Multi-link and fault detection method and device, storage medium

Abstract

The present invention relates to the field of fault detection technology, and specifically discloses a multi-split machine and its fault detection method and device, and a storage medium, wherein the multi-split machine fault detection method comprises the following steps: traversing the operation data set accumulated by the multi-split machine day by day, so as to store the operation data set into a corresponding data pool day by day according to the operation conditions of the multi-split machine; when the amount of data stored in the data pool reaches a lower limit length, training the data in the data pool to obtain a clustering model, an autoencoder model and a reconstruction error threshold; inputting the operation data set of the same type as the trained data pool into the clustering model and the autoencoder model to output a reconstruction error test value; detecting the fault state of the multi-split machine according to the reconstruction error test value and the reconstruction error threshold, which can reduce the fault detection cost and after-sales maintenance cost, and at the same time can also improve the fault detection accuracy and user experience, and has broad commercial application prospects.

Description

Technical Field

[0001] The present invention relates to the technical field of fault detection, and in particular to a multi-link machine and a fault detection method and device thereof, and a storage medium. Background Art

[0002] With the rapid development of industrial technology, multi-split systems have been widely used in modern commercial and residential buildings due to their high efficiency, energy saving, flexibility and ease of use. However, like all complex systems, if multi-split air conditioners encounter various faults during operation, the detection and diagnosis of faults becomes particularly complicated.

[0003] In the related technology, maintenance personnel are generally relied on to conduct on-site inspection and judgment, which places great demands on manpower and material resources and requires the maintenance personnel to have rich work experience. Otherwise, the inspection accuracy is difficult to guarantee and the maintenance cost is also high. Summary of the invention

[0004] The present invention aims to solve one of the technical problems in the related art to at least a certain extent. To this end, the first purpose of the present invention is to propose a multi-connected fault detection method, which can reduce the fault detection cost, thereby reducing the after-sales maintenance cost, and at the same time can improve the fault detection accuracy and user experience, and has broad commercial application prospects.

[0005] A second object of the present invention is to provide a computer-readable storage medium.

[0006] The third objective of the present invention is to provide a multi-connection fault detection device.

[0007] A fourth object of the present invention is to provide a multi-link system.

[0008] To achieve the above-mentioned purpose, the first aspect of the embodiment of the present invention proposes a multi-connected machine fault detection method, which includes the following steps: obtaining an operating data set accumulated by the multi-connected machine; traversing the operating data set day by day to store the daily operating data set into a corresponding data pool according to the operating conditions of the multi-connected machine; when the amount of data stored in at least one data pool reaches a lower limit length, training the data in the at least one data pool to obtain a clustering model, an autoencoder model and a reconstruction error threshold; when the operating conditions corresponding to the operating data set of the current day are the same as the operating conditions corresponding to the trained data pool, inputting the operating data set of the current day into the clustering model and the autoencoder model to output a reconstruction error test value; and detecting the fault state of the multi-connected machine according to the reconstruction error test value and the reconstruction error threshold.

[0009] The multi-connected machine fault detection method of the embodiment of the present invention first obtains the operation data set accumulated by the multi-connected machine, and then traverses the operation data set day by day, and stores it in the corresponding data pool according to the operation conditions of the multi-connected machine. When the amount of data stored in the data pool reaches the lower limit length, the amount of data reaching the length is trained to obtain a clustering model, an autoencoder model and a reconstruction error threshold. When traversing the operation data set again, if the working condition corresponding to the operation data set of the current day is the same as the working condition corresponding to the trained data pool, then the operation data set of the current day can be input into the trained clustering model and the autoencoder model to output a reconstruction error test value. By comparing the reconstruction error test value and the reconstruction error threshold, it is detected whether the multi-connected machine has a fault on the current day. It is convenient and fast, and can also reduce the fault detection cost, thereby reducing the after-sales maintenance cost. At the same time, it can also improve the fault detection accuracy and user experience, and has broad commercial application prospects.

[0010] In some embodiments of the present invention, the training of the data in the at least one data pool includes: performing cluster analysis on the data in the data pool to obtain the clustering model and a plurality of clustering sub-cluster training data; establishing a corresponding autoencoder model based on the plurality of clustering sub-cluster training data, and determining the reconstruction error threshold.

[0011] In some embodiments of the present invention, the clustering algorithms used for clustering analysis of the data in the data pool include: EAC-AL (evidence accumulation integrated clustering algorithm based on average connection hierarchical clustering), Kmeans (K-means clustering algorithm), Birch (balanced iterative condensation algorithm), SOM (self-organizing map network) and FCM (fuzzy C-means clustering) algorithms.

[0012] In some embodiments of the present invention, the method further includes: when it is determined that the multi-machine is in a faulty state based on the detection of the operating data set of the current day, the operating data set of the current day is recorded; when it is determined that the multi-machine is not in a faulty state based on the detection of the operating data set of the current day, and the multi-machines corresponding to the operating data sets of a first preset number of days before the current day and the operating data sets of a second preset number of days after the current day are not in a faulty state, the operating data set of the current day is stored in a corresponding data pool.

[0013] In some embodiments of the present invention, the method also includes: when the amount of data stored in at least one data pool reaches an update length, deleting data of a second preset length in the data pool, and updating the clustering model, the autoencoder model and the reconstruction error threshold based on the remaining data in the data pool.

[0014] In some embodiments of the present invention, a corresponding autoencoder model is established based on the plurality of cluster subcluster training data, including: standardizing the plurality of cluster subcluster training data; obtaining a preset amount of data from the plurality of cluster subcluster training data as a validation set; and using the root mean square error of the validation set as an indicator, repeatedly training a plurality of autoencoder models using the plurality of cluster subcluster training data until the indicator stops decreasing and starts to increase.

[0015] In some embodiments of the present invention, the method also includes: obtaining the maximum reconstruction error and the median reconstruction error output by each autoencoder model; if the maximum reconstruction error is greater than a preset multiple of the median reconstruction error, the autoencoder model is determined to be abnormal, and the autoencoder model is eliminated.

[0016] In some embodiments of the present invention, determining the reconstruction error threshold includes: determining the number of reconstruction error scatter points corresponding to each cluster sub-cluster training data; and determining the weighted average of the number of reconstruction error scatter points within a preset interval as the reconstruction error threshold.

[0017] In some embodiments of the present invention, after obtaining the multi-online accumulated operating data set, the method further includes: preprocessing the operating data set, wherein the preprocessing includes at least one of data merging, data deduplication, outlier processing, operating segment extraction, feature calculation, time resampling, missing value processing and feature extraction.

