Time scale based alerting method and apparatus

Abstract

The application relates to a time scale-based alarm method and device, which comprises the following steps: obtaining vibration data of a target device; determining predicted sample data of the target device from the vibration data according to a time scale; inputting the predicted sample data into a support vector machine model to output distribution characteristics corresponding to the predicted sample data; performing abnormality detection on the distribution characteristics to determine an abnormality index of the target device; and executing a pre-warning strategy when the abnormality index exceeds an alarm threshold. The application can reduce the burden of data processing and improve the calculation efficiency by applying an incremental and decremental learning strategy to optimize a specific type of support vector machine, thereby supporting real-time online monitoring, accurately tracking the dynamic changes of the distribution region of the monitoring parameters by using the time scale, and not only realizing comprehensive data collection but also improving the accuracy of the alarm.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of safety detection, and in particular to an alarm method and device based on time scale. BACKGROUND

[0002] With the rapid development of industrial internet and intelligent manufacturing, the safety, stability and reliability of complex mechanical and electrical equipment are increasingly concerned. These devices usually work in complex environments and variable working conditions, and their monitoring parameters are affected by various factors such as device state, operation history and working conditions. Therefore, for the health monitoring of complex mechanical and electrical equipment, it is particularly important to automatically process data in real time, adapt to changes in working conditions, reduce false positives and false negatives, and so on.

[0003] However, in the case of stable device state, it is difficult to accurately study the small changes of monitoring parameters. Once the device state is abnormal, the probability distribution region of the monitoring parameters will change accordingly. In the face of such changes in the distribution region of monitoring parameters, it is often difficult to accurately predict and warn.

[0004] Therefore, there is an urgent need to develop an alarm method based on time scale to solve one or more of the above problems. SUMMARY

[0005] In view of this, in order to solve the above technical problems or part of the technical problems, the embodiments of the present application provide an alarm method and device based on time scale.

[0006] In a first aspect, the present application provides an alarm method based on time scale, which comprises:

[0007] obtaining vibration data of a target device;

[0008] determining predicted sample data of the target device from the vibration data according to time scale;

[0009] inputting the predicted sample data into a support vector machine model to output distribution features corresponding to the predicted sample data;

[0010] performing anomaly detection on the distribution features to determine an anomaly index of the target device;

[0011] if the anomaly index exceeds an alarm threshold, executing a warning strategy.

[0012] In one possible implementation, the determination of the predicted sample data of the target device from the vibration data according to time scale comprises:

[0013] determining a target period corresponding to a prediction time by using time scale;

[0014] determine a data sequence corresponding to the target period from the vibration data;

[0015] use the data sequence as the prediction sample data of the target device.

[0016] In one possible implementation, the vibration data includes vibration sub-data at different vibration time points;

[0017] The determining of the data sequence corresponding to the target period from the vibration data includes:

[0018] creating a corresponding time label according to the vibration time point and marking the time label to the corresponding vibration sub-data;

[0019] determining a start time point and an end time point corresponding to the target period;

[0020] determining a vibration sub-data sequence corresponding to the start time point to the end time point from the vibration data by using the time label, as the data sequence corresponding to the target period.

[0021] In one possible implementation, the method further includes:

[0022] obtaining historical vibration data of a target device;

[0023] performing data preprocessing on the historical vibration data to obtain sample support vectors;

[0024] establishing a vector machine model by using the sample support vectors as an initial vector machine model;

[0025] updating the initial vector machine model to obtain a support vector machine model.

[0026] In one possible implementation, the performing of the data preprocessing on the historical vibration data to obtain sample support vectors includes:

[0027] performing vector extraction on the historical vibration data to obtain an initial support vector set;

[0028] selecting sample support vectors from the initial support vector set according to features of initial support vectors in the initial support vector set.

[0029] In one possible implementation, the updating of the initial vector machine model to obtain a support vector machine model includes:

[0030] reselecting support vectors from the initial support vector set according to features of initial support vectors in the initial support vector set;

[0031] The reselected support vectors are taken as new sample support vectors to perform the step of establishing a vector machine model by using the sample support vectors, to obtain a new vector machine model, and the vector machine model is added to a support vector machine model set;

[0032] A support vector machine model is selected from the support vector machine model set.

[0033] In one possible implementation, the method further comprises:

[0034] It is determined whether there is an unselected support vector in the initial support vector set;

[0035] In the case where there is an unselected support vector in the initial support vector set, the step of reselecting a sample support vector is continuously performed.

[0036] In one possible implementation, before reselecting a support vector from the initial support vector set according to the characteristics of the initial support vectors in the initial support vector set, the method further comprises:

[0037] The historical vibration data is input into a vector machine model, and a distribution characteristic corresponding to the historical vibration data is output;

[0038] A state parameter of the target device is obtained according to the distribution characteristic;

[0039] A parameter difference between the state parameter and a historical state parameter recorded in a state label of the historical vibration data is obtained;

[0040] In the case where the parameter difference does not satisfy a preset difference rule, the sample support vector is deleted from the initial support vector set.

[0041] In a second aspect, the application provides an alarm device based on a time scale, the device comprising:

[0042] A first acquisition module is configured to acquire vibration data of a target device;

[0043] A sample determination module is configured to determine predicted sample data of the target device from the vibration data according to a time scale;

[0044] An output module is configured to input the predicted sample data into a support vector machine model, and output a distribution characteristic corresponding to the predicted sample data;

[0045] An index determination module is configured to perform anomaly detection on the distribution characteristic, and determine an anomaly index of the target device;

[0046] An execution module is configured to perform a pre-warning strategy in the case where the anomaly index exceeds an alarm threshold.

[0047] In one possible implementation, the apparatus further includes:

[0048] a second obtaining module, configured to obtain historical vibration data of the target device;

[0049] a preprocessing module, configured to perform data preprocessing on the historical vibration data to obtain sample support vectors;

[0050] a establishing module, configured to establish a vector machine model by using the sample support vectors as an initial vector machine model;

[0051] an updating module, configured to update the initial vector machine model to obtain a support vector machine model.

[0052] The above technical solution provided by the embodiments of the present application has the following advantages compared with the prior art: the method provided by the embodiments of the present application optimizes a specific type of support vector machine by applying an incremental and decremental learning strategy, which can reduce the burden of data processing and improve the calculation efficiency, thereby supporting real-time online monitoring; in addition, the dynamic changes of the distribution region of the monitoring parameters are accurately tracked by using the time scale; not only comprehensive data collection is realized, but also the accuracy of the alarm is improved, which ensures that the complex mechanical and electrical equipment can be monitored in real time and warned accurately under the condition of complex working environment and variable working conditions. BRIEF DESCRIPTION OF DRAWINGS

[0053] The accompanying drawings, which are incorporated in and constitute a part of this specification, illustrate embodiments consistent with the present application and serve to explain the principles of the present application together with the specification.

[0054] In order to more clearly illustrate the technical solutions in the embodiments of the present application or the prior art, the accompanying drawings needed to be used in the embodiments or prior art description will be briefly introduced here. Obviously, for those skilled in the art, other drawings can also be obtained based on these drawings without any creative effort.

