A sensor self-calibration method for variable air volume air conditioning system based on autoencoder

Abstract

The present invention discloses a variable air volume air conditioning system sensor self-calibration method based on an autoencoder, comprising the following steps: S1, clustering measurement values ​​of different operating conditions of the variable air volume air conditioning system based on a Gaussian mixture model; S2, using the autoencoder to establish its own sensor model based on the internal relationship between its own measurement values; S3, constructing the functions required in the sensor calibration process; S4, introducing the distance function driven by the autoencoder into the Bayesian theorem, and obtaining the sensor offset size through random sampling. The present invention adopts the method of dividing the working conditions by Gaussian mixture clustering and establishing the sensor self-model by the autoencoder, and proposes a variable air volume air conditioning system sensor self-calibration method based on the autoencoder, calibrating the sensors involved in the basic control of the variable air volume air conditioning system, increasing the accuracy of the sensor measurement values, ensuring the normal adjustment of the control loop, and enhancing the applicability of the sensor calibration method.

Description

Technical Field

[0001] The present invention relates to a sensor self-calibration method, in particular to a variable air volume air conditioning system sensor self-calibration method based on an autoencoder, and belongs to the technical field of underlying sensor data processing of building energy consumption monitoring systems. Background Art

[0002] With economic development and rising living standards, people's requirements for indoor air quality are becoming increasingly stringent, and the application of air conditioning systems in buildings is becoming increasingly widespread. Variable air volume (VAV) air conditioning systems have been widely adopted in many countries and regions due to their advantages, including high energy efficiency, excellent thermal comfort, flexible control, and ease of expansion. VAV systems operate by varying the air volume delivered to each room to balance load variations within the conditioned area, enabling comprehensive control of a building's thermal and humidity environment and air quality. To further enhance energy efficiency, various intelligent control strategies, such as supply air temperature control, static pressure control, and differential pressure control, are being widely applied to VAV systems. However, the proper operation of these control loops relies on the accuracy of the measurements from temperature, pressure, and differential pressure sensors. Failures in these sensors not only increase energy consumption for the pumps and fans but can even cause the entire control system to crash. Automatic control of VAV systems is crucial for achieving energy-efficient operation, and this automation relies on data from various sensors. When these sensors fail, simply replacing them would disrupt the system's operation and be time-consuming and labor-intensive due to their complex installation environment and their constant involvement in the control process. Therefore, it is imperative to find a reasonable and effective method to automatically diagnose and repair sensors. In recent years, numerous experts and scholars have employed various methods to diagnose air conditioning system sensor failures. Existing research focuses on three main approaches for sensor fault detection and diagnosis: analysis-based, knowledge-based, and data-driven. While these methods can diagnose and calibrate sensor failures to a certain extent, their application in practical air conditioning systems still faces several challenges: 1) Sensors operate under varying operating conditions, potentially resulting in varying offsets. Direct calibration without distinguishing these differences will not guarantee accurate calibration results; 2) Due to cost and installation location constraints, only some sensors involved in system control are typically installed, while monitoring sensors are insufficiently installed. This makes it impossible to construct sensor models using previous methods, resulting in unsuccessful calibration; 3) Model construction during the calibration process requires significant expertise and is relatively difficult to implement. These issues will pose challenges and difficulties in future practical applications. Summary of the Invention

[0003] In response to the above problems, the present invention adopts a method of dividing working conditions by Gaussian mixture clustering and establishing a sensor self-model by autoencoder, and provides a method for sensor fault self-calibration that can effectively improve the data quality of variable air volume air conditioning systems, thereby improving the applicability of sensor calibration methods.

