Water chilling unit fault detection and classification method, system and device and medium

Abstract

The invention provides a water chilling unit fault detection and classification method, system and device and a medium, and the method comprises the steps: collecting the operation data of a water chilling unit, and recording the operation data as characteristic variables; preprocessing the feature variables, extracting fault detection feature variables in the feature variables to generate time sequence data, and storing the time sequence data in a time sequence database; collecting time sequence data in the normal operation period of the water chilling unit in the time sequence database, and constructing a training sample set; carrying out algorithm modeling on the time sequence data by using a MoE algorithm to generate a MoE model, and carrying out parameter configuration; based on the training sample set, using an Adam optimizer and a CosineLR scheduler to train a MoE model; acquiring real-time time sequence data of the water chilling unit, inputting the real-time time sequence data into the MoE model, and identifying abnormal time sequence data; and performing unsupervised classification on the abnormal time sequence data by using a K-means clustering algorithm to determine the fault type of the water chilling unit.

Description

Technical Field

[0001] The present invention relates to the technical field of predictive maintenance of chiller equipment, and more particularly to a method, system, device and medium for detecting and classifying chiller faults. Background Art

[0002] With the development of smart buildings and green buildings, more and more HVAC systems are integrated into building automation management systems. However, due to the complex working environment of chillers, they are easily affected by many factors, such as load changes, ambient temperature fluctuations, improper operation, etc., resulting in complex and diverse failure modes. Traditional maintenance methods often rely on expert experience for regular inspections and troubleshooting. This method is not only costly and time-consuming, but also prone to missed faults or misjudgments due to different expert levels. In addition, traditional rule-based methods are difficult to deal with complex multi-condition faults, and it is often difficult to perform predictive maintenance on chiller faults in a timely and accurate manner.

[0003] In recent years, with the advancement of the Internet of Things and big data technology, real-time data collection of chillers has become easier, and a large amount of sensor data provides a data basis for fault detection and diagnosis of chillers. However, due to the imbalance of normal and fault state samples of chillers, many fault samples are relatively scarce in practical applications. The traditional supervised learning method requires a large amount of labeled data, which is difficult to meet the actual needs of engineering. Therefore, how to use a large amount of unlabeled data to classify chiller faults through unsupervised learning methods has become a hot issue in current research.

[0004] In the existing technology, the fault detection and classification of chillers are mostly based on traditional machine learning or deep learning models, and common methods include support vector machine (SVM), random forest (RF), long short-term memory network (LSTM), etc. Although these methods have improved the performance of fault detection to a certain extent, they often have problems such as insufficient generalization ability and overly strict assumptions on data distribution. Summary of the invention

[0005] In view of the above problems, the purpose of the present invention is to provide a method, system, device and medium for chiller fault detection and classification, which uses the MoE model to realize unsupervised fault classification of chillers, avoids the dilemma of insufficient labeled data, and improves the timeliness and accuracy of fault detection.

[0006] In order to achieve the above object, the present invention is implemented through the following technical solutions: In a first aspect, the present invention discloses a chiller fault detection and classification method, comprising: Collect the operating data of the chiller and record it as characteristic variables; pre-process the characteristic variables, extract the fault detection characteristic variables to generate time series data, and store them in the time series database; Collect the time series data of the chiller during normal operation in the time series database to construct a training sample set; Use the MoE algorithm to perform algorithmic modeling on time series data, generate a MoE model, and configure model parameters; Based on the training sample set, the MoE model is trained using the Adam optimizer and the CosineLR scheduler; Obtain the real-time time series data of the chiller in the time series database, input it into the MoE model, and identify abnormal time series data; The K-means clustering algorithm is used to perform unsupervised classification on abnormal time series data to determine the fault type of the chiller.

[0007] Further, the operation data of the chiller is collected and recorded as characteristic variables; the characteristic variables are preprocessed, fault detection characteristic variables are extracted therein to generate time series data, and stored in a time series database, including: The operating data of the chiller is collected by sensors installed in the refrigeration circuit and recorded as a preset number of characteristic variables; the characteristic variables include pressure, cooling capacity, temperature and power; Perform data cleaning and data normalization on feature variables; Extract 16 characteristic variables for fault detection from the characteristic variables and generate time series data; The six characteristic variables for fault detection include: evaporator water inlet temperature, evaporator water outlet temperature, condenser water inlet temperature, condenser water outlet temperature, compressor power, evaporator temperature difference characterization value, condenser temperature difference characterization value, evaporation temperature, condensation temperature, refrigerant subcooling, compressor suction temperature, compressor suction superheat, compressor exhaust temperature, compressor exhaust superheat, oil tank temperature and oil supply pressure; The time series data is uploaded to the IoT platform via ModBus and MQTT protocols, and the time series data is stored in the built-in time series database via the IoT platform.

