Sensor fault online intelligent self-correction method for building hvac system

Abstract

The application discloses an online intelligent self-correction method for sensor faults of a building heating, ventilation and air conditioning system, and belongs to the technical field of intelligent operation and maintenance of building energy systems. The method learns the constraint relationship among multiple parameters of the heating, ventilation and air conditioning system by using the feature compression capability of a self-encoder, and reconstructs system operation data by using high-density constraint relationship. In the reconstruction process, the sensor faults are corrected. The model input of the self-encoder is original operation data, and the output is corrected data. The data automatically corrects the sensor faults and does not affect normal parameter values. Meanwhile, an artificial data-driven model training strategy is proposed. Sensor fault samples are randomly generated by using a Monte Carlo algorithm for model training, and the dependence on labeled samples is completely eliminated. The method can be deployed on a cloud system or a local server, and is used for preprocessing of collected data. The processed data realizes self-correction of the sensor faults and does not affect the characteristics of the original data.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of intelligent operation and maintenance of building energy systems, and particularly relates to an online intelligent self-correction method for sensor faults of a building heating, ventilation and air conditioning (HVAC) system. BACKGROUND

[0002] Sensor faults are a common type of fault in building HVAC systems, which can seriously affect the normal health monitoring and operation control of the system, and cause a significant decline in system energy efficiency. Accurate identification of sensor faults and timely correction are of great significance to maintaining the stable and efficient operation of the HVAC system. Due to the complex structure of the HVAC system and the involvement of various heat and mass exchange processes, the number of sensors is large and related coupling is associated, so it is not easy to identify and correct sensor faults in a timely manner.

[0003] Most of the existing solutions focus on the identification and diagnosis of HVAC systems, but only fault identification without timely correction of faults cannot solve the root problem. The present application aims to ensure efficient operation of the system, skips the fault identification step, and directly corrects the data with sensor faults, ensuring that the data used for health monitoring and operation control does not deviate from the measurement. The present application can be deployed online in a cloud monitoring system to preprocess real-time collected system operation data, correct measurement parameters with sensor faults in a timely manner, and at the same time will not change the values of normal parameters. The data after preprocessing does not contain sensor faults, which maximizes the quality of subsequent control and monitoring. SUMMARY

[0004] In view of the above problems existing in the prior art, the purpose of the present application is to provide an online intelligent self-correction method for sensor faults of a building HVAC system.

[0005] The present application provides the following technical solution: an online intelligent self-correction method for sensor faults of a building HVAC system, comprising the following steps:

[0006] S1: Data acquisition: collecting operation data of the HVAC system to be corrected, ensuring that the system does not have sensor faults during data collection;

[0007] S2: Data segmentation, classification and preprocessing: processing the collected data into the required format type according to the model requirements;

[0008] S3: Constructing a sensor fault correction model based on an autoencoder;

[0009] S4: Artificial data construction and model training: simulating sensor faults to realize artificial construction of sensor fault samples;

[0010] S5: Deployment and online use of the self-correction model.

[0011] Further, the specific process of step S2 is as follows:

[0012] S21: Select and divide the collected data parameters, including control parameters (such as valve opening, press speed, etc.), state parameters (such as evaporation pressure, press exhaust temperature, supercooling degree, etc.), and environmental parameters (such as indoor dry bulb temperature, outdoor dry bulb temperature, indoor humidity, etc.). Invalid parameters such as set values, long-term constant values, and system structure information are removed, and the number of remaining valid parameters is denoted as N;

[0013] S22: Select the length of the time sequence of the system parameters for single analysis: according to the number of parameters included in the parameter set, the sequence length and the number of parameters are kept consistent as much as possible, but if the number of parameters is too small or too large, the sequence length should be ensured to be between 10-30, and the sequence length is denoted as M;

[0014] S23: Normalize the parameters for preprocessing, calculate the mean u and variance sigma of all samples, and use the formula x1 = ((x / 2)-u) / 2sigma to obtain the normalized parameters.

[0015] Further, the specific process of step S3 is as follows:

[0016] The convolutional autoencoder structure is adopted, with N x M as the input data structure, and the input data should be the original parameter sequence of the HVAC. The output structure of the autoencoder is N x M, and the output data is the corrected data. The middle hidden layer of the autoencoder is used to learn the highly compressed feature vector containing the constraint relationship, and is composed of fully connected neural unit layers. The neuron data is determined according to the size of N and M, and N and M take values in the range of 32-128.

[0017] Further, the specific process of step S4 is as follows:

[0018] S41: Set the search boundary of Monte Carlo, including all sensor parameters of the HVAC system, sensor parameters that may have faults, the maximum number of sensors that may have faults at the same time, and the possible deviation range of the sensors;

[0019] S42: Set the number of samples to be generated, which is related to the amount of original data. The generated samples should be 2-4 times the amount of original data. Random faults are generated under boundary conditions using the Monte Carlo algorithm to obtain the fault parameter set and its deviation value;

[0020] S43: Add the fault parameter deviation value to the healthy data to generate artificial samples. The samples are artificially constructed data samples containing various sensor faults;

[0021] S44: training the auto-encoder by using artificial data samples as input and corresponding health data as output to realize artificial data driving.

