Data quality

Abstract

A method of automatically identifying data quality issues in background data for laser diffraction particle characterization is provided. The method includes receiving background data corresponding to a light intensity measured by each of a plurality of detectors in a laser diffraction based particle size analysis system. A processor is used to automatically determine whether the background data contains at least one artifact indicative of an error source in the particle size analysis system. If it is found that the background data contains at least one artifact, an indication is provided that the background data contains at least one artifact. Further, the at least one artifact is classified, and the result of the classification is reported.

Description

Technical Field

[0001] The present invention relates to particle size analysis based on laser diffraction. Background Art

[0002] Laser diffraction is a well-established technique for characterizing the size of particles in a sample. Particles in a sample can be illuminated using a light source (a laser) and the resulting diffraction pattern captured at one or more detectors. The diffraction pattern depends on the size of the particles, so the size of the particles in the sample as well as the distribution of particle sizes can be determined.

[0003] Particle size analysis is an important technique used in a variety of industries and processes. For example, in the food and pharmaceutical industries, the technique can be used to analyze suspensions and emulsions by evaluating how well the particles are dispersed in the liquid medium and whether there is excessive flocculation or agglomeration. Many manufacturing processes, such as 3D printing, use fine powders. Particle size analysis can be used on "dry" media to ensure that the particle size allows for efficient powder flow.

[0004] Equipment configured to perform laser diffraction based particle size analysis is commercially available. One such equipment is the Mastersizer 3000 from Malvern Panalytical.

[0005] However, obtaining accurate and reliable data with any such device can be challenging. When making measurements, an accurate and stable background measurement (i.e., data collected by the detector when no particle sample is present) is required. A poor background measurement can result in inaccurate particle size measurements, or no particle size measurements can be generated at all due to insufficient signal-to-noise ratio.

[0006] Currently, most data quality checks are performed after particle size measurements. This requires users to analyze and interpret data quality issues, which can be time consuming and require additional training. Even for trained users, it can be difficult to identify background data issues in the measurement data.

[0007] Furthermore, checking the quality of measurement data may require multiple measurements, especially when it is difficult for the user to determine the cause of data quality issues. This can be time consuming, result in a waste of sample and dispersant, and lead to inefficient use of resources, energy, and manpower.

[0008] An effective method is needed to identify various problems and artifacts in background data for laser diffraction-based particle size analysis. Summary of the invention

[0009] According to a first aspect, there is provided a method for automatically identifying data quality issues in background data for laser diffraction particle characterization, the method comprising:

[0010] receiving background data corresponding to light intensity measured by each of a plurality of detectors in a laser diffraction-based particle size analysis system;

[0011] automatically determining, using a processor, whether the background data includes at least one artifact indicative of a source of error in the particle size analysis system;

[0012] In the event that it is determined that the background data does include at least one artifact, an indication is provided, using the processor, that the background data includes at least one artifact.

[0013] The method may include classifying the at least one artifact in the background data. The classification may identify one or more classes of artifacts present in the background data (e.g., misalignment, sample pool contamination, dispersant contamination, thermal instability). The method may include reporting artifact classification results to a user. Reporting artifact classification results may include suggesting corrective actions to the user. The method may employ a plurality of different algorithms, each algorithm detecting a particular class of artifacts. Artifact classification may be determined at least in part by the algorithm that detected the artifact.

[0014] The method may further include:

[0015] generating background data by: irradiating a sample cell including a control sample with a light beam from a light source to generate scattered light by interaction of the light beam with the sample cell and / or the control sample, and detecting the scattered light with the plurality of detectors;

[0016] Wherein the control sample consists nominally entirely of the dispersant, and wherein the plurality of detectors are positioned at a plurality of different angles relative to a direction of propagation of the light beam.

[0017] Automatically determining whether the background data contains at least one artifact includes applying at least one of the following algorithms to the background data:

[0018] a dynamic algorithm configured to identify transient fluctuations within the background data; and

[0019] A static algorithm is configured to identify static artifacts in the background data.

[0020] The dynamic algorithm may include at least one of the following algorithms:

[0021] (i) a peak detection algorithm configured to identify peaks within the time series background data and calculate a percentage of data in the time series background data that contains peaks with intensities greater than a threshold;

[0022] (ii) an adaptive diffraction algorithm configured to identify transient events in background data and calculate characteristics of the transient events; and

[0023] (iii) a standard deviation algorithm, wherein the maximum standard deviation of the time series background data for each detector is calculated; and

[0024] The dynamic algorithm is configured to identify the time series background data as containing dynamic fluctuations if at least one of the following conditions occurs:

[0025] a) the calculated percentage is above the threshold;

[0026] b) the calculated characteristic of the transient event is above a threshold; and

[0027] c) The calculated maximum standard deviation of the time series background data is higher than the threshold.

[0028] The static algorithm may include at least one of the following algorithms:

[0029] (i) a hump algorithm configured to identify humps within static background data; and

[0030] (ii) a spike algorithm configured to identify excessive differences between static background data measured between at least one pair of detectors; and

[0031] The static algorithm is configured to identify the static background data as containing static artifacts if a hump is identified or an excessively large difference between detectors is identified.

[0032] The static algorithm may include:

[0033] (i) a numerical algorithm configured to compare static background data measured by at least two of the detectors and identify whether a first detector of the two or more detectors is greater than a second detector of the two or more detectors by a threshold value;

[0034] (ii) A curve fitting algorithm configured to fit an exponential decay function and a Gaussian function to the static background data.

