6, it can be concluded that under the unbalanced data set, the diagnosis accuracy of RO and RL faults has a significant downward trend compared to the previous embodiment. Especially for RL faults, the diagnosis rate of each model has decreased significantly. For the four types of local faults, the fault recognition degree of the four models has decreased slightly.

[0128] From Table 6, it can be seen that for the ResNet_6 model with the most obvious decrease in overall evaluation index, the accuracy rate decreased by 5.9%, the precision rate decreased by 5.5%, the recall rate decreased by 7.3%, and the F1 decreased by 6.8%.

[0129] Table 7 Overall evaluation index of water chiller fault model under different oversampling rates of LOF-ADASYN

[0130] SA in the above table 7 represents a self-attention mechanism, and MHSA represents a multi-head self-attention mechanism. As can be seen from table 7, the seven models are all tested under four different oversampling rates. It can be seen that under each oversampling rate, the ResNet_6-CBAM model proposed in the embodiment is superior to other control models in three overall evaluation indexes. Under the oversampling rate of 200%, compared with the GRU with the lowest index, the diagnostic accuracy rate is increased by 6.1%, the recall rate is increased by 5.8%, and the F1 is increased by 6.0%. FIG. 7 is an accuracy rate iteration curve of each detection model under an oversampling rate of 200%, and the accuracy rate curve of the ResNet_6-CBAM model is the most stable relative to the other three models. FIG. 8 is a diagnostic result diagram of each detection model on the test set under an oversampling rate of 200%. Therefore, it can be concluded that the ResNet_6-CBAM model has the highest robustness and the strongest generalization.

[0131] Table 8 Comparison of water chiller fault diagnosis results of ResNet_6-CBAM model under different conditions

[0132] Embodiment 2

[0133] A kind of refrigeration unit fault detection method based on improved deep learning model, the difference between the embodiment and embodiment 1 lies in that the measured refrigeration unit is specifically air-cooled heat pump unit.

[0134] This embodiment relies on the existing air-cooled heat pump unit experimental equipment in the laboratory, and designs and builds a fault simulation test bench. Based on the enthalpy difference test test bench, the diagnosed object is introduced, various sensors are arranged on the tested air-cooled heat pump unit, the original system is appropriately modified, and a scheme for simulating the fault is developed to simulate the fault state of the air-cooled heat pump unit.

[0135] The tested machine is a split type heat pump floor air conditioner, the model is KFR-72LW, the rated refrigerating capacity of the unit is 7.2kW, the heating capacity is 8kW, the evaporator and the condenser both use air fin tube heat exchanger, the throttling device uses a thermal expansion valve, and the compressor inlet is equipped with a thermal expansion valve temperature bulb.

[0136] C# upper computer software is designed on the small air-cooled heat pump unit test bench in the laboratory, 18 temperature sensors and 4 pressure sensors are arranged, and data acquisition is performed every 10 seconds, and 22 channels of sensor data are successfully acquired.

[0137] Table 9 Air-cooled heat pump unit fault types

[0138] The embodiment under balanced data set selects 400 pieces of each state, constitutes a sample data set ω of a 1600x22 matrix, and is randomly divided into a training set, a verification set and a test set according to a 6:2:2 ratio;

[0139] The embodiment under unbalanced data set selects 400 pieces of normal data and 200 pieces of each fault data, performs experiments under an unbalanced ratio of 2:1, and the data set division method is consistent with that of the embodiment under balanced data set;

[0140] In the embodiment of expanding the fault data set by using LOF-ADASYN, finally under the balanced data set, the data of each state in the training set and the verification set is 400 pieces. For this purpose, 400 pieces of data in the normal state are taken, the abnormal points in the fault data set are removed by using the LOF algorithm for each fault state, and then the ADASYN method is used to collect different amounts of data according to different oversampling rates (when the oversampling rate is 100%, the original fault data of each type is 200 pieces; and when the oversampling rate is 200%, the original fault data of each type is 133 pieces). Finally, the data is balanced, the total number of data in the training set and the verification set is about 1600, and the ratio of the training set and the verification set is divided as 5:1. The test set is another 500 pieces of original data of the air-cooled heat pump data set, that is, 125 pieces of each type, which is used to verify the generalization ability of the deep learning model.

