Industrial Inspection Systems

Abstract

To provide an industrial inspection system for reducing cost and improving determination accuracy.SOLUTION: An industrial inspection system 1 includes a quality determination device 4 and an abnormality determination device 5. The quality determination device 4 acquires determination data from an inspection object by a determination data acquisition unit 41, and determines the quality of the inspection object on the basis of the determination data and a learned model A. The abnormality determination device 5 stores first evaluation data based on intermediate data in the learned model A, and acquires second evaluation data based on data in the middle of processing the determination data in a quality determination unit 43 of the quality determination device 4. Then, it calculates inter-data distance L between the first evaluation data and the second evaluation data, and determines whether a data acquisition condition is abnormal when the determination data acquisition unit 41 acquires the determination data on the basis of a comparison result obtained by comparing the inter-data distance L with a threshold P.SELECTED DRAWING: Figure 1

Description

[Technical Field]

[0001] The present invention relates to industrial inspection systems. [Background technology]

[0002] Conventionally, in the manufacturing process of industrial products, an industrial inspection system has been used that judges the quality of a product by machine learning using trained data related to the quality judgment of the product. Patent Document 1 discloses a configuration in which image data of the product's appearance is acquired as judgment data, and a trained model by a neural network created from the trained data is used to judge the quality of the product. [Prior art documents] [Patent documents]

[0003] [Patent Document 1] Japanese Patent Publication No. 2021-189039 Summary of the Invention [Problem to be solved by the invention]

[0004] However, in pass / fail judgments using the trained model, if the judgment data deviates from the trained data, the accuracy of the pass / fail judgment of the product decreases. For example, in the above-mentioned industrial inspection system, when acquiring appearance image data as judgment data, if acquisition conditions such as the shooting environment change, the acquired judgment data may also change significantly. If the judgment data changes significantly, the judgment data may deviate from the trained data. As a result, even if the product is pass / fail, it may be judged as defective or as indeterminable based on the judgment data that deviates from the trained data, which may reduce the judgment accuracy.

[0005] In order to prevent and improve such a decrease in judgment accuracy, it is conceivable to directly compare the judgment data with the learned data and determine whether the judgment data deviates from the learned data based on the comparison result. However, when comparing the judgment data with the learned data, T 2 If known statistics such as the number of dimensions is used, the number of dimensions becomes extremely large, resulting in a huge amount of calculation, which is inefficient, disadvantageous in terms of cost, and unrealistic. Also, while it is conceivable to visually inspect products that are judged as defective or unclassifiable by the system, or to re-inspect them using a different quality judgment device, this is also inefficient and disadvantageous in terms of cost.

[0006] The present invention has been made in view of the above problems, and aims to provide an industrial inspection system that reduces costs and improves judgment accuracy. [Means for solving the problem]

[0007] One aspect of the present invention is an industrial inspection system including a quality determination device that determines the quality of industrially used products, and an abnormality determination device that determines an abnormality in a data acquisition condition of the determination data used for the quality determination, The quality determination device is a judgment data acquisition unit that acquires the judgment data from the product as an inspection target; a trained model storage unit that stores a trained model created in advance by machine learning the judgment data of the product and the pass / fail judgment result of the product; a quality determination unit that determines quality based on the determination data of the inspection target and the trained model, The abnormality determination device a first evaluation data storage unit in which first evaluation data based on intermediate data in the trained model is stored; a threshold value storage unit that stores a threshold value related to the inter-data distance from the first evaluation data; a second evaluation data acquisition unit that acquires second evaluation data based on data during processing of the judgment data in the pass / fail judgment unit; a distance calculation unit that calculates a data distance between the first evaluation data and the second evaluation data; an abnormality determination unit that determines whether or not the data acquisition conditions when the judgment data acquisition unit acquired the judgment data are abnormal based on a comparison result obtained by comparing the inter-data distance calculated by the distance calculation unit with the threshold value; a determination result output unit that outputs a determination result by the abnormality determination unit, in industrial inspection systems. [Effects of the Invention]

[0008] The industrial inspection system includes a quality determination device and an abnormality determination device. The quality determination device acquires judgment data from an inspection object using a judgment data acquisition unit, and determines the quality of the inspection object based on the judgment data and a trained model. The abnormality determination device stores first evaluation data based on intermediate data in the trained model, and acquires second evaluation data based on data during processing of the judgment data by the quality determination unit of the quality determination device. The abnormality determination device then calculates a data distance between the first evaluation data and the second evaluation data, and determines whether the data acquisition conditions when the judgment data acquisition unit acquired the judgment data are abnormal based on the comparison result obtained by calculating the data distance and a threshold.

