Abnormality detection method and system for air conditioning unit, electronic equipment and storage medium

Abstract

The embodiment of the invention provides an anomaly detection method and system for an air conditioning unit, electronic equipment and a storage medium, and belongs to the field of artificial intelligence and the field of predictive maintenance of industrial equipment. The method comprises the steps that operation parameters of the air conditioning unit are collected and input into an anomaly detection model, anomaly time sequence parameters in the operation parameters of the air conditioning unit are output, and the anomaly detection model is constructed based on a TimesNet algorithm; and inputting the abnormal time sequence parameters into a fault classifier to obtain a fault classification result of the air conditioning unit. A time sequence model of normal operation of the unit is established by using a Times Net algorithm, and multi-sensor operation parameters of the air conditioning unit are detected in real time. And inputting the time sequence interval in which the problem is found into a fault classifier which is jointly constructed based on a mode matcher and a multi-classifier to classify the type of the fault. Abnormality detection and classification of the air conditioning unit equipment are achieved, and the equipment operation efficiency is effectively improved.

Description

Technical Field

[0001] The present invention relates to the fields of artificial intelligence and predictive maintenance technology for industrial equipment, and in particular to an abnormality detection method, system, electronic device and storage medium for an air-conditioning unit. Background Art

[0002] In recent years, especially in the field of deep learning, thanks to the powerful nonlinear modeling capabilities of deep models, many studies have proposed methods to capture complex changes in real-world time series. One type of method uses recurrent neural networks (RNNs) to model continuous time points based on the Markov assumption. However, these methods are usually unable to effectively capture long-term dependencies and are inefficient due to the sequence computing paradigm. Another type of method uses convolutional neural networks (TCNs) along the time dimension to extract change information. However, due to the local characteristics of one-dimensional convolution kernels, these methods can only model changes between adjacent time points, so they are still insufficient in capturing long-term dependencies. Recently, Transformers modeling based on attention mechanisms has been widely used in time series. This type of method uses attention mechanisms or their variants to capture pairwise temporal dependencies between time points. However, it is difficult for attention mechanisms to directly identify reliable dependencies from scattered time points because temporal dependencies may be deeply buried in complex temporal patterns.

[0003] As a core component of the heating, ventilation and air conditioning system, the central air-conditioning chiller is mainly responsible for the effective output of cooling capacity and provides the basic conditions for the stable operation of the entire system. The chiller is mainly composed of an evaporator, a condenser, a compressor and a throttling device. Under their joint action, the refrigerant can continuously cool the designated space. However, in actual applications, due to the very complex structure of the chiller and the harsh working environment, it is easy to have various types of failures. Chiller failure will not only have an adverse effect on the effectiveness of the heating, ventilation and air conditioning system, but also cause energy waste and shortened equipment life. Chiller failures are mainly divided into hard failures and soft failures. Among them, hard failures often have a great impact on the equipment. For example, a motor burnout may cause the system to completely deviate from its original state. Although the probability of hard faults is relatively low and difficult to prevent, the symptoms of hard faults are usually obvious when they occur, so hard faults are easier to detect and do not require complex fault diagnosis methods to identify. In contrast, the symptoms of soft faults in chillers are not obvious, and users and maintenance personnel usually find it difficult to detect, so soft faults may persist for a long time, which will cause long-term damage to the equipment and cause unnecessary energy loss. A survey of mainstream chiller products found that the probability of refrigerant leakage failures reached 18.5%, while the probability of condenser failures reached 8.5%. At the same time, the survey report also listed 7 common types of faults, namely, reduced condenser water flow (FWC), reduced evaporator water flow (FWE), refrigerant leakage / insufficient (RL), excessive refrigerant (RO), excessive oil (EO), condenser scaling (CF), and non-condensable gas (NC) in the refrigerant.

[0004] Through the analysis of the above-mentioned chiller failure problems, it can be seen that each failure will cause a relatively serious adverse effect. It can also be found that under the same conditions, different types of failures may occur, and different types of failures may also cause similar effects. At the same time, the adverse effects caused by a certain failure may cause another failure. In short, there is a very close connection between various failures, which greatly increases the difficulty of building a fault diagnosis model. Summary of the invention

[0005] The purpose of the embodiments of the present invention is to provide an abnormality detection method, system, electronic device and storage medium for an air-conditioning unit, which are used to completely or at least partially solve the technical problems existing in the above-mentioned prior art.

[0006] In order to achieve the above object, an embodiment of the present invention provides an abnormality detection method for an air conditioning unit, comprising: Collect the operating parameters of the air conditioning unit; Inputting the operating parameters into a pre-built anomaly detection model, and outputting abnormal time series parameters in the operating parameters of the air-conditioning unit, wherein the pre-built anomaly detection model is built based on the TimesNet algorithm; The abnormal time series parameters are input into a fault classifier to obtain a fault classification result of the air-conditioning unit, wherein the fault classifier is constructed based on a combination of a pattern matcher and a multi-classifier.

[0007] Optionally, the operating parameters of the air-conditioning unit include but are not limited to the core equipment parameters of the air-conditioning unit and the parameters obtained at the inlet and outlet positions.