[0055] One or more embodiments are exemplarily illustrated by the pictures in the drawings corresponding to the embodiments, and these exemplary illustrations do not constitute a limitation on the embodiments. Elements with the same reference numerals in the drawings represent similar elements, unless otherwise specified. The drawings in the drawings do not constitute a proportional limitation.

[0056] Figure 1 A flowchart of an alarm method based on a time scale provided by the embodiments of the present application;

[0057] Figure 2 A flowchart of a support vector machine model alarm provided by the embodiments of the present application;

[0058] Figure 3 A flow chart of a time scale based alarm method provided by the embodiment of the present application is shown in the figure;

[0059] Figure 4 A flow chart of another time scale based alarm method provided by the embodiment of the present application is shown in the figure;

[0060] Figure 5 A geometric description of a support vector machine provided by the embodiment of the present application is shown in the figure;

[0061] Figure 6 A step chart of establishing a support vector machine model provided by the embodiment of the present application is shown in the figure;

[0062] Figure 7a A support vector increment provided by the embodiment of the present application is shown in the figure;

[0063] Figure 7b A support vector decrement provided by the embodiment of the present application is shown in the figure;

[0064] Figure 8 A normal and stable running state of a device provided by the embodiment of the present application is shown in the figure;

[0065] Figure 9 A normal and stable running state of a device provided by the embodiment of the present application is shown in the figure;

[0066] Figure 10 A sudden change state of a device provided by the embodiment of the present application is shown in the figure;

[0067] Figure 11 A slow change state of a device provided by the embodiment of the present application is shown in the figure;

[0068] Figure 12 A structure chart of a time scale based alarm device provided by the embodiment of the present application is shown in the figure. DETAILED DESCRIPTION

[0069] In order to make the purpose, technical solutions and advantages of the embodiments of the present application clearer, the technical solutions of the embodiments of the present application will be described clearly and completely below with reference to the drawings in the embodiments of the present application. Obviously, the described embodiments are part of the embodiments of the present application, rather than all the embodiments of the present application. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative work fall within the scope of protection of the present application.

[0070] The following disclosure provides a number of different embodiments or examples for implementing different aspects of the present application. In order to simplify the disclosure of the present application, the components and arrangements of the specific examples below are described. Of course, they are merely examples and are not intended to limit the present application. Furthermore, the present application can repeat reference numerals and / or letters in different examples. Such repetition is for the purpose of simplicity and clarity and does not in itself dictate a relationship between the various embodiments and / or configurations discussed.

[0071] To solve the technical problem in the prior art that the device monitoring method often relies on periodic inspection and fixed threshold setting, which has the problems of fixed alarm threshold setting, difficulty in adapting to changes in the running state of different devices, inability to realize real-time monitoring and comprehensive data collection of the device state, resulting in insufficient timeliness and accuracy of fault early warning, and ultimately failing to meet the requirements of the complexity and variability of devices in industrial production, and although the details of the change of the monitoring parameter cannot be accurately studied, the distribution area of the monitoring parameter must be stable in the sense of probability, and once the state quantity changes abnormally, the distribution area of the monitoring parameter will change. In the face of the change of the distribution area of the monitoring parameter, the current dynamic monitoring method cannot accurately predict and warn, the present application selects prediction sample data after obtaining vibration data, and then inputs the sample data into a vector machine model to output the distribution characteristics of the sample data, so as to realize abnormal detection of the device, and in the case that the abnormal index exceeds the alarm threshold, a warning strategy is executed; by applying the incremental and decremental learning strategy to optimize the specific type of support vector machine, the burden of data processing can be reduced and the calculation efficiency can be improved, thereby supporting real-time online monitoring. In addition, the time scale is used to accurately track the dynamic change of the distribution area of the monitoring parameter; not only comprehensive data collection is realized, but also the accuracy of the alarm is improved, and real-time monitoring and accurate early warning of the state of the complex mechanical and electrical equipment under the condition of complex working environment and variable working conditions are ensured.

[0072] Figure 1 A flowchart of an alarm method based on a time scale provided for an embodiment of the present application is shown in FIG. Figure 1 The method specifically comprises the following steps.

[0073] S101, obtaining vibration data of a target device;

[0074] In this embodiment, the vibration of the device is monitored in real time by the vibration sensor pre-installed on the target device; then, the vibration signal collected by the sensor is transmitted to the central processing unit by the data acquisition system, and various signal processing techniques such as Fourier transform and wavelet transform are used to convert the time domain signal into a frequency domain signal, so as to better analyze the vibration characteristics of the device. Through the analysis of the vibration signal, the abnormal vibration mode in the operation of the device is identified, and potential faults and problems are discovered in time, so that corresponding maintenance measures can be taken to ensure the normal operation of the device and prolong its service life.

[0075] S102、According to the time scale, determine the prediction sample data of the target device from the vibration data;

[0076] Time scale is a standard or unit for dividing and defining time, used to measure and record the time range or interval of an event or phenomenon. In the field of data processing and analysis, it plays a crucial role. Through accurate time scale, the running state of mechanical and electrical equipment can be tracked and analyzed in detail, so as to realize the rapid response and accurate prediction of the change of equipment state. In this device, the use of time scale is not limited to the data collection stage, but also throughout the whole monitoring and alarm process, ensuring the timeliness and accuracy of data processing. For example, when analyzing vibration data, time scale can help determine the sampling frequency of data, and then affect the efficiency and effect of signal processing and feature extraction. Time scale is conducive to understanding and organizing data, making it more meaningful and interpretable.

[0077] In this embodiment, when extracting the prediction sample data of the target device, first, the vibration data needs to be divided according to the predetermined time scale, including determining the sampling frequency of data, selecting appropriate time window and defining the time sequence of data, etc. For example, the vibration data can be divided according to the time scale of hour, day, week or month, so as to better observe and analyze the running state of the equipment in different time periods.

[0078] The process of extracting data by using time scale is also the process of screening and sorting the original data. Specifically, from a large amount of vibration data, the key information related to the target device is extracted, such as the vibration frequency, amplitude and vibration mode of the equipment. Using time scale to determine sample data not only improves the efficiency and accuracy of data processing, but also provides a solid foundation for subsequent prediction analysis. By comparing and analyzing the data under different time scales, we can better understand the running rules and trends of the equipment, so as to discover potential problems in time and take corresponding maintenance measures.

[0079] S103、Input the prediction sample data into the support vector machine model, and output the distribution characteristics corresponding to the prediction sample data;

[0080] Support vector machine model has strong classification and regression function, can process high-dimensional data and effectively extract features; It is widely used in the field of machine learning, especially in text classification, image recognition and bioinformatics; Support vector machine model finds the best classification hyperplane by maximizing the boundary between different classes, so as to realize the effective segmentation of data, which not only improves the accuracy of classification, but also avoids overfitting to a certain extent; It has good generalization ability and can perform well on unknown data.

