A putty particle size abnormality early warning system fusing multi-source data

Abstract

The present application relates to the technical field of abnormal early warning, in particular to a putty powder granularity abnormal early warning system fusing multi-source data, which comprises a particle size and shape extraction module, a shape deviation identification module, a data weight calibration module, a multi-dimensional risk scoring module and a trend-driven grading module. The present application combines particle size sequences of multiple stages of discharging, mixing and stirring and packaging, constructs a continuous fluctuation characteristic track based on time series, quantitatively compares the differences of particle size statistical indexes between processes, realizes the phased identification of abnormal change trend of granularity, simultaneously performs difference calculation and correction factor adjustment on the particle size mean deviation section, and quantitatively determines the risk by pairing the particle boundary shape parameters and the granularity variation coefficient, thereby enhancing the multi-dimensional stability of abnormal identification. On this basis, the risk level is driven and graded through the consistency of particle size change trend and scoring result, thereby improving the accuracy of particle risk identification and the timeliness of early warning result.

Description

TECHNICAL FIELD

[0001] The application relates to the technical field of anomaly early warning, and in particular to a putty powder granularity anomaly early warning system fusing multi-source data. BACKGROUND

[0002] The anomaly early warning technical field belongs to the cross-application direction of intelligent monitoring and prediction control, mainly uses methods such as sensor data acquisition, statistical analysis, machine learning modeling, pattern recognition and multi-dimensional feature correlation calculation to realize real-time identification and early warning of abnormal states appearing in production processes, equipment operation, quality detection or environment monitoring. The core technologies involved in this field include multi-source acquisition of abnormal feature data, data noise filtering, trend modeling based on time series analysis, dynamic threshold setting, abnormal pattern matching and risk grading response mechanism. The application range covers industrial production quality control, equipment health management, environment safety monitoring, energy system stability control and other scenes, aiming to reduce the loss caused by abnormalities through intelligent analysis, and realize early intervention and resource optimization.

[0003] Among them, the putty powder granularity anomaly early warning system is a system for quality monitoring and risk management of the putty powder production process, mainly used for real-time detection and early warning of abnormal fluctuations of putty powder granularity in the production link. The system realizes automatic identification of the deviation of granularity from the normal range by collecting granularity data in the production process, combining the set particle size threshold and statistical analysis, and issuing an early warning signal when an abnormality is detected. Its purpose is to improve the quality stability of putty powder production, reduce the product rejection rate, and help enterprises quickly respond to abnormalities in the production process, improve overall production efficiency and product consistency.

[0004] The traditional early warning system lacks correlation analysis between multi-channel data under different processes, which easily causes fluctuation abnormal information to be limited to a local perspective, and it is difficult to accurately identify the abnormal source in the case of inconsistent particle size changes in multiple stages. Moreover, the dynamic evolution trend of particle size distribution is not quantitatively evaluated, which leads to the fact that part of the small trend deviation cannot be responded in real time. For example, although the particle size fluctuation in the mixing and stirring section does not exceed the single-point threshold, the edge aggregation phenomenon in the continuous time period may not be discovered in time, which will cause the quality hidden danger to be not fully revealed, and affect the comprehensive perception and response to the particle abnormality in the production link. SUMMARY

[0005] The purpose of the present application is to solve the problems existing in the prior art, and to provide a putty powder granularity anomaly early warning system fusing multi-source data.

[0006] In order to achieve the above purpose, the present application adopts the following technical scheme: a putty powder granularity anomaly early warning system fusing multi-source data, the system comprises:

[0007] The particle size morphology extraction module obtains putty powder granularity monitoring data, calls particle size data collected in the same time window, analyzes fluctuation interval distribution in the time period, and combines trajectories according to time sequences to generate a particle size statistical feature trajectory set;

[0008] The morphology offset identification module calls a discharge end sensor as a reference channel according to the particle size statistical feature trajectory set, compares skewness difference values and standard deviation difference values in the same time window in the mixing and stirring area and the packaging area sensors, judges offset states of each channel, and generates a channel particle size feature offset marker;

[0009] The data weight calibration module calls particle size sequences at an abnormal marker position in the offset channel according to the channel particle size feature offset marker, analyzes particle size mean deviation of the channel in the corresponding time period, applies a correction factor to sampling points in the offset section for numerical adjustment, and generates a calibrated particle size distribution record;

[0010] The multi-dimensional risk score module obtains a gray-scale image in a corresponding sampling period in a particle image sensor in the mixing and stirring area, calls particle size sequences in a synchronous interval in the calibrated particle size distribution record, calculates a risk score value corresponding to a particle image and a particle size synchronous feature, and generates particle risk section score information.

[0011] The present application improves that the particle size statistical feature trajectory set includes a time sequence distribution map extracted based on a sampling window, a particle size feature combination structure of each monitoring position, and a three-channel particle size morphology time alignment map, the channel particle size feature offset marker specifically is a monitoring channel offset state identification label, a fluctuation abnormality number corresponding to the offset channel, and a morphology index offset record, the calibrated particle size distribution record includes a weighted and corrected particle size sequence, a reconstructed interval value of the offset section, and a particle size comparison structure after channel synchronization, and the particle risk section score information specifically is a particle image feature aggregation score value, a time period particle size variation combination score, and a risk factor weight mapping table.

[0012] The present application improves that the particle size morphology extraction module includes:

[0013] The data acquisition submodule obtains putty powder granularity monitoring data, including particle size distribution sequences collected by granularity sensors at the discharge end, the mixing and stirring area, and the packaging area, calls particle size sampling data of each channel in the same time window, and extracts particle size original point data in each channel in three continuous sampling windows to generate a multi-channel particle size basic data group;

[0014] The morphology calculation submodule respectively calls original point data in a corresponding sampling window of each channel according to the multi-channel particle size basic data group, calculates standard deviation values, skewness values, and range values, takes the three statistics as a morphology parameter set, constructs a particle size morphology time slice sequence in channel order, and generates a particle size window statistical parameter sequence;

[0015] The trajectory construction submodule calls a set of shape parameters arranged in time in each channel according to the particle size window statistical parameter sequence, connects the window shape change path according to the order of the statistical parameter values, and uniformly constructs a three-channel shape alignment structure to generate a particle size statistical feature trajectory set.

