Vibration monitoring for identifying faults in process automation

Abstract

The invention relates to a computer-implemented monitoring method (1), which is more reliable in fault detection, for vibration analysis for detecting machine faults in process automation, comprising statistical evaluation (4) of at least one group or at least one part thereof in order to determine deviations of measured sensor data (5) in each case on the basis of at least three criteria (K, I, II, III), the statistical evaluation provides a zero hypothesis and an expected value, and calculates a statistical significance of a deviation from a predetermined zero hypothesis in order to issue an alarm when the statistical significance deviates from the zero hypothesis and / or a probability of deviating from the zero hypothesis at a predetermined significance level alpha is greater than the significance level alpha.

Description

Technical Field

[0001] The invention relates to a computer-implemented monitoring method for vibration analysis in process automation in order to identify machine faults or hardware faults. Background Art

[0002] It is known in the prior art to monitor processes performed by machines to determine whether errors or malfunctions have occurred in the machines, computers, etc. used. This is typically done using sensors that continuously record data during the process and evaluate the data according to predetermined criteria to determine whether a fault is present. For example, DE 10 2018 222 562 A1 discloses detecting fault conditions in the excitation circuit of a motor, wherein the excitation current and control signals are filtered and compared with reference values ​​as comparison signals detected by the sensors. EP 0 934 567 A1 discloses a method for classifying statistical correlations in measurable time series, which describes iterative testing of various null hypotheses. DE 10 2019 107 363 A1 also discloses a computer-implemented method for outputting a wear signal for a machine tool using a distance metric, known as "change point detection." However, this method is generally not sufficiently reliable for long-term, 24 / 7 monitoring.

[0003] In addition, data quality assessment with the help of null hypothesis has also been mentioned in the paper "Fractional Dynamics of PMU Data" by L. Shalalfeh et al. (IEEE Transactions on Smart Grid, IEEE, USA, Vol. 12, No. 3, pp. 1-11, May 2021), paper number XP011850407, ISSN: 1949-3053, DOI: 10.1109 / TSG.2020.3044903).

[0004] Furthermore, EP 0 907 913 B1 discloses a method for diagnostic differentiation, which uses a histogram to measure control deviations, wherein various sources of interference can be inferred. Summary of the Invention

[0005] The object of the present invention is to provide a monitoring method for fault detection which allows particularly reliable fault detection.

[0006] This object is achieved based on the monitoring method previously known from the prior art by the features of claim 1 .

[0007] Advantageous embodiments and further developments of the invention are achieved by means of the measures mentioned in the dependent claims.

[0008] The monitoring method according to the invention is intended for vibration analysis in process automation. For example, a machine with a rotor (e.g. a fan) can be checked for possible faults by detecting vibrations and their deviations, such as fault deviations that develop slowly according to a trend. Therefore, the monitoring method according to the invention for identifying machine faults or hardware faults in a process performed by a machine, which is derived from vibration analysis, first also includes detecting at least one set of sensor data measured during the course of the process. These data detected by the sensors are usually recorded in a time series. However, it is also conceivable, for example, that this set of sensor data can represent a local sequence or a local process if the sensor is to detect whether the same distance is always maintained between two components, or whether a component has not left its predetermined storage position due to heating, imbalance or other abnormal conditions.

[0009] The teaching of the present invention is characterized in that the fault identification sought can also include fault prediction. The present invention recognizes that machine or hardware faults are usually "forecasted", especially when monitoring vibrations or rotations, that is, sudden spontaneous failures of the machine rarely occur without any signs, but rather that the fault usually has some symptomatic phenomena that are indicative of the fault and can be measured, because, for example, the measured variables related to the machine / hardware gradually change. The changing measured variables can usually be detected by sensors, such as output current, input current, rotational speed (for example for blowers), etc. Usually, the machine or hardware can continue to perform its work, sometimes initially even without or without noticeable loss of production quality, until a major fault occurs in the machine.

