A computer room abnormality detection method, terminal equipment and storage medium

Abstract

The embodiment of the present application provides a method for detecting abnormalities in a computer room, a terminal device, and a storage medium. The method includes: converting a vibration signal into a spectrum signal and converting infrared image data into optical flow field image data; respectively calculating a first correlation coefficient between a business operation and a vibration signal, and a second correlation coefficient between a business operation and infrared image data; loading an abnormality detection network; inputting the vibration signal and the spectrum signal into a first backbone structure to extract target vibration features; inputting infrared image data and optical flow field image data into a second backbone structure to extract target image features; splicing the first correlation coefficient, the second correlation coefficient, the target vibration feature, and the target image feature into a target operation feature; inputting the target operation feature into a detection head structure to detect a target detection frame that characterizes the abnormal operation of a device group; and performing an alarm operation on the device group according to the target detection frame. Provide diversified monitoring for container clusters.

Description

Technical Field

[0001] The embodiments of the present application belong to the technical field of computer processing, and in particular, relate to a method for detecting anomalies in a computer room, a terminal device, and a storage medium. Background Art

[0002] With the rapid development of new-generation information technologies such as cloud computing and big data, the scale of data center computer rooms has expanded rapidly, the number of computing devices in the computer rooms has continued to increase, and the data processing capabilities of data centers have continued to improve.

[0003] A large number of computing devices may experience abnormalities during large-scale operation. In order to ensure the normal operation of the data center, a monitoring program is usually deployed in the data center to monitor the status of each computing device.

[0004] At present, the main monitoring methods are threshold method and deep learning. The threshold method refers to setting a threshold for indicators such as processor occupancy rate and memory occupancy rate in computing devices. When the indicator exceeds the threshold, it is considered that an abnormality has occurred. Deep learning refers to training a neural network and inputting indicators such as processor occupancy rate and memory occupancy rate in computing devices into the neural network to detect whether an abnormality has occurred.

[0005] Both the threshold method and deep learning are monitoring methods at the system level of computing devices. They rely heavily on data collection and data transmission at the system level and have limitations. In some cases, when an abnormality occurs at the driver layer or hardware layer at the system level, it will affect operations such as data collection and data transmission, causing these monitoring methods to fail. Summary of the invention

[0006] In view of this, an embodiment of the present application provides a computer room anomaly detection method, a terminal device and a storage medium to improve the comprehensiveness of abnormal monitoring of computing devices in a computer room.

[0007] A first aspect of an embodiment of the present application provides a method for detecting anomalies in a computer room, including:

[0008] The computing devices in the computer room are divided into multiple device groups based on cabinets; the computing devices in the same device group are deployed with the same group of containers in the container cluster Kubernetes, and the same group of containers has microservices with dependencies on each other deployed;

[0009] When the microservices in the same device group perform business operations, vibration signals and infrared image data are collected from the device group, and the vibration signals are converted into frequency spectrum signals and the infrared image data are converted into optical flow field image data;

[0010] respectively calculating a first correlation coefficient between the business operation and the vibration signal, and a second correlation coefficient between the business operation and the infrared image data;

[0011] Loading an anomaly detection network; the anomaly detection network has a first backbone structure, a second backbone structure and a detection head structure;

[0012] Inputting the vibration signal and the spectrum signal into the first backbone structure to extract target vibration features;

[0013] Inputting the infrared image data and the optical flow field image data into the second backbone structure to extract target image features;

[0014] splicing the first correlation coefficient, the second correlation coefficient, the target vibration feature and the target image feature into a target operation feature;

[0015] Inputting the target operation feature into the detection head structure to detect a target detection frame representing abnormal operation of the equipment group;

[0016] An alarm operation is performed on the device group according to the target detection frame.

[0017] A second aspect of an embodiment of the present application provides a device for detecting abnormalities in a computer room, including:

[0018] The device grouping module is used to divide the computing devices in the computer room into multiple device groups based on cabinets; the computing devices in the same device group deploy the same group of containers in the container cluster Kubernetes, and the same group of containers deploys microservices with mutual dependencies;

[0019] A data acquisition module, used for collecting vibration signals and infrared image data from the device group when the microservices in the same device group perform business operations, converting the vibration signals into frequency spectrum signals and converting the infrared image data into optical flow field image data;

[0020] A correlation coefficient calculation module, used to calculate a first correlation coefficient between the business operation and the vibration signal, and a second correlation coefficient between the business operation and the infrared image data;

[0021] An anomaly detection network loading module, used to load an anomaly detection network; the anomaly detection network has a first backbone structure, a second backbone structure and a detection head structure;

[0022] A target vibration feature extraction module, used for inputting the vibration signal and the spectrum signal into the first backbone structure to extract the target vibration feature;

[0023] A target image feature extraction module, used for inputting the infrared image data and the optical flow field image data into the second backbone structure to extract target image features;

[0024] A target operation feature splicing module, used for splicing the first correlation coefficient, the second correlation coefficient, the target vibration feature and the target image feature into a target operation feature;

[0025] A target detection frame generation module, used for inputting the target operation feature into the detection head structure to detect a target detection frame representing abnormal operation of the equipment group;

[0026] An alarm operation execution module is used to execute an alarm operation on the device group according to the target detection frame.

[0027] A third aspect of an embodiment of the present application provides a terminal device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein when the processor executes the computer program, the computer room anomaly detection method as described in the first aspect above is implemented.

[0028] A fourth aspect of an embodiment of the present application provides a computer-readable storage medium, wherein the computer-readable storage medium stores a computer program, and when the computer program is executed by a processor, the computer room anomaly detection method as described in the first aspect above is implemented.

[0029] A fifth aspect of the embodiments of the present application provides a computer program product. When the computer program product is run on a computer, the computer executes the computer room anomaly detection method described in the first aspect.

[0030] Compared with the background technology, the embodiments of the present application have the following beneficial effects:

[0031] In this embodiment, each computing device in the computer room is divided into multiple device groups based on cabinets; the computing devices in the same device group deploy the same group of containers in the container cluster Kubernetes, and microservices with mutual dependencies are deployed in the same group of containers; when the microservices in the same device group perform business operations, vibration signals and infrared image data are collected for the device group, the vibration signals are converted into spectrum signals, and the infrared image data are converted into optical flow field image data; the first correlation coefficient between the business operation and the vibration signal, and the second correlation coefficient between the business operation and the infrared image data are calculated respectively; the anomaly detection network is loaded; the anomaly detection network has a first backbone structure, a second backbone structure, and a detection head structure; the vibration signal and the spectrum signal are input into the first backbone structure together to extract the target vibration feature; the infrared image data and the optical flow field image data are input into the second backbone structure together to extract the target image feature; the first correlation coefficient, the second correlation coefficient, the target vibration feature, and the target image feature are spliced ​​into the target operation feature; the target operation feature is input into the detection head structure to detect the target detection frame that characterizes the abnormal operation of the device group; and the alarm operation is performed on the device group according to the target detection frame. This embodiment uses the execution of business operations by microservices in the same device group as a reference, and uses information on vibration and heat to jointly evaluate whether computing devices in the same device group have any abnormalities. It mainly monitors the hardware level, and incorporates the multimodal correlation between the system level and the hardware level of the container cluster Kubernetes. With the correlation as prior knowledge, the accuracy of anomaly detection can be guaranteed, and the dependence on data collection and data transmission at the system level can be reduced. It provides diversified monitoring for the container cluster Kubernetes, and when anomalies occur at the driver layer or hardware layer at the system level, the monitoring is still guaranteed to be effective. BRIEF DESCRIPTION OF THE DRAWINGS

[0032] In order to more clearly illustrate the technical solutions in the embodiments of the present application, the following briefly introduces the drawings required for use in the embodiments or prior art descriptions. Obviously, the drawings described below are only some embodiments of the present application, and for ordinary technicians in this field, other drawings can be obtained based on these drawings without creative work.