[0008] Furthermore, the time series data during the normal operation of the chiller is collected in the time series database to construct a training sample set, including: The time series data of the chiller during normal operation is collected in the time series database. According to the corresponding equipment, the time series data of the same equipment in different time periods are spliced, and the data is smoothed to generate a training sample set.

[0009] Furthermore, the MoE algorithm is used to perform algorithm modeling on the time series data, generate a MoE model, and configure model parameters, including: Time series data are time-aligned to generate multivariate time series data with T consecutive timestamps and K metrics. The data consists of an ordered sequence, which is specifically expressed as follows:

[0010] in, Defined as a specific timestamp The set of observed values ​​of all metrics at time : , in the collection In each data point Indicates The value of a metric at that point in time.

[0011] Construct the MoE model, which includes an expert network, a gating mechanism, and a tower network; The MoE model The tasks are represented as:

[0012] in, represents the dot product, represents the set of intermediate results generated by experts, represents the gated structure, Indicates The model of downstream tasks is called tower network.

[0013] The temporal dependency and cross-metric dependency are extracted by configuring the convolutional layers of the MoE model.

[0014] A dual gating mechanism is set up for the MoE model by combining shared gating and personalized gating.

[0015] Further, the MoE model is trained based on the training sample set using the Adam optimizer and the CosineLR scheduler, including: Generate a training set and a validation set based on the training sample set; Set the training parameters, including window size, number of expert networks, horizon value, number of cores in each expert network, and weight coefficient of shared gating; Train the MoE model using the Adam optimizer and the CosineLR scheduler; During training, the initial learning rate is set to 0.001. The batch size is set to 128. The maximum number of training rounds is set to 10.

[0016] Furthermore, the real-time time series data of the chiller is obtained from the time series database, inputted into the MoE model, and abnormal time series data is identified, including: Obtain the real-time time series data of the chiller in the time series database and input it into the MoE model to generate prediction results; When the prediction result is 0, the corresponding time series data is normal; When the prediction result is 1, the corresponding time series data is abnormal.

[0017] Further, the K-means clustering algorithm is used to perform unsupervised classification on the abnormal time series data to determine the fault type of the chiller, including: The chiller faults are set into eight categories, including reduced chilled water flow, reduced cooling water flow, excessive refrigerant charge, non-condensable gas, refrigerant leakage, excessive lubricating oil and regulating valve failure; Based on eight types of faults, the initial number of centroids of the K-means clustering algorithm is set to k=8; Generate time series data sets based on abnormal time series data , where each time series data corresponds to an abnormal feature; For each pair of abnormal features ( ), and use the dynamic time warping algorithm to calculate the distance between them; in, ; Constructing the local distance matrix ,in and They are time series data and In the and The value at a time point, Use dynamic programming to find the alignment path for two time series data and select the path with the minimum cost; The cumulative minimum distance matrix is ​​recursively calculated by the following formula :

[0018] Among them, the lower right corner value of the matrix Representing time series data and The DTW distance between them; Calculate the DTW distance between all abnormal time series data and construct a The distance matrix ,in Indicates an exception With exception The DTW distance between them; Use multidimensional scaling to convert the distance matrix Convert to low-dimensional Euclidean space coordinates; The K-means clustering algorithm is applied to perform unsupervised classification based on the low-dimensional Euclidean space coordinates, and according to the result of the K-means clustering algorithm, the corresponding abnormal features are divided into eight categories.

[0019] In a second aspect, the present invention further discloses a chiller fault detection and classification system, comprising: The multi-sensor data processing module is used to collect the operating data of the chiller and record it as characteristic variables; pre-process the characteristic variables, extract the fault detection characteristic variables therein to generate time series data, and store it in the time series database; The data set sample construction module is used to collect the time series data of the chiller during normal operation in the time series database and construct a training sample set; The model building module is used to use the MoE algorithm to perform algorithmic modeling on time series data, generate the MoE model, and configure the model parameters; Model training module, used to train the MoE model based on the training sample set using the Adam optimizer and the CosineLR scheduler; The online anomaly detection module is used to obtain the real-time time series data of the chiller in the time series database, input it into the MoE model, and identify abnormal time series data; The unsupervised fault classification module is used to perform unsupervised classification on abnormal time series data using the K-means clustering algorithm to determine the fault type of the chiller.

[0020] In a third aspect, the present invention further discloses a chiller fault detection and classification device, comprising: a memory for storing a chiller fault detection and classification program; A processor is used to implement the steps of the chiller fault detection and classification method as described in any one of the above when executing the chiller fault detection and classification program.

[0021] In a fourth aspect, the present invention further discloses a readable storage medium, on which a chiller fault detection and classification program is stored. When the chiller fault detection and classification program is executed by a processor, the steps of the chiller fault detection and classification method as described in any one of the above items are implemented.

