Vacuum plug hole machine fault early warning method based on data feature recognition

Abstract

The present application relates to the technical field of fault early warning, and especially relates to a vacuum plug hole machine fault early warning method based on data feature recognition, which comprises the following steps: obtaining negative pressure time series data and performing baseline removal to obtain a negative pressure fluctuation sequence; calculating a preliminary oscillation degree, a non-periodicity degree and a micro trend factor, and calculating a span optimization factor according to the three factors; calculating an adaptive differential span based on the span optimization factor, and applying a Teager energy operator to the negative pressure fluctuation sequence by using the adaptive differential span to obtain a target energy value; and early warning when the target energy value meets a condition. The present application can adaptively adjust the operator span by fusing multi-dimensional features, suppress mechanical noise, enhance hidden crack signal features, and improve the accuracy of fault early warning.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of fault early warning, and in particular to a vacuum plug hole machine fault early warning method based on data feature recognition. BACKGROUND

[0002] As core equipment in semiconductor packaging and printed circuit board manufacturing processes, vacuum plug hole machines mainly use vacuum chucks to transfer various substrates at high speed between stations. Ultra-thin glass substrates are widely used due to their excellent physical properties. During the adsorption and handling process, if the substrate has a small hidden crack, the negative pressure effect will induce a small amount of gas leakage. Although such leakage does not cause a sharp drop in pressure, it can easily cause the substrate to break under the action of high-speed inertial force, thereby contaminating the vacuum cavity and damaging the precision mesh plate. Therefore, accurate monitoring of the adsorption state is crucial to ensure production safety.

[0003] Existing monitoring methods determine adsorption failure by comparing whether the real-time negative pressure value is lower than the preset warning line. To improve the detection sensitivity of weak leakage signals, the Teager energy operator method is introduced to process the negative pressure fluctuation sequence. This algorithm reflects energy fluctuations by calculating the product of the instantaneous amplitude and frequency of the signal, attempting to capture the transient high-frequency characteristics caused by gas leakage to detect abnormal conditions before the negative pressure value reaches the macro alarm threshold.

[0004] However, when processing negative pressure data accompanied by complex mechanical background noise, the traditional Teager energy operator method has limitations. Specifically, during operation, the vacuum pump pulsation and mechanical arm vibration of the vacuum plug hole machine form background noise with specific frequency patterns. The traditional algorithm usually uses a fixed difference span parameter to calculate energy, lacking the ability to adaptively adjust the frequency components of the signal. When faced with mixed signals, the fixed parameter operator has difficulty distinguishing between regular mechanical noise and high-frequency micro-vibration induced by hidden cracks. If the parameter is not set properly, the algorithm may amplify normal mechanical pulsation as an abnormal signal, causing false positives, or the smoothing effect may mask the weak hidden crack oscillation characteristics, leading to false negatives, making it difficult to effectively warn before the substrate breaks. SUMMARY

[0005] To solve the technical problem that the traditional Teager energy operator method has difficulty distinguishing between high-frequency micro-vibration of hidden cracks and regular mechanical pulsation from complex mechanical noise of the vacuum plug hole machine due to the use of a fixed difference span parameter, the present application provides a vacuum plug hole machine fault early warning method based on data feature recognition, which includes the following steps:

[0006] The negative pressure time sequence data of the vacuum plug hole machine chuck in the adsorption and carrying process is acquired, the negative pressure time sequence data is subjected to baseline removal processing to obtain a negative pressure fluctuation sequence, the instantaneous energy value of each sampling point in the negative pressure fluctuation sequence is calculated, the preliminary shock degree of each sampling point is determined based on the instantaneous energy value, the numerical difference between each sampling point and the corresponding sampling point within a preset delay window is calculated to determine the non-periodicity degree of each sampling point, the numerical drop amplitude and fluctuation degree of each sampling point within a preset long-time observation window are calculated to determine the micro trend factor of each sampling point, the span optimization factor of each sampling point is calculated according to the preliminary shock degree, non-periodicity degree and micro trend factor, the span optimization factor is positively correlated with the preliminary shock degree, non-periodicity degree and micro trend factor, the adaptive differential span of each sampling point is calculated based on the span optimization factor, the Teager energy operator is applied to the negative pressure fluctuation sequence by using the adaptive differential span to obtain the target energy value of each sampling point, and fault early warning is triggered when the target energy value meets a preset alarm condition.

[0007] The present application aims at the problem that the mechanical background noise is complex and the micro crack signal is difficult to extract in the adsorption and carrying process of the vacuum plug hole machine, and proposes a fault early warning method based on data feature recognition, the preliminary shock degree, non-periodicity degree and micro trend factor are respectively obtained from the energy shock, time domain waveform randomness and micro pressure change through in-depth analysis of the negative pressure fluctuation sequence, and these features are fused into the span optimization factor to reflect the crack risk level of the current signal, and then the differential span of the Teager energy operator is adaptively adjusted by using the factor, so that the algorithm can automatically adopt a smaller span to focus on the high-frequency transient characteristics in the area with high crack risk, and a larger span is adopted in the area dominated by mechanical noise to play a smoothing filtering role, thereby enhancing the energy expression of the weak crack signal while effectively suppressing the interference of the periodic mechanical noise, and improving the accuracy and robustness of the system for glass substrate crack fault early warning.

