A Smart Fault Diagnosis Method for Heat Pump Systems Based on Human Data

By training the diagnostic network with manually labeled data that generates directional tags, the problem of insufficient training data in the fault diagnosis of heat pump systems is solved, achieving efficient and reliable fault diagnosis and improving the quality of operation and maintenance.

CN115753160BActive Publication Date: 2026-11-10ZHEJIANG UNIV OF TECH
View PDF 2 Cites 0 Cited by

Patent Information

Application Number
CN202210953909.3
Authority / Receiving Office
CN · China
Patent Type
Patents(China)
Current Assignee / Owner
Filing Date
2022-08-10
Publication Date
2026-11-10
Estimated Expiration
2042-08-10

AI Technical Summary

Technical Problem

Existing fault diagnosis methods for heat pump systems rely on data-driven approaches, but are limited by the quality and quantity of training data. Industries lacking labeled data find it difficult to promote and apply these methods, and it is also difficult to achieve accurate quantitative analysis of deviation characteristics.

Method used

By generating artificially labeled data with directional tags, training a diagnostic network using a deep model, and combining this with generative adversarial networks to enrich residual data information, a data-driven fault diagnosis method is constructed, avoiding the need to obtain real fault label data.

Benefits of technology

This improved the reliability and interpretability of fault diagnosis for heat pump systems, reduced resource consumption, and enhanced the quality of operation and maintenance.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN115753160B_ABST
    Figure CN115753160B_ABST
Patent Text Reader

Abstract

A data-driven intelligent fault diagnosis method for heat pump systems belongs to the field of heat pump / air conditioning system fault diagnosis technology. It includes the following steps: 1. Acquiring physical information of the heat pump equipment; 2. Acquiring healthy operation data; 3. Dividing healthy data parameters; 4. Constructing a fault deviation vector; 5. Obtaining deviation data samples; 6. Constructing a data-driven benchmark model; 7. Constructing a generative adversarial network to generate virtual residual data samples; 8. Superimposing the deviation data samples and residual data to obtain artificially generated labeled data samples; 9. Constructing and training a convolutional neural network to achieve real-time fault diagnosis of the actual system. This invention generates artificially labeled data with directional labels by combining qualitative deviations of the fault system with a data generation strategy. The diagnostic network is trained using the generated artificially labeled data to achieve fault diagnosis, improving reliability and interpretability.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] This invention belongs to the field of heat pump / air conditioning system fault diagnosis technology, specifically relating to an intelligent fault diagnosis method for heat pump systems based on artificial data. Background Technology

[0002] Currently, the industry's operation and management strategy for heat pump systems still mainly relies on a "regular inspection + alarm emergency repair" model. This strategy is highly unpredictable and resource-intensive. Real-time monitoring based on artificial intelligence technology is significant for reducing operation and maintenance costs and improving maintenance quality. Due to the inherent complexity of heat pump systems, it is difficult to establish accurate physical models, leading to current mainstream diagnostic methods being data-driven—that is, learning diagnostic rules from data without relying on physical models. However, data-driven approaches are severely constrained by the quality and quantity of training data, making it difficult for industries lacking labeled data to widely adopt them.

[0003] For the fault diagnosis of heat pump systems, the fault characteristics are manifested in certain specific thermodynamic parameters of the system deviating from the normal range, such as excessively high / low temperature and increased / decreased pressure. For a given heat pump system, based on its thermodynamic cycle structure, it is easy to qualitatively analyze the deviating parameters and their directions. The difficulty lies in accurately and quantitatively analyzing the deviation characteristics. It is easy to describe the deviation patterns of the heat pump system with only a certain rough range of deviation characteristics. Based on this conclusion, this invention proposes to integrate qualitative deviation knowledge and express it in a data-driven form to generate artificial data with physical laws. The diagnostic network is trained using this artificial data to achieve the purpose of fault diagnosis. Among them, relying on artificial data to train a data-driven model and utilizing the adaptability of the deep model within its range to realize the diagnosis of real fault scenarios is called artificial data-driven. Summary of the Invention

[0004] To address the aforementioned problems in the existing technology, the purpose of this invention is to provide an intelligent fault diagnosis method for heat pump systems based on artificial data, which has high reliability and interpretability.

[0005] This invention provides the following technical solution:

[0006] A data-driven intelligent fault diagnosis method for heat pump systems generates artificially labeled data with directional tags by combining qualitative deviations of the fault system with a data generation strategy. The generated artificially labeled data is then used to train a diagnostic network to achieve fault diagnosis. The method includes the following steps:

[0007] Step 1: Obtain physical information of heat pump equipment, specifically including physical information such as the circulation structure of the heat pump system, the number and distribution of sensors, and the working fluid used in the system.