[0081] In this embodiment, the prediction sample data obtained according to the time scale is input into the support vector machine model, and the distribution characteristics corresponding to the prediction sample data are obtained by indicating the processing and analysis of the vector machine model. Specifically, the support vector machine model will classify and feature extract the input prediction sample data according to the classification boundary and decision function learned in the training process, thereby revealing the distribution rule and characteristics of the data in high-dimensional space, and revealing the internal structure of the prediction sample data, providing an important reference for subsequent data analysis and decision-making.

[0082] S104, performing anomaly detection on the distribution characteristics to determine the anomaly index of the target device;

[0083] In this embodiment, the anomaly detection is performed by detailed analysis and evaluation of the distribution characteristics, so as to determine whether the target device has an anomaly index. Specifically, first, the relevant data of the target device is collected, and the distribution characteristics of the prediction sample data are deeply analyzed and evaluated, and then advanced algorithms and models are used to systematically detect the distribution characteristics to identify any anomaly index deviating from the normal range, which can effectively identify the potential problems of the target device, so that corresponding measures are taken for repair and optimization.

[0084] S105, executing a warning strategy when the anomaly index exceeds the alarm threshold.

[0085] In this embodiment, when the anomaly index exceeds the alarm threshold, a warning strategy is executed, for example, when the vibration data generated during the operation of the device exceeds the pre-set alarm threshold, the system will automatically trigger and execute a series of warning strategies.

[0086] Specifically, an alarm threshold of vibration intensity is set, for example, when the vibration intensity exceeds a certain specific value (such as 5g), it is considered abnormal, and once it is determined through analysis that the device vibration data reaches or exceeds this alarm threshold, the system will immediately start the warning strategy; the specific warning strategy can be sending an alarm message to the relevant personnel, recording the abnormal data for subsequent analysis, or starting the self-protection mechanism of the device, such as automatic shutdown to avoid possible further damage, by setting the alarm threshold, the possible problems in the operation of the device can be found and handled in time, and the stable operation of the device is ensured.

[0087] The time scale based alarm method provided by the embodiments of the present application first acquires vibration data of a target device, analyzes the data according to time series to identify a sample data set for prediction, then inputs the prediction sample data into a support vector machine model to extract corresponding distribution features, further performs an anomaly detection algorithm on the distribution features to determine an anomaly index of the target device, and once the anomaly index exceeds a preset alarm threshold, a warning mechanism is immediately started, and a specific type of support vector machine is optimized by applying an incremental and decremental learning strategy, which can reduce the burden of data processing and improve the calculation efficiency, thereby supporting real-time online monitoring. In addition, the time scale is also used to accurately track the dynamic changes of the monitoring parameter distribution region. Not only comprehensive data acquisition is realized, but also the accuracy of the alarm is improved, and real-time monitoring and accurate warning of the state of complex mechanical and electrical equipment under the condition of complex working environment and variable working conditions are ensured.

[0088] As shown in Figures 2-3 , the process of the device state dynamic self-adaptive warning method based on the analysis time scale obtained by combining the support vector machine with the time scale analysis of the sliding window is as follows:

[0089] Step one: through the online algorithm of the support vector machine, the distribution D k of the target sample set {x k-m+1 ,…,x k-1 ,x k} in the feature space F is obtained at time k.

[0090] Step two: when k+1, new monitoring data x k+1 is acquired, I k+1 is calculated, and compared with the alarm threshold η to determine whether to alarm.

[0091] Step three: x k+1 is added to the target sample set, and x k-m+1 is deleted from the target sample set, and the new distribution D k+1 is calculated online.

[0092] Step four: k←k+1, repeat steps two to three.

[0093] The anomaly index is a relative index of a spatial distance, and the threshold η is set to determine when the new data distance the distribution region D k needs to be alarmed.

[0094] Suppose that the target sample set X k at time k satisfies the normal distribution N(μ k ,υ k ), then I k+1 ≥η is equivalent to:

[0095]

[0096] wherein C(σy k ,ν) is a certain boundary threshold of normal distribution.

[0097] The device state dynamic self-adaptive early warning method based on the analysis time scale can better meet the dynamic condition. Specifically, with the continuous update of the monitoring data, the distribution region is continuously estimated under the analysis scale, so that the change of the device state can be adapted. By monitoring the operation data of the device in real time, the early warning parameters are dynamically adjusted to ensure the sensitivity and accuracy of the early warning system. In addition, the personalized device monitoring parameters are used as target samples to solve the personalized alarm problem, so that the specific operation characteristics and historical data of each device are fully considered, thereby realizing more accurate and targeted early warning. Through the above alarm method, the early warning system can discover potential faults and abnormalities in time according to the actual operation state of the device, so as to take measures in advance to avoid the occurrence of device faults and improve the reliability and safety of the device.

[0098] When the state of the device remains stable, although the details of the change of the monitoring parameters cannot be accurately studied, the distribution region of the monitoring parameters must be stable in the sense of probability. Once the state quantity changes abnormally, the distribution region of the monitoring parameters will change accordingly. The application establishes an analysis time scale on the monitoring parameters through a sliding window, performs kernel-based statistical analysis on the data segment in a time scale, estimates the distribution region of the data segment in the feature space, and continuously performs the process to obtain the distribution region information dynamically updated in time sequence.

[0099] In an optional scheme of the embodiment of the application, the determination of the prediction sample data of the target device from the vibration data according to the time scale comprises:

[0100] The target time period corresponding to the prediction moment is determined by using the time scale;

[0101] In the embodiment, the specific target time period corresponding to the prediction moment is determined from the time axis by using the time scale. Specifically, the time scale determines the time period corresponding to the prediction moment on the time axis;

[0102] For example, in order to ensure that the state prediction of the target device can accurately reflect its actual running condition, it is necessary to evaluate the performance of the device in the past week; at this time, the time scale is set to a time span of one week, and by analyzing the data in this week, a more comprehensive device state evaluation can be obtained, and the target period corresponding to the prediction moment is defined as the time range of one week before the prediction moment, to ensure the accuracy and reliability of the prediction result. By analyzing the state of the device in the previous week, the state of the device at the prediction moment is predicted, thereby providing strong data support for the maintenance and operation of the device.

[0103] determining a data sequence corresponding to the target period from the vibration data;

[0104] In this embodiment, after determining the target period corresponding to the prediction moment, the vibration data is first preprocessed, including filtering, denoising and other steps, to ensure the accuracy and reliability of the data; then, the data is matched with the target period by using the time stamp or other time markers, so as to extract the vibration data sequence in the period; thereby obtaining the vibration data sequence corresponding to the target period and having representativeness, providing a solid foundation for further data analysis and processing.

[0105] using the data sequence as the prediction sample data of the target device.

[0106] In this embodiment, the data sequence is used as the prediction sample data of the target device, so that more accurate prediction and decision can be made in subsequent analysis and processing; then the data sequence is used as input to train and verify the support vector machine model, thereby generating reliable prediction results, better understanding the behavior pattern and running state of the target device, and further optimizing the performance and improving the reliability of the device.

