Device anomaly detection method and device and storage medium

Abstract

A device anomaly detection method, apparatus, and storage medium are disclosed. The method comprises: obtaining a feature data set within a first time period of the device; determining m interval points based on a comparison of a first time difference between the collection times of adjacent feature data information with a set collection period; and dividing the feature data set into (|m+1|) subsets based on the m interval points, where m ≥ 0; constructing a plurality of first sample sets for each subset based on a tracing period size n; inputting a sample subset of each first sample set into a convolutional neural network; and obtaining prediction information corresponding to each sample subset output by the convolutional neural network; screening out abnormal information from the prediction information corresponding to all sample subsets; interpreting the cause of the abnormality for the first sample set corresponding to the abnormal information using a Shapley additive model interpreter, and obtaining an interpretation result; and determining the device abnormality point based on the interpretation result. This method can solve the problem of the inability to timely discover and locate the cause of the device abnormality and the abnormality point.

Description

Technical Field

[0001] The present application relates to the field of equipment monitoring technology, and in particular to a method and apparatus for detecting equipment anomaly, and a storage medium. Background Art

[0002] Production equipment in factories (such as ice machines, air compressors, etc.) may experience equipment abnormalities during the production process, and equipment abnormalities will seriously affect the achievement of production goals and cause losses to manufacturers.

[0003] During the production process, regular maintenance of production equipment by technicians is not an effective way to prevent equipment failures. In some application scenarios, monitoring models can be used to monitor equipment data and detect anomalies based on this data. However, when the monitored data indicates an equipment anomaly, the cause of the anomaly cannot be discovered promptly, and the location of the anomaly cannot be accurately located, resulting in a high rate of equipment failure. Summary of the Invention

[0004] The embodiments of the present application provide a device anomaly detection method and apparatus as well as a storage medium, which can classify data based on time difference, construct a training set (X, Y) for similar data, and directly splice and merge different classes, which can effectively solve the time gap problem of time series data.

[0005] In a first aspect, an embodiment of the present application provides a device abnormality detection method, the method comprising: obtaining a feature data set within a first time period of the device, the feature data set comprising multiple feature data information sorted by collection time, and each feature data information comprising the collection time of the current feature data information and feature values ​​of several features, wherein the several features are several types of monitoring information during the operation of the device; determining m interval points based on a comparison result of a first time difference between the collection times of adjacent feature data information and a set collection period, and dividing the feature data set into (|m+1|) subsets based on the m interval points, where m≥0; constructing several first sample sets for each subset based on a tracing period size n, wherein each first sample set comprises a sample subset, and the sample subset comprises n feature data information with continuous collection times, wherein a first time difference exists between any two adjacent feature data information, and the continuous collection time means that the first time difference and the second time difference between the set collection period are not greater than a time interval threshold;

[0006] Inputting the sample subset of each of the first sample sets into a convolutional neural network, and obtaining prediction information corresponding to each of the sample subsets output by the convolutional neural network; screening out abnormal information in the prediction information corresponding to all of the sample subsets; interpreting the cause of the abnormality for the first sample set corresponding to the abnormal information based on a Shapley additive model interpreter, and obtaining an interpretation result; and determining the abnormal point of the device based on the interpretation result.

[0007] Furthermore, the determining of m interval points based on a comparison result of the first time difference between the acquisition times of adjacent feature data information and a set acquisition period, and dividing the feature data set into (|m+1|) subsets based on the m interval points includes: traversing the feature data set, determining whether the first time difference between the acquisition times of adjacent feature data information is greater than the set acquisition period, and if so, determining that there is an interval point between the adjacent feature data information, determining m interval points based on the comparison result, and dividing the feature data set into (|m+1|) subsets based on the m interval points; if not, determining that there is no interval point between the adjacent feature data information.

[0008] Furthermore, before constructing a plurality of first sample sets for each subset based on the tracing period size n, the method further includes: determining the tracing period size n based on the first time period and the collection period; wherein, n represents the size of the tracing period, T represents the first period, and s represents the collection period.

[0009] Furthermore, each first sample set also includes focus data, and constructing several first sample sets for each subset based on the tracing period size n includes: according to the order of the feature data information in the subset, starting with the first feature data information in the subset, and sequentially using the feature data information as the starting element of the sample subset to construct multiple rounds of first sample sets until all the subsets have completed the construction of the first sample set; wherein, the construction process of one round of first sample sets includes: starting from the starting element, obtaining a total of n feature data information with continuous acquisition time as the sample subset of the current first sample set, and obtaining the (n+1)th feature data information as the focus data of the current first sample set; wherein, starting from the current starting element, when the (n+1)th feature data information is the last feature data information in the subset, the construction of the last round of first sample sets of the current subset is completed.

