Knowledge-data dual-drive-based air processor fault diagnosis method

Abstract

The invention discloses an air handling unit fault diagnosis method based on knowledge-data dual drive, and relates to the field of fault diagnosis of a heating ventilation air conditioning system.In the method, training of an air handling unit fault diagnosis model comprises the steps that operation data of a unit in different states are obtained, data preprocessing is conducted, and a data set is constructed; comprising a training set and a test set. The training set comprises data without labels and data with labels. An energy balance model is used to screen key physical features, and a stack auto-encoder is used to extract depth features. The method comprises the following steps: carrying out pre-training through self-supervised learning and a Transform encoder, and training a fault classification model by using labeled data; and finally, verifying the model by using a test set to ensure that the accuracy reaches the standard.

Description

Technical Field

[0001] The present application relates to the field of HVAC system fault diagnosis, and in particular to a knowledge-data dual-driven air handling unit fault diagnosis method. Background Art

[0002] With economic development and rising demands for a better quality of life, heating, ventilation, and air conditioning (HVAC) systems have become an indispensable component of modern buildings, accounting for 40% to 80% of a building's total energy consumption. Air handling units (AHUs), as a key component of HVAC systems, are a major source of energy consumption. Due to their complex structure and variable operating conditions, AHUs are prone to various failures during long-term use. These failures not only severely impact indoor comfort but can also lead to decreased building energy efficiency and even shorten the equipment's service life. Therefore, research on AHU fault diagnosis is crucial for ensuring indoor environmental comfort, improving building energy efficiency, and conserving energy.

[0003] Benefiting from the rapid development of technologies such as artificial intelligence and data mining, data-driven fault diagnosis methods have gradually become the research focus in the field of HVAC. Most existing data-driven fault diagnosis methods are based on supervised learning, and their diagnostic performance depends to a large extent on the quantity and quality of labeled samples. However, in practical applications, obtaining accurate labels for AHU operating data is an expensive and time-consuming task, resulting in only a small part of the collected data being labeled data, and most of the data being unlabeled data. Due to the limited amount of labeled data, traditional deep learning-based fault diagnosis methods are prone to overfitting and have weak generalization capabilities, making it difficult to obtain ideal diagnostic results in real scenarios. Therefore, in the case of insufficient labeled samples, how to achieve accurate, stable and efficient identification of AHU faults has become a key issue that needs to be solved urgently. Summary of the Invention

[0004] The purpose of this application is to provide an air handling unit fault diagnosis method based on knowledge-data dual drive, which can improve the accuracy of fault diagnosis when label data is insufficient.

[0005] To achieve the above objectives, this application provides the following solutions:

[0006] In a first aspect, the present application provides an air handling unit fault diagnosis method based on knowledge-data dual drive, comprising:

[0007] Obtaining the operating data of the air handling unit to be tested;

[0008] Inputting the operating data into a trained air handling unit fault diagnosis model to obtain the fault type of the air handling unit to be detected;

[0009] Among them, the training process of the air handling unit fault diagnosis model is:

[0010] Obtain operating data of air handling units in normal status and various fault conditions;

[0011] performing data preprocessing on the operation data to obtain preprocessed operation data;

[0012] Constructing an air handling unit operation data set based on the preprocessed operation data; the air handling unit operation data set includes a training set and a test set divided in a set ratio; the training set includes unlabeled training data and labeled training data in a set ratio;

[0013] Based on the energy balance model of the air handling unit, the top k physical features with the strongest correlation with the fault type of the air handling unit are selected to form a priori feature set.

[0014] Extracting deep features of the unlabeled training data in the training set based on a stacked autoencoder model to obtain a deep feature set;

[0015] Based on the self-supervised learning framework, a Transformer-based basic encoder is constructed and pre-trained using unlabeled training data and proxy labels to obtain a pre-trained Transformer encoder; the proxy labels are a fusion feature set that integrates prior features and deep features;

[0016] The fault classification model is trained based on labeled training data to obtain an air handling unit fault diagnosis model; the fault classification model is a model in which a classification module is connected to a pre-trained Transformer encoder;

[0017] Based on the test set, the air handling unit fault diagnosis model is tested. When the test accuracy meets the preset requirements, a trained air handling unit fault diagnosis model is obtained.