[0033] Figure 1 It is a schematic diagram of a method for detecting anomalies in a computer room provided in an embodiment of the present application;

[0034] Figure 2 It is a structural diagram of an anomaly detection network provided in an embodiment of the present application;

[0035] Figure 3 It is a structural diagram of a timing block structure provided in an embodiment of the present application;

[0036] Figure 4is a structural schematic diagram of a first convolution block structure provided in an embodiment of the present application;

[0037] Figure 5 is a structural schematic diagram of a second convolution block structure provided in an embodiment of the present application;

[0038] Figure 6 is a structural schematic diagram of a third convolution block structure provided in an embodiment of the present application;

[0039] Figure 7 is a schematic diagram of a device for detecting abnormalities in a computer room provided in an embodiment of the present application;

[0040] Figure 8 It is a schematic diagram of a terminal device provided in an embodiment of the present application. DETAILED DESCRIPTION

[0041] In the following description, specific details such as specific system structures, technologies, etc. are proposed for the purpose of illustration rather than limitation, so as to provide a thorough understanding of the embodiments of the present application. However, it should be clear to those skilled in the art that the present application can also be implemented in other embodiments without these specific details. In other cases, detailed descriptions of well-known systems, devices, circuits, and methods are omitted to prevent unnecessary details from hindering the description of the present application.

[0042] The technical solution of the present application is described below through specific embodiments.

[0043] Reference Figure 1 , shows a schematic diagram of a computer room anomaly detection method provided by an embodiment of the present application, which may specifically include the following steps:

[0044] Step 101: Divide each computing device in the computer room into multiple device groups based on cabinets.

[0045] In the computer room of the data center, multiple independent container clusters Kubernetes can be combined on each computing device according to business needs.

[0046] Among them, the container cluster Kubernetes is a platform for automated container operations, providing functions such as application deployment, maintenance, and expansion mechanisms, and can be used to manage containerized applications running across clusters. The container cluster Kubernetes contains a group of nodes, which can be computing devices in the computer room.

[0047] Generally speaking, the container cluster Kubernetes has components such as Master, Node, and Pod. Among them, the instance pod is the smallest basic unit created or deployed by the container cluster Kubernetes. An instance Pod represents a process running on the container cluster Kubernetes. An instance Pod can encapsulate one or more containers, storage resources, an independent IP (Internet Protocol) address, and policy options for managing and controlling the operation mode of the container.

[0048] A container is an abstraction of program code and its dependencies at the application layer. Multiple containers run as independent processes in the user space on the same computing device and share the operating system kernel.

[0049] The container is loaded with microservices with different business functions. Dynamic isolation management is implemented for the containers in the container cluster Kubernetes. At this time, all containers are divided into multiple groups, and some microservices with dependencies are created in the containers of the same group. The so-called dependency means that the data input to one of the microservices is the data output by another microservice. For example, in the login business function, microservice A responsible for maintaining user sessions depends on microservice B responsible for maintaining user devices. If microservice B outputs that the user device has been changed, microservice A will synchronize all the latest records of the session between the current user and all other users. If microservice B outputs that the user device has not been changed, microservice A will synchronize the latest incremental records of the session between the current user and all other users in the user device, and so on.

[0050] In this embodiment, when deploying the container cluster Kubernetes, each group of containers can be selectively deployed according to the distribution of computer devices, thereby dividing the various computing devices in the computer room into multiple device groups, so that the computing devices in the same device group deploy the same group of containers in the container cluster Kubernetes, and the same group of containers are deployed with microservices that have dependencies on each other.

[0051] Furthermore, computing devices are usually installed on cabinets. In this embodiment, the computing devices in the computer room can be divided into multiple device groups based on the cabinets.

[0052] Step 102: When the microservices in the same device group perform business operations, vibration signals and infrared image data are collected from the device group, and the vibration signals are converted into spectrum signals and the infrared image data are converted into optical flow field image data.

[0053] In actual applications, microservices with dependencies on the same set of containers are generally called in a mixed serial and parallel manner to perform large-scale business operations, such as training large language models (LLMs), using large language models for reasoning, and so on.

[0054] At this time, computing devices of the same device group are distributed on the same cabinet. The CPU (Central Processing Unit), GPU (Graphics Processing Unit), memory, fan and other hardware resources of the computing devices in the same device group follow the calls of microservices with dependencies on the same group of containers to support the microservices with dependencies on the same group of containers to perform large-scale business operations.

[0055] The hardware resources of the computing devices in the same device group on the same cabinet will dynamically generate vibrations and heat to the outside during operation. The computing devices have a self-regulating mechanism that can control the vibration and heat within a certain range. Therefore, vibration and heat can reflect the operating status of the computing devices to a certain extent. If vibration and heat are generated abnormally, the computing device may have faults at the system level and / or hardware level.

[0056] In this embodiment, if Figure 2 As shown, a vibration sensor can be set in the frame of the cabinet. The vibration sensor can continuously collect vibration signals generated by computing devices of the same device group on the same cabinet, and use the FFT (Fast Fourier Transformation) algorithm to convert the vibration signal from the time domain to the frequency domain, thereby converting the vibration signal into a spectrum signal.

[0057] In addition, an infrared camera can be set above the cabinet. The infrared camera can continuously collect infrared image data of computing devices in the same device group on the same cabinet, and use optical flow to convert the infrared image data into optical flow field image data.

[0058] In the optical flow method, the difference in grayscale distribution of different infrared image data can be used to obtain the motion field description of the moving object (such as air flow fluctuations generated by heat), and the motion field in space can be transferred to the optical flow field image data.

[0059] Step 103: Calculate a first correlation coefficient between the business operation and the vibration signal, and a second correlation coefficient between the business operation and the infrared image data respectively.

[0060] Generally speaking, there is a positive correlation between the hardware resources of the computing devices of the same device group on the same cabinet and the business operations performed by the microservices with dependencies on the same group of containers. Therefore, there is a positive correlation between the vibration signals generated by the computing devices of the same device group on the same cabinet and the business operations performed by the microservices with dependencies on the same group of containers. There is a positive correlation between the infrared image data generated by the computing devices of the same device group on the same cabinet and the business operations performed by the microservices with dependencies on the same group of containers. At this time, the correlation function can be called to calculate the first correlation coefficient between the business operation and the vibration signal and the second correlation coefficient between the business operation and the infrared image data.

[0061] In one embodiment of the present application, step 103 may include the following steps:

[0062] Step 10301: Determine multiple time points for sampling.

[0063] In this embodiment, a plurality of time points for adopting business operations, vibration signals and infrared image data may be determined on the time axis, and the time intervals between two adjacent time points are the same.

[0064] Step 10302: at each time point, query the load of each microservice in the container cluster Kubernetes.

[0065] At each point in time, you can query various system-level indicators of various microservices with dependencies on the same group of containers on the control end such as the master node of the container cluster Kubernetes, such as CPU occupancy, GPU occupancy, memory occupancy, I / O (Input / Output), etc., and use linear methods (i.e. weighted summation) to merge various indicators into load.