[0010] Furthermore, filtering out abnormal information from the prediction information corresponding to all the sample subsets includes: comparing the prediction information corresponding to each of the sample subsets with the focus data corresponding to the corresponding sample subset, and obtaining an abnormal score; and obtaining abnormal focus data, wherein the abnormal focus data is the focus data whose abnormal score exceeds a score threshold.

[0011] Furthermore, the score threshold is 3 times the standard deviation of all the abnormal scores.

[0012] Furthermore, the Shapley additive model interpreter is used to interpret the cause of the abnormality of the first sample set corresponding to the abnormal information, and the interpretation result obtained includes: calculating the characteristic error between each eigenvalue in each of the abnormal focus data and the corresponding eigenvalue in the prediction information of the corresponding sample subset; calculating the total error corresponding to each of the abnormal focus data, and the total error is the sum of the characteristic errors corresponding to each eigenvalue in the abnormal focus data; sorting the characteristic errors of each eigenvalue in each of the abnormal focus data from large to small based on the error value, and obtaining the target characteristic information in which the sum of the characteristic errors from large to small in the sorting exceeds the proportion threshold of the total error; calculating the Shapley value corresponding to each of the target characteristic information based on the Shapley additive model interpreter; and determining the abnormal factor of the abnormal focus data based on the Shapley value corresponding to each of the target characteristic information in each of the abnormal focus data.

[0013] Furthermore, the determining of the abnormal factor of the abnormal focus data based on the Shapley value corresponding to each target feature information in each of the abnormal focus data includes: calculating a weighted average based on the Shapley value corresponding to each target feature information in each of the abnormal focus data; obtaining the weighted averages of the k targets before the weighted averages among the weighted averages corresponding to all the abnormal focus data; and determining the abnormal factor of the abnormal focus data based on the Shapley values ​​corresponding to the k target weighted averages.

[0014] Furthermore, determining the abnormal point of the device based on the interpretation result includes: after determining the abnormal factor of the abnormal focus data, highlighting the abnormal factor of the abnormal focus data and the Shapley value of the abnormal factor; obtaining the abnormal factor with a positive Shapley value among the abnormal factors, and determining that the abnormal factor with a positive Shapley value represents the current abnormal point of the device.

[0015] Furthermore, when the device is an ice machine device, the several monitoring information during the operation of the device includes at least one or more of the following: ice water inlet temperature, oil pressure difference, oil tank outlet temperature, motor power and saturated evaporation temperature during the operation of the ice machine device.

[0016] In a second aspect, an embodiment of the present application further provides a device for detecting anomalies of an equipment, which includes: a processor and a memory, wherein the memory is used to store at least one instruction, and when the instruction is loaded and executed by the processor, the device anomaly detection method provided in the first aspect is implemented.

[0017] In a third aspect, an embodiment of the present application further provides a device abnormality detection apparatus, the apparatus comprising:

[0018] an acquisition module, configured to acquire a feature data set of a device within a first time period, the feature data set comprising a plurality of feature data information sorted by collection time, and each feature data information comprising a collection time of the current feature data information and feature values ​​of several features, the several features being several types of monitoring information during operation of the device;

[0019] The analysis module is used to perform the following steps: determining m interval points based on a comparison result of a first time difference between the acquisition times of adjacent feature data information and a set acquisition period, and dividing the feature data set into (|m+1|) subsets based on the m interval points, where m≥0; constructing a plurality of first sample sets for each subset based on a tracing period size n, wherein each first sample set contains a sample subset, and the sample subset contains n feature data information with continuous acquisition times, wherein a first time difference exists between any two adjacent feature data information, and the continuous acquisition time means that a second time difference between the first time difference and the set acquisition period is not greater than a time interval threshold; inputting the sample subset of each first sample set into a convolutional neural network, and obtaining prediction information corresponding to each sample subset output by the convolutional neural network; screening out abnormal information in the prediction information corresponding to all the sample subsets; interpreting the cause of the abnormality for the first sample set corresponding to the abnormal information based on a Shapley additive model interpreter, and obtaining an explanation result; and determining the abnormal point of the device according to the explanation result.

[0020] In a fourth aspect, an embodiment of the present application further provides a computer-readable non-volatile storage medium having a computer program stored thereon, which, when executed by a processor, implements the device anomaly detection method provided in the first aspect. BRIEF DESCRIPTION OF THE DRAWINGS

[0021] In order to more clearly illustrate the embodiments of the present application or the technical solutions in the prior art, a brief introduction will be given below to the drawings required for use in the embodiments or the description of the prior art. Obviously, the drawings described below are some embodiments of the present application. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative labor.

