Method for estimating wear of a component

Abstract

This disclosure relates to a computer-implemented method for estimating component wear. At least one cluster analysis can be applied to measurements obtained during component use to identify clusters and generate alerts.

Description

Technical Field

[0001] This invention relates to a computer-implemented method for estimating component wear. At least one clustering analysis can be applied to measurements obtained during component use to identify clusters and generate alerts. Background Technology

[0002] When mechanical parts are used, parts of the parts wear down. Wear may include degradation or shape change of the part, most commonly occurring at surfaces in contact with another part of the machine or with external surfaces. Wear can be chemical, mechanical, or both, and can be exacerbated by conditions around the contact area, such as pressure, temperature, and lubrication.

[0003] Wear and tear necessitates the inspection and replacement of parts, as continuing to use worn parts is inefficient and potentially unsafe.

[0004] For example, frequently replaced parts in a vehicle include windshield wiper blades, filters, brake pads, tires, belts, and cylinders. All of these components are susceptible to frictional and thermal wear. Wear on any of these components makes driving the vehicle less safe because the component will be less efficient as it wears down (e.g., worn brake pads will provide poor braking performance), and because the eventual failure of the component removes a critical function from the vehicle (e.g., a faulty windshield wiper blade will not provide visibility to the driver in adverse conditions).

[0005] Wear is also detrimental in machining processes such as metal cutting, where material is removed in the form of chips to create a net shape through the relative movement between the tool and the workpiece. Indeed, one of the most important and unavoidable challenges in machining is the continuous wear of the tool. In this case, tool wear is the loss of material from the cutting surface due to the interaction between the tool and the workpiece. Tool wear is well known to cause many problems with machined parts, including morphological and dimensional variations, vibration and chatter, and insufficient surface finish.

[0006] Machining with worn or damaged tools will result in surface defects and cracks that propagate over time, leading to the component's rejection or premature retirement. Therefore, the industry employs a time-based maintenance strategy when machining safety-critical components, discarding the tool at a predetermined time regardless of its wear condition. Typically, tools may be retired when they retain 50-80% of their service life.

[0007] Assessing component wear during use can be challenging because the component may be inaccessible and / or in motion, rendering visual inspection ineffective. In the vehicle example, inspecting brake pad wear might require removing the wheel, a task that the average vehicle user might not be able to perform. Belts and other internal engine components can be even more difficult to inspect.

[0008] Accessibility can also be a challenge in machining. Attempting to inspect components such as tools used for machining has the added disadvantage that the tool must be taken out of service while the inspection is taking place. During the tool design and testing phases, designers may opt for offline wear measurements using a run-to-failure method. This method intermittently interrupts the cutting process and requires the tool to be temporarily removed from the machine to perform wear measurements using an optical microscope. This is a time-consuming and expensive method that also introduces confounding influences on the process, such as differences in tool position between the first and second machining cycles. These positional differences can affect cutting conditions, and therefore the wear progress may differ for machining cycles within the same time period.

[0009] Figure 1 This illustrates a typical breakdown of the time a tool spends in different states during its lifecycle. Less than half of the lifecycle of this exemplary typical tool is spent on machining, with the remainder of its lifecycle spent idle due to workpiece or tool measurements.

[0010] TCM aims to shift the industrial focus from preventative approaches to predictive maintenance strategies to reduce waste. Ideally, to improve process efficiency, a TCM system should also help reduce time spent on measurements to increase machining time.

[0011] Supervised machine learning can be used to assist TCM techniques. As with any supervised learning technique, a model can be built using data associated with many previously used tools, where the model attempts to define the relationship between the measured degradation of previously used tools and the tool's characteristics. For example, a simple model generated by supervised learning can discover and define the relationship between machining time to failure and tool diameter, workpiece hardness, and tool revolutions per minute (rpm). Then, for a new tool, such a TCM technique can use tool diameter, workpiece hardness, and tool revolutions per minute (rpm) as input parameters to the model derived from the machine learning to predict the time to failure. However, due to the complexity of machining, varying operating conditions, and the availability of descriptive labels for wear states, there are still situations where TCM may be inaccurate.

[0012] There is a need for improved estimation of component wear in order to reduce component downtime and improve safety. Summary of the Invention

[0013] This invention is defined by the appended independent claims. Embodiments of the invention are defined in the dependent claims.

[0014] In a first aspect, a computer-implemented method for estimating component wear is provided, comprising: performing a first cluster analysis on a first plurality of measurements to identify one or more clusters in the first measurements, wherein the first measurements include measurements of parameters of the component obtained during use of the component; and generating an alarm when a new cluster is identified in the first measurements.

[0015] In this way, the method provides an estimate of component wear by highlighting when the measured parameters have changed significantly enough to ensure that new clusters are identified.

[0016] This method uses measurements taken during the use of the relevant component, thus avoiding the harmful effects associated with using the test component or with dependence on data related to components other than those in use. Therefore, this method is more accurate because it does not rely on the assumption that components in use wear in a manner similar to methods used to build or train wear models (e.g., trained classifiers). The wear rate of components can vary considerably depending on the component's category and the conditions under which it is used, as well as on properties that cannot be easily measured or accurately predicted due to their randomness (e.g., inclusions and microstructure). Therefore, relying on assumptions about uniform tool characteristics can lead to reduced accuracy in wear estimation. In contrast, the method of this invention uses current data measured from the specific tool whose wear needs to be estimated. More accurate estimates of component wear can reduce component downtime by avoiding unnecessary inspections, improve safety by indicating the need for inspections where other estimates might not so indicate, and reduce wasted lifespan remaining when the tool is discarded.

[0017] Cluster analysis performed during tool use and updated as new measurements are obtained is advantageous because components typically exhibit significant changes in behavior when a certain level of wear is reached. For example, the appearance of cracks in a machining tool or the onset of abrasion in a drive belt can lead to changes in acoustic emission from the component, forces around the component, or temperatures generated in the friction zone of the contacting component. Cluster analysis can identify such changes in measurements. Furthermore, cluster analysis is advantageous compared to simple threshold measurements (where alarms are generated at simple thresholds) because threshold-based alarms are more susceptible to triggering outliers. For example, a very brief spike in temperature between the drive belt and sprocket (e.g., a single measurement) might be caused by factors unrelated to belt wear (e.g., debris entering and rapidly leaving the friction zone). Cluster analysis may not identify this outlier as the start of a new cluster, thus avoiding alarms that could lead to unnecessary inspections, maintenance, or component waste.

[0018] Alarm generation can be an alarm to the controller of a component and / or associated machinery to stop operation in preparation for inspection or as a safety measure. Alarm generation can also include audible or visual signals to the operator, or alarms sent as notifications to external devices.

[0019] The method may further include: performing a second cluster analysis on the second plurality of measurement results to identify one or more clusters in the second measurement results, wherein the second measurement results include measurement results of parameters of the component obtained after an alarm has been generated and during further use of the component; and generating an alarm when a new cluster is identified in the second measurement results.

