CN113378908B

CN113378908B - Heating ventilation air conditioning system fault diagnosis method based on LightGBM algorithm and grid search algorithm

Info

Publication number: CN113378908B
Application number: CN202110626412.6A
Authority: CN
Inventors: 陆玲霞; 韩宝慧; 于淼; 季文献
Original assignee: Zhejiang University ZJU
Current assignee: Zhejiang University ZJU
Priority date: 2021-06-04
Filing date: 2021-06-04
Publication date: 2023-04-18
Anticipated expiration: 2041-06-04
Also published as: CN113378908A

Abstract

The invention discloses a heating ventilation air conditioning system fault diagnosis method based on a LightGBM algorithm and a grid search algorithm, which comprises the following steps: 1) Data acquisition and pretreatment, namely collecting fault data of an air conditioner water chilling unit, and carrying out pretreatment such as data cleaning and normalization on the data; 2) Constructing a LightGBM-based air conditioner fault diagnosis model, and determining an over-parameter to be optimized of the model and a value range of the over-parameter; 3) And (3) training and optimizing the hyper-parameters by using a grid search method and combining a five-fold cross validation method, and determining the optimal hyper-parameter combination of the model. By adopting the technology, compared with the prior art, the invention overcomes the defects of the prior art, provides the fault diagnosis method of the heating ventilation air conditioning system based on the LightGBM algorithm and the grid search algorithm, and improves the diagnosis and prediction effect of the air conditioning fault by utilizing the characteristics of rapidness and high-performance distribution of the LightGBM.

Description

Heating ventilation air conditioning system fault diagnosis method based on LightGBM algorithm and grid search algorithm

Technical Field

The invention relates to the technical field of air conditioner fault diagnosis, in particular to a heating ventilation air conditioning system fault diagnosis method based on a LightGBM algorithm and a grid search algorithm.

Background

The air conditioning system breaks down due to various factors such as equipment aging, natural wear, unreasonable design and regulation and control and the like in the operation process, and the system efficiency is reduced due to the faults, so that not only is the energy consumption waste and the carbon emission increased, but also the economic loss is caused, and the environment is damaged. Therefore, the method has important practical significance for timely detecting and accurately diagnosing the faults of the heating, ventilating and air conditioning system, can save energy, reduce shutdown, avoid irreversible damage of the air conditioning system, reduce carbon emission and reduce the time required for planning spare parts and inventory.

Most of the conventional fault diagnosis methods are based on statistics or threshold setting methods, but the threshold setting method depends on the experience of workers, so that the error reporting rate is high, and the diagnosis effect is poor. In recent years, the structure of an air conditioning system is more and more complex, the fault types are more and more diversified, and the air conditioning fault diagnosis through a machine learning and data mining method is also a current research hotspot. The gradient hoist (GBM) model is based on a decision tree algorithm, and related improved algorithms have good effects on multi-classification prediction problems, such as an extreme gradient hoist (XGboost), a light gradient hoist (LightGBM) and the like, have the advantages of high training speed, low memory occupation and the like, and are optimal algorithms for predicting medium and small-sized structure data.

However, the fault diagnosis model constructed based on the complaint method still has a large number of uncertain parameters, which causes large fluctuation of the model diagnosis accuracy. Different hyper-parameters have larger influence on the diagnosis effect of the fault diagnosis model, so that the superiority of the fault diagnosis model can be brought into play by selecting a good hyper-parameter combination, and the accuracy of fault diagnosis is improved.

Disclosure of Invention

In order to improve the accuracy of air conditioner fault diagnosis and solve the problem of poor diagnosis effect in the prior art, the invention provides a heating ventilation air conditioning system fault diagnosis method based on a LightGBM algorithm and a grid search algorithm, and the accurate prediction and diagnosis of the faults of a water chilling unit of a heating ventilation air conditioning system are realized by utilizing machine learning and data mining technologies.

The technical scheme adopted by the invention is as follows:

a fault diagnosis method for a heating, ventilating and air conditioning system based on a LightGBM algorithm and a grid search algorithm comprises the following steps:

(1) Data acquisition and pretreatment, namely collecting state data of a water chilling unit of a heating ventilation air conditioning system during normal operation and different faults, carrying out pretreatment such as data cleaning and normalization on the data, and marking fault labels to obtain training sample data;

(2) Constructing a LightGBM model as an air conditioner fault diagnosis model, and determining an over-parameter to be optimized of the model and a value range of the over-parameter; the input of the LightGBM model is preprocessed state data, and the output of the LightGBM model is a predicted value of a fault label.