[0022] Compared with the prior art, the present invention has the following beneficial effects: 1. The present invention collects real-time data from various sensors of the chiller and pre-processes the data, including data cleaning, normalization, feature extraction, etc., to ensure the quality and consistency of the data. These data provide a reliable basis for subsequent anomaly detection and fault classification.

[0023] 2. The present invention uses multiple expert models to model the status of the chiller under different working conditions. Through the gated network, the most appropriate expert model is adaptively selected according to the characteristics of the input data to improve the accuracy of anomaly detection. This module can detect anomalies in the chiller in a timely manner and provide input data for subsequent fault classification.

[0024] 3. After detecting the abnormality of the chiller, the present invention uses the KMeans clustering algorithm to classify the abnormal data to determine the type of fault. This unsupervised learning method can effectively solve the problem of insufficient fault samples and improve the flexibility and generalization ability of fault classification.

[0025] 4. The present invention can fully explore the potential relationship of the multi-sensor time series data generated inside the chiller, improve the accuracy and real-time performance of abnormal detection of chiller equipment, thereby reducing the equipment failure time of the air-conditioning chiller, reducing the energy consumption of the equipment, and increasing the service life of the equipment.

[0026] It can be seen that compared with the prior art, the present invention has outstanding substantive features and significant progress, and the beneficial effects of its implementation are also obvious. BRIEF DESCRIPTION OF THE DRAWINGS

[0027] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. Obviously, the drawings described below are only embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on the provided drawings without paying creative work.

[0028] Figure 1 It is a method flow chart of a specific implementation mode of the present invention.

[0029] Figure 2 It is a network structure diagram of the MoE model according to a specific implementation manner of the present invention.

[0030] Figure 3 It is a system structure diagram of a specific implementation mode of the present invention.

[0031] In the figure, 1. Multi-sensor data processing module; 2. Data set sample construction module; 3. Model construction module; 4. Model training module; 5. Online anomaly detection module; 6. Unsupervised fault classification module. DETAILED DESCRIPTION

[0032] In order to enable those skilled in the art to better understand the scheme of the present invention, the present invention is further described in detail below in conjunction with the accompanying drawings and specific implementation methods. Obviously, the described embodiments are only part of the embodiments of the present invention, rather than all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative work are within the scope of protection of the present invention.

[0033] See also Figure 1 As shown, this embodiment provides a chiller fault detection and classification method, comprising the following steps: S1: Collect the operating data of the chiller and record it as feature variables; preprocess the feature variables, extract the fault detection feature variables to generate time series data, and store them in the time series database.

[0034] In a specific implementation, the operation data of the chiller is collected by various sensors installed in the refrigeration circuit and recorded as 64 characteristic variables; the characteristic variables include parameters such as pressure, cooling capacity, temperature and power. Among them, 48 items are parameters directly collected by sensors, and 16 items are parameters calculated based on real-time data. Among these 64 characteristic variables, 16 are particularly important for fault detection. The specific characteristic variables are described in Table 1 below: Table 1: Chiller parameter list

[0035] After the collection is completed, the characteristic variables are cleaned and normalized. Referring to Table 1, 16 characteristic variables for fault detection are extracted from the characteristic variables, and time series data are generated; At this point, the time series data is uploaded to the Internet of Things (IoT) platform through the ModBus and MQTT protocols. The IoT platform plays a key role in the data management process. It not only connects and manages various sensor devices, but is also responsible for the initial processing and routing of data.

[0036] The IoT platform then stores the collected data in a time series database, which is specially designed to process data arranged in chronological order and can efficiently store and query a large amount of air conditioning unit equipment data.

[0037] This step achieves efficient coordination of data collection and storage by establishing a complete data link from sensors to IoT platform and then to time series database. This mechanism not only ensures the integrity and real-time nature of the data, but also provides basic support for subsequent chiller anomaly detection and fault classification.

[0038] S2: Collect the time series data of the chiller during normal operation in the time series database and construct a training sample set.

[0039] In a specific implementation, the training input data is mainly the normal operation data of the unit for multiple consecutive periods of time. This type of normal operation data has the following characteristics: 1. The data comes from the time series data of each specific device in the chiller; 2. The frequency of data collection is different, so the amount and characteristics of data in each time series may vary greatly, and the starting and ending points of each data segment may change significantly.

[0040] In view of the characteristics of the data set samples, this method adopts the method of data splicing and aggregation when constructing the training sample set. Specifically, by directly splicing the time series data of the same device in different time periods and performing smoothing at the splicing point, the impact of data mutation on the model is effectively avoided, thereby generating a higher quality training sample set. The advantage of this solution is that the sample size is significantly increased after data splicing, allowing the MoE network to more fully learn the operation mode of the device at different time periods.