[0008] Preferably, the calculation of the instantaneous energy value of each sampling point in the negative pressure fluctuation sequence and the determination of the preliminary shock degree of each sampling point based on the instantaneous energy value comprises: constructing a short-time neighborhood window with any sampling point in the negative pressure fluctuation sequence as the center; applying the Teager energy operator to each sampling point in the short-time neighborhood window based on a preset fixed span to obtain the instantaneous Teager energy value of each sampling point; calculating the statistical mean of the instantaneous Teager energy values of all sampling points in the short-time neighborhood window, and taking the statistical mean as the preliminary shock degree of the corresponding center sampling point.

[0009] This invention achieves a smooth assessment of the local energy level of a negative pressure fluctuation sequence by constructing a short-time neighborhood window and calculating the statistical mean of the instantaneous Teager energy values ​​of each sampling point within the window. This processing method not only preserves the energy strength information of the signal, but also reduces the accidental influence of single-point random noise on the initial oscillation degree assessment through statistical averaging, providing a robust energy dimension reference for the subsequent calculation of the span optimization factor.

[0010] Preferably, the degree of aperiodicity satisfies the following relationship:

[0011] ;

[0012] in, It is the first The degree of non-periodicity of each sampling point; It is the length of the delay window; , They are the first , Negative pressure fluctuation values ​​at each sampling point; It is the first The local mean of each sampling point within the delay window; It is the preset first minute value.

[0013] This invention utilizes the local mean and the ratio of covariance to variance within a delay window to assess the degree of non-periodicity of a signal. This method can effectively distinguish between highly repetitive mechanical background noise and turbulent signals with random and chaotic characteristics from the time-domain waveform. Under similar energy levels, the algorithm can accurately identify potential hidden crack risk areas based on the regularity differences in the waveforms, reducing misjudgments caused by excessive mechanical vibration energy.

[0014] Preferably, the micro-trend factors satisfy the following relationship:

[0015] ;

[0016] in, It is the first Micro-trend factors at each sampling point; , They are the first , Negative pressure fluctuation values ​​at each sampling point; It is the radius of the long-term observation window, which is the radius of the first long-term observation window. A neighborhood window is constructed centered on each sampling point; It is the first The standard deviation of each sampling point within a long-term observation window; It is a maximum value function; It is a natural exponential function; It is the preset second minute value.

[0017] This invention introduces a long-term observation window and uses local standard deviation to normalize the pressure drop, eliminating the interference of basic fluctuation amplitude on trend judgment. At the same time, combined with the mapping characteristics of the natural exponential function, the micro trend factor is restricted to a specific range. Thus, it can keenly capture the continuous pressure decay trend caused by microcracks that are hidden under background fluctuations when the macro negative pressure value has not fallen below the safety alarm threshold.

[0018] Preferably, the step of calculating the span optimization factor for each sampling point based on the initial oscillation degree, the degree of non-periodicity, and the micro-trend factor includes: adding the micro-trend factor of each sampling point to a preset benchmark constant to obtain a trend correction term; calculating the product of the initial oscillation degree, the degree of non-periodicity, and the trend correction term for each sampling point, and normalizing the product to obtain the span optimization factor for each sampling point.

[0019] This invention achieves deep integration of multi-dimensional features by multiplying and normalizing the trend correction term with the initial oscillation degree and the degree of non-periodicity. This logic ensures that the span optimization factor will only have a large value when the signal has high energy, strong non-periodicity and leakage trend. Thus, the mutual verification of multiple features reduces the risk of misjudgment that may be caused by a single feature.

[0020] Preferably, the adaptive difference span satisfies the following relationship:

[0021] ;

[0022] in, It is the first Adaptive differential span of each sampling point; It is the baseline span; It is the adjustment range coefficient; It is the first Span optimization factor for each sampling point; It is the rounding function; It is a maximum value function.

[0023] This invention establishes a negative correlation mapping relationship between adaptive differential span and span optimization factor, and uses the adjustment amplitude coefficient to constrain the range of change, so that in areas with high risk of hidden cracks, the adaptive differential span automatically converges to a smaller value to improve the ability to capture high-frequency features, while in areas dominated by mechanical noise, the adaptive differential span is maintained at a larger value to exert a low-pass filtering effect, thereby achieving targeted processing of different signal components.

[0024] Preferably, the target energy value of each sampling point is obtained by applying the Teager energy operator to the negative pressure fluctuation sequence using the adaptive differential span, comprising: taking any sampling point as a current sampling point, selecting a previous sampling point and a next sampling point with a span of the adaptive differential span from the current sampling point; calculating the product of the negative pressure fluctuation values of the previous sampling point and the next sampling point, denoted as a span product term; calculating the difference between the square of the negative pressure fluctuation value of the current sampling point and the span product term, and normalizing the difference to obtain the target energy value of the current sampling point.

