Wear detection device

Abstract

[Problem] To provide a wear assessment device capable of performing automatic re-learning of machine learning without needing a specialist. [Solution] A wear assessment device 100 for a tool in a processing machine is characterized by comprising: a feature amount extraction unit 130 for acquiring a feature amount from an output of a sensor; a dynamic learning unit 140 for constructing a learning model on the basis of the feature amount; an inference unit 160 for inferring a degree of wear on the basis of the learning model constructed by the dynamic learning unit; a threshold value storage unit 180 for storing a threshold value of wear; and a wear assessment unit 190 for assessing wear on the basis of the degree of wear and the threshold value. The wear assessment device is also characterized in that: the inference unit determines whether inference based on the current learning model is appropriate, and if determining that the foregoing is not appropriate, then transmits a re-learning instruction to the dynamic learning unit; and the dynamic learning unit, upon receiving control data from the inference unit, performs re-learning on the basis of the threshold value and an additional feature amount different from the feature amount that was used when constructing the current learning model, and updates the learning model.

Description

[Technical Field]

[0001] The present invention relates to a wear determination device that applies machine learning. [Background technology]

[0002] Many techniques that apply machine learning to tool wear assessment or lifespan prediction have been proposed. For example, a lifespan prediction device disclosed in Patent Document 1 observes lifespan-related data to create a probabilistic model of the replacement lifespan of consumable parts, and predicts the replacement lifespan of consumable parts based on the observed lifespan-related data using the created probabilistic model. Patent Document 1 states that it is possible to predict the lifespan of consumable parts of manufacturing machines with a certain degree of accuracy even when there is little collected data. [Prior art documents] [Patent documents]

[0003] [Patent Document 1] Japanese Patent Application Publication No. 2019-207576 Summary of the Invention [Problem to be solved by the invention]

[0004] However, in conventional machine learning, including that disclosed in Patent Document 1, wear conditions are first learned under specific conditions to construct a learning model, and then the degree of wear is inferred under the same conditions based on the learning model. While this method can infer with high accuracy when the conditions remain unchanged, there is a problem in that the accuracy drops significantly when conditions such as the machining method, tools, or workpiece change, making it impossible to make appropriate inferences (generalizability issue). Therefore, even if learning is performed once, re-learning is required every time the conditions change. However, re-learning is extremely difficult unless you are a machine learning expert, and there is a problem in that it cannot be handled by on-site workers.

[0005] Therefore, an object of the present invention is to provide a wear determination device that can automatically perform machine learning relearning without the need for an expert. [Means for solving the problem]

[0006] A typical configuration of the present invention is a tool wear determination device for a processing machine, comprising a sensor that measures the condition of a workpiece or tool, a feature extraction unit that acquires features from the output of the sensor, a dynamic learning unit that constructs a learning model based on the features, an inference unit that infers the degree of wear based on the learning model constructed by the dynamic learning unit, a threshold memory unit that stores a wear threshold, and a wear determination unit that determines wear based on the degree of wear and the threshold, wherein the inference unit determines whether the inference based on the current learning model is appropriate, and if it determines that it is not appropriate, sends a re-learning instruction to the dynamic learning unit, and upon receiving control data from the inference unit, the dynamic learning unit performs re-learning based on additional features and thresholds that are different from the features used when constructing the current learning model, thereby updating the learning model.

[0007] When conditions such as the machining method, tools, and workpiece change, the degree of wear inferred from the features changes significantly. As a result, the judgment by the wear judgment unit also changes significantly earlier or later than the actual wear limit. However, by dynamically updating the learning model using the features at the same time as the inference, it becomes possible to automatically retrain the machine learning without the need for retraining by an expert (improved generalizability).

[0008] Another representative configuration of the present invention is a tool wear determination device for a processing machine, comprising: a sensor that measures the state of a workpiece or tool; a feature extraction unit that acquires features from the output of the sensor; a dynamic learning unit that constructs a learning model based on the features; an inference unit that infers the degree of wear based on the learning model constructed by the dynamic learning unit; a threshold memory unit that stores a wear threshold; a dynamic threshold setting unit that stores a threshold input by an operator in the threshold memory unit; and a wear determination unit that determines wear based on the degree of wear and the threshold, wherein the inference unit determines whether the threshold has been changed, and if it determines that it has been changed, sends a re-learning instruction to the dynamic learning unit, and upon receiving control data from the inference unit, the dynamic learning unit performs re-learning based on the features and the changed threshold and updates the learning model.