[0041] S3: Use the MoE algorithm to perform algorithmic modeling on time series data, generate a MoE model, and configure model parameters.

[0042] In a specific implementation, algorithm modeling is performed on time series data of 47 sensors based on MoE.

[0043] First, 47 time series data are time-aligned, each of which represents a continuous observation of a long-term metric. The multivariate time series data with T consecutive timestamps and K metrics is generated by an ordered sequence, which is specifically represented as follows:

[0044] in, Defined as a specific timestamp The set of observed values ​​of all metrics at time : , in the collection In each data point Indicates The value of a metric at that point in time.

[0045] Then, construct the MoE model as Figure 2 As shown, the MoE model includes an expert network, a gating mechanism, and a tower network.

[0046] The MoE model uses ensemble learning technology to introduce multiple gating modules, each of which corresponds to a task with a feature, so that it can use shared embeddings while dealing with complex task correlation issues. This can also make the network more sparse, so that more parameters can be introduced without increasing computational overhead. The kth task of this MoE model can be expressed as:

[0047] in, represents the dot product, represents the set of intermediate results generated by experts, represents the gated structure, Indicates The model of downstream tasks is called tower network.

[0048] The temporal dependency and cross-metric dependency are extracted by configuring the convolutional layers of the MoE model.

[0049] refer to Figure 2 It can be seen that the expert network uses convolutional layers to extract temporal dependencies and cross-metric dependencies to generate a set of candidate embeddings. The gating mechanism includes shared gating and personalized gating, which are used to map inputs to the weights of candidate embeddings. The tower network compresses the weighted candidate embeddings into the final prediction output.

[0050] In a typical multi-indicator MoE work scenario, there is no obvious correlation between the downstream task and the input data. That is, each dimension of the input data is equally important to the current task, so each gate calculates the expert's weighted combination based on all the input data. However, this assumption does not hold true in multi-time series anomaly detection.

[0051] In this case, for the subtask of judging whether a specific indicator is abnormal, it is unreasonable to give equal weight to the historical information in its own feature space and the information in the overall input feature space composed of other indicators. More indicators mean more factors that may interfere with the limited capacity structure, especially when there are conflicts and reasonable drifts between indicators.

[0052] To address this issue, this step splits the input at the indicator level, using its own local time window instead of the window of all indicators to facilitate the gating structure to learn a personalized mapping relationship from time series to expert distribution.

[0053] The embedding of the original input 𝐸 at time 𝑡 consists of the outputs of all expert networks:

[0054] in, Indicates An expert network is given an input The outputs of these expert networks constitute the embedded representation of the original input.

[0055] Embed Based on the gate Given the weights, embed the original input Combination of:

[0056] in, is a local time window based on the kth indicator The calculated gating weight, ⊙, represents the element-level multiplication. This personalized gating based on indicators can better solve the conflict and drift problems between indicators.

[0057] In addition, it should be noted that in the MoE model, the convolution layer contains N convolution kernels with a width of l (equal to the window size) and a height of 1. When a convolution kernel sweeps over the input w_t containing all K indicators, the window of each indicator is convolved into a single value, so the layer produces a vector of size N×1×K. In this way, we obtain temporal dependencies through indicator-level convolution operations, and dependencies between indicators through shared filters. The vector is then flattened and further input into a two-layer fully connected network to form an embedding of the original input, and all embeddings constitute a candidate set.

[0058] If the patterns between indicators are very different, different indicators can utilize different expert networks, which is reflected by the weight combination given by the gating. This allows the model to resolve conflicts and irrelevant drifts between indicators. In addition, a dual gating mechanism is designed. Shared gating Receive all selected windows, while personalizing the gating Only receive The shared gating has a window of indicators. Due to the large number of indicators, if all gates have the same weight, it will make the back-propagation gradient more confusing for the expert network. Shared gating has a larger weight and can learn a robust mapping relationship between expert fusion and input from more data, which is more likely to guide the expert to converge to the dominant feature. This method combines shared gating and personalized gating, taking advantage of both, so that personalized gating plays an auxiliary role in fine-tuning the subtle differences between indicators.

[0059]

[0060]

[0061] in, and ( = 1, 2, ..., K) is a trainable matrix, >0.5 is the weight coefficient of shared gating. This hybrid gating mechanism can help the model better resolve conflicts and irrelevant drift issues between indicators.

[0062] In the framework of this MoE model, the purpose of the tower network is to compress the embedding into a final prediction value. We use two dense layers with ReLU activation function and a dropout layer to reduce the dimensionality. The predictions of all towers are added to the set Since the true observation value at time t is ,The final training objective L aims to minimize the squared L2 norm:

[0063] in, represents the L2 norm. It is also used in the inference phase as anomaly score and further compared with the threshold set in the previous online detection scenario.

[0064] S4: Based on the training sample set, the MoE model is trained using the Adam optimizer and the CosineLR scheduler.