[0018] In some embodiments of the present invention, the method further comprises: performing feature classification and redundancy processing on the operating data set to generate multiple groups of operating data sets.

[0019] In some embodiments of the present invention, the operating conditions include cooling conditions, heating non-enthalpy injection conditions and heating enthalpy injection conditions. The data pool corresponding to the cooling conditions is the cooling conditions data pool, the data pool corresponding to the heating non-enthalpy injection conditions is the heating non-enthalpy injection conditions data pool, and the data pool corresponding to the heating enthalpy injection conditions is the heating enthalpy injection conditions data pool.

[0020] In some embodiments of the present invention, before inputting the current day's operating data set into the clustering model and the autoencoder model, the method also includes: determining that the data pool corresponding to the current day's operating data set has completed training, determining that the data volume of the current day's operating data set is greater than or equal to the test lower limit length, and determining that the data pool matching the working conditions corresponding to the current day's operating data set has completed training.

[0021] To achieve the above objectives, a second aspect of the present invention provides a computer-readable storage medium on which a multi-machine fault detection program is stored. When the program is executed by a processor, the multi-machine fault detection method described in any one of the above embodiments is implemented.

[0022] The computer-readable storage medium of the embodiment of the present invention can reduce the fault detection cost and thus reduce the after-sales maintenance cost by executing the multi-connection fault detection program stored thereon through a processor, while also improving the fault detection accuracy and user experience, and has broad commercial application prospects.

[0023] To achieve the above-mentioned purpose, the third aspect of the present invention proposes a multi-connected machine fault detection device, which includes: an acquisition module, used to acquire the operating data set accumulated by the multi-connected machine; a data processing module, used to traverse the operating data set day by day, so as to store the daily operating data set in the corresponding data pool according to the operating conditions of the multi-connected machine; a training module, used to train the data in the at least one data pool when the amount of data stored in the at least one data pool reaches the lower limit length, so as to obtain a clustering model, an autoencoder model and a reconstruction error threshold; a testing module, used to input the operating data set of the current day into the clustering model and the autoencoder model when the operating conditions corresponding to the operating data set of the current day are the same as the operating conditions corresponding to the trained data pool, so as to output a reconstruction error test value; a detection module, used to detect the fault state of the multi-connected machine according to the reconstruction error test value and the reconstruction error threshold.

[0024] The multi-connected machine fault detection device of the embodiment of the present invention includes an acquisition module, a data processing module, a training module, a test module and a detection module. First, the acquisition module acquires the operation data set accumulated by the multi-connected machine. Then, the data processing module traverses the operation data set day by day and stores it in the corresponding data pool according to the operation conditions of the multi-connected machine. When the amount of data stored in the data pool reaches the lower limit length, the amount of data reaching the length is trained to obtain a clustering module, an autoencoder module and a reconstruction error threshold. Then, when traversing the operation data set again, if the working condition corresponding to the operation data set of the current day is the same as the working condition corresponding to the trained data pool, the test module can input the operation data of the current day into the trained clustering model and the autoencoder model to output a reconstruction error test value. The detection module then detects whether the multi-connected machine has a fault on the current day by comparing the reconstruction error test value with the reconstruction error threshold. It is convenient and fast, and can also reduce the fault detection cost, thereby reducing the after-sales maintenance cost. At the same time, it can also improve the fault detection accuracy and user experience, and has broad commercial application prospects.

[0025] To achieve the above-mentioned purpose, a fourth aspect of the present invention provides a multi-split system, which includes the multi-split system fault detection device described in the above-mentioned embodiment.

[0026] The multi-split system of the embodiment of the present invention can reduce the fault detection cost and thus reduce the after-sales maintenance cost through the multi-split system fault detection device of the above embodiment, and can also improve the fault detection accuracy and user experience, and has broad commercial application prospects.

[0027] Additional aspects and advantages of the present invention will be given in part in the following description and in part will be obvious from the following description, or will be learned through practice of the present invention. BRIEF DESCRIPTION OF THE DRAWINGS

[0028] Figure 1 is a flow chart of a multi-connection fault detection method according to an embodiment of the present invention;

[0029] Figure 2 is a schematic diagram of feature extraction in a specific embodiment of the present invention;

[0030] Figure 3 is a schematic diagram of a data pool in a specific embodiment of the present invention;

[0031] Figure 4 is a schematic diagram of a data pool in another specific embodiment of the present invention;

[0032] Figure 5 is a schematic diagram of calculation of reconstruction error in a specific embodiment of the present invention;

[0033] Figure 6 is a schematic diagram of a data pool in another specific embodiment of the present invention;

[0034] Figure 7 is a schematic diagram of an autoencoder model in one embodiment of the present invention;

[0035] Figure 8 is a schematic diagram of merging reconstruction errors in one embodiment of the present invention;

[0036] Fig. 9 is a flow chart of a multi-connection fault detection method according to a specific embodiment of the present invention;

[0037] Fig.10 It is a structural block diagram of a multi-connection fault detection device in an embodiment of the present invention;

[0038] Fig.11 It is a structural block diagram of a multi-link in an embodiment of the present invention. DETAILED DESCRIPTION

[0039] Embodiments of the present invention are described in detail below, examples of which are shown in the accompanying drawings, wherein the same or similar reference numerals throughout represent the same or similar elements or elements having the same or similar functions. The embodiments described below with reference to the accompanying drawings are exemplary and are intended to be used to explain the present invention, and should not be construed as limiting the present invention.

[0040] The following describes a multi-connected machine and its fault detection method and device, and storage medium according to an embodiment of the present invention with reference to the accompanying drawings.

[0041] The present invention is applicable to multi-split air-conditioning equipment and also to other multi-connection equipment. When applied to different types of multi-split equipment, the operating data involved in fault detection can be adaptively modified, which can be specifically formulated based on corresponding experimental data or expert knowledge in the same field. To avoid redundancy, the present invention is illustrated by taking multi-split air-conditioning equipment as an example.

[0042] Figure 1 The present invention is a flowchart of a multi-connection fault detection method according to an embodiment of the present invention.

[0043] like Figure 1 As shown, the present invention proposes a multi-connection fault detection method, which includes the following steps:

[0044] S10, obtaining a multi-online accumulated running data set.

[0045] Specifically, the operation data accumulated by the multi-split system can be directly obtained through sensors, or can be calculated based on the data obtained by the sensors. For example, the operation data set in this embodiment can include but is not limited to saturated condensing temperature, saturated evaporating temperature, outdoor heat exchanger outlet temperature, outdoor heat exchanger inlet temperature, compressor frequency, compressor exhaust temperature. In addition, the accumulated time of the multi-split system can be several consecutive months.