[0066] The indicators of various microservices with dependencies on the same group of containers at the system level are diverse. When an exception occurs in the driver layer or hardware layer at the system level, the unaffected indicators can be adaptively integrated into the load of the microservice, which has a relatively slight impact on the monitoring method based mainly on the hardware level. Even if the indicator is empty, the monitoring method based mainly on the hardware level can still be executed normally, which can reduce the dependence on data collection and data transmission at the system level.

[0067] Step 10303: The load of all microservices at the same time point is integrated into the computing volume of the business operation.

[0068] In this embodiment, the load of all microservices at the same time point is integrated into the computing volume of the business operation in a linear or nonlinear manner.

[0069] Due to the mixed serial and parallel calls between microservices, the call links of the same business operation vary in length in different situations. Although linear fusion is simple to operate, it may amplify the load caused by the call, thereby reducing the accuracy of the calculation amount of the business operation. Therefore, in some methods, the load can be converted nonlinearly to increase the calculation amount of the business operation.

[0070] In this method, the load of all microservices at the same time point is added together to obtain the total load value.

[0071] For a given business operation, you can count the average number of microservice calls from historical log files, record it as the average number of calls, and filter out the maximum value from all the call times, record it as the longest number of calls. When calling one microservice from another, the number of calls can be accumulated by 1.

[0072] On the one hand, the inverse of the average number of calls of the microservice when performing business operations is taken to obtain the serial ratio. On the other hand, the ratio between the average number of calls of the microservice when performing business operations and the longest number of calls of the microservice is calculated to obtain the call ratio, and then the average value between the serial ratio and the call ratio is calculated as the call rate.

[0073] At this time, the product of the total load value and the call rate is calculated as the computational load of the business operation.

[0074] Step 10304: organize the computational complexity of the business operations at each time point into a first sequence.

[0075] The computational volume of the same business operation at each time point is arranged in sequence to form a first sequence.

[0076] Step 10305: organize the amplitudes of the vibration signal at each time point into a second sequence.

[0077] In the vibration signal, the amplitudes at various time points are sorted in chronological order to form a second sequence.

[0078] Step 10306: At each time point, locate multiple coordinate points pointing to the computing device in the infrared image data.

[0079] When the installation position of the infrared camera is fixed and parameters such as the focal length are fixed, the technician can pre-calibrate multiple coordinate points pointing to the computing device on the infrared image data used as a sample.

[0080] Then, for the infrared image data corresponding to each current time, the pre-calibrated coordinate points can be mapped to the infrared image data corresponding to each current time.

[0081] Step 10307: At the same time point, calculate the average value of the temperature values ​​at the coordinate points in the infrared image data.

[0082] At the same time point, the temperature value at each coordinate point can be queried in the infrared image data, and the average value of the temperature value at each coordinate point can be calculated to characterize the overall temperature of the computing devices of the same device group on the same cabinet.

[0083] Step 10308: The average values ​​at each time point are combined into a third sequence.

[0084] The average values ​​at each time point are sorted in chronological order to form the third sequence.

[0085] Step 10309: Calculate the Pearson correlation coefficient between the first sequence and the second sequence as the first correlation coefficient between the business operation and the vibration signal.

[0086] In this embodiment, the amount of business operation calculation and the vibration signal are used as paired continuous variables, and the Pearson correlation coefficient between the first sequence and the second sequence can be calculated as the first correlation coefficient between the business operation and the vibration signal, so as to measure the linear correlation degree of the two paired continuous variables.

[0087] Step 10310: Calculate the Pearson correlation coefficient between the first sequence and the third sequence as the second correlation coefficient between the business operation and the infrared image data.

[0088] In this embodiment, the amount of business operation calculation and the average temperature representing heat (i.e., the average value of temperature values) are used as paired continuous variables, and the Pearson correlation coefficient between the first sequence and the third sequence can be calculated as the first correlation coefficient between the business operation and the infrared image data, so as to measure the linear correlation degree between the two paired continuous variables.

[0089] Step 104: Load the anomaly detection network.

[0090] In this embodiment, an anomaly detection network can be pre-built and trained based on deep learning for the task of detecting whether computing devices in the same device group on the same cabinet are running abnormally.

[0091] Among them, Figure 2 As shown, the anomaly detection network has a first backbone structure Backbone_1, a second backbone structure Backbone_2 and a detection head structure Head.

[0092] The first backbone structure Backbone_1 and the second backbone structure Backbone_2 both belong to the backbone network in the anomaly detection network. The backbone network is usually composed of multiple convolutional layers or other feature extraction layers, which are used to gradually extract high-level features from the input data.

[0093] The first backbone structure Backbone_1 and the second backbone structure Backbone_2 are the main parts of the anomaly detection network. The first backbone structure Backbone_1 is responsible for jointly extracting features from vibration signals and spectrum signals, and the second backbone structure Backbone_2 is responsible for jointly extracting features from infrared image data and optical flow field image data.

[0094] The detection head structure Head is the top part of the anomaly detection network, responsible for performing specific tasks. The detection head structure Head is usually composed of fully connected layers or convolutional layers, which are used to map the features extracted by the first backbone structure Backbone_1 and the second backbone structure Backbone_2 to the final output space, generating the final prediction results of the anomaly detection network.

[0095] Step 105: Input the vibration signal and the spectrum signal into the first backbone structure to extract the target vibration feature.

[0096] In this embodiment, if Figure 2 As shown, the vibration signal and the spectrum signal are input into the first backbone structure Backbone_1. The first backbone structure Backbone_1 processes the vibration signal and the spectrum signal according to its structure to extract the target vibration feature. The vibration signal and the spectrum signal are two different dimensions of information. Combining them to extract the target vibration feature can improve the quality of the target vibration feature.

[0097] In the specific implementation, Figure 2 As shown, the first backbone structure Backbone_1 includes the first convolution layer Conv_1, the second convolution layer Conv_2, the bidirectional long short-term memory network BiLSTM, the self-attention mechanism layer Self-Attention, the first multi-layer perceptron MLP_1 and three timing block structures TimeBlock.

[0098] Among them, Figure 3 As shown, each time block structure TimeBlock is sequentially distributed with a long short-term memory network LSTM, a third convolutional layer Conv_3 and a maximum pooling layer MaxPooling.

[0099] In one of the branches of the first backbone structure Backbone_1:

[0100] The vibration signal is divided into a plurality of segment signals, wherein the length of the segment data is a preset value.

[0101] Each segment signal is input into the first convolutional layer Conv_1 to perform a convolution operation, and the first candidate vibration feature is extracted therefrom.

[0102] Each first candidate vibration feature is input into a bidirectional long short-term memory network BiLSTM for processing, so as to output a second candidate vibration feature in the forward long short-term memory network and output a third candidate vibration feature in the bidirectional long short-term memory network.

[0103] Among them, the bidirectional long short-term memory network BiLSTM includes two long short-term memory networks LSTM, one long short-term memory network LSTM processes the input sequence forward after multiple time steps (i.e., the first candidate vibration feature), and the other long short-term memory network LSTM processes the sequence backward after multiple time steps (i.e., the first candidate vibration feature). After the processing is completed, the outputs of the two long short-term memory network LSTMs are spliced ​​together to obtain the final BiLSTM output result of the bidirectional long short-term memory network (i.e., the third candidate vibration feature), which better captures the bidirectional semantic dependency. The second candidate vibration feature output by the forward long short-term memory network and the third candidate vibration feature output by the backward long short-term memory network can both be used to extract the target vibration feature, which not only utilizes the multi-level features, but also simplifies the structure of the first backbone structure Backbone_1.