[0018] According to the specific embodiments provided in this application, this application discloses the following technical effects:

[0019] The present application provides an air handling unit fault diagnosis method based on knowledge-data dual drive. First, by introducing the energy balance model of the air handling unit, the top k physical features with the strongest correlation with the fault type are screened out to form a priori feature set. Secondly, a stacked autoencoder model is used to extract deep features from unlabeled training data to obtain a deep feature set. This process enhances the model's ability to represent data, enabling the model to learn useful feature information from complex operating data, further improving the model's diagnostic accuracy. Furthermore, based on the self-supervised learning framework, a basic encoder based on Transformer is constructed, and the basic encoder is pre-trained using unlabeled training data and proxy labels that fuse prior features and deep features. This process enables the model to learn the intrinsic structure and laws of the data through self-supervised learning in the absence of sufficient labeled samples, thereby improving the generalization ability of the model. Finally, a classification module is connected to the pre-trained Transformer encoder to form a fault classification model, and it is trained using limited labeled training data. This application achieves the goal of maintaining high diagnostic accuracy even when there are insufficient labeled samples by combining multiple advanced technologies such as knowledge-data dual-driven strategy, stacked autoencoders, self-supervised learning and Transformer models. BRIEF DESCRIPTION OF THE DRAWINGS

[0020] In order to more clearly illustrate the embodiments of the present application or the technical solutions in the prior art, the following briefly introduces the drawings required for use in the embodiments. Obviously, the drawings described below are only some embodiments of the present application. For ordinary technicians in this field, other drawings can be obtained based on these drawings without creative work.

[0021] Figure 1 A flow chart of a knowledge-data dual-driven air handling unit fault diagnosis method provided in one embodiment of the present application;

[0022] Figure 2 A flow chart of the training process of the air handling unit fault diagnosis model provided in one embodiment of the present application;

[0023] Figure 3 A schematic diagram of a test bench provided in one embodiment of the present application;

[0024] Figure 4 A schematic diagram of a confusion matrix of a self-supervised learning model provided in one embodiment of the present application;

[0025] Figure 5 A schematic diagram of a confusion matrix for an air handling unit fault diagnosis method based on knowledge and data dual drive provided in one embodiment of the present application. DETAILED DESCRIPTION

[0026] The following will be combined with the drawings in the embodiments of this application to clearly and completely describe the technical solutions in the embodiments of this application. Obviously, the embodiments described are only part of the embodiments of this application, not all of the embodiments. Based on the embodiments in this application, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of this application.

[0027] In order to make the above-mentioned purposes, features and advantages of the present application more obvious and easy to understand, the present application is further described in detail below with reference to the accompanying drawings and specific implementation methods.

[0028] Example 1

[0029] This embodiment provides an air handling unit fault diagnosis method based on knowledge and data dual drive, including:

[0030] Step 101: Acquire the operating data of the air handling unit to be tested;

[0031] Step 102: Inputting the operating data into a trained air handling unit fault diagnosis model to obtain the fault type of the air handling unit to be detected;

[0032] Among them, Figure 2 As shown in Figure 2, the training process of the air handling unit fault diagnosis model is as follows:

[0033] Step 201: Acquire operating data of the air handling unit in normal state and various fault states;

[0034] Step 202: performing data preprocessing on the operation data to obtain preprocessed operation data;

[0035] Step 203: constructing an air handling unit operation data set based on the preprocessed operation data; the air handling unit operation data set includes a training set and a test set divided in a set ratio; the training set includes unlabeled training data and labeled training data in a set ratio;

[0036] Step 204: Based on the energy balance model of the air handling unit, the top k physical features with the strongest correlation with the fault type of the air handling unit are selected to form a priori feature set;