(3) Training and optimizing the hyper-parameters of the LightGBM model by using the training sample data in the step (1) and combining a grid search method with a five-fold cross validation method, and determining the optimal hyper-parameter combination of the model to obtain the optimal air conditioner fault diagnosis model.

(4) And collecting the running state data of the water chilling unit of the heating, ventilating and air conditioning system in real time, carrying out preprocessing such as data cleaning and normalization, and inputting the preprocessed data into an optimal air conditioner fault diagnosis model to carry out air conditioner fault diagnosis.

Further, the state data at least includes evaporation temperature, condensation temperature, suction temperature, discharge temperature, evaporation pressure, condensation pressure, flow rate, valve position information, compressor power, and the like.

Further, in the step (1), for a missing value occurring in the data acquisition process, a lagrange interpolation method is adopted for processing, and an interpolation polynomial is as follows:

where θ represents the position of the missing value data, L _n (theta) is interpolated x, n is n data points adjacent to missing value column, k is index, x _k Taking the value of the kth data adjacent to the missing value, l _k (θ) is a basis function, and the mathematical expression is:

wherein, theta _k Representing the k-th position adjacent to theta. And removing abnormal values occurring in the data acquisition process and processing the abnormal values by adopting a Lagrange interpolation method.

Further, the determined hyper-parameters of the LightGBM model that need to be optimized include: bagging _ freq, bagging _ fraction, feature _ fraction, num _ leaves, max _ bin, leaving _ rate, max _ depth, min _ data _ in _ leaf. Wherein, the value ranges of the three parameters of the bagging _ freq, the bagging _ fraction and the feature _ fraction are all set as [0.5,1.0], and the step length is 0.1. The num _ leaves has the value range of [30,300] and the step length of 10. The max _ bin range is set to 10,50, with a step size of 5. The learning _ rate value range is set to [0.01,0.2], and the step size is 0.01. The value range of max _ depth is [3,11], and the step length is 1. The value range of min _ data _ in _ leaf is set to [20,120] and the step size is 10.

Further, the step (3) specifically includes the following sub-steps:

3.1 Determine max _ depth and num _ leaves;

3.2 Determine min _ data _ in _ leaf and max _ bin;

3.3 Determine feature _ fraction, bagging _ freq;

3.4 Reduce the learning _ rate, increase the number of iterations, train and validate the model using the quintuple cross validation method. And (5) repeating the steps (3.1) to (3.4) until the optimal hyper-parameter combination is obtained, so as to obtain the optimal air conditioner fault diagnosis model.

Preferably, the LightGBM model is implemented by adopting python language, and the fault diagnosis model of the warm air conditioning system based on the LightGBM algorithm and the grid search algorithm is implemented by using a Lightgbm machine learning library.

The invention has the beneficial technical effects that:

the invention provides a fault diagnosis method of a heating, ventilating and air conditioning system based on a LightGBM algorithm and a grid search algorithm, and compared with the prior art, the fault diagnosis method has the following advantages: aiming at the characteristics of large group data volume and multiple data dimensions of a cold water unit of the heating, ventilation and air conditioning system, the optimal hyper-parameter is found out by utilizing the distributed and efficient characteristics of LightGBM through a grid search method, and the problems of difficulty in fault diagnosis and modeling and low accuracy of the heating, ventilation and air conditioning system are well solved, so that the type of the fault occurrence is accurately judged, and support is provided for solving the fault.

Drawings

FIG. 1 is a LightGBM model framework;

fig. 2 is a flow chart based on the LightGBM algorithm and the grid search algorithm;

FIG. 3 is a LightGBM histogram;

FIG. 4 is a five-fold cross-validation graph.

Detailed Description

In order to make the technical solutions of the present invention more clear and definite for those skilled in the art, the following embodiments are combined with the accompanying drawings. The present invention is described in further detail, but the embodiments of the present invention are not limited thereto.