[0104] In another branch of the first backbone structure Backbone_1:

[0105] The spectrum signal is sequentially input into three time sequence block structures TimeBlock for processing to obtain the fourth candidate vibration feature.

[0106] The so-called sequentially means that in the sorting order, the feature map output by the previous timing block structure TimeBlock is the feature map input by the next timing block structure TimeBlock, the feature map initially input to the first timing block structure TimeBlock is the spectrum signal, and the feature map output by the last timing block structure TimeBlock is the fourth candidate vibration feature.

[0107] Furthermore, in each time series block structure TimeBlock, the feature map is processed in the long short-term memory network LSTM, the third convolution layer Conv_3, and the maximum pooling layer MaxPooling in turn, wherein the long short-term memory network LSTM provides an operation of extracting time series features, the third convolution layer Conv_3 provides a convolution operation, and the maximum pooling layer MaxPooling provides a maximum pooling operation.

[0108] In this first backbone structure Backbone_1 converges the backbone:

[0109] The second candidate vibration feature is concatenated with the fourth candidate vibration feature to obtain a fifth candidate vibration feature.

[0110] The fifth candidate vibration feature is input into the self-attention mechanism layer to extract the sixth candidate vibration feature.

[0111] Among them, the self-attention mechanism layer, also known as the internal attention mechanism, is an attention mechanism that associates different positions of a single sequence to calculate the representation of the same sequence, thereby improving the intrinsic correlation between the vibration signal and the spectrum signal.

[0112] The sixth candidate vibration feature is concatenated with the third candidate vibration feature to obtain the seventh candidate vibration feature. The shallow features of the vibration signal and the high-level features fused between the vibration signal and the spectrum signal are connected. In addition to the main position of the original enhanced vibration signal, the residual can also be learned to improve the performance of the first backbone structure Backbone_1.

[0113] The seventh candidate vibration feature is input into the second convolutional layer Conv_2 to extract the eighth candidate vibration feature.

[0114] The eighth candidate vibration feature is input into the first multi-layer perceptron MLP_1 and mapped into a target vibration feature, so that the target vibration feature is aligned and fused with the first correlation coefficient, the second correlation coefficient, and the target image feature.

[0115] Step 106: Input the infrared image data and the optical flow field image data into the second backbone structure to extract target image features.

[0116] In this embodiment, if Figure 2 As shown, the infrared image data and the optical flow field image data are input into the second backbone structure Backbone_2. The second backbone structure Backbone_2 processes the infrared image data and the optical flow field image data according to its structure to extract the target image features. The infrared image data and the optical flow field image data are two different dimensions of information. Combining them to extract the target image features can improve the quality of the target image features.

[0117] In the specific implementation, Figure 2As shown, the second backbone structure Backbone_2 includes three first convolution block structures ConvBlcok_1, three second convolution block structures ConvBlcok_2, three third convolution block structures ConvBlcok_3, an atrous spatial pyramid pooling layer (Atrous Spatial Pyramid Pooling, ASPP) and a second multi-layer perceptron MLP_2.

[0118] Among them, Figure 4 As shown, each first convolution block structure ConvBlcok_1 is sequentially distributed with a plurality of fourth convolution layers Conv_4 and a first average pooling layer AvgPooling_1.

[0119] like Figure 5 As shown, each second convolution block structure ConvBlcok_2 is sequentially distributed with a plurality of fifth convolution layers Conv_5 and a second average pooling layer AvgPooling_2.

[0120] like Figure 6 As shown, each third convolution block structure ConvBlcok_3 is sequentially distributed with a bottleneck layer Bottleneck and multiple sixth convolution layers Conv_6.

[0121] In one of the branches of the second backbone structure Backbone_2:

[0122] The infrared image data is sequentially input into three first convolution block structures ConvBlcok_1 for processing, so that the first infrared image feature is output in the second first convolution block structure ConvBlcok_1 and the second infrared image feature is output in the third first convolution block structure ConvBlcok_1.

[0123] The so-called sequentially means that in the sorting order, the feature map output by the previous first convolution block structure ConvBlcok_1 is the feature map input by the next first convolution block structure ConvBlcok_1, the feature map initially input to the first first convolution block structure ConvBlcok_1 is the infrared image data, the feature map output by the second first convolution block structure ConvBlcok_1 is the first infrared image feature, and the feature map output by the third first convolution block structure ConvBlcok_1 is the second infrared image feature.

[0124] Furthermore, in each first convolution block structure ConvBlcok_1, the feature map is processed in turn in multiple fourth convolution layers Conv_4 and the first average pooling layer AvgPooling_1, wherein the fourth convolution layer Conv_4 provides a convolution operation, and the first average pooling layer AvgPooling_1 provides an average pooling operation.

[0125] In another branch of the second backbone structure Backbone_2:

[0126] The optical flow field image data is sequentially input into the three second convolution block structures ConvBlcok_2 for processing to obtain the optical flow field image features.

[0127] The so-called sequentially means that in the sorting order, the feature map output by the previous second convolution block structure ConvBlcok_2 is the feature map input by the next second convolution block structure ConvBlcok_2, the feature map initially input to the first second convolution block structure ConvBlcok_2 is the optical flow field image data, and the feature map output by the third second convolution block structure ConvBlcok_2 is the optical flow field image feature.

[0128] Furthermore, in each second convolution block structure ConvBlcok_2, the feature map is processed in a plurality of fifth convolution layers Conv_5 and a second average pooling layer AvgPooling_2 in sequence, wherein the fifth convolution layer Conv_5 provides a convolution operation, and the second average pooling layer AvgPooling_2 provides an average pooling operation.

[0129] In this second backbone structure Backbone_2 converges with the backbone:

[0130] The second infrared image feature and the optical flow field image feature are concatenated to form the first candidate image feature.

[0131] The first candidate image feature is input into the first third convolution block structure ConvBlcok_3 for processing to obtain the second candidate image feature.

[0132] Among them, in the first third convolution block structure ConvBlcok_3, the first candidate image feature is processed in the distribution bottleneck layer Bottleneck and multiple sixth convolution layers Conv_6 in sequence, wherein the distribution bottleneck layer Bottleneck provides a dimensionality reduction operation, and the multiple sixth convolution layers Conv_6 provide a convolution operation.

[0133] The second candidate image feature is input into the second third convolution block structure ConvBlcok_3 for processing to obtain the third candidate image feature.

[0134] Among them, in the second third convolution block structure ConvBlcok_3, the second candidate image features are processed in the distribution bottleneck layer Bottleneck and multiple sixth convolution layers Conv_6 in sequence, wherein the distribution bottleneck layer Bottleneck provides a dimensionality reduction operation, and the multiple sixth convolution layers Conv_6 provide a convolution operation.

[0135] The first infrared image feature and the third candidate image feature are concatenated into the fourth candidate image feature, and the shallow features of the infrared image data and the high-level features fused between the infrared image data and the optical flow field image data are connected. In addition to the main position of the original infrared image data, the residual can also be learned to improve the performance of the second backbone structure Backbone_2.

[0136] The fourth candidate image feature is input into the third convolution block structure ConvBlcok_3 for processing to obtain the fifth candidate image feature.

[0137] Among them, in the third third convolution block structure ConvBlcok_3, the fourth candidate image feature is processed in the distribution bottleneck layer Bottleneck and multiple sixth convolution layers Conv_6 in sequence, wherein the distribution bottleneck layer Bottleneck provides a dimensionality reduction operation, and the multiple sixth convolution layers Conv_6 provide a convolution operation.