[0065] In a specific implementation, the MoE model training strategy is as follows: The window size is set to 16, the number of expert networks is set to 9, and the horizon is set to 1. The number of cores in each expert network is set to 16. The weight coefficient ε of the shared gating is set to 0.7.

[0066] The MoE model is trained using the Adam optimizer and the CosineLR scheduler, with an initial learning rate of 0.001. The batch size is set to 128. The maximum number of training rounds is set to 10. At the beginning of the model training phase, 80% of the dataset is selected as the training set and 20% of the dataset is selected as the validation set for model training.

[0067] S5: Obtain the real-time time series data of the chiller in the time series database, input it into the MoE model, and identify abnormal time series data.

[0068] In a specific implementation, real-time time series data of the chiller is obtained from a time series database and input into the MoE model to generate a prediction result; because the MoE model has modeled and learned the time series laws of the normal operating status of the central air-conditioning unit during the training phase, it can predict and output a detection result of whether there is an abnormality.

[0069] Normally, if the prediction result is normal, it is represented as "0", and if the result is abnormal, it is represented as "1".

[0070] S6: Use K-means clustering algorithm to perform unsupervised classification on abnormal time series data to determine the fault type of the chiller.

[0071] In a specific implementation, the chiller fault is first set to eight types of faults, including: (1) Refrigerant leakage: Refrigerant leakage will cause the evaporation temperature to drop, the COP and cooling capacity of the unit to decrease, the compressor power to decrease, and the condensing temperature and heat exchange efficiency to decrease.

[0072] (2) Excessive refrigerant filling: Excessive refrigerant will reduce the effective heat exchange area of ​​the condenser, resulting in increased condensation temperature and pressure, increased compressor power, increased exhaust pressure, and prone to liquid hammer, which will damage the compressor.

[0073] (3) Insufficient chilled water flow: Improper water pump selection or system failure will cause the chilled water inlet and outlet temperature difference to increase, the evaporation temperature and pressure to decrease, and the heat exchange efficiency to decrease.

[0074] (4) Insufficient cooling water flow: Improper water pump selection or system problems will cause the condensation temperature and pressure to rise, and the condensation efficiency to decrease.

[0075] (5) Condenser scaling: Long-term scaling will increase the heat exchange resistance, reduce the condenser efficiency, increase the condensation temperature and pressure, and reduce the cooling water temperature difference.

[0076] (6) Non-condensable gases: Non-condensable gases in the condenser will increase pressure, reduce heat exchange area, cause condensation temperature and pressure to increase, and reduce system COP.

[0077] (7) Excessive lubricating oil: Excessive lubricating oil will reduce the suction pressure of the compressor, resulting in a decrease in the system COP.

[0078] (8) Regulating valve failure.

[0079] This step uses the dynamic time warping (DTW) algorithm to calculate the distance between these time series data. DTW can effectively solve the errors caused by the asynchronous occurrence time of the fault. It measures the similarity of two time series through nonlinear alignment, and is particularly suitable for processing the time offset or deformation that may exist in the time series data.

[0080] The specific steps are as follows: Step 1: Feature extraction and dimensionality reduction. First, generate a time series data set based on the abnormal time series data. , each time series data corresponds to an abnormal feature.

[0081] Step 2: Calculate the DTW distance of each pair of anomalies. For each pair of anomaly features ( ), and their DTW distance is calculated by the dynamic time warping algorithm. Specifically, the abnormal feature is defined as , because the failure may occur at different time points, directly comparing the two time series may produce deviations. The DTW algorithm finds their similarities by calculating the optimal alignment path between the time series. Then, the local distance matrix is ​​constructed ,in and Sequential and In the and The value of each time point, Euclidean distance is used to measure the local distance. The alignment path of the two sequences is solved by dynamic programming method. The path with the minimum cost is selected, and the cumulative minimum distance matrix is ​​recursively calculated. :

[0082] Finally, the value in the lower right corner of the matrix Representing time series and The DTW distance between them.

[0083] Then build a distance matrix: calculate the DTW distance between all anomalies and build a The distance matrix ,in Indicates an exception With exception The DTW distance between them.

[0084] Step 3: Unsupervised classification. By calculating the DTW distance matrix, all anomalies can be classified unsupervised. After obtaining the low-dimensional Euclidean space coordinates through MDS, the K-means clustering algorithm is applied to perform unsupervised classification of these anomaly features. The core steps of the K-means clustering algorithm are as follows: A. Initial centroid selection: K initial centroids are randomly selected. In this patent, there are 8 types of faults, so K=8 here.

[0085] B. Calculate distances: For each data point, calculate the distance between it and each centroid and assign the data point to the centroid with the closest distance.

[0086] C Update the centroid: Calculate the centroid of each cluster, that is, the mean of all data points in the cluster, and then update the position of the centroid.