[0008] Optionally, the process of building an anomaly detection model based on the TimesNet algorithm includes: Constructing a data set sample, wherein the data set sample is data of the air-conditioning unit operating normally for multiple consecutive periods of time; Building an anomaly detection model based on the TimesNet algorithm: The air conditioning unit data is converted into two-dimensional data, and a set of two-dimensional tensors at different frequencies and periods is obtained by discrete Fourier transform; The TimesNet model stacks multiple TimesBlocks in a residual manner. For each TimesBlock, the Inception module is used in two-dimensional space to capture various time changes in a two-dimensional tensor set, and various time changes are fused according to the normalized amplitude value to output the information representation in the two-dimensional tensor set. Train the anomaly detection model: Configure the parameters of the anomaly detection model. At the beginning of the training phase of the anomaly detection model, select samples from the data set as the training set to train the anomaly detection model, and select samples from the data set as the verification set to verify the anomaly detection model to obtain the anomaly detection model.

[0009] Optionally, the abnormal time series parameters are input into a fault classifier to obtain a fault classification result of the air conditioning unit, including: The abnormal time series parameters are input into the pattern matcher for feature extraction to obtain the abnormal type of the abnormal time series; For each abnormal time series abnormal type, a multi-classifier is used to determine the fault type corresponding to each abnormal time series abnormal type, wherein the multi-classifier is formed by stacking the random forest algorithm, the AdaBoost algorithm and the XGBoost algorithm.

[0010] Optionally, the abnormal time series parameters are input into the pattern matcher for feature extraction to obtain abnormal time series abnormal types, including: Use sliding windows to extract abnormal time series, and perform regularization on the abnormal time series to obtain regularized abnormal detection fragments; One-dimensional convolution is used to extract features in the anomaly detection segment, and a fully connected layer is used to classify the features to obtain the time series anomaly type.

[0011] Optionally, the construction process of the multi-classifier includes training and prediction of base learners and training and prediction of meta learners, wherein the base learners include three types of base learners: random forest, AdaBoost and XGBoost, and the meta learner adopts a decision tree model: Divide the training data set into a training set and a validation set, train each base learner on the training set, and make predictions on the validation set; Use the trained random forest, AdaBoost and XGBoost three types of base learners to predict the validation set, record the prediction results of each base learner, and combine the prediction results into a new secondary feature set, where each row of the new secondary feature set corresponds to a sample in the validation set, and each column corresponds to the prediction result of a base learner; The new secondary feature set is used as input and the true classification target label is used as output to train the meta-learner and obtain a multi-classifier composed of the random forest algorithm, AdaBoost algorithm and XGBoost algorithm stacked together.

[0012] Optionally, the abnormal time series parameters in the operating parameters of the air-conditioning unit include a single abnormal point, a single-classification continuous abnormal point, and a multi-classification continuous abnormal point; wherein a single abnormal point is characterized by both the before and after of the abnormal point being detected as a normal data point; a single-classification continuous abnormal point is characterized by both the before and after of the abnormal point being detected as an abnormal data point, and the type of the abnormal point before or after the point is the same as the current abnormal point; a multi-classification continuous abnormal point is characterized by both the before and after of the abnormal point being detected as an abnormal data point, and the type of the abnormal point before or after the point is different from the current abnormal point.

[0013] On the other hand, the present invention also provides an abnormality detection system for an air conditioning unit, comprising: Data acquisition module, used to collect operating parameters of the air-conditioning unit; An anomaly detection module, used for inputting the operating parameters into a pre-built anomaly detection model, and outputting an abnormal time series parameter in the operating parameters of the air-conditioning unit, wherein the pre-built anomaly detection model is built based on the TimesNet algorithm; The abnormal classification module is used to input the abnormal time series parameters into the fault classifier to obtain the fault classification result of the air-conditioning unit, wherein the fault classifier is constructed based on the combination of a pattern matcher and a multi-classifier.

[0014] On the other hand, the present invention also provides an electronic device, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor implements the steps of the above-mentioned method for detecting abnormalities of an air-conditioning unit when executing the program.

[0015] On the other hand, the present invention further provides a non-transitory computer-readable storage medium having a computer program stored thereon, and when the computer program is executed by a processor, the steps of the above-mentioned method for detecting abnormality of an air-conditioning unit are implemented.

[0016] Through the above technical solution, the TimesNet network is used to extract the characteristics of the normal operation of the unit equipment within and between cycles, and the time series model of the normal operation of the unit equipment is established to perform real-time detection of the multi-sensor operating parameters of the air-conditioning unit. The time series interval where the problem is found is input into the fault classifier based on the joint construction of the pattern matcher and the multi-classifier to classify the type of fault. The abnormal detection and classification of the air-conditioning unit equipment is realized, which effectively improves the efficiency of equipment operation.