As shown in fig. 1 to 4, the air conditioner fault diagnosis method provided in this embodiment is based on a LightGBM algorithm of a decision tree gradient boosting method, and combines a grid search method and five-fold cross validation to find an optimal combination of hyper-parameters, thereby improving accuracy and stability of the model.

In this embodiment, the LightGBM model is implemented by using python language, the air conditioner fault diagnosis model based on the LightGBM algorithm and the grid search algorithm is implemented by using a LightGBM machine learning library, and the trained model hardware environment is intel corei5 and nvidiageferenrtx 2080Ti.

Further, in this embodiment, as shown in fig. 1 to 4, the method for diagnosing faults of an hvac system based on the LightGBM algorithm and the grid search algorithm includes the following steps:

the method comprises the following steps of (1) data acquisition and pretreatment, namely collecting fault data of a water chilling unit of the heating, ventilating and air conditioning system, carrying out pretreatment such as data cleaning and normalization on the data, and dividing the pretreated data into a training set, a verification set and a test set, wherein the method specifically comprises the following steps:

1) The fault data of the water chilling unit in a 1043-RP project of the American society of heating, refrigeration and air conditioning engineers (ASHRAE) is selected, the project is that under the laboratory condition, the running condition of the unit is changed, the unit is made to run under different working conditions and fault conditions, corresponding data is collected, and relatively authoritative data information is provided for fault detection diagnosis and performance research of the water chilling unit.

The ASHRAE-1043-RP data acquisition interval is 10s, and 64 parameters are acquired in total. Wherein, there are 48 parameters directly gathered through the sensor, including 29 temperature parameters, 7 valve position parameters, 5 pressure parameters, 2 flow parameters, 1 power parameter, 1 current parameter etc. and the rest 16 parameters are calculated in real time by means of VisSim software.

In this embodiment, 7 typical failures and normal data packets of the chiller are selected, each failure containing 4 severity levels, as shown in table 1.

TABLE 1 ASHRAE 1043-RP typical failure

Type of failure	Class	1	Class 2	Class 3	Class 4
						Insufficient flow of freezing water	-10％	-20％	-30％	-40％
Insufficient cooling water flow	-10％	-20％	-30％	-40％
					Leakage of refrigerant	-10％	-20％	-30％	-40％
Excess refrigerant	+10％	+20％	+30％	+40％
					With non-condensable gases	+1.0％	+1.7％	+2.4％	+5.7％
Excessive lubricating oil	+14％	+32％	+50％	+68％
					Condenser fouling	+12％	+20％	+30％	+45％

2) The data of 4 grades are labeled with faults and then mixed as an input data set of the embodiment, and the state data at least comprises characteristics such as evaporation temperature, condensation temperature, suction temperature, exhaust temperature, evaporation pressure, condensation pressure, flow, valve position information, compressor power and the like which are expressed as the characteristics

Where s represents the number of samples and r represents the number of features in each sample. Incorporating a corresponding fault type label y _i And forming training sample data.

In the data acquisition process, missing values or abnormal values may occur, and for the abnormal values, the original data can be deleted and processed as the missing values; for the missing value, a Lagrange interpolation method is adopted for processing, and the interpolation polynomial is as follows:

where θ represents the position of the missing value data, L _n (theta) is interpolated x, n is n data points adjacent to missing value column, k is index, x _k Value l at this location for the corresponding function, i.e. the original fault data _k (θ) is a basis function, and the mathematical expression is:

where θ represents the position of the missing value data, and θ _k Representing the k-th position adjacent to theta.

In order to reduce the operation amount of subsequent model processing, each characteristic of all samples is normalized by adopting a min-max normalization method, the normalized data range is between [0,1], and the calculation formula is as follows:

where δ' is the normalized value, δ is the original data value, δ _max Is the maximum value, δ, of the corresponding feature in the data sample _min Is the minimum value of the corresponding feature in the data sample.

Step (2), a LightGBM-based model is constructed to serve as an air conditioner fault diagnosis model, and the hyper-parameters and the value range of the hyper-parameters, which need to be optimized, of the model are determined, and the method specifically comprises the following steps:

establishing LightGBM model

Combining M weak regression trees into a strong regression tree linearly, wherein the calculation formula is as follows:

in the formula: f (x) is the final output value; f. of _m (x) The output value of the m weak regression tree.