[0009] In conventional machine learning, the threshold value had to be set before the inference program started, and changing the threshold required stopping and restarting the inference program. However, with the above configuration, the threshold value can be dynamically updated, allowing the operator to reset it without stopping the inference program, improving convenience.

[0010] The sensor is preferably a vibration sensor that acquires vibration acceleration as a state of the tool or workpiece, and the feature quantity extraction unit acquires the feature quantity using vibration acceleration. Vibration acceleration is highly dependent on the degree of wear and is robust against changes in other conditions, making it a preferred state quantity for acquiring the feature quantity. [Effects of the Invention]

[0011] According to the present invention, it is possible to provide a wear determination device that can automatically perform machine learning relearning without the need for an expert. [Brief explanation of the drawings]

[0012] [Figure 1] 1 is a block diagram of a wear determination device according to an embodiment of the present invention. [Figure 2]10 is a flowchart illustrating the operation of an inference program of the wear determination device. [Figure 3] FIG. 10 is a diagram illustrating an example of feature amounts. [Figure 4] FIG. 10 is a diagram illustrating an example of feature amounts. DETAILED DESCRIPTION OF THE INVENTION

[0013] Preferred embodiments of the present invention will be described in detail below with reference to the accompanying drawings. Dimensions, materials, and other specific values ​​shown in the embodiments are merely examples for facilitating understanding of the invention and, unless otherwise specified, do not limit the present invention. In this specification and drawings, elements having substantially the same functions and configurations are designated by the same reference numerals to avoid redundant explanation, and elements not directly related to the present invention are not shown.

[0014] FIG. 1 is a block diagram of a wear determination device according to this embodiment, and FIG. 2 is a flowchart illustrating the operation of an inference program of the wear determination device.

[0015] The wear determination device 100 is a device for determining wear on the tool 12 of the processing machine 10 that processes the workpiece 20. The processing machine 10 is a processing machine that mainly performs cutting. It is assumed that one tool is subjected to, for example, 100 to 150 machining operations (machining 100 to 150 workpieces), and the tool is replaced when it reaches its wear limit. The flowchart in FIG. 2 is executed for each machining operation (each workpiece). First, the inference program determines whether or not machining is in progress (step 300). If machining is in progress, data is acquired by the data acquisition unit 120 (step 302).

[0016] The data acquisition unit 120 acquires outputs from the sensor 110 attached to the tool spindle and the sensor 112 attached to the work spindle. Specifically, the data acquisition unit 120 includes an amplifier and a logger.

[0017] In this embodiment, sensors 110 and 112 are vibration sensors. Other than vibration sensors, temperature sensors, displacement sensors, and the like are also possible. However, the values ​​of temperature sensors and displacement sensors are significantly affected by factors other than wear. In contrast, vibration acceleration is highly dependent on the degree of wear and is robust to changes in other conditions, making it a preferable state quantity for acquiring feature quantities. It should be noted that equivalent functionality can be achieved by inputting data such as "position deviation" and "integrated power value" of the tool spindle and work spindle obtained from NC (Numerical Control) in addition to acceleration data obtained from the vibration sensors into the data acquisition unit.

[0018] The feature extraction unit 130 extracts and stores feature values ​​from the state values ​​acquired by the sensors 110 and 112 (step 304). In this embodiment, feature values ​​are acquired using vibration acceleration. Feature values ​​that can be acquired from vibration acceleration include the absolute average, RMS (Root Mean Square), STFT (Short-time Fourier Transform), the fundamental frequency (cutting edge passing frequency) of the fluctuating cutting force, and its harmonics.

[0019] Next, the inference program determines whether to perform inference or to perform initial learning (step 308) (step 306). Inference is not performed when a learning model does not yet exist. Inference is performed when a learning model has already been constructed.

[0020] If inference is not performed (if a learning model does not yet exist), the operator 30 determines the wear limit and ends one machining operation. The inference program then performs learning using the accumulated feature values ​​(step 308) and constructs the first learning model.

[0021] The dynamic learning unit 140 can perform both static learning, which constructs a learning model only from the feature quantities output by the feature extraction unit 130, and dynamic learning, which performs re-learning in response to a re-learning instruction (flag) from the inference unit. This dynamic learning is a feature of the present invention. Known algorithms such as search algorithms and genetic algorithms can be used as machine learning algorithms. The dynamic learning unit 140 stores the constructed learning model in the learning memory unit 150.