[0017] Other features and advantages of the embodiments of the present invention will be described in detail in the subsequent detailed description. BRIEF DESCRIPTION OF THE DRAWINGS

[0018] The accompanying drawings are used to provide a further understanding of the embodiments of the present invention and constitute a part of the specification. Together with the following specific embodiments, they are used to explain the embodiments of the present invention, but do not constitute a limitation on the embodiments of the present invention. In the accompanying drawings: Figure 1 is an implementation flow chart of an abnormality detection method for an air conditioning unit provided by an embodiment of the present invention; Figure 2 A TimesNet network structure diagram provided by an embodiment of the present invention; Figure 3 is a schematic diagram of feature extraction based on a pattern matcher provided by an embodiment of the present invention; Figure 4 (am) is a schematic diagram of a time series anomaly type provided by an embodiment of the present invention; Figure 5 It is a time series anomaly type classification network structure diagram provided by an embodiment of the present invention; Figure 6 This is an overall architecture diagram of an abnormality detection of an air-conditioning unit provided by an embodiment of the present invention; Figure 7 It is a structural schematic diagram of an abnormality detection system for an air-conditioning unit provided by an embodiment of the present invention. DETAILED DESCRIPTION

[0019] The specific implementation of the embodiment of the present invention is described in detail below in conjunction with the accompanying drawings. It should be understood that the specific implementation described here is only used to illustrate and explain the embodiment of the present invention, and is not used to limit the embodiment of the present invention.

[0020] See also Figure 1 As shown, it is a flowchart of an implementation method of an abnormality detection method of an air conditioning unit provided by an embodiment of the present invention, comprising the following execution steps: Step 100: Collect operating parameters of the air conditioning unit.

[0021] Specifically, the operating parameters of the air-conditioning unit include, but are not limited to, the core equipment parameters of the air-conditioning unit and the parameters obtained at the inlet and outlet positions.

[0022] In some implementations, multiple types of sensor data refer to relevant data obtained from the core equipment and inlet and outlet locations of the air-conditioning unit, mainly including: chilled water inlet and outlet temperatures (measured by resistance measuring devices and thermistors, respectively), cooling water inlet and outlet temperatures (measured by resistance measuring devices and thermistors, respectively), cooling water loop heat exchanger inlet and outlet temperatures, chilled water loop heat exchanger near outlet temperatures, compressor power, and other 47 specific parameters. Since the time intervals of the data collected by each sensor are exactly the same, the 47 types of data are directly input into the TimesNet time series analysis algorithm without any type of preprocessing according to their respective units, value ranges, etc. The input parameters can be expressed as The specific parameters are shown in Tables 1-1 and 1-2, where Table 1-2 follows Table 1-1: Table 1-1 Core parameters of central air conditioning units

[0023] Table 1-2 Core parameters of central air conditioning units

[0024] Step 101: input the operating parameters into a pre-built anomaly detection model, and output abnormal time series parameters in the operating parameters of the air-conditioning unit.

[0025] The pre-built anomaly detection model is built based on the TimesNet algorithm.

[0026] In some implementations, the process of building an anomaly detection model based on the TimesNet algorithm includes: Constructing a data set sample, wherein the data set sample is data of the air-conditioning unit operating normally for multiple consecutive periods of time; Building an anomaly detection model based on the TimesNet algorithm: The air conditioning unit data is converted into two-dimensional data, and a set of two-dimensional tensors at different frequencies and periods is obtained by discrete Fourier transform; The TimesNet model stacks multiple TimesBlocks in a residual manner. For each TimesBlock, the Inception module is used in two-dimensional space to capture various time changes in a two-dimensional tensor set, and various time changes are fused according to the normalized amplitude value to output the information representation in the two-dimensional tensor set. Train the anomaly detection model: Configure the parameters of the anomaly detection model. At the beginning of the training phase of the anomaly detection model, select samples from the data set as the training set to train the anomaly detection model, and select samples from the data set as the verification set to verify the anomaly detection model to obtain the anomaly detection model.

[0027] In some embodiments, the anomaly detection of air-conditioning unit equipment based on the TimesNet anomaly detection model mainly includes two steps: TimesNet training phase and TimesNet reasoning phase. Among them, the TimesNet training phase refers to training TimesNet with a normal time series data set, so that TimesNet learns the time series regularity information of the central air-conditioning unit under normal operation mode. The TimesNet reasoning phase refers to inputting the sensor data in the production system operation into the trained TimesNet to realize the detection of abnormal time series.

[0028] Among them, the TimesNet training stage includes: (1) Dataset sample construction and model training strategy: The training input data samples are mainly the normal operation data of the units for multiple consecutive periods of time. This type of normal operation data has the following characteristics: ① Operation data refers to data obtained from a specific type of equipment; ② The running data is collected in different time periods, so the parameters of each time series may be completely different. In addition, there may be a large jump between the starting point and the end point of each data segment.

[0029] In view of the current situation of data set samples, this method adopts two methods when constructing the test data set. By comparing the actual application effects of the two methods in a specific equipment anomaly detection task, the method of training the data set construction method in the production environment is selected.

[0030] Method ① uses data splicing and aggregation: This method constructs a training sample set by directly splicing the data of different time periods of the same device. The advantage of this method is that the data set after splicing is large in sample size, which allows the TimesNet network to better learn the operating rules of the device in different time periods. The disadvantage of this method is that there are jump points at the splicing of time series, which can easily affect the modeling of TimesNet to a certain extent.

[0031] Method ② uses the same data comparison method: This method trains the TimesNet network by using the same or similar data set as the training data set for the same device and the current operating environment. The data set that is the same or similar to the current operating environment refers to the data information generated by the normal operation of the air-conditioning unit under the same or similar temperature, humidity, heat load and other conditions. The advantage of this method is that the data information generated under the same or similar conditions can facilitate TimesNet to model the situation more accurately and enhance the impact of anomaly detection. However, the disadvantage of this method is that the universality of TimesNet algorithm modeling is not strong, and TimesNet in different environments needs to be adjusted and matched in time according to the situation.