The main improvements of the LightGBM model include the histogram algorithm (fig. 3) and the Leaf growth with depth limitation (Leaf-wise) strategy. The histogram algorithm divides the continuous data into K integers and constructs a histogram of width K. And during traversal, the discretized value is used as an index to be accumulated in the histogram, and then the optimal decision tree segmentation point is searched.

While the Leaf-wise strategy with depth limitation means that at each split, the Leaf that finds the greatest gain is split and cycled through. Meanwhile, the complexity of the model is reduced and overfitting is prevented by limiting the depth of the tree and the number of leaves.

The hyper-parameters that the LightGBM model needs to optimize are shown in table 2:

TABLE 2 LightGBM model hyper-parameters

Step (3), optimizing the hyper-parameters of the LightGBM model by using a grid search method combined with a five-fold cross validation method, and determining the optimal hyper-parameter combination of the model, wherein the method specifically comprises the following steps:

3.1 Set a higher learning rate first, accelerate convergence speed, set learning _ rate =0.1, and then select Boosting as the gradient Boosting decision tree gbdt;

3.2 Secondly, optimizing max _ depth and num _ leaves, wherein if the max _ depth is too large, the model is over-fitted, and if the max _ depth is too small, the learning capacity of the model is reduced, so that the value range of the max _ depth is selected to be [3,11], and the step length is 1; the value of num _ leaves is generally less than 2 (max _ depth), so the value range of num _ leaves is [30,300], and the step length is 10;

3.3 Min _ data _ in _ leaf and max _ bin are optimized, the two parameters are adjusted to reduce overfitting, the value range of min _ data _ in _ leaf is set to [20,120], and the step size is 10; the value range of max _ bin is set to [10,50], and the step length is 5;

3.4 Optimizing feature _ fraction, bagging _ fraction and bagging _ fraction, aiming at improving training speed and reducing overfitting, wherein the value ranges of three parameters are set to be 0.5,1.0, and the step length is 0.1;

3.5 Reduce the learning rate, set the learning rate learning _ rate value range to [0.01,0.2], and the step size to 0.01. The number of iterations is increased and the number of iterations,

3.6 Setting initial hyper-parameters of the model according to the hyper-parameters determined in the steps 3.1) to 3.5), using the training sample data obtained in the step 1, and performing iterative training on the model by adopting a five-fold cross validation method through a minimum loss function;

3.7 Judging whether the model reaches an early stop condition, if so, ending the training in advance; if not, returning to the step 3.6) to continue iteration;

and repeating the steps 3.1) -3.7) until the optimal hyper-parameter combination is obtained, and obtaining the optimal air conditioner fault diagnosis model.

The LightGBM model optimal hyper-parameters are shown in table 3.

TABLE 3 LightGBM model optimal hyperparameters

Hyper-parameter
		max_depth	10
num_leaves	200
		min_data_in_leaf	100
max_bin	35
		feature_fraction	0.8
bagging_fraction	0.8
		bagging_freq	1
learning_rate	0.06

Wherein, the five-fold cross validation mode is as follows (fig. 4): firstly, disorganizing the acquired original data, then randomly dividing the data into five parts, training a model by sequentially using four parts of data, detecting the identification and classification accuracy of the fault diagnosis model by using the remaining part of data as a verification set, and finally, evaluating the diagnosis performance of the model on fault problems by taking the average value of the five parts of accuracy obtained on the verification set.

In this embodiment, after the LightGBM model parameters reach the optimum, model test and evaluation are performed on the cross validation sample set and the test set by using a common model evaluation algorithm, which specifically comprises the following steps:

the accuracy _ score (accuracy) is used to evaluate the model, and the accuracy is calculated by comparing the predicted fault label with the real fault label, and the calculation formula is as follows:

in the formula, s is the number of samples,

to predict the label, y _i For the actual label, the larger the accuracy _ score is, the closer the diagnosed fault label and the actual fault label are, the more accurate the classification isAnd (8) determining.

The LightGBM model of the present invention was compared to some common models as shown in table 4.

TABLE 4 comparison of different models

Through comparison in table 3, it can be seen that the Decision Tree-based models such as Random Forest, decision Tree, lightGBM, etc. have better diagnostic effects, which indicates that the Decision Tree-based models have advantages in processing such tabular data, and it can be seen that the LightGBM model of the present invention has the highest accuracy and the best diagnostic effect.