[0010] In this respect, the invention also enables optimization of maintenance intervals, that is, for an existing machine inventory, it is possible to predict when maintenance will be required and how long or short the maintenance intervals should be.

[0011] According to the present invention, the following difficulty is recognized: the measured sensor data are masked by statistical errors, in particular statistical noise, which makes it difficult to evaluate the individual measured sensor data because they are affected by statistical errors and may deviate from predetermined target values ​​even in the absence of errors.

[0012] It is recognized that, for example, a variable detected by a sensor does not change or changes slowly on average, but the statistical fluctuations of the variable around an expected value may increase.

[0013] Furthermore, the present invention takes into account for the first time that occurring faults often affect the detected sensor data in various ways. Therefore, during the evaluation and investigation process, statistical anomalies are primarily addressed, regardless of their cause, as simple comparison and detection of deviations in individual values ​​have proven insufficient as a reliable criterion for error detection. Consequently, the method according to the present invention is applicable to a wide range of applications, as no special requirements for the monitored machine are required and, generally, only the generally expected behavior of the measured variable detected by the sensor is required, even if, for example, the measured variable has a single constant value over time.

[0014] According to the present invention, at least two, preferably at least three, criteria are checked before a warning signal is issued. A basic criterion considered essential for fault detection according to the present invention is a check of whether the measured data is constant within a predetermined tolerance range. To enable largely parallel processing, these criteria can, in one embodiment, be determined and checked simultaneously, if necessary, before a warning is issued.

[0015] Because the basic criterion is crucial and represents a more universal test than other criteria, such as trend analysis or breakout investigations, and serves as the foundation for trend or breakout deviations, it is also checked first in the first step. Only if this basic criterion reveals a deviation is the additional criteria checked. These subsequent criteria can also be evaluated in parallel for faster evaluation.

[0016] Advantageously, one embodiment of the present invention also provides a multi-stage (at least three-stage) warning system, wherein individual warning levels are displayed depending on the number of criteria met when a fault is detected. This also provides a monitoring method that weights occurring faults, so that the operator is informed of the severity of the anomaly and can specifically assess whether and what measures should be taken to intervene in the execution of the process or machine.

[0017] If, in a first step, a basic criterion is initially checked and in a subsequent step further criteria are checked, the warning system can also advantageously adapt to this situation by issuing an at least two-stage warning that takes this division into account.

[0018] The present invention is particularly suitable for process automation of systems that are continuously operated and whose vibration behavior or vibration characteristics are closely related to the operation. As examples, the monitoring of vacuum pumps for clean room applications or the monitoring of fans that are operated 24 hours a day, 7 days a week should be mentioned. By analyzing the vibration signal according to the invention, it is advantageous to provide an indication of initial damage very early, so that the machine or system can be repaired or maintained in time before major damage and failure occur. Based on a variety of such indications, maintenance intervals can also be adjusted, that is, if it is possible to estimate how often such damage occurs during operation and when it is expected to occur. The system can be adapted to various applications very quickly for long-term monitoring. In a preferred improvement, a vibration sensor is used as the sensor, for example, to monitor continuously running fans, bearings, pumps, vacuum pumps or other motors.

[0019] For example, an advantageous application can be found in the field of medical technology, where technical and biological processes interact, which can favor the generation of statistical deviations and make the identification of faults requiring intervention more difficult to distinguish from simple statistical deviations. Another advantageous application is in the field of cooling in production processes (ventilators, fans, etc.).

[0020] According to the present invention, expected values ​​for the measured sensor data are calculated and provided. These expected values ​​are required, for example, for statistical evaluations. Deviations from the measured sensor data can be compared to the expected values. The expected value describes the average value of a random variable (in this case, the sensor data) over at least a specific time period. Even if a trend exists, it is possible to check, for example, whether the deviation from the expected value increases over time. The expected value can also be used to determine sensitivity to fluctuations.