[0025] Preferably, when the target energy value meets the preset alarm condition, triggering a fault warning, comprising: smoothing the target energy values of the sampling points to obtain a smoothed energy sequence; when there are a continuous preset number of sampling points in the smoothed energy sequence, the values of which all exceed a preset safety alarm threshold, it is determined that the preset alarm condition is met, and a fault warning is triggered.

[0026] Preferably, the preset safety alarm threshold is obtained by: obtaining historical negative pressure data of the vacuum plug machine in a fault-free running state to construct a background noise energy sequence; smoothing the background noise energy sequence to obtain a background risk index sequence; calculating the statistical mean and standard deviation of the background risk index sequence; and the preset safety alarm threshold is the sum of the statistical mean and the standard deviation multiplied by a preset factor.

[0027] Preferably, the baseline removal processing is high-pass filtering or sliding average detrending processing.

[0028] The beneficial effects of the present application are: the present application comprehensively considers the energy oscillation, waveform non-periodicity and microscopic pressure attenuation trend of the negative pressure fluctuation sequence, constructs a span optimization factor through the fusion of multi-dimensional features, overcomes the limitation that a single feature cannot accurately distinguish mechanical noise and hidden crack signals, and improves the comprehensiveness and accuracy of feature recognition. The present application uses the span optimization factor to adaptively adjust the differential span of the Teager energy operator, retains the smoothing filtering effect in the mechanical noise region, improves the extraction ability of the high-frequency transient features in the hidden crack region, and improves the signal-to-noise ratio of the weak fault signal in the strong noise background. The present application combines energy sequence smoothing processing and a dynamic threshold determination mechanism based on statistics to establish an early warning logic containing time duration verification, reduces false alarms caused by incidental interference, and adapts to the complex industrial site environment of the vacuum plug machine. BRIEF DESCRIPTION OF DRAWINGS

[0029] Figure 1 A flowchart of the vacuum plug machine fault early warning method based on data feature recognition provided by the embodiments of the present application is shown in the figure.

[0030] Figure 2The prior art provided for the embodiments of the present application and the corresponding fault early warning effect of the method of the present application are compared in the schematic diagram. DETAILED DESCRIPTION

[0031] The embodiments of the present application provide a vacuum hole plugging machine fault early warning method based on data feature recognition, as shown in the schematic diagram, the method comprises steps S100-S400: Figure 1

[0032] Step S100, acquiring negative pressure time sequence data of a vacuum hole plugging machine suction cup in the suction and carrying process, performing baseline removal processing on the negative pressure time sequence data to obtain a negative pressure fluctuation sequence.

[0033] It should be noted that when the vacuum hole plugging machine utilizes the vacuum suction cup to transfer the ultra-thin glass substrate, the suction process is a dynamic negative pressure establishment and maintenance process. In this process, the slight fluctuation of the gas supply pressure of the gas source, the inherent mechanical pulsation of the vacuum pump and the possible hidden crack leakage of the glass substrate will be superimposed in the negative pressure signal. In order to capture the microsecond-level transient pressure fluctuation caused by the hidden crack, it is necessary to ensure sufficient data sampling frequency and eliminate low-frequency drift interference irrelevant to the core features of the suction state.

[0034] Specifically, a high-sensitivity negative pressure sensor is installed on the suction cup air path of the vacuum hole plugging machine, and the sampling frequency is set to 2kHz. After the suction cup suctions the glass substrate and establishes a steady negative pressure, the original negative pressure data is continuously collected at a preset time interval during the negative pressure maintenance stage before the mechanical arm starts to place the glass substrate. All collected original negative pressure data is arranged in the order of sampling time to obtain an original negative pressure sequence. The original negative pressure sequence is processed by removing the ultra-low frequency drift component caused by the slow fluctuation of the gas supply pressure of the gas source, and only the dynamic fluctuation component reflecting the instantaneous change of the suction state is retained, so as to obtain the negative pressure fluctuation sequence. Preferably, high-pass filtering or sliding average detrending method is adopted, which is the prior art and will not be described here.

[0035] At this point, the negative pressure fluctuation sequence is obtained.

[0036] Step S200, calculating the instantaneous energy value of each sampling point in the negative pressure fluctuation sequence, and determining the preliminary shock degree of each sampling point based on the instantaneous energy value.

[0037] ​It should be noted that during the high-speed adsorption and handling process of the vacuum plugging machine, if there are microcracks in the glass substrate, the turbulence of the airflow at the crack will cause high-frequency transient oscillations in the negative pressure signal. Unlike gas source fluctuations or conventional low-frequency mechanical vibrations, this oscillation contains high instantaneous energy within a very short time. Therefore, this invention is used to evaluate the instantaneous energy performance of the signal. By calculating the intensity of the oscillation at each sampling point, the high-energy region where microcracks may exist can be initially identified, providing a data basis for subsequently distinguishing mechanical noise from actual microcrack signals.