[0020] After an alarm has been generated and the component has undergone its first inspection, it may be deemed suitable for further use and returned to its associated machinery. During further use, a second cluster analysis is performed in the same manner as the first cluster analysis, thereby generating an alarm in the same way. The second cluster analysis benefits from the same advantages as the first cluster analysis.

[0021] The method may further include: continuing a first cluster analysis on the first measurement result, and introducing a second measurement result into the first cluster analysis to identify one or more clusters in the combined first and second measurement results; and generating an alert when a new cluster is identified in the combined first and second measurement results.

[0022] There are advantages associated with both operating cluster analysis across the component's lifecycle (global analysis) and operating cluster analysis using only measurements obtained since the component was most recently removed from its associated machinery (e.g., during inspection) (local analysis). Trends in a global analysis can provide a more accurate estimate of overall component wear, i.e., an estimate of the component's remaining lifespan. This is because the cluster analysis can access measurements of the component under more diverse conditions (including when the component is new) and over longer time spans. However, global analysis may be less sensitive to changes in component behavior that are significant at a local scale (perhaps significant enough to warrant inspection) but less significant in the context of the component's entire lifespan. In particular, global analysis will be particularly insensitive to significant changes in wear characteristics exhibited in these measurements if locational differences after inspection result in measurements that are highly different from those obtained before inspection.

[0023] Parallel operation of global and local analyses (both of which provide alerts indicating new clusters) can thus provide a more comprehensive estimate of component wear. Alerts generated by the first and second cluster analyses can be distinct from each other, as they can be differentiated by the controller or operator. For example, if local analysis is used to monitor when inspection is necessary, the alert could be a notification to the operator recommending an inspection. Alerts generated by global analysis could be notifications to both the operator and manager that the tool is nearing the end of its lifespan.

[0024] The first cluster analysis may be referred to herein as global cluster analysis, and the second cluster analysis may be referred to herein as local cluster analysis. It should be understood that although only two parallel analyses are described (one including lifecycle measurements, and the other including measurements since the last inspection), any number of parallel analyses can be performed according to embodiments of the invention. For example, cluster analysis based on measurement results between the second and fifth inspections of a component can be performed.

[0025] The first and / or second cluster analysis can be performed using unsupervised machine learning algorithms, and the number of clusters to be determined during the analysis does not need to be predefined.

[0026] In this way, the estimation of tool wear becomes more accurate again because the method does not rely on a manually selected number of clusters, which may fail to reflect the behavior of parts in use. A predefined number of clusters instructs cluster analyses how many clusters they must place data points (in this case, measurements) into. This number can be determined based on the number of clusters that leads to a valid and accurate estimate for similar parts previously used. However, this also suffers from the same aforementioned problem associated with differences (sometimes imperceptible) between parts, which can limit the accuracy of predictions based on prior knowledge.

[0027] Unsupervised machine learning algorithms can include Dirichlet process mixture models (DPMMs). DPMM properties have been found to be particularly suitable for estimating component wear, especially when global and local DPMMs are executed in parallel.

[0028] Alarms can be a suggestion to perform component checks.

[0029] The parameters of a component can be one or more of the following: the component's temperature; the forces acting on the component; and acoustic emissions from the component. In this way, by measuring parameters other than the direct measurement of wear itself, these measurements can occur during the component's use. Therefore, removing the component from its associated machinery can be avoided, thereby reducing component downtime and eliminating the possibility of repositioning errors.

[0030] The parameter can be acoustic emission from the component, and the first cluster analysis and / or the second cluster analysis can use multiple adjacent frequency ranges as candidate features.

[0031] Acoustic emission has been found to be a particularly effective parameter to be measured when analyzing measured data using cluster analysis. While unsupervised cluster analysis advantageously does not rely on a model trained on the training data and optionally does not have a predefined number of clusters, it can be beneficial to provide the algorithm with some guidance related to candidate features. Candidate features are variables of interest to the algorithm; therefore, in the case of unsupervised clustering algorithms, candidate features are variables related to the data points to be clustered. When the variable in question is the acoustic emission of a component, frequency bins are effective candidate features.

[0032] The component can be a machining tool. It should be understood that while the component is a machining tool in the preferred embodiment, the methods described herein for estimating component wear are suitable and advantageous for any component of machinery. In particular, any mechanical component that generates parameters suitable for indirect measurement during use, including but not limited to temperature, acoustics, force, pressure, light of any wavelength, etc. Examples relating to vehicle components have also been given, but are not limiting.

[0033] Therefore, the associated machinery as referred to herein may include workpieces, robot control components, motors, etc.; components other than the tool itself used for machining workpieces. In the previous vehicle example, if the component is a drive belt, the associated machinery may include driven components, such as alternators or pumps, sprockets on which the belt is mounted, etc.

[0034] Each frequency range can correspond to a predetermined harmonic of chip formation in the tool, optionally within a 4kHz band. Chip formation is a particularly relevant factor when the part in question is a machining tool, as the high slip distance and rpm of machining tools make them especially susceptible to chip or crack formation. By using frequency ranges based on chip harmonics, new clusters are more likely to be identified when new harmonics of chip formation are reached (i.e., as chip formation becomes increasingly likely).

[0035] In a second aspect, a method for inspecting wear on a component is provided, comprising: performing a computer-implemented method according to the first aspect; and inspecting the component for wear in response to an alarm output from the method executed from the computer.

[0036] In a third aspect, a data processing apparatus is provided, including a manner for implementing the steps of the first aspect. The data processing apparatus can be any processing mode capable of or configured to perform the steps of the methods described herein. The data processing apparatus can be a single processing unit, a distributed system with a wired connection, or a distributed system with a wireless connection. Alternatively, the data processing apparatus can be hosted on a cloud server, and the analysis can be performed on the cloud server.

[0037] In a fourth aspect, a computer program including instructions is provided that, when executed by a computer, causes the computer to perform the steps of the first aspect. The computer program may, for example, be a mobile application configured to operate on a mobile device.

[0038] In a fifth aspect, a computer-readable storage medium including instructions is provided, which, when executed by a computer, cause the computer to perform the steps of the first aspect. Attached Figure Description

[0039] The invention can be more fully understood when considered in conjunction with the accompanying drawings, wherein similar reference characters designate the same or similar parts throughout several views, and wherein:

[0040] Figure 1 The breakdown of time spent in a typical tool wear test setup is depicted.

[0041] Figure 2 The process of Gibbs sampling is described.

[0042] Figure 3 A machining setup according to an embodiment of the present invention is described.

[0043] Figure 4 The spectrum of the PcBN tool under specific cutting conditions was depicted.

[0044] Figure 5 The DPMM clustering results of the testing tool A1 are depicted, using the harmonics of the debris formation frequency from AE as input features.

[0045] Figure 6 The DPMM clustering results of the testing tool A2 are depicted, using the harmonics of the debris formation frequency from AE as input features.

[0046] Figure 7 The DPMM clustering results of the testing tool A3 are depicted, using the harmonics of the debris formation frequency from AE as input features.