[0004] The technical solutions of the present invention are as follows:

[0005] A self-calibration method for variable air volume air conditioning system sensors based on an autoencoder, the steps are as follows:

[0006] S1. Cluster the measured values ​​of different operating conditions of the variable air volume air conditioning system based on the Gaussian mixture model. The specific steps are as follows:

[0007] S1.1. Analyze and study the variable air volume air conditioning system to obtain the basic control loop of the sensor: supply air temperature control, supply air static pressure control and chilled water pump pressure difference control. Therefore, it is determined that the supply air temperature sensor T s , air supply static pressure sensor P s Calibrate the differential pressure sensor DP before and after the chilled water pump;

[0008] S1.2. For the real-time data obtained by the variable air volume air conditioning monitoring system, outliers and transient measurement values ​​are eliminated to ensure that the measurement data is in a steady-state condition;

[0009] S1.3. Based on the operating characteristics of the variable air volume air conditioning system, indoor and outdoor temperature and humidity are selected as clustering input parameters. Gaussian mixture clustering is performed on the measured values ​​under the steady-state condition in S1.2 to ensure that the measured values ​​with the same offset are in the same cluster, so that the calibration results converge to the same value.

[0010] S1.4. Measurement value T under the same working condition s 、P s , DP is divided into training data set, test data set and calibration data set;

[0011] S2. Use the autoencoder to build its own sensor model through the internal relationship between its own measurement values. The specific steps are as follows:

[0012] S2.1. For the training dataset of the sensors divided in S1.4, determine the optimal parameters of the autoencoder model using a step-by-step manual testing method;

[0013] S2.2, according to the optimal parameters, T s 、P s The sensor model based on the autoencoder is trained by taking the three variables DP as input parameters, as shown in formulas (1) and (2);

[0014] Z=g1(b1+W1X) (1)

[0015] X * =g2(b2+W2Z) (2)

[0016] Among them, Z is the latent variable, g is the nonlinear transfer function, W is the weight, b is the bias, X is the input variable, X * is the output variable;

[0017] S2.3. Substitute the sensor test data set into the sensor model obtained in S2.2. Generate a new data set based on the model. Determine the accuracy of the sensor model based on the reconstruction error between the test data and the new data set. If the reconstruction error is less than the set threshold, the established sensor model meets the requirements and can be used for further calibration. Otherwise, discard all training data sets divided in S1.4 and select a new data set for training.

[0018] S2.4. The definition of the sensor error threshold adopts the 3σ principle, and the threshold is calculated as shown in the following equations (3)-(5):

[0019]

[0020]

[0021]

[0022] Where Th is the threshold of reconstruction error, Reconstruct the mean error for the training data, σ R The standard deviation of the reconstruction error of the training data, x j is the reconstruction error of the jth training data;

[0023] S3. Construction of the functions required in the sensor calibration process. The specific steps are as follows:

[0024] S3.1. Based on the sensor calibration data set T s 、P s , DP and their corresponding offsets, and establish the sensor calibration function based on the basic relationships of addition, subtraction, multiplication and division, as shown in formula (6). Before calibration, the specific offset is unknown, so it is represented by the variable x, which will be solved later.

[0025] Calibration function

[0026] Among them, T c,s ,P c,s ,DP c is the calibration function of the sensor to be calibrated, T s ,P s,DP is the measurement value of each type of sensor, x DP is the offset of each sensor, and f is the calibration function model.

[0027] S3.2. Bring the calibration function in S3.1 into the sensor model based on the autoencoder established in S2 to construct the reference function required for calibration, as shown in formula (7):

[0028] Benchmark Function

[0029] in They are T in the sensor model based on the autoencoder. s 、P s , the benchmark function corresponding to DP, AE is the autoencoder model;

[0030] S3.3. Subtract the calibration function constructed in S3.1 and S3.2 from the reference function to establish the distance function required for calibration, as shown in formula (8):

[0031]

[0032] Among them D AE (X) is the distance function driven by the autoencoder, and N is the number of calibration datasets;

[0033] S4. Substitute the distance function driven by the autoencoder obtained in S3 into the Bayesian theorem and obtain the sensor offset size through random sampling. The specific steps are as follows:

[0034] S4.1, multiple calibration data sets T of the same operating conditions in S1 s 、P s , DP is brought into the distance function in S3;

[0035] S4.2. Substitute the distance function obtained in S4.1 into the Gaussian distribution to form the likelihood function required for sensor calibration, as shown in Equation (9):

[0036]

[0037] Where P(Y|X) is the likelihood function of the offset, σ is the covariance matrix;

[0038] S4.3. For Bayesian theorem, based on the likelihood function and prior distribution, the sensor measurement data T is obtained by random sampling using Monte Carlo Markov chain. s 、P s , the posterior distribution of the DP offset is shown in formula (10); in addition, the prior distribution of the target sensor is obtained based on the empirical knowledge before sensor calibration;