[0141] In the ResNet_6-CBAM model in this embodiment, the input dimension of the ResNet module and the channel reduction ratio in the multi-layer perception (MLP) in CBAM are changed. The input dimension of the ResNet module is 22, the channel reduction ratio in the multi-layer perception (MLP) in CBAM is the input dimension factor of the air-cooled heat pump unit fault data, and after trying two reduction ratios 2 and 11, it is determined through experiments that r=11 can provide the best diagnostic performance. The other hyperparameters of the ResNet_6-CBAM model are consistent with those of the water chiller unit fault diagnosis embodiment.

[0142] Table 10 Model results of the air-cooled heat pump unit fault diagnosis in the balanced embodiment

[0143] Table 11 Overall evaluation index of the model results of the air-cooled heat pump unit fault diagnosis in the balanced embodiment

[0144] Table 12 Model results of the air-cooled heat pump unit fault diagnosis in the unbalanced embodiment

[0145] Table 13 Overall evaluation index of the model results of the air-cooled heat pump unit fault diagnosis in the unbalanced embodiment

[0146] Table 14 Overall evaluation index of air-cooled heat pump unit fault model under different oversampling rates of LOF-ADASYN

[0147] Table 15 Comparison of air-cooled heat pump unit fault diagnosis results under different conditions of ResNet_6-CBAM model

[0148] To make the purposes, technical solutions and advantages of the embodiments of the present application clearer, those skilled in the art should understand that the discussion of any of the above embodiments is only exemplary and is not intended to imply that the scope (including claims) of the present disclosure is limited to these examples. Under the idea of the present application, the parameters (such as network layers) in the above embodiments or different embodiments, hyperparameters (such as batch size, learning rate, dropout, etc.), optimization algorithms, activation functions, etc. can be modified according to the actual situation, and there are many other changes of different aspects of the present application as described above. In order to be brief, they are not provided in details.

[0149] Embodiments of the present application are intended to cover all such alternatives, modifications and variations that fall within the broad scope of the appended claims. Therefore, anything within the scope of the present application is within the scope of the present application.

Claims

1. A method for refrigeration unit fault detection based on an improved deep learning model, characterized in that: The method comprises the following steps: Step one, obtaining the operating parameters of the refrigeration unit in normal operating state and different fault types as a data set; Step two, detecting the local outliers in the data set by using the local outlier factor algorithm and eliminating them, and then expanding the data set by using the adaptive synthetic sampling algorithm; Step three, normalizing the data set for pretreatment; Step four, constructing a fault detection model; the fault detection model comprises a ResNet module and a CBAM module; the ResNet module comprises three residual blocks connected in sequence; each residual block introduces a Dropout layer; the Dropout layer randomly shuts down part of the neurons of the 1DCNN layer connected thereto at a dropout rate p; the input of the residual block is added to the output of the Dropout layer in a constant mapping to serve as the output of the residual block; The output of the ResNet module serves as the input of the CBAM module; The CBAM module comprises a channel attention module and a spatial attention module connected in sequence; In the spatial attention module, the channel attention weights output by the channel attention module are subjected to maximum pooling and average pooling on the channel of each feature point, to obtain feature maps To the feature map The channel-based splicing is performed, and the obtained features are subjected to three different scale convolution operations; the three convolution operation results are fused and activated; the convolution kernel sizes n1, n2 and n3 of the three different scale convolution operations satisfy the following constraint condition: n1+2=n2=n3-2, and n1, n2 and n3 are positive integers; The output features of the CBAM module are input to a fully connected layer; the fault detection model is trained by using the data set obtained in step two; Step five, inputting the parameters of the measured refrigeration unit into the fault detection model to determine whether the measured refrigeration unit has a fault and the type of the fault.