[0008] Step 2: Acquisition of health operation data: For the heat pump object to be diagnosed, select a healthy heat pump system of the same model and collect its operation data under healthy conditions; preferably, the data sample needs to be more than 20,000, and the sampling interval is between 15 seconds and 3 minutes, depending on the specific experimental conditions; the experimental heat pump can be set up separately, or a newly deployed heat pump can be selected on site, or it can be set up using simulation software.

[0009] The third step is to divide the collected health data parameters into two categories: independent parameters and dependent parameters. Independent parameters are those that are not affected by factors such as operating conditions but are only affected by human control factors, such as ambient temperature, valve opening, and fan speed. Dependent parameters are those that are affected by independent parameters, such as compressor exhaust pressure, subcooling temperature, and superheat. Dependent parameters must include fault-sensitive characteristic parameters.

[0010] Step 4: Construct a fault deviation vector based on the physical information of the heat pump system. The deviation vector contains the changing trends and magnitudes of all dependent parameters under specific fault categories. The deviation vector includes all dependent parameters of the heat pump system, and these parameters contain the required sensitive features for all faults. The construction process of the deviation vector is as follows: Through prior domain knowledge, the deviation direction and magnitude of the sensitive features under the corresponding fault can be obtained. Based on the proportional relationship between the deviation amount and the health value, the deviation magnitude is classified into levels, ultimately resulting in a one-dimensional vector with a length equal to the dependent parameter data and a value representing the deviation level. Preferably, the process of classifying the deviation magnitude levels is as follows:

[0011] The deviation is divided into four levels, from 1 to 4. Level 1 is below 10%, Level 2 is 10% to 20%, Level 3 is 20% to 30%, and Level 4 is more than 30%.

[0012] Step 5: Using a random scaling strategy, assign values ​​to the deviation vector and scale it up or down proportionally according to the selected scaling factor range to obtain a data-represented deviation data sample. The random scaling strategy is a strategy for randomly generating scaling factors, and the process is as follows:

[0013] First, the deviation vector is assigned a value according to the deviation level. Preferably, levels 1-4 are assigned values ​​of 0.1-0.4 respectively. Then, a limited scaling range is selected according to the specific heat pump system application scenario, preferably 0.6-3. A scaling factor is randomly generated within this range, and the factor is multiplied with the assigned deviation vector to obtain a data sample of the deviation amplitude fluctuating within a certain range, simulating the changes in deviation characteristics under different operating conditions and environments.

[0014] Step 6: Construct a data-driven benchmark model using healthy heat pump operation data; the benchmark model is built using a deep neural network, with the main body being a recurrent neural network and the feature extraction part being a convolutional network, and the two networks are connected using an encoder and decoder.

[0015] The model takes independent parameters as input and dependent parameters as output, obtaining the model's predicted dependent values ​​for the corresponding independent parameters. The difference between these predicted and measured dependent values ​​of the health system is then calculated to obtain the residual data for the health system. Theoretically, this residual data represents the fitting deviation of the benchmark model relative to the real health system. Before training the model, the running data is normalized so that all data are within the range of -1 to 1.

[0016] Step 7: Construct a Generative Adversarial Network (GAN) to learn the distribution of residual data and generate virtual residual data samples to enrich the residual data information. The purpose of training the GAN is to enrich the information content of the artificial data and fit the self-error of the benchmark model. The training process requires segmenting the residual data according to time periods. Preferably, the time step is consistent with the number of dependent parameters. For example, if the residual contains 15 dependent parameters, the step size is also selected as 15, forming a 15x15 matrix data.

[0017] Step 8: Overlay the deviation data samples generated in Step 5 and the residual data generated in Step 7 to obtain artificially generated labeled data samples. The labels of the samples are consistent with the labels of the deviation data samples.

[0018] Step 9: Construct a convolutional neural network, train the convolutional network with artificial data using supervised learning methods to obtain a fault classifier, and deploy the classifier to achieve real-time fault diagnosis of the actual system.

[0019] By employing the above-described technology, the beneficial effects of the present invention compared to the prior art are as follows:

[0020] This invention generates artificially labeled data with directional tags by combining qualitative deviations from fault systems with a data generation strategy. The diagnostic network is then trained using this artificially labeled data to achieve fault diagnosis. This fully leverages the advantages of data-driven and supervised learning algorithms in artificial intelligence, while avoiding the need to acquire fault-labeled data, completely eliminating the huge human and material resources required in engineering practice to obtain a large number of fault-labeled samples. At the same time, the artificial data-driven method transforms fault mechanism knowledge into data samples, using artificial samples instead of real samples. Compared with methods that rely entirely on model self-supervised training, this method has stronger reliability and interpretability. Attached Figure Description

[0021] Figure 1 This is a flowchart of the artificial data-driven diagnostic process of the present invention;

[0022] Figure 2 This is a schematic diagram of the data random scaling strategy of the present invention;

[0023] Figure 3 This is a structural diagram of the reference model of the heat pump system of the present invention;

[0024] Figure 4 This is a schematic diagram of the target system in an embodiment of the present invention. Detailed Implementation

[0025] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention.