[0016] The present application improves that the shape offset identification module comprises:

[0017] The reference setting submodule calls the particle size standard deviation value and the skewness value corresponding to the particle size sensor at the discharge end according to the particle size statistical feature trajectory set, and marks them as reference channel shape data, respectively, and generates a channel comparison shape parameter group by comparing the shape parameters in the same time window in the mixing and stirring area and the packaging area.

[0018] The difference calculation submodule calls the standard deviation value and the skewness value of the mixing and stirring area and the packaging area according to the channel comparison shape parameter group, and obtains the shape offset value of the offset channel in the unit time window by operation, to obtain a shape offset sequence.

[0019] The offset judgment submodule calls the offset value in each channel according to the shape offset sequence, and performs interval judgment with the set offset tolerance threshold, filters the out-of-limit channel, and gives an identification number, to obtain a channel particle size feature offset mark.

[0020] The present application improves that the data weight calibration module comprises:

[0021] The deviation extraction submodule calls the target channel particle size sequence at the abnormal mark position and the particle size sequence of the reference channel in the same time period according to the channel particle size feature offset mark, respectively calculates the particle size mean value of each channel, performs difference calculation on the target channel particle size mean value and the reference channel particle size mean value, and generates a particle size mean deviation information.

[0022] The interval screening submodule calls the sampling point sequence at the deviation corresponding position in the channel particle size sequence according to the particle size mean deviation information, calculates the three standard deviation intervals of each group of sampling points as the confidence range, judges whether the particle size mean deviation falls outside the interval range, screens the sampling point index deviating from the interval, and generates an offset sampling point index set.

[0023] The sampling correction submodule calls the original particle size value at the corresponding position and the reference channel mean value and the target channel mean value obtained in the deviation calculation according to the offset sampling point index set, calculates the ratio of the two as a correction factor, and adjusts the product of the offset sampling point corresponding value to obtain a sampling segment corrected particle size value sequence, and establishes a calibrated particle size distribution record.

[0024] The present application improves that the multi-dimensional risk score module comprises:

[0025] The image extraction submodule obtains the gray scale image corresponding to the sampling period in the particle image sensor of the mixed stirring zone, extracts the pixel boundary points of the particle edge in the image, sequentially constructs the particle boundary set and calculates the profile curvature and edge gray scale gradient of the set, and obtains the particle image boundary feature set;

[0026] The feature pairing submodule obtains the particle size sequence corresponding to the time period in the calibrated particle size distribution record according to the particle image boundary feature set, divides the intervals according to the image sampling section, combines the profile curvature value and the particle size coefficient of variation in each interval, obtains the synchronous feature parameter pair under each interval, and obtains the particle synchronous feature matching information;

[0027] The risk calculation submodule calculates the combined offset degree after mapping according to the paired profile curvature value, particle size coefficient of variation and edge gradient value in each section, obtains the image particle size coupling risk score value of each sampling section, integrates the risk score value of each section, and generates the particle risk section score information.

[0028] The system further comprises:

[0029] The trend-driven grading module judges whether the score value is consistent with the change direction of the particle size slope according to the particle risk section score information and the particle size average slope of the continuous three time periods in the calibrated particle size distribution record, divides the corresponding early warning level label of each time period, and generates the particle risk grading early warning output result;

[0030] The particle risk grading early warning output item comprises a risk level label sequence, a trend score overrun time point and an abnormal trend trigger factor record table.

[0031] The trend-driven grading module comprises:

[0032] The score linkage submodule obtains the time period label of the score and particle size trend consistency according to the score sequence in each time period and the particle size average in the continuous three time periods in the calibrated particle size distribution record according to the time sequence, respectively calculates the score change value and the particle size average slope between adjacent two time periods, and performs symbol comparison on the score change direction and the particle size slope direction, generates the score trend consistency label set;

[0033] The trend construction submodule obtains the time period number marked as consistent, extracts the score change value and the particle size slope amplitude of the time period, performs absolute value calculation on the product of the score change value and the particle size slope according to the section order, constructs the trend factor value sequence, and generates the section risk trend information according to the score trend consistency label set.

[0034] The grade determination sub-module calls a trend factor value according to the section risk trend information, and performs interval comparison with a risk grading threshold value, obtains a corresponding grade label according to the threshold section, establishes a corresponding risk grade number sequence, and generates a particle risk grading early warning output result.

[0035] Compared with the prior art, the application has the advantages and positive effects that:

[0036] In the application, by combining the multiple-stage particle size sequences of discharging, mixing and stirring and packaging, a continuous fluctuation characteristic trajectory is constructed based on a time sequence, the particle size statistical index difference between processes is quantitatively compared, the abnormal change trend of granularity is identified in stages, the difference value of the particle size mean value offset section is calculated and a correction factor is adjusted, the granularity distribution of the fluctuation abnormal section is normalized, the particle boundary shape parameter and the granularity variation coefficient are paired for risk quantification, the multi-dimensional stability of abnormal identification is enhanced, and on this basis, the risk grade is driven by the consistency of the particle size change trend and the score result, so that the early warning grades of different time periods have continuity and directionality, and the accuracy of particle risk identification and the timeliness of early warning results are improved. BRIEF DESCRIPTION OF DRAWINGS

[0037] Figure 1 The system flowchart of the application is shown in the figure;

[0038] Figure 2 The flowchart of the particle size shape extraction module of the application is shown in the figure;

[0039] Figure 3 The flowchart of the shape offset identification module of the application is shown in the figure;

[0040] Figure 4 The flowchart of the data weight calibration module of the application is shown in the figure;

[0041] Figure 5 The flowchart of the multi-dimensional risk scoring module of the application is shown in the figure;

[0042] Figure 6 The flowchart of the trend driven grading module of the application is shown in the figure. DETAILED DESCRIPTION

[0043] In order to make the purpose, technical scheme and advantages of the application clearer, the application will be further described in detail below in combination with the drawings and examples. It should be understood that the specific examples described herein are only used to explain the application, and are not used to limit the application.