2. The method of claim 1, wherein the method comprises: In the ResNet module, the input of the first residual block is the operating data of the refrigerating unit, and the output of the last residual block is the output of the ResNet module; the output of the ResNet module is output to the intermediate hidden layer; the hidden layer output features h = [h1, h2, …, h N ] obtained by the ResNet module extracting relevant features are sequentially subjected to the channel attention module and the spatial attention module; Output features Y of the Dropout layer in the residual block d The expression is as follows: Wherein, x is the input of the Dropout layer, and p is the dropout rate.

3. The method of claim 2, wherein the method comprises: The value of the dropout rate p is 0.

2.

4. The method of claim 1, wherein the method comprises: The residual block comprises a first 1DCNN layer, an activation function layer, a BN layer, a second 1DCNN layer, a Dropout layer and a constant mapping layer connected in sequence; the input of the first 1DCNN layer and the output of the Dropout layer are jointly input to the constant mapping layer for superposition.

5. The method of claim 1, wherein the method comprises: In the channel attention module, the hidden layer output features output by the ResNet module are subjected to global maximum pooling and global average pooling operations to obtain feature maps feature maps respectively. In the multi-layer perceptron, the channel dimension of the feature maps is reduced by channel reduction. The reduction ratio r of the channel reduction is a factor of the number of operating parameters, and r>

1. The feature map 1 and the feature map 2 are subjected to element splicing to obtain a new feature map. The new feature map is activated by a Sigmoid function to assign a new weight to each element in the new feature map. The weight is a channel attention weight M C (F).

6. The method of claim 1, wherein the method comprises: The feature vector Y obtained by weighting the CBAM module is input to a fully connected layer; the weight parameters in the fully connected layer combine the input features to generate high-dimensional global features Z; the weight parameters are learned and updated through the back propagation algorithm during the training process; the global features Z are input to the output layer of the fully connected layer; the output layer uses a Softmax function; the output layer comprises n neurons; n is the number of refrigeration unit states; the n neurons are respectively used to determine n-1 fault states and the normal state of the refrigeration unit.

7. The method of claim 1, wherein the method comprises: The specific process of step two is as follows: 1) reach_d of any point b in the constructed dataset k The expression of (b, o) is as follows: reach_d k (b,o) = max{d k (o), d(b,o)} (1) wherein point o is an arbitrary one of the neighboring points of point b; d k (o) is the distance between the kth neighboring point and point o Distance; d(b, o) represents the Euclidean distance between points b and o; 2) calculating the local reachable density Ird k (b) as follows: where N k (b) is the set of K nearest neighbor data points of point b 3) calculating the local outlier factor LOF of point b k (b): 4) Calculate the local outlier factor LOF k (b) After culling, determine the number of samples G that need to be generated: G = (m l -m s )*β (4) wherein m l denotes the number of existing majority class samples; m s denotes the number of existing minority class samples, β ∈ [0, 1]; 5) Calculate the proportion of majority class samples r in K nearest neighbors of each minority class sample i : r i = Δ i / K (5) where Δ i is the number of samples belonging to the majority class among the K nearest neighbors of the ith minority sample; 6) Calculate the normalized fraction r i * As follows: 7) Calculate the number of new samples g each minority class sample needs to generate i As follows: g i = r i * *G (7) 8) For each minority sample x i , generate g i new samples: s i = x i + (x zi - x i )*γ (8) where s i is the newly generated sample; x zi is a minority class sample randomly selected from the Kth nearest neighbors of the minority class sample x i is a random number.

8. The method of claim 5, wherein the method further comprises: The identified refrigeration unit is a chiller; the fault types of the chiller include one or more of excessive lubricating oil, condenser fouling, refrigerant leakage, excessive refrigerant, non-condensable gas, insufficient cooling water and insufficient chilled water; the operating parameters include one or more temperature parameters, one or more valve position parameters, one or more pressure parameters, one or more flow parameters, one or more power parameters and one or more current parameters.

9. A computer device comprising a memory, a processor, and a computer program stored on the memory and executable on the processor, characterized in that: The memory stores a computer program; and the processor executes the refrigerating unit fault detection method based on the improved deep learning model according to any one of claims 1-8.

10. A readable storage medium, storing a computer program; characterized in that: The computer program is executed by the processor to implement the refrigerating unit fault detection method based on the improved deep learning model according to any one of claims 1-8.

Images