[0038] First, a short-time neighborhood window is constructed. It's important to note that to accurately assess the energy state at each moment and avoid random interference from single-point noise signals, we cannot rely solely on single-point values; we need to combine the signal performance within its neighborhood for statistical smoothing. Therefore, a short-time neighborhood window needs to be defined. The size of this window determines the time range of energy statistics, thus affecting the sensitivity of feature extraction.

[0039] As a preferred implementation, the sampling points are centered on each other and extend forward and backward respectively. 100 sampling points, thus obtaining a total of 100 sampling points. A short-time neighborhood window for each data point, where... It is the radius of the short-term neighborhood window.

[0040] It should be noted that the radius of the short-time neighborhood window can be adjusted according to the actual sampling frequency and the expected signal characteristics: for higher sampling frequencies, such as 5kHz, or scenarios where the expected oscillation duration is longer, the radius of the short-time neighborhood window can be appropriately increased. One sampling point; for low sampling rate or extremely short pulse detection, the short-time neighborhood window radius can be appropriately reduced to [number missing]. Each sampling point. In this embodiment, the window radius... The preferred setting is 10 sampling points, which can not only fully cover the typical microcrack oscillation cycle and ensure the integrity of energy calculation, but also avoid introducing too much irrelevant mechanical operation cycle noise due to an excessively large short-time neighborhood window.

[0041] Then, the initial oscillation degree at each sampling point is calculated. It should be noted that for energy feature extraction, this invention preferably uses the Teager energy operator (TEO) to construct this index. TEO is a nonlinear operator capable of extracting instantaneous energy changes in a signal; its output is approximately equal to the square of the product of the instantaneous amplitude and instantaneous frequency. This invention chooses TEO because it possesses high sensitivity and excellent temporal resolution in vacuum negative pressure detection scenarios. Compared to the root mean square calculation method, which focuses on average energy, TEO is more suitable for capturing high-frequency component changes caused by turbulence. Therefore, even before a weak microcrack signal causes a significant change in macroscopic pressure, anomalies can be identified through the joint frequency-amplitude response.

[0042] Based on the above logic, the first The initial oscillation degree at each sampling point satisfies the following relationship:

[0043] ;

[0044] in, It is the first The initial degree of oscillation at each sampling point; , , It is the first , , Negative pressure fluctuation values ​​at each sampling point; It is the radius of the short-time neighborhood window; It is a standard normalized function; It is the absolute value symbol.

[0045] In this relation This is a standard Teager energy operator calculation term used to calculate the instantaneous energy at a single sampling point. A larger value indicates a higher energy level at the [number]th sampling point. Within a local area of ​​a sampling point, the larger the product of the instantaneous amplitude and frequency of the signal, the stronger its energy. This relationship is expressed through a window. Taking the average within the range can smooth out the interference of single-point noise, thus accurately evaluating the first... The degree of local oscillation at each sampling point.

[0046] At this point, the initial oscillation level of each sampling point has been obtained.

[0047] Step S300: Calculate the numerical difference between each sampling point and its corresponding sampling point within a preset delay window to determine the degree of non-periodicity of each sampling point; calculate the magnitude of numerical decrease and fluctuation of each sampling point within a preset long-term observation window to determine the micro-trend factor of each sampling point; calculate the span optimization factor of each sampling point based on the initial oscillation degree, non-periodicity, and micro-trend factor.

[0048] It should be noted that in the running scene of the vacuum plug machine, in addition to paying attention to the shock energy, the time domain form and micro trend of the signal also need to be fully considered. This is because the periodic pulsation of the vacuum pump and the servo jitter of the mechanical arm as background noise will also produce a higher energy response, but such mechanical noise usually has strong periodicity and regularity, that is, the signal has high autocorrelation; while the airflow turbulence caused by the hidden crack shows obvious randomness and disorder. In addition, the hidden crack is usually accompanied by a small amount of gas leakage, even if the negative pressure value does not drop below the safety alarm threshold in the macroscopic sense, it will still show a continuous pressure decay trend in the microscopic sense. Therefore, the present application is used to evaluate the non-periodicity degree and the micro pressure decay trend of the signal, combined with the preliminary shock degree, to construct a multi-dimensional criterion to distinguish noise from hidden cracks.

[0049] Specifically includes steps S310-S330:

[0050] Step S310, acquire the non-periodicity degree of each sampling point.

[0051] It should be noted that the present application is used to distinguish regular mechanical vibration from chaotic hidden crack turbulence through mathematical means. In order to accurately reflect the local signal confusion degree at the current time, a historical backtracking interval is introduced, and the similarity of the signal after delay comparison is evaluated by using the statistical characteristics in the interval.

[0052] First, taking any sampling point as the current sampling point, a delay window immediately before the current sampling point is constructed, and its length is set to The delay window covers the data range from the th sampling point to the th sampling point, that is, it contains historical data before the current time, and does not contain the th sampling point.