[0035] The hump algorithm may include a machine learning algorithm, including a machine learning model trained to identify humps in static background data.

[0036] The machine learning model may include a convolutional neural network. The convolutional neural network may be a deep convolutional neural network having at least three, four or five convolutional layers.

[0037] The step of automatically determining whether the background data contains at least one artifact may be performed prior to performing a particle analysis measurement on the particle sample, and the particle analysis measurement is aborted if it is determined that the background data contains at least one artifact.

[0038] The particle analysis measurement may include: irradiating a sample with a light beam from a light source, thereby generating scattered light through the interaction of the light beam with particles of the sample; detecting raw measurement data, which includes the distribution of scattered light intensity at a series of different scattering angles; and determining particle characteristics based on the raw measurement data using a processor.

[0039] The method may further include:

[0040] Generating corrected measurement data by subtracting background data from raw measurement data; and performing a negative value data check, wherein the corrected measurement data is identified as faulty if the number of detectors with negative corrected measurement values ​​is above a threshold or the maximum negative value among all corrected measurement data is above a threshold.

[0041] The step of automatically determining whether the background data contains at least one artifact may be performed at least once as the background data is received. For example, the background data may be checked for artifacts once per second as the background data is received. A live display may be provided to the user, which is updated to flag any problems in the background data as the background data is checked (before any sample is consumed by the measurement).

[0042] The method may also include automatically determining a corrective action in response to the determination of the artifact type, and displaying an indication of the corrective action to a user.

[0043] According to a second aspect, there is provided a non-volatile machine-readable medium comprising instructions for configuring a processor to perform a method comprising:

[0044] receiving background data corresponding to light intensity measured by each of a plurality of detectors in a laser diffraction-based particle size analysis system;

[0045] automatically determining, using a processor, whether the background data includes at least one artifact indicative of a source of error in the particle size analysis system;

[0046] In the event that it is determined that the background data does include at least one artifact, an indication is provided, using the processor, that the background data includes at least one artifact.

[0047] The method may be according to the first aspect, including any optional features thereof.

[0048] According to a third aspect, a laser diffractometer is provided, comprising:

[0049] Sample pool;

[0050] a light source configured to illuminate the sample cell with a light beam, thereby generating scattered light through interaction of the light beam with particles in the sample cell;

[0051] a plurality of light detectors configured to detect scattered light intensity distributions at a range of different scattering angles;

[0052] a processor configured to determine a particle characteristic based on a distribution of scattered light intensity at a range of different scattering angles;

[0053] The processor is further configured to:

[0054] receiving background data corresponding to the light intensity measured by each of the plurality of detectors;

[0055] automatically determining whether the background data includes at least one artifact indicative of a source of error in the particle size analysis system;

[0056] In the event that it is determined that the background data does include at least one artifact, an indication is provided that the background data includes at least one artifact.

[0057] The processor may be configured to perform the method of the first aspect, including any optional features thereof. BRIEF DESCRIPTION OF THE DRAWINGS

[0058] Embodiments of the present invention will be described by way of example with reference to the accompanying drawings, in which:

[0059] Figure 1 shows a schematic diagram of a particle analyzer based on laser diffraction;

[0060] Figure 2 Examples of particle size distribution data and background measurements are shown;

[0061] Figure 3 An example of background data without any artifacts is shown;

[0062] Figure 4 An example of background data with artifacts is shown;

[0063] Figure 5 An example of background data with artifacts is shown;

[0064] Figure 6 An example of background data with artifacts is shown;

[0065] Figure 7 An example of background data with artifacts is shown;

[0066] Figure 8 A block diagram of a laser diffraction-based particle analysis system is shown. The particle analysis system includes a Figure 1 The analyzer shown;

[0067] Fig. 9shows a flow chart illustrating a method for determining whether background data contains any artifacts according to the present invention;

[0068] Fig.10 shows an example of results generated according to the present invention;

[0069] Fig.11 shows an example of results generated according to the present invention;

[0070] Fig.12 shows an example of results generated according to the present invention;

[0071] Fig.13 shows an example of results generated according to the present invention;

[0072] Fig.14 shows an example of results generated according to the present invention;

[0073] Fig.15 shows an example of results generated according to the present invention;

[0074] Fig.16 An example of background data with artifacts is shown;

[0075] Fig.17 shows an example of results generated according to the present invention;

[0076] Fig.18 An example of results generated according to the present invention is shown. DETAILED DESCRIPTION

[0077] See also Figure 1 , a schematic diagram of a particle size analyzer 1 is shown. The particle analyzer 1 uses laser diffraction to measure the particle size and particle size distribution (PSD) in a sample material. For understanding the present invention, it is not necessary to consider the detailed operation of the laser diffraction-based particle size analysis, but the general principles are described below.

[0078] A sample cell 10 containing sample material is provided within the analyzer 1. The sample cell 10 may include a flow cell, a cuvette, or a surface on which a droplet of sample is received. The sample includes particles 11 suspended in a dispersant. The analysis may be performed using a wet dispersion (i.e., a liquid dispersant such as water, isopropanol, or oil) or a dry dispersion (e.g., an air or dry nitrogen dispersant).