[0046] After obtaining the operating data set, the multi-online fault detection method further includes: preprocessing the operating data set, the preprocessing including at least one of data merging, data deduplication, outlier processing, operating segment extraction, feature calculation, time resampling, missing value processing and feature extraction.

[0047] Specifically, when the sensor acquires the operating data of multiple machines, the acquisition time may be different or the time sequence of the data may be disordered after acquisition. Therefore, the embodiment may first merge the data and normalize it using the time series data as an index, so as to convert the data into my country's UTC+8 date and time in milliseconds.

[0048] Data deduplication is to solve the problem of repeated sampling of sensors and reduce the storage pressure of the storage pool. Specifically, it can be done by comparing data of the same type and using time data as an index to determine the data sampled at the same time, and then deleting one of them to complete the data deduplication operation.

[0049] In abnormal value processing, taking the temperature sensor as an example, a threshold of 80°C can be set for the indoor return air temperature point, a threshold of 150°C can be set for the indoor unit inlet and outlet pipe temperature points, and a threshold of 150°C can be set for each temperature point of the outdoor unit. If data exceeding the corresponding temperature threshold is obtained later, it means that the acquisition is wrong, so the data needs to be eliminated to maintain the correctness of the data.

[0050] In the operation segment processing, it is necessary to process the abnormal data and missing data in the multi-link operation segment, including multiple important parameters such as system high pressure, system low pressure, compressor exhaust temperature, etc., and define that if half or more of the data in these important parameters are missing, the corresponding data can be listed as missing state data. Furthermore, if a large area is missing in a single operation segment (for example, 80% of the data is missing), then the data of the entire operation segment can be eliminated, because the data of this operation segment is incomplete and it is easy to make mistakes when participating in subsequent calculations. In addition, in order to ensure the accuracy of subsequent detection, the points where the bypass valve is opened a small amount (such as the opening is less than 1%) and the data is unstable are screened out. For machines with a large number of bypasses, they can be directly considered as faulty machines, and no subsequent fault detection is performed. Use the runModeEx (operation mode) point to screen out the scattered points of operating conditions such as oil return, defrosting, and preheating. In view of the problem that some defrosting scattered points at this point still remain, a small algorithm is designed to further screen out such scattered points. In addition, the rule that the supercooling degree is greater than 20 and the supercooling degree is greater than 5 times the variance plus the mean is designed to screen out the related supercooling degree abnormal value points. For the temporary interruption scattered points under some injection enthalpy conditions, similar screening is performed to meet the continuous injection enthalpy for 4 moments (in this embodiment, 1 moment can be 5 minutes), 1-2 injection enthalpy interruption points appear, and the injection enthalpy is restored after the interruption to reach 4 moments. Then, the 1-2 injection enthalpy interruption points that appear in the middle are screened out; after the processing is completed, each operation section is merged to form a complete pure operation section data.

[0051] The extraction of the above-mentioned operation segment data, first of all, the operation segment can be defined as the total compressor frequency of the equipment being greater than zero for 3 consecutive moments, that is, the multi-split machine is started and running, and when the multi-split machine is in operation, once the frequency is zero, the operation segment is terminated. The rest is the multi-split shutdown segment. After determining the operation segment, the mean, maximum value, maximum first-order difference and operation segment length of the compressor frequency in each operation segment are counted. Among them, the mean value of the compressor frequency represents the average usage intensity of a single use by the user, the maximum value of the compressor frequency represents the peak cooling and heating load demand of a single use by the user, the maximum first-order difference aims to explore the urgency of changes in the demand side of the machine, and the operation segment length aims to explore the running time of the machine after a single start.

[0052] For feature calculation, other feature data is obtained by calculation through the acquired data. For example, the subcooling degree under refrigeration conditions is calculated: saturated condensing temperature - outdoor heat exchanger outlet temperature; superheat degree under heating conditions is: outdoor heat exchanger inlet temperature - saturated evaporation temperature; superheat degree under refrigeration conditions: first calculate the difference between the outlet temperature of each indoor evaporator and the saturated evaporation temperature, and then take the average of the indoor units in the turned-on state; subcooling degree under heating conditions is: saturated condensing temperature minus condenser outlet temperature, and the average is calculated in the same way (saturated evaporation and condensation temperatures are calculated by the current system pressure and refrigerant type); compressor exhaust temperature (multi-split units may have multiple compressors): use the weighted average method, if the frequency of each compressor is not less than 1, calculate the weight of each compressor frequency (the proportion of the compressor frequency to the sum of the compressor frequencies), and the exhaust temperature is weighted and summed according to the compressor frequency weight; the same is true for compressor suction temperature; compressor exhaust superheat degree: compressor exhaust temperature - saturated condensing temperature; compressor frequency sum: sum of multiple compressor frequencies; calculate the average expansion valve opening of the indoor unit; calculate the average return air temperature of the indoor unit, etc.

[0053] For time resampling processing, for example, each hour can be divided into 12 equal parts, each of which is a 5-minute left-closed and right-open interval. The data with the earliest timestamp within each 5 minutes is retained, and the remaining data within the 5-minute segment is removed to solve data redundancy and possible imbalance problems in the time series.

[0054] For feature extraction processing, see Figure 2 , use sliding window to extract the running data set, filter the continuous data segments of specified length N, create sliding windows of specified length M in each continuous data segment, calculate the mean of each variable in the sliding window, replace the data of the first data point in the window, and slide with L as the step length to realize feature extraction, so that the data has time series information. Here, the selection of M and L can be customized by the user. Figure 2Taking the sliding window in as an example, M is set to 3 and L is set to 2, that is, each feature extraction point contains the information of three original data points. Taking 2 as the sliding step size means that the amount of data after feature extraction becomes half of the original, and the time interval before and after the data becomes twice of the original.

[0055] in addition, Figure 2 The three sets of data of time series 88, 89, and 90 are deleted because they are only three sets of continuous data. If a sliding window with a length of 3 is used, sliding cannot be performed. Of course, the three sets of data can also be averaged and the average can be used as a set of data to replace the three sets of data. Time series 1, 2, 3, 4, 5, 6 and time series 106, 107, 108, 109, 110, 111, and 112 can be processed normally.

[0056] It should be noted that, in practical applications, one or more of the above descriptions of various preprocessing methods may be adopted, and the specific preprocessing methods adopted are not specifically limited here.

[0057] In some embodiments, after the operating data set is preprocessed, feature classification and redundancy processing may be performed on it to generate multiple groups of operating data sets.