[0138] The fifth candidate image feature is input into the atrous spatial pyramid pooling layer ASPP for processing to obtain the sixth candidate image feature, wherein the atrous spatial pyramid pooling layer ASPP performs atrous convolution and parallel sampling on the fifth candidate image feature at different sampling rates, concatenates the obtained results together, expands the number of channels, and reduces the number of channels to an expected value through a 1×1 convolution, thereby capturing the context of the fifth candidate image feature at multiple scales.

[0139] The sixth candidate image feature is input into the second multi-layer perceptron MLP_2 and mapped into a target image feature, so that the target image feature is aligned and fused with the first correlation coefficient, the second correlation coefficient, and the target vibration feature.

[0140] Step 107: concatenate the first correlation coefficient, the second correlation coefficient, the target vibration feature and the target image feature into a target operation feature.

[0141] In this embodiment, if Figure 2 As shown, the first correlation coefficient between the business operation and the vibration signal, the second correlation coefficient between the business operation and the infrared image data, the target vibration feature extracted from the vibration signal and the spectrum signal, and the target image feature extracted from the infrared image data and the optical flow field image data are concatenated into the target operation feature.

[0142] When deploying the container cluster Kubernetes, computing devices and microservices can be regarded as a whole combination of a system and hardware. The hardware provides operation support for the system. The operating status of the system has a certain impact on the operating status of the hardware. The operating status of the system can be regarded as the prior knowledge of the operating status of the hardware. The first correlation coefficient, the second correlation coefficient, the target vibration feature and the target image feature are spliced ​​into the target operation feature. The hardware features of the computing devices of the same device group on the same cabinet can be associated with the system features of the microservices with dependencies on the same group of containers, providing prior knowledge assistance for anomaly detection.

[0143] For example, when microservices with dependencies on the same group of containers execute business operations at a high concurrency, computing devices in the same device group on the same cabinet will generate excessive vibration and heat, which is within a controllable range and is not abnormal.

[0144] In the target operation characteristics, the amount of information of the target vibration characteristics and the target image characteristics is significantly higher than the amount of information of the first correlation coefficient and the second correlation coefficient, so that the target operation characteristics are mainly based on the hardware characteristics of the computing devices in the same device group on the same cabinet, supplemented by the system characteristics of the microservices with dependencies on the same group of containers. In this way, it can complement the monitoring method implemented at the system level in the container cluster Kubernetes, and provide more comprehensive monitoring for the container cluster Kubernetes.

[0145] Step 108: Input the target operation feature into the detection head structure to detect the target detection frame representing the abnormal operation of the equipment group.

[0146] In this embodiment, if Figure 2 As shown, the target operation features are input into the detection head structure Head, and the head structure Head processes the target operation features according to its structure, thereby detecting the target detection frame representing the abnormal operation of the equipment group.

[0147] In the specific implementation, Figure 2 As shown in the figure, the detection head structure Head includes multiple seventh convolutional layers Conv_7, candidate region networks (Region Proposal Networks, RPN) and multiple fully connected layers (Fully Connected, FC).

[0148] In this detection head structure Head, the target running features are sequentially input into three seventh convolutional layers Conv_7 for processing to obtain candidate detection features.

[0149] Among them, the three seventh convolutional layers Conv_7 can all provide convolution operations. The so-called in sequence means that in the sorting order, the feature map output by the previous seventh convolutional layer Conv_7 is the feature map input by the next seventh convolutional layer Conv_7, the feature map initially input to the first seventh convolutional layer Conv_7 is the target running feature, and the feature map output by the last seventh convolutional layer Conv_7 is the candidate detection feature.

[0150] The candidate detection features are input into the candidate region network RPN to generate candidate detection boxes, and the candidate detection boxes are sequentially input into multiple fully connected layers FC to be mapped into target detection boxes that represent abnormal operation of the equipment group.

[0151] Among them, multiple fully connected layers FC can provide mapping operations. The so-called sequentially means that in the sorting order, the feature map output by the previous fully connected layer FC is the feature map input by the next fully connected layer FC. The feature map initially input to the first fully connected layer FC is the candidate detection box, and the feature map output by the last fully connected layer FC is the target detection box.

[0152] Step 109: Perform an alarm operation on the device group according to the target detection frame.

[0153] If the target detection box output by the anomaly detection network is empty, it means that the operation of the computing devices in the same device group on the same cabinet is normal, with reference to the running status of the microservices with dependencies on the same group of containers.

[0154] If the target detection box output by the anomaly detection network is not empty, it means that the operation of the computing devices in the same device group on the same cabinet is abnormal with reference to the running status of the microservices with dependencies on the same group of containers. At this time, an alarm operation can be performed on the computing devices in the same device group on the same cabinet for the target detection box that may have an abnormality, so as to verify it in conjunction with other monitoring methods, and prompt the operation and maintenance personnel to conduct an inspection.

[0155] In one embodiment of the present invention, step 109 may include the following steps:

[0156] Step 1091: Load the three-dimensional virtual model created for the computer room.

[0157] In this embodiment, an engine such as Three.js can be used to construct a three-dimensional virtual model based on the distribution of each cabinet and its computing equipment in the computer room, wherein the three-dimensional virtual model has a first device model representing each computing device, and the computing devices of the same device group on the same cabinet correspond to the first device model to form a second device model representing the device group.

[0158] Step 1092: construct the contour area of ​​each computing device in the infrared image data.

[0159] When the installation position of the infrared camera is fixed and parameters such as the focal length are fixed, the technician can pre-calibrate the contour area of ​​each computing device on the infrared image data used as a sample.

[0160] Then, for the current infrared image data, the pre-calibrated contour area can be mapped onto the current infrared image data.

[0161] Step 1093: If the relationship between the contour area and the target detection frame is intersection or inclusion, the contour area is set as the target area.

[0162] The contour areas of each computing device are traversed, and the contour areas of the computing devices are compared with the target detection frame. If the relationship between the contour area and the target detection frame is an intersection, or the contour area contains the target detection frame, or the target detection frame contains the contour area, then the contour area can be set as the target area.

[0163] Step 1094: Calculate the intersection-over-union ratio between the target area and the target detection frame.

[0164] Traverse each target area, calculate the intersection over union (IOU) between the target area and the target detection box, and compare the IOU with the preset overlap threshold.

[0165] Step 1095: If the intersection-over-union ratio is greater than or equal to a preset overlap threshold, determine that the computing device indicated by the target area is the target device.

[0166] If the intersection-and-union ratio between a target area and the target detection frame is greater than or equal to the overlap threshold, it means that the probability of an abnormality occurring in the computing device corresponding to the target area is high, and the computing device indicated by the target area can be determined to be the target device.

[0167] Step 1096: Mark the first device model corresponding to the target device as the first target model, and mark the second device model where the target model is located as the second target model.

[0168] The three-dimensional virtual model of the computer room is traversed to find a first device model corresponding to the target device, and the first device model is marked as a first target model.

[0169] In addition, in addition to the target device, other computing devices adjacent to the target device also have a certain probability of failure, so the second device model where the first target model is located is marked as the second target model.

[0170] Step 1097: Perform rendering operations that comply with the alarm operation on the first target model and the second target model in the three-dimensional virtual model respectively.

[0171] In the three-dimensional virtual model, the rendering operation that complies with the alarm operation can be performed on the first target model and the second target model respectively by highlighting or the like, thereby realizing the alarm operation on the first target model and the second target model.

[0172] Furthermore, if the monitoring method implemented at the system level in the container cluster Kubernetes also triggers an alarm operation on the same target device, a fuse operation can be performed on the microservice in the target device to enable the target device to perform other computing operations in addition to business operations. The fuse operation can fuse the microservice to perform business operations. Based on the load balancing strategy of the container cluster Kubernetes, other computing operations will be dynamically scheduled for execution by the microservice after a period of time to improve resource utilization.