[0032] (2) Algorithm modeling based on TimesNet: Algorithm modeling is performed on the time series data of 47 sensors based on TimesNet. The specific steps are as follows: ① Time series representation transformation: This step converts the 47 time series parameters collected by the sensor from a one-dimensional time series into a two-dimensional data representation, and uses discrete Fourier transform to model the changes within the time series data period (continuous changes in adjacent points) and between cycles (long-term regular changes).

[0033] For a one-dimensional time series with a length of T and a number of channels of C , for long time series, its periodicity can be calculated by Fourier transform, as shown in formulas (1)-(3): (1) (2) (3) Among them, A represents the intensity of each frequency component calculated after fast Fourier transform in the one-dimensional time series collected by the sensor, and the k frequencies with the largest intensity are Corresponding to the most significant k period lengths Therefore, the result obtained by performing fast Fourier transform on the one-dimensional time series collected by the sensor can be simplified as formula (4): (4) Based on the above calculations, different two-dimensional tensor representations can be calculated according to different period lengths: (5) Among them, the Padding operation is to maintain the consistency of tensor dimensions; Indicates the frequency , cycle length The i-th 2D tensor to transform based on. The rows and columns represent the changes within a cycle and between cycles, respectively. The one-dimensional time series data Converted into a two-dimensional tensor set with k different frequencies and periods . This structure can use two-dimensional convolution kernels for feature extraction and processing.

[0034] ② Two-dimensional convolution operation of time series: refer to the ResNet neural network structure Figure 2 As shown in Figure 1, TimesNet stacks multiple TimesBlocks in a residual manner, which can capture various temporal 2D changes in k different reshaped tensors using Inception modules in 2D space and fuse them together according to the normalized amplitude values. Specifically, for the central air conditioning unit with a time length of T, , through the embedding layer Projecting the original input to the deep features For the lth layer of TimesNet , the input is , then the residual structure can be expressed as: (6) For the lth TimesBlock, the whole process consists of two consecutive parts: capturing the two-dimensional changes in time and adaptively aggregating representations of different periods. Capturing the two-dimensional changes in time: Similar to the above equation, deep features can be estimated by period(⋅) Based on the estimated period length, the one-dimensional time series can be transformed into a two-dimensional space and a set of two-dimensional tensors can be obtained. Information representation can be easily obtained from these tensors through parameter-efficient seating blocks. This process is formalized as: (7) (8) (9) (10) Model training strategy: For the datasets constructed in the two ways, the TimesNet model training is performed in the same way and with the same parameter configuration. In order to further improve the universality of the model during training, cross-validation is used for training. For example, at the beginning of model training, 80% of the dataset is selected as the training set and 20% of the dataset is selected as the validation set for model training. The parameter configuration of TimesNet in the training stage is shown in Table 2: Table 2 TimesNet parameter configuration table

[0035] TimesNet reasoning stage: The TimesNet prediction reasoning stage is the same as the prediction reasoning stage of learning algorithms such as deep learning networks. The time series data collected by each sensor is input into the TimesNet anomaly detection module. Since the TimesNet time series model has modeled and learned the time series law of the normal operation of the central air-conditioning unit during the training stage, it can predict and output a detection result of whether there is an anomaly. Usually, if the prediction result is normal, it is represented as "0", and if the result is abnormal, it is represented as "1".

[0036] In some embodiments, the abnormal time series parameters in the operating parameters of the air-conditioning unit include a single abnormal point, a single-classification continuous abnormal point, and a multi-classification continuous abnormal point; wherein a single abnormal point is characterized as the abnormal point before and after being detected as a normal data point; a single-classification continuous abnormal point is characterized as the abnormal point before or after being detected as an abnormal data point, and the type of the abnormal point before or after the point is the same as the current abnormal point; a multi-classification continuous abnormal point is characterized as the abnormal point before or after being detected as an abnormal data point, and the type of the abnormal point before or after the point is different from the current abnormal point.

[0037] Step 102: Input the abnormal time series parameters into the fault classifier to obtain the fault classification result of the air conditioning unit.

[0038] Wherein, the fault classifier is constructed based on the combination of pattern matcher and multi-classifier.

[0039] Specifically, when executing step 102, the following steps may be specifically performed: S1020: Input the abnormal time series parameters into the pattern matcher for feature extraction to obtain the abnormal type of the abnormal time series.

[0040] Specifically, when executing step S1020, the following sub-steps may be specifically performed: SA: Use sliding windows to extract abnormal time series and regularize the abnormal time series to obtain regularized anomaly detection fragments.

[0041] SB: Use one-dimensional convolution to extract features in the anomaly detection segment, and use a fully connected layer to classify the features to obtain the time series anomaly type.