[0039] P(X|Y)∝P(Y|X)×π(X)(10)

[0040] Among them, P(X|Y) is the sensor measurement data T s 、P s , the posterior distribution of DP, π(X) is the prior distribution of the offset;

[0041] S4.4. Obtain sensor measurement data T through S4.1-S4.3 s 、P s , the posterior distribution mean of DP, which represents the offset size after sensor calibration; compensating the offset to the original measurement realizes the supply air temperature sensor T s , air supply static pressure sensor P s And calibration of the differential pressure sensor DP before and after the chilled water pump.

[0042] Beneficial effects of the present invention: The present invention provides a method for self-calibration of sensor faults that can effectively improve the data quality of a variable air volume air conditioning system, calibrates sensors involved in the basic control of the variable air volume air conditioning system, increases the accuracy of sensor measurement values, ensures the normal adjustment of the control loop, and at the same time enhances the applicability of the sensor calibration method. BRIEF DESCRIPTION OF THE DRAWINGS

[0043] Figure 1 This is a schematic diagram of the sensor self-calibration of the variable air volume air conditioning system;

[0044] Figure 2 This is a basic control diagram of a variable air volume air conditioning system. DETAILED DESCRIPTION

[0045] The specific embodiments of the present invention are described in detail below in conjunction with the summary of the invention and the accompanying drawings.

[0046] Basic control of variable air volume air conditioning system Figure 2 As shown, refer to Figure 1 The present invention provides a variable air volume air conditioning system sensor self-calibration method based on an autoencoder, which comprises the following steps:

[0047] S1. Cluster the measured values ​​of different operating conditions of the variable air volume air conditioning system based on the Gaussian mixture model. The specific steps are as follows:

[0048] S1.1. Analyze and study the variable air volume air conditioning system to obtain the basic control loop of the sensor: supply air temperature control, supply air static pressure control and chilled water pump pressure difference control. Therefore, it is determined that the supply air temperature sensor T s , air supply static pressure sensor P s Calibrate the differential pressure sensor DP before and after the chilled water pump;

[0049] S1.2. For the real-time data obtained by the variable air volume air conditioning monitoring system, outliers and transient measurement values ​​are eliminated to ensure that the measurement data is in a steady-state condition;

[0050] S1.3. Based on the operating characteristics of the variable air volume air conditioning system, indoor and outdoor temperature and humidity are selected as clustering input parameters. Gaussian mixture clustering is performed on the measured values ​​under the steady-state condition in S1.2 to ensure that the measured values ​​with the same offset are in the same cluster, so that the calibration results converge to the same value.

[0051] S1.4. Measurement value T under the same working condition s 、P s , DP is divided into, 70% of the measured values ​​are training data set, 20% are test data set, and 10% are calibration data set;

[0052] S2. Use the autoencoder to build its own sensor model through the internal relationship between its own measurement values. The specific steps are as follows:

[0053] S2.1. For the training dataset of the sensors divided in S1.4, determine the optimal parameters of the autoencoder model using a step-by-step manual testing method;

[0054] S2.2. Based on the optimal parameters, these three variables are used as input parameters to train the sensor model based on the autoencoder, as shown in equations (1) and (2);

[0055] Z=g1(b1+W1X) (1)

[0056] X * =g2(b2+W2Z) (2)

[0057] Among them, Z is the latent variable, g is the nonlinear transfer function, W is the weight, b is the bias, X is the input variable, X * is the output variable;

[0058] S2.3. The sensor test data set is introduced into the sensor model. A similar new data set is generated based on the model. The accuracy of the sensor model is determined based on the error between the test data and the similar new data set. If the reconstruction error is less than a pre-set threshold, the sensor model meets the requirements and can be used for further calibration. Otherwise, all training data sets divided in S1.4 are discarded and a new data set is selected for training.