[0022] Further, the specific process of step S5 is as follows:

[0023] The trained self-correcting model is deployed in an online system, and after data collection, the data is preprocessed by the self-correcting model before being used for subsequent control and monitoring.

[0024] Compared with the prior art, the present application has the following advantages:

[0025] 1) The method of the present application completely utilizes data-driven mode to realize constraint learning between sensor parameters, does not need to rely on physical law analysis between HVAC system parameters, and can be better applied to large and complex energy and thermal systems.

[0026] 2) The artificial data-driven method of the present application solves the dependence on labeled samples of the traditional data-driven method, and realizes data-driven model training under the condition of no labeled sample;

[0027] 3) The self-correcting model of the present application corrects the sensor fault data without affecting the values of normal data, and when applied online, all data can be processed by the self-correcting model without worrying about the influence of normal data. BRIEF DESCRIPTION OF DRAWINGS

[0028] Figure 1 is a self-correcting model structure diagram of the present application, wherein: Fig. a is a self-correcting model structure diagram based on an auto-encoder, and Fig. b is a principle diagram of the self-correcting model;

[0029] Figure 2 is a model training flowchart of the present application;

[0030] Figure 3 is a method flowchart of the present application;

[0031] Figure 4 is a principle diagram of a target system of an embodiment of the present application;

[0032] Figure 5 is a data distribution evaluation diagram before and after self-correction of an embodiment of the present application, wherein: Fig. a is a distribution evaluation diagram based on R2, and Fig. b is a distribution evaluation diagram based on RMSE. DETAILED DESCRIPTION

[0033] 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.

[0034] 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.

[0035] Example:

[0036] This implementation case uses a chiller unit from the ASHRAE RP-1043 dataset as the target system. ASHRAE RP-1043 is a project initiated by the Society of Heating, Refrigerating and Air-Conditioning Engineers (SHEAE) in 1999, officially named "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 is a 90-ton centrifugal chiller unit installed in a 70°F constant temperature chamber. The project includes six types of thermal fault states and one normal state. The specific implementation steps are as follows:

[0037] S1: As Figure 4 The system schematic shown illustrates a 90-ton chiller unit as the target system in this case study. The health category data in this dataset can be used as the raw operational data for this implementation. The data sampling interval is 10 seconds, and there are over 15,000 data samples.

[0038] S21: There are 65 parameters in the data set, excluding invalid parameters such as set values, and selecting 33 valid parameters, including: evaporator inlet water temperature; evaporator outlet water temperature; condenser inlet water temperature; condenser outlet water temperature; HX side inlet water temperature (condensing circulating water); HX side outlet water temperature (condensing circulating water); building side inlet water temperature (evaporating circulating water); building side inlet water temperature (evaporating circulating water); compressor power; condenser water flow; evaporator water flow; evaporation temperature; evaporation pressure; condensation temperature; condensation pressure; subcooling temperature; suction temperature; suction superheat; discharge temperature; discharge superheat; small steam valve opening; large steam valve opening; hot water valve opening; three-way valve opening; condensing circulating water valve opening; evaporating 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, condensing circulating water inlet and outlet temperature difference; evaporating circulating water inlet and outlet temperature difference.

[0039] S22: Considering the model learning ability, data complexity, and computational complexity, the time sequence length M is selected as 25, and the final format of the data is a 33x25 matrix. Finally, in order to use the data for training of the convolutional neural network, the two-dimensional matrix data is converted to 33x25x1 three-dimensional data through Python language.

[0040] S23: The mean u and variance sigma of the 15000 samples of the 33 parameters are calculated respectively, and then the parameter normalization value is calculated by the formula x1 = ((x / 2)-u) / 2sigma.

[0041] S3: A convolutional autoencoder is used to construct a self-correcting model for sensor faults, with the input and output sizes of the model being 33x25x1, and the number of neurons in the intermediate feature layer being 128. The specific structure of the model is shown in Table 1.

[0042] Table 1 Network structure of self-correcting model

[0043]

[0044]

[0045] S41: The number of sensor parameters is 33, and all parameters in this case are listed as correctable parameters, i.e. any sensor failure should be corrected. At the same time, considering the actual situation that there may be multiple sensor failures in the HVAC system, artificial data construction must consider this situation, and application can ensure the correction effect of multiple sensor failures. Here, the number of sensor parameters that may exist faults is set to 33, the maximum number of simultaneous faults is 3, and the possible deviation range is -20%-20%.

[0046] S42: According to the total amount of health data, set the number of samples to be generated to 30000 groups. Use the Monte Carlo random generation algorithm to generate deviation vectors one by one, in the form as follows:

[0047] [0, 0, -0.15, 0, 0, 0, 0.08, 0, 0, 0, 0, 0, 0, 0, 0, 0.19, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0]

[0048] The deviation vector indicates that the 3rd, 7th and 16th parameters produce deviations of -15%, 8% and 19% respectively. 30000 vectors are generated in this form.