[0173] After the fuse operation is completed, the anomaly detection network can be continuously called to generate a target detection frame, calculate the position deviation between the target detection frame before the fuse operation and the target detection frame after the fuse operation, and compare the deviation with a preset distance threshold.

[0174] Among them, the deviation can be expressed in the form of the Euclidean distance between the midpoint of the target detection box before the fuse operation and the midpoint of the target detection box after the fuse operation. If the target detection box after the fuse operation is empty, the deviation can be set to a default value, which is much larger than the distance threshold.

[0175] If the deviation is greater than or equal to the preset distance threshold, it means that the exception has a significant offset with the fuse, and the microservice on the target device may have system-level exceptions such as avalanches. In this case, the first target model can be rendered to prompt the microservice of a chain exception by highlighting and prompting words.

[0176] If the deviation is less than the preset distance threshold, it means that the abnormality has not produced a significant offset due to the blowing of the fuse, and the device group where the target device is located may have hardware-level abnormalities such as fan motor failure. In this case, the second target model can be rendered in a manner consistent with the hardware abnormality prompt using highlighting and prompt words.

[0177] In this embodiment, each computing device in the computer room is divided into multiple device groups based on cabinets; the computing devices in the same device group deploy the same group of containers in the container cluster Kubernetes, and microservices with mutual dependencies are deployed in the same group of containers; when the microservices in the same device group perform business operations, vibration signals and infrared image data are collected for the device group, the vibration signals are converted into spectrum signals, and the infrared image data are converted into optical flow field image data; the first correlation coefficient between the business operation and the vibration signal, and the second correlation coefficient between the business operation and the infrared image data are calculated respectively; the anomaly detection network is loaded; the anomaly detection network has a first backbone structure, a second backbone structure, and a detection head structure; the vibration signal and the spectrum signal are input into the first backbone structure together to extract the target vibration feature; the infrared image data and the optical flow field image data are input into the second backbone structure together to extract the target image feature; the first correlation coefficient, the second correlation coefficient, the target vibration feature, and the target image feature are spliced ​​into the target operation feature; the target operation feature is input into the detection head structure to detect the target detection frame that characterizes the abnormal operation of the device group; and the alarm operation is performed on the device group according to the target detection frame. This embodiment uses the execution of business operations by microservices in the same device group as a reference, and uses information on vibration and heat to jointly evaluate whether computing devices in the same device group have any abnormalities. It mainly monitors the hardware level, and incorporates the multimodal correlation between the system level and the hardware level of the container cluster Kubernetes. With the correlation as prior knowledge, the accuracy of anomaly detection can be guaranteed, and the dependence on data collection and data transmission at the system level can be reduced. It provides diversified monitoring for the container cluster Kubernetes, and when anomalies occur at the driver layer or hardware layer at the system level, the monitoring is still guaranteed to be effective.

[0178] It should be noted that the size of the serial numbers of the steps in the above embodiments does not mean the order of execution. The execution order of each process should be determined by its function and internal logic, and should not constitute any limitation on the implementation process of the embodiments of the present application.

[0179] Reference Figure 7 , shows a schematic diagram of a computer room anomaly detection device provided in an embodiment of the present application, which may specifically include the following modules:

[0180] The device grouping module 701 is used to divide the computing devices in the computer room into multiple device groups based on cabinets; the computing devices in the same device group deploy the same group of containers in the container cluster Kubernetes, and the same group of containers deploys microservices with mutual dependencies;

[0181] The data acquisition module 702 is used to collect vibration signals and infrared image data from the device group when the microservices in the same device group perform business operations, convert the vibration signals into frequency spectrum signals and convert the infrared image data into optical flow field image data;

[0182] A correlation coefficient calculation module 703, used to calculate a first correlation coefficient between the business operation and the vibration signal, and a second correlation coefficient between the business operation and the infrared image data;

[0183] Anomaly detection network loading module 704, used to load anomaly detection network; the anomaly detection network has a first backbone structure, a second backbone structure and a detection head structure;

[0184] A target vibration feature extraction module 705, configured to input the vibration signal and the spectrum signal into the first backbone structure to extract the target vibration feature;

[0185] A target image feature extraction module 706, configured to input the infrared image data and the optical flow field image data into the second backbone structure to extract target image features;

[0186] A target operation feature splicing module 707, configured to splice the first correlation coefficient, the second correlation coefficient, the target vibration feature and the target image feature into a target operation feature;

[0187] A target detection frame generation module 708 is used to input the target operation feature into the detection head structure to detect a target detection frame representing abnormal operation of the device group;

[0188] The alarm operation execution module 709 is used to execute an alarm operation on the device group according to the target detection frame.

[0189] Optionally, the correlation coefficient calculation module 703 is further used for:

[0190] Determine multiple time points for sampling;

[0191] At each of the time points, query the load of each of the microservices in the container cluster Kubernetes;

[0192] Merging the load of all the microservices at the same time point into the computing volume of the business operation;

[0193] Grouping the computational amounts of the business operations at the various time points into a first sequence;

[0194] Combining the amplitudes of the vibration signal at each of the time points into a second sequence;

[0195] At each of the time points, locating a plurality of coordinate points pointing to the computing device in the infrared image data;

[0196] At the same time point, calculating an average value of the temperature values ​​at the coordinate points in the infrared image data;

[0197] Combining the average values ​​at each of the time points into a third sequence;

[0198] calculating a Pearson correlation coefficient between the first sequence and the second sequence as a first correlation coefficient between the business operation and the vibration signal;

[0199] A Pearson correlation coefficient between the first sequence and the third sequence is calculated as a second correlation coefficient between the business operation and the infrared image data.

[0200] Optionally, the correlation coefficient calculation module 703 is further used for:

[0201] Add the loads of all the microservices at the same time point to obtain a total load value;

[0202] The reciprocal of the average number of calls of the microservice when executing the business operation is obtained to obtain a serial ratio;

[0203] Calculate the ratio between the average number of calls of the microservice when executing the business operation and the longest number of calls of the microservice to obtain a call ratio;

[0204] Calculate the average value between the serial ratio and the call ratio as the call rate;

[0205] The product of the total load value and the call rate is calculated as the computation amount of the business operation.

[0206] Optionally, the first backbone structure includes a first convolutional layer, a second convolutional layer, a bidirectional long short-term memory network, a self-attention mechanism layer, a first multi-layer perceptron and three sequential block structures; each of the sequential block structures is sequentially provided with a long short-term memory network, a third convolutional layer and a maximum pooling layer;

[0207] The target vibration feature extraction module 705 is also used for:

[0208] Dividing the vibration signal into a plurality of segment signals;

[0209] Inputting each of the segment signals into the first convolutional layer to extract the first candidate vibration feature;

[0210] Input each of the first candidate vibration features into the bidirectional long short-term memory network for processing, so as to output a second candidate vibration feature in the forward long short-term memory network and a third candidate vibration feature in the bidirectional long short-term memory network;

[0211] Inputting the spectrum signal into the three time sequence block structures in sequence for processing to obtain a fourth candidate vibration feature;

[0212] concatenating the second candidate vibration feature with the fourth candidate vibration feature to obtain a fifth candidate vibration feature;

[0213] Inputting the fifth candidate vibration feature into the self-attention mechanism layer to extract a sixth candidate vibration feature;

[0214] concatenating the sixth candidate vibration feature with the third candidate vibration feature to obtain a seventh candidate vibration feature;

[0215] Inputting the seventh candidate vibration feature into the second convolutional layer to extract an eighth candidate vibration feature;

[0216] The eighth candidate vibration feature is input into the first multi-layer perceptron and mapped into a target vibration feature.