[0022] Figure 1 A flowchart of a device anomaly detection method provided in one embodiment of the present application;

[0023] Figure 2 A schematic diagram of a feature data set provided for one embodiment of the present application;

[0024] Figure 3 A schematic diagram of sample subset training provided for one embodiment of the present application;

[0025] Figure 4 A schematic diagram of abnormal factors of device abnormalities provided in one embodiment of the present application;

[0026] Figure 5 A schematic diagram of the structure of a device anomaly detection apparatus provided in one embodiment of the present application. DETAILED DESCRIPTION

[0027] To make the purpose, technical solutions, and advantages of the embodiments of this application more clear, the technical solutions in the embodiments of this application will be clearly and completely described below in conjunction with the drawings in the embodiments of this application. Obviously, the described embodiments are part of the embodiments of this application, not all of the embodiments. Based on the embodiments in this application, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of this application.

[0028] Figure 1 A flowchart of a device anomaly detection method provided in accordance with an embodiment of the present application.

[0029] Reference Figure 1 As shown, the detection method may include:

[0030] Step 101: Acquire a feature data set of a device within a first period of time.

[0031] In one embodiment, obtaining a feature data set of a device within a first period of time may be obtaining a data set of device operation data of one or more target devices within 30 minutes. For example, the target devices may be a refrigerator and an air compressor (hereinafter referred to as an air compressor).

[0032] In one embodiment, the feature data set includes a plurality of feature data information sorted by acquisition time. Figure 2 A schematic diagram of a feature data set provided for an embodiment of the present application. Figure 2 As shown, the acquired multiple feature data information is used as time series samples for training: x0, x1, x2, x3, x4, x5, x6, x7, x8, x9, x10, x11, x12, x13... Among them, the multiple feature data information can be collected according to a preset collection period, such as Figure 2 As shown, the preset collection period is 3 minutes, and thus a feature data information can be collected every 3 minutes. For example, after starting to collect feature data information, the first data collected is taken as x0. After 3 minutes, it is collected again, and the next data after x0 is collected as x1, and so on until the collection is completed.

[0033] In one embodiment, each feature data item includes the collection time of the current feature data item and feature values ​​of several features, where the several features are several types of monitoring information during the operation of the device. For example, the several types of monitoring information during the operation of the device may include one or more of the following: chilled water inlet temperature, oil pressure difference, oil tank outlet temperature, motor power, and saturated evaporation temperature during the operation of the ice machine.

[0034] Step 102: Determine m interval points based on a comparison result of a first time difference between collection times of adjacent feature data information and a set collection period, and divide the feature data set into (|m+1|) subsets based on the m interval points.

[0035] In the process of collecting characteristic data information, time gaps may occur. Figure 2 As shown, the collection time of the two adjacent feature data information shown in the dotted box is "02:06" and "03:39" respectively, and the time difference between the collection time of the two adjacent feature data information is greater than the collection period (3min), and it can be determined that there is a time gap between the two adjacent feature data information.

[0036] Feature data information with time faults as training samples will affect the accuracy of the data. In order to reduce the impact of time faults on data accuracy, in one embodiment, the feature data set can be divided based on time faults, and then the divided feature data set can be processed accordingly to reduce the impact of time faults on data accuracy.

[0037] In one embodiment, m interval points can be determined based on the comparison result of the first time difference between the acquisition times of adjacent feature data information and the set acquisition period, and the feature data set can be divided into (|m+1|) subsets based on the m interval points. Specifically, the feature data set can be traversed to determine whether the first time difference between the acquisition times of adjacent feature data information is greater than the set acquisition period. If so, it is determined that there is an interval point between the adjacent feature data information, that is, each time it is determined that the first time difference between the acquisition times of a pair of adjacent feature data information is greater than the set acquisition period, 1 interval point is determined. After traversing the feature data set, m interval points can be determined, and then the feature data set can be divided into (|m+1|) subsets based on the m interval points; if not, it is determined that there is no interval point between the adjacent feature data information. Exemplarily, after traversing the feature data set, 1 interval point is determined, that is, m=1, and then the feature data set can be divided into (|1+1|) subsets, that is, the feature data set can be divided into 2 subsets based on the interval point. Reference Figure 2As shown, the feature data set contains only one time fault (i.e., one interval point), which is the interval point between the acquisition times "02:06" and "03:39" of the two adjacent feature data information shown in the dotted box. The feature data set can then be divided into a first subset and a second subset, wherein the first feature data information of the feature data set serves as the initial data of the first subset, and the feature data information corresponding to the acquisition time "02:06" serves as the cutoff data of the first subset. The feature data information corresponding to the acquisition time "03:39" serves as the initial data of the second subset, and the original cutoff data in the feature data set serves as the cutoff data of the second subset. The division of the feature data set is completed in the above manner. It should be noted that if there is no time fault in the feature data set, i.e., m=0, the feature data set is divided into (|0+1|) subsets. In other words, there is no need to divide the feature data set.