[0042] In some embodiments, see Figure 3 As shown, it is a schematic diagram of feature extraction based on a pattern matcher provided by an embodiment of the present invention. For the abnormal time series detected by TimesNet, it is proposed to use a pattern matcher to further extract features from the detected suspected abnormal time series. The feature extraction operation mainly includes three main steps: (1) Using a sliding window to extract the time series: There is diversity in the suspected abnormal time series detected by TimesNet. A single abnormal point refers to a point where both the front and back of the abnormal point are detected as normal data points; a single-classification continuous abnormal point refers to a point where both the front and back of the abnormal point are detected as abnormal data points, and the type of the abnormal point before or after the point is the same as the current abnormal point; a multi-classification continuous abnormal point refers to a point where both the front and back of the abnormal point are detected as abnormal data points, and the type of the abnormal point before or after the point is different from the current abnormal point. For the above three types of abnormal situations, and it is known that the TimesNet anomaly detection network has a certain degree of detection error, in order to further combine the characteristics of the time series, it is proposed to use a sliding window with a width of T to process the time series and extract a time series with a length of T. . (2) Filtering normal working time series: First, pattern matching seeks to comprehensively utilize the hypothesis test of two samples as a coarse-grained anomaly detection algorithm, which has the characteristics of high efficiency and accuracy. This method can filter out indicators that perform normally when an event occurs, thereby significantly reducing the search space. (3) Performing abnormal pattern classification: Abnormal pattern classification can classify the abnormal time series abnormal model after filtering the normal time series, and classify it into one of the thirteen types of abnormalities from the perspective of time series, such as Figure 4 (am), where Figure 4 (a) is a schematic diagram showing a sudden rise anomaly. Figure 4 (b) is a schematic diagram showing a sudden drop anomaly. Figure 4 (c) is a schematic diagram showing an upward displacement anomaly. Figure 4 (d) is a schematic diagram showing the downward displacement anomaly. Figure 4 (e) shows a schematic diagram of a steady growth anomaly. Figure 4 (f) is a schematic diagram showing a stable descent anomaly. Figure 4 (g) is a schematic diagram of the instantaneous peak anomaly. Figure 4 (h) is a schematic diagram of the instantaneous valley value anomaly. Figure 4 (i) is a schematic diagram showing the abnormal instantaneous level rise. Figure 4 (j) is a schematic diagram of the abnormal instantaneous level drop. Figure 4 (k) is a schematic diagram of multiple instantaneous peak anomalies. Figure 4 (l) is a schematic diagram showing multiple instantaneous valley value anomalies. Figure 4 (m) is a schematic diagram of fluctuation anomaly. Furthermore, the feature extraction of time series is also achieved through the classification of this anomaly type. This method is a new type of anomaly pattern classification method based on a one-dimensional convolutional neural network (see Figure 5 As shown in Figure 2, the regularized anomaly detection fragment is firstly extracted using a one-dimensional convolutional neural network, and then the fully connected layer is used for feature classification. The calculation process is as follows: (11) in, represents the extracted abnormal time series features; Represents the vectors corresponding to the thirteen categories; , and They represent batch normalization regularization calculation, one-dimensional convolutional neural network and 2-layer fully connected network structure respectively.

[0043] S1021: For each abnormal time series abnormal type, a multi-classifier is used to determine the fault type corresponding to each abnormal time series abnormal type, wherein the multi-classifier is formed by stacking a random forest algorithm, an AdaBoost algorithm, and an XGBoost algorithm.

[0044] In some embodiments, the construction process of the multi-classifier includes training and prediction of base learners and training and prediction of meta learners, wherein the base learners include three types of base learners: random forest, AdaBoost and XGBoost, and the meta learner adopts a decision tree model: Divide the training data set into a training set and a validation set, train each base learner on the training set, and make predictions on the validation set; Use the trained random forest, AdaBoost and XGBoost three types of base learners to predict the validation set, record the prediction results of each base learner, and combine the prediction results into a new secondary feature set, where each row of the new secondary feature set corresponds to a sample in the validation set, and each column corresponds to the prediction result of a base learner; The new secondary feature set is used as input and the true classification target label is used as output to train the meta-learner and obtain a multi-classifier composed of the random forest algorithm, AdaBoost algorithm and XGBoost algorithm stacked together.

[0045] In some implementations, a multi-classifier consisting of random forest, XGboost, and AdaBoost joint modeling is used to implement classification of extracted abnormal time series features.

[0046] (1) Construction of fault classification sample set: Since the pattern matcher was used to extract features from a set of sensor data in the previous task, the input of the multi-classifier anomaly detection is to extract abnormal time series features. To this end, the ground truth of the classifier needs to be preprocessed to a certain extent. Select a row of data of a certain fault type, perform sliding window calculation according to the parameters consistent with the time length T of the feature matcher, extract the time series features corresponding to the fault point, and construct a pair of samples that can be trained under supervision (time series features of a certain fault point parameter, fault type). The preprocessing of a fault point sample is completed in the above way, and a fault classification sample set is constructed from all types of data of the seven major types of faults.