[0044] In the description of the present application, it should be understood that the terms "length", "width", "upper", "lower", "front", "rear", "left", "right", "vertical", "horizontal", "top", "bottom", "inner", "outer" and the like indicate the orientation or positional relationship based on the orientation or positional relationship shown in the drawings, and are only for the convenience of describing the present application and simplifying the description, and do not indicate or imply that the device or element referred to must have a particular orientation, be constructed and operated in a particular orientation, and therefore cannot be understood as a limitation of the present application. In addition, in the description of the present application, the meaning of "a plurality of" is two or more, unless otherwise explicitly and specifically limited.

[0045] Referring to Figure 1 The present application provides a technical solution: a putty powder particle size abnormality early warning system fusing multi-source data, the system comprising a particle size morphology extraction module, a morphology deviation identification module, a data weight calibration module, a multi-dimensional risk scoring module and a trend-driven classification module;

[0046] The particle size morphology extraction module acquires putty powder particle size monitoring data, including particle size distribution sequences collected by particle size sensors at the discharge end, the mixing and stirring zone and the packaging zone, calls particle size data collected in the same time window, calculates particle size standard deviation, skewness and range in three consecutive sampling windows respectively, obtains fluctuation interval distribution information in the time period, and combines to form a trajectory according to time sequence to generate a particle size statistical feature trajectory set;

[0047] The morphology deviation identification module calls the sensor at the discharge end as a reference channel according to the particle size statistical feature trajectory set, compares the skewness difference and the standard deviation difference in the same time window in the sensors at the mixing and stirring zone and the packaging zone, calculates the morphology deviation, judges whether the morphology deviation of each channel is out of limit according to the set deviation tolerance threshold, acquires the identification state of the deviation channel, and generates a channel particle size feature deviation marker;

[0048] The data weight calibration module calls the particle size sequence at the abnormal marker position in the deviation channel according to the channel particle size feature deviation marker, performs difference calculation on the particle size mean value in the corresponding time window in the reference channel, acquires the particle size mean deviation of the channel in the corresponding time period, judges whether it coincides with the confidence interval range, applies a correction factor to the sampling points in the deviation section for numerical adjustment if it is out of the interval, the correction factor is the ratio of the reference channel mean value to the target channel mean value, and generates a calibrated particle size distribution record;

[0049] The multi-dimensional risk scoring module acquires a gray scale image in the corresponding sampling period in the particle image sensor at the mixing and stirring zone, calls the particle size sequence in the synchronous interval in the calibrated particle size distribution record, calculates the contour curvature and edge gradient of the particle boundary in the image, and performs numerical pairing with the coefficient of variation of the particle size in the time period, calculates the risk score value corresponding to the particle image and the particle size synchronous feature, and generates particle risk section score information;

[0050] The trend-driven grading module calls the risk score sequence of each section according to the particle risk section score information, judges whether the score value is consistent with the change direction of the particle size slope of the three consecutive time periods in the calibrated particle size distribution record, constructs a trend factor according to the score gradient and the slope amplitude, compares it with the risk grading threshold, divides the corresponding early warning level label of each time period, and generates the particle risk grading early warning output result;

[0051] The particle size average slope represents the increase or decrease rate of the particle size change per unit time;

[0052] The particle size statistical feature trajectory set includes a time series distribution graph extracted based on a sampling window, a particle size feature combination structure of each monitoring position, and a three-channel particle size morphology time alignment graph. The channel particle size feature offset label specifically refers to a monitoring channel offset state identification label, a fluctuation anomaly number corresponding to the offset channel, and a morphology index offset record. The calibrated particle size distribution record includes a weighted corrected particle size sequence, a reconstructed interval value of the offset section, and a channel-synchronized particle size comparison structure. The particle risk section score information specifically refers to a particle image feature aggregation score value, a time period particle size variation combination score, and a risk factor weight mapping table. The particle risk grading early warning output item includes a risk level label sequence, a trend score over-limit time point, and an abnormal trend trigger factor record table.

[0053] Please refer to Figure 2 , the particle size morphology extraction module includes:

[0054] The data acquisition sub-module acquires putty powder granularity monitoring data, including particle size distribution sequences collected by granularity sensors at the discharge end, the mixing and stirring zone, and the packaging zone. It calls particle size sampling data of each channel under the same time window and extracts particle size original point data in each channel within three consecutive sampling windows to generate a multi-channel particle size basic data group.

[0055] The putty powder granularity monitoring data is acquired by deploying online laser particle size analyzers at the discharge end (channel 0), the mixing and stirring zone (channel 1), and the packaging zone (channel 2). Within the production cycle from to , the particle size data of 10 sampling windows ( ) is continuously collected with a sampling interval of . In each window, 5 particle size original point data are collected for each channel, with a unit of microns ( ), forming the original particle size sequence shown in Table 1. In the execution, the particle size original point data of the first to third consecutive sampling windows ( ) of the first channel (discharge end) are called first, specifically window , the window of and the window of , then, the raw point data of the second channel (mixing stirring area) in the same three sampling windows are called, specifically the window of , the window of and the window of , finally, the raw point data of the third channel (wrapping area) in the same three sampling windows are called, specifically the window of , the window of and the window of all the raw point data collected in the three continuous sampling windows of the three channels are combined to generate a multi-channel particle size basic data group.