[0107] In addition, since the support vector machine is established on the basis of the structural risk minimization theory, it can effectively coordinate the contradiction between the complexity and generalization of the learning machine, and the introduction of the parameter v makes a compromise between the volume of the distribution region and the number of sample points contained, so as to avoid the influence of a small number of abnormal points on the estimation result, and does not need special processing of abnormal data, thereby ensuring the continuity of dynamic estimation.

[0108] For example, sample points x1, x2,…, x l ∈R n , l is the number of samples, α i ≥ 0 is the Lagrange coefficient, K(x i ,x j ) = <φ(x i ), φ(x j )> = K i,j is a kernel function.

[0109] When a new monitoring parameter x k+1 arrives, it is determined whether it belongs to the distribution region determined by the target samples in the sliding window at time k, and the distance between x k+1 and the distribution is quantitatively expressed, and the decision equation is as follows:

[0110]

[0111] According to the above decision equation, if x k+1 belongs to the distribution at time k; x k+1 is located on the boundary of the distribution; x k+1 is located outside the distribution, the smaller I is, the farther x is from the distribution at time k.

[0112] In order to conform to the conventional habit, the negative logarithmic form of I is taken as an anomaly index:

[0113]

[0114] When I k+1 <0, x k+1 belongs to the distribution at time k; I k+1 =0 is located on the boundary of the distribution; I k+1 >0 is located outside the distribution.

[0115] In an optional solution of the embodiment of the application, the vibration data includes vibration sub-data at different vibration times;

[0116] The method for determining the data sequence corresponding to the target period from the vibration data comprises:

[0117] According to the vibration time, a corresponding time label is created, and the time label is marked in the corresponding vibration sub-data;

[0118] The start time and the end time corresponding to the target period are determined;

[0119] Using the time label, the vibration sub-data sequence corresponding to the start time to the end time is determined from the vibration data as the data sequence corresponding to the target period.

[0120] ​In the embodiment, before obtaining the data sequence corresponding to the target period, first, according to the time of occurrence of vibration, the corresponding time label is created, and the time label is accurately labeled to the corresponding vibration sub-data; each vibration event can be more conveniently identified and tracked. Then, the starting time and the ending time of the target period are determined, and the vibration sub-data sequence occurring between the starting time and the ending time of the target period is filtered out from the entire vibration data set as the input data for studying and analyzing the vibration in the target period through the time label created.

[0121] Figure 4 Another flowchart of the alarm method based on time scale provided in the embodiment is shown in FIG. 4, and the method specifically includes the following steps. Figure 4

[0122] S401, obtaining historical vibration data of a target device;

[0123] In the embodiment, in order to accurately predict the state of the target device, first, the historical vibration data of the device in the past period of time is obtained, and the selected historical vibration data includes the vibration characteristics of the device in different running states, thereby helping to obtain the running law of the device, and the vibration of the device in the normal operation, abnormal situation and fault occurrence is recorded by collecting these historical vibration data. Subsequently, a support vector model (SVM) is established by using these historical vibration data, and the model will be used for state prediction of the target device.

[0124] S402, data preprocessing is performed on the historical vibration data to obtain sample support vectors;

[0125] In the embodiment, when processing the historical vibration data, a series of data preprocessing steps, such as data cleaning, feature selection and the like, are required to ensure the quality and availability of the data, and then the preprocessed data is used to find the corresponding sample support vectors through a training algorithm.

[0126] S403, a vector machine model is established by using the sample support vectors as an initial vector machine model;

[0127] The specific process of establishing a vector machine model by using the sample support vectors is used as an initial vector machine model

[0128] In the embodiment, the sample support vectors are data points located on the boundary; when constructing the model, the support vectors will be used as the input of the model, and the vector machine algorithm is trained to obtain a model capable of accurately classifying data as an initial vector machine model.

[0129] S404, updating the initial vector machine model to obtain a support vector machine model.​

[0130] In the embodiment, after the initial vector machine model is established, the initial vector machine model is verified according to historical vibration data. Specifically, whether the support vector of the current sample needs to be updated is determined according to the difference between the model output state and the real state, so as to identify the sample points, i.e., the support vectors, which have greater influence on the model performance by comparing the difference between the model prediction result and the actual result. The sample support vectors are adjusted and optimized according to the difference between the output result and the real value, so as to improve the initial vector machine model, improve the accuracy of the model, and obtain a more efficient and reliable support vector machine model.

[0131] The alarm method based on time scale provided in the embodiment of the application uses historical vibration data of a target device to establish a support vector machine model suitable for the target device, and in the process of creating the support vector machine model, the difference between the output state and the real state recorded in the historical vibration data of the target device is used to update the support vector model, so that an accurate and reliable support vector machine model is finally obtained, and the model can predict the future state of the device according to the current vibration characteristics, so as to discover potential faults or abnormal conditions in advance. The device fault prevention is realized, and the reliability of the device is significantly improved.

[0132] As shown in Figure 5 , the support vector machine is used to find the optimal distribution region D in the feature space F, which can be geometrically described by the following figure. In order to simplify the problem, it is assumed that a normalized kernel function is selected, which satisfies the condition K(x, x) = 1, such as a radial basis kernel function. Satisfying this condition means that the norm of all kernel functions is equal to 1, and the sample is projected into a possibly infinite-dimensional unit hypersphere F. ω and ρ define a hyperplane W, which is orthogonal to ω and has a distance from the center of the sphere equal to ρ / ||ω|. D is the part of the hypersphere away from the center divided by W. The purpose of optimization is to find the maximum value of ρ / ||ω| under the given conditions, at which time the volume of D is the smallest.

[0133] In Figure 5 , the decision function f(x i ) of the non-support vector set (denoted as NSV) is greater than 0, which is located in the distribution region D, while the decision function f(x i ) of the support vector is less than or equal to 0, which is located on the boundary of D or outside the boundary; the support vector set f(x i ) = 0 is called the boundary support vector set, denoted as MSV, and the subscript set is S = {s1, s2, …, s ls}, ls is the number of support vectors; the support vector set f(x i ) < 0 is called the non-boundary support vector set, denoted as NMSV, and the corresponding ξ i> 0, in order to ensure that the target class distribution area is compact, the samples in the set are considered as outliers outside the distribution area. Through such division, the target sample set is composed of three subsets, X = MSV U NMSV U NSV. The parameter v is both the upper bound of the non-boundary support vector and the lower bound of the support vector. When training the same sample set X, adjusting the parameter can cause the points in NSV, MSV and NMSV to transform. The smaller v is, the more points contained in NSV, and the fewer points contained in NMSV; the larger v is, the fewer points contained in NSV, and the more points contained in NMSV.

[0134] Figure 6 The steps of establishing the support vector machine model provided in the embodiments of the present application are shown in the schematic diagram as shown in the figure. Figure 6 The establishment process of the support vector machine model is as follows:

[0135] First, the historical data is preprocessed to extract the parameters required for initializing the model, and then based on the state information of the device in the historical data, the corresponding sample label is constructed. In the model training process, the difference between the true value and the predicted value of each sample is calculated, the current performance of the model is evaluated by analyzing these errors, and the model is adjusted accordingly.