[0058] Specifically, the features of the preprocessed operating data set can be defined first, and after the data of each feature group is standardized, the standardized data and the mean and standard deviation of the corresponding group of operating data sets after standardization can be stored, so that when the subsequent test set is passed in, the same parameters can be used for standardization. The feature data in each operating data set can be first formulated based on engineering experience and expert knowledge, such as determining the feature data group according to indoor and outdoor meteorological parameters and determining the feature group according to the thermophysical parameters of the outdoor unit, and then clustering using the same clustering algorithm and using the mutual information index to determine the redundancy of the feature group. If the mutual information index between groups is large, the feature groups are relatively redundant, and the more redundant feature groups are discarded, and the more independent feature groups are retained.

[0059] S20, traversing the operation data set day by day to store the daily operation data set into a corresponding data pool according to the operation conditions of the multi-connection.

[0060] Specifically, after the operation data set is obtained through the above processing, the operation data set can be traversed day by day, and then the daily operation data set can be stored in the corresponding data set according to the operation conditions of the multi-link, wherein different operation conditions correspond to different data pools. More specifically, the operation conditions include cooling conditions, heating non-injection enthalpy conditions, and heating injection enthalpy conditions, wherein the data pool corresponding to the cooling conditions is the cooling conditions data pool, the data pool corresponding to the heating non-injection enthalpy conditions is the heating non-injection enthalpy conditions data pool, and the data pool corresponding to the heating injection enthalpy conditions is the heating injection enthalpy conditions data pool.

[0061] It should be noted that the creation of a data pool can realize unattended fault detection and improve intelligence. Specifically, after the data pool is created, the data pool and its state variables can be initialized. If the state variable is False, it means that the initialization has been completed. Then define the hyperparameters of each data pool. Figure 3 As shown in the figure, define the minimum test data length (i.e., how many data are required to be tested on the same day), the minimum test length of a single condition (the minimum number of data required for cooling, heating without enthalpy injection, and heating with enthalpy injection on the same day), and the minimum test ratio of a single condition (the minimum test ratio required for cooling, heating without enthalpy injection, and heating with enthalpy injection on the same day). It should be noted that the lower limit lengths corresponding to different data pools can be different or the same. Figure 3 The lower limit length of the cooling pool, i.e., the cooling condition data pool, is 6000, the lower limit length of the heating non-spraying enthalpy pool, i.e., the heating non-spraying enthalpy condition data pool, is also 6000, and the lower limit length of the heating spraying enthalpy pool, i.e., the heating spraying enthalpy condition data pool, is 1500. It should be noted that, in this embodiment, the lower limit length is the minimum length of the data in the data pool for training.

[0062] S30, when the amount of data stored in at least one data pool reaches a lower limit length, training the data in at least one data pool to obtain a clustering model, an autoencoder model and a reconstruction error threshold.

[0063] Specifically, in the process of storing data in the data pool, the amount of data in the storage pool can be compared with the data length defined above, and then determine whether the currently stored data can be trained. The lower limit length in this embodiment is the same as the minimum test data length of the above embodiment. After the amount of data reaches the lower limit length, the data in the data pool can be clustered to obtain a clustering model and cluster sub-cluster training data. In addition, in order to improve the detection accuracy, different working conditions correspond to different clustering models.

[0064] S40, when the working condition corresponding to the current day's operating data set is the same as the working condition corresponding to the trained data pool, the current day's operating data set is input into the clustering model and the autoencoder model to output a reconstruction error test value.

[0065] Specifically, after the data pool corresponding to a working condition completes training, in the subsequent process of traversing the operating data set day by day, if the operating data set of the current day is the same as the working condition corresponding to the trained data pool, it means that the current working condition already has a corresponding clustering model and an autoencoder model. Therefore, the operating data set of the current day can be input into the clustering model and the autoencoder model for analysis to obtain the corresponding reconstruction error test value.

[0066] S50: Detect the fault status of the multi-connection system according to the reconstruction error test value and the reconstruction error threshold.

[0067] Specifically, the reconstruction error threshold in this embodiment is determined after the clustering model training is completed, and the autoencoder model is trained according to the clustering sub-cluster training data, and the reconstruction error threshold is determined in the process of training the autoencoder model. After the test data is input into the autoencoder model, the reconstruction error test value can be output after being processed by the autoencoder model, and the reconstruction error test value is then compared with the reconstruction error threshold, so as to determine whether the multi-connected machine is in a faulty state. For example, if the reconstruction error test value is greater than the reconstruction error threshold, it can be determined that the multi-connected machine is in a faulty state, and if the reconstruction error test value is less than or equal to the reconstruction error threshold, it can be determined that the multi-connected machine is not in a faulty state.

[0068] In some embodiments of the present invention, training data in at least one data pool includes: performing cluster analysis on the data in the data pool to obtain a clustering model and a number of cluster sub-cluster training data; establishing a corresponding autoencoder model based on the several cluster sub-cluster training data, and determining a reconstruction error threshold.

[0069] Specifically, Figure 4 As shown in the figure, if the data volume of the day reaches the lower limit length, the training state is in the ready state, that is, the Ready state, and the data pool has not completed the training, then the data in the data pool can be trained to finally obtain the clustering model and the autoencoder model. During the training process, the data needs to be clustered first to obtain the clustering model and several cluster sub-cluster training data, and then the several cluster sub-cluster training data are used to establish each independent autoencoder model, where one cluster sub-cluster training data can establish an autoencoder model.

[0070] In a specific embodiment, determining the reconstruction error threshold may include: determining the number of reconstruction error scatter points corresponding to each cluster sub-cluster training data; and determining a weighted average of the number of reconstruction error scatter points within a preset interval as the reconstruction error threshold.

[0071] Specifically, in the process of establishing the autoencoder model, each reconstruction error scatter point can be recorded, such as Figure 5 As shown, Figure 5 Each star-shaped point in the figure represents a reconstruction error scatter point, and the ordinate represents the reconstruction error value. The mean of the 25%-75% quantile scatter points in the reconstruction error scatter points can be used as the reconstruction error threshold, thereby reducing the impact of abnormal reconstruction error scatter points and further improving fault detection accuracy.