[0056] Table 1 original particle size monitoring data table

[0057]

[0058] As shown in Table 1, the table lists the original particle size data collected by the three channels in the three continuous sampling windows.

[0059] The shape calculation sub-module calls the raw point data in the sampling window corresponding to each channel according to the multi-channel particle size basic data group, calculates the standard deviation value, skewness value and range value, takes the three statistics as the shape parameter set, constructs the particle size shape time slice sequence in the order of the channel, and generates the particle size window statistical parameter sequence;

[0060] Based on the multi-channel particle size basic data group, first, the raw point data in the sampling window of channel 0 is called , the mean value is calculated, and the standard deviation value is calculated , the skewness value is calculated by calculating the ratio of the third central moment to the cube of the standard deviation , the range value is calculated by calculating the difference between the maximum value and the minimum value , and the range value is calculated , the three statistics are taken as the shape parameter set of channel 0 in the window , then, the raw point data of channel 1 in the sampling window is called , the mean value is calculated, and the standard deviation value is calculated , the skewness value is calculated The range is , constitutes the set of morphological parameters Next, channel 2 is invoked in the sampling window. Original point data within Calculate its mean as And based on this, the standard deviation value is calculated. skewness value The range is , constitutes the set of morphological parameters Following this process, the set of morphological parameters for all channels within all sampling windows is calculated sequentially. Then, according to the channel order and time order of {channel 0 (W1), channel 1 (W1), channel 2 (W1) 3}.{channel 0 (Wz), channel 1 (Wz), channel 2 (W2)}, ..., a sequence of particle size morphology time slices consisting of standard deviation, skewness, and range is constructed, generating a sequence of particle size window statistical parameters.

[0061] The trajectory construction submodule calls the set of morphological parameters arranged by time in each channel according to the sequence of statistical parameters of the particle size window, connects the window morphological change path according to the order of the statistical parameter values, and uniformly constructs a three-channel morphological alignment structure to generate a set of particle size statistical feature trajectories.

[0062] Based on the particle size window statistical parameter sequence, the morphological parameter set arranged by time in channel 0 is first called, specifically the window... set ,window set and windows set The standard deviations of these three sets , , Connecting them in chronological order forms the standard deviation variation path of channel 0, and the skewness values... , , Connecting them in chronological order forms the skewness variation path of channel 0, which will then represent the range. , , Connecting them chronologically, the range variation path for channel 0 is constructed. Subsequently, using the same method, the morphological parameter sets arranged chronologically in channels 1 and 2 are connected to construct the variation paths for each statistical parameter in channels 1 and 2. Finally, the standard deviation variation paths, skewness variation paths, and range variation paths of the three channels are structurally aligned. Specifically, the standard deviation variation path of channel 0... Path of standard deviation variation with channel 1 and the standard deviation variation path of Channel 2 Aligning in the time dimension, and so on, uniform structural alignment is performed on all morphological parameter paths to generate a particle size statistical feature trajectory set.

[0063] Please refer to Figure 3 , the morphological offset identification module comprises:

[0064] The reference setting submodule calls the particle size standard deviation value and the skewness value corresponding to the particle size sensor at the discharge end from the particle size statistical feature trajectory set, respectively marks them as reference channel morphological data, and compares the morphological parameters in the same time window in the mixing and stirring zone and the packaging zone to generate a channel comparison morphological parameter group;

[0065] Based on the particle size statistical feature trajectory set, the particle size standard deviation value trajectory and the skewness value trajectory corresponding to the discharge end (channel 0) are called, and the two trajectories are respectively marked as reference channel morphological data. Then, the morphological parameter set of the mixing and stirring zone (channel 1) in the first time window is compared with the morphological parameter set of the reference channel in the same time window , and the morphological parameter set of the packaging zone (channel 2) in the first time window is compared with the morphological parameter set of the reference channel in the same time window , and the morphological parameter set of the packaging zone (channel 2) in the first time window is compared with the morphological parameter set of the reference channel in the same time window In this way, the morphological parameters of the mixing and stirring zone and the packaging zone in all subsequent time windows are compared with the morphological parameters of the reference channel in the corresponding time windows one by one to generate a channel comparison morphological parameter group.

[0066] The difference calculation submodule calls the standard deviation value and the skewness value of the mixing and stirring zone and the packaging zone according to the channel comparison morphological parameter group, combines the kurtosis of the particle distribution of each channel, and uses the formula:

[0067] ;

[0068] The operation obtains the morphological offset value of the offset channel in the unit time window to obtain a morphological offset sequence.

[0069] Wherein, represents the morphological offset value of the first channel, reflecting the deviation of the channel from the reference channel in the particle size statistical morphology, represents the normalized value of the particle size standard deviation of the first channel, which is obtained by dividing the original standard deviation by the maximum standard deviation value of the channel under standard working conditions, represents the normalized value of the particle size standard deviation of the reference channel, Normalized value representing the length of the sampling time window, obtained by dividing the current sampling window length by the maximum length of the sampling period, Normalized value representing the skewness of the particle size distribution of the first channel, derived from the real-time sampling sequence output by the particle size sensor, calculated by dividing the third central moment in the skewness definition formula by the cube of the standard deviation, and then normalized by dividing by the maximum skewness value in the full cycle of the channel, Normalized value representing the skewness of the particle size distribution of the reference channel, Normalized value representing the kurtosis of the particle size distribution of the first channel, obtained by dividing the fourth central moment of the particle size sequence by the fourth power of the standard deviation to obtain the kurtosis value, and then dividing the kurtosis value by the maximum kurtosis value in the historical operation of the channel, calculated from the fourth central moment data structure, used to represent the sharpness of the particle size distribution curve, the offset tolerance threshold is a static preset value, which is set based on the mean of all values in the shape offset sequence of each channel under stable operating conditions plus three times its standard deviation, and the upper limit obtained is used as the judgment threshold;