[0009] As a result, if there is a discrepancy between the judgment data and the learned data (training data) used to create the trained model, it can be detected as an anomaly in the data acquisition conditions for the judgment data in the judgment data acquisition unit. This can reduce erroneous judgments and reduce the number of dimensions and the amount of calculations compared to when the judgment data and the learned data are directly compared. Furthermore, since there is no need for a human to visually judge products judged as defective or unclassifiable by the system or to re-inspect them using a different quality judgment device, the first-pass rate of the production line can be improved. As a result, costs can be reduced and judgment accuracy can be improved.

[0010] As described above, according to the above aspect, it is possible to provide an industrial inspection system that reduces costs and improves determination accuracy. [Brief explanation of the drawings]

[0011] [Figure 1] FIG. 1 is a block diagram showing the configuration of an industrial inspection system according to a first embodiment. [Figure 2] (a) Conceptual diagram of a trained model in embodiment 1, (b) Conceptual diagram for explaining the layer structure in the trained model. [Figure 3] FIG. 3 is a conceptual diagram showing the relationship between first evaluation data and second evaluation data in feature amount space in the first embodiment. [Figure 4] FIG. 4 is a conceptual diagram showing the relationship between the distance between data and a threshold value in the first embodiment. [Figure 5] FIG. 2 is a flow chart showing a usage mode in the first embodiment. [Figure 6] FIG. 4 is a diagram showing image data as determination data in the first embodiment. DETAILED DESCRIPTION OF THE INVENTION

[0012] (Embodiment 1) 1. Industrial Inspection System 1 Configuration The configuration of an industrial inspection system 1 according to this embodiment will be described with reference to FIGS. 1 to 4. As shown in FIG. 1, the industrial inspection system 1 according to this embodiment includes a detection unit 2, a learning processing device 3, a quality determination device 4, and an abnormality determination device 5. The detection unit 2 detects industrial products that are the target of inspection by the industrial inspection system 1. In this embodiment, the product is a rack bar used in a vehicle steering mechanism, and the inspection target is image data of the luminescent state of the rack bar's exterior in a magnetic particle inspection. The detection unit 2 is composed of a camera that detects an exterior image of a magnetized rack bar with fluorescent magnetic particles adsorbed thereto in a magnetic particle inspection. If the rack bar has an abnormality in its surface texture due to scratches or the like, the adhesion state of the fluorescent magnetic particles contained in the inspection reagent on the magnetized rack bar will change from that of a normal portion. In the magnetic particle inspection, scratches or the like on the rack bar are detected by detecting this change.

[0013] In the industrial inspection system 1 of this embodiment, the product is not limited to the rack bar described above, and can be any product used in industry. The inspection target is also not limited, and can be, for example, appearance image data from penetrant inspection, ultrasonic observation data from ultrasonic inspection, or temperature distribution data from induction heating inspection. Furthermore, the inspection target is not limited to inspection by flaw detection, and can also be appearance image data for detecting abnormalities in the appearance of the product, or image data reflecting the operating state of the product.

[0014] The learning processing device 3 shown in FIG. 1 creates a trained model. The learning processing device 3 includes a training data acquisition unit 31, a quality acquisition unit 32, a feature calculation unit 33, and a model creation unit 34. The training data acquisition unit 31 acquires training data for creating a trained model A using data detected by the detection unit 2. In this embodiment, the training data acquisition unit 31 acquires training data for rack bars whose quality is known, indicating whether the product is good or defective. Then, the quality acquisition unit 32 acquires quality information linked to the training data. The quality information is input by an operator, who has previously determined whether the product is good or defective. The feature calculation unit 33 calculates multiple feature amounts related to quality from the training data. These are configured by a predetermined arithmetic unit (not shown).

[0015] Then, a trained model is created in the model creation unit 34 based on the feature amounts and the pass / fail information. The model creation unit 34 is also configured by a predetermined calculation device (not shown). The trained model created by the model creation unit 34 is stored in the trained model storage unit 42 (described later).

[0016] The trained model created by the model creation unit 34 may be created by machine learning based on training data. In this embodiment, the trained model may be configured as a neural network. As shown in FIG. 2(a), the trained model A in this embodiment is a convolutional neural network ( C onvolutionary N eural N The number of hidden layers ML is not limited and can be more than one, for example, 2 or more, 10 or more, 50 or more, or 70 or more. In this embodiment, a total of 70 hidden layers ML are provided.