[0037] Step 205: extracting deep features of the unlabeled training data in the training set based on the stacked autoencoder model to obtain a deep feature set;

[0038] Step 206: Based on the self-supervised learning framework, a basic Transformer-based encoder is constructed, and the basic encoder is pre-trained using unlabeled training data and proxy labels to obtain a pre-trained Transformer encoder; the proxy labels are fused feature sets that fuse prior features and deep features;

[0039] Step 207: Train the fault classification model based on the labeled training data to obtain an air handling unit fault diagnosis model; the fault classification model is a model that is connected to the classification module based on the pre-trained Transformer encoder;

[0040] Step 208: Based on the test set, the air handling unit fault diagnosis model is tested. When the test accuracy meets the preset requirement, a trained air handling unit fault diagnosis model is obtained.

[0041] In some embodiments, Figure 1 As shown, when executing steps 201-208, the specific steps may be as follows:

[0042] Step 1: Collect operating data of the air handling unit in normal operation and various fault conditions, with each data type containing S samples. Perform preprocessing operations on the data, including removing outliers and supplementing missing values.

[0043] Specifically, the ASHRAR 1312-RP project involves two AHUs with the same structure and configuration, named AHU-A and AHU-B, corresponding to the experimental group and the control group. Figure 3 As shown. AHU-A operates under conditions of artificially introduced faults (such as valve jamming, damper jamming, and pipeline leakage) to generate various fault data. AHU-B operates under normal conditions to collect normal operating data. All experiments are carried out under real climate environments and building load conditions, and the data covers typical seasons such as spring, summer, and winter. This application uses summer experimental data as an implementation example, selects operating data under normal conditions and 6 typical fault conditions, and collects 1,440 samples under each working condition. In the data processing stage, outliers are first eliminated and missing values ​​are filled in to ensure data quality and model training effect. The specific fault information is shown in Table 1.

[0044] Table 1. Fault information description

[0045] Fault Category Fault description F1 Exhaust damper stuck (fully open) F2 Fan failure F3 Unstable cooling coil valve control F4 Cooling coil valve partially closed F5 Outdoor damper leak F6 Pipeline leak

[0046] Step 2: Randomly divide the preprocessed data into a training set and a test set. The training set contains only a small number of labeled samples, while the majority of samples are unlabeled. Specifically, the preprocessed data is divided into training and test sets at a ratio of 0.7:0.3. To simulate a real-world scenario with many unlabeled samples and few labeled samples, this application sets the number of labeled training data to less than 100, for example, 5, 10, 20, 30, 40, 50, or 100.

[0047] Step 3: Based on the energy balance model of the air handling unit, select the top k physical features with the highest correlation with the fault type to form a priori feature set: p = {p1, p2, ..., p k}.

[0048] Specifically, based on the energy balance model of the air handling unit, the top k physical features with the strongest correlation with the fault type are screened out. In this application, k=11 is taken to form the prior feature set: p={p1, p2, ... p 11}; The specific prior feature set information is shown in Table 2.

[0049] Table 2 Prior feature set of air handling units

[0050] Serial number Feature variables describe 1 Tsa Supply air temperature 2 Tra Return air temperature 3 Toa Outdoor air temperature 4 Tma Mixed air temperature 5 Hsa Supply air humidity 6 Hra Return air humidity 7 Psf Blower power 8 Vsa Air supply volume 9 Vra Return air volume 10 Vea Exhaust volume 11 Ecc Cooling coil load

[0051] Step 4: Build a stacked autoencoder model to extract k-dimensional deep features of unlabeled training data. This deep feature set may contain fault-related information not covered by the prior feature set.