[0072] In a specific embodiment, different clustering algorithms are used to cluster each data set to form several sub-cluster data. Specifically, the algorithms include EAC-AL, Kmeans, Birch, SOM and FCM, a total of five kinds. Among them, EAC-AL and KMeans are selected. Choosing EAC-AL can obtain more stable clustering results, but the computing resources are consumed more, while KMeans is the opposite. The user can choose it by himself. When the EAC-AL algorithm is selected, the number of clusters K is automatically determined by the EAC-AL algorithm, and K is passed to the Birch, SOM, and FCM algorithms. For the Birch and SOM algorithms, the K is first used to optimize the remaining hyperparameters, and then the optimal hyperparameters are used for clustering. If the KMeans algorithm is selected, the KMean algorithm hyperparameters are first optimized to determine the optimal K value, and then K is passed to the Birch, SOM, and FCM algorithms. For the Birch and SOM algorithms, the K is first used to optimize the remaining hyperparameters, and then the optimal hyperparameters are used for clustering.

[0073] Taking a data set as an example, the steps of cluster analysis are described as follows: First, the data set is passed into the EAC-AL algorithm for clustering to obtain the labels of each data point under the algorithm, and then the number of categories of the labels is counted to obtain the number of clusters K. The number of clusters K is then used as a known parameter and passed into the hyperparameter optimization module in the SOM algorithm for grid search hyperparameter optimization to obtain the optimal hyperparameter combination, and passed into the hyperparameter optimization module in the Birch algorithm for grid search hyperparameter optimization to obtain the Birch optimal hyperparameter combination. The optimization index is a combination of the CH (Calinski Harabasz) index and the DB (Davies Bouldin) index, and the purpose is to find a hyperparameter value with the CH index as large as possible and the DB index as small as possible. The number of clusters K and the optimal hyperparameter combination are passed into the SOM algorithm, the number of clusters K and the Birch optimal hyperparameter combination are passed into the Birch algorithm, and the number of clusters K is passed into the FCM algorithm to obtain the labels of each data point and the corresponding clustering models under these three algorithms.

[0074] In some embodiments of the present invention, the multi-online fault detection method also includes: when the amount of data stored in at least one data pool reaches an update length, deleting data of a second preset length in the data pool, and updating the clustering model, autoencoder model and reconstruction error threshold based on the remaining data in the data pool.

[0075] Specifically, in order to improve the fault detection accuracy, this embodiment also updates the clustering model, the autoencoder model and the reconstruction error threshold, such as Figure 6 As shown, in at least one data pool ( Figure 6When the amount of data stored in the data pool (where the data pool is the cooling pool) reaches the update length, the data in the data pool is used for updating. Optionally, the update length may be twice the lower limit length.

[0076] More specifically, when the amount of data in the data pool reaches the update length, the earliest 25% of the data stored in the storage pool can be removed first, and then the remaining 75% of the data can be used to update the clustering model, autoencoder model and reconstruction error threshold to increase the update speed.

[0077] In some embodiments of the present invention, a corresponding autoencoder model is established based on a number of cluster subcluster training data, including: standardizing the number of cluster subcluster training data; obtaining a preset amount of data from the number of cluster subcluster training data as a validation set; using the root mean square error of the validation set as an indicator, repeatedly training a number of autoencoder models using the number of cluster subcluster training data until the indicator increases.

[0078] Specifically, the autoencoder is an unsupervised learning neural network model, including an encoder and a decoder, which is mainly used for data compression and feature learning. Figure 7 As shown, where x 1 、x 2 、x 3 , …, x d represents the input signal of the autoencoder model, , , , …, Represents the output signal of the autoencoder model. z in the autoencoder model 1 , …, z p Represents the signal learned by the autoencoder. In order to reduce the negative impact of volatility and randomness of autoencoder training on prediction performance, this module trains multiple autoencoders and checks whether the quality of each autoencoder model has a large deviation. If not, the model is saved locally and several autoencoder models are integrated to improve the generalization ability of the model.

[0079] The specific steps of the autoencoder model training are as follows: First, several cluster sub-cluster training data are standardized. The standardization process can be completed in combination with the mean value and the standard deviation. The specific implementation process is not specifically limited in this embodiment. After completing the standardization process, 10% of the data in the several cluster sub-cluster training data can be used as a validation set, and then the root mean square error of the validation set is used as an indicator. In the process of training the autoencoder model with several sub-cluster training data, the size of the indicator can be observed. If the indicator no longer decreases and begins to increase, the training can be stopped, and the parameters of the autoencoder model are saved in the local memory to complete the training of the autoencoder model.

[0080] In this embodiment, the method also includes: obtaining the maximum reconstruction error and the median reconstruction error output by each autoencoder model; if the maximum reconstruction error is greater than a preset multiple of the median reconstruction error, the autoencoder model is determined to be abnormal and the autoencoder model is eliminated.

[0081] Specifically, after the autoencoder model is trained, in order to improve the availability of the autoencoder model, the reconstruction error value output by the autoencoder can also be used to determine whether the autoencoder model is abnormal. Specifically, the maximum reconstruction error and the median reconstruction error output by the autoencoder are first determined. If the maximum reconstruction error is greater than 10 times the median reconstruction error, it means that the training effect of the autoencoder is extremely poor, so the autoencoder result can be proposed to reduce the negative impact of the volatility of the autoencoder on the fault detection results. It should be noted that the preset multiple in this specific embodiment is 10, and in other embodiments, it can be replaced by other values, which are not specifically limited here.

[0082] In some embodiments of the present invention, each data set will generate various reconstruction errors after being processed by cluster analysis and autoencoder model, see Figure 8 In the figure, Cluster-A-1, Cluster-A-2, Cluster-B-1, Cluster-B-2, Cluster-B-3, Cluster-C-1, Cluster-C-2 and Cluster-C-3 represent different subclusters, and the total reconstruction error and item reconstruction error of each subcluster are merged respectively. Among them, the total reconstruction error is the sum of squares of the reconstruction errors of each dimensional data. For one data, there is only one total reconstruction error value greater than or equal to 0; the item reconstruction error is the original reconstruction error of each dimensional data. For one data, there are the same number of item reconstruction errors as the number of dimensions. In the merging process, the importance of different clustering algorithms can also be considered, and different weights can be defined, and then weighted merging can be performed.

[0083] After using the reconstruction error to judge the fault of the multi-connected machine, the operation data is processed differently according to different fault states. When it is determined that the multi-connected machine is in a fault state according to the operation data set detection of the current day, the operation data set of the current day is recorded; when it is determined that the multi-connected machine is not in a fault state according to the operation data set detection of the current day, and the multi-connected machine corresponding to the operation data sets of the first preset number of days before the current day and the operation data sets of the second preset number of days after the current day are not in a fault state, the operation data set of the current day is stored in the corresponding data pool.