[0053] It should be noted that the value of the length of the delay window should be set according to the frequency characteristics of the mechanical noise in the actual working condition: for scenes with low mechanical vibration frequency, such as large vacuum pump running environment, the value of can be set to a larger value, such as 5 to 8 sampling points, to cover a sufficient phase change range and effectively capture the periodicity of low-frequency noise; for scenes with high mechanical vibration frequency or short waveform, such as high-frequency servo motor jitter, the value of should be appropriately reduced, such as 2 to 3 sampling points, to prevent the delay time from exceeding the correlation time scale of the noise, resulting in failure of periodicity identification. In this embodiment, considering the typical mechanical noise frequency of the vacuum plug machine, the value is preferably set to 5 sampling points, that is, corresponding to 2.5 ms under a sampling rate of 2 kHz.

[0054] Then, the degree of local non-periodicity at each sampling point is calculated. It should be noted that this invention uses mathematical methods to distinguish between regular mechanical vibrations and chaotic hidden crack turbulence. In order to accurately reflect the local signal disorder at the current moment, this index needs to be based on a delay window. The statistical characteristics within the delay window are used to evaluate the similarity of the signals after the delay comparison. If the similarity is low, it indicates that it lacks regularity and the possibility of hidden cracks is higher.

[0055] Based on the above logic, the first The degree of non-periodicity of each sampling point satisfies the following relationship:

[0056] ;

[0057] in, It is the first The degree of non-periodicity of each sampling point; It is the length of the delay window; , They are the first , Negative pressure fluctuation values ​​at each sampling point; It is the first The local mean of each sampling point within the delay window; It is a preset first tiny value used to prevent the denominator from being 0, and can be set to 0.001.

[0058] In this relationship, the molecule The denominator represents the delayed autocorrelation term of adjacent data points relative to the local mean within the delay window. This represents the variance of the signal within the delay window. The ratio of the two values ​​reflects the autocorrelation coefficient of the local signal within a very short time. A larger fraction indicates a stronger correlation between the local signal and the delayed signal, higher waveform repeatability, and a corresponding periodic mechanical noise characteristic. Conversely, a smaller fraction indicates a chaotic signal with little correlation between consecutive time points. Subtracting this fraction from 1 makes... The larger the value, the stronger the non-periodicity of the signal, meaning the higher the credibility of the oscillation being caused by a hidden crack.

[0059] Thus, the degree of non-periodicity at each sampling point was obtained.

[0060] Step S320: Calculate the micro-trend factor for each sampling point.

[0061] It should be noted that in addition to the oscillation pattern, the most essential physical characteristic of the hidden crack is the leakage of gas. Although this leakage is very weak in the early stage and is not enough to trigger the macro threshold alarm, it will statistically show a downward trend in the center of gravity of the local data. In order to capture this microscopic downward trend, the present application introduces a long observation window, which compares the pressure difference between the current time and the future time, and combines the standard deviation of local fluctuations for standardized evaluation, so as to identify the leakage signs hidden in the fluctuations.

[0062] Firstly, a long observation window is constructed for trend identification. It should be noted that micro-leakage is usually a slow accumulation process, and its downward amplitude is often submerged in transient mechanical noise fluctuations. If only the aforementioned short window is used, it is easy to be disturbed by a single mechanical pulse, resulting in the inability to correctly reflect the overall downward trend. Therefore, a long observation window with a wider range needs to be constructed for trend identification.

[0063] Specifically, the radius of the long observation window is set as , and a long observation window with a total length of is constructed based on the current sampling point . When the latest data point is collected, the center point is calculated. At this time, the long observation window covers the data range from the historical time to the future time .

[0064] It should be noted that the radius of the long observation window needs to be greater than the radius of the short neighborhood window , because is concerned about the transient oscillation envelope at the millisecond level, and a shorter window is needed to maintain the sensitivity to mutations; while is concerned about the continuous pressure decay at the tens of milliseconds level, and only a long enough window can smooth out the high-frequency mechanical running cycle noise, so as to highlight the low-frequency leakage trend. It should be further noted that the value of the radius of the long observation window needs to be set according to the actual working condition: for the scene where the expected leakage speed is fast and the pressure drops significantly, the can be set to a smaller value, such as 20 sampling points, to enhance the response speed to sudden leakage; for the scene where the hidden crack is extremely small, the gas leakage is slow and easy to be masked by noise, the needs to be appropriately increased, such as 80 to 100 sampling points, to accumulate the downward trend through a longer time span observation, so as to avoid being misled by short-term fluctuations. In the present embodiment, considering the leakage characteristics of the conventional glass hidden crack, the is preferably 50 sampling points.

[0065] Then, the micro trend factor of each sampling point is calculated. It should be noted that simply calculating the pressure difference at both ends of the long observation window is easily affected by the amplitude of the equipment foundation vibration, and in the case of severe vibration, normal random fluctuations can also produce a larger instantaneous difference, thereby leading to misjudgment. Therefore, the invention introduces the idea of standardization, and normalizes the drop amplitude using the standard deviation in the local window. This logic is used to measure whether the current pressure drop is significantly greater than the conventional fluctuation amplitude of the region. If the drop amplitude is much greater than the standard deviation of the background fluctuation, the confidence that the drop is caused by leakage is higher.