[0047] Figure 8 The DPMM clustering results of the testing tool A4 are depicted, using the harmonics of the debris formation frequency from AE as input features.

[0048] Figure 9 The DPMM clustering results of the testing tool B1 are depicted, using the harmonics of the debris formation frequency from AE as input features.

[0049] Figure 10 The DPMM clustering results of the testing tool B2 are depicted, using the harmonics of the debris formation frequency from AE as input features.

[0050] Figure 11 The DPMM clustering results of the testing tool B3 are depicted, using the harmonics of the debris formation frequency from AE as input features.

[0051] Figure 12 The DPMM clustering results of the testing tool B4 are depicted, using the harmonics of the debris formation frequency from AE as input features.

[0052] Figure 13 The DPMM clustering results of the testing tool A5 are depicted, using the harmonics of the debris formation frequency from AE as input features.

[0053] Figure 14 The DPMM clustering results of the testing tool A5 are depicted, using the harmonics of the debris formation frequency from AE as input features.

[0054] Figure 15 The number of clusters formed by two example tools (right-A4 and left-B3) for a range of α values ​​is depicted.

[0055] Figure 16 A system 100 according to an embodiment of the present invention is described.

[0056] Figure 17 The DPMM clustering results of the testing tool A1 are depicted, using the harmonics of the debris formation frequency from AE as input features.

[0057] Figure 18 The DPMM clustering results of the testing tool A2 are depicted, using the harmonics of the debris formation frequency from AE as input features.

[0058] Figure 19 The DPMM clustering results of the testing tool A3 are depicted, using the harmonics of the debris formation frequency from AE as input features.

[0059] Figure 20 The DPMM clustering results of the testing tool A4 are depicted, using the harmonics of the debris formation frequency from AE as input features.

[0060] Figure 21The DPMM clustering results of the testing tool B1 are depicted, using the harmonics of the debris formation frequency from AE as input features.

[0061] Figure 22 The DPMM clustering results of the testing tool B2 are depicted, using the harmonics of the debris formation frequency from AE as input features.

[0062] Figure 23 The DPMM clustering results of the testing tool B3 are depicted, using the harmonics of the debris formation frequency from AE as input features.

[0063] Figure 24 The DPMM clustering results of the testing tool B4 are depicted, using the harmonics of the debris formation frequency from AE as input features.

[0064] Figure 25 The DPMM clustering results of the testing tool A5 are depicted, using the harmonics of the debris formation frequency from AE as input features.

[0065] Figure 26 The DPMM clustering results of the testing tool A5 are depicted, using the harmonics of the debris formation frequency from AE as input features. Detailed Implementation

[0066] To illustrate and explain embodiments of the invention, detailed examples are presented in the field of machining. In this example, the component is a machining tool, and the measured parameter is acoustic emission. It should be understood that this is merely exemplary, and the methods and systems according to embodiments of the invention are applicable to any number of fields in which components are worn down by use, exposure, or by any other means. Similarly, parameters other than acoustic emission, measured during component use and preferably indirectly so as not to interrupt said use, can be used as input data for the described clustering algorithm.

[0067] For example, cluster analysis can be performed on temperature measurements taken near a fan (component) in a desktop computer (as part of the machine). The clustering algorithm, as described further in this document, can be applied to the temperature measurements and to alarms generated when new clusters are identified. For instance, an alarm might prompt the replacement or inspection of the fan.

[0068] What is described now is a realistic TCM system using the Dirichlet Process Hybrid Model (DPMM), which is able to handle confounding effects while preserving the accuracy of reasoning about previously unseen tools (unseen tools are typically new tools that have not yet been used in machining or tools that have not yet been directly measured to obtain measurements of tool wear). Therefore, the resulting system can be more general than previous attempts, which allows for easier implementation in industry.

[0069] An unsupervised clustering method is employed, utilizing a Dirichlet process mixture model to detect changes in cutting process characteristics online for diagnostic purposes. Besides providing a useful monitoring tool, this method has the potential to reduce the need for associated, exhaustive wear measurements required for prediction. The model is well-suited to the unstable and unpredictable nature of tool wear progression because the number of clusters required to determine possible damage states is not a priori. Therefore, this method has the ability to handle variations across homogeneous and heterogeneous tool material composition groups.

[0070] The method presented here is demonstrated as a means to reduce the time required for trials of new tool wear characteristics. In the example shown, results indicate that this method results in a (average) reduction of approximately 30% in test time during the outer diameter turning of surface-hardened steel using 10 polycrystalline cubic boron nitride tools from two different material compositions.

[0071] The ability to monitor and predict tool wear during machining is a crucial objective, as wear conditions significantly impact the surface quality of machined parts. However, establishing a comprehensive condition monitoring system for diagnosis and prediction requires extensive knowledge of measurement and tool wear. Collecting labeled datasets including damage information for this purpose can be expensive and time-consuming.

[0072] In this work, DPMM is used as an unsupervised method to detect changes in properties in data for online damage detection. The inventors applied DPMM to an acoustic emission (AE) dataset collected during turning operations. DPMM allows clustering as the data is collected, eliminating the need for prior settings of the possible number of clusters, thus reducing the need for in-depth prior knowledge of the machining process. The need for pre-labeled training data is also eliminated, thereby reducing the costs associated with data collection.

[0073] This work addresses the confounding effects introduced by fluctuating operations during experimentation by using two DPMMs in parallel. The intention here is that a novel cluster initiation will be used to prompt interventions (i.e., manual checks of the tools). Therefore, it is hoped that the number of process interruptions can be reduced. The aim is to increase the time spent on machining and reduce the time spent on tool measurements. Another objective, by tracking the characteristics of the data over time, is to stop the machining process before failure and reduce the number of premature tool disposals. Although not covered in this work, a semi-supervised approach could be adopted by adding damage labels to the model, thereby reducing the number of false positives.

[0074] Dirichlet Process

[0075] DPMM is used across many research fields for clustering data. In natural language processing, research in translation and text generation has led researchers to use DPMM to cluster verbs. This technique is applied to organize and predict words. In medical research, DPMM has been used to classify brain tissue from magnetic resonance imaging (MRI) scans. While GMM performs efficiently for well-defined tissue types, researchers are able to classify anomalous brain data by not having to pre-set the number of clusters in DPMM.

[0076] This section introduces the Gaussian mixture model of the Dirichlet process, where Gauss is used as the fundamental distribution.

[0077] DPMM can be used to cluster Gaussian and non-Gaussian data. DPMM can be viewed as an infinite Gaussian mixture model (IGMM), where a mixture of Gaussian distributions is used to cluster the data. Here, the number of Gaussian distributions utilized can approach infinity, allowing modeling of any non-Gaussian dataset; a cloud of data that does not follow a Gaussian distribution can be divided into countless small clusters, which are themselves Gaussian. IGMM can also be used to learn information about the probability that each data point belongs to each cluster. The idea is to find clusters within the data, and also to find the parameters (size, shape) and labels of these clusters (cluster 1, ..., cluster n). Finding the labels and parameters of the clusters in one step is impossible; therefore, Gibbs sampling is used in this work to infer the joint distribution. Before the algorithm is applied, DPMM does not need to know the number of possible clusters within the data; if the data already evaluated by the algorithm is sufficiently different from the current data point, then the Gibbs sampler can initiate new clusters.