[0217] Optionally, the second backbone structure includes three first convolution block structures, three second convolution block structures, three third convolution block structures, a void space pyramid pooling layer and a second multi-layer perceptron; each of the first convolution block structures is sequentially distributed with a plurality of fourth convolution layers and a first average pooling layer; each of the second convolution block structures is sequentially distributed with a plurality of fifth convolution layers and a second average pooling layer; each of the third convolution block structures is sequentially distributed with a bottleneck layer and a plurality of sixth convolution layers;

[0218] The target image feature extraction module 706 is also used for:

[0219] Inputting the infrared image data into three first convolution block structures in sequence for processing, so as to output a first infrared image feature in the second first convolution block structure and output a second infrared image feature in the third first convolution block structure;

[0220] Inputting the optical flow field image data into three second convolution block structures in sequence for processing to obtain optical flow field image features;

[0221] splicing the second infrared image feature and the optical flow field image feature into a first candidate image feature;

[0222] Inputting the first candidate image feature into the first third convolutional block structure for processing to obtain a second candidate image feature;

[0223] Inputting the second candidate image feature into the second third convolution block structure for processing to obtain a third candidate image feature;

[0224] splicing the first infrared image feature and the third candidate image feature into a fourth candidate image feature;

[0225] Inputting the fourth candidate image feature into the third third convolution block structure for processing to obtain a fifth candidate image feature;

[0226] Inputting the fifth candidate image feature into the atrous spatial pyramid pooling layer for processing to obtain a sixth candidate image feature;

[0227] The sixth candidate image feature is input into the second multi-layer perceptron and mapped into a target image feature.

[0228] Optionally, the detection head structure includes a plurality of seventh convolutional layers, a candidate region network and a plurality of fully connected layers;

[0229] The target detection frame generation module 708 is also used for:

[0230] Inputting the target running features into the three seventh convolutional layers in sequence for processing to obtain candidate detection features;

[0231] Inputting the candidate detection features into the candidate region network to generate a candidate detection frame;

[0232] The candidate detection frames are sequentially input into a plurality of the fully connected layers and mapped into target detection frames representing abnormal operation of the device group.

[0233] Optionally, the alarm operation execution module 709 is further used to:

[0234] Loading a three-dimensional virtual model created for the computer room; the three-dimensional virtual model has a first device model representing each of the computing devices, and the first device models constitute a second device model representing the device group;

[0235] constructing a contour area of ​​each of the computing devices in the infrared image data;

[0236] If the relationship between the contour area and the target detection frame is intersection or inclusion, setting the contour area as the target area;

[0237] Calculating the intersection-over-union ratio between the target area and the target detection frame;

[0238] If the intersection-over-union ratio is greater than or equal to a preset overlap threshold, determining that the computing device indicated by the target area is a target device;

[0239] Marking the first device model corresponding to the target device as a first target model, and marking the second device model where the first target model is located as a second target model;

[0240] A rendering operation that complies with an alarm operation is performed on the first target model and the second target model in the three-dimensional virtual model respectively.

[0241] Optionally, the alarm operation execution module 709 is further used to:

[0242] Performing a fuse operation on the microservice in the target device so that the target device performs other computing operations except the business operation;

[0243] Calculating a positional deviation between the target detection frame before the fuse operation and the target detection frame after the fuse operation;

[0244] If the deviation is greater than or equal to a preset distance threshold, a rendering operation is performed on the first target model to prompt that a chain exception occurs in the microservice;

[0245] If the deviation is less than a preset distance threshold, a rendering operation that is consistent with prompting a hardware abnormality is performed on the second target model.

[0246] An embodiment of the present application provides a device for detecting abnormalities in a computer room. By using the device, each step in the aforementioned method embodiments can be implemented.

[0247] As for the device embodiment, since it is basically similar to the method embodiment, the description is relatively simple, and the relevant parts can be referred to the description of the method embodiment part.

[0248] Reference Figure 8 , shows a schematic diagram of a terminal device provided by an embodiment of the present application. Figure 8 As shown, the terminal device 800 in the embodiment of the present application includes: a processor 810, a memory 820, and a computer program 821 stored in the memory 820 and executable on the processor 810. When the processor 810 executes the computer program 821, the steps in each embodiment of the above-mentioned computer room abnormality detection method are implemented. Alternatively, when the processor 810 executes the computer program 821, the functions of each module / unit in the above-mentioned device embodiments are implemented.

[0249] Exemplarily, the computer program 821 may be divided into one or more modules / units, which are stored in the memory 820 and executed by the processor 810 to complete the present application. The one or more modules / units may be a series of computer program instruction segments capable of completing specific functions, which may be used to describe the execution process of the computer program 821 in the terminal device 800.

[0250] The terminal device 800 may include, but is not limited to, a processor 810 and a memory 820. Those skilled in the art will appreciate that Figure 8 It is only an example of the terminal device 800 and does not constitute a limitation of the terminal device 800. It may include more or fewer components than shown in the figure, or a combination of certain components, or different components. For example, the terminal device 800 may also include input and output devices, network access devices, buses, etc.

[0251] The processor 810 may be a central processing unit (CPU), or other general-purpose processors, digital signal processors (DSP), application-specific integrated circuits (ASIC), field-programmable gate arrays (FPGA) or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. A general-purpose processor may be a microprocessor or any conventional processor, etc.

[0252] The memory 820 may be an internal storage unit of the terminal device 800, such as a hard disk or memory of the terminal device 800. The memory 820 may also be an external storage device of the terminal device 800, such as a plug-in hard disk, a smart media card (SMC), a secure digital (SD) card, a flash card, etc. equipped on the terminal device 800. Further, the memory 820 may also include both an internal storage unit of the terminal device 800 and an external storage device. The memory 820 is used to store the computer program 821 and other programs and data required by the terminal device 800. The memory 820 may also be used to temporarily store data that has been output or is to be output.

[0253] An embodiment of the present application also discloses a terminal device, including a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, the computer room anomaly detection method as described in the aforementioned embodiments is implemented.

[0254] The embodiment of the present application further discloses a computer-readable storage medium, wherein the computer-readable storage medium stores a computer program, and when the computer program is executed by a processor, the computer room anomaly detection method as described in the above-mentioned embodiments is implemented.

[0255] The embodiment of the present application further discloses a computer program product. When the computer program product is run on a computer, the computer is enabled to execute the computer room anomaly detection method described in the above-mentioned embodiments.

[0256] The above-described embodiments are only used to illustrate the technical solutions of the present application, rather than to limit them. Although the present application is described in detail with reference to the above-mentioned embodiments, a person skilled in the art should understand that the technical solutions described in the above-mentioned embodiments can still be modified, or some of the technical features can be replaced by equivalents; and these modifications or replacements do not make the essence of the corresponding technical solutions deviate from the spirit and scope of the technical solutions of the embodiments of the present application, and should be included in the protection scope of the present application.