[0087] D Repeated iteration: Repeat the process of calculating the distance and updating the center of mass until the center of mass position no longer changes or the preset maximum number of iterations is reached.

[0088] Through the above steps, according to the results of K-means clustering algorithm, the corresponding abnormal features are divided into eight categories.

[0089] The present invention provides a method for anomaly detection and unsupervised fault classification of chillers based on MoE, by collecting real-time data from various sensors of the chiller and performing preprocessing, including data cleaning, normalization and feature extraction. Multiple expert models are used to model the state of the chiller under various working conditions, and a gated network is used to improve the accuracy of detection. Once an anomaly is detected, the K-means clustering algorithm is used to classify the fault, overcoming the challenge of scarce labeled fault data and improving the generalization ability. This method improves the accuracy and timeliness of anomaly detection of chillers, reduces downtime, reduces energy consumption, extends equipment life, and provides an effective predictive maintenance solution.

[0090] See also Figure 3 As shown, the present invention also discloses a chiller fault detection and classification system, including: a multi-sensor data processing module 1, a data set sample construction module 2, a model construction module 3, a model training module 4, an online anomaly detection module 5 and an unsupervised fault classification module 6. The multi-sensor data processing module 1 is used to collect the operating data of the chiller and record it as characteristic variables; pre-process the characteristic variables, extract the fault detection characteristic variables therein to generate time series data, and store it in a time series database.

[0091] The data set sample construction module 2 is used to collect the time series data of the chiller during normal operation in the time series database to construct a training sample set.

[0092] The model building module 3 is used to use the MoE algorithm to perform algorithmic modeling on the time series data, generate the MoE model, and configure the model parameters.

[0093] Model training module 4 is used to train the MoE model based on the training sample set using the Adam optimizer and the CosineLR scheduler.

[0094] The online anomaly detection module 5 is used to obtain the real-time time series data of the chiller in the time series database, input it into the MoE model, and identify abnormal time series data.

[0095] The unsupervised fault classification module 6 is used to perform unsupervised classification on the abnormal time series data using a K-means clustering algorithm to determine the fault type of the chiller.

[0096] The specific implementation of the chiller fault detection and classification system of this embodiment is basically consistent with the specific implementation of the chiller fault detection and classification method described above, and will not be repeated here.

[0097] The present invention also discloses a chiller fault detection and classification device, comprising a processor and a memory; wherein the processor implements the steps of the chiller fault detection and classification method as described in any one of the above items when executing the chiller fault detection and classification program stored in the memory.

[0098] Furthermore, the chiller fault detection and classification device in this embodiment may also include: The input interface is used to obtain the chiller fault detection and classification program imported from the outside, and save the obtained chiller fault detection and classification program to the memory, and can also be used to obtain various instructions and parameters transmitted by external terminal devices, and transmit them to the processor so that the processor can use the above various instructions and parameters to carry out corresponding processing. In this embodiment, the input interface can specifically include but is not limited to a USB interface, a serial interface, a voice input interface, a fingerprint input interface, a hard disk reading interface, etc.

[0099] The output interface is used to output various data generated by the processor to the terminal device connected thereto, so that other terminal devices connected to the output interface can obtain various data generated by the processor. In this embodiment, the output interface may specifically include but is not limited to a USB interface, a serial interface, etc.

[0100] The communication unit is used to establish a remote communication connection between the chiller fault detection and classification device and the external server so that the chiller fault detection and classification device can mount the image file to the external server. In this embodiment, the communication unit may specifically include but is not limited to a remote communication unit based on wireless communication technology or wired communication technology.

[0101] The keyboard is used to obtain various parameter data or instructions input by the user by tapping the keycaps in real time.

[0102] Display for real-time display of relevant information on the chiller fault detection and classification process.

[0103] The mouse can be used to assist users in inputting data and simplify user operations.

[0104] The present invention also discloses a readable storage medium, wherein the readable storage medium includes a random access memory (RAM), a memory, a read-only memory (ROM), an electrically programmable ROM, an electrically erasable programmable ROM, a register, a hard disk, a removable hard disk, a CD-ROM or any other form of storage medium known in the technical field. The readable storage medium stores a chiller fault detection and classification program, and when the chiller fault detection and classification program is executed by a processor, the steps of the chiller fault detection and classification method described in any one of the above items are implemented.

[0105] In summary, the present invention uses the MoE model to realize unsupervised fault classification of chillers, which can avoid the dilemma of insufficient labeled data and improve the timeliness and accuracy of fault detection.

[0106] In this specification, each embodiment is described in a progressive manner, and each embodiment focuses on the differences from other embodiments. The same or similar parts between the embodiments can be referred to each other. As for the method disclosed in the embodiment, since it corresponds to the system disclosed in the embodiment, the description is relatively simple, and the relevant parts can be referred to the method part.