[0017] In this embodiment, the trained model X, which is made up of a CNN, has an intermediate layer ML configured by repeating a unit consisting of a convolution layer Conv(n), an activation layer ReLU, and a pooling layer Pool(n) multiple times. In this embodiment, as shown in FIG. 2(b), ten layers, each consisting of two repeats of a convolution layer Conv(n), an activation layer ReLU, a convolution layer Conv(n+1), an activation layer ReLU, and a pooling layer Pool(n+1), are defined as one unit U1 to U7. The intermediate layer ML then repeats the unit seven times and then has a fully connected layer FC (Fc).

[0018] Of the intermediate layer ML, the convolution layer Conv(n) is also called a convolution layer, and performs convolution processing by applying a predetermined filter to feature quantities of the dimension corresponding to the number of channels in the calculation results of the input layer or the previous pooling layer.Then, it outputs feature quantities of the dimension corresponding to the number of channels according to the filter as the calculation results of the convolution processing.

[0019] In the activation layer, the calculation results in the convolution layer are applied to an activation function for processing. In this embodiment, the activation function is ReLU( Re ctified L inear U nits) is used. By using ReLU, the differential value does not saturate, making it easier to increase the number of layers. In addition, instead of ReLU, Leaky ReLU or PReLU ( P arametric Re ctified L inear U nits), GELU ( G aussian E rror L inear U In these cases, negative inputs can be handled, reducing data loss and improving accuracy.

[0020] The pooling layer performs pooling, which is a process for obtaining a feature map that has been downsampled from only the representative value within the local range. Examples of pooling include maximum pooling, which obtains the maximum value within the local range, and average pooling, which obtains the average value within the local range, as the representative value. The fully connected layer FC performs dimensional adjustment with the final output layer OL. In this embodiment, the fully connected layer FC flattens the data to reduce the number of dimensions in order to classify products into good and bad products in the output layer and determine whether they are good or bad.

[0021] Next, the quality determination device 4 shown in Fig. 1 determines the quality of rack bars as products used in industry. The quality determination device 4 includes a determination data acquisition unit 41, a trained model storage unit 42, a quality determination unit 43, and a quality determination output unit 44. The determination data acquisition unit 41 acquires determination data from the detection results of the detection unit 2. In this embodiment, image data captured by the detection unit 2 is acquired as the determination data. The trained model storage unit 42 stores the trained model A described above.

[0022] The pass / fail judgment unit 43 then inputs the judgment data acquired by the judgment data acquisition unit 41 into the trained model A stored in the trained model storage unit 42, thereby obtaining a judgment result on the pass / fail of the product subjected to inspection. The pass / fail judgment output unit 44 outputs the judgment result. The output form of the radius result is not limited, and can be configured, for example, by a display unit that displays the judgment result to notify the operator of the pass / fail judgment result of the product. Furthermore, the inspection object that has been judged to be defective can be labeled or tagged in a way that allows it to be distinguished from inspection objects judged to be pass, so that the inspection object can be reinspected or not shipped.

[0023] In the pass / fail judgment device 4, the judgment data acquisition unit 41 and the pass / fail judgment unit 43 are configured by a calculation device (not shown). The trained model storage unit 42 is configured by a storage device. The pass / fail judgment output unit 44 is configured by a display device or labeling device according to the output mode.

[0024] 1 determines whether the determination data used in the pass / fail determination device 4 is appropriate. The abnormality determination device 5 has a first evaluation data storage unit 51, a second evaluation data acquisition unit 52, a distance calculation unit 53, a threshold value storage unit 54, an abnormality determination unit 55, and a determination result output unit 56. The first evaluation data storage unit 51 stores first evaluation data, which is intermediate data in the trained model A.

[0025] In this embodiment, the first evaluation data storage unit 51 stores the intermediate data as first evaluation data calculated based on calculation results in one or more layers of the hidden layer ML shown in FIGS. 2(a) and 2(b). In this specification, the hidden layer ML used to calculate the first evaluation data is referred to as the evaluation hidden layer MLX. When there are multiple hidden layers ML, there is no limitation on which hidden layer ML to use as the evaluation hidden layer MLX. However, it is preferable that the evaluation hidden layer MLX include a hidden layer ML closer to the input layer IL than the output layer OL. Since the calculation results of the hidden layer ML closer to the input layer IL do not sufficiently distinguish between good products and bad products, there remains an area in the feature space where the calculation results of the training data for good products and the calculation results of the training data for bad products are mixed, as shown in FIG. 3. As shown in Figure 3, the data distance L between the first evaluation data and the second evaluation data described below is easy to calculate in an area where the calculation results of training data for good products and the calculation results of training data for defective products are mixed, so it is suitable for evaluating whether the judgment data deviates from the training data based on the data distance L.