[0070] Based on the comparison of the shape parameters of the channels, first, the offset tolerance threshold is set based on the shape offset values of the continuous 100 sampling windows of each channel under standard conditions , forming an offset sequence containing 200 values, for example, the mean of the offset sequence of channel 1 is , the standard deviation is , the mean of the offset sequence of channel 2 is , the standard deviation is , the offset tolerance threshold of channel 1 is set to , and the offset tolerance threshold of channel 2 is set to , this threshold is a static preset value, then the standard deviation value of the mixed mixing zone (channel 1) in window is called, combined with the skewness value of the particle distribution of the channel, and the kurtosis value , normalized processing is performed, wherein the maximum standard deviation value of channel 1 under standard conditions is , the maximum skewness value is , and the maximum kurtosis value is , the normalized standard deviation value is , the normalized skewness value is , and the normalized kurtosis value is , the standard deviation value of the reference channel (channel 0) in window is , the skewness value is , the normalization method is the same as that of channel 1, and the maximum value is set to be the same as that of channel 1, then , Normalized value of sampling time window length The method for obtaining this is to set the duration of the current sampling window. Divide by the set maximum sampling period duration ,get Substitute these values ​​into the formula:

[0071] ;

[0072] This formula is used to quantify the morphological offset value of the offset channel within a unit time window, where... Indicates the first The shape offset value of each channel. Indicates the first Normalized value of the standard deviation of the particle size of each channel. This represents the normalized value of the standard deviation of the reference channel particle size. The normalized value representing the length of the sampling time window. Indicates the first Normalized value of the particle size skewness of each channel. This represents the normalized value of the reference channel particle size skewness. Indicates the first The normalized value of the kurtosis of each channel particle size is calculated using the following logic: first, the normalized difference of the standard deviation is calculated. and divide by the normalized time window length The rate of change of standard deviation is obtained, and then multiplied by a factor determined by kurtosis. A logarithmic amplification factor is determined, which amplifies the effect of the rate of change of the standard deviation when the kurtosis deviates from the norm (i.e., the distribution is sharper or flatter). This result is then squared and the absolute difference between it and the skewness-normalized value. The summation and square root calculation combine the rate of change of standard deviation and the absolute difference in skewness—two different dimensions of morphological deviation—to obtain a comprehensive morphological offset value. The calculation process is as follows:

[0073] ;

[0074] The calculation result That is, channel 1 in the window The morphological offset values ​​are then used to calculate the offset values ​​of all channels across all time windows, resulting in a morphological offset sequence. The advantage of this formula lies in the introduction of kurtosis. As a logarithmic amplification factor, when the particle distribution morphology is abnormally sharp or flat, even if the absolute change in the standard deviation is not large, the effect of its rate of change will be significantly amplified. At the same time, combined with the direct difference in skewness, it has a higher sensitivity to potential process anomalies (such as agglomeration or over-grinding).

[0075] The offset judgment submodule calls the offset value under each channel according to the morphology offset sequence, and interval judgment is performed with the set offset tolerance threshold, the out-of-limit channel is screened, and an identification number is assigned to obtain the channel particle size characteristic offset marker;

[0076] Based on the morphology offset sequence, the offset value of the mixing and stirring zone (channel 1) under the window is called , and interval judgment is performed with the set offset tolerance threshold of channel 1 . Since , the value is in the out-of-limit interval, the channel is screened, and an identification number is assigned, for example, channel 1 is marked as "offset-01" under the window . Subsequently, the offset value of the wrapping zone (channel 2) under the window is called, and the calculated value is , and interval judgment is performed with the offset tolerance threshold of channel 2 . Since , the value is not in the out-of-limit interval, no screening and marking are performed, and according to this process, interval judgment and screening are performed on the offset value of each channel in the morphology offset sequence to obtain the channel particle size characteristic offset marker.

[0077] Referring to Figure 4 , the data weight calibration module comprises:

[0078] The deviation extraction submodule calls the target channel particle size sequence at the abnormal marker position and the particle size sequence of the reference channel in the same time period according to the channel particle size characteristic offset marker, calculates the average particle size of each channel, and performs difference calculation of the average particle size of the target channel and the average particle size of the reference channel to generate particle size average deviation information.

[0079] Based on the channel particle size characteristic offset marker, the position marked as "offset-01", i.e., the target channel particle size sequence of the mixing and stirring zone (channel 1) in the time window is called , and the particle size sequence of the reference channel (channel 0) in the same time period is called . First, the average particle size of the target channel is calculated as , then the average particle size of the reference channel is calculated as , then the difference calculation of the average particle size of the target channel and the average particle size of the reference channel is performed, i.e. , and particle size average deviation information is generated.

[0080] The interval filtering submodule calls the sampling point sequence corresponding to the deviation position in the channel particle size sequence based on the particle size mean deviation information, calculates the three-times standard deviation interval of the channel where each group of sampling points is located as the confidence range, determines whether the particle size mean deviation falls outside the interval range, and filters the sampling point index that deviates from the interval to generate an offset sampling point index set.

[0081] Based on the particle size mean deviation information, channel 1 is called in the window. Particle size sequence With channel 0 in the window Particle size sequence First, calculate the standard deviation of the channel 1 sequence. Then its three-standard-deviation interval is ,Right now Then the standard deviation of the channel 0 sequence was calculated as follows: Then its three-standard-deviation interval is ,Right now Then, determine the deviation of the average particle size obtained in the aforementioned steps. Whether it falls outside these two confidence ranges is determined by judging the mean of the target channel. Is it within the confidence interval of the baseline channel? The mean of the inner and base channels. Is it within the confidence interval of the target channel? Inside, due to exist Within the interval, and Also there Within the interval, the mean deviation does not exceed the confidence range. If it does exceed the range, the main sampling points causing the mean deviation are selected. For example, if a sampling point in channel 1 is... If the mean deviates significantly, the index of the sampling point (such as the first point) is selected to generate an offset sampling point index set.