[0078] The generative model of DPMM is shown in equations (1a) to (1e). In the Gaussian mixture model, each data point x is plotted... i (i = 1, ..., N, where N is the number of data points) are derived from a Gaussian distribution (with labels c). i The assumption of sampling (equation (1a)) in data clustering.

[0079] For each Gaussian distribution, the conjugate priors (a normal inverse-Wishart(N|W) distribution with hyperparameters (μ0,Σ0,κ0,ν0) (Equation (1b)) and covariance (Equation (1c)) are used to realize the closed-form solution of the posterior. In this work, the data are normalized and the prior clusters are Gaussian with zero mean and unit variance.

[0080]

[0081]

[0082]

[0083] c i |π~Mult(π) (1d)

[0084] π~Dir(a) (1e)

[0085] Cluster labels are sampled from a multinomial distribution (Equation (1d)). The mixing proportion π is the probability that a data point belongs to each cluster. To calculate these probabilities, a Dirichlet distribution is used because it is the conjugate of the multinomial distribution (Equation (1e)). π is controlled by the intensity parameter α (i.e., the Dirichlet process prior, which can also be called the scaling or scaling parameter)). The scaling parameter α determines the probability that a new cluster will be identified; that is, the chance of a new cluster being identified for a new data point (rather than the chance of assigning a new data point to an existing cluster) is proportional to α.

[0086] Ultimately, the goal of this process is to find the posterior distribution over the cluster labels from which the most likely labels can be selected. Finding the probability of all cluster labels given all the data (Equation (2)) is very difficult because it is impossible to sample each cluster simultaneously to find the clustering parameter (μ). c i,Σ c i)) At the same time, the mixing ratio of all clusters is also found.

[0087] p(c|X) (2)

[0088] This can be addressed by implementing a collapsed Gibbs sampler. The collapsed Gibbs sampler sequentially samples a new set of clustering parameters based on the labeled samples and samples a new set of labels based on the parameters (Equation (3)). Figure 2 The steps followed by the Gibbs sampler are illustrated. Each of these sampling steps can be performed in a closed-form manner due to the choice of conjugate priors. In other words, it is an efficient Markov chain Monte Carlo method that sequentially finds the distributions on cluster labels and cluster parameters. The folded Gibbs sampler is executed on a data window whose length is specified based on available computational power.

[0089] p(c i |x i X -i c -i (3)

[0090] Equation (3) represents the first probability distribution of interest. Assume the model has seen all the data X. -i All other clusters c-i And with the new data point, it is point x. i With cluster label c i The probability of is given. This posterior probability has a multinomial distribution. Here, -i represents all data points except data point i.

[0091] To compute it, the Gibbs sampler randomly assigns data to clusters and sequentially removes data points to update the clustering parameters and find the best cluster for that data point. The prior likelihood of extracting data from existing clusters (k = 1, ..., K) can be found in equation (4), where N -i,k It is the number of data points in the current category, and N is the total number of data points if all data points under consideration have been removed.

[0092]

[0093] New clustering (k) * The priors of ) are calculated using equation (5).

[0094]

[0095] Given data values, the current cluster in which the data belongs, all data already in that cluster, and some hyperparameters, the posterior prediction likelihood of a point belonging to each cluster should now be calculated, which is the probability of assigning the current data point to cluster k.

[0096] To calculate the likelihood of DPMM, equation (6) is used, referred to as the posterior predicted distribution. This indicates that the data points are clustered based on the likelihood of the posterior distribution defined on the parameters, where D... k This represents all the data that the algorithm has seen in a given cluster. For the DPMM model, this is represented by a multivariate t-distribution, which has heavier tails than the Gaussian distribution, thus promoting smaller clusters to accept new data points.

[0097] p(x|D k (6)

[0098] The posterior is normalized using marginal likelihood (i.e., the sum of all calculated posteriors ensures it is an efficient probability distribution) and the cluster label c for point i is found. i The data follows a multinomial distribution. If a point is assigned to a new cluster, the N|W prior is used to initialize that cluster, and the number of clusters is increased by 1. This process is repeated until all data in the window has been re-evaluated.

[0099] There are three main advantages to using this model. First, it eliminates the need for the operator to set the number of clusters (or, in this case, the damage states) before using the algorithm, thus removing the need for potentially unavailable prior knowledge of the process. Second, the entire model is controlled by hyperparameters; therefore, threshold tuning and calibration are unnecessary. Third, DPMM allows the covariance function to vary with the input data, resulting in models that can handle different datasets, which can be useful when detecting damage to multiple tools worn differently from each other.

[0100] Experimental setup

[0101] The test setup used here to explore the use of DPMM is... Figure 3 The diagram shows PcBN tools being used to machine surface-hardened workpieces until catastrophic failure. These are referred to as "end-of-life" tests. Since it is impossible to measure tool wear during machining, these tests are segmented so that four consecutive cuts (referred to as a pass) are performed at a time, followed by tool inspection. The inspection involves removing the tool from the machine to measure flank and crater wear using both 2D optical and 3D scanning microscopes, respectively. On average, completing four passes and the associated tool wear measurements takes more than 8 minutes. Over 42% of the time spent completing the tool wear tests is spent on actions associated with these measurements.

[0102] When selecting a model for TCM, limitations of this setup must be considered. First, tool removal for measurement is unavoidable, as it's the only way to collect damage tags. Once the tool is removed, it's impossible to fix it back in its precise pre-measurement position. Therefore, the depth of cut value for the next pass is uncertain. Depth of cut is crucial for machining. Variations in depth of cut can affect the forces, stresses, and dynamics of the process. Second, during the four-pass intervals mentioned above, the workpiece is replaced by a new one. This is because surface hardening only penetrates a small layer of the workpiece's outer diameter and is likely to be machined away after four passes.

[0103] Acoustic emission was measured here, and a Mistras Micro-30D differential sensor with a sampling rate of 1 MHz was used. This results in a frequency range of 0-500 kHz when the Nyquist criterion is considered. An Okuma Space Turn lathe LB3000EXII was used for this experiment. To gain a deeper understanding of the generation mechanism of AE, chips were collected.

[0104] AE signals were collected from two different types of tools, with the only variation being the material composition or the material's "grade"; the tool grade specifies the percentage of cubic boron nitride (cBN) particles in its composition. The grade has a direct impact on the tool's performance and defines its suitability for a particular type of machining operation. The two grades used in this work are designated as Grade A (tools A1-A5) and Grade B (tools B1-B5).