[0107] Professionals may further appreciate that the units and algorithm steps of each example described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, computer software, or a combination of the two. In order to clearly illustrate the interchangeability of hardware and software, the composition and steps of each example have been generally described in the above description according to function. Whether these functions are performed in hardware or software depends on the specific application and design constraints of the technical solution. Professionals and technicians may use different methods to implement the described functions for each specific application, but such implementation should not be considered to be beyond the scope of the present invention.

[0108] In the several embodiments provided by the present invention, it should be understood that the disclosed systems, systems and methods can be implemented in other ways. For example, the system embodiments described above are only schematic. For example, the division of the units is only a logical function division. There may be other division methods in actual implementation, such as multiple units or components can be combined or integrated into another system, or some features can be ignored or not executed. Another point is that the mutual coupling or direct coupling or communication connection shown or discussed can be through some interfaces, indirect coupling or communication connection of systems or units, which can be electrical, mechanical or other forms.

[0109] The units described as separate components may or may not be physically separated, and the components shown as units may or may not be physical units, that is, they may be located in one place or distributed on multiple network units. Some or all of the units may be selected according to actual needs to achieve the purpose of the solution of this embodiment.

[0110] In addition, each functional module in each embodiment of the present invention may be integrated into one processing unit, or each module may exist physically separately, or two or more modules may be integrated into one unit.

[0111] Similarly, each processing unit in each embodiment of the present invention may be integrated into one functional module, or each processing unit may exist physically, or two or more processing units may be integrated into one functional module.

[0112] The steps of the method or algorithm described in conjunction with the embodiments disclosed herein may be implemented directly using hardware, a software module executed by a processor, or a combination of the two. The software module may be placed in a random access memory (RAM), a memory, a read-only memory (ROM), an electrically programmable ROM, an electrically erasable programmable ROM, a register, a hard disk, a removable disk, a CD-ROM, or any other form of storage medium known in the art.

[0113] Finally, it should be noted that, in this article, relational terms such as first and second, etc. are only used to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Moreover, the terms "include", "comprise" or any other variants thereof are intended to cover non-exclusive inclusion, so that a process, method, article 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, article or device. In the absence of further restrictions, the elements defined by the sentence "comprise a ..." do not exclude the presence of other identical elements in the process, method, article or device including the elements.

[0114] The above is a detailed introduction to the chiller fault detection and classification method, system, device and readable storage medium provided by the present invention. Specific examples are used herein to illustrate the principles and implementation methods of the present invention. The description of the above embodiments is only used to help understand the method of the present invention and its core idea. It should be pointed out that for ordinary technicians in this technical field, without departing from the principles of the present invention, several improvements and modifications can be made to the present invention, and these improvements and modifications also fall within the scope of protection of the present invention.

Claims

1. A chiller fault detection and classification method, characterized in that: include: Collect the operating data of the chiller and record it as characteristic variables; Preprocess the characteristic variables, extract the fault detection characteristic variables to generate time series data, and store them in the time series database; Collect the time series data of the chiller during normal operation in the time series database to construct a training sample set; Use the MoE algorithm to perform algorithmic modeling on time series data, generate a MoE model, and configure model parameters; Based on the training sample set, the MoE model is trained using the Adam optimizer and the CosineLR scheduler; Obtain the real-time time series data of the chiller in the time series database, input it into the MoE model, and identify abnormal time series data; The K-means clustering algorithm is used to perform unsupervised classification on abnormal time series data to determine the fault type of the chiller.

2. The chiller fault detection and classification method according to claim 1, characterized in that: The operation data of the chiller is collected and recorded as characteristic variables; Preprocess the characteristic variables, extract the fault detection characteristic variables to generate time series data, and store them in the time series database, including: The operating data of the chiller is collected by sensors installed in the refrigeration circuit and recorded as a preset number of characteristic variables; the characteristic variables include pressure, cooling capacity, temperature and power; Perform data cleaning and data normalization on feature variables; Extract 16 characteristic variables for fault detection from the characteristic variables and generate time series data; The six characteristic variables for fault detection include: evaporator water inlet temperature, evaporator water outlet temperature, condenser water inlet temperature, condenser water outlet temperature, compressor power, evaporator temperature difference characterization value, condenser temperature difference characterization value, evaporation temperature, condensation temperature, refrigerant subcooling, compressor suction temperature, compressor suction superheat, compressor exhaust temperature, compressor exhaust superheat, oil tank temperature and oil supply pressure; The time series data is uploaded to the IoT platform via ModBus and MQTT protocols, and the time series data is stored in the built-in time series database via the IoT platform.

3. The chiller fault detection and classification method according to claim 2, characterized in that: The time series data during the normal operation of the chiller are collected in the time series database to construct a training sample set, including: The time series data of the chiller during normal operation is collected in the time series database. According to the corresponding equipment, the time series data of the same equipment in different time periods are spliced, and the data is smoothed to generate a training sample set.