[0082] The sampling correction submodule calls the original particle size value at the corresponding position and the mean value of the reference channel and the mean value of the target channel obtained in the deviation calculation according to the offset sampling point index set. It calculates the ratio of the two as a correction factor, multiplies it with the corresponding value of the offset sampling point, obtains the particle size value sequence after sampling segment correction, and establishes a calibrated particle size distribution record.

[0083] Based on the offset sampling point index set, the selected offset sampling points are set as channel 1 in the window. The first point has an original particle size value of And call the baseline channel mean obtained in the deviation calculation. With the target channel mean First, calculate the ratio of the two as a correction factor, that is... Then the correction factor is multiplied with the corresponding value of the offset sample point to obtain the corrected value , and the corrected particle size value is replaced with the original value to obtain the sequence of corrected particle size values of the sampling section, for example, the sequence of channel 1 in the window is updated to All corrected sequences are integrated to establish the calibrated particle size distribution record.

[0084] Please refer to Figure 5 The multi-dimensional risk score module comprises:

[0085] The image extraction submodule obtains the gray scale image corresponding to the sampling period in the mixed stirring zone particle image sensor, extracts the pixel boundary points of the particle edges in the image, constructs the particle boundary set in order and calculates the contour curvature and edge gray scale gradient of the set to obtain the particle image boundary feature set;

[0086] Based on the gray scale image corresponding to the sampling period in the mixed stirring zone particle image sensor, specifically a gray scale image of pixels collected in the window period, first, the image is edge detected to extract all the pixel boundary points of the particles in the image, for a typical particle, its boundary point set is , the boundary set of the particle is constructed in clockwise order, then, for the boundary set, the three-point method is used to calculate the contour curvature of each boundary point, for example, for points and its adjacent points and , the geometric angle formed thereby is calculated, and the curvature value is derived to obtain the curvature sequence of the particle, and the pixel gray scale gradient of each boundary point position is calculated, for example, the gray scale value of point is 150, the gray scale values of the pixels around it are 145 and 155 respectively, and the gradient is 10, the edge gray scale gradient sequence of the particle is obtained, and the contour curvature sequence and the edge gray scale gradient sequence of all particles are integrated to obtain the particle image boundary feature set.

[0087] The feature pairing submodule calls the particle size sequence of the corresponding time period in the calibrated particle size distribution record according to the particle image boundary feature set, divides the intervals according to the image sampling section, combines the contour curvature value and the particle size coefficient of variation in each interval to obtain the synchronous feature parameter pair under each interval, and obtains the particle synchronous feature matching information;

[0088] Based on the particle image boundary feature set, the particle size sequence of channel 1 in the window period in the aforementioned calibrated particle size distribution record is called and the image is divided into 5 sampling sections, each section corresponding to a particle size sampling point, for example, the mean value of the curvature of all particle outlines in the first section is The value is combined with the particle size variation coefficient calculated from the first particle size sampling point, that is, the value obtained by dividing the standard deviation by the mean value, to form a synchronous feature parameter pair. According to this process, the numerical combination of the curvature of the outline and the particle size variation coefficient is performed for each section in the image, and the synchronous feature parameter pair under each interval is obtained to obtain the particle synchronous feature matching information.

[0089] The risk calculation sub-module calls the paired curvature of the outline, particle size variation coefficient and edge gradient value in each section according to the particle synchronous feature matching information, calculates the combined offset degree after mapping, and uses the formula:

[0090] ;

[0091] The operation obtains the image particle size coupling risk score value of each sampling section, integrates the risk score values of each section, and generates the particle risk section score information;

[0092] wherein, represents the image particle size coupling risk score value of the first sampling section, which is used to represent the cooperative offset degree between the image boundary and the particle size statistical feature of the section, represents the normalized value of the particle outline curvature of the first section, which is derived from the ratio of the curvature value obtained by the geometric angle constructed by the image boundary three-point method to the maximum curvature value, represents the normalized value of the image edge gray gradient of the first section, which is derived from the ratio of the maximum image gray change rate to the corresponding image gray histogram average gradient value, represents the normalized value of the particle size variation coefficient of the first section, which is calculated by normalizing the particle size standard deviation of the section to [0, 1] divided by the particle size mean value, represents the normalized value of the image distribution uniformity index of the first section, which is defined as the ratio of the information entropy value of the image partition to the maximum information entropy, reflecting the density dispersion degree of the particle distribution in the section image, is the base of natural logarithm;

[0093] Based on particle synchronization feature matching information, the risk grading threshold is set by statistically analyzing the particle risk segment scoring information of a large number of standard production batches. The scoring value sequence is divided according to percentiles. For example, a score value below the 70th percentile is defined as "low risk" (label 1), between the 70th and 95th percentiles is defined as "medium risk" (label 2), and above the 95th percentile is defined as "high risk" (label 3). This establishes the risk grading threshold. Subsequently, the paired contour curvature value in the first segment is used. Particle size variation coefficient Mean of the edge gradient of this section And the image distribution uniformity index (information entropy value) of this segment. Normalization is performed; for example, the maximum value of all curvatures is... The maximum gradient is The maximum coefficient of variation is The maximum information entropy is The normalized value is , , , Substitute these values ​​into the formula:

[0094] ;