[0105] Feature selection in DPMM

[0106] An AE signal can be visualized as a spectrum in the frequency domain. Figure 4 In the example, the spectrogram of the tool is displayed as a function of the sliding distance (the distance the tool travels). Figure 4 In the diagram, frequencies are plotted in kilohertz on the y-axis, sliding distances in kilometers on the x-axis, and shading is used to plot signal power as average power per unit bandwidth in decibels per hertz. Important frequency ranges are highlighted here; sawtooth formation is assumed to occur at frequencies on the debris (also referred to as debris-forming frequencies in this work) and the harmonics of these frequencies, which may be due to nonlinearities within the system.

[0107] Ultimately, DPMM has the potential to be applied in industrial environments to automatically cluster data while running online. To achieve this, the input features should behave similarly across a set of homogeneous tools. From the spectrogram, it is clear that the power of the harmonics at the chip formation frequency increases with the progress of tool wear. Subsequently, for each tool grade, four frequency ranges (in the form of 4kHz segments) with the harmonics at the chip formation frequency (throughout the tool life) were selected as candidate features. It is important to note that prior knowledge of the process is required to identify these features that increase in intensity at the end of tool life. Since the process runs online, it is impossible to identify this behavior a priori for new tools. Here, the frequency bands used reduce the requirement for specific prior knowledge; however, the identification of these frequency bands is still achieved through initial trials using any new tool grade.

[0108] It should be understood that similar preliminary tests can be performed on any part of the machinery to determine candidate features. Such preliminary tests do not need to focus on acoustic emission; rather, they can focus on any measurable parameter described herein. Furthermore, the clustering algorithm described herein can be performed without using candidate features determined from preliminary tests. For example, arbitrary ranges can be used in the initial use of the part and iteratively refined. As an example only, if temperature is chosen as the measurement parameter during the use of the part, a temperature range starting at 50°C and incrementing in 25°C, 50°C, 75°C, or 100°C increments can be used. Similarly, force, light emission wavelength, and pressure can be used as candidate features using Newton, wavelength, and Pascal ranges, with or without preliminary testing.

[0109] Parallel DPMM

[0110] Clustering in DPMM arises from changes in the mean and variance of the input features. During AE generation in machining operations, these changes in mean and variance are assumed to be caused by the onset of tool wear. However, unavoidable confounding effects (such as changes in the workpiece at specified intervals or tool holder positioning after measurement) can also cause variations in the AE signal. Distinguishing between these two causes of clustering formation when using DPMM is crucial to avoid false positives.

[0111] In structural health monitoring (SHM), confounding effects may manifest themselves as both short-term and long-term trends, which must be distinguished from damage for accurate prediction. Among other methods, principal component analysis can be used to address confounding effects, where principal components with low variance can be used as damage-sensitive features that are not hindered by prominent environmental trends. While these techniques successfully suppress unwanted trends, they require a representative set of prior training data.

[0112] This work proposes a method that considers confounding effects while retaining sensitivity to potential tool damage. It also proposes running two DPMMs in parallel:

[0113] 1. The first DPMM starts at the beginning of each tool test and runs throughout the tool's lifespan. This DPMM is thereafter referred to as the global DPMM.

[0114] 2. The second DPMM is started at the beginning of each tool test and reset at the beginning of every four passes of the workpiece (i.e., at the point where workpiece changes and tool positioning changes occur). This DPMM will be named the Local DPMM.

[0115] The next section will explore the application of parallel DPMM to the aforementioned dataset.

[0116] Experimental results

[0117] For each tool, features are fed into two parallel DPMMs, where the Gibbs sampler is set with a window length of 200 (200 seconds of data) to achieve convergence to the target distribution.

[0118] Figures 5 to 12 The results for tools A1-A4 and B1-B4 are shown. Figures 5 to 8 Parallel clustering for tools A1 through A4 is shown respectively, and Figures 9 to 12 Parallel clustering for tools B1 through B4 is shown separately. Each of these plots contains results from both the global DPMM and the local DPMM. In all cases, the top four plots are associated with the global DPMM, and the bottom four plots are associated with the local DPMM. In each case, harmonics of the debris-forming frequency from acoustic emission are used as input features F1 through F4. As mentioned above, features F1 through F4 are frequency “ranges,” which in this exemplary work are 4 kHz segments carrying the harmonics of the debris-forming frequency throughout the tool life. F1 is the lowest Hz frequency range and F4 is the highest, and the scale in the y-axis for each feature has been normalized before plotting.

[0119] exist Figures 5 to 14 and Figures 17 to 26 In the method shown, new clusters are identified in both the global DPMM and the local DPMM. New clusters in the global DPMM are identified by vertical lines, while those in the local DPMM are identified by color changes. In both cases, when a new cluster is identified for any of the candidate / input features F1 through F4, that new cluster is activated for all other candidate features. For example, in Figure 10 and Figure 22 In the diagram, points 10.1 and 22.1 of interest show the identification of new global clusters based on data points in the fourth frequency interval (feature F4). Based on this identification result, features F1 to F3 are updated and begin to classify data points into this new cluster (see points 10.2 and 22.2 of interest), even though the new cluster identification thresholds may not yet be met for the data points of features F1 to F3 themselves.

[0120] Local clustering homogenizes input features in the same way. However, unlike global clustering, local clustering analysis restarts on each tool check. In other words, a local clustering graph comprises multiple local clustering analyses, each using a dataset that includes data points between vertical dashed lines.

[0121] It should be understood that this method of applying a clustering identifier based on a single feature to all other features in the relevant DPMM (i.e., within the global DPMM or within the local DPMM) is one possible embodiment, but other methods are also possible according to the invention. Features can remain completely separate, and new alerts are generated individually for each feature. Alternatively, a threshold number of features can be identified before alert generation. Similarly, an alert can only be generated when a new cluster is identified for all features. As used herein, using cluster analysis to identify new clusters aims to cover all such options.

[0122] In the method according to the invention, an alarm is triggered when any one of the four candidate features used globally can identify a new cluster.

[0123] Figures 17 to 26 It shows the relationship with Figures 5 to 12 The same result, but presented in color. The correspondence between color and grayscale legend can be found from... Figure 5 and Figure 17 The legend in the diagram is derived from this.

[0124] For the global DPMM results, clusters are presented in different colors, with the initiation of each cluster indicated by a vertical line in the same color as the cluster label. The first cluster is initiated upon the arrival of the first data point based on a threshold. Here, the threshold is set to 1, specifying the number of data points a given cluster needs before it can be identified as a new cluster. Since this is an unsupervised approach and does not require training on similar data, DPMM restarts from cluster 1 for each new tool. In the local DPMM results, changes to the tool are indicated by black vertical dashed lines, where DPMM is reset to 1. Although the cluster initiation lines are not shown for clarity, the cluster colors maintain the same order as in the global DPMM.

[0125] As it applies Figures 17 to 26 The following analysis also applies. Figures 5 to 12 The results show how cluster identifiers can be used as triggers to generate alerts that provide guidance on the condition of components.