[0084] Specifically, in order to further improve the accuracy of the model, this embodiment also stores the corresponding operating data set that has been judged to be fault-free in the data pool, so that the model can be trained using the fault-free operating data set in the future. In order to avoid misjudgment, this embodiment also limits the operating data set stored in the data pool to not only the corresponding multi-connection state needs to be in a fault-free state, but also requires that the multi-connection has no faults in the three days before the current day and has no faults in the seven days after the current day, before the operating data of the current day can be stored in the storage pool.

[0085] In a specific embodiment of the present invention, Fig. 9 As shown, the running data set is first input into the preprocessing module for preprocessing, and then the preprocessed running data set is input into the feature processing module. After the feature processing module classifies the features, the processed data is dynamically pooled through the data processing module. Specifically, the data can be traversed day by day, stored in different data pools according to different working conditions, and the data volume can be judged. When the data volume reaches the lower limit length, the model training is carried out. Specifically, the training data is standardized and sequentially input into the clustering module and the autoencoder module. After that, the reconstruction error can be output, and the reconstruction error threshold is obtained after weighted processing and abnormal elimination processing by the detection module. In the subsequent day-by-day traversal process, if the data pool corresponding to the current day's operating data set has completed training, the amount of data in the current day's operating data set is greater than or equal to the test lower limit length, and the data pool matching the working condition corresponding to the current day's operating data set has completed training, then the current day's operating data set can be tested and processed through the trained clustering model and autoencoder model to obtain the reconstruction error test value, and then the reconstruction error test value is compared with the reconstruction error threshold to determine whether there is a fault in the current day's multi-connection machine. When the current day's operating data set passes the test without fault, it is stored in the corresponding data pool. If the data in the data pool reaches the update length, the clustering model and the autoencoder model are updated to ensure the accuracy of the fault detection results. In the detection module, the total reconstruction error of the sub-cluster of the same feature group and the sub-item reconstruction error of each sub-cluster are obtained through the weighted algorithm and the abnormal information integration, and then the total reconstruction error and the sub-item reconstruction error of each sub-cluster are merged and processed respectively to output the reconstruction error, and the reconstruction error is compared with the reconstruction error threshold to obtain the detection result. In addition, it should be noted that Fig. 9The dotted box in the middle represents the specific structure of the data processing module, wherein the arrow on the left extending outside the dotted box indicates that the data is processed by the feature processing module and then input into the data processing module for query detection, that is, the data set in the daily traversal data set is the data set processed by the feature processing module, and the arrow on the right extending outside the dotted box indicates the data processed by the data processing module and then input into the clustering module for clustering analysis.

[0086] In summary, the multi-connected fault detection method of the embodiment of the present invention can reduce the fault detection cost, thereby reducing the after-sales maintenance cost, and can also improve the fault detection accuracy and user experience, and has broad commercial application prospects.

[0087] Furthermore, the present invention proposes a computer-readable storage medium on which a multi-machine fault detection program is stored. When the program is executed by a processor, the multi-machine fault detection method of any one of the above embodiments is implemented.

[0088] The computer-readable storage medium of the embodiment of the present invention can reduce the fault detection cost and thus reduce the after-sales maintenance cost by executing the multi-connection fault detection program stored thereon through a processor, while also improving the fault detection accuracy and user experience, and has broad commercial application prospects.

[0089] Fig.10 It is a structural block diagram of a multi-connection fault detection device in an embodiment of the present invention.

[0090] Furthermore, if Fig.10 As shown, the present invention proposes a multi-connected fault detection device 100, which includes an acquisition module 101, a data processing module 102, a training module 103, a testing module 104 and a detection module 105.

[0091] Among them, the acquisition module 101 is used to obtain the operating data set accumulated by the multi-connected machine; the data processing module 102 is used to traverse the operating data set day by day to store the daily operating data set in the corresponding data pool according to the operating conditions of the multi-connected machine; the training module 103 is used to train the data in at least one data pool when the amount of data stored in at least one data pool reaches the lower limit length to obtain the clustering model, the autoencoder model and the reconstruction error threshold; the testing module 104 is used to input the operating data set of the current day into the clustering model and the autoencoder model when the operating conditions corresponding to the operating data set of the current day are the same as the operating conditions corresponding to the trained data pool, so as to output the reconstruction error test value; the detection module 105 is used to detect the fault state of the multi-connected machine according to the reconstruction error test value and the reconstruction error threshold.

[0092] In some embodiments of the present invention, the training module 103 is specifically used to: perform cluster analysis on the data in the data pool to obtain a clustering model and several cluster sub-cluster training data; establish a corresponding autoencoder model based on the several cluster sub-cluster training data, and determine a reconstruction error threshold.

[0093] In some embodiments of the present invention, the clustering algorithms used to perform cluster analysis on the data in the data pool include: multiple algorithms among EAC-AL, Kmeans, Birch, SOM and FCM.

[0094] In some embodiments of the present invention, the data processing module 102 is further used for: when it is determined that the multi-connected machine is in a faulty state based on the detection of the operation data set of the current day, recording the operation data set of the current day; when it is determined that the multi-connected machine is not in a faulty state based on the detection of the operation data set of the current day, and the multi-connected machines corresponding to the operation data sets of the first preset number of days before the current day and the operation data sets of the second preset number of days after the current day are not in a faulty state, then storing the operation data set of the current day in the corresponding data pool.

[0095] In some embodiments of the present invention, the data processing module 102 is used to: when the amount of data stored in at least one data pool reaches the update length, delete the data of a second preset length in the data pool; the training module 103 is used to: update the clustering model according to the remaining data in the data pool, and update the autoencoder model and the reconstruction error threshold according to the remaining data in the data pool.

[0096] In some embodiments of the present invention, the training module 103 is also used to: standardize a number of cluster sub-cluster training data; obtain a preset amount of data from a number of cluster sub-cluster training data as a validation set; and use the root mean square error of the validation set as an indicator to repeatedly train a number of autoencoder models using a number of cluster sub-cluster training data until the indicator stops decreasing and starts to increase.

[0097] In some embodiments of the present invention, the data processing module 102 is also used to: obtain the maximum reconstruction error and the median reconstruction error output by each autoencoder model; if the maximum reconstruction error is greater than a preset multiple of the median reconstruction error, the autoencoder model is determined to be abnormal, and the autoencoder model is eliminated.

[0098] In some embodiments of the present invention, the training module 103 is used to: determine the number of reconstruction error scatter points corresponding to each cluster sub-cluster training data; and determine the weighted average of the number of reconstruction error scatter points within a preset interval as the reconstruction error threshold.