In conclusion, the invention provides the fault diagnosis method for the heating ventilation air conditioning system based on the LightGBM algorithm and the grid search algorithm, the LightGBM algorithm improves the model performance, meanwhile, the calculation speed is 10 times of that of the original GBDT method, and the size of the memory occupied by the training model is reduced by 3 times. The major improvements of the LightGBM model over the original GBDT method include the histogram algorithm and the Leaf growth with depth limitation (Leaf-wise) strategy. The histogram algorithm divides the continuous data into K integers and constructs a histogram of width K. And during traversal, the discretized value is used as an index to be accumulated in the histogram, so that the optimal decision tree segmentation point is searched, and the prediction precision is improved while the speed is improved.

The above are merely preferred examples of the present invention and do not limit the scope of the present invention. It should be noted that, for those skilled in the art, the equivalent substitutions or changes made according to the technical solutions or concepts of the present invention belong to the protection scope of the present invention.

Claims

1. A heating ventilation air conditioning system fault diagnosis method based on a LightGBM algorithm and a grid search algorithm is characterized in that: the method comprises the following steps:

(1) Data acquisition and pretreatment, namely collecting state data of a water chilling unit of a heating ventilation air conditioning system during normal operation and different faults, cleaning and carrying out normalized pretreatment on the data, and marking fault labels to obtain training sample data; the state data at least comprises evaporation temperature, condensation temperature, suction temperature, exhaust temperature, evaporation pressure, condensation pressure, flow, valve position information and compressor power;

(2) Constructing a LightGBM model as an air conditioner fault diagnosis model, and determining an over-parameter to be optimized of the model and a value range of the over-parameter; the input of the LightGBM model is preprocessed state data, and the output of the LightGBM model is a predicted value of a fault label;

(3) Training and optimizing the hyper-parameters of the LightGBM model by using the training sample data in the step (1) and combining a grid search method with a five-fold cross validation method, and determining the optimal hyper-parameter combination of the model to obtain an optimal air conditioner fault diagnosis model; the determined hyper-parameters of the LightGBM model needing to be optimized comprise: bagging _ freq, bagging _ fraction, feature _ fraction, num _ leaves, max _ bin, leaving _ rate, max _ depth, and min _ data _ in _ leaf; wherein, the value ranges of the three parameters of the bagging _ freq, the bagging _ fraction and the feature _ fraction are all set as [0.5,1.0], and the step length is 0.1; the value range of num _ leaves is [30,300], and the step length is 10; the value range of max _ bin is set to [10,50], and the step length is 5; setting the value range of the learning _ rate as [0.01,0.2], and setting the step length as 0.01; the value range of max _ depth is [3,11], and the step length is 1; setting the value range of min _ data _ in _ leaf as [20,120], and the step length is 10;

the step (3) specifically comprises the following substeps:

(3.1) determining max _ depth and num _ leaves;

(3.2) determining min _ data _ in _ leaf and max _ bin;

(3.3) determining feature _ fraction, bagging _ fraction and bagging _ freq;

(3.4) reducing the learning _ rate, increasing the iteration times, and training and verifying the model by using a five-fold cross-validation method;

repeating the steps (3.1) - (3.4) until the optimal hyper-parameter combination is obtained, and obtaining an optimal air conditioner fault diagnosis model;

(4) Collecting the running state data of the water chilling unit of the heating, ventilating and air conditioning system in real time, carrying out data cleaning and normalization preprocessing, and inputting the data into an optimal air conditioner fault diagnosis model for air conditioner fault diagnosis.

2. The method for diagnosing faults of heating, ventilating and air conditioning system based on LightGBM algorithm and grid search algorithm as claimed in claim 1, wherein: in the step (1), a lagrange interpolation method is adopted to process missing values occurring in the data acquisition process, and the interpolation polynomial is as follows:

where θ represents the position of the missing value data, L _n (theta) is x after interpolation, n is n same state data points adjacent to the missing value, k is an index, and x _k Taking the value of the kth data adjacent to the missing value, l _k (θ) is a basis function, and the mathematical expression is:

wherein, theta _k Represents the k-th position adjacent to θ; and removing abnormal values occurring in the data acquisition process and processing the abnormal values by adopting a Lagrange interpolation method.