[0021] According to the present invention, a statistical evaluation of at least one recording sequence or a portion of at least one recording sequence is provided. In one advantageous embodiment, at least three criteria are also checked so that a corresponding warning level can be assigned depending on how many of these criteria are simultaneously met. The monitoring method according to the present invention is typically implemented as a computer-implemented method, in which the sensor data is evaluated electronically or computer-assisted. This overcomes the technical disadvantage that signals containing statistical errors or noise cannot be statistically evaluated with respect to error and plausibility.

[0022] The statistical evaluation according to the present invention is performed by establishing a null hypothesis. If the null hypothesis is deviated from, or the probability of deviation from the null hypothesis is higher than a predetermined significance level α, a warning is issued.

[0023] In a particularly preferred development of the invention, for example, at least one of the following three criteria can be checked in the statistical evaluation:

[0024] Typical errors that can be identified using statistical methods are jumps in the average sensor data level within a temporal or spatial series. These jumps must be distinguished from individual deviations, which are caused by noise and are therefore purely statistical and do not represent true errors. Such jump-like errors can occur, for example, when the measured variable actually undergoes sudden but persistent changes (i.e., deviations from stable behavior). However, errors that result in such level deviations, for example, are also conceivable due to sensor-induced errors. Consequently, the average value (at least averaged over larger temporal or spatial intervals) can vary over time.

[0025] Another statistically detectable error can be when the measured sensor data continuously changes according to a consistent trend ("drift"). In this case, there is also a deviation from a stable behavior. The time average value changes continuously.

[0026] - Even if the mean value remains constant on average over a large time or space interval, the sensitivity or fluctuation of the fluctuation may change significantly or especially increase when errors occur.

[0027] In one embodiment of the present invention, the expected behavior can be described by a model function consisting of the sum of a function describing stationary behavior, a function representing random walk for noise modeling, and a linear function describing trend behavior over time / spatial distance, thereby reflecting typical effects such as background noise or measurement drift. Therefore, the measurement value can be described as follows:

[0028] y t =c t +δt+u t ,

[0029] Among them, y t represents the change of the modeled measurement data with time t (or position), u t Describes the stationary behavior, c t represents random walk (for example, used to simulate noise), and δt represents the trend, that is, the "deviation" or "drift" of the measured value.

[0030] These model descriptions allow for a mathematically concrete statistical assessment of the expected errors for possible errors. This model description allows the calculation of statistical quantities such as the expected value, standard deviation, and variance. For stationary behavior, the expected value can be determined by averaging. The model also allows the calculation of the deviation of the measured value from the model function, the residual.

[0031] Since the trend behavior of the measured values, in particular a continuous drift of the measured values, can represent an error, δ=0 is initially assumed in the model.

[0032] The first criterion is to check whether there are sudden jumps in the average level of the measured data, so-called level shifts. In the case of stationary behavior, the variance is ideally equal to 0. If the measured values ​​are affected by noise, the variance lies around a certain expected value (ideally the mean value). According to the null hypothesis, the random walk component c t The volatility of is equal to zero and deviates from zero only in the case of overall non-stationary behavior (even taking into account potential noise). In the case of non-stationary behavior, the expected value essentially changes whenever it is redefined in successive time intervals. In this case, statistical observations can include, for example, the occurrence of jumps that shift the measured values ​​so strongly overall that, given a certain variance, the deviation from the expected value due to potential noise can no longer be explained by the noise behavior.

[0033] For the null hypothesis, the stationary KPSS test (KPSS: Kwiatkowski, Phillips, Schmidt, and Shin) can be used:

[0034] For a number T of measurements, the test statistic for the null hypothesis can then be considered to be

[0035]

[0036] Among them, S t It is the residual e at each measurement point t (usually each time point t) relative to the model function or regression curve t sum

[0037] S t =e1+e2+e3+...+e t ,

[0038] s 2 is the variance relative to the expected value. When checking the first criterion (whether there is a horizontal shift), it is usually advantageous to use a relatively small amount of measurement data in order to form the sum in the test function, since it is not necessary for the point at which the potential jump occurs in the data to be located within the time or position interval to be checked when making the comparison. The test statistic follows a certain distribution. In order to test the null hypothesis, it is now possible to check how probable a given test statistic is. Below a certain probability or significance, the null hypothesis can be considered incorrect. The significance level α, which should not be exceeded, can be set to 1%, for example; otherwise, the deviation from the test statistic is too large and a warning needs to be issued.