[0026] Conversely, this invention encompasses any substitutions, modifications, equivalent methods, and solutions made within the spirit and scope of the invention as defined in the claims. Furthermore, to provide a better understanding of the invention, certain specific details are described in detail below. However, those skilled in the art will fully understand the invention even without these detailed descriptions.

[0027] Refer to the instruction manual. Figure 1-4 Using the RP-1043 dataset from ASHRAE in the United States as an example, the specific implementation of the method of the present invention will be further explained.

[0028] ASHRAE RP-1043 is a project initiated by the Society of Heating, Refrigerating and Air-Conditioning Engineers (SHEAE) in 1999. Its full name is Fault Detection and Diagnostic (FDD) Requirements and Evaluation Tools for Chillers. The project aims to establish a dataset for research and evaluation of fault diagnosis methods in refrigeration systems. This dataset includes transient and steady-state operating data of refrigeration units under different loads and fault conditions. The test system for the project is a 90-ton centrifugal chiller unit installed in a 70°F constant temperature chamber. The project includes six fault states and one normal state, with four different severity levels assigned to each fault state, and operating data is recorded for each. The specific implementation steps are as follows:

[0029] Step 1) The physical information required for the target system can be obtained from the project report. The system is a water-cooled heat exchange system that simulates the actual load through self-exchange and external water sources.

[0030] Step 2) The system sampling interval is 10 seconds, and a total of 20,000 healthy operating samples are selected. The faulty system and the healthy system in this project are the same chiller unit, and the fault simulation is achieved by artificially adding faults.

[0031] Step 3) Based on the structural principle of the test system, select 17 independent parameters, including: evaporator inlet water temperature; condenser inlet water temperature; building inlet water temperature; building outlet water temperature; compressor power; condenser water flow rate; evaporator water flow rate; small steam valve opening; large steam valve opening; three-way valve opening; condensate circulating water valve opening; evaporator circulating water valve opening; external water valve opening; external inlet water temperature; external outlet water temperature; hot water inlet water temperature; hot water outlet water temperature. Select 7 dependent parameters, including: evaporator inlet and outlet water temperature difference TEI-TEO, condenser inlet and outlet water temperature difference TCO-TCI, evaporator pressure PRE, condenser pressure PRC, subcooling TRCsub, suction superheat Tshsuc, and exhaust superheat Tshdis.

[0032] Step 4) There are 6 types of faults in the dataset, as shown in Table 1. Based on the characteristics of the 6 faults, select the sensitive characteristics and analyze their changing trends to construct the deviation vector, as shown in Table 2. In the table, the direction of the arrows indicates the direction of deviation, upward arrows indicate an increase, the number of arrows indicates the level of change, and parameters that do not deviate are represented by the origin.

[0033] Step 5) Observation of random scaling strategy Figure 2 First, the deviation vectors need to be assigned values. Based on the deviation direction and magnitude in Table 2, values ​​of 0.1-0.4 are assigned according to the deviation level, and 0 is assigned to non-deviation vectors, thus converting the deviation vectors into deviation samples. The generation range of the random scaling factor is selected as [0.6, 3]. Within this range, a scaling factor is randomly generated and multiplied by the deviation sample to obtain a set of artificial deviation data. Different numbers of deviation data are generated according to the training requirements of the diagnostic model. In this case, 2000 samples are generated for each type of fault.

[0034] Table 1. Detailed Description of Case Fault Categories

[0035]

[0036] Table 2 Multidimensional Deviation Vector Table of Faults

[0037]

[0038] Step 6) The baseline model structure of the heat pump system is as follows: Figure 3 As shown, health data is used as the independent variable for input and the dependent variable for output to train a baseline model, achieving a fit to the thermodynamic characteristics of the heat pump system. Before retraining, the operating data is first normalized, according to... Normalization is achieved by using the mean and three times the variance.

[0039] Step 7) Generative Adversarial Network (GAN) uses a convolutional neural network as the backbone network and adopts the WGAN structure. The input layer is a vector of length 100, and the output is a 7x7 residual data sample. The GAN is trained with 5,000 sets of real residuals for a total of 100,000 training rounds.