[0079] An incident light beam 30 from a light source 4 is directed onto a sample cell 10, and diffracted / scattered light 31 from particles 11 is measured using a plurality of detectors positioned at different angles around the measurement cell 10 (relative to the direction of propagation of the incident laser beam 30). The plurality of detectors may include focal plane detectors 20 that detect light diffracted by particles 11 at small angles, side scatter detectors 21 that detect light diffracted / scattered by particles 11 at large angles, and backscatter detectors 22 that detect light backscattered by particles 11 in a direction generally opposite to the direction of propagation of the incident laser beam 30.

[0080] The intensity of scattered light detected at each of the plurality of detectors is provided to a processor (see Figure 8 The processor is configured to analyze the distribution of scattered light intensity to measure sample characteristics, such as particle size and / or particle size distribution. Generally, small particles will diffract light at larger angles (more isotropically) than large particles. Therefore, by analyzing the scattered light intensity received at each of a plurality of detectors located at different angles around the measuring cell 10, the particle size and particle size distribution can be determined.

[0081] Figure 1 The particle analyzer 1 shown is a simplified version of a real system. The system may include two light sources, such as a red (e.g., 633 nm) laser and a blue (e.g., 470 nm) LED. Using two light sources may be advantageous because a wider range of particle sizes can be analyzed. The shorter wavelength of the blue light source enables small particles (e.g., submicron particles) to produce a higher scattering intensity, which can allow smaller particles to be detected and analyzed compared to using a single wavelength light source. The system may include mirrors and beam splitters so that both lasers can be emitted along the same wavelength as the sample. Figure 1 The light beam 30 is shown being directed along the same incident path.

[0082] Each of the detectors 20, 21, 22 may include multiple detector elements or be otherwise divided into sub-areas. Figure 1 The focal plane detector 20 shown in FIG. 5 may include multiple detectors, each positioned at a different angle relative to the illumination beam propagation direction. The angular displacement between each detector element may be small enough to enable accurate characterization of particle size.

[0083] See also Figure 2 , showing the use of something like Figure 1Example data 100 of a typical particle size analysis experiment performed on the system shown. Data 100 shows a graph of light energy or light intensity (expressed in nominal units) versus detector number. Higher detector numbers are associated with larger scattering angles. As discussed above, each detector is positioned at a different angle relative to the direction of beam propagation. For example, detector number 1 may be offset by 1° relative to the direction of beam propagation, detector number 2 may be offset by 2°, and so on. Knowing how each detector is distributed (which is a predetermined characteristic of the system), it means that Figure 2 The graph shown is similar to a graph of energy level versus diffraction angle (except that the x-axis is not necessarily linear because the increment in scattering angle between consecutive detector numbers is not necessarily fixed).

[0084] Figure 2 The data 100 of FIG. 1 shows a background data graph 101. The background data graph shows the background measurement of the system when there are no particles. The corrected data 102 can be calculated by obtaining the raw data (not shown) when measuring the sample and subtracting the background data 101.

[0085] The corrected data 102 shows that the light energy presents a relatively smooth "hump" distribution near detector 30, indicating the particle size distribution (the particle size can be determined based on the corresponding detector angle and the wavelength of the illumination beam). A small secondary peak can be seen at about detector 53. This is due to the system switching to a blue light source, which increases the diffracted light energy generated by small particles with larger diffraction angles.

[0086] See also Figure 3 , shows an example of "good" background data 101. The background data 101 is shown this time as a bar graph. Since the corrected data from which the particle size characteristics are inferred depends largely on the background data 101, it is very important that the background data 101 does not contain any artifacts or contamination. The background data can be acquired at a sampling rate of 5kHz or 10kHz, depending on whether a wet or dry sample is used.

[0087] Background data 101 can be considered "good" data if it meets the following criteria: the energy level of detector No. 1 is less than 100; the energy level of detector No. 20 is less than 20; the gradient amplitude tends to decrease as the detector number increases; and the energy level fluctuations between detectors are limited.

[0088] Background data that meets these criteria indicates that: the windows and dispersant of the measurement cell are clean and free of contaminants; the system has good optical alignment; and the dispersant is stable (e.g., the dispersant does not contain a large number of bubbles). The following discusses various types of artifacts that may appear in the background data 101, what they may correspond to, and how to automatically identify them. Artifacts can be classified as being caused by the following factors: measurement cell contamination, optical misalignment, thermal instability (causing a non-uniform refractive index of the dispersant), and dispersant contamination (caused by particles or bubbles).

[0089] Ensuring that background data 101 is free of artifacts before taking any measurements can help improve the signal-to-noise ratio of the sample data, the accuracy of the particle size characterization, and the stability and repeatability of any measurements. It can also save time and reduce sample waste because the measurement can be aborted (e.g., automatically aborted) if the background contains artifacts (e.g., before adding dispersant to the sample).

[0090] See also Figure 4 , shows an example of background data 101 where there is motion artifact within the pool. This can be seen in the background data as a "hump" or intermittent peak within region 102, rather than as Figure 3 This is due to light scattering caused by contamination within the dispersant (e.g., residue from previous samples left in the dispersant) or bubbles within the dispersant (which can be removed by, for example, degassing). Since dispersant contaminants and bubbles move within the cell, they cause artifacts in the background data 101 that appear to fluctuate over time. These dynamic fluctuations can be detected by detecting outlier peaks in the time series of background measurements taken over a given time interval. A "hump" caused by a transient contamination may be larger than the averaged background measurement below. Figure 5 is smaller in , because the data after averaging will contain some mixture of contaminated and uncontaminated measurements in the time series.