[0126] Generally, cluster 1 is usually the largest (ignoring the effects of confounding), indicating that the energy change of the AE harmonics is insufficient to guarantee new clusters under the current hyperparameters. Physically, this may mean that the early wear progress of the tool does not significantly affect the harmonics of the AE. In most cases, the energy of the harmonic frequencies increases rapidly near the end of the tool's life due to tool wear. Small dislocations and microcracks in the grain boundaries within the tool material release stress waves due to the large forces acting on the tool, also altering the tool's wear profile. Therefore, the fragmentation process may be permanently altered, leading to an increase in the energy of the AE signal. In global DPMM, these structural and conditional differences are likely to present themselves as non-Gaussian clusters, prompting the algorithm to separate into smaller Gaussian clusters, thus resulting in an increase in the number of clusters near the end of the tool's life.

[0127] In practice, global cluster initiation can now be used as a warning of tool degradation for operators; the emergence of new clusters can be used to trigger inspections instead of after a set sliding distance. Therefore, cluster initiation can be considered a novel method for setting detection thresholds. Observations from tool inspections can be used to determine whether the tool has been damaged beyond a predetermined tolerance threshold. In the latter case, the new cluster can be considered an undamaged state, thus allowing similar observations to be classified into the same cluster in the future.

[0128] However, if cluster initiation occurs at the point in time of workpiece change and tool measurement, the operator should refer to local cluster initiation. Local clustering is able to detect corresponding changes in the AE (Automatic Effect) in cases where the tool may be damaged due to a collision with the workpiece at the start of the pass. In other words, resetting the clusters upon workpiece change essentially means that the local DPMM will only trigger a new cluster if a significant change is seen in the data during the machining of each workpiece. Therefore, it can be assumed that these changes are not due to anything other than tool wear (and associated effects such as increased temperature and force), as all other operational effects remain constant.

[0129] The results from hyperparameter tuning indicate that smaller α values ​​can be used for this dataset because the clusters are well separated. This result may be beneficial because in practice, if many clusters are formed, it is disadvantageous because cluster initiation is a warning sign for tool checks. If a large number of clusters are initiated, the number of tool checks will also increase, which will reduce the time savings that DPMM can achieve.

[0130] Some tools show that after a new cluster is initiated, data can sometimes be grouped back into previous clusters. For example, in... Figure 11 and Figure 23After tool B3 initiates a third cluster at approximately 4 km (denoted as points of interest 11.1 and 23.1), some data is still assigned to clusters 1 and 2 (data identified by points of interest 11.2 and 23.2 includes most of the data points classified as clusters 1 and 2). One explanation for this behavior might be altered cutting edge. As the tool traverses along the workpiece, it continuously loses material from its cutting edge. Some particles ejected from the tool may be larger than others, causing temporary holes in the cutting edge, thus causing a momentary change in chip formation characteristics. The method of this invention can achieve a short delay before generating an alarm, accompanied by analysis of data points after identifying a new cluster to determine the proportion of subsequent data points that are identified as belonging to that new cluster. A threshold proportion of data points belonging to the newly identified cluster before the alarm is generated can be introduced.

[0131] As a result, DPMM clusters the data into new categories. However, once the tool moves laterally along the workpiece after this event, it may smooth out the hole and revert to a cutting edge similar to the earlier case. In this case, the AE data will behave similarly to the previous clustering, resulting in the behavior shown in the figure.

[0132] Another reason for the aforementioned clustering may be the increased temperature at the cutting edge. In some cases, the generated chips may envelop the workpiece and accumulate at the cutting edge. In this situation, the temperature at the cutting edge rises rapidly, thus burning the chips around the tool. Therefore, the chip formation characteristics (and thus the generated AEs) can vary as workpiece and tool temperatures affect how the material behaves. This is one of the reasons, among others, that temperature is another measurement parameter from which clustering identifiers can be used to generate alerts.

[0133] As mentioned above, the regression of certain data points to earlier clusters may be useful for operators using DPMM to monitor tools, and this information may also be useful for research and development purposes. This is especially true near the end of tool life, where data is clustered back to cluster 1 or 2 after tool failure, indicating that the AE behaves similarly to that at the start of tool life. The reduced contact area between the tool and the workpiece may be the reason for this observation. To avoid unnecessary inspections due to confounding effects, increasing the threshold makes it possible to require a larger number of data points before initiating a new cluster. Therefore, the method of the present invention not only benefits from increasing tool life and reducing the remaining useful tool life at tool retirement, but also helps guide further research on component wear estimation.

[0134] However, this threshold should be carefully considered, as a large value might mean that clustering initiation is less sensitive to tool damage; clustering will not initiate until the threshold is met, at which point the tool could be catastrophically damaged. This example can be found in... Figure 12 and Figure 24 As seen in the figure, global cluster 2 was initiated (marked as points of interest 12.1 and 24.1 in the figure, respectively), possibly due to confounding effects: changes in the workpiece. It can be seen that global cluster 2 was initiated abruptly for all four features (F1 to F4). However, at this time, a second cluster also immediately initiated in the local DPMM, indicating that the data had been affected during this pass (marked as points of interest 12.2 and 24.2 in the figure). This cluster initiation may be due to the interaction between the tool and the workpiece as it enters the cut, where the sudden contact affects the chip formation characteristics of the tool. The initiation of cluster 4 is likely due to confounding effects (marked as points of interest 12.3 and 24.3 in the figure). Similarly, for global cluster 4, the significant and transient changes in acoustic emission that caused the cluster initiation suggest confounding effects. Here, global clusters 3, 5, and 6 can be used as warnings to the operator regarding damage inspection.

[0135] For several reasons, this parallel clustering method is better suited than binary methods (such as outlier analysis) for tool damage detection problems. In outlier analysis, the data is assumed to be sampled from a Gaussian distribution in order to apply Gaussian statistics.

[0136] Any deviation from a Gaussian distribution is considered an outlier indicating abnormal data. As previously mentioned, DPMM can automatically separate larger non-Gaussian clusters into smaller Gaussian clusters without requiring a priori setting of the number of clusters and thus repeated detection of outliers. By assigning data to multiple clusters, it is also possible to understand the various stages of degradation of the tool throughout time, while remaining robust to confounding effects. For example, the second cluster initiation of tools A2 and A3, and the fourth cluster initiation of tools B1 and B4, occurred due to a large change in AE energy compared to previous data. From the tool wear check, it is clear that these initiations were not caused by damage. By using the parallel DPMM method, operator intervention should be avoided in these cases, whereas this is not the case for outlier analysis.

[0137] Given the lack of continuous wear measurements, it is difficult to directly verify that DPMM performs clustering at salient points. One way to partially evaluate the success of DPMM is to consider experiments that notice anomalous or unconventional behavior. In the next section, clustering by different tools will be explored to further investigate this point.

[0138] Clustering of different tools

[0139] Predicting tool wear status for tools that differ from other tools in the dataset can lead to low accuracy because the training set is not representative of the test set. Since DPMM does not require a training phase, the feasibility of detecting AE changes for tools that behave differently from other tools is investigated here.