[0095] This formula is used to quantify the image grain size coupling risk score for each sampling segment, where, Indicates the first Risk score of each sampling segment, Indicates the first Normalized values ​​of particle profile curvature in each segment Indicates the first The normalized value of the gray-level gradient at the edge of each image segment. Indicates the first Normalized values ​​of the coefficient of variation of particle size in each segment Indicates the first The normalized value of the image distribution uniformity index for each segment. Given the base of the natural logarithm, the formula's operational logic is to use the contour curvature, which characterizes the irregularity of particle morphology, as the base. Edge gradients characterizing the clarity of particle surface states Multiply by, then multiply by a factor determined by the uniformity of the image distribution. The determined exponential amplification factor increases with increasing particle size distribution to amplify the risk score. Finally, this result is divided by a factor proportional to the particle size variation coefficient. The related denominator, which is greater when the particle size distribution is more uneven, plays a role in adjusting the risk score. This operation combines the microscopic morphology, macroscopic distribution statistics, and image clarity of the particles to obtain a coupled risk score. The example calculation process is as follows:

[0096] ;

[0097] The calculation result is the image particle size coupled risk score value of the first sampling section. The risk score values of all sections are calculated according to this process, and the integrated particle risk section score information is generated. The advantage of the formula is that the introduction of the exponential term makes the uniformity of the particle distribution in the image a sensitive amplification factor. It can nonlinearly amplify the local aggregation or sparseness problem caused by uneven mixing and stirring, thereby having higher risk warning capability than analyzing particle size or morphology alone.

[0098] Please refer to Figure 6 , the trend-driven classification module includes:

[0099] The score linkage submodule calls the score sequence in each time period and the particle size average in the continuous three periods in the calibrated particle size distribution record according to the particle risk section score information, calculates the score change value and the particle size average slope between adjacent two periods in time order, and performs symbol comparison on the score change direction and the particle size slope direction to obtain the time period label with consistent score and particle size trends, and generates the score trend consistency label set.

[0100] Based on the particle risk section score information, the score sequence in the window time period is called, which is set to , and the particle size average in the window three periods in the calibrated particle size distribution record is called, which is set to . First, the score change value between adjacent two periods is calculated according to time order, i.e. and . Then, the particle size average slope between adjacent two periods is calculated, i.e. and . Then, for the first period, the score change direction is positive and the particle size slope direction is positive, and the symbols are the same, so the time period is marked for consistency. For the second period, the score change direction is negative and the particle size slope direction is negative, and the symbols are the same, so the time period is also marked for consistency. The time period label with consistent score and particle size trends is obtained, and the score trend consistency label set is generated.

[0101] The trend building submodule extracts the score change value and the particle size slope amplitude of the time period according to the score trend consistency marking set, performs absolute value calculation of the product of the score change value and the particle size slope in the order of the section, builds a trend factor value sequence, and generates section risk trend information;

[0102] Based on the score trend consistency marking set, the first time period marked as consistent is called, and its score change value and the particle size slope amplitude are extracted , the absolute value calculation of the product of the score change value and the particle size slope is performed in the order of the section, and the trend factor value is obtained , then the second marked time period is called, and its score change value and the particle size slope amplitude are extracted , and the trend factor value is calculated as , all the calculated trend factor values are arranged in time order to build a trend factor value sequence, and section risk trend information is generated.

[0103] The grade determination submodule calls the trend factor value according to the section risk trend information and compares it with the risk grading threshold value, obtains the corresponding grade label according to the threshold segment where it is located, establishes a corresponding risk grade number sequence, and generates a particle risk grading early warning output result;

[0104] Based on the section risk trend information, the first trend factor value is called, and interval comparison is performed with the aforementioned set risk grading threshold value, the risk grading threshold is set as: low risk , medium risk , and high risk , since is located in the interval , the corresponding grade label “medium risk” (number 2) is obtained, then the second trend factor value is called, and interval comparison is performed with the threshold value, since is located in the interval , the grade label “low risk” (number 1) is obtained, all the obtained grade labels are arranged in order to establish a corresponding risk grade number sequence, and a particle risk grading early warning output result is generated.

[0105] The above is only a preferred embodiment of the present application, and does not limit the form of the present application, any skilled person in the art can modify or change the above disclosed technical content to equivalent embodiments applied to other fields, but any simple modification, equivalent change and modification made according to the technical essence of the present application to the above embodiments without departing from the technical solution content of the present application still belongs to the protection scope of the present application.

Claims

1. A putty powder particle size anomaly early warning system integrating multi-source data, characterized in that, The system includes: The particle size extraction module acquires putty powder particle size monitoring data, calls particle size data collected within the same time window, analyzes the distribution of fluctuation intervals within the time period, and combines them according to the time series to form a trajectory, generating a set of particle size statistical feature trajectories. The particle size and morphology extraction module includes: The data acquisition submodule acquires putty powder particle size monitoring data, including particle size distribution sequences collected by particle size sensors at the discharge end, mixing zone, and packaging zone. It calls the particle size sampling data of each channel under the same time window and extracts the original particle size point data within three consecutive sampling windows in each channel to generate a multi-channel particle size basic data set. The morphology calculation submodule calls the original point data in the sampling window corresponding to each channel according to the multi-channel particle size basic data group, calculates the standard deviation, skewness and range, uses the three statistics as a set of morphological parameters, constructs a particle size morphology time slice sequence according to the channel order, and generates a particle size window statistical parameter sequence. The trajectory construction submodule calls the set of morphological parameters arranged in time in each channel according to the sequence of statistical parameters of the particle size window, connects the window morphological change path according to the order of statistical parameter values, and uniformly constructs a three-channel morphological alignment structure to generate a set of particle size statistical feature trajectories. The morphological offset recognition module uses the discharge end sensor as the reference channel based on the particle size statistical feature trajectory set. It compares the skewness difference and standard deviation difference in the same time window between the sensors in the mixing zone and the packaging zone to determine the offset state of each channel and generate a channel particle size feature offset mark. The data weight calibration module calls the particle size sequence at the abnormal marker position in the offset channel according to the channel particle size feature offset marker, analyzes the mean particle size deviation of the channel in the corresponding time period, applies a correction factor to the sampling point of the offset segment to adjust the value, and generates a calibrated particle size distribution record. The multidimensional risk scoring module acquires the grayscale image of the particle image sensor in the mixing zone corresponding to the sampling time period, calls the particle size sequence of the synchronous interval in the calibrated particle size distribution record, calculates the risk score value corresponding to the particle image and particle size synchronization feature, and generates particle risk segment score information.