[0052] Specifically, a stacked autoencoder model is constructed to extract k-dimensional deep features of unlabeled training data. The designed stacked autoencoder model consists of an encoder and a decoder, where the encoder is used to extract k-dimensional deep features d = {d1, d2, ..., d k}, the decoder reconstructs the sample by minimizing the mean square error between the reconstructed sample and the original sample. Assume that the unlabeled dataset is in represents the i-th unlabeled sample, N is the number of unlabeled samples, and the objective function of the stacked autoencoder in this application is as follows:

[0053]

[0054] Where E and D represent the encoder and decoder of the stacked autoencoder, respectively. By solving the above equation through the backpropagation algorithm, the deep features of the unlabeled data can be obtained:

[0055] Step 5: Within the self-supervised learning framework, a Transformer-based basic encoder is constructed and pre-trained using a large amount of unlabeled data. During this process, a fused feature set that combines prior features and deep features is used as a proxy label in the pre-training process.

[0056] Specifically, a Transformer-based basic encoder is constructed within a self-supervised learning framework and pre-trained using a large amount of unlabeled data. A fused feature set that combines prior features and deep features is used as a proxy label during pre-training. The specific steps are as follows:

[0057] 1) Construct a basic encoder based on Transformer under the self-supervised learning framework.

[0058] The constructed basic Transformer-based encoder consists of a multi-head attention mechanism, a fully connected feedforward layer, layer normalization, and a residual connection. The multi-head attention mechanism is the core part of the Transformer encoder. Given unlabeled data xi, its multi-head self-attention calculation process is as follows:

[0059] O i =concat(head1,head2,...,head H )W O ;

[0060]

[0061] Q i ,K i ,V i =Linear(X i );

[0062] Among them, W o is the weight matrix, Q i , K i 、V i are query matrix, key matrix and value matrix respectively, X i is x i After residual connection and layer normalization, multi-head self-attention is transformed into:

[0063] O A =LayerNorm( i +X i );

[0064] Then, x i The multi-head attention value is input into the fully connected feedforward layer to obtain the corresponding output:

[0065]

[0066] Where n is the number of fully connected feedforward layers, W (i) and b (i) are the weight and bias of the i-th layer, g (i) is the activation function used in layer i. After residual connection and layer normalization, the final output of the Transformer encoder is obtained:

[0067]

[0068] Among them O FF The dimension is 2k.

[0069] 2) The fused feature set consisting of the prior feature p and the deep feature d is used as the proxy label in the pre-training process.

[0070] For a given unlabeled data The corresponding proxy tag can be expressed as:

[0071]

[0072] 3) Use a large amount of unlabeled data to pre-train it.

[0073] Unlabeled data The input is the Transformer encoder, which is calculated through modules such as the multi-head attention mechanism, the fully connected feedforward layer, the layer normalization, and the residual connection to obtain the representation information: Then, the characterization information The fusion feature set F that combines prior features and deep features i For comparison, the objective function of self-supervised pre-training is constructed:

[0074]

[0075] By optimizing the above objective function through the back-propagation algorithm, all the parameters of the Transformer encoder can be obtained.

[0076] Step 6: Integrate the classification module based on the pre-trained Transformer encoder to build a complete fault classification model, and use a small amount of labeled data for training (also called fine-tuning) to obtain the final air handling unit fault diagnosis model.

[0077] Specifically, a classification module is connected to the pre-trained Transformer encoder to build a complete fault classification model, which is then trained using a small amount of labeled data to obtain the final air handling unit fault diagnosis model. The specific steps are as follows:

[0078] Given a labeled dataset in represents the i-th labeled sample, y i is its sample label, M is the number of labeled samples. First, the labeled samples Input to the Transformer encoder, extract its feature representation, then input the extracted features into the classification module and output the predicted label Finally, by minimizing the true label y i and predicted labels The cross entropy loss between is used to train the fault diagnosis model. The cross entropy loss function is as follows:

[0079]

[0080] During the parameter training process, the Transformer encoder parameters are kept unchanged (i.e., the parameters are frozen), and only the classification module parameters are optimized.

[0081] Step 7: Use test samples to verify the constructed knowledge-data dual-driven self-supervised air handling unit fault diagnosis model to evaluate its fault diagnosis performance.