[0047] (2) Construction of multiple classifiers: Three commonly used classifier algorithms are selected: random forest algorithm, AdaBoost algorithm and XGBoost algorithm. The three classifiers are stacked to finally classify the central air-conditioning unit faults, which can further improve the accuracy of classifier classification. The following introduces the three classification algorithms and the integration ideas of the three algorithms: ① Random forest is an integrated model composed of multiple decision trees. Each tree is trained on a random subset of the training data. This method is called "bagging". Random forest uses the idea of ​​ensemble learning to reduce overfitting and improve the generalization ability of the model by combining the prediction results of multiple models. ② AdaBoost is an effective boosting algorithm that constructs a strong classifier by weighted combination of multiple weak classifiers. A weak classifier refers to a classifier that performs slightly better than random guessing on a certain classification task. Its adaptive weight adjustment mechanism enables it to focus on samples that are difficult to classify, thereby improving the overall performance. ③ XGBoost is a powerful machine learning algorithm that has been widely used in many fields for its efficiency, flexibility and superior performance. It combines the idea of ​​gradient boosting with regularization technology, can effectively process large-scale data sets, and provide good prediction capabilities. Through reasonable parameter tuning, XGBoost can perform well in various complex tasks. ④ Model stacking of ensemble learning: Stacking is an ensemble learning method, in which three types of base learners, random forest, AdaBoost and XGBoost, are trained and their outputs are used as input features to train a meta-learner to obtain more accurate predictions. The basic idea of ​​stacking is to use the complementarity of different models to improve the accuracy and robustness of predictions.

[0048] (3) Training and prediction of multiple classifiers: The training and prediction of the classification of central air-conditioning unit faults by stacking three types of classifiers can be divided into two parts: training and prediction of base learners and training and prediction of meta learners: ① Training base learners: Divide the training data set into two parts: training set and validation set. The commonly used division method is cross-validation. Train each base learner on the training set and make predictions on the validation set. This can avoid overfitting. ② Generate predictions of base learners: Use the trained base learners to predict the validation set and record the prediction results of each base learner. In addition, these prediction results are combined into a new data set called the "secondary feature set". Each row of this data set corresponds to a sample in the validation set, and each column corresponds to the prediction result of a base learner. ③ Training meta learners: Use the "secondary feature set" as input and the actual classification target label as output to train a meta learner. The meta learner uses a decision tree model. ④ Final prediction: For new test data, first use all base learners to make predictions to generate the prediction results of the base learners. These prediction results are input into the meta-learner as features to obtain the final prediction results.

[0049] In some implementations, the failures of the central air-conditioning unit are divided into two categories: hard failures and soft failures. Hard failures mainly refer to the situation where the central air-conditioning unit stops working directly due to equipment shutdown, explosion, etc. This type of situation is very easy to identify, but it will cause significant economic losses and safety risks, and should be avoided as much as possible in actual production. Soft failures mainly involve problems such as decreased flow rate in the pipe network, unrelieved pressure, and untimely heat exchange. Such problems will not cause the central air-conditioning unit to stop working, but will cause problems such as increased workload of the compressor and increased energy consumption of the chiller.

[0050] 1. Description of fault classification: For soft faults, these unit fault problems will cause relatively serious adverse effects. Under the same conditions, different types of faults may occur, and different types of faults may also cause similar effects. At the same time, the adverse effects of a certain fault may cause another fault. In short, there is a very close connection between various faults, which will greatly increase the difficulty of building a fault diagnosis model. At present, there are seven aspects of faults: (1) Reduced condenser water flow: In the water circulation system where the condenser is located, if there are problems such as improper water pump selection or failure, filter blockage, abnormal valve adjustment or air in the pipeline, the water flow entering the condenser will be reduced.

[0051] (2) Evaporator water flow reduction: The reasons for the reduction of evaporator water flow are similar to those for the reduction of condenser water flow.

[0052] (3) Refrigerant leakage / insufficiency: In a chiller, the refrigerant, which is the main medium, circulates in complex pipes. There are many connecting components in the circulating pipes. Once there are cracks in the joints or holes in the valves of the components, refrigerant leakage will occur.

[0053] (4) Excessive refrigerant: In the case of excessive refrigerant, the refrigerant usually cannot be completely evaporated, which will cause liquid hammer in the compressor and damage the impeller. Similarly, the excess refrigerant cannot be completely liquefied in the condenser, and the unliquefied refrigerant vapor will affect the heat exchange efficiency between the liquid refrigerant and the cooling medium.

[0054] (5) Excessive oil: Excessive oil refers to excessive lubricating oil in the compressor, which is generally caused by staff errors during installation or maintenance. Lubricating oil mainly plays a role in reducing component wear and transferring heat. However, in the case of excessive lubricating oil, it is easy to cause compressor power loss due to viscosity effect.

[0055] (6) Condenser scaling: If the cooling water itself is very hard or the cooling water quality deteriorates after contact with air, the cooling water will form scale inside the condenser when the temperature changes. When a certain amount of scale forms in the condenser, the effective heat exchange area between the refrigerant and the cooling water will decrease, and the heat exchange efficiency will also decrease.

[0056] (7) Non-condensable gases in the refrigerant: During refrigerant charging, equipment maintenance or unit failure, non-condensable gases such as air or gaseous lubricating oil are likely to be mixed into the refrigerant. As the refrigerant circulates, such gases will accumulate on the inner walls of the evaporator and condenser, which will lead to a reduction in heat exchange area and heat exchange efficiency.