2. The putty powder particle size anomaly early warning system integrating multi-source data according to claim 1, characterized in that, The particle size statistical feature trajectory set includes a time series distribution map extracted based on the sampling window, a particle size feature combination structure at each monitoring location, and a three-channel particle size morphology time alignment map. The channel particle size feature offset markers specifically include a monitoring channel offset status identification label, a fluctuation anomaly number corresponding to the offset channel, and a morphological index offset record. The calibrated particle size distribution record includes a weighted and corrected particle size sequence, a reconstructed interval value of the offset segment, and a particle size comparison structure after channel synchronization. The particle risk segment scoring information specifically includes a particle image feature aggregation score value, a time period particle size variation combination score, and a risk factor weight mapping table.

3. The putty powder particle size anomaly early warning system integrating multi-source data according to claim 1, characterized in that, The morphological offset recognition module includes: The baseline setting submodule calls the standard deviation and skewness values ​​of the particle size corresponding to the particle size sensor at the discharge end according to the particle size statistical feature trajectory set, and marks them as baseline channel morphology data respectively. It also compares the morphology parameters in the mixing and stirring zone and the packaging zone channel under the same time window to generate a channel comparison morphology parameter group. The difference calculation submodule uses the standard deviation and skewness values ​​of the mixing zone and the packaging zone to calculate and obtain the morphological offset value of the offset channel within a unit time window, based on the channel comparison morphological parameter group, and obtains the morphological offset sequence. The offset judgment submodule calls the offset value of each channel according to the morphological offset sequence, and performs interval judgment with the set offset tolerance threshold to filter out channels that exceed the limit, assign an identification number, and obtain the channel particle size feature offset mark.

4. The putty powder particle size anomaly early warning system integrating multi-source data according to claim 3, characterized in that, The data weight calibration module includes: The deviation extraction submodule, based on the channel particle size feature offset marker, calls the particle size sequence of the target channel and the particle size sequence of the reference channel within the same time period at the abnormal marker position, calculates the mean particle size of each channel respectively, performs the difference calculation between the mean particle size of the target channel and the mean particle size of the reference channel, and generates particle size mean deviation information. The interval filtering submodule calls the sampling point sequence corresponding to the deviation position in the channel particle size sequence based on the particle size mean deviation information, calculates the three-times standard deviation interval of the channel where each group of sampling points is located as the confidence range, determines whether the particle size mean deviation falls outside the interval range, and filters the sampling point index that deviates from the interval to generate an offset sampling point index set. The sampling correction submodule, based on the offset sampling point index set, calls the original particle size value at the corresponding position and the mean value of the reference channel and the mean value of the target channel obtained in the deviation calculation, calculates the ratio of the two as a correction factor, multiplies it with the corresponding value of the offset sampling point, obtains the particle size value sequence after sampling segment correction, and establishes a calibrated particle size distribution record.

5. The putty powder particle size anomaly early warning system integrating multi-source data according to claim 4, characterized in that, The multidimensional risk scoring module includes: The image extraction submodule acquires the grayscale image of the particle image sensor in the mixing zone during the corresponding sampling period, extracts the pixel boundary points of the particle edges in the image, constructs the particle boundary set in sequence, and calculates the contour curvature and edge grayscale gradient of the set to obtain the particle image boundary feature set. The feature matching submodule calls the particle size sequence of the corresponding time period in the calibrated particle size distribution record according to the particle image boundary feature set, divides the interval according to the image sampling segment, and combines the contour curvature value and particle size variation coefficient in each interval to obtain the synchronous feature parameter pair under each interval, thus obtaining particle synchronous feature matching information. The risk calculation submodule, based on the particle synchronization feature matching information, calls the paired contour curvature value, particle size variation coefficient and edge gradient value in each segment, calculates the degree of combined offset after mapping, calculates the image particle size coupling risk score value of each sampling segment, integrates the risk score values ​​of each segment, and generates particle risk segment score information.

6. The putty powder particle size anomaly early warning system integrating multi-source data according to claim 5, characterized in that, The system also includes: The trend-driven classification module, based on the particle risk segment scoring information, calls the risk score sequence of each segment and the average slope of particle size in three consecutive time periods in the calibrated particle size distribution record, determines whether the score value is consistent with the direction of change of the particle size slope, divides the warning level label corresponding to each time period, and generates particle risk classification warning output results. The particle risk classification and early warning output items include a risk level label sequence, trend score exceeding the limit time point, and an abnormal trend triggering factor record table.

7. The putty powder particle size anomaly early warning system integrating multi-source data according to claim 6, characterized in that, The trend-driven hierarchical module includes: The scoring linkage submodule calls the scoring sequence under each time period and the average particle size in three consecutive time periods in the calibrated particle size distribution record according to the particle risk segment scoring information. It calculates the score change value and the slope of the average particle size between two adjacent time periods in chronological order, and compares the sign of the score change direction with the particle size slope direction to obtain the time period label that is consistent with the score and particle size trend, and generates a scoring trend consistency tag set. The trend construction submodule calls the time period number marked as consistent according to the scoring trend consistency mark set, and extracts the score change value and particle size slope amplitude of the time period. It performs the absolute value calculation of the product of the score change value and particle size slope in the segment order to construct the trend factor value sequence and generate segment risk trend information. The risk level determination submodule calls the trend factor value and compares it with the risk level threshold value according to the risk trend information of the segment. It obtains the corresponding level label according to the threshold segment, establishes the corresponding risk level number sequence, and generates the granular risk level warning output result.

Images