[0099] In some embodiments of the present invention, the data processing module 102 is also used to: after obtaining the multi-online accumulated operating data set, preprocess the operating data set, and the preprocessing includes at least one of data merging, data deduplication, outlier processing, operating segment extraction, feature calculation, time resampling, missing value processing and feature extraction.

[0100] In some embodiments of the present invention, the data processing module 102 is further used to perform feature classification and redundancy processing on the operating data set to generate multiple groups of operating data sets.

[0101] In some embodiments of the present invention, the operating conditions include cooling conditions, heating non-enthalpy injection conditions and heating enthalpy injection conditions. The data pool corresponding to the cooling conditions is the cooling conditions data pool, the data pool corresponding to the heating non-enthalpy injection conditions is the heating non-enthalpy injection conditions data pool, and the data pool corresponding to the heating enthalpy injection conditions is the heating enthalpy injection conditions data pool.

[0102] In some embodiments of the present invention, the test module 104 is also used to: before inputting the current day's operating data set into the clustering model and the autoencoder model, determine that the data pool corresponding to the current day's operating data set has completed training, determine that the data volume of the current day's operating data set is greater than or equal to the test lower limit length, and determine that the data pool matching the working conditions corresponding to the current day's operating data set has completed training.

[0103] It should be noted that the specific implementation of the multi-connected machine fault detection device in the embodiment of the present invention can refer to the specific implementation of the multi-connected machine fault detection method in the above embodiment, and will not be described again here to avoid redundancy.

[0104] In summary, the multi-connected fault detection device of the embodiment of the present invention can reduce the fault detection cost, thereby reducing the after-sales maintenance cost, and can also improve the fault detection accuracy and user experience, and has broad commercial application prospects.

[0105] Fig.11 It is a structural block diagram of a multi-link in an embodiment of the present invention.

[0106] Furthermore, if Fig.11 As shown, the present invention proposes a multi-split machine 200, which includes the multi-split machine fault detection device 100 in the above embodiment.

[0107] The multi-split system of the embodiment of the present invention can reduce the fault detection cost and thus reduce the after-sales maintenance cost through the multi-split system fault detection device of the above embodiment, and can also improve the fault detection accuracy and user experience, and has broad commercial application prospects.

[0108] In addition, other structures and functions of the multi-split system in the embodiment of the present invention are known to those skilled in the art and will not be described in detail here to reduce redundancy.

[0109] It should be noted that the logic and / or steps represented in the flowchart or otherwise described herein, for example, can be considered as a sequenced list of executable instructions for implementing logical functions, and can be specifically implemented in any computer-readable medium for use by an instruction execution system, device or apparatus (such as a computer-based system, a system including a processor, or other system that can fetch instructions from an instruction execution system, device or apparatus and execute instructions), or in combination with these instruction execution systems, devices or apparatuses. For the purposes of this specification, "computer-readable medium" can be any device that can contain, store, communicate, propagate or transmit a program for use by an instruction execution system, device or apparatus, or in combination with these instruction execution systems, devices or apparatuses. More specific examples (non-exhaustive list) of computer-readable media include the following: an electrical connection portion with one or more wirings (electronic device), a portable computer disk box (magnetic device), a random access memory (RAM), a read-only memory (ROM), an erasable and programmable read-only memory (EPROM or flash memory), an optical fiber device, and a portable compact disk read-only memory (CDROM). In addition, the computer-readable medium may even be paper or other suitable medium on which the program is printed, since the program may be obtained electronically, for example, by optically scanning the paper or other medium and then editing, interpreting or processing in other suitable ways if necessary, and then stored in a computer memory.

[0110] It should be understood that the various parts of the present invention can be implemented by hardware, software, firmware or a combination thereof. In the above-mentioned embodiments, multiple steps or methods can be implemented by software or firmware stored in a memory and executed by a suitable instruction execution system. For example, if implemented by hardware, as in another embodiment, it can be implemented by any one of the following technologies known in the art or their combination: a discrete logic circuit having a logic gate circuit for implementing a logic function for a data signal, a dedicated integrated circuit having a suitable combination of logic gate circuits, a programmable gate array (PGA), a field programmable gate array (FPGA), etc.

[0111] In the description of this specification, the description with reference to the terms "one embodiment", "some embodiments", "examples", "specific examples", or "some examples" means that the specific features, structures, materials or characteristics described in conjunction with the embodiment or example are included in at least one embodiment or example of the present invention. In this specification, the schematic representation of the above terms does not necessarily refer to the same embodiment or example. Moreover, the specific features, structures, materials or characteristics described may be combined in any one or more embodiments or examples in a suitable manner.

[0112] In the description of the present invention, it is to be understood that the terms “center”, “longitudinal”, “lateral”, “length”, “width”, “thickness”, “up”, “down”, “front”, “back”, “left”, “right”, “vertical”, “horizontal”, “top”, “bottom”, “inside”, “outside”, “clockwise”, “counterclockwise”, “axial”, “radial”, “circumferential”, etc., indicating orientations or positional relationships based on the orientations or positional relationships shown in the accompanying drawings, are only for the convenience of describing the present invention and simplifying the description, and do not indicate or imply that the referred device or element must have a specific orientation, be constructed and operated in a specific orientation, and therefore should not be understood as limiting the present invention.

[0113] In addition, the terms "first", "second", etc. used in the embodiments of the present invention are only used for descriptive purposes and should not be understood as indicating or implying relative importance, or implicitly indicating the number of technical features indicated in the present embodiment. Therefore, the features defined by the terms "first", "second", etc. in the embodiments of the present invention can explicitly or implicitly indicate that the embodiment includes at least one of the features. In the description of the present invention, the word "multiple" means at least two or two or more, such as two, three, four, etc., unless otherwise clearly and specifically defined in the embodiments.

[0114] In the present invention, unless otherwise clearly specified or limited in the embodiments, the terms "installed", "connected", "connected" and "fixed" etc. in the embodiments should be understood in a broad sense. For example, the connection can be a fixed connection, a detachable connection, or an integrated connection. It can be understood that it can also be a mechanical connection, an electrical connection, etc.; of course, it can also be a direct connection, or an indirect connection through an intermediate medium, or it can be the internal connection of two elements, or the interaction relationship between two elements. For ordinary technicians in this field, the specific meanings of the above terms in the present invention can be understood according to the specific implementation situation.