[0039] If the volatility of the model function = 0, there is a stationary behavior that may be accompanied by noise.

[0040] To check the second criterion—whether a trend curve exists or whether the measured data consistently "deviates"—it usually requires evaluating and comparing more measurement points. Otherwise, the KPSS test can also be used for this criterion. As a condition, δ = 0 is still assumed in the model function, as otherwise a trend must be assumed and autocorrelation must be ignored.

[0041] If the significance is now less than a certain predetermined value and a large number of measurement data are used, a trend curve can be assumed which meets the second criterion.

[0042] However, the checks for the first and second criteria differ only in the number of test data or the measured time intervals. The checks are performed accordingly:

[0043] - whether the volatility of the random walk term is zero or greater in magnitude than a specified value; and / or

[0044] -The test statistic of the stationary KPSS test satisfies a certain predetermined probability.

[0045] To check the third criterion, whether the fluctuation range or fluctuation sensitivity is steadily increasing, it is sufficient to compare the variances in two different test periods or measurement intervals. The null hypothesis to be tested is that the variance remains constant or at least deviates only with a predetermined probability. If the fluctuation range increases, the variance will also increase in subsequent periods or measurement intervals. The amount of test data may be smaller than when checking the second criterion of the trend curve.

[0046] Therefore, from a statistical point of view, the significance can also be determined relative to a predetermined probability value, wherein if the significance is below the probability value, the fluctuation range is increased and the third criterion is met. BRIEF DESCRIPTION OF THE DRAWINGS

[0047] Embodiments of the present invention are shown in the accompanying drawings and will be described in detail below, along with other details and advantages. Specifically:

[0048] Figure 1 is a schematic diagram of a monitoring method for fault identification and fault prediction according to the present invention;

[0049] Figure 2-4 Various measurement data sequences with superimposed noise are shown to illustrate fault identification;

[0050] Figure 5 shows a series of measurement data over a longer period of time, containing various errors, including a representation of the error terms;

[0051] Figure 6 is a schematic diagram of a system for machine monitoring of multiple vacuum pumps;

[0052] Figure 7 is a schematic diagram of the three criteria that are evaluated in parallel before a warning signal is issued; and

[0053] Figure 8 It is a schematic diagram of the step-by-step assessment before a warning signal is issued. DETAILED DESCRIPTION

[0054] Figure 1 The schematic diagram shows a computer-implemented monitoring method 1 for fault identification and prognosis according to the present invention. A device 2 is monitored using sensors 3, which transmit the measured data to a computer-controlled evaluation unit 4. This evaluation unit 4 can be, for example, a control unit or controller within the machine, or it can also run separately on a computer.

[0055] Evaluation unit 4 performs statistical evaluation of the measurement data:

[0056] The evaluation unit first specifies a statistical significance level, for example 1%, and also specifies at least one criterion. Here, the evaluation unit 4 examines the basic criterion and three other criteria I, II and III to determine:

[0057] - According to the basic standard, are there any deviations that could lead to failure, because the measured data are unstable within the predefined tolerance range?

[0058] -First Criteria I:

[0059] Is there a horizontal shift in the curve of the measured data?

[0060] -Second Standard II:

[0061] Is there a trend-curve?

[0062] -Third Standard III:

[0063] Has volatility or volatility sensitivity changed?

[0064] The so-called null hypothesis is used to determine whether one of the criteria is met.

[0065] This is based on a model function that accounts for statistical errors, e.g. via random walks, but also contains a term that can describe fundamentally undesirable trend behavior.

[0066] Trend behavior is initially considered to be absent, i.e., the terms describing the trend are set to zero.