[0040] Step 8) The artificially labeled data generation process is fully automated. A scaling factor is generated using a random scaling strategy to obtain deviation samples. Simultaneously, GAN is used to generate residual data. The two are superimposed to obtain a set of artificial data samples. According to training needs, 2000 sets of samples are generated for each fault and health condition.

[0041] Step 9) Train the convolutional diagnostic model shown in Table 3 using manually labeled data, using cross-entropy as the loss function, and train for 60 rounds. During diagnosis, first obtain the residual data of the target system using the baseline model, then divide the residual data into 7-step time intervals, input them into the convolutional network, and obtain the diagnostic results.

[0042] Table 3. Structure of Convolutional Neural Network Diagnostic Model

[0043]

[0044] Following the steps described in the above case, data-driven fault diagnosis can be achieved without relying on labeled data. The diagnostic model used can be deployed on a cloud system or a local system for real-time fault monitoring.

[0045] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A method for intelligent fault diagnosis of heat pump systems based on artificial data, characterized in that: By combining qualitative deviations from faulty systems with data generation strategies, artificially labeled data with directional tags is generated. The diagnostic network is then trained using the generated artificially labeled data to achieve fault diagnosis. It includes the following steps: Step 1: Obtaining physical information of the heat pump equipment; Step 2: Acquiring health operation data; Step 3: Divide the collected health data parameters into two categories: independent parameters and dependent parameters; Step 4: Construct a fault deviation vector based on the physical information of the heat pump system; the deviation vector contains all dependent parameters of the heat pump system, and the dependent parameters contain all the required sensitive features of the fault; the construction process of the deviation vector is as follows: through prior domain knowledge, the deviation direction and magnitude of the sensitive features under the corresponding fault can be obtained; according to the proportional relationship between the deviation amount and the health value itself, the deviation magnitude is divided into levels, and finally a one-dimensional vector with a length equal to the dependent parameter data and a value equal to the deviation level is obtained. Step 5: Using a random scaling strategy, assign values ​​to the deviation vector and scale it up or down proportionally according to the range of scaling factors to obtain a data-represented deviation data sample. Step 6: Build a data-driven benchmark model using healthy heat pump operating data; Step 7: Construct a Generative Adversarial Network (GAN) to learn the distribution of residual data and generate virtual residual data samples to enrich the residual data information; Step 8: Overlay the deviation data samples generated in Step 5 and the residual data generated in Step 7 to obtain artificially generated labeled data samples. The labels of the samples are consistent with the labels of the deviation data samples. Step 9: Construct a convolutional neural network, train the convolutional network with artificial data using supervised learning methods to obtain a fault classifier, and deploy the classifier to achieve real-time fault diagnosis of the actual system.

2. The intelligent fault diagnosis method for a heat pump system based on artificial data as described in claim 1, characterized in that... The independent parameter is a parameter that is not affected by operating conditions but only by human control factors, while the dependent parameter is a parameter that is affected by the independent parameter.

3. The intelligent fault diagnosis method for a heat pump system based on artificial data as described in claim 2, characterized in that... The random scaling strategy in step five is a scaling factor random generation strategy, and the process is as follows: First, the deviation vector is assigned a value according to the deviation level. Then, a limited scaling range is selected according to the specific heat pump system application scenario. Within this range, a scaling factor is randomly generated. This factor is multiplied with the assigned deviation vector to obtain a data sample in which the deviation amplitude fluctuates within a certain range, simulating the changes in deviation characteristics under different operating conditions and environments.

4. The intelligent fault diagnosis method for a heat pump system based on artificial data as described in claim 3, characterized in that... In the sixth step, the baseline model is built using a deep neural network, with the main body being a recurrent neural network and the feature extraction part being a convolutional network. The two networks are connected using an encoder-decoder.

5. The intelligent fault diagnosis method for a heat pump system based on artificial data as described in claim 4, characterized in that... The driving benchmark model takes the independent variable parameters as input and the dependent variable parameters as output to obtain the dependent variable value predicted by the model under the corresponding independent variable parameters. The difference between the model prediction and the measured dependent variable value of the health system is used to obtain the residual data of the health system. Theoretically, this data is the fitting deviation of the benchmark model relative to the real health system.

6. The intelligent fault diagnosis method for a heat pump system based on artificial data as described in claim 5, characterized in that... In the seventh step, the purpose of training the generative adversarial network is to enrich the information content of the artificial data and fit the self-error of the benchmark model; the training process requires segmenting the residual data according to time periods.

Citation Information

Patent Citations

  • Fault diagnosis method for centrifugal water chilling unit

    CN110986407A

  • Fault detection method and system based on generative adversarial network and computer program

    US20200202221A1