[0136] Specifically, if the error of a certain vector sample is large, a new sample vector is selected to improve the accuracy of the model. Through continuous iteration of this process, all the samples in the data stream are selected, and the final optimized support vector machine model is obtained.

[0137] In an optional solution of the embodiments of the present application, the data preprocessing of the historical vibration data to obtain sample support vectors comprises:

[0138] Vector extraction is performed on the historical vibration data to obtain an initial support vector set;

[0139] According to the characteristics of the initial support vectors in the initial support vector set, sample support vectors are selected from the initial support vector set.

[0140] In the embodiments, the key feature information of the historical vibration data is extracted by detailed analysis and processing of the historical vibration data, the historical vibration data is converted into a vector form according to the key feature information, and an initial support vector set is obtained. Then, the vector selection is performed in time sequence, and the sample support vectors are selected from the initial support vector set.

[0141] In an optional solution of the embodiments of the present application, the updating processing of the initial vector machine model to obtain the support vector machine model comprises:

[0142] reselect support vectors from the initial support vector set according to the features of the initial support vectors in the initial support vector set;

[0143] execute the step of establishing a vector machine model using the sample support vectors with the reselected support vectors as new sample support vectors to obtain a new vector machine model, and add the vector machine model to a support vector machine model set;

[0144] select a support vector machine model from the support vector machine model set.

[0145] In this embodiment, after the initial vector machine model is established by selecting sample support vectors from the initial support vector set, new support vectors are continuously reselected from the initial support vector set according to the feature information possessed by each initial support vector in the initial support vector set, the step of establishing a vector machine model using sample support vectors is executed again with the new support vectors as sample support vectors, and the initial vector machine model is updated; and the new vector machine model is recorded in the support vector machine model set in this process.

[0146] Specifically, when new support vectors are continuously selected from the initial support vector set, those elements suitable for being support vectors can be reselected from the initial set according to the time sequence; it is ensured that all vectors in the set have a chance to be selected and each vector is selected only once, and repetition is avoided; and meanwhile, this is also helpful to maintain the representativeness and diversity of the support vectors.

[0147] In an optional solution of the embodiment of the application, the method further includes:

[0148] determining whether there is an unselected support vector in the initial support vector set;

[0149] in the case where there is an unselected support vector in the initial support vector set, the step of reselecting sample support vectors is continuously executed.

[0150] In this embodiment, after the initial vector machine model is updated and the new vector machine model is added to the support vector machine model set, it is further determined whether all support vectors in the initial support vector set are selected to construct a support vector machine model, so as to ensure that all support vectors have a chance to participate in the construction process of the model; if there is an unselected support vector, the selection process is continuously executed until all support vectors are considered, and a loop iteration method is used to ensure that the model can fully use the information in the initial support vector set, thereby improving the accuracy and generalization ability of the model.

[0151] The above is the implementation step of the embodiment, and the sliding window will be specifically applied to support vectors and described in detail.

[0152] The core operation of SVM is to solve the quadratic programming problem. If the length of the sliding window is m, and the quadratic programming problem is solved once every time the monitoring parameter slides, then the single calculation amount is o(m 3 ), which cannot meet the needs of real-time alarm. The application proposes an improved SVM algorithm through incremental and decremental learning methods, as follows:

[0153] In the continuous estimation process, the monitoring parameters in the current sliding window are taken as the target samples. The addition of new data is called the incremental process, and the deletion of old data is called the decremental process. The total number of samples in the sliding window remains m. When the data slides, the incremental process is processed first, and then the decremental process is processed. By adjusting the parameters α i and g i , it is ensured that the KKT condition still holds in the target sample set.

[0154] Let the new data x c be added to the target sample set, and α c be its corresponding Lagrange coefficient. After x c is added, the adjustment amount of α c is Δα c , and the differential form of the KKT condition equality is as follows:

[0155]

[0156]

[0157] For ease of analysis, first assume that non-support vectors, boundary support vectors, and non-boundary support vectors do not transform into each other. For x i ∈MSV, g i ≡0, i.e., Δg i =0, only α i changes, and the following matrix operation can be obtained:

[0158]

[0159] where K M is the kernel matrix of the boundary support vector MSV. Decomposition is performed to obtain:

[0160] Δρ=-βΔα c ;

[0161] Δα j =-β j Δα c j∈S;

[0162] In the formula, R=-Q M-1 ; for x i ∈ NMSV and x i ∈ NSV, α i respectively remain 0 and unchanged, i.e. Δα i = 0, only g i changes, which can be obtained from the differential form equation:

[0163] Δg i = γ i Δα c ;

[0164]

[0165] Thus the change relationship between Δα c and Δρ, the change relationship between Δα c and the boundary support vector set parameter α i , and the change relationship between Δα c and the non-boundary support vector and the non-support vector corresponding g i are determined, and the influence of α c change on other target sample parameters is described when the non-support vector, the boundary support vector and the non-boundary support vector do not change.

[0166] In actual cases, Δα c may cause the data between the three data sets to change, in order to reduce the complexity of the calculation, α c needs to change gradually according to the process of data set transformation until the balance state, which is that x c is determined to be one of the non-support vector, the boundary support vector or the non-boundary support vector, and the entire target sample set satisfies the KKT condition.

[0167] As shown in Figure 7a , it is assumed that there are sample points x1, x2,..., x l ∈ R n , and l is the number of samples. Let a new sample x c be added to the incremental operation process of the target sample set as follows:

[0168] Step one: calculate g c , if g c > 0, it indicates that x c ∈ NSV (non-support vector set, denoted as NSV), the whole process is ended, g c calculation formula is as follows:

[0169]

[0170] Where ρ is the boundary threshold, v is the tradeoff parameter 0≤v≤1, and K(x) is the threshold value. i ,x j )=<φ(x i ),φ(x j )>=K i,j Let α be the kernel function. i ≥0 represents the Lagrange coefficient.

[0171] Step 2: If g c ≤0, according to the following Figure 7a Calculate Δα when all datasets change. c Search for Δα c Minimal conversion operations;

[0172] Step 3: Adjust ρ and α j j∈S, adjust

[0173] That is: Δρ=-βΔα c , Δα j =-β j Δα c j∈S、Δg i =γ i Δα c ;

[0174] Step 4: Adjust each dataset;

[0175] Step 5: If x c If the system enters either the boundary support vector set or the non-boundary support vector set, the process stops; otherwise, proceed to step two.

[0176] When the boundary support vector set MSV changes, R also changes. To avoid repeated matrix inversion operations, the Woodbury formula is used. When sample x... d When adding MSV:

[0177]

[0178] In the formula When sample x in MSV o When removing from the set:

[0179] R ij ←R ij -R oo -1 R io R oj , where i,j∈S,i,j≠o.