[0082] Specifically, the constructed knowledge-data dual-driven self-supervised air handling unit fault diagnosis model is verified through test samples to evaluate its fault diagnosis performance.

[0083] 1) Set the training parameters. Set the learning rate of the stacked autoencoder to 0.001 and train for 65 epochs. Set the learning rate of the Transformer encoder to 0.003 and the number of pre-training epochs to 65. Finally, train the fault classification module for 100 epochs with a learning rate of 0.0008.

[0084] 2) The performance of fault diagnosis is evaluated using diagnostic accuracy, which is defined as the ratio of the number of correctly classified test samples to the total number of test samples;

[0085] 3) To fully verify the performance of this application, existing fault diagnosis methods for air handling units based on supervised learning (SVM, DT) and self-supervised learning (Self-CNN, MSFormer) are selected for comparative experiments;

[0086] 4) To avoid the randomness of the comparison results, all experiments were repeated 10 times. For each experiment, a new dataset was generated by random sampling from the preprocessed original dataset. The experimental comparison results are the average of the 10 repeated experiments.

[0087] The experimental comparison results are shown in Table 3:

[0088] Table 3. Accuracy of different models with different numbers of labels (%)

[0089]

[0090] As can be seen from Table 3, the model proposed in this application is superior to existing supervised and self-supervised learning methods in terms of fault diagnosis accuracy. Due to the ability to learn fault representations from unlabeled data, the overall performance of the fault diagnosis methods based on self-supervised learning (Self-CNN and MSFormer) is better than that of traditional supervised learning methods (SVM, DT). Among the comparison methods, MSFormer performed best, but its classification accuracy was always lower than that of the method proposed in this application. When there are only 5 labeled samples, the fault diagnosis accuracy of the model proposed in this application can reach 87.50%, while that of MSFormer is 83.70%. When the number of labeled samples increases to 100, the fault diagnosis accuracy of the model proposed in this application is increased to 97.84%, which is 2.57% higher than that of MSFormer. The above results show that this application has good fault diagnosis performance, especially when there is extremely limited labeled data.

[0091] In addition, this application further compares the diagnosis results of the MSFormer model and the method proposed in this application when the number of labeled data is 20, and plots the confusion matrix on Figure 4 and Figure 5 In. According to Figure 4 It shows that MSFormer achieves satisfactory diagnosis results on most data categories, but there is a type of fault whose diagnosis accuracy is only 79%. Figure 5 This shows that under the same conditions, the proposed method achieves higher diagnostic accuracy for all fault categories, with the lowest category accuracy reaching 88%. These results further verify that the proposed method outperforms existing methods in fault identification capabilities in scenarios with limited labeled samples.

[0092] In summary, this application has the following technical effects:

[0093] Compared with traditional deep learning fault diagnosis methods, this application successfully solves the problem of insufficient labeled data and difficulty in training high-precision fault diagnosis models, and has higher economic benefits and practical value.

[0094] The technical features of the above embodiments can be combined arbitrarily. To make the description concise, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0095] This document uses specific examples to illustrate the principles and implementation methods of this application. The description of the above examples is only intended to help understand the method and core concept of this application. At the same time, for those skilled in the art, based on the concept of this application, there may be changes in the specific implementation methods and application scope. In summary, the content of this specification should not be understood as limiting this application.