Claims

1. A method for detecting abnormality in a computer room, characterized in that: include: Divide the computing equipment in the computer room into multiple equipment groups based on cabinets; The computing devices in the same device group deploy the same group of containers in the container cluster Kubernetes, and the same group of containers has microservices with dependencies on each other deployed; When the microservices in the same device group perform business operations, vibration signals and infrared image data are collected from the device group, and the vibration signals are converted into frequency spectrum signals and the infrared image data are converted into optical flow field image data; respectively calculating a first correlation coefficient between the business operation and the vibration signal, and a second correlation coefficient between the business operation and the infrared image data; The respectively calculating a first correlation coefficient between the business operation and the vibration signal and a second correlation coefficient between the business operation and the infrared image data comprises: Determine multiple time points for sampling; At each of the time points, query the load of each of the microservices in the container cluster Kubernetes; Merging the load of all the microservices at the same time point into the computing volume of the business operation; Grouping the computational amounts of the business operations at the various time points into a first sequence; Combining the amplitudes of the vibration signal at each of the time points into a second sequence; At each of the time points, locating a plurality of coordinate points pointing to the computing device in the infrared image data; At the same time point, calculating an average value of the temperature values ​​at the coordinate points in the infrared image data; Combining the average values ​​at each of the time points into a third sequence; calculating a Pearson correlation coefficient between the first sequence and the second sequence as a first correlation coefficient between the business operation and the vibration signal; Calculating a Pearson correlation coefficient between the first sequence and the third sequence as a second correlation coefficient between the business operation and the infrared image data; Loading an anomaly detection network; the anomaly detection network has a first backbone structure, a second backbone structure and a detection head structure; Inputting the vibration signal and the spectrum signal into the first backbone structure to extract target vibration features; Inputting the infrared image data and the optical flow field image data into the second backbone structure to extract target image features; splicing the first correlation coefficient, the second correlation coefficient, the target vibration feature and the target image feature into a target operation feature; Inputting the target operation feature into the detection head structure to detect a target detection frame representing abnormal operation of the equipment group; An alarm operation is performed on the device group according to the target detection frame.

2. The method according to claim 1, characterized in that The merging of the loads of all the microservices at the same time point into the computing volume of the business operation includes: Add the loads of all the microservices at the same time point to obtain a total load value; The reciprocal of the average number of calls of the microservice when executing the business operation is obtained to obtain a serial ratio; Calculate the ratio between the average number of calls of the microservice when executing the business operation and the longest number of calls of the microservice to obtain a call ratio; Calculate the average value between the serial ratio and the call ratio as the call rate; The product of the total load value and the call rate is calculated as the computation amount of the business operation.

3. The method according to claim 1, characterized in that The first backbone structure includes a first convolutional layer, a second convolutional layer, a bidirectional long short-term memory network, a self-attention mechanism layer, a first multi-layer perceptron and three sequential block structures; each of the sequential block structures is sequentially distributed with a long short-term memory network, a third convolutional layer and a maximum pooling layer; The step of inputting the vibration signal and the spectrum signal together into the first backbone structure to extract the target vibration feature comprises: Dividing the vibration signal into a plurality of segment signals; Inputting each of the segment signals into the first convolutional layer to extract the first candidate vibration feature; Input each of the first candidate vibration features into the bidirectional long short-term memory network for processing, so as to output a second candidate vibration feature in the forward long short-term memory network and a third candidate vibration feature in the bidirectional long short-term memory network; Inputting the spectrum signal into the three time sequence block structures in sequence for processing to obtain a fourth candidate vibration feature; concatenating the second candidate vibration feature with the fourth candidate vibration feature to obtain a fifth candidate vibration feature; Inputting the fifth candidate vibration feature into the self-attention mechanism layer to extract a sixth candidate vibration feature; concatenating the sixth candidate vibration feature with the third candidate vibration feature to obtain a seventh candidate vibration feature; Inputting the seventh candidate vibration feature into the second convolutional layer to extract an eighth candidate vibration feature; The eighth candidate vibration feature is input into the first multi-layer perceptron and mapped into a target vibration feature.

4. The method according to claim 3, characterized in that: The second backbone structure includes three first convolutional block structures, three second convolutional block structures, three third convolutional block structures, a void space pyramid pooling layer and a second multi-layer perceptron; each of the first convolutional block structures is sequentially distributed with a plurality of fourth convolutional layers and a first average pooling layer; each of the second convolutional block structures is sequentially distributed with a plurality of fifth convolutional layers and a second average pooling layer; each of the third convolutional block structures is sequentially distributed with a bottleneck layer and a plurality of sixth convolutional layers; The step of inputting the infrared image data and the optical flow field image data into the second backbone structure to extract target image features comprises: Inputting the infrared image data into three first convolution block structures in sequence for processing, so as to output a first infrared image feature in the second first convolution block structure and output a second infrared image feature in the third first convolution block structure; Inputting the optical flow field image data into three second convolution block structures in sequence for processing to obtain optical flow field image features; splicing the second infrared image feature and the optical flow field image feature into a first candidate image feature; Inputting the first candidate image feature into the first third convolutional block structure for processing to obtain a second candidate image feature; Inputting the second candidate image feature into the second third convolution block structure for processing to obtain a third candidate image feature; splicing the first infrared image feature and the third candidate image feature into a fourth candidate image feature; Inputting the fourth candidate image feature into the third third convolution block structure for processing to obtain a fifth candidate image feature; Inputting the fifth candidate image feature into the atrous spatial pyramid pooling layer for processing to obtain a sixth candidate image feature; The sixth candidate image feature is input into the second multi-layer perceptron and mapped into a target image feature.

5. The method according to claim 4, characterized in that The detection head structure includes a plurality of seventh convolutional layers, a candidate region network and a plurality of fully connected layers; The step of inputting the target operation feature into the detection head structure to detect a target detection frame representing abnormal operation of the equipment group includes: Inputting the target running features into the three seventh convolutional layers in sequence for processing to obtain candidate detection features; Inputting the candidate detection features into the candidate region network to generate a candidate detection frame; The candidate detection frames are sequentially input into a plurality of the fully connected layers and mapped into target detection frames representing abnormal operation of the device group.

6. The method according to any one of claims 1 to 5, characterized in that The performing an alarm operation on the device group according to the target detection frame includes: Loading a three-dimensional virtual model created for the computer room; the three-dimensional virtual model has a first device model representing each of the computing devices, and the first device models constitute a second device model representing the device group; constructing a contour area of ​​each of the computing devices in the infrared image data; If the relationship between the contour area and the target detection frame is intersection or inclusion, setting the contour area as the target area; Calculating the intersection-over-union ratio between the target area and the target detection frame; If the intersection-over-union ratio is greater than or equal to a preset overlap threshold, determining that the computing device indicated by the target area is a target device; Marking the first device model corresponding to the target device as a first target model, and marking the second device model where the first target model is located as a second target model; A rendering operation that complies with an alarm operation is performed on the first target model and the second target model in the three-dimensional virtual model respectively.

7. The method according to claim 6, characterized in that The performing rendering operations that comply with the alarm operation on the first target model and the second target model in the three-dimensional virtual model respectively includes: Performing a fuse operation on the microservice in the target device so that the target device performs other computing operations except the business operation; Calculating a positional deviation between the target detection frame before the fuse operation and the target detection frame after the fuse operation; If the deviation is greater than or equal to a preset distance threshold, a rendering operation is performed on the first target model to prompt that a chain exception occurs in the microservice; If the deviation is less than a preset distance threshold, a rendering operation that is consistent with prompting a hardware abnormality is performed on the second target model.

8. A terminal device, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that: When the processor executes the computer program, the computer room anomaly detection method as described in any one of claims 1-7 is implemented.

9. A computer-readable storage medium storing a computer program, characterized in that: When the computer program is executed by a processor, the computer room anomaly detection method as described in any one of claims 1 to 7 is implemented.

Images