[0091] There are several possible actions that can be taken to correct these artifacts, including degassing the dispersant to remove air bubbles, or performing a cleaning cycle with fresh dispersant to remove contaminants.

[0092] See also Figure 5 , shows an example of background data 101, where the window of the measurement cell is subject to some contamination. Figure 5 The measurements shown may include the average light intensity obtained over a relatively long measurement period. Contamination of the cell window may be caused by, for example, fingerprints, dust, ink or oily residue. Such contamination may occur on the outer window of the cell, for example if someone touches the cell. Alternatively, contamination may occur on the inner window of the cell, for example due to residue from a previous experiment adhering to the inner window. This can be seen in the background data 101 as a "hump" at 103, rather than as Figure 3 A gradual decrease is seen. This is due to light scattering caused by contamination being measured on the detector. This hump can only be considered to be caused by dirt on the measuring cell window if there are no dynamic fluctuations in the background, as dirt on the measuring cell window does not appear to move over time. The corrective action for this is to inform the user to clean the measuring cell window. The detection of a relatively stationary hump can be used to classify the artifact as being caused by sample cell contamination.

[0093] See also Figure 6 , shows an example of background data 101 where the focal plane detector is misaligned with the incident laser beam. This can be seen in the background data 101 as spikes in some detector levels in region 104 (i.e., measurements recorded at low numbered, small angle detectors of the focal plane detector). This is caused by the detectors on one side of the incident beam receiving more light than the detectors on the opposite side. This may be caused by contamination on the cell window. If such an artifact is detected, the user should clean the cell window before re-measuring the background signal and / or checking the alignment and settings of the analyzer itself.

[0094] In this case, the background data 101 shows a spike (i.e., a high value of light energy) on every other detector in region 104. This is due to the detector layout of the particle size analysis system used to collect the data. The system can be roughly understood as having detector numbers that are alternately distributed on each side of the incident beam direction as the angle increases. For example, relative to the beam propagation direction, where 0° is the detector directly in front of the beam, the system may have detector No. 1 at +1°, detector No. 2 at -2°, detector No. 3 at +3°, and so on. This may be due to the spatial limitations of the system, where it is not possible to have detectors at both the positive and negative angles of each increment. Such an arrangement may be more efficient than setting detectors at corresponding positive and negative angles. Ideally, scattering is symmetrical around the beam propagation direction, so it is expected that the measurements at corresponding positive and negative angles are the same. Therefore, this arrangement can measure twice the number of angles using the same number of detectors, or can measure the same number of angles using half the number of detectors. This detector layout is more common in laser diffractometers.

[0095] If there is a misalignment causing more light to be scattered on one side than on the other, this will cause a spike on every other detector in the arrangement, e.g. Figure 6 This type of artifact is characteristic of optical misalignment and can therefore be classified as being caused by optical misalignment.

[0096] See also Figure 7, shows an example of background data 101 where there is an excessive thermal gradient within the instrument or dispersant (US2018 / 038782 gives an example method for identifying scattering caused by thermal gradients in dispersants, which is incorporated herein by reference). This can be seen in background data 101 as high background energy levels or large fluctuations that are asymmetric about the illumination axis, or fluctuations with specific time characteristics. The thermal gradient may be caused by the system temperature being higher than the dispersant. The thermal gradient may cause the refractive index of the dispersant to change within the volume of the cell.

[0097] As the dispersant temperature stabilizes and the refractive index gradient decreases, the background signal should decrease. For more volatile dispersants, the time required for the dispersant temperature to stabilize may be longer. If this artifact is encountered, the user should allow time for the background measurement to stabilize. Artifacts caused by thermal instabilities are classified as such based on artifact detection performed by an algorithm that detects thermal instabilities.

[0098] See also Figure 8 , an example particle characterization system 200 is shown. The system 200 includes a particle analyzer 1, such as Figure 1 The particle analyzer shown. The system also includes a computer or controller 201. The controller 201 is configured to receive data from each detector of the analyzer 1. The controller 201 may include a processor configured to execute methods for analyzing data collected from the detectors. These methods may include methods for checking the quality of background data measurement results. These methods may also include: once it is determined that the analyzer 1 is configured correctly and the data is recorded as "good", calculating the particle size and particle size distribution. The controller 201 may include a storage device for electronically storing data received from the detectors or any calculated results.

[0099] The system 200 also includes a user interface 202. The user interface may include a visual display and a user input device (such as a keyboard and / or a mouse). The visual display may display information output from the controller 201 to the user. This may include results calculated by the controller 201 (including real-time results), and / or instructions and warnings prompting the user to take steps to correctly perform the particle size analysis. The user input device allows the user to interact with the graphical user interface (GUI) displayed on the visual display.

[0100] See also Fig. 9 , an example method 300 for automated background data quality assessment is shown. The background assessment method 300 can be used to identify and correct the various artifacts discussed above.

[0101] A user can set up a particle size analysis system and take background measurements. Background measurements can be taken over a given period of time to observe and evaluate dynamic fluctuations in the background data. Method 300 can be applied to pre-recorded background data (e.g., a user takes background measurements and then applies method 300), or method 300 can be applied to background data continuously and in near real time as the background data is being recorded.

[0102] First, the dynamic algorithm can be applied to the background data. The dynamic algorithm can be used to determine if there are any transient instabilities within the time series data of the background. The dynamic algorithm can detect time-varying contamination caused by particles or bubbles, as well as undesirable background caused by artifacts (e.g., using low stirring speeds in the case of high viscosity or volatile dispersants).