[0038] After the feature data set is divided in the above manner, the divided feature data set (including the undivided feature data set) can be subjected to corresponding data splicing and merging processing to reduce the impact of time gaps on data accuracy. This can be achieved specifically by the following step 103 .

[0039] Step 103: constructing a plurality of first sample sets for each subset based on the tracing period size n.

[0040] In one embodiment, the following steps may be further performed before performing step 103:

[0041] Determine the tracing period size n based on the first time period and the acquisition period;

[0042] in, n represents the lookback period size (lookback_size), T represents the first time period, and s represents the collection period.

[0043] For example, the first time period is 30 minutes, the collection period is 3 minutes, and the tracing period size can be calculated to be 10, that is, n=10.

[0044] After determining the tracing period size n, step 103 is executed, i.e., constructing a plurality of first sample sets for each subset based on the tracing period size n. In one embodiment, each first sample set also includes focal data. Constructing a plurality of first sample sets for each subset based on the tracing period size n includes: starting with the first feature data information in the subset, sequentially using the feature data information as the starting element of the sample subset according to the order of the feature data information in the subset, until all the subsets have completed the construction of the first sample set; wherein the process of constructing a round of the first sample set includes: starting from the starting element, obtaining a total of n pieces of feature data information with continuous acquisition time as the sample subset of the current first sample set, and obtaining the (n+1)th feature data information as the focal data of the current first sample set; wherein, starting from the current starting element, when the (n+1)th feature data information is the last feature data information in the subset, the construction of the last round of the first sample set for the current subset is completed. wherein, the continuous acquisition time refers to that the first time difference between any two adjacent feature data information and the second time difference between the set acquisition periods are not greater than the time interval threshold. Exemplarily, if the tracing period size n=10, taking the first feature data information (x0) in the first subset as the current starting element, obtain the feature data information of the first 10 consecutive acquisition times, that is, (x0, x1, x2, x3, x4, x5, x6, x7, x8, x9) as the sample subset X0 of the current first sample set, obtain the current (10+1)th feature data information, that is, (x10) as the focus data Y0 of the current first sample set, and then obtain the current first sample set (X0, Y0). After obtaining a first sample set, the collection order of the feature data information is postponed one position as the current starting element. If the starting element of the previous round is (x0), the starting element of this round (i.e., the current starting element) is (x1), and the feature data information of the first 10 consecutive collection times obtained are (x1, x2, x3, x4, x5, x6, x7, x8, x9, x10), that is, the sample subset X1 of the current first sample set = (x1, x2, x3, x4, x5, x6, x7, x8, x9, x10), and accordingly, the current (10+1)th feature data information is (x11), that is, the focus data Y1 of the first sample set = (x11), that is, the current first sample set (X1, Y1) is obtained. This process is repeated until the (n+1)th feature data information is the last feature data information in the subset, and the first sample set of the current subset is constructed, and the first sample set of all the subsets is constructed. In any sample subset, there is a first time difference between any two adjacent feature data information among the 10 feature data information with consecutive collection times, and the first time difference and the second time difference between the set collection period are not greater than the time interval threshold.Here, taking the sample subset X1=(x1, x2, x3, x4, x5, x6, x7, x8, x9, x10) as an example, the collection period is set to 3 minutes, and there is a first time difference Δt1 between the collection times of any two adjacent feature data information, and the second time difference Δt2 between each first time difference Δt1 and the set collection period (3 minutes) is not greater than the time interval threshold, which can be 0. It should be noted that after step 102, the feature data set is divided into (|m+1|) subsets, and each subset constructs a plurality of first sample sets. In the scenario where the collection period is set to 3 minutes, the maximum first time difference Δt1 between any two adjacent feature data information in the sample subset of each first sample set is 3 minutes, and thus the difference between the maximum first time difference Δt1 (3 minutes) and the set collection period (3 minutes) (i.e., the second time difference) is 0, and thus the time interval threshold can be set to 0.

[0045] Step 104: Input each of the sample subsets of the first sample set into a convolutional neural network, and obtain prediction information corresponding to each of the sample subsets output by the convolutional neural network.