[0059] S2.4. The definition of the sensor error threshold adopts the 3σ principle, and the threshold is calculated as shown in the following equations (3)-(5):

[0060]

[0061]

[0062]

[0063] Where Th is the threshold of reconstruction error, Reconstruct the mean error for the training data, σ R The standard deviation of the reconstruction error of the training data, x j is the reconstruction error of the jth training data;

[0064] S3. Construction of the functions required in the sensor calibration process. The specific steps are as follows:

[0065] S3.1. Based on the sensor calibration data set T s 、P s , DP and offset, and establish the sensor calibration function based on the basic relationship of addition, subtraction, multiplication and division, as shown in equations (6)-(8). Before calibration, the specific offset is unknown, and it is represented by the variable x, which will be solved later.

[0066]

[0067]

[0068] DP c =DP+x DP (8)

[0069] Among them, T c,s ,P c,s ,DP c is the calibration function of the sensor to be calibrated, T s ,P s ,DP is the measurement value of each type of sensor, x DP is the offset constant of each sensor, and f is the calibration function model.

[0070] S3.2. Bring the calibration function into the sensor model based on the autoencoder established in S2 to construct the reference function required for calibration, as shown in formula (9):

[0071]

[0072] in They are T in the sensor model based on the autoencoder. s 、P s , the benchmark function corresponding to DP, AE is the autoencoder model;

[0073] S3.3. Subtract the calibration function constructed in S3.1 and S3.2 from the reference function to establish the distance function required for calibration, as shown in formula (10):

[0074]

[0075] Among them D AE (X) is the distance function driven by the autoencoder, and N is the number of calibration datasets;

[0076] S4. Substitute the distance function driven by the autoencoder obtained in S3 into the Bayesian theorem and obtain the sensor offset size through random sampling. The specific steps are as follows:

[0077] S4.1, multiple calibration data sets T of the same operating conditions in S3 s 、P s , DP is brought into the distance function;

[0078] S4.2. Substitute the distance function into the Gaussian distribution to form the likelihood function required for sensor calibration, as shown in formula (11):

[0079]

[0080] Where P(Y|X) is the likelihood function of the offset, σ is the covariance matrix;

[0081] S4.3. For Bayesian theorem, based on the likelihood function and prior distribution, the sensor measurement data T is obtained by random sampling using Monte Carlo Markov chain. s 、P s , the posterior distribution of the DP offset is shown in Equation (12); In addition, the prior distribution of the target sensor is obtained based on the empirical knowledge before sensor calibration;

[0082] P(X|Y)∝P(Y|X)×π(X) (12)

[0083] Where P(X|Y) is the sensor measurement data T s 、P s , the posterior distribution of DP, π(X) is the prior distribution of the offset;

[0084] S4.4. Obtain sensor measurement data T through S4.1-S4.3 s 、P s , the posterior distribution mean of DP, which represents the offset size after sensor calibration; compensating the offset to the original measurement realizes the supply air temperature sensor T s , air supply static pressure sensor P sThe calibration results of this embodiment are shown in Table 1. Based on the autoencoder, different types of sensors establish their own models for calibration. After calibration, the accuracy of these sensors involved in basic control is improved by more than 95.69%, ensuring the accuracy of the sensor measurements and maintaining normal regulation of the control loop.

[0085] Table 1 Sensor calibration results

[0086] sensor True value Measurements Offset Calibration value <![CDATA[T s ]]> 39.15 37.15 2.09 39.24 <![CDATA[P s ]]> 45.27 40.27 4.97 45.24 <![CDATA[D P ]]> 105.17 97.17 8.27 105.44

Claims

1. A self-calibration method for variable air volume air conditioning system sensors based on autoencoder, characterized in that: Here are the steps: S1. Cluster the measured values ​​of different operating conditions of the variable air volume air conditioning system based on the Gaussian mixture model. The specific steps are as follows: S1.

1. Analyze and study the variable air volume air conditioning system to obtain the basic control loop of the sensor: supply air temperature control, supply air static pressure control and chilled water pump pressure difference control. Therefore, it is determined that the supply air temperature sensor T s , air supply static pressure sensor P s Calibrate the differential pressure sensor DP before and after the chilled water pump; S1.

2. For the real-time data obtained by the variable air volume air conditioning monitoring system, outliers and transient measurement values ​​are eliminated to ensure that the measurement data is in a steady-state condition; S1.