[0103] The dynamic algorithm may also be able to use the time series background data to distinguish different causes of instability. The dynamic algorithm may be able to identify whether the background data is being negatively impacted by bubbles and then instruct or prompt the user to take corrective action (e.g., degassing). A short transient duration may indicate bubbles and / or a long transient duration may indicate the absence of bubbles.

[0104] The dynamic algorithm may also be able to identify particulate contamination within the dispersant and subsequently instruct or prompt the user to take corrective action (e.g., perform a cleaning cycle with fresh dispersant to remove contaminants).

[0105] The dynamic algorithm can be a peak detection algorithm. The peak detection algorithm can be configured to identify contamination within the time series background data by detecting the percentage of time series that contain abnormal peaks in the time series. The first step is to smooth the time series of each detector using the following normalized Gaussian kernel:

[0106]

[0107] where σ is the standard deviation. This acts as a low pass filter, removing any high frequency components in the signal.

[0108] The summation in the denominator is performed over the local range of the time series, from -(2lσ+0.5) to (2lσ+0.5). Initial analysis of the algorithm found that l=4 and σ=50 are suitable values ​​to use. The time unit in the above formula is the number of "snapshots" or individual points in the time series data. In a measurement with a sampling rate of 10kHz (probably a wet measurement), l=4 means a time window of 0.4ms. Time windows of 0.1ms to 1ms can be used.

[0109] The smoothed signal can then be convolved with a 1D Sobel filter

[0110] Sob(t=0,1,2)=[-1,0,1] (2)

[0111] This filter acts as a peak / edge detection algorithm that amplifies any peaks that remain within the time series background data after filtering out the high frequency components. The Sobel filter can be viewed as a gain or edge detection stage.

[0112] Any suitable low pass filter and gain or edge detection algorithm may be used instead of the Gaussian and Sobel functions described above.

[0113] After this convolution, any data less than a specified threshold can be set to zero. The percentage of remaining non-zero data can then be calculated. If the calculated percentage is greater than a specified value, the background data is identified as containing dynamic fluctuation artifacts. The peak threshold value can be selected to be 0.5 (or a number between 0.4 and 0.7), and the peak percentage value can be set to 0.5% (or a percentage between 0.25% and 1.5%). These values ​​can be changed to produce the best results, and the user may be able to specify the values ​​of these values.

[0114] Alternatively, the dynamic algorithm can be an adaptive diffraction algorithm. An adaptive diffraction algorithm may also be referred to as a transient event algorithm. An adaptive diffraction algorithm may be configured to detect dynamic noise sources. The adaptive algorithm toolbox is a set of algorithms that analyze time series scattering data (i.e., a vector of scattering intensities at each of a plurality of detectors) and calculate the principal components of the data. The principal components may be determined with reference to a set of reference scattering data that is calculated for a range of particle sizes and / or particle size distributions. Determining the principal components of the scattering data may provide a compact representation that best characterizes the scattering pattern. The algorithm may use the principal time series of principal components determined from the scattering data time series to determine a "ground state" - a set of closely clustered points with similar principal components that describe most instances in the measurement, and "transient events" - other outliers that represent particles or scattering events that are different from the ground state. For example, a clustering algorithm may be used to determine the ground state and identify states that are not in the "ground state" cluster.

[0115] These transient events may represent, for example, a single large particle in a sample dominated by small particles. Alternatively, they represent any smaller population within the overall population of the sample. Only the first 2 principal components can be used, making this a 2D problem.

[0116] To reduce noise in the data and make data processing faster, the data can be averaged. The data can be divided into blocks of 100 snapshots (or at least 50 snapshots, or at least 20 snapshots) and then averaged. This effectively reduces the data sampling rate of the measurement (including "wet" measurements) from 10kHz to 100Hz.

[0117] Rather than examining peaks based on percentage of background measurements (similar to the peak detection algorithm described above), the adaptive diffraction algorithm uses the percentage of data that contains transient events to classify the background data. Other statistics can also be calculated for these transient events, such as the mean and standard deviation of the Euclidean distance of all data points in the principal component space. Euclidean distance is a form of vector distance or two-dimensional distance. Euclidean distance is perhaps the simplest form of principal component-based vector distance between transient events and the center of the ground state, but other measures of vector distance can also be used.

[0118] The Euclidean distance d is defined by the following formula, where x and y represent the distances in the principal component space, and n is the number of principal components considered in the analysis.

[0119]

[0120] The algorithm may use a Gaussian classifier, which effectively fits a 2D Gaussian probability distribution to the spread of data points in the principal component space. The threshold of the classifier may be, for example, 0.01, which means that the outer 1% of the data points in the Gaussian distribution are classified as transient events.

[0121] Alternatively, the standard deviation can be used to analyze the variation of the background data. It is also possible to calculate the maximum standard deviation of the time series for each detector for each measurement. This allows an empirically derived standard deviation limit to be derived, which can be used to distinguish unstable, fluctuating data from stable data.

[0122] Any of the above algorithms can be used as a dynamic algorithm to analyze the time series background data and identify whether there is a dynamic fluctuation in the background data. If it is determined that such a dynamic fluctuation exists, the controller can instruct or prompt the user to take corrective measures accordingly.