[0046] By constructing the first sample sets of all subsets through the above step 103, all the obtained first sample sets can be trained. In one embodiment, a DeepCNN (deep learning convolutional neural network) model can be used to perform autoencoding training on all first sample sets. In the above autoencoding training, a sample subset of each first sample set can be input into a convolutional neural network, and the prediction information corresponding to each sample subset output by the convolutional neural network is obtained. Figure 3 A sample subset training diagram provided for an embodiment of the present application. Figure 3 As shown, the sample subsets of all the first sample sets obtained can be input into the convolutional neural network. For example, h first sample sets are obtained through step 103, and then in the autoencoder training, the sample subsets of the h first sample sets [X0, X1, X2...X h-1 , X h ] input into the convolutional neural network, through which the prediction information corresponding to each sample subset [Y'0, Y'1, Y'2…Y' h-1 , Y' h ].

[0047] Step 105: Filter out abnormal information in the prediction information corresponding to all the sample subsets.

[0048] In one embodiment, after predicting the prediction information corresponding to all sample subsets through a convolutional neural network, the abnormal information in the prediction information corresponding to all sample subsets can be further screened out. In one embodiment, the abnormal information can be screened out by comparing the prediction information corresponding to each sample subset with the focus data corresponding to the corresponding sample subset, and obtaining an abnormal score, and then obtaining abnormal focus data, wherein the abnormal focus data is the focus data whose abnormal score exceeds the score threshold. It should be noted that for each first sample set, it contains a sample subset and focus data, the sample subset contains the operation data of the device for a previous period of time, and the focus data is the actual operation data of the device at a subsequent time point. The above-mentioned deep learning convolutional neural network can predict the prediction information of the sample subset (the operation data of the device for a previous period of time), that is, calculate the predicted operation data of the device at a subsequent time point. Furthermore, the focus data (actual operation data) can be compared with the prediction information (predicted operation data) to obtain the abnormal focus data. Reference Figure 3 As shown, abnormal information screening can be to filter the prediction information [Y'0, Y'1, Y'2...Y' h-1 , Y' h ] and the corresponding focus data [Y0, Y1, Y2…Y h-1 , Y h ] and compare them to obtain an anomaly score (loss). After obtaining the anomaly score, the anomaly score can be filtered. In one embodiment, the focus data with an anomaly score exceeding a score threshold can be filtered out, and the focus data with an anomaly score exceeding the score threshold can be considered as abnormal focus data. In one embodiment, the score threshold can be 3 times the standard deviation of all anomaly scores.

[0049] After the abnormal focus data is determined, the abnormal cause of the abnormal focus data may be explained, which may be specifically achieved through step 106 .

[0050] Step 106: Based on the Shapley additive model interpreter, the cause of the abnormality is interpreted for the first sample set corresponding to the abnormal information, and an interpretation result is obtained.

[0051] In one embodiment, the specific steps of explaining the cause of the abnormality may include:

[0052] Calculate the characteristic error between each eigenvalue in each abnormal focus data and the corresponding eigenvalue in the prediction information of the corresponding sample subset; the characteristic error calculation formula between each eigenvalue and the corresponding eigenvalue in the prediction information of the corresponding sample subset is |Y' (i，j) -|Y (i，j) |, j = 0, 1,…, h.

[0053] Calculate the total error corresponding to each abnormal focus data. The total error is the sum of the characteristic errors corresponding to each eigenvalue in the abnormal focus data.

[0054] Based on the error values, the feature errors of each feature value in each abnormal focus data are sorted from large to small, and target feature information is obtained in which the sum of the feature errors from large to small in the sorting exceeds a percentage threshold of the total error. In one embodiment, the percentage threshold can be 80%, that is, target feature information is obtained in which the sum of the feature errors from large to small in the sorting exceeds 80% of the total error. Among them, the corresponding features with a large error percentage in each feature value can be focused on. For example, several features during equipment operation include: ice water inlet temperature, oil pressure difference, oil tank outlet temperature, motor power, and saturated evaporation temperature during operation of the ice machine equipment. In an application scenario with a percentage threshold of 80%, if the error values ​​corresponding to the oil tank outlet temperature and saturated evaporation temperature account for 82% of the total error, priority attention can be given to the equipment's oil tank outlet temperature and saturated evaporation temperature, so that the larger abnormal points during equipment operation can be prioritized, thereby ensuring the basic stability of equipment operation.

[0055] The Shapley value (shap value) corresponding to each target feature information is calculated based on the Shapley additivity model interpreter (shap interpreter); in one embodiment, the dimension of the Shapley value corresponding to each target feature information is (lookback_size=10, features=h).