[0022] When inference is performed in step 310, the inference unit 160 infers the degree of wear based on the learning model constructed by the dynamic learning unit 140 (step 310). That is, inference is performed using the learning model read out from the learning memory unit 150 and the features sequentially acquired from the feature extraction unit 130. The degree of wear is a probabilistic model, and is intermediate data that indicates the degree of wear with a certain probability.

[0023] The inferred value processing unit 162 performs weighting to prevent erroneous determination when the sensors 110, 112 output outliers. Specifically, if the amount of machining since the tool change is small, a coefficient that greatly underestimates the output of the inferred value processing unit 160 is applied. Conversely, if the amount of machining since the tool change is large, the underestimation coefficient is reduced. This improves the validity of the inferred value and prevents erroneous determination.

[0024] The wear determination unit 190 determines the wear of the tool using the wear degree (inferred value) obtained via the inferred value processing unit 162 and the threshold value obtained from the threshold storage unit 180 (step 318). The display device 200 displays the wear degree output by the inference unit 160 and the determination result (whether the wear limit has been reached) output by the wear determination unit 190 (step 318) (step 320). Here, a set value is input in advance to the threshold storage unit 180 before the first learning (step 308). However, even while the inference program is looping, the value can be updated at any time by the operator 30 inputting it via the dynamic threshold setting unit 170.

[0025] Next, re-learning will be explained. After performing the inference in step 310, the inference unit 160 determines whether the inference was appropriate (step 312). The determination of whether the inference is appropriate can be made, for example, when the inference value calculated by the current learning model loses correlation with the number of processes. If the inference is inappropriate, the inference unit 160 sends a re-learning instruction 161 (control data) to the dynamic learning unit 140 (step 313).

[0026] The inference program of the wear determination device 100 also determines whether the threshold value has been changed (step 314). If a new threshold value has been set in the threshold value storage unit 180 by the operator 30 through the dynamic threshold value setting unit 170, the inference unit 160 transmits a re-learning instruction 161 (control data) to the dynamic learning unit 140 (step 315).

[0027] When inference is performed (the flow from step 310), the inference program also determines whether the current tool has reached its wear limit (step 322). If the wear limit has been reached, a signal to stop the processing machine 10 is sent (step 323), and the program returns to step 300. If the wear limit has not yet been reached, the program determines whether a re-learning command 161 has been issued (step 324). If a re-learning command 161 has not been issued, the program returns to step 300.

[0028] If a re-learning instruction 161 has been sent, the dynamic learning unit 140 performs re-learning using the accumulated features (step 326). The features used for re-learning at this time are features (hereinafter referred to as "additional features") that are different from the features used when constructing the current learning model (the latest learning model). The additional features may be sent from the inference unit 160 to the dynamic learning unit 140 together with the re-learning instruction 161, or may be sent directly from the feature extraction unit 130 to the dynamic learning unit 140.

[0029] For example, suppose an initial learning model is constructed for tool A. In conventional technology, this learning model is used to make inferences about tools B and C (step 310 onwards), and no further updates to the learning model are made. However, in the present invention, when a re-learning instruction 161 is issued, the learning model can be updated using the features (additional features) of tools B and C as well.

[0030] Furthermore, if the number of machining operations (e.g., 120 times) when initially constructing a learning model for tool A is not close enough to the wear limit, learning will be insufficient and correct inference will not be possible. In such cases, the operator 30 can make a judgment while looking at the trend of the judgment results in the dynamic threshold setting unit 170, and can cause re-learning to be performed for an additional number of machining operations (e.g., the feature amounts for 121-150 times are the additional feature amounts).

[0031] When conditions such as the machining method, tool, or workpiece change, the degree of wear inferred from the feature values ​​changes significantly. This causes the wear determination unit 190 to make a determination significantly earlier or later than the actual wear limit. However, as explained above, by dynamically determining whether the inference is appropriate using the feature values ​​at the same time as the inference and updating the learning model as appropriate, it becomes possible to automatically retrain the machine learning system without the need for expert retraining (improved generalizability).

[0032] In addition, in conventional machine learning, the threshold value had to be set before the inference program started, and changing the threshold value required stopping and restarting the inference program. However, with the above configuration, the threshold value can be dynamically updated, allowing the operator to reset it without stopping the inference program, improving convenience.