[0067] For increased volatility III, it is sufficient to examine the variances of two (time-sequentially) detected subsequences by the sensor. Ideally, their ratio = 1. For this behavior, it is also possible to consider using a test statistic and check whether the measured value behavior of the test statistic is sufficiently close.

[0068] For horizontal excursion behavior (criterion I), changes in the value are expected. Here, too, the test statistic can be used to address this. A large number of measurements is not required, as the horizontal excursion is expected to occur quite quickly, so averaging over a long period of time is not necessary to reliably detect the signal change.

[0069] For trending behavior, which can also occur very slowly, measurements must be taken over a longer period of time to identify changes.

[0070] The significance may be assessed and the deviation from the corresponding null hypothesis determined in execution block 4a.

[0071] If the measured data are not stable enough within the tolerance limits (basic criterion K) and one of the criteria I, II, or III is met, a first warning level WI is issued by the evaluation unit 4. Similarly, if two criteria are met, a more severe second warning level WII is issued; if three criteria are met, a particularly high warning level WIII is issued. Figure 1 5 shows the time course of the measured data, wherein the ranges WI, WII, WIII that meet one to three criteria are shown. It is also conceivable to determine a deviation measure from the respective null hypothesis and select the warning level WI, WII or WIII accordingly.

[0072] Warning levels WI, WII, WIII can be displayed simply, for example, by a signal lamp or signal column 6, but depending on the embodiment it is also conceivable that the evaluation unit 4 sends corresponding control commands to the device 2. These commands can also be adjusted to certain error sources that are typically present with standards I, II and III.

[0073] It must generally be assumed that the recorded measurement data of a measurement data series are superimposed with noise, e.g. Figure 2-4 This is the case in . Due to the superimposed noise, the measured data is so widely distributed in various parts of the measured data series that errors such as trend curves or level shifts are difficult to identify. In particular, if the variance of the measured data caused by noise is relatively high, and the measured values ​​are equally likely to move above or below the mean or expected value, fluctuation sensitivity or fluctuations can be difficult to identify graphically alone.

[0074] Figure 2A signal with noticeable superposition of noise is shown, plotted as a series of measured data. However, on average, the measured data remain at a constant level, shown as a dashed line. Despite the large variance of the measured values, this also exhibits constant volatility. This is because the variance does not increase over time; instead, the measured data in this section maintain a relatively constant distribution around the mean / expected value. Therefore, when observed over this period, apart from statistical outliers in the measured values, there are, on average, no signal fluctuations. The evaluation according to the third criterion, III, reveals constant volatility.

[0075] Figure 3 This also shows a signal with superimposed noise. However, the measured values ​​continue to increase over time, forming a trend curve. The dashed line representing the average value of the measured data is a straight line with a positive slope, indicating an upward trend. However, due to the large statistical fluctuations or variance of the data, it is difficult to identify this trend without evaluating the average value or expected value. This trend curve is identified based on the second criterion II.

[0076] Figure 4 Shown with Figure 2-3 A similar situation occurs when a noisy signal is present. When evaluating the mean or expected value, it is apparent that the measured values ​​in the curve suddenly jump (horizontally shift) to a higher level. The mean or expected value is represented by a dashed line with a sudden jump (horizontally shift). Before and after this jump, the measured data remain stable on average within these time periods. Therefore, in this case, the evaluation according to the first criterion (I) is highly relevant.

[0077] exist Figure 5 The first row ("Values") shows the curve of a set of measurement data over a longer period of 19 days. The subsequent rows describe the terms in the model function on a scale from 0 to 1, which describe the trend-curve (criterion II), the horizontal deviation (criterion I) and the volatility (criterion III). It can be seen in the curve of the measured data ("Values") that statistical noise is superimposed on the measured values ​​over the entire time period. The intensity of this noise also fluctuates, i.e. the variance in the range around, for example, July 4 is significantly greater than the variance in the range around June 17 to June 20. However, on average, the values ​​in the range around June 17 to approximately June 20 are constant, resulting in statistically smaller fluctuations, which are not apparent in the original text. Figure 5 This is reflected in the last row of the chart: the volatility was at the value 0 before June 20. For the trend-curve in the second row, the value was also constant at 0 before June 20, since the signal level remained unchanged on average.