3. Based on the operating characteristics of the variable air volume air conditioning system, indoor and outdoor temperature and humidity are selected as clustering input parameters. Gaussian mixture clustering is performed on the measured values ​​under the steady-state condition in S1.2 to ensure that the measured values ​​with the same offset are in the same cluster, so that the calibration results converge to the same value. S1.

4. Measurement value T under the same working condition s 、P s , DP is divided into training data set, test data set and calibration data set; S2. Use the autoencoder to build its own sensor model through the internal relationship between its own measurement values. The specific steps are as follows: S2.

1. For the training dataset of the sensors divided in S1.4, determine the optimal parameters of the autoencoder model using a step-by-step manual testing method; S2.2, according to the optimal parameters, T s 、P s The sensor model based on the autoencoder is trained by taking the three variables DP as input parameters, as shown in formulas (1) and (2); Z=g1(b1+W1X) (1) X * =g2(b2+W2Z) (2) Among them, Z is the latent variable, g is the nonlinear transfer function, W is the weight, b is the bias, X is the input variable, X * is the output variable; S2.

3. Substitute the sensor test data set into the sensor model obtained in S2.

2. Generate a new data set based on the model. Determine the accuracy of the sensor model based on the reconstruction error between the test data and the new data set. If the reconstruction error is less than the set threshold, the established sensor model meets the requirements and can be used for further calibration. Otherwise, discard all training data sets divided in S1.4 and select a new data set for training. S2.

4. The definition of the sensor error threshold adopts the 3σ principle, and the threshold is calculated as shown in the following equations (3)-(5): Where Th is the threshold of reconstruction error, Reconstruct the mean error for the training data, σ R The standard deviation of the reconstruction error of the training data, x j is the reconstruction error of the jth training data; S3. Construction of the functions required in the sensor calibration process. The specific steps are as follows: S3.

1. Based on the sensor calibration data set T s 、P s , DP and their corresponding offsets, and establish the sensor calibration function based on the basic relationships of addition, subtraction, multiplication and division, as shown in formula (6); Calibration function Among them, T c，s , P c，s , DP c is the calibration function of the sensor to be calibrated, T s , P s , DP is the measurement value of each type of sensor, is the offset of each sensor, and f is the calibration function model; S3.

2. Bring the calibration function in S3.1 into the sensor model based on the autoencoder established in S2 to construct the reference function required for calibration, as shown in formula (7): Benchmark Function in They are T in the sensor model based on the autoencoder. s 、P s , the benchmark function corresponding to DP, T c，s , P c，s , DP c is the corresponding calibration function, AE is the autoencoder model; S3.

3. Subtract the calibration function constructed in S3.1 and S3.2 from the reference function to establish the distance function required for calibration, as shown in formula (8): Among them D AE (X) is the distance function driven by the autoencoder, and N is the number of calibration datasets; S4. Substitute the distance function driven by the autoencoder obtained in S3 into the Bayesian theorem and obtain the sensor offset size through random sampling. The specific steps are as follows: S4.1, multiple calibration data sets T of the same operating conditions in S1 s 、P s , DP is brought into the distance function in S3; S4.

2. Substitute the distance function obtained in S4.1 into the Gaussian distribution to form the likelihood function required for sensor calibration, as shown in Equation (9): Where P(Y|X) is the likelihood function of the offset, σ is the covariance matrix; S4.

3. For Bayesian theorem, based on the likelihood function and prior distribution, the sensor measurement data T is obtained by random sampling using Monte Carlo Markov chain. s 、P s , the posterior distribution of the DP offset is shown in formula (10); in addition, the prior distribution of the target sensor is obtained based on the empirical knowledge before sensor calibration; P(X|Y)∝P(Y|X)×π(X) (10) Among them, P(X|Y) is the sensor measurement data T s 、P s , the posterior distribution of DP, π(X) is the prior distribution of the offset; S4.

4. Obtain sensor measurement data T through S4.1-S4.3 s 、P s , the posterior distribution mean of DP, which represents the offset size after sensor calibration; compensating the offset to the original measurement realizes the supply air temperature sensor T s , air supply static pressure sensor P s And calibration of the differential pressure sensor DP before and after the chilled water pump.

Images