[0056] Determining the abnormal factor of the abnormal focus data based on the Shapley value corresponding to each target feature information in each abnormal focus data. In one embodiment, determining the abnormal factor of the abnormal focus data based on the Shapley value includes:

[0057] Calculate the weighted average based on the Shapley value corresponding to each target feature information in each abnormal focus data;

[0058] The weighted average values ​​corresponding to all abnormal focus data are obtained, and the weighted average values ​​of the first k targets are weighted averaged; in one embodiment, k=50, that is, the weighted average values ​​of the first 50 targets can be weighted averaged.

[0059] The abnormal factor of the abnormal focus data is determined based on the Shapley value corresponding to the weighted average of k targets.

[0060] After determining the abnormal factors of the abnormal focus data, the interpretation results are obtained.

[0061] Figure 4 A schematic diagram of abnormal factors of device abnormalities provided in one embodiment of the present application.

[0062] Reference Figure 4As shown in the figure, ECW_tt27 represents the variable ECW from 27 minutes ago. Shaded boxes indicate positive shap values ​​for this feature, positively impacting the results; vertically shaded boxes indicate negative shap values, negatively impacting the results. Positive factors that may contribute to abnormal results include: LCDW (chilled water inlet temperature), KW (motor power), OIL_PD (oil pressure differential), and CTRL_PNT (oil tank outlet temperature) at different times in the past. These factors should be immediately investigated to prevent equipment abnormalities, production disruptions, and company losses.

[0063] Step 107: Determine abnormal points of the device according to the interpretation result.

[0064] After obtaining the interpretation results, the abnormal point of the device may be determined based on the interpretation results. In one embodiment, the abnormal point may be determined by:

[0065] After determining the abnormal factors of the abnormal focus data, the abnormal factors of the abnormal focus data and the Shapley values ​​of the abnormal factors are highlighted;

[0066] The abnormal factors with positive Shapley values ​​among the abnormal factors are obtained, and it is determined that the abnormal factors with positive Shapley values ​​represent the current abnormal points of the equipment.

[0067] Through the above steps 101 to 107, no data filling method is used to solve the problem of time gaps between adjacent feature data information. In combination with the autoencoding training results of the DeepCNN model and based on the SHAP interpreter, the cause of equipment abnormality is discovered in time and the abnormal point of the equipment is accurately located. This provides a guarantee for timely finding the cause of the abnormality when the equipment is abnormal, reducing the occurrence rate of equipment failure and improving production efficiency.

[0068] Figure 5 This is a schematic diagram of the structure of a device abnormality detection device provided by an embodiment of the present application. Figure 5 As shown, the device may include a processor 501 and a memory 502, where the memory 502 is used to store at least one instruction, which, when loaded and executed by the processor 501, implements the device anomaly detection method provided in any embodiment of the present application.

[0069] Another embodiment of the present application further provides a device abnormality detection apparatus, which may include:

[0070] an acquisition module, configured to acquire a feature data set of the device within a first period of time, the feature data set comprising a plurality of feature data information sorted by collection time, and each feature data information comprising a collection time of the current feature data information and feature values ​​of several features, wherein the several features are several types of monitoring information during the operation of the device;

[0071] An analysis module that performs the following steps:

[0072] Determine m interval points based on a comparison result of a first time difference between collection times of adjacent feature data information and a set collection period, and divide the feature data set into (|m+1|) subsets based on the m interval points, where m≥0;

[0073] Based on the tracing period size n, a plurality of first sample sets are constructed for each subset, wherein each first sample set contains a sample subset, and the sample subset contains n pieces of feature data information with continuous collection time, wherein a first time difference exists between any two adjacent pieces of feature data information, and the continuous collection time means that the first time difference and a second time difference between the set collection period are not greater than a time interval threshold;

[0074] Inputting each sample subset of the first sample set into a convolutional neural network, and obtaining prediction information corresponding to each sample subset output by the convolutional neural network;

[0075] Filter out abnormal information in the prediction information corresponding to all sample subsets;

[0076] Based on the Shapley additive model interpreter, the cause of the abnormality is explained for the first sample set corresponding to the abnormal information, and an explanation result is obtained;

[0077] Determine the abnormal points of the equipment based on the interpretation results.

[0078] Another embodiment of the present application further provides a computer-readable storage medium having a computer program stored thereon. When the computer program is executed by a processor, the device anomaly detection method provided in any embodiment of the present application is implemented.

[0079] Another embodiment of the present application further provides a computer program product, including a computer program or instructions, which, when executed by a processor, implements the device anomaly detection method provided in any embodiment of the present application.

[0080] It should be noted that the terminals involved in the embodiments of the present application may include but are not limited to personal computers (PCs), personal digital assistants (PDAs), wireless handheld devices, tablet computers, mobile phones, MP3 players, MP4 players, etc.