[0180] like Figure 7b As shown, unlike the incremental process, when the old sample x cWhen x is removed from the target sample set, if c If a vector is not a support vector, it can be directly deleted; otherwise, the parameters of other target samples need to be adjusted gradually until x... c When entering a non-support vector set, the corresponding reduction operation process is as follows:

[0181] Step 1: If α c =0 indicates that x c ∈NSV, the entire process ends;

[0182] Step 2: If α c >0, according to Figure 7b Calculate Δα when all datasets change. c Search for Δα c Minimal conversion operations;

[0183] Step 3: Adjust ρ and α j j∈S, adjust

[0184] Step 4: Adjust each dataset;

[0185] Step 5: If x c If the input enters a non-support vector set, stop; otherwise, proceed to step two.

[0186] Furthermore, during the transformation between non-support vectors, boundary support vectors, and non-boundary support vectors, it is possible for the boundary support vector set to be empty. According to the differential form formula 2... It can be seen that, due to Δα c It can only equal 0, and cannot be used to adjust other sample parameters, causing the algorithm to stop. However, from the differential form of formula 1... It can be seen that at this time Δg i = -Δρ, therefore -Δρ is used to replace Δα. c , to obtain γ i =1, according to Δg i =γ i Δα c Update the parameter g of the current target sample i The algorithm was able to continue.

[0187] The core idea of ​​the online support vector machine algorithm is to ensure that the KKT conditions are satisfied on the current target sample set through local adjustments; when a new sample is introduced and an old sample is removed, the number of target samples in the sliding window remains m, therefore α i The upper bound remains unchanged at 1 / vm, and the form of the dual problem remains unchanged. If the sliding step size s>1, it can be decomposed into s single-in-single-out processes for processing, and the result is still an accurate global optimal solution for quadratic programming.

[0188] In addition, the online SVM algorithm makes full use of the results of the previous time, and recursively calculates the distribution region of the current target sample, thereby reducing the calculation amount at each data update. When the added or deleted sample points are non-support vector points, the distribution of the data in the two adjacent sliding windows does not change, and only the g c The calculation results of the calculation formula are judged, i.e., all operations are completed. Therefore, for a stable and unchanged data set, in addition to the first sliding window requiring more operations, the calculation amount at the subsequent time only depends on the number of support vectors, i.e., one kernel function operation is performed for each support vector.

[0189] When a boundary support vector or a non-boundary support vector appears in the new data, or the deleted sample point is a boundary support vector or a non-boundary support vector, then the target sample in the sliding window needs to be ensured to maintain the KKT condition through a finite-step analysis calculation. Since each distribution change will generate an analysis calculation, therefore, under the same number of data points, the observation sequence with continuous distribution changes has the largest calculation amount.

[0190] For example, in order to verify the calculation efficiency of the online support vector machine algorithm and the influence of data distribution on the calculation efficiency, a two-dimensional data set is constructed, which is composed of random points subject to two two-dimensional Gaussian functions, where μ1=[0.2, 0.2], σ1=[0.2, 0.2], μ2=[-0.2, -0.2], σ2=[0.2, 0.2], and the appearance order of the data points is divided into two types: random arrangement and arrangement according to the horizontal coordinate. Taking the standard support vector machine as a comparison object, the sliding window length m=50, the sliding step s=1, ν=0.1, and the kernel parameter σ=0.33 are taken, and the same parameters are used for the two algorithms to continuously slide and estimate the data distribution region. The CPU is 2.4G, Matlab is selected as the software platform, the operation time when the data points are randomly arranged is shown in Table 1, and the operation time when the data points are arranged according to the horizontal coordinate is shown in Table 2:

[0191] Table 1

[0192]

[0193] Table 2

[0194]

[0195] Through analysis, it can be found that under two different data arrangements, the time used by the online algorithm is less than that of the standard support vector machine, because the online algorithm only adjusts part of the parameters when the sample window slides, while the standard algorithm needs to recalculate the distribution of 50 samples in the feature space. The arrangement order of data points has no effect on the standard algorithm, but has a greater effect on the online algorithm. When the data set is arranged in the order of the abscissa, the distribution of the target samples in the sliding window is constantly changing, and the calculation time is about twice that of the random distribution. By averaging the times in the first two tables and removing the calculation effect of the first sliding window, the following conclusions are obtained: for a two-dimensional data set and under the current parameter settings, the standard algorithm takes an average of 0.018 seconds per sliding sample window, the algorithm takes an average of 0.005 seconds when the samples are randomly arranged, and the algorithm takes an average of 0.009 seconds when the samples are sequentially arranged. Since the single calculation amount of the standard algorithm is o(m 3 ), and the single calculation amount of the online algorithm is independent of the length of the sliding window, therefore, when the window length increases, the efficiency gap between the online algorithm and the original algorithm will further increase.

[0196] In addition, the efficiency of the algorithm is also reflected in its dynamic management of support vectors. Under the sliding window mechanism, only part of the data points are retained, and these data points are the points that have the greatest impact on the model in the current window. In this way, the algorithm can reduce unnecessary calculations and focus on data points that have a significant impact on the current decision boundary, not only improving the calculation efficiency of the algorithm and reducing the impact of outdated data on the model, but also to some extent improving the generalization ability of the model.

[0197] In practical applications, the performance of the algorithm is also affected by the size of the sliding window and the sliding step. The window size determines the amount of data that the algorithm can handle, while the step affects the frequency of algorithm updates. Choosing the right window size and step is crucial to balance calculation efficiency and model accuracy. If the window is too small, it may lead to frequent model updates and increase the computational burden. If the window is too large, it may not be able to capture changes in data distribution in a timely manner, affecting the accuracy of the model.

[0198] To further optimize the performance of the online support vector machine algorithm, an adaptive mechanism can be introduced to automatically adjust the size and step of the sliding window according to the dynamic changes in data distribution. For example, when a significant change in data distribution is detected, the algorithm can reduce the sliding step to improve the response speed of the model. Conversely, if the data distribution is relatively stable, the step can be increased appropriately to reduce the calculation amount.

[0199] The time scale based alarm method provided in the embodiments of the application, in combination with an online support vector machine algorithm, provides an efficient and adaptive solution, which is suitable for scenarios requiring real-time monitoring and rapid response, and can effectively process large-scale data streams and maintain the accuracy and stability of the model when the data distribution changes through dynamic adjustment of model parameters and optimization of the calculation process.

[0200] In an optional solution of the embodiments of the application, before the support vectors are reselected from the initial support vector set according to the features of the initial support vectors in the initial support vector set, the method further includes:

[0201] inputting the historical vibration data into a vector machine model to output distribution features corresponding to the historical vibration data;

[0202] obtaining state parameters of the target device according to the distribution features;

[0203] obtaining a parameter difference between the state parameters and historical state parameters recorded in the state labels of the historical vibration data;

[0204] in a case where the parameter difference does not satisfy a preset difference rule, deleting the sample support vectors from the initial support vector set.

[0205] In the embodiments, the historical vibration data is input into a vector machine model, and the distribution features corresponding to the historical vibration data can be obtained through processing of the model. Then, the state parameters of the target device are obtained according to the distribution features. Further, the parameter difference between the state parameters and the historical state parameters recorded in the historical vibration data is obtained. If the parameter difference does not satisfy the preset difference rule, the data serving as the sample support vectors are deleted from the initial support vector set.