[0140] Figure 13 and Figure 14 The results of global and local DPMM for tools A5 and B5 are shown respectively. Figure 25 and 26 This is also true. For comparison, the same features and hyperparameters as all other tools within each level were used for these tools. It is clear that in both cases, the AE features behave differently from the other features in the dataset. It can be seen that for tools A5 and B5, the global DPMM detects cluster 3 at approximately 1.6km and 3km, respectively, with the AE energy increasing at these specified frequencies due to unknown behavior. Due to this increased energy in the middle of the tool's lifetime, the global DPMM does not detect new clusters near 2.7km (A5) and 5.2km (B5), even though the signal energy has increased relative to the beginning of the tool's lifetime. However, at this point, local clustering has created a second cluster to reflect this energy increase, which can be used to intervene and avoid machining with damaged tools. Since DPMM does not rely on data from other tools for training, it is able to cluster different data in a way that still avoids machining with damaged tools. It is clear that DPMM is robust to a wide range of tool behaviors, demonstrating its suitability and effectiveness.

[0141] Parallel DPMM Guidelines for Avoiding Tool Failures

[0142] To accurately identify when a tool may be nearing failure and to avoid false positives, the parallel DPMM guidelines are recommended:

[0143] Global clustering should be preferred over local clustering unless global clustering has been initiated when the artifact changes. This is because global clustering considers the entire dataset and is therefore sensitive to large changes in the dataset. Since artifact changes may lead to new clusters, clustering during artifact changes should be handled with care.

[0144] • If global clustering is initiated when the artifact changes, and local clustering is initiated during the artifact, then the initiation of local clustering should be treated as a warning.

[0145] • If global clustering is initiated when the artifact changes, but local clustering is not initiated during that artifact, then the initiation of global clustering should be ignored without warning.

[0146] Modify the DPMM method when applied to tools

[0147] Various modifications can be made to the detailed embodiments described above. As previously stated, this method can be used for any mechanical component susceptible to wear, and any parameter can be measured and analyzed. Furthermore, the method itself can be modified based on any of the steps detailed below.

[0148] Because DPMM does not rely on learning damage labels, the dataset can be of any size. Since larger datasets are more informative, the frequency with which data arrives at clusters can be increased by averaging the AE data more frequently. In the detailed example above, a data point represents one second of data comprising 1 million raw AE data points. However, in alternative embodiments, the data per second can be represented by a much larger number of data points (i.e., one data point can represent one-tenth of a second, thus increasing the amount of information input to DPMM tenfold), thus producing a much more informative data trend or data cloud for clustering. As more data becomes available, DPMM may be able to initiate clustering earlier than currently is, as the cloud will be better defined at a higher resolution, reducing the time spent on tool testing. It should be understood that the specific representation of the data points described above is not limiting, and the input data points can represent any number of seconds of measurement data including any number of raw parameter measurement data points; the appropriate number of seconds and raw parameter measurement data points will depend on the components and parameters to be analyzed. The selection of such an appropriate number of seconds or data points can be made by those skilled in the art and can be modified using preliminary testing and / or real-time results.

[0149] The number of frequency intervals can be increased. By doing so, higher frequency resolution can be achieved, which may result in more features with increased intensity near tool failure. Four parameter intervals (here, frequencies) are proposed as input features or candidate features in the detailed embodiments, but the number of input features can be between 1 and 100, between 2 and 50, between 3 and 25, and preferably 4.

[0150] Here, DPMM has been used as an unsupervised technique where data is clustered without a target value. In the future, DPMM could be used as a semi-supervised learning technique, where damage labels can also be included in the model when detecting changes.

[0151] Interestingly, in this work, the increase in (harmonic) AE energy caused the final clustering start-up prior to failure to typically occur during the penultimate pass of the workpiece. This information could be further investigated to develop predictive systems; by studying the nature of the final clustering start-up prior to catastrophic failure, it is possible to predict the remaining useful life of the tool.

[0152] Transfer learning can also be applied to this dataset; in transfer learning, clusters (along with known labels) are projected onto a space that already contains unlabeled clusters. The idea is to minimize the distance between cluster means. For the dataset collected here, inductive transfer learning can be applied, where the algorithm has information about the current clusters and requires a small amount of data from the new clusters to make a similarity judgment. However, it is unclear whether this method will be robust to tool breakage or wear.

[0153] The method of this invention is compatible with specialized microscopes that can capture 3D images of the tool while it is still in its support, thus eliminating the need to remove the tool. By doing so, it is possible to eliminate positional changes that have a confounding effect on this current dataset. Therefore, the relative importance and necessity of parallel DPMMs are reduced, and a single DPMM can be utilized accurately.

[0154] Furthermore, settings such as those shown here can be programmed using tool wear tolerances and thresholds that can be used to evaluate tools at cluster initiation without operator intervention, resulting in fully online tool wear testing.

[0155] in conclusion

[0156] This work explores the idea of ​​unsupervised learning, where changes in monitored data can be detected using an incomplete set of damage labels without a training phase. DPMM and Gibbs sampling are used in this work to cluster AE data collected during the turning process. Since this method can be viewed as an infinite Gaussian mixture model, the number of clusters does not need to be set a priori, resulting in a fully online implementation of the algorithm. By using unsupervised cluster initiation as a warning / hint for operators to conduct tool condition surveys, DPMM can be successfully used to reduce the time spent measuring tools and to provide points where tools can be retired. This method using cluster initiation as a detection threshold has not been observed in previous literature.

[0157] The model's input features are taken from the frequency domain of the acoustic emission signal, which is measured throughout the tool's lifespan. Harmonics of the debris-forming frequency energy are used as input features because they are generally found to increase in intensity as the tool approaches failure.

[0158] Although the clustering analysis used throughout the detailed embodiments described herein is DPMM, it should be understood that alternative clustering analyses can be utilized to achieve similarly efficient results. Alternative clustering algorithms capable of identifying clusters in data related to component wear conditions include, but are not limited to: affinity propagation, k-means clustering, mean shift, agglomerative clustering, and spectral clustering.

[0159] Save time through parallel DPMM

[0160] Due to the confounding effects on this dataset, a combination of parallel global DPMM and local DPMM is implemented here for the first time in the literature; the global DPMM samples from the entire dataset collected during the tool's lifespan, while the local DPMM samples only from data collected during each workpiece (the point of structural change in the process). By using these parallel DPMMs, it is possible to avoid machining with damaged tools while reducing the time spent on tool measurements. A complete list of time savings is presented in Table 1. Promising initial results show that tool measurement time can be significantly reduced from 42% to 13.2% of the average test time when using this method. Here, the results are also compared with a preventative maintenance strategy, in which the operator stops machining when the tool meets a predefined threshold (determined based on the machining operation).

[0161] The results presented in this work may be beneficial for industrial applications for several reasons. First, this method does not require training data encompassing all operating conditions to safely predict unseen tools. Second, it reduces the number of routine tool wear measurements currently performed to collecting damage labels only when prompted by the algorithm. By running DPMM in parallel, the effects of confounding factors should be avoided. This method also works with AE features that can be used online without preprocessing such as dimensionality reduction. Furthermore, since DPMM clusters based on changes in features, it has the potential to detect damage on tools with different material compositions and on tools that behave differently from other tools. In short, this method can be applied to any tool as long as the input features are enhanced as they approach tool failure. In fact, this method can be applied to any part of a machine and any input features (i.e., measurable parameters) that change with the use and wear of the part.