[0026] From this perspective, in this embodiment, the first evaluation data includes the calculation results of the hidden layer ML that are closer to the input layer OL than the center position between the input layer IL and the output layer OL in the entire hidden layer ML shown in Fig. 2(b). Specifically, as shown in Fig. 2(b), the evaluation hidden layer MLX includes one layer selected from each of the first unit U1, third unit U3, and fifth unit U5 on the input layer IL side of the seven units U1 to U7.

[0027] Furthermore, it is preferable that the evaluation hidden layer MLX includes a convolution layer Conv among the hidden layers ML shown in Fig. 2(b). This is because the calculation results of the convolution layer Conv are before pooling and therefore retain more information related to feature quantities, which contributes to improving the accuracy of the determination. When multiple hidden layers ML are used as the evaluation hidden layer MLX, the average value of the calculation results of each hidden layer ML or the maximum or minimum value of the calculation results of the multiple hidden layers ML can be used as the first evaluation data.

[0028] Furthermore, it is preferable that the evaluation hidden layer MLX includes layers that are spaced apart from each other in the hidden layer ML shown in Fig. 2(b). Because the calculation results of adjacent layers in the hidden layer ML tend to be similar to each other, the inter-data distance L between the first evaluation data and the second evaluation data (described later) based on the calculation results of the adjacent layers also becomes similar to each other, and is not effective in evaluating whether the judgment data deviates from the training data.

[0029] From the above perspective, in this embodiment, the evaluation hidden layer MLX includes one convolution layer Conv3 in each of the first unit U1, the third unit U3, and the fifth unit U5 on the input layer IL side, among the seven units U1 to U7 constituting the hidden layer ML. Furthermore, in this embodiment, the first evaluation data is obtained by reducing the dimension corresponding to coordinate information, which is the horizontal and vertical positions of the image, from the calculation results of the evaluation hidden layer MLX. This reduces the amount of calculation required for calculating the distance between data, which will be described later. While there are no limitations on the method for reducing the dimension corresponding to the coordinate information, for example, the dimension corresponding to the coordinate information can be reduced by using the average, maximum, or minimum value of the calculated values ​​for each coordinate in the image as the first evaluation data. As shown in FIG. 3, the first evaluation data includes both training data for non-defective products and training data for defective products as learned data.

[0030] The second evaluation data acquisition unit 52 shown in FIG. 1 acquires second evaluation data based on data during processing of the judgment data in the pass / fail judgment unit 43. The data during processing of the judgment data in the pass / fail judgment unit 43 for acquiring the second evaluation data is a calculation result of the evaluation intermediate layer MLX, which is the same intermediate layer ML as the first evaluation data described above. In this embodiment, the evaluation intermediate layer MLX for acquiring the second evaluation data is one convolution layer Conv3 in each of the first unit U1, third unit U3, and fifth unit U5 on the input layer IL side of the seven units U1 to U7 constituting the intermediate layer ML, as in the case of the first evaluation data described above, in the process when the judgment data is input to the trained model A. Then, as in the case of the first evaluation data described above, the average value of the calculation results of the evaluation intermediate layer MLX is calculated and the dimension corresponding to the coordinate information is reduced, and the result is acquired as the second evaluation data.

[0031] Next, as shown in FIG. 3, the distance calculation unit 53 in the abnormality determination device 5 shown in FIG. 1 calculates the inter-data distance L between the first evaluation data and the second evaluation data in a feature space based on the feature quantities included in the first evaluation data and the second evaluation data. The method for calculating the inter-data distance L is not limited, and it can be calculated as any one of Euclidean distance, Mahalanobis distance, Manhattan distance, and Chebyshev distance in the feature space. Alternatively, the inverse of a probability value calculated using a probability density function of the second evaluation data with respect to the first evaluation data may be calculated as the inter-data distance L between the first evaluation data and the second evaluation data. Note that, in calculating the inter-data distance L, all data included in the first evaluation data may be used. Alternatively, the inter-data distance L may be calculated using a data group obtained by selecting, from the first evaluation data, a plurality of data items that are located close to the second evaluation data in the feature space.