[0123] A static algorithm or a static contamination algorithm can be applied to the background data. The static algorithm can be applied after the dynamic algorithm. The static algorithm can check for the presence of artifacts observed in the static background measurement. The static background measurement can be the average of the background measurements taken over a given time interval. While the dynamic contamination check algorithm can be used to detect contamination within the cell, the static contamination algorithm can determine if there are artifacts that are not caused by dispersants within the cell, and these artifacts do not cause dynamic background effects. This can include contamination on the cell window or misalignment of the detector and light source.

[0124] Static algorithms may include algorithms for identifying hump-containing data. Hump-containing data may indicate contamination on the pool window.

[0125] The hump-containing data algorithm can be a numerical check algorithm. The hump algorithm can check, for example, whether the scattered signal on detector 15 is greater than signal detector 1, or whether the scattered signal on detector 1 is greater than 200, or whether the scattered signal on detector 20 is greater than 30. Data collected from other detectors can be analyzed to determine whether there are humps in other areas of the data.

[0126] Alternatively, the hump algorithm may include a curve fitting algorithm. The curve fitting algorithm may include fitting the background data according to two functions simultaneously. The two functions may be: an exponential decay function (exponential decay functions appear in good, clean background data); a Gaussian function (Gaussian functions approximate the hump caused by light scattering caused by dirt or other static contamination). The function used to fit the scattering distribution may be

[0127]

[0128] Where d is the detector number, k is the exponential decay constant, σ is the standard deviation of the Gaussian function, x is the center position of the Gaussian function, and A and B are scaling constants. The contribution of these two functions to the overall fit can be evaluated. If the contribution or amplitude of the Gaussian function is above a predetermined threshold, the algorithm can determine that the background data has a hump artifact.

[0129] Alternatively, the hump algorithm may be a machine learning algorithm configured to identify hump-containing data. With proper training, the machine learning algorithm may be able to pick out and identify poor quality data. Simulated data may be used to generate a large number of curves with and without humps to train the machine learning algorithm on this annotated data set.

[0130] A range of machine learning algorithms are evaluated to determine which performs best. A simulated data set is generated and used to train different types of machine learning models on the data, and the performance of these trained models is compared. At least some of the machine learning algorithms tested include artificial neural networks (ANNs).

[0131] The ANN tested does not depend on the type of data provided as input. The model may include a convolutional neural network. The model is trained using a simulated dataset and the Tensorflow / Keras deep learning library in Python.

[0132] Simulated data (used to train machine learning algorithms) can be modeled by combining a decay curve (e.g., a 1 / x change is used here) with a Gaussian peak to model how the hump appears. Background data with and without the hump can be generated and appropriately labeled for use in training.

[0133] The formula and table below describe how to generate such a data set. Reasonable ranges for setting the random values ​​are given in Table 1 below. These variables can be adjusted for use on a specific system (e.g., a system with different detector numbers).

[0134]

[0135] Pseudo-random background with hump = g(x) + f(x)

[0136] Pseudo-random background without hump = f(x)

[0137] Additionally, a small amount of random noise and / or instrument response functions may be added to the data to more accurately reflect the experimental setup used to obtain the light scattering data. This can make the training data and subsequent machine learning algorithms more accurate. The specific form of these random noises and / or instrument response functions may be determined empirically based on the experimental setup.

[0138] Alternatively, training data for machine learning can include experimentally obtained data using a light scattering experimental setup and then reviewed for artifacts and annotated accordingly.

[0139]

[0140] Table 1: Examples of suitable variables for generating simulated background data.

[0141] During the algorithm testing, it was found that the sequential convolutional neural network performed best. The layers of the convolutional neural network model are shown in Table 2 and can be reproduced in any modern deep learning library.

[0142]

[0143]

[0144] Table 2: Layers of the convolutional neural network found to perform best in identifying data containing humps. Static algorithms may also include algorithms for identifying data containing spikes. Data containing spikes may indicate that the analyzer system optics are misaligned. For example, the focal plane detector may be offset from the center of the beam, or appear to be offset due to contamination.

[0145] The spike data algorithm can be a value checking algorithm. For example, the spike algorithm can check whether the scattered signal on detector 1 is more than 2.5 times the signal on detector 2, or whether the scattered signal on detector 2 is more than 1.3 times the signal on detector 1, or whether the scattered signal on detector 3 is higher than 1.3 times the signal on detector 2.

[0146] The algorithm for spikey data can also be a machine learning algorithm, similar to the algorithm described for camel-humped data.

[0147] Any of the above algorithms may be used as a static algorithm (or part of a static algorithm) to analyze static background metrics and identify any static artifacts. If such static artifacts are determined to be present, the controller may indicate or prompt the user that one or more artifacts are present, what those artifacts are, and may further prompt the user to take corrective action accordingly.

[0148] If both the dynamic and static algorithms determine that the background data is "good", the user may proceed with the measurement and may be prompted to do so via the display. A particle sample should be placed in the analyzer and raw measurement data collected. Corrected data may be obtained by subtracting the background data from the raw measurement data as discussed above. Corrected data may be updated in near real time and displayed to the user as "live" data.

[0149] A negative data check may be performed on the corrected data. The actual data should fluctuate around zero. If there are significant negative values ​​on some or all detectors, this may indicate that the background data has not stabilized before the background data is measured. The negative data check may include a numerical check algorithm. The negative data check may check whether any of the first 15 detectors has a scattering signal less than a predetermined negative value (e.g., -5), or whether more than ten of the first 15 detectors have a scattering signal less than a predetermined negative value (e.g., -0.5).