[0049] S43: Add the faulted parameter deviation value to the health data to generate artificial samples. This process randomly selects a set of health data samples and combines them with the deviation vector. The vector addition constitutes an artificially constructed data sample containing sensor faults.

[0050] S44: Use the artificially constructed data sample as input and the corresponding health data as output to train the autoencoder. This process uses a supervised training algorithm. The total number of artificial data and original data is 45000 groups, which is divided into training set, validation set and test set according to the ratio of 6:6:4. The loss function of model training is MSE, the optimization function is Adam, the number of samples per batch is 128, and the number of training rounds is 100.

[0051] S5: For the system that has realized cloud integration monitoring, the self-correcting model can be integrated into the cloud server to pre-process all collected data. For local monitoring systems that are not integrated with the cloud, the self-correcting model can be deployed to the central server to realize online monitoring.

[0052] To verify the effect of the self-correction model, 9 groups of sensor fault experiment scenarios are designed to analyze the data space distribution before and after correction. The 9 groups of experiment settings are as follows: No. 1-evaporation into water temperature +10; No. 2-condensing into water temperature +15; No. 3-evaporation pressure +5; No. 4-condensing pressure-20; No. 5-subcooling degree +7; No. 6-exhaust gas superheat +15; No. 7-condensing into water +15, condensing pressure-20; No. 8-evaporation into water +10, condensing pressure-20; No. 9-evaporation pressure +5, subcooling degree +7, exhaust gas superheat +15. The 9 groups of experiments contain the high-frequency sensor fault types, and No. 7-No. 9 belong to multiple sensor faults. Two evaluation indexes of R2 and RMSE are used for data analysis. The definitions of R2 and RMSE are shown in formula 1 and formula 2. The distribution distance between each data sample and healthy data sample is evaluated by using the above two indexes, wherein the closer the R2 index is to 1, the closer the distribution of the two data sets is, that is, the better the correction effect is, and the lower the RMSE is, the smaller the distribution difference of the two data sets is, that is, the better the correction is.

[0053]

[0054]

[0055] As shown in Figure 5 , the distribution difference of the corrected fault data and healthy data is respectively shown, and the distribution difference of the uncorrected sensor fault data is also shown. As can be seen from the figure, the distribution difference of the uncorrected sensor fault data and the healthy data is obvious, which verifies that the data characteristics are changed. After correction, whether the data contains faults or not, the distribution difference with the healthy data is very small, which indicates that the correction model of the present application can well realize the data correction of the sensor fault and will not change the space distribution characteristics of the healthy data, and has good self-correction effect.

[0056] The above only describes the preferred embodiments of the present application and is not used to limit the present application, and any modification, equivalent replacement and improvement made within the spirit and principle of the present application shall be included in the protection scope of the present application.

Claims

1. A method for online intelligent self-correction of sensor faults for building HVAC systems, characterized in that: Comprise the following steps: S1: data collection: the operation data collection of the HVAC system to be corrected, need to ensure that there is no sensor failure in the system when collecting; S2: data segmentation, classification and preprocessing: the collected data is processed into the required format type according to the model requirements; S3: constructing a sensor fault correction model based on autoencoder; the specific process of the step S3 is as follows: The convolutional autoencoder structure is adopted, and NXM is used as the input data structure. The input data should be the original parameter sequence of HVAC. The output structure of the autoencoder is NXM, and the output data is the corrected data. The middle hidden layer of the autoencoder is used to learn the highly compressed feature vector containing the constraint relationship, which is composed of fully connected neural unit layer. The neuron data is determined according to the size of N and M, and the value range of N and M is between 32 and 128; S4: artificial data construction and model training: simulate sensor failure to realize artificial construction of sensor fault samples; the specific process of the step S4 is as follows: S41: set the search boundary of Monte Carlo, including all sensor parameters of HVAC system, possible fault sensor parameters, maximum possible number of simultaneous faults of sensor and possible deviation range of sensor; S42: set the number of samples to be generated, which is related to the amount of original data. The generated sample should be 2-4 times the amount of original data; random fault generation is carried out under the boundary condition by using Monte Carlo algorithm to obtain the parameter set and its deviation value of the generated fault; S43: add the parameter deviation value of the generated fault to the healthy data to generate artificial samples. The sample is an artificially constructed data sample containing various sensor faults; S44: use the artificially constructed data sample as input and the corresponding healthy data as output to train the autoencoder and realize artificial data driving S5: deployment and online use of self-correction model.

2. The method for online intelligent self-correction of sensor faults in building HVAC systems as claimed in claim 1 wherein The specific process of the step S2 is as follows: S21: select and divide the collected data parameters, eliminate invalid parameters, and record the number of remaining valid parameters as N; S22: select the length of the time sequence of system parameters for single analysis; S23: normalize and preprocess the parameters.

3. The method for online intelligent self-correction of sensor faults in building HVAC systems as claimed in claim 1 wherein The specific process of the step S5 is as follows: Deploy the trained self-correction model to the online system. After running data collection, preprocess the data through the self-correction model first, and then use it for subsequent control and monitoring.

Images