[0078] The behavior of the data measured from June 30th to July 5th differs. Deviations increased, while the signal level remained constant on average. Variance increased. The accumulation of deviations increased volatility starting on June 30th. The constant average signal level was reflected by trend values ​​within the 0 range during this period.

[0079] At two discrete points, around June 22 / 23 and June 30, the signal level experienced sudden changes, which can be seen in the third row by isolated peaks at these times. Trend and volatility also change at these times, and the unusual peaks in the signal curve also change, because the jumps also shift the signal curve (trend) or cause deviations that exceed the variance.

[0080] Between June 22 / 23 and June 30, the measured signal continued to increase and rose slightly. Figure 5 In the second row of , the trend value is a value within the range of 1, because there is an upward trend curve.

[0081] Figure 5 The third row shows the horizontal deviation term, with particularly noticeable peaks in the ranges around June 22 / 23 and June 30, as mentioned above. Beyond these two peaks, the horizontal term remains close to zero, but exhibits irregular deviations. This is because the measured signal ("value") is not always statistically constant on average, depending on the range under consideration. This means that the signal level, already affected by noise, can vary slightly.

[0082] In addition, higher statistical fluctuations and / or level changes that occur in the meantime (such as here around June 22 / 23 and June 30) may cause the values ​​within the fluctuation term to temporarily change and not be 0. This is because the signal undergoes changes in the form of jumps, which are also greater than the variance caused by noise, that is, the fluctuation sensitivity of the signal also changes here.

[0083] Figure 6 The schematic diagram shows a system for monitoring vacuum pumps, which uses vibration sensors with diagnostic electronics for process monitoring. The manufacturer's model number for these vibration sensors is "ifm VSE." The signals are sent to a processor for evaluation. Multiple cycles of a cyclic process are combined into a single machine cycle. With the support of numerical evaluation software, the measured data is continuously analyzed (continuous mode monitoring) to analyze fluctuations, step changes (level deviations), or trend curves. Mean values ​​or expected values ​​are also calculated. Probabilistic calculation methods are used to extrapolate recent measured data to predict data trends or to check whether the measured data lie within or outside a defined confidence interval with a certain probability.

[0084] Optionally, parameterization can be employed, which provides a more detailed analysis based on certain criteria or parameters (detailed parameterization options), for example:

[0085] - Select reference data from the dataset; and / or

[0086] - configuring objects during vibration analysis; and / or

[0087] -Analysis with the help of narrow confidence intervals to identify even small deviations or stability of the measured data.

[0088] There are generally two possible ways for users to train the pattern monitor, depending on the design of the present invention:

[0089] Option A: "Supervised": The user manually marks existing patterns in the historical data, such as trends (or similar fluctuations or jumps) within a specific time period. The algorithm then sets its own parameters to accurately identify that time period and fail to identify all other time periods (if the testing method can achieve this). A check is also performed to see if the user has forgotten to mark any areas or marked areas where there are no patterns at all.

[0090] Option B: "Unsupervised": The user loads a static time period without any patterns. The algorithm is then trained not to issue an alarm or warning signal during the selected time period, but to be as sensitive as possible to the sensor parameters (e.g., narrow confidence intervals) so that it can react to even the slightest changes or patterns. The data is also checked to see if there are any patterns that the user simply forgot to label.

[0091] Figure 7 and Figure 8 Two approaches to determining the criteria are shown. Figure 7 , three criteria are determined in parallel, including basic criterion K, trend analysis II and mutation I. Depending on which criterion is relevant, one of warning prompts WK, WI, WII (or multiple warning prompts if necessary) is issued.