[0033] Once the re-learning (step 326) is complete, the process immediately returns to step 310 and restarts the inference. In this case, if inappropriate inference continues in step 312, an infinite loop will occur, so an upper limit is set on the number of times re-learning can be performed in the case of inappropriate inference. When the upper limit is reached, a message to that effect is displayed to the worker 30 and the process returns to step 300.

[0034] 3 and 4 are diagrams illustrating examples of feature quantities. As described above, in this embodiment, feature quantities are acquired using vibration acceleration, which is a state quantity that is robust against conditions other than wear.

[0035] Figure 3(a) shows the average absolute value of vibration acceleration versus the number of processing steps as a feature. Alternatively, the sum of the areas created by the amplitude can be calculated. Figure 3(b) shows the root mean square (RMS) of the amplitude of vibration acceleration versus the number of processing steps.

[0036] Referring to Figures 3(a) and (b), it can be seen that in both cases, the average absolute value or RMS increases almost proportionally as the number of cuts increases. In this way, there is a correlation between the number of cuts and the average absolute value of vibration acceleration, or the number of cuts and the RMS of the amplitude value of vibration acceleration. By using this correlation as a basis for machine learning to characterize the relationship between the average absolute value or RMS and the degree of wear, it is possible to infer the degree of wear from the average absolute value or RMS for each cut.

[0037] Figure 4 shows the STFT (short-time Fourier transform) of vibration acceleration during five-pass machining of a workpiece (gear). No. 1 in Figure 4(a) is data for a new tool, and No. 18 in Figure 4(b) is data for a tool that has reached its wear limit.

[0038] Referring to Figure 4, the horizontal axis represents time, and we can see that a spectrum appears for each machining pass. The fifth pass is the finishing process, which takes a long time. Overall, No. 18 in Figure 4(b) has a higher frequency intensity, which indicates that it has been detected that vibration increases as wear progresses. Furthermore, Figure 4(b), which shows more wear, has more spectra appearing around 800 Hz in the fifth pass. By using machine learning to characterize these differences in the frequency distribution of the STFT, it is possible to distinguish between new parts and those at the wear limit. The difference in frequency distribution can be detected by quantifying the distribution, or by image processing the distribution image as a pattern.

[0039] Furthermore, instead of using a single index as a feature, multiple indexes may be combined. This can further improve the reliability of the inference. Also, since each index should have a correlation with the number of edits, it is preferable to constantly evaluate the correlation between the number of edits and the feature, and not display the judgment result if there is no correlation. This can prevent erroneous judgments from being displayed.

[0040] While the preferred embodiments of the present invention have been described above with reference to the accompanying drawings, it goes without saying that the present invention is not limited to such examples. It is clear that those skilled in the art can conceive of various modifications and alterations within the scope of the claims, and it is understood that these modifications and alterations also fall within the technical scope of the present invention. [Industrial Applicability]

[0041] The present invention can be used as a wear determination device that applies machine learning. [Explanation of symbols]

[0042] 10...machining machine, 12...tool, 20...workpiece, 30...operator, 100...wear determination device, 110...sensor, 112...sensor, 120...data acquisition unit, 130...feature extraction unit, 140...dynamic learning unit, 150...learning memory unit, 160...inference unit, 161...relearning instruction, 162...inference value processing unit, 170...dynamic threshold setting unit, 180...threshold memory unit, 190...wear determination unit, 200...display unit

Claims

1. A tool wear determination device for a processing machine, a sensor for measuring the condition of a workpiece or tool; a feature extraction unit that acquires feature amounts from the output of the sensor; a dynamic learning unit that constructs a learning model based on the feature amount; and an inference unit that is independent of the dynamic learning unit and that infers a wear level based on the learning model constructed by the dynamic learning unit; a threshold value storage unit that stores a wear threshold value; a dynamic threshold setting unit that stores a threshold input by an operator in the threshold storage unit; a wear determination unit that determines wear based on the wear degree and the threshold value, the inference unit determines whether the threshold has been changed, and if it determines that the threshold has been changed, sends a re-learning instruction to the dynamic learning unit; The wear determination device is characterized in that, when the dynamic learning unit receives control data from the inference unit, it performs re-learning based on the feature amount and the changed threshold value and updates the learning model.

2. 2. The wear determination device according to claim 1, wherein the sensor is a vibration sensor that acquires vibration acceleration as the state of the tool or workpiece, and the feature extraction unit acquires the feature using the vibration acceleration.

Images