[0066] According to the above logic, the micro trend factor satisfies the relationship:

[0067] ;

[0068] Wherein, is the micro trend factor of the i th sampling point; , , are the negative pressure fluctuation values of the i th and j th sampling points, respectively; is the radius of the long observation window; is the standard deviation of the i th sampling point in the long observation window; is the maximum value function; is the natural exponential function; is a preset second infinitesimal value, used to prevent the denominator from being 0, which can be set to 0.001. In this relationship, the numerator of the i th sampling point measures the pressure drop amplitude in the long observation window, and the denominator normalizes it using the local standard deviation, eliminating the influence of the background fluctuation amplitude, so that the evaluation focuses on the degree of drop relative to the current fluctuation level. The maximum value function ensures that only when the pressure drops will it contribute, and the output of the maximum value function is used as the negative exponential term of the natural exponential function, which uses the monotonic increasing property of the 1-negative natural exponential function when

[0069] to map the micro trend factor to the interval: when the normalized drop amplitude is larger, the independent variable of the natural exponential function is more negative, and the value of the natural exponential function is closer to 0, so that is closer to 1, indicating that the possibility of micro leakage trend in this region is greater; on the contrary, if there is no drop, then is 0.

[0070] Step S330, obtain the span optimization factor of each sampling point.

[0071] ​​​​​It should be noted that a single feature may lead to misjudgment. For example, simple oscillations may be noise, and a simple slight decrease may be gas source fluctuations. Therefore, this invention constructs a span optimization factor by integrating three dimensions: oscillation energy, non-periodicity, and micro-trend. Only when the signal simultaneously possesses high energy, disorder, and a leakage trend is it identified as a high-risk hidden crack signal, thereby guiding the subsequent adjustment of algorithm parameters.

[0072] Based on the above logic, the span optimization factor satisfies the following relationship:

[0073] ;

[0074] in, It is the first Span optimization factor for each sampling point; It is the first The initial degree of oscillation at each sampling point; It is the first The degree of non-periodicity of each sampling point; It is the first Micro-trend factors at each sampling point; It is a standard normalization function, and minimum-maximum normalization is preferred.

[0075] This relationship integrates the characteristics of three dimensions—oscillation intensity, disorder, and leakage trend—through a product. The constant 1 in the formula serves as a fundamental weight, ensuring that even if the micro-pressure decay trend is not significant, the product of the oscillation intensity and the degree of non-periodicity will not be zeroed out. Conversely, when a leakage trend exists, The value increases, It has a gain effect, further amplifying the span optimization factor. This highlights the risk of hidden cracks. In summary, The larger the value, the more likely the current signal is a hidden crack signal; conversely, it tends to be mechanical noise or a normal state.

[0076] Thus, the span optimization factor for each sampling point has been obtained.

[0077] Step S400: Calculate the adaptive differential span of each sampling point based on the span optimization factor, apply the Teager energy operator to the negative pressure fluctuation sequence using the adaptive differential span, and obtain the target energy value of each sampling point; when the target energy value meets the preset alarm condition, trigger a fault warning.

[0078] It should be noted that in the Teager energy operator method, the differential span parameter determines the operator's frequency sensitivity range. A smaller span parameter allows the algorithm to focus on high-frequency micro-oscillations, suitable for capturing transient microcrack signals; a larger span parameter has a smoothing effect, filtering out low-frequency or fixed mechanical jitter. Therefore, by adaptively adjusting this parameter according to the aforementioned optimization factors, accurate extraction of microcrack features can be achieved.

[0079] First, calculate the adaptive difference span for each sampling point. Based on the above logic, the adaptive difference span satisfies the following relationship:

[0080] ;

[0081] in, It is the first Adaptive differential span of each sampling point; It is the baseline span; It is the adjustment range coefficient; It is the first Span optimization factor for each sampling point; It is the rounding function; It is a maximum value function.

[0082] In this relation, As the attenuation coefficient, its value is determined by the span optimization factor. Decision. When The larger the value, the higher the probability that the current signal has hidden crack characteristics. At this time, the attenuation coefficient decreases, making the calculated adaptive differential span more accurate. Converging towards smaller values ​​improves the sensitivity of the Teager energy operator to high-frequency transient characteristics, ensuring that weak signals are captured. Conversely, when... The smaller the value, the more likely the signal is dominated by mechanical noise, and the attenuation coefficient approaches 1, making it more likely to be affected by mechanical noise. Maintaining a span close to the baseline The larger value of the function is used to reduce the sensitivity to background noise by utilizing the smoothing effect of the large span, thus playing a filtering role; the maximum value function ensures that the adaptive difference span is greater than 0.