[0092] Figure 8 The step-by-step determination of the criteria is shown: first, the basic criterion K is checked. If a deviation is detected here, two further criteria I and II are executed in parallel. Here, it is sufficient to output two-stage warning signals WI and WII, depending on whether the unstable measurement data exhibits a sudden change or a trend.

[0093] The monitoring method according to the invention advantageously enables a statistical evaluation of measurement data, some of which are also susceptible to statistical errors, for example due to superimposed noise, and in this way identifies any errors. The method according to the invention primarily uses statistical methods and is largely universal, regardless of the type of machine / hardware to be monitored, since it is possible to fundamentally distinguish between the occurrence of statistical errors and errors caused by real faults. According to the invention, this is achieved by determining the deviation from the null hypothesis and evaluating the statistical significance relative to a predetermined significance level. Depending on the number of error criteria met and / or the degree of deviation, they are divided into different categories, and in a further development, a classification of the severity of the situation can be achieved.

[0094] List of Reference Numerals

[0095] 1 Monitoring methods

[0096] 2 Equipment to be monitored

[0097] 3 sensors

[0098] 4 evaluation units

[0099] 4a Calculation of significance relative to the null hypothesis

[0100] 5 Measurement data and error classification

[0101] 6 signal posts

[0102] I Standard (Volatility)

[0103] II standard (horizontal offset)

[0104] III Criterion (Trend Behavior)

[0105] WI First Warning Level

[0106] WII Second Warning Level

[0107] WIII Warning Level 3

[0108] WK basic warning level

[0109] α significance level

[0110] K basic standard.

Claims

1. A computer-implemented monitoring method (1) for vibration analysis for fault identification of machine faults and / or hardware faults in process automation during machine operation and / or supported production processes, the method comprising: Detecting at least one set of sensor data (5) measured in a temporal and / or spatial sequence during the process, wherein: performing a statistical evaluation (4) of the at least one group or at least a portion of the at least one group in order to determine a deviation of the measured sensor data (5) according to at least three criteria (K, I, II, III), wherein the statistical evaluation: Provide a null hypothesis; Providing expected values ​​for the measured sensor data; Calculating the statistical significance of deviations from a predetermined null hypothesis; and If the statistically significant deviation from the null hypothesis occurs, and / or if the probability of a deviation from the null hypothesis at a predetermined significance level α is greater than the significance level α, a warning is issued, It is characterized by: Before issuing the warning, the deviation of the measured sensor data is determined in the statistical evaluation (4) according to the following criterion (K): in the statistical evaluation (4), in order to test the basic criterion (K), it is checked whether the average sensor data level remains constant in the temporal and / or spatial sequence within a tolerance range predetermined by the basic criterion (K) and is also determined according to at least one of the following criteria (I, II, III): i. in the statistical evaluation (4), in order to test the second criterion (II), checking whether there is a trend of continuous change in the average sensor data level in the time and / or space series; and / or ii. in the statistical evaluation (4), in order to test the first criterion (I), checking whether there is a jump in the temporal and / or spatial sequence of the mean sensor data levels, wherein, in particular, a jump has occurred if the jump deviates the measured values ​​as a whole so strongly that the deviation from the expected value due to underlying noise can no longer be explained by the noise behavior with a certain variance; and / or iii. In order to test the third criterion (III) during the statistical evaluation (4), it is checked whether the fluctuation sensitivity in the group changes in the temporal and / or spatial sequence.

2. The computer-implemented monitoring method for vibration analysis (1) according to claim 1, characterized in that The basic criterion (K) is always checked during the statistical evaluation (4), wherein deviations of the measured sensor data are determined during the statistical evaluation (4) first according to the basic criterion (K) and then according to two or three of the further criteria (I, II, III) simultaneously or with temporal overlap.

3. The computer-implemented monitoring method (1) for vibration analysis according to any one of the preceding claims, characterized in that During the statistical evaluation (4), deviations of the measured sensor data are determined according to the criteria (K, I, II, III) simultaneously and / or with temporal overlap before the warning is issued.