[0081] It is understandable that the application may be an application program (nativeApp) installed on the terminal, or may be a web page program (webApp) of a browser on the terminal, and this embodiment of the present application does not limit this.

[0082] Those skilled in the art will clearly understand that, for the convenience and brevity of description, the specific working processes of the systems, devices and units described above can refer to the corresponding processes in the aforementioned method embodiments and will not be repeated here.

[0083] In the several embodiments provided in this application, it should be understood that the disclosed systems, devices and methods can be implemented in other ways. For example, the device embodiments described above are merely schematic. For example, the division of the units is merely a logical function division. In actual implementation, there may be other division methods. For example, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed. Another point is that the mutual coupling or direct coupling or communication connection shown or discussed may be an indirect coupling or communication connection through some interface, device or unit, which may be electrical, mechanical or other forms.

[0084] The units described as separate components may or may not be physically separate, and the components shown as units may or may not be physical units, that is, they may be located in one place or distributed across multiple network units. Some or all of these units may be selected to achieve the purpose of this embodiment according to actual needs.

[0085] In addition, the functional units in the various embodiments of the present application may be integrated into a single processing unit, or each unit may exist physically separately, or two or more units may be integrated into a single unit. The aforementioned integrated units may be implemented in the form of hardware or in the form of hardware plus software functional units.

[0086] The above-mentioned integrated unit implemented in the form of a software functional unit can be stored in a computer-readable storage medium. The above-mentioned software functional unit is stored in a storage medium and includes a number of instructions for causing a computer device (which can be a personal computer, server, or network device, etc.) or a processor to perform some steps of the method described in each embodiment of the present application. The aforementioned storage medium includes: a USB flash drive, a mobile hard disk, a read-only memory (ROM), a random access memory (RAM), a magnetic disk or an optical disk, and other media that can store program code.

[0087] The above description is only a preferred embodiment of the present application and is not intended to limit the present application. Any modifications, equivalent replacements, improvements, etc. made within the spirit and principles of the present application shall be included in the scope of protection of the present application.

[0088] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present application, rather than to limit them. Although the present application has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the aforementioned embodiments, or make equivalent replacements for some or all of the technical features therein. These modifications or replacements do not deviate the essence of the corresponding technical solutions from the scope of the technical solutions of the embodiments of the present application.

Claims

1. A device anomaly detection method, characterized in that: The method comprises: Acquire a feature data set of the device within a first time period, the feature data set comprising a plurality of feature data information sorted by collection time, and each feature data information comprising a collection time of the current feature data information and feature values ​​of several features, the several features being several types of monitoring information during operation of the device; Determine m interval points based on a comparison result of a first time difference between the collection times of adjacent feature data information and a set collection period, and divide the feature data set into (|m+1|) subsets based on the m interval points, where m≥0; Determine a tracing period size n based on the first time period and the collection period; in, n represents the size of the tracing period, T represents the first period, and s represents the collection period; Based on the tracing period size n, a plurality of first sample sets are constructed for each subset, wherein each first sample set includes a sample subset, and the sample subset includes n pieces of feature data information collected at consecutive times, wherein the first time difference exists between any two adjacent pieces of feature data information, and the consecutive collection time means that the first time difference and a second time difference between the set collection period are not greater than a time interval threshold; Inputting the sample subset of each of the first sample sets into a convolutional neural network, and obtaining prediction information corresponding to each of the sample subsets output by the convolutional neural network; Screening out abnormal information in the prediction information corresponding to all the sample subsets; Performing an abnormality cause interpretation on the first sample set corresponding to the abnormal information based on a Shapley additive model interpreter, and obtaining an interpretation result; and An abnormal point of the equipment is determined according to the interpretation result.

2. The method according to claim 1, characterized in that The determining of m interval points based on a comparison result of the first time difference between the acquisition times of adjacent feature data information and a set acquisition period, and dividing the feature data set into (|m+1|) subsets based on the m interval points includes: Traverse the feature data set to determine whether the first time difference between the acquisition times of adjacent feature data information is greater than the set acquisition period; if so, determine that there is an interval point between the adjacent feature data information, determine m interval points based on the comparison result, and divide the feature data set into (|m+1|) subsets based on the m interval points; if not, determine that there is no interval point between the adjacent feature data information.

3. The method according to claim 1 or 2, characterized in that Each first sample set also includes focus data. The step of constructing a plurality of first sample sets for each subset based on the tracing period size n includes: According to the order of the characteristic data information in the subsets, starting with the first characteristic data information in the subsets, the characteristic data information is sequentially used as the starting element of the sample subsets to construct the first sample set for multiple rounds until all the subsets have completed the construction of the first sample set; The process of constructing the first sample set in a round includes: Starting from the starting element, a total of n pieces of feature data information with continuous acquisition time are obtained as a sample subset of the current first sample set, and the (n+1)th feature data information is obtained as the focus data of the current first sample set; Wherein, starting from the current starting element, when the (n+1)th feature data information is the last feature data information in the subset, the construction of the last round of the first sample set of the current subset is completed.