[0206] Based on the above, the specific embodiments of the time scale based alarm method are described in detail.

[0207] Specifically, in the embodiments, to evaluate the alarm effect of the dynamic monitoring method on the abnormal state of the device, four kinds of data distribution situations are processed, including stable distribution, stable distribution with sparse singular points, distribution region mutation and distribution region slow change, to simulate four situations in the device running process, including normal smooth running, smooth running process containing sparse abnormal data, state mutation and slow deterioration of state.

[0208] The primary research objective of Support Vector Machines (SVMs) is the distribution in high-dimensional space, thus possessing unique advantages in learning high-dimensional data. However, for intuitive description, one-dimensional simulated data is used for analysis. Points where the decision equation equals zero are the boundary support vector points. The boundary support vectors at time step time determine the distribution boundary of the data within the sliding window at that time step; therefore, the boundary can be directly represented in the one-dimensional case. A sliding window length of m = 50, a sliding step size of s = 1, and a parameter v = 0.1 are selected. A Gaussian function is chosen as the kernel function, with a kernel parameter σ = 0.3.

[0209] Figures 8-11 The diagrams show the boundary support vectors and anomaly indicators calculated at different times under four different scenarios. At time k, the data within the sliding window is {x}. k-m+1 ,…,x k-1 ,x k The distribution D is calculated. k The boundary support vector (NSV) and data points are plotted together on the left side of the diagram, along with the anomaly index I. k+1 It is drawn on the right side of the diagram.

[0210] like Figure 8 As shown, when the device is running smoothly, the state variables remain stable and the distribution of the monitoring parameters is stable. Therefore, the boundary support vectors are roughly the same at different times, and the abnormal indicators only fluctuate within a very small area.

[0211] In two scenarios, a stationary distribution dataset with sparse singularities may be obtained: one is when outliers appear during monitoring, and the other is when there is an adjustment process between the two data acquisition points, allowing the equipment status to recover in a timely manner. When a singularity appears in the monitoring data at time k+1, due to the distribution D at time k... k The middle does not contain such a point, therefore I k+1 Increase the value and continue calculating D. k+1 x k+1 The point enters the non-boundary support vector set, falling outside the distribution region. At the next time step, if no new outliers appear, the point remains in the non-boundary support vector set until it is removed from the sliding window.

[0212] like Figure 9 As shown, sparse outliers only cause changes in the outlier indicators at the corresponding time point, and do not affect the boundary support vectors and the identification of other data points in the distribution area, thus exhibiting good robustness.

[0213] like Figure 10 As shown, when the device state undergoes a sudden change, the distribution area of ​​the feature data changes accordingly. At the critical point, I k+1 It increases rapidly, and subsequent data points continue to fall on D. kIn addition, data points in the non-boundary support vector set gradually move to the boundary support vector set, and the distribution boundary changes. With the sliding of the time window, the distribution area re-stabilizes and corresponds to a new state. Similarly, after passing through an area with a sharp change, the anomaly index re-tends to be stable. Compared with a data set with sparse singular points, the anomaly index is in a concentrated area, and the fluctuation range of the anomaly index also changes significantly after the state changes.

[0214] As shown in Figure 11 , when the gradual deterioration of the device state is accompanied by a slow transition of the feature data distribution area, although part of the data may be outside the distribution range at the previous moment, the deviation is usually small. In addition, the change speed of the target sample distribution in the subsequent data in the sliding window is faster than the evolution speed of the device state itself. Therefore, the anomaly index is maintained within a limited range and does not change significantly.

[0215] The support vector machine is committed to finding the best balance point between the compactness of the data distribution and the inclusion of a sufficient number of samples, resulting in some training data points being excluded from the target category. It is this balance strategy that ensures that the device state dynamic adaptive warning method based on the analysis time scale remains sensitive to abnormal states while not being disturbed by a small number of abnormal points.

[0216] In addition, the method can continuously adapt to changes in data and estimate and characterize changes in state quantities through distribution areas with strong generalization ability. The established anomaly index not only has important theoretical significance, but also simplifies the complexity of threshold setting in practical applications and can achieve satisfactory warning effects.

[0217] In summary, the device state dynamic adaptive warning method based on the analysis time scale can accurately predict and warn the state of complex mechanical and electrical equipment in a complex working environment and under varying conditions.

[0218] Figure 12 A structure diagram of an alarm device based on a time scale provided for an embodiment of the present application is shown in Figure 12 , and the device specifically includes:

[0219] A first acquisition module 1201 is configured to acquire vibration data of a target device.

[0220] A sample determination module 1202 is configured to determine predicted sample data of the target device from the vibration data according to a time scale.

[0221] An output module 1203 is configured to input the predicted sample data into a support vector machine model and output distribution features corresponding to the predicted sample data.

[0222] The index determination module 1204 is configured to perform anomaly detection on the distribution feature, and determine an anomaly index of the target device.

[0223] The execution module 1205 is configured to execute a pre-warning strategy if the anomaly index exceeds an alarm threshold.

[0224] In a possible implementation, the sample determination module 1202 is further configured to determine a target time period corresponding to the prediction time point by using a time scale, determine a data sequence corresponding to the target time period from the vibration data, and use the data sequence as the prediction sample data of the target device.

[0225] In a possible implementation, the sample determination module 1202 is further configured to create a corresponding time label according to a vibration time point, mark the time label to corresponding vibration sub-data, determine a start time point and an end time point corresponding to the target time period, and determine a vibration sub-data sequence corresponding to the start time point to the end time point from the vibration data by using the time label, as the data sequence corresponding to the target time period.

[0226] In a possible implementation, the apparatus further includes the apparatus further includes:

[0227] The second acquisition module 1206 is configured to acquire historical vibration data of the target device.

[0228] The preprocessing module 1207 is configured to perform data preprocessing on the historical vibration data, to obtain a sample support vector.

[0229] The establishment module 1208 is configured to establish a vector machine model by using the sample support vector, as an initial vector machine model.

[0230] The update module 1209 is configured to perform update processing on the initial vector machine model, to obtain a support vector machine model.

[0231] In a possible implementation, the preprocessing module 1207 is further configured to perform vector extraction on the historical vibration data, to obtain an initial support vector set, and select a sample support vector from the initial support vector set according to a feature of an initial support vector in the initial support vector set.

[0232] In a possible implementation, the preprocessing module 1207 is further configured to reselect a support vector from the initial support vector set according to a feature of an initial support vector in the initial support vector set; perform the step of establishing a vector machine model by using the sample support vector by taking the reselected support vector as a new sample support vector, to obtain a new vector machine model, and add the vector machine model to a support vector machine model set; and select a support vector machine model from the support vector machine model set.

[0233] In a possible implementation, the apparatus further includes a judging module 1210 (not shown in the figure) configured to judge whether there is an unselected support vector in the initial support vector set; and continue to perform the step of reselecting a sample support vector in a case where there is an unselected support vector in the initial support vector set.