4. The computer-implemented monitoring method for vibration analysis (1) according to claim 1 or 2, characterized in that During the statistical evaluation (4), deviations of the measured sensor data are determined firstly according to the basic criterion (K) and then according to two of the further criteria (I, II, III) simultaneously or with temporal overlap.

5. The computer-implemented monitoring method (1) for vibration analysis according to any one of the preceding claims, characterized in that The statistical evaluation provides for the output of at least three, in particular four, warning levels (WK, WI, WII, WIII), wherein any number of the criteria (K, I, II, III) that are met is respectively assigned a warning level (WK, WI, WII, WIII).

6. The computer-implemented monitoring method (1) for vibration analysis according to any one of the preceding claims, characterized in that The statistical evaluation provides for the output of at least two levels of warning, depending on whether, in addition to the basic criterion, the first criterion or the second criterion also exhibits deviations.

7. The computer-implemented monitoring method (1) for vibration analysis according to any one of the preceding claims, characterized in that As a model function of the behavior of the sensor data (5) in the group, the following function is used: y t =c t +δt+u t , Where, when the sensor data of the group is plotted against t, y t Describes the modeled sensor data associated with t, c t describes random wandering, δ describes trend, δt describes linear trend behavior, and u t Describes stationary behavior, where t describes the temporal and / or spatial variation of the process.

8. A computer-implemented monitoring method (1) for vibration analysis according to any one of the preceding claims, characterized in that To check the first criterion (I), the null hypothesis is used, that is, the variance is taken as the mean squared deviation from the expected value and the random walk term c t The volatility of is zero, wherein, in particular for the model function, the trend δ=0 is assumed.

9. The computer-implemented monitoring method (1) for vibration analysis according to any one of the preceding claims, characterized in that To check the first and / or second criterion (I), a stationary KPSS test is performed, wherein it is checked whether a test statistic is met with a predetermined probability as the null hypothesis, wherein in particular: As part of the sequence T and S t The sum of squares and variance S 2 The ratio of the product of the square of the number of sequence parts T, where S t This is again the sum of the residuals of the regression curve relative to the measured sensor data at each point t, wherein a warning is issued if the test statistic is not met with less than a predetermined probability.

10. The computer-implemented monitoring method (1) for vibration analysis according to any one of the preceding claims, characterized in that To check the first and / or second criterion and / or the basic criterion (I, II, K), autocorrelations are used, in particular in the case of signal delays in the form of delay-affected sensor data, which are taken into account and / or subtracted and / or ignored.

11. The computer-implemented monitoring method (1) for vibration analysis according to any one of the preceding claims, characterized in that For checking the second criterion (II), a larger portion of the sequence with more sensor data is used than for checking the first criterion (I).

12. The computer-implemented monitoring method (1) for vibration analysis according to any one of the preceding claims, characterized in that To check the third criterion (III), at least two partial sequences of sensor data are recorded.

13. The computer-implemented monitoring method (1) for vibration analysis according to any one of the preceding claims, characterized in that To check the third criterion (III), the test statistic F is used as the null hypothesis, which constitutes the ratio of the variances of the two test sequences, and as the null hypothesis, F=1 is assumed.

14. The computer-implemented monitoring method (1) for vibration analysis according to any one of the preceding claims, characterized in that The warning is transmitted to a signal post, wherein: assigning a different illuminated symbol to each warning level on the signal column, depending on the number of warnings; and / or • Assigning different illuminated symbols on the signal column depending on the deviation from the null hypothesis.

15. Computer-implemented monitoring method (1) for vibration analysis according to any one of the preceding claims, characterized in that During the detection, data from a fixed time period, in particular one that does not contain any patterns, are loaded, wherein the algorithm is subsequently trained so that it does not issue an alarm and / or warning signal within the selected time period, but is parameterized as sensitively as possible, for example by means of a narrow confidence band, i.e. it reacts to correspondingly small changes and / or patterns.

Images