4. The method according to claim 1, wherein The step of filtering out abnormal information in the prediction information corresponding to all the sample subsets includes: Comparing the prediction information corresponding to each of the sample subsets with the focus data corresponding to the corresponding sample subset, and obtaining an anomaly score; and Abnormal focus data is acquired, where the abnormal focus data is focus data whose abnormal score exceeds a score threshold.

5. The method according to claim 4, characterized in that The score threshold is 3 times the standard deviation of all the anomaly scores.

6. The method according to claim 4, characterized in that The Shapley additive model-based interpreter performs an abnormality cause interpretation on the first sample set corresponding to the abnormal information, and obtains an interpretation result including: Calculating a characteristic error between each characteristic value in each of the abnormal focus data and a corresponding characteristic value in the prediction information of the corresponding sample subset; Calculating a total error corresponding to each abnormal focus data, where the total error is the sum of the feature errors corresponding to each feature value in the abnormal focus data; Sort the feature errors of each feature value in each of the abnormal focus data based on the error values ​​from large to small, and obtain target feature information in which the sum of the feature errors from large to small in the sorting accounts for a total error that exceeds a percentage threshold; Calculating the Shapley value corresponding to each target feature information based on the Shapley additive model interpreter; The abnormal factor of the abnormal focus data is determined based on the Shapley value corresponding to each target feature information in each abnormal focus data.

7. The method according to claim 6, characterized in that Determining the abnormal factor of the abnormal focus data based on the Shapley value corresponding to each target feature information in each abnormal focus data includes: Calculating a weighted average based on the Shapley value corresponding to each target feature information in each abnormal focus data; Obtain the weighted average values ​​of the top k targets among the weighted average values ​​corresponding to all abnormal focus data; And the abnormal factor of the abnormal focus data is determined based on the Shapley value corresponding to the k target weighted average values.

8. The method according to claim 6 or 7, characterized in that Determining the abnormal point of the device according to the interpretation result includes: After determining the abnormal factor of the abnormal focus data, highlighting the abnormal factor of the abnormal focus data and the Shapley value of the abnormal factor; An abnormal factor with a positive Shapley value among the abnormal factors is obtained, and it is determined that the abnormal factor with a positive Shapley value represents a current abnormal point of the device.

9. The method according to any one of claims 1, 2, 4-7, characterized in that: When the device is an ice machine, the monitoring information during operation of the device includes at least one or more of the following: ice water inlet temperature, oil pressure difference, oil tank outlet temperature, motor power, and saturated evaporation temperature during operation of the ice machine.

10. A device for detecting abnormality of equipment, characterized in that: The device comprises: A processor and a memory, wherein the memory is used to store at least one instruction, and when the instruction is loaded and executed by the processor, the method according to any one of claims 1 to 9 is implemented.

11. A device for detecting abnormality of equipment, characterized in that: The device comprises: an acquisition module, configured to acquire a feature data set of a device within a first time period, the feature data set comprising a plurality of feature data information sorted by collection time, and each feature data information comprising a collection time of the current feature data information and feature values ​​of several features, the several features being several types of monitoring information during operation of the device; An analysis module that performs the following steps: Determine m interval points based on a comparison result of a first time difference between the collection times of adjacent feature data information and a set collection period, and divide the feature data set into (|m+1|) subsets based on the m interval points, where m≥0; Determine a tracing period size n based on the first time period and the collection period; in, n represents the size of the tracing period, T represents the first period, and s represents the collection period; Based on the tracing period size n, a plurality of first sample sets are constructed for each subset, wherein each first sample set includes a sample subset, and the sample subset includes n pieces of feature data information collected at consecutive times, wherein the first time difference exists between any two adjacent pieces of feature data information, and the consecutive collection time means that the first time difference and a second time difference between the set collection period are not greater than a time interval threshold; Inputting the sample subset of each of the first sample sets into a convolutional neural network, and obtaining prediction information corresponding to each of the sample subsets output by the convolutional neural network; Screening out abnormal information in the prediction information corresponding to all the sample subsets; Performing an abnormality cause interpretation on the first sample set corresponding to the abnormal information based on a Shapley additive model interpreter, and obtaining an interpretation result; An abnormal point of the equipment is determined according to the interpretation result.

12. A computer-readable non-volatile storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the method according to any one of claims 1 to 9 is implemented.

Images