[0057] In some embodiments, see Figure 6 As shown, it is an overall architecture diagram of anomaly detection of an air-conditioning unit provided by an embodiment of the present invention. By inputting the acquired one-dimensional time series parameters of the water temperature meter, flow meter, pressure gauge, voltmeter and ammeter of the air-conditioning unit into the TimesNet convolution model, it is determined whether the data has an anomaly. If not, the data collection continues. If an anomaly exists, the single-point anomaly and continuous anomaly (multi-point continuous, single-point continuous anomaly) output by the TimesNet model (i.e., the anomaly detection model) are respectively input into the fault classifier stacked by three types of base learners: random forest, AdaBoost and XGBoost, and finally the standard anomaly is obtained.

[0058] This method integrates heterogeneous data such as core components of the unit collected by multiple sensors, uses TimesNet network to extract the characteristics of the unit equipment that is operating normally within and between cycles, establishes a time series model for the normal operation of the unit equipment, and performs real-time detection of the multi-sensor operating parameters of the air-conditioning unit. The time series interval where the problem is found is input into the fault classifier based on the joint construction of pattern matcher and multi-classifier to classify the type of fault, the degree of fault, etc. This realizes the abnormal detection and classification of air-conditioning unit equipment, effectively improving the efficiency of equipment operation.

[0059] See also Figure 7 FIG. 1 is a schematic diagram of a structure of an abnormality detection system for an air-conditioning unit provided by an embodiment of the present invention, comprising: A data acquisition module 70, for collecting operating parameters of the air conditioning unit; An anomaly detection module 71 is used to input the operating parameters into a pre-built anomaly detection model and output an abnormal time series parameter in the operating parameters of the air-conditioning unit, wherein the pre-built anomaly detection model is built based on the TimesNet algorithm; The abnormal classification module 72 is used to input the abnormal time series parameters into the fault classifier to obtain the fault classification result of the air-conditioning unit, wherein the fault classifier is constructed based on the combination of the pattern matcher and the multi-classifier.

[0060] On the other hand, the present invention also provides an electronic device, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein when the processor executes the program, the steps of the abnormality detection method for the air-conditioning unit described in any one of the above embodiments are implemented.

[0061] On the other hand, the present invention further provides a non-transitory computer-readable storage medium having a computer program stored thereon, and when the computer program is executed by a processor, the steps of the abnormality detection method for an air-conditioning unit described in any of the above embodiments are implemented.

[0062] Those skilled in the art will appreciate that the embodiments of the present application may be provided as methods, systems, or computer program products. Therefore, the present application may adopt the form of a complete hardware embodiment, a complete software embodiment, or an embodiment combining software and hardware. Moreover, the present application may adopt the form of a computer program product implemented on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program codes.

[0063] The present application is described with reference to the flowcharts and / or block diagrams of the methods, devices (systems), and computer program products according to the embodiments of the present application. It should be understood that each process and / or box in the flowchart and / or block diagram, as well as the combination of the processes and / or boxes in the flowchart and / or block diagram, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, a special-purpose computer, an embedded processor, or other programmable data processing device to generate a machine, so that the instructions executed by the processor of the computer or other programmable data processing device generate instructions for implementing the processes in the flowchart and / or block diagram. Figure 1 A process or multiple processes and / or boxes Figure 1 A device that provides the functions specified in a block or multiple blocks.

[0064] These computer program instructions may also be stored in a computer-readable memory capable of directing a computer or other programmable data processing device to operate in a specific manner, so that the instructions stored in the computer-readable memory produce an article of manufacture comprising an instruction device, which implements the process Figure 1 A process or multiple processes and / or boxes Figure 1 A function specified in one or more boxes.

[0065] These computer program instructions can also be loaded onto a computer or other programmable data processing device so that a series of operating steps are executed on the computer or other programmable device to produce a computer-implemented process, thereby providing instructions for implementing the process. Figure 1 A process or multiple processes and / or boxes Figure 1 The steps for the functions specified in one or more boxes.

[0066] In a typical configuration, a computing device includes one or more processors (CPU), input / output interfaces, network interfaces, and memory.

[0067] The memory may include non-permanent memory in a computer-readable medium, random access memory (RAM) and / or non-volatile memory in the form of read-only memory (ROM) or flash RAM. The memory is an example of a computer-readable medium.

[0068] Computer readable media include permanent and non-permanent, removable and non-removable media that can be implemented by any method or technology to store information. Information can be computer readable instructions, data structures, program modules or other data. Examples of computer storage media include, but are not limited to, phase change memory (PRAM), static random access memory (SRAM), dynamic random access memory (DRAM), other types of random access memory (RAM), read-only memory (ROM), electrically erasable programmable read-only memory (EEPROM), flash memory or other memory technology, compact disk read-only memory (CD-ROM), digital versatile disk (DVD) or other optical storage, magnetic cassettes, magnetic tape magnetic disk storage or other magnetic storage devices or any other non-transmission media that can be used to store information that can be accessed by a computing device. As defined herein, computer readable media does not include temporary computer readable media (transitory media), such as modulated data signals and carrier waves.

[0069] It should also be noted that the terms "include", "comprises" or any other variations thereof are intended to cover non-exclusive inclusion, so that a process, method, commodity or device including a series of elements includes not only those elements, but also other elements not explicitly listed, or also includes elements inherent to such process, method, commodity or device. In the absence of more restrictions, the elements defined by the sentence "comprises a ..." do not exclude the existence of other identical elements in the process, method, commodity or device including the elements.