[0083] It should be noted that the benchmark span and adjustment range coefficient The value needs to be set according to the noise characteristics and detection sensitivity requirements of the actual equipment: for older equipment scenarios with strong mechanical background noise and large vacuum pump pulsation amplitude, the value can be appropriately increased. If set to 7 or 8, and appropriately reduced Setting it to 0.6 enhances the smooth suppression of low-frequency mechanical noise and prevents false alarms; for scenarios where the equipment operates smoothly and the accuracy of detecting minute microcracks is extremely high, it can be appropriately reduced. If set to 3 or 4, and increased appropriately. For example, setting it to 0.9 improves the algorithm's sensitivity to capturing high-frequency transient signals. In this embodiment, considering both versatility and detection performance, the preferred value is... Set to 5, preferred Set it to 0.8.

[0084] Then, using the adaptive differential span, the target energy value for each sampling point is calculated. It should be noted that calculating the adaptive differential span is not the final goal, but rather a means to perform an adaptive secondary energy assessment of the original signal using this parameter. By optimizing the... Substituting the values ​​into the operator, for the hidden crack region, the operator will sensitively capture high-frequency transients with a smaller span; for the mechanical noise region, the operator will smooth the signal with a larger span. This processing method can enhance the energy characteristics of the hidden crack signal while effectively suppressing background noise, thereby generating a fault criterion with a high signal-to-noise ratio.

[0085] Based on the above logic, the target energy value of the sampling point satisfies the following relationship:

[0086] ;

[0087] in, It is the first Target energy value at each sampling point; It is the first Negative pressure fluctuation values ​​at each sampling point; , It is the first indivual, Negative pressure fluctuation values ​​at each sampling point; It is the standard normalized function.

[0088] In this relation, Characterized the first The instantaneous power of the signal at each sampling point; Utilizing adaptive span The autocorrelation of the signal at a specific time interval was calculated, and the difference between the two values ​​reflects the signal's autocorrelation at that time interval. The corresponding nonlinear energy fluctuation at the frequency scale. This invention introduces dynamically changing... It can maximize energy output when microcracks occur and minimize energy output during normal mechanical vibration, thereby achieving feature enhancement.

[0089] Finally, a fault warning is given based on the target energy value. It should be noted that although the target energy value highlights the hidden crack signal, a single-point energy pulse can still be caused by accidental electromagnetic interference or sensor jitter. In order to avoid false positives caused by such transient noise and to evaluate the persistence of the hidden crack risk, the energy sequence needs to be smoothed to convert high-frequency energy fluctuations into an index sequence reflecting the steady-state risk.

[0090] Specifically, first, the target energy value of each sampling point is obtained to construct a Teager energy sequence, and the sequence is smoothed by moving average to generate a final hidden crack risk index sequence. Then, a safety alarm threshold is set If the hidden crack risk index at a certain time during the monitoring process exceeds the threshold and the duration exceeds 20 ms, it is determined that the currently adsorbed glass substrate has a hidden crack risk, and an alarm is triggered in time and the mechanical arm is controlled to stop moving at high speed.

[0091] It should be noted that the value of the safety alarm threshold needs to be set according to the actual working conditions and product characteristics: for a single piece of glass substrate with extremely high value or extremely strong fragility, in order to ensure absolute safety, the value of can be set to a lower value, such as the 99.5% quantile of the risk index of historical normal batch data, to improve the detection sensitivity, and a small amount of false positives is acceptable; for old equipment scenes with extremely fast production rhythm and large mechanical background noise fluctuations, in order to avoid frequent downtime affecting productivity, the value of can be appropriately increased, such as the 99.9% quantile of the risk index of historical normal batch data, to enhance the tolerance to noise. In the present embodiment, the detection safety and production efficiency are considered comprehensively, and the value of is preferably set to the 99.7% quantile of the risk index of historical normal batch data. In one possible implementation, the safety alarm threshold is obtained by: obtaining historical negative pressure data of the vacuum plug hole machine in a fault-free running state to construct a background noise energy sequence; smoothing the background noise energy sequence to obtain a background risk index sequence; calculating the statistical mean and standard deviation of the background risk index sequence; and the preset safety alarm threshold is the sum of the statistical mean and the standard deviation multiplied by a preset multiple.

[0092] Figure 2The figure is a comparative diagram of fault early warning effects of the prior art and the method of the present application. In the interval of 0ms to 100ms, the curve of the prior art breaks through the safe alarm threshold due to the error of energy amplitude, resulting in false alarm, because the periodic mechanical interference cannot be removed. The present application effectively suppresses the background noise by adaptively increasing the differential span parameter and using the smoothing effect of large span, so that the amplitude of the curve corresponding to the present application always maintains below the threshold, thereby avoiding false alarm. In the interval of 120ms to 140ms, the curve of the prior art fails to reach the threshold due to over-smoothing, resulting in missed alarm. In contrast, the method of the present application adaptively reduces the differential span parameter in this interval, and specifically enhances the high-frequency transient energy of the weak hidden crack signal, so that the curve corresponding to the present application clearly rises and clearly exceeds the threshold at this point.