[0162]

[0163]

[0164] Table 1: Summary of time savings when using DPMM compared to other monitoring strategies. Since the run-to-failure method results in over 42% of test time being spent on tool measurements, DPMM leads to significant time savings.

[0165] In the examples presented in this paper, the hyperparameter α is fixed a priori at 20; a low α value limits the number of clusters initiated. Tuning the hyperparameter α and the threshold can be used to modify the number of samples in each cluster. In the general SHM case, the presence of cracks is considered damage, and cross-validation can be used to check for misclassification in DPMM, thus aiding in the tuning of the hyperparameter α. If the clusters are well-separated, the number of clusters generated is insensitive to the α parameter.

[0166] Since DPMM is used in this work to cluster data online, the value of α cannot be learned prior to find the optimal distribution. To understand the effect of the α parameter, an optimal distribution is found prior. Tools A4 and B3 (respectively in...) were used to obtain the optimal distribution to understand the effect of the α parameter. Figure 8 and Figure 11 The features of the results (in the middle) are fed through an algorithm for multiple α values. In each case (i.e., for each α value), the algorithm is repeated 100 times to obtain the average number of clusters initiated. The results have been used Figure 15 The box plot is presented in the image. The average number of clusters across 100 runs is shown in red, and outliers are indicated by red crosses that deviate beyond ±2.7σ (annotated by a fine whisker). The top and bottom of the boxes represent the 75th and 25th percentiles, respectively.

[0167] from Figure 15 It is clear from the data that when α is small, the number of clusters is independent of the α value. This means that the clusters are well separated. When considering α values ​​between 0.1 and 50, the average number of clusters changes very little. Therefore, the suggestion here is that any α value between 1 and 50 can be used without significantly affecting the formation of the number of clusters.

[0168] The window length of the Gibbs sampler can also affect the results of DPMM. Although it is not a hyperparameter, it can influence the number of clusters and the value of the cluster initiation point. This is because the window length governs the number of data points the algorithm re-evaluates. In practice, it is useless to the operator if the Gibbs sampler re-evaluates data points from earlier sliding distances and reassigns them to a different cluster than the one the data point was in at that time.

[0169] However, the window length should be large enough to ensure that the Markov chain converges to the target distribution. Nevertheless, increasing the window length may reduce computational speed, so engineering judgment should be used again to find a reasonable window length.

[0170] Example System

[0171] Figure 16A system 100 according to an embodiment of the present invention is depicted. The system 100 includes associated machinery 102, components 104, sensors 106, processing devices 108, and mobile devices 110.

[0172] The associated mechanism 102 is depicted as including component 104 and sensor 106, but this is merely exemplary. Component 104 is removable and replaceable; this is why estimating its wear is advantageous. Sensor 106 may be included within the associated mechanism 102 or may be attached to and / or separate from it. In some embodiments, component 104 includes sensor 106. This may be advantageous when physical contact between component 104 and sensor 106 facilitates the measurement of relevant parameters during use of component 104 (e.g., if the force applied by or to component 104 is to be measured). Sensor 106 may be a multipurpose sensor configured to measure multiple different parameters during use of component 104, or it may be a sensor dedicated to a single parameter (such as acoustic emission). Sensor 106 may be configured to measure parameters of component 104 indirectly, i.e., during use of component 104 and without interfering with the use of component 104.

[0173] Processing device 108 is any computer processor suitable for performing the steps of the methods of the present invention described herein. Processing device 108 may be included within associated machinery 102, having a wired connection to that machinery, or having a wireless connection to that machinery. In the depicted embodiment, processing device 108 receives measurement results 112 of parameters of the component 104 in use via a wireless connection from sensor 106. Wireless communication protocols suitable for such transmissions will be apparent to those skilled in the art. Furthermore, measurement results 112 may be transmitted to processing device 108 via the Internet.

[0174] Processing device 108 performs at least one clustering analysis according to embodiments of the present invention. Measurement results 112 can be received continuously from sensor 106 or can be received in one or more data packets. Data processing device 108 uses the received measurement results 112 to perform one or more clustering analyses and generates an alarm 114 after identifying a new cluster in the measurement results 112. Alarm 114 can be processed by data processing device 108 itself and can include one or more calculated instructions sent to associated machinery 102 to stop operation and recommendations for component inspection or replacement for user review. Data processing device 108 may alternatively or additionally transmit alarm 114 to a third device (e.g., mobile device 110) for user review. Data processing device 108 may be included within mobile device 110 or may be detached from mobile device (as shown). Communication between processing device 108 and mobile device 110 can be wired or wireless, as described with respect to communication between sensor 106 and processing device 108.

Claims

1. A computer-implemented method for estimating component wear, comprising: A first cluster analysis is performed on a plurality of first measurement results to identify one or more clusters among the first measurement results, wherein the first measurement results include measurement results of parameters of the component obtained during use of the component; An alert is generated when a new cluster is identified in the first measurement result; A second cluster analysis is performed on multiple second measurement results to identify one or more clusters among the second measurement results, wherein the second measurement results include measurement results of the parameters of the component obtained after the alarm has been generated and during further use of the component; An alert is generated when a new cluster is identified in the second measurement result; Continue performing the first cluster analysis on the first measurement result, and incorporate the second measurement result into the first cluster analysis to identify one or more clusters of the combined first and second measurement results; as well as An alarm is generated when a new cluster is identified in the combined first and second measurement results.

2. The method as described in claim 1, wherein, The first clustering analysis and / or the second clustering analysis are performed using unsupervised machine learning algorithms.

3. The method as described in claim 2, wherein, The unsupervised machine learning algorithm does not predefine the number of clusters to be determined during the analysis.

4. The method of claim 3, wherein, The unsupervised machine learning algorithm is the Dirichlet process hybrid model.

5. The method of claim 1, wherein, The alert is a suggestion to perform a component check.

6. The method of claim 1, wherein, The parameters of the component are one or more of the following: the temperature of the component; the force on the component; and the acoustic emission from the component.

7. The method of claim 6, wherein, The parameters are acoustic emissions from the component, and the first cluster analysis and / or the second cluster analysis use multiple adjacent frequency ranges as candidate features.

8. The method of claim 7, wherein, The component is a machining tool.

9. The method of claim 8, wherein, Each frequency range corresponds to a predetermined harmonic formed by debris in the tool.

10. The method of claim 9, wherein each frequency range is a 4kHz band.

11. A method for inspecting wear on a component, comprising: Perform the computer-implemented method as described in claim 1; as well as In response to an alarm output from the method implemented in the computer, the component is inspected for wear.

12. A data processing apparatus comprising components for carrying out the method as described in any one of claims 1-11.

13. A computer-readable storage medium comprising instructions that, when executed by a computer, cause the computer to perform the method as described in any one of claims 1-11.

Images