[0150] If any of the above conditions are met, the negative data check will indicate that the data is erroneous and instruct or prompt the user to take corrective action accordingly (e.g., check the sample and analyzer, and repeat method 300). If the corrected data passes the negative data check, it can be passed to another algorithm that is configured to calculate a characteristic of the sample particle size, such as the modal particle size, the mean particle size, or the median particle size, or a particle size distribution (PSD).

[0151] See also Fig.10 , shows the results of a dynamic algorithm based on standard deviation. For this purpose, a large amount of background data known to contain fluctuations and a large amount of background data marked as good are used. The results show the value of the maximum standard deviation of all detectors and are graphically illustrated by a box plot. Fig.11 Shows Fig.10 Boxplot of the subset of the data included in , showing the range of “good” data.

[0152] These box plots show that there is only a small overlap between the maximum standard deviations measured for the "good" cases and the maximum standard deviations measured for the cases with fluctuations. Therefore, this algorithm can provide a simple and fast way to identify fluctuations. In this case, a maximum standard deviation threshold of 10 across all detectors is sufficient to flag most cases that contain fluctuations.

[0153] See also Fig.12 , shows the results of the adaptive diffraction dynamics algorithm, where the average Euclidean distance of transient events is used as the comparison statistic. Data are marked as "good" and containing fluctuations, and the results are similar to Fig.10 The box plot of is shown. Similarly, Fig.13 Shows Fig.12 Boxplot of the subset of the data included in , showing the range of “good” data.

[0154] Fig.12 and Fig.13 It is shown that the average Euclidean distance of transient events is also an acceptable criterion for identifying fluctuations, as there is almost no overlap between the “good” and fluctuating datasets on this measure.

[0155] See also Fig.14 , shows the results of a curve fitting method for detecting humps within static data. Although it is difficult to classify good data and data containing humps, this can be achieved if the combination of variables is considered. Fig.14 A graph showing the relationship between the hump center and the hump height obtained by the curve fitting algorithm is shown.

[0156] In this case, the hump can be filtered out by setting a threshold of height greater than 10 and center position greater than 20, removing the three outlier data points. Hums may be detected even on good data because the assumed good background does not have a pure exponential decay curve. The low hump in the inner detector is usually required to modify the Gaussian curve to achieve a good fit. Contamination is expected to result in larger humps, or humps in the outer detectors, which are different from the small Gaussian fit observed in the good data.

[0157] See also Fig.15 , shows the confusion matrix of the convolutional neural network model described in Table 2. The model was trained using the simulated data described in Formula 6 and Formula 7. The left image shows the counts and the right image shows the corresponding percentages.

[0158] The overall accuracy of the model is about 84%. From the percentage point of view, the good data is more likely to be misclassified than the bad data (i.e., the data containing the hump), which may be due to the fact that the good data does not completely conform to the exponential decay, as discussed above.

[0159] Machine learning algorithms can also be used to identify data containing spikes. However, based on the data available during testing, the simple numerical method described earlier proved sufficient, where the differences between elements in the background array data were analyzed to determine if there were any large changes.

[0160] See also Fig.16 , shows an example of background data containing spikes. The spikes (seen as alternating peaks on low detector numbers (e.g., detectors 1 to 10)) can be caused by focal plane detector misalignment. In this example, the spikes are on the order of 10 3 A threshold of, for example, 50 can be used to easily identify datasets containing spikes.

[0161] See also Fig.17 and Fig.18 , shows the classification results of dynamic fluctuation sources. Fig.17 The results of classifying fluctuations using the standard deviation algorithm discussed earlier are shown. Fig.18 The results of classification using the average transient duration, which can be derived by using the adaptive diffraction algorithm discussed previously and calculating the duration of the identified transient events, are shown.

[0162] The effects of particle contamination of the dispersant as well as the bubbles within the dispersant were studied.

[0163] Table 3 shows examples of different contaminants used during the studies.

[0164]

[0165] Table 3: Examples of different dispersants and particle contaminants used in the study of fluctuation source classification.

[0166] Two methods were used to create bubbles during the study. One method was to create bubbles using carbonated liquids and surfactants (the corrective measure was to degas the dispersant) and the other method was to create bubbles mechanically through agitator speed adjustment (the corrective measure was to correct the agitator speed). The liquids studied were soda water, sparkling water, deionized water with Tween surfactant, and oil, and the agitation speed ranged from 1000 rpm to 3500 rpm. The agitation speed experiments used ethanol, isopropanol, and isooctane dispersants, and the agitation speed ranged from 500 rpm to 3500 rpm.

[0167] Fig.17 shows that while standard deviation may be a good measure for detecting fluctuations (e.g. Fig.10), but may not be the best metric for classifying the source of fluctuations. It appears that particle contamination shows a higher maximum standard deviation than the bubble or stirring speed experiments, but there is overlap between these groups, which may make effective classification difficult. It also appears that bubbles show less fluctuation than particles, perhaps due to some hydrodynamic effect that limits bubble formation.

[0168] Furthermore, it appears that the duration of transient events in the bubble data is lower than that of other sources of fluctuations, probably due to the fact that the bubbles rise faster in the flow cell than the flow rate. Fig.18 It is shown that the average transient duration of bubbles is lower than other sources of fluctuations. A threshold can be set to classify bubbles. If the transient duration is below the threshold, bubbles are likely to be a source of fluctuations. If the transient duration is greater than the threshold and the stirrer speed is high, the system can suggest reducing the stirrer speed. If the transient duration is greater than the threshold and the stirrer speed is low enough, particles are likely to be a source of fluctuations.