4. The chiller fault detection and classification method according to claim 3, characterized in that: The MoE algorithm is used to perform algorithmic modeling on time series data, generate a MoE model, and configure model parameters, including: Time series data are time-aligned to generate multivariate time series data with T consecutive timestamps and K metrics. The data consists of an ordered sequence, which is specifically expressed as follows: in, Defined as a specific timestamp The set of observed values ​​of all metrics at time : , in the collection In each data point Indicates The value of a metric at that point in time; Construct the MoE model, which includes an expert network, a gating mechanism, and a tower network; The MoE model The tasks are represented as: in, represents the dot product, represents the set of intermediate results generated by experts, represents the gated structure, Indicates A model for downstream tasks, namely a tower network; Extract temporal dependency and cross-metric dependency by configuring the convolutional layers of the MoE model; A dual gating mechanism is set up for the MoE model by combining shared gating and personalized gating.

5. The chiller fault detection and classification method according to claim 4, characterized in that: The MoE model is trained based on the training sample set using the Adam optimizer and the CosineLR scheduler, including: Generate a training set and a validation set based on the training sample set; Set the training parameters, including window size, number of expert networks, horizon value, number of cores in each expert network, and weight coefficient of shared gating; Train the MoE model using the Adam optimizer and the CosineLR scheduler; During training, the initial learning rate is set to 0.001, the batch size is set to 128, and the maximum number of training rounds is set to 10.

6. The chiller fault detection and classification method according to claim 5, characterized in that: The real-time time series data of the chiller is obtained from the time series database, inputted into the MoE model, and abnormal time series data is identified, including: Obtain the real-time time series data of the chiller in the time series database and input it into the MoE model to generate prediction results; When the prediction result is 0, the corresponding time series data is normal; When the prediction result is 1, the corresponding time series data is abnormal.

7. The chiller fault detection and classification method according to claim 6, characterized in that: The K-means clustering algorithm is used to perform unsupervised classification on abnormal time series data to determine the fault type of the chiller, including: The chiller faults are set into eight categories, including reduced chilled water flow, reduced cooling water flow, excessive refrigerant charge, non-condensable gas, refrigerant leakage, excessive lubricating oil and regulating valve failure; Based on eight types of faults, the initial number of centroids of the K-means clustering algorithm is set to k=8; Generate time series data sets based on abnormal time series data , where each time series data corresponds to an abnormal feature; For each pair of abnormal features ( ), and use the dynamic time warping algorithm to calculate the distance between them; in, ; Constructing the local distance matrix ,in and They are time series data and In the and The value at a time point, Use dynamic programming to find the alignment path for two time series data and select the path with the minimum cost; The cumulative minimum distance matrix is ​​recursively calculated by the following formula : Among them, the lower right corner value of the matrix Representing time series data and The DTW distance between them; Calculate the DTW distance between all abnormal time series data and construct a The distance matrix ,in Indicates an exception With exception The DTW distance between them; Use multidimensional scaling to convert the distance matrix Convert to low-dimensional Euclidean space coordinates; The K-means clustering algorithm is applied to perform unsupervised classification based on the low-dimensional Euclidean space coordinates, and according to the result of the K-means clustering algorithm, the corresponding abnormal features are divided into eight categories.

8. A chiller fault detection and classification system, characterized in that: The system adopts the chiller fault detection and classification method as claimed in any one of claims 1 to 7; The system comprises: The multi-sensor data processing module is used to collect the operating data of the chiller and record it as characteristic variables; pre-process the characteristic variables, extract the fault detection characteristic variables therein to generate time series data, and store it in the time series database; The data set sample construction module is used to collect the time series data of the chiller during normal operation in the time series database and construct a training sample set; The model building module is used to use the MoE algorithm to perform algorithmic modeling on time series data, generate the MoE model, and configure the model parameters; Model training module, used to train the MoE model based on the training sample set using the Adam optimizer and the CosineLR scheduler; The online anomaly detection module is used to obtain the real-time time series data of the chiller in the time series database, input it into the MoE model, and identify abnormal time series data; The unsupervised fault classification module is used to perform unsupervised classification on abnormal time series data using the K-means clustering algorithm to determine the fault type of the chiller.

9. A chiller fault detection and classification device, characterized in that: include: a memory for storing a chiller fault detection and classification program; A processor is used to implement the steps of the chiller fault detection and classification method as described in any one of claims 1 to 7 when executing the chiller fault detection and classification program.

10. A readable storage medium, characterized in that: The readable storage medium stores a chiller fault detection and classification program, which, when executed by a processor, implements the steps of the chiller fault detection and classification method according to any one of claims 1 to 7.

Images