[0115] In the present invention, unless otherwise clearly specified and limited, a first feature being "above" or "below" a second feature may mean that the first and second features are in direct contact, or the first and second features are in indirect contact through an intermediate medium. Moreover, a first feature being "above", "above" or "above" a second feature may mean that the first feature is directly above or obliquely above the second feature, or simply means that the first feature is higher in level than the second feature. A first feature being "below", "below" or "below" a second feature may mean that the first feature is directly below or obliquely below the second feature, or simply means that the first feature is lower in level than the second feature.

[0116] Although the embodiments of the present invention have been shown and described above, it is to be understood that the above embodiments are exemplary and are not to be construed as limitations of the present invention. A person skilled in the art may change, modify, replace and vary the above embodiments within the scope of the present invention.

Claims

1. A multi-connection fault detection method, characterized in that: The method comprises: Acquire the multi-connection accumulated operation data set; Traversing the operation data set on a daily basis to store the daily operation data set into a corresponding data pool according to the operation conditions of the multi-link; When the amount of data stored in at least one data pool reaches a lower limit length, training the data in the at least one data pool to obtain a clustering model, an autoencoder model, and a reconstruction error threshold; When the operating condition corresponding to the current day's operating data set is the same as the operating condition corresponding to the trained data pool, the current day's operating data set is input into the clustering model and the autoencoder model to output a reconstruction error test value; Detecting a fault state of the multi-connection machine according to the reconstruction error test value and the reconstruction error threshold; The method further includes: when it is determined that the multi-connected machine is in a fault state according to the operation data set detection of the current day, recording the operation data set of the current day; when it is determined that the multi-connected machine is not in a fault state according to the operation data set detection of the current day, and the multi-connected machines corresponding to the operation data sets of the first preset number of days before the current day and the operation data sets of the second preset number of days after the current day are not in a fault state, storing the operation data set of the current day in a corresponding data pool; The step of training the data in the at least one data pool includes: Performing cluster analysis on the data in the data pool to obtain the cluster model and a plurality of cluster sub-cluster training data; Corresponding autoencoder models are established according to the plurality of cluster sub-cluster training data, and the reconstruction error threshold is determined.

2. The multi-connection fault detection method according to claim 1, characterized in that: The clustering algorithms used for clustering analysis of the data in the data pool include: multiple algorithms among EAC-AL, Kmeans, Birch, SOM and FCM.

3. The multi-connection fault detection method according to claim 1, characterized in that: The method further comprises: When the amount of data stored in at least one data pool reaches the update length, the data of the second preset length in the data pool is deleted, and the clustering model, the autoencoder model and the reconstruction error threshold are updated according to the remaining data in the data pool.

4. The multi-connection fault detection method according to claim 1, characterized in that: A corresponding autoencoder model is established according to the plurality of cluster sub-cluster training data, including: Performing standardization processing on the plurality of cluster sub-cluster training data; Acquire a preset amount of data from the plurality of cluster sub-cluster training data as a verification set; Taking the root mean square error of the validation set as an indicator, the plurality of cluster sub-cluster training data are used to repeatedly train the plurality of autoencoder models until the indicator stops decreasing and starts to increase.

5. The multi-connection fault detection method according to claim 4, characterized in that: The method further comprises: Get the maximum value and median value of the reconstruction error output by each autoencoder model; If the maximum value of the reconstruction error is greater than a preset multiple of the median value of the reconstruction error, it is determined that the autoencoder model is abnormal and the autoencoder model is eliminated.

6. The multi-connection fault detection method according to claim 1, characterized in that: Determining the reconstruction error threshold includes: Determine the number of reconstruction error scatter points corresponding to each cluster sub-cluster training data; The weighted average value of the number of reconstruction error scatter points within a preset interval is determined as the reconstruction error threshold.

7. The multi-connection fault detection method according to claim 1, characterized in that: After acquiring the multi-connection accumulated operation data set, the method further includes: The running data set is preprocessed, and the preprocessing includes at least one of data merging, data deduplication, outlier processing, running segment extraction, feature calculation, time resampling, missing value processing and feature extraction.

8. The multi-connection fault detection method according to claim 1, characterized in that: The method further comprises: Feature classification and redundancy processing are performed on the operation data set to generate multiple groups of operation data sets.

9. The multi-connection fault detection method according to claim 1, characterized in that: The operating conditions include cooling conditions, heating non-enthalpy injection conditions and heating enthalpy injection conditions. The data pool corresponding to the cooling conditions is the cooling conditions data pool, the data pool corresponding to the heating non-enthalpy injection conditions is the heating non-enthalpy injection conditions data pool, and the data pool corresponding to the heating enthalpy injection conditions is the heating enthalpy injection conditions data pool.

10. The multi-connection fault detection method according to claim 1, characterized in that: Before inputting the current day's operating data set into the clustering model and the autoencoder model, the method further includes: Determine that the data pool corresponding to the current day's operating data set has completed training, determine that the data volume of the current day's operating data set is greater than or equal to the test lower limit length, and determine that the data pool matching the operating condition corresponding to the current day's operating data set has completed training.

11. A computer-readable storage medium, characterized in that: A multi-connection fault detection program is stored thereon, and when the program is executed by a processor, the multi-connection fault detection method described in any one of claims 1-10 is implemented.

12. A multi-connection fault detection device, characterized in that: The device comprises: An acquisition module, used for acquiring the operation data set accumulated by the multi-connection; A data processing module, used for traversing the operation data set on a daily basis, so as to store the daily operation data set into a corresponding data pool according to the operation conditions of the multi-link; A training module, used for training the data in the at least one data pool when the amount of data stored in the at least one data pool reaches a lower limit length, so as to obtain a clustering model, an autoencoder model and a reconstruction error threshold; A testing module, used for inputting the operating data set of the current day into the clustering model and the autoencoder model to output a reconstruction error test value when the operating condition corresponding to the operating data set of the current day is the same as the operating condition corresponding to the trained data pool; A detection module, used for detecting a fault state of the multi-link according to the reconstruction error test value and the reconstruction error threshold; The data processing module is further used for: when it is determined that the multi-connected machine is in a fault state according to the detection of the operation data set of the current day, recording the operation data set of the current day; when it is determined that the multi-connected machine is not in a fault state according to the detection of the operation data set of the current day, and the multi-connected machines corresponding to the operation data sets of the first preset number of days before the current day and the operation data sets of the second preset number of days after the current day are not in a fault state, storing the operation data set of the current day in the corresponding data pool; The training module is also used to: perform cluster analysis on the data in the data pool to obtain the cluster model and several cluster sub-cluster training data; establish corresponding autoencoder models according to the several cluster sub-cluster training data, and determine the reconstruction error threshold.

13. A multi-link system, characterized in that: It includes the multi-connection fault detection device as described in claim 12.

Images