[0234] In a possible implementation, the preprocessing module 1207 is further configured to input the historical vibration data into a vector machine model, and output a distribution feature corresponding to the historical vibration data; obtain a state parameter of the target device according to the distribution feature; obtain a parameter difference between the state parameter and a historical state parameter recorded in a state label of the historical vibration data; and delete the sample support vector from the initial support vector set in a case where the parameter difference does not satisfy a preset difference rule.

[0235] The time-scale-based alarm device provided in this embodiment can be a time-scale-based alarm device as shown in Figure 12 , can perform all steps of the time-scale-based alarm as shown in Figures 1-11 , and thus achieves the technical effects of the time-scale-based alarm as shown in Figures 1-11 . For details, refer to the related description, which will not be repeated here for brevity. Figures 1-11

[0236] The apparatus embodiments described above are merely schematic, wherein the units described as separate components can or can not be physically separated, and the components shown as units can or can not be physical units, that is, can be located in one place or distributed on multiple network units. Part or all of the modules can be selected according to actual needs to achieve the purpose of the embodiment scheme.

[0237] ​Those skilled in the art can clearly understand that each embodiment can be implemented by means of software plus a general hardware platform, and of course can be implemented by hardware. Based on such an understanding, the technical solutions described above essentially or in other words make contributions to the related art, and can be embodied in the form of a software product. The computer software product can be stored in a computer readable storage medium, such as a ROM / RAM, a magnetic disk, or an optical disk, and includes a plurality of instructions used to cause a computer device (which can be a personal computer, a server, or a network device, etc.) to execute the methods described in each embodiment or some parts of the embodiments.

[0238] It is to be understood that the terminology used herein is for the purpose of describing particular example embodiments only and is not intended to be limiting. As used herein, the singular forms "a", "an" and "the" are intended to include the plural forms as well, unless the context clearly indicates otherwise. The terms "comprises", "comprising", "includes", "including" and "has" are inclusive and therefore specify the presence of stated features, steps, operations, elements, and / or components, but do not preclude the presence or addition of one or more other features, steps, operations, elements, components, and / or groups thereof. The method steps, processes, and operations described herein are not to be construed as necessarily requiring their performance in the particular order in which they are described unless specifically indicated as such. It is also to be understood that additional or alternative steps can be employed.

[0239] The above description is merely that of specific embodiments of the present application, and thus is not intended to limit the present application. Various modifications to these embodiments can be apparent to those skilled in the art, and the generic principles defined herein can be applied to other embodiments without departing from the spirit or scope of the application. Accordingly, the present application is not intended to be limited to the embodiments shown herein but is to be accorded the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. A time-scale-based alarm method, characterized in that, include: Acquire vibration data of the target equipment; Predictive sample data for the target device is determined from the vibration data according to a time scale; Acquire historical vibration data of the target device; perform data preprocessing on the historical vibration data to obtain sample support vectors; A vector machine model is built using the sample support vectors, which serves as the initial vector machine model. The process of updating the initial vector machine model to obtain a support vector machine model includes: reselecting support vectors from the initial support vector set based on the characteristics of the initial support vectors in the initial support vector set; using the reselected support vectors as new sample support vectors to perform the step of building a vector machine model using the sample support vectors to obtain a new vector machine model, and adding the new vector machine model to the support vector machine model set; and selecting a support vector machine model from the support vector machine model set. Before reselecting support vectors from the initial support vector set based on the characteristics of the initial support vectors in the initial support vector set, the method further includes: inputting the historical vibration data into a vector machine model and outputting the distribution characteristics corresponding to the historical vibration data; obtaining the state parameters of the target device based on the distribution characteristics; obtaining the parameter difference between the state parameters and the historical state parameters recorded in the state labels of the historical vibration data; and deleting the sample support vector from the initial support vector set if the parameter difference does not meet a preset difference rule. The predicted sample data is input into the support vector machine model, and the distribution features corresponding to the predicted sample data are output. Anomaly detection is performed on the distribution characteristics to determine the abnormal indicators of the target device; If the abnormal indicator exceeds the alarm threshold, an early warning strategy will be implemented.

2. The method according to claim 1, characterized in that, The step of determining the predicted sample data of the target device from the vibration data according to the time scale includes: Use time scales to determine the target time period corresponding to the prediction time; Determine the data sequence corresponding to the target time period from the vibration data; The data sequence is used as the prediction sample data for the target device.

3. The method according to claim 2, characterized in that, The vibration data includes vibration sub-data at different vibration moments; Determining the data sequence corresponding to the target time period from the vibration data includes: Create a corresponding time tag based on the vibration time, and mark the time tag in the corresponding vibration sub-data; Determine the start and end times corresponding to the target time period; Using the time stamp, the vibration sub-data sequence corresponding to the start time to the end time is determined from the vibration data, and used as the data sequence corresponding to the target time period.

4. The method according to claim 1, characterized in that, The process of preprocessing the historical vibration data to obtain sample support vectors includes: Vector extraction is performed on historical vibration data to obtain an initial set of support vectors; Based on the characteristics of the initial support vectors in the initial support vector set, sample support vectors are selected from the initial support vector set.

5. The method according to claim 4, characterized in that, The method further includes: Determine whether there are any unselected support vectors in the initial support vector set; If there are unselected support vectors in the initial support vector set, the step of reselecting sample support vectors continues.

6. An alarm device based on a time scale, characterized in that, include: The first acquisition module is used to acquire vibration data of the target device; A sample determination module is used to determine predicted sample data of the target device from the vibration data according to a time scale; The second acquisition module is used to acquire historical vibration data of the target device; The preprocessing module is used to preprocess the historical vibration data to obtain sample support vectors; A module is established to build a vector machine model using the sample support vectors, which serves as the initial vector machine model. The update module is used to update the initial vector machine model to obtain a support vector machine model, including: reselecting support vectors from the initial support vector set based on the features of the initial support vectors in the initial support vector set; using the reselected support vectors as new sample support vectors to perform the step of building a vector machine model using the sample support vectors to obtain a new vector machine model, and adding the new vector machine model to the support vector machine model set; and selecting a support vector machine model from the support vector machine model set. Before reselecting support vectors from the initial support vector set based on the characteristics of the initial support vectors in the initial support vector set, the method further includes: inputting the historical vibration data into a vector machine model and outputting the distribution characteristics corresponding to the historical vibration data; obtaining the state parameters of the target device based on the distribution characteristics; obtaining the parameter difference between the state parameters and the historical state parameters recorded in the state labels of the historical vibration data; and deleting the sample support vector from the initial support vector set if the parameter difference does not meet a preset difference rule. The output module is used to input the predicted sample data into the support vector machine model and output the distribution features corresponding to the predicted sample data. The indicator determination module is used to perform anomaly detection on the distribution characteristics and determine the abnormal indicators of the target device. The execution module is used to execute the early warning strategy when the abnormal indicator exceeds the alarm threshold.

Images