[0070] The above are only embodiments of the present application and are not intended to limit the present application. For those skilled in the art, the present application may have various changes and variations. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present application should be included within the scope of the claims of the present application.

Claims

1. A method for detecting abnormality of an air conditioning unit, characterized in that: include: Collect the operating parameters of the air conditioning unit; Inputting the operating parameters into a pre-built anomaly detection model, and outputting abnormal time series parameters in the operating parameters of the air-conditioning unit, wherein the pre-built anomaly detection model is built based on the TimesNet algorithm; The abnormal time series parameters are input into a fault classifier to obtain a fault classification result of the air-conditioning unit, wherein the fault classifier is constructed based on a combination of a pattern matcher and a multi-classifier.

2. The abnormality detection method of the air conditioning unit according to claim 1, characterized in that: The operating parameters of the air conditioning unit include but are not limited to the core equipment parameters of the air conditioning unit and the parameters obtained at the inlet and outlet locations.

3. The abnormality detection method for an air conditioning unit according to claim 1, characterized in that: The process of building an anomaly detection model based on the TimesNet algorithm includes: Constructing a data set sample, wherein the data set sample is data of the air-conditioning unit operating normally for multiple consecutive periods of time; Building an anomaly detection model based on the TimesNet algorithm: The air conditioning unit data is converted into two-dimensional data, and a set of two-dimensional tensors at different frequencies and periods is obtained by discrete Fourier transform; The TimesNet model stacks multiple TimesBlocks in a residual manner. For each TimesBlock, the Inception module is used in two-dimensional space to capture various time changes in a two-dimensional tensor set, and various time changes are fused according to the normalized amplitude value to output the information representation in the two-dimensional tensor set. Train the anomaly detection model: Configure the parameters of the anomaly detection model. At the beginning of the training phase of the anomaly detection model, select samples from the data set as the training set to train the anomaly detection model, and select samples from the data set as the verification set to verify the anomaly detection model to obtain the anomaly detection model.

4. The abnormality detection method for an air conditioning unit according to claim 1, characterized in that: The abnormal time series parameters are input into the fault classifier to obtain the fault classification results of the air-conditioning unit, including: The abnormal time series parameters are input into the pattern matcher for feature extraction to obtain the abnormal type of the abnormal time series; For each abnormal time series abnormal type, a multi-classifier is used to determine the fault type corresponding to each abnormal time series abnormal type, wherein the multi-classifier is formed by stacking the random forest algorithm, the AdaBoost algorithm and the XGBoost algorithm.

5. The abnormality detection method of the air conditioning unit according to claim 4, characterized in that: The abnormal time series parameters are input into the pattern matcher for feature extraction to obtain the abnormal time series abnormal types, including: Use sliding windows to extract abnormal time series, and perform regularization on the abnormal time series to obtain regularized abnormal detection fragments; One-dimensional convolution is used to extract features in the anomaly detection segment, and a fully connected layer is used to classify the features to obtain the time series anomaly type.

6. The abnormality detection method of the air conditioning unit according to claim 4, characterized in that: The construction process of the multi-classifier includes the training and prediction of the base learner and the training and prediction of the meta-learner. The base learners include three types of base learners: random forest, AdaBoost and XGBoost. The meta-learner uses a decision tree model: Divide the training data set into a training set and a validation set, train each base learner on the training set, and make predictions on the validation set; Use the trained random forest, AdaBoost and XGBoost three types of base learners to predict the validation set, record the prediction results of each base learner, and combine the prediction results into a new secondary feature set, where each row of the new secondary feature set corresponds to a sample in the validation set, and each column corresponds to the prediction result of a base learner; The new secondary feature set is used as input and the true classification target label is used as output to train the meta-learner and obtain a multi-classifier composed of the random forest algorithm, AdaBoost algorithm and XGBoost algorithm stacked together.

7. The abnormality detection method for an air conditioning unit according to claim 1, characterized in that: The abnormal time series parameters in the operating parameters of the air-conditioning unit include single abnormal points, single-classification continuous abnormal points, and multi-classification continuous abnormal points; among them, a single abnormal point is characterized by the detection of normal data points before and after the abnormal point; a single-classification continuous abnormal point is characterized by the detection of abnormal data points before or after the abnormal point, and the type of the abnormal point before or after the point is the same as the current abnormal point; a multi-classification continuous abnormal point is characterized by the detection of abnormal data points before or after the abnormal point, and the type of the abnormal point before or after the point is different from the current abnormal point.

8. An abnormality detection system for an air conditioning unit, characterized in that: include: Data acquisition module, used to collect operating parameters of the air-conditioning unit; An anomaly detection module, used for inputting the operating parameters into a pre-built anomaly detection model, and outputting an abnormal time series parameter in the operating parameters of the air-conditioning unit, wherein the pre-built anomaly detection model is built based on the TimesNet algorithm; The abnormal classification module is used to input the abnormal time series parameters into the fault classifier to obtain the fault classification result of the air-conditioning unit, wherein the fault classifier is constructed based on the combination of a pattern matcher and a multi-classifier.

9. An electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that: When the processor executes the program, the steps of the abnormality detection method for the air conditioning unit as described in any one of claims 1 to 5 are implemented.

10. A non-transitory computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the steps of the abnormality detection method for an air-conditioning unit according to any one of claims 1 to 5 are implemented.

Images