Claims

1. A knowledge-data dual-driven air handling unit fault diagnosis method, characterized in that: include: Obtaining the operating data of the air handling unit to be tested; Inputting the operating data into a trained air handling unit fault diagnosis model to obtain the fault type of the air handling unit to be detected; Among them, the training process of the air handling unit fault diagnosis model is: Obtain operating data of air handling units in normal status and various fault conditions; performing data preprocessing on the operation data to obtain preprocessed operation data; Constructing an air handling unit operation data set based on the preprocessed operation data; the air handling unit operation data set includes a training set and a test set divided in a set ratio; the training set includes unlabeled training data and labeled training data in a set ratio; Based on the energy balance model of the air handling unit, the top k physical features with the strongest correlation with the fault type of the air handling unit are selected to form a priori feature set. Extracting deep features of the unlabeled training data in the training set based on a stacked autoencoder model to obtain a deep feature set; Based on the self-supervised learning framework, a Transformer-based basic encoder is constructed and pre-trained using unlabeled training data and proxy labels to obtain a pre-trained Transformer encoder; the proxy labels are a fusion feature set that integrates prior features and deep features; The fault classification model is trained based on labeled training data to obtain an air handling unit fault diagnosis model; the fault classification model is a model in which a classification module is connected to a pre-trained Transformer encoder; Based on the test set, the air handling unit fault diagnosis model is tested. When the test accuracy meets the preset requirements, a trained air handling unit fault diagnosis model is obtained.

2. The air handling unit fault diagnosis method based on knowledge-data dual drive according to claim 1 is characterized in that: The Transformer-based basic encoder consists of a multi-head attention mechanism, a fully connected feedforward layer, layer normalization, and a residual connection. The multi-head self-attention calculation process in the multi-head attention mechanism is: O i =concat(head1,head2,...,head H )W O ; Q i ,K i ,V i =Linear(X i ); Among them, W o is the weight matrix, Q i , K i 、V i are query matrix, key matrix and value matrix respectively, X i is x i The embedded version of , where k is the dimension.

3. The air handling unit fault diagnosis method based on knowledge-data dual drive according to claim 2 is characterized in that: The formula expression of the stacked autoencoder model is: Where E and D represent the encoder and decoder of the stacked autoencoder, respectively, and x i represents the i-th unlabeled sample, and N is the number of unlabeled samples.

4. The air handling unit fault diagnosis method based on knowledge-data dual drive according to claim 3 is characterized in that: The formula expression of the proxy tag is: Among them, p k (i) is the kth prior feature, d k (i) is the kth depth feature, f 2k (i) is the 2kth fusion feature.

5. The air handling unit fault diagnosis method based on knowledge-data dual drive according to claim 4 is characterized in that: Based on the self-supervised learning framework, a basic Transformer-based encoder is constructed. The basic encoder is pre-trained using unlabeled training data and proxy labels to obtain a pre-trained Transformer encoder. Specifically, the following steps are performed: The unlabeled training data is input into the Transformer encoder, and the representation information is obtained through the multi-head attention mechanism, fully connected feedforward layer, layer normalization and residual connection module: The characterization information Comparing with the proxy label constitutes the objective function of self-supervised pre-training: The objective function of self-supervised pre-training is optimized through the back-propagation algorithm to obtain the pre-trained Transformer encoder.

6. The air handling unit fault diagnosis method based on knowledge-data dual drive according to claim 5 is characterized in that: The fault classification model is trained based on labeled training data to obtain the air handling unit fault diagnosis model, which specifically includes: Based on the pre-trained Transformer encoder, a classification module is connected to build a fault classification model; Input the labeled training data into the Transformer encoder to obtain the feature representation of the labeled training data; Input the feature representation into the classification module to obtain the predicted label Based on the true label y i and predicted labels The minimum cross entropy loss between them is used to determine the fault diagnosis model of the air handling unit.

7. The air handling unit fault diagnosis method based on knowledge-data dual drive according to claim 6 is characterized in that: The formula expression of the labeled dataset is: in represents the i-th labeled sample, y i is the sample label, and M is the number of labeled samples.

8. The air handling unit fault diagnosis method based on knowledge-data dual drive according to claim 7 is characterized in that: The formula expression of the cross entropy loss function is: in, is the predicted label.

9. The air handling unit fault diagnosis method based on knowledge-data dual drive according to claim 1, characterized in that: Performing data preprocessing on the operating data to obtain preprocessed operating data specifically includes: Outliers are eliminated and missing values ​​are filled in for the operating data to obtain preprocessed operating data.

Images