[0169] Various algorithms can be combined or used in combination to effectively identify artifacts, classify them, and alert the user to the class of artifacts that are present. Dynamic artifacts can be identified using standard deviation and then classified as bubbles or dispersant contamination, for example, using a transient algorithm.

[0170] The exemplary embodiments are not intended to limit the scope of the invention, which should be determined with reference to the appended claims.

Claims

1. A method for automatically identifying data quality issues in background data for laser diffraction particle characterization, the method comprising: receiving background data corresponding to light intensity measured by each of a plurality of detectors in a laser diffraction-based particle size analysis system; automatically determining, using a processor, whether the background data includes at least one artifact indicative of a source of error in the particle size analysis system; upon determining that the background data does include at least one artifact, providing, using the processor, an indication that the background data includes at least one artifact; The at least one artifact is classified and a result of the classification is reported.

2. The method according to claim 1, further comprising: generating background data by: irradiating a sample cell including a control sample with a light beam from a light source to generate scattered light through interaction of the light beam with the sample cell and / or the control sample, and detecting the scattered light with the plurality of detectors; wherein the control sample is nominally composed entirely of dispersant, and wherein the plurality of detectors are positioned at a plurality of different angles relative to a direction of propagation of the light beam.

3. The method of claim 1 or 2, wherein automatically determining whether the background data contains at least one artifact comprises applying at least one of the following algorithms to the background data: a dynamic algorithm configured to identify transient fluctuations within the background data; and A static algorithm is configured to identify static artifacts in the background data.

4. The method according to claim 3, wherein the dynamic algorithm comprises at least one of the following algorithms: (i) a peak detection algorithm configured to identify peaks within the time series background data and calculate a percentage of data in the time series background data that contains peaks with intensities greater than a threshold; (ii) an adaptive diffraction algorithm configured to identify transient events in the background data and calculate characteristics of the transient events; as well as (iii) a standard deviation algorithm, wherein the maximum standard deviation of the time series background data for each detector is calculated; and The dynamic algorithm is configured to identify the time series background data as containing dynamic fluctuations if at least one of the following conditions occurs: a) the calculated percentage is above a threshold value; b) the calculated characteristic of the transient event is above a threshold; and c) The calculated maximum standard deviation of the time series background data is higher than a threshold.

5. The method according to claim 3 or 4, wherein the static algorithm comprises at least one of the following algorithms: (i) a hump algorithm configured to identify humps within static background data; and (ii) a spike algorithm configured to identify excessive differences between the static background data measured between at least one pair of detectors; and The static algorithm is configured to identify the static background data as containing static artifacts if a hump is identified or if an excessive difference between detectors is identified.

6. The method according to claim 5, wherein the camel case method comprises: A machine learning algorithm comprising a machine learning model trained to identify humps in the static background data.

7. The method of claim 6, wherein the machine learning model comprises a convolutional neural network having at least five convolutional layers.

8. A method according to any one of the preceding claims, wherein a step of automatically determining whether the background data contains at least one artifact is performed before performing a particle analysis measurement on the particle sample, and if it is determined that the background data contains an artifact, the particle analysis measurement is aborted.

9. The method of claim 8, wherein the particle analysis measurements include: illuminating the sample with a light beam from a light source, thereby generating scattered light through interaction of the light beam with particles of the sample; Detecting raw measurement data, the raw measurement data including scattered light intensity distribution at a series of different scattering angles; and determining particle characteristics based on the raw measurement data using a processor.

10. The method according to claim 9, further comprising: generating corrected measurement data by subtracting the background data from the raw measurement data; as well as A negative value data check is performed, wherein the corrected measurement data are identified as faulty if the number of detectors having negative corrected measurement values ​​is above a threshold or the maximum negative value of all the corrected measurement data is above a threshold.

11. The method according to any of the preceding claims, wherein the step of automatically determining whether the background data contains at least one artifact is performed at least once upon receiving the background data.

12. The method of any preceding claim, further comprising automatically determining a corrective action in response to the determination of the artifact type, and displaying an indication of the corrective action to a user.

13. A non-transitory machine-readable medium comprising instructions for configuring a processor to perform a method comprising: receiving background data corresponding to light intensity measured by each of a plurality of detectors in a laser diffraction-based particle size analysis system; automatically determining, using a processor, whether the background data includes at least one artifact indicative of a source of error in the particle size analysis system; In the event that it is determined that the background data does include at least one artifact, an indication is provided using the processor that the background data includes at least one artifact.

14. A laser diffractometer, comprising: Sample pool; a light source configured to illuminate the sample cell with a light beam, thereby generating scattered light through the interaction of the light beam with particles in the sample cell; a plurality of light detectors configured to detect scattered light intensity distributions at a range of different scattering angles; a processor configured to determine a particle characteristic based on the scattered light intensity distribution at a series of different scattering angles; The processor is further configured to: receiving background data corresponding to light intensity measured by each of the plurality of detectors; automatically determining whether the background data includes at least one artifact indicative of a source of error in the particle size analysis system; In the event that it is determined that the background data does include at least one artifact, an indication is provided that the background data includes at least one artifact.

15. The laser diffractometer of claim 14, wherein the processor is configured to perform the method of any one of claims 1 to 12.

Images