[0093] At this point, the fault early warning of the vacuum plug hole machine is completed.

[0094] The above are preferred embodiments of the present application, and do not limit the protection scope of the present application. Therefore, any equivalent changes made in terms of structure, shape, principle, etc. according to the present application should be covered within the protection scope of the present application.

Claims

1. A vacuum plug hole machine fault early warning method based on data feature recognition, characterized in that, The method comprises the following steps: acquiring negative pressure time sequence data of a vacuum plug hole machine suction chuck in a suction and carrying process, and performing baseline removal processing on the negative pressure time sequence data to obtain a negative pressure fluctuation sequence; calculating instantaneous energy values of each sampling point in the negative pressure fluctuation sequence, and determining a preliminary shock degree of each sampling point based on the instantaneous energy values; The numerical difference between each sampling point and its corresponding sampling point within the preset delay window is calculated to determine the non-periodicity degree of each sampling point, satisfying the relationship: , is the non-periodicity degree of the first sampling point, is the non-periodicity degree of the first sampling point, is the length of the delay window, , are negative pressure fluctuation values of the first and second sampling points, , are negative pressure fluctuation values of the first and second sampling points, is the local mean value of the first sampling point within the delay window, is the local mean value of the first sampling point within the delay window, is a preset first infinitesimal value; Calculate the magnitude of numerical decrease and fluctuation of each sampling point within a preset long-term observation window, determine the micro-trend factor of each sampling point, and satisfy the following relationship: , It is the first Micro-trend factors at each sampling point , They are the first , Negative pressure fluctuation values ​​at each sampling point; It is the radius of the long-term observation window, which is based on the first... A neighborhood window is constructed centered on each sampling point; It is the first The standard deviation of each sampling point within a long-term observation window. It is a maximum value function. It is a natural exponential function. It is the preset second minute value; calculating a span optimization factor of each sampling point according to the preliminary shock degree, aperiodicity degree and a micro trend factor, including: adding the micro trend factor of each sampling point to a preset reference constant to obtain a trend correction term; calculating the product of the preliminary shock degree, the aperiodicity degree and the trend correction term of each sampling point, and normalizing the product to obtain the span optimization factor of each sampling point; The adaptive differential span of each sampling point is calculated based on the span optimization factor, satisfying the following relationship: , It is the first Adaptive differential span of each sampling point It is the baseline span. It is the adjustment range coefficient. It is the first Span optimization factor for each sampling point It is the rounding function; applying a Teager energy operator to the negative pressure fluctuation sequence by using an adaptive differential span to obtain a target energy value of each sampling point; and triggering a fault early warning when the target energy value meets a preset alarm condition.

2. The method of claim 1, wherein, The method of calculating the instantaneous energy values of each sampling point in the negative pressure fluctuation sequence and determining the preliminary shock degree of each sampling point based on the instantaneous energy values comprises the following steps: constructing a short-time neighborhood window with any sampling point in the negative pressure fluctuation sequence as the center; applying a Teager energy operator to each sampling point in the short-time neighborhood window based on a preset fixed span to obtain an instantaneous Teager energy value of each sampling point; calculating the statistical mean of the instantaneous Teager energy values of all sampling points in the short-time neighborhood window, and taking the statistical mean as the preliminary shock degree of the corresponding sampling point.

3. The method of claim 1, wherein, The method of applying a Teager energy operator to the negative pressure fluctuation sequence by using an adaptive differential span to obtain a target energy value of each sampling point comprises the following steps: taking any sampling point as a current sampling point, and selecting a previous sampling point and a next sampling point which are both spaced apart from the current sampling point by an adaptive differential span; calculating the product of the negative pressure fluctuation values of the previous sampling point and the next sampling point, denoted as a span product term; calculating the difference between the square of the negative pressure fluctuation value of the current sampling point and the span product term, and normalizing the difference to obtain the target energy value of the current sampling point.

4. The method for vacuum plug hole machine fault early warning based on data feature recognition according to claim 1, characterized in that, The method of triggering a fault early warning when the target energy value meets a preset alarm condition comprises the following steps: performing smoothing processing on the target energy values of the sampling points to obtain a smoothed energy sequence; when there are a continuous preset number of sampling points in the smoothed energy sequence, the values of which all exceed a preset safety alarm threshold, it is determined that the preset alarm condition is met, and a fault early warning is triggered.

5. The method of claim 4, wherein, The method of obtaining the preset safety alarm threshold comprises the following steps: acquiring historical negative pressure data of the vacuum plug hole machine in a fault-free running state to construct a background noise energy sequence; performing smoothing processing on the background noise energy sequence to obtain a background risk index sequence; calculating the statistical mean and the standard deviation of the background risk index sequence; the preset safety alarm threshold is the sum of the statistical mean and the standard deviation multiplied by a preset multiple.

6. The method for vacuum plug hole machine fault early warning based on data feature recognition according to claim 1, characterized in that, The baseline removal processing is high-pass filtering or sliding average detrending processing.

Images