[0032] The threshold value storage unit 54 in the abnormality determination device 5 stores a threshold value for the distance between the data, which is used in the abnormality determination unit 55 described later. The threshold value can be set as appropriate.

[0033] The abnormality determination unit 55 in the abnormality determination device 5 determines whether the data acquisition conditions when the determination data acquisition unit 41 acquired data acquisition are abnormal or not based on the comparison result of the data distance L calculated by the distance calculation unit 53 and the threshold value P.

[0034] The data acquisition conditions when the determination data acquisition unit 41 acquires the determination data include the environmental conditions, test conditions, and composition of the test object when the determination data is acquired. For example, when an image is acquired as the determination data, examples of the data acquisition conditions include environmental conditions such as brightness and temperature when the image was captured, the input sensitivity and placement of the image capture device, and the configuration of the image capture device. Furthermore, when the determination data is related to test data, examples of the data acquisition conditions include the configuration of the test device, the configuration and lot number of the test reagent, etc. Furthermore, the data acquisition conditions can include conditions related to the test object, and for example, the data acquisition conditions may include the lot number of the material constituting the test object.

[0035] In this embodiment, if the inter-data distance L is greater than the threshold P, the abnormality determination unit 55 determines that the data acquisition conditions when the judgment data acquisition unit 41 acquired the data acquisition are abnormal. Also, if the inter-data distance L is not greater than the threshold P, the abnormality determination unit 55 determines that the data acquisition conditions when the judgment data acquisition unit 41 acquired the data acquisition are normal. Note that, in this embodiment, the abnormality determination unit 55 determines that the data acquisition conditions when the judgment data acquisition unit 41 acquired the data are abnormal when a state in which the inter-data distance L is greater than the threshold P continues for a predetermined period of time, that is, when the time period from t1 to t2 in FIG. 4 is reached. Alternatively or additionally, the abnormality determination unit 55 may determine that the data acquisition conditions when the judgment data acquisition unit 41 acquired the data are abnormal when a state in which the inter-data distance L is greater than the threshold P is detected a predetermined number of times in succession.

[0036] The determination result output unit 56 in the abnormality determination device 5 outputs the determination result of the abnormality determination unit 55. The manner in which the determination result is output is not limited, and for example, whether the data acquisition conditions are abnormal or normal can be displayed on a display unit (not shown). Alternatively, the test objects determined to have abnormal data acquisition conditions can be labeled or tagged in a way that makes them distinguishable, so that the test objects determined to have abnormal data acquisition conditions can be retested or prevented from being shipped.

[0037] In the abnormality determination device 5, the first evaluation data storage unit 51 and the threshold value storage unit 54 are configured by a calculation device (not shown). Also, the second evaluation data acquisition unit 52, the distance calculation unit 53, and the abnormality determination unit 55 are configured by a calculation device (not shown). Also, the determination result output unit 56 is configured by a display device or a labeling device according to the output mode.

[0038] 2.Inspection flow of Industrial Inspection System 1 Next, the inspection flow of the industrial inspection system 1 of this embodiment will be described with reference to Figs. 5 and 6. As shown in Fig. 5, first, in step S1, a rack bar to be inspected is prepared. Next, in step S2, magnetic powder is scattered on the inspection object to start the magnetic particle inspection. After that, in step S3, the detection unit 2 takes an image of the inspection object. Then, the judgment data acquisition unit 41 acquires the image shown in Fig. 6 as judgment data.

[0039] 5, the pass / fail judgment unit 43 inputs the judgment data to the trained model A. Then, in step S5 as a first parallel process, a pass / fail judgment result is obtained as an output from the trained model A to which the judgment data has been input. Then, in step S6, the pass / fail judgment result is displayed by the pass / fail judgment output unit 44.

[0040] On the other hand, after step S4 shown in Fig. 5 as the second parallel processing, the process proceeds to step S7. In step S7, the second evaluation data acquisition unit 52 calculates and acquires second evaluation data based on data in the middle of processing after the judgment data is input to the trained model A. The second evaluation data is obtained by calculating the average value of the calculation results of each evaluation intermediate layer MLX in the trained model A shown in Fig. 2(b) and reducing the dimension corresponding to the coordinate information.

[0041] 5, the distance calculation unit 53 calculates the inter-data distance L in the feature space between the first evaluation data and the second evaluation data stored in the first evaluation data storage unit 51. Note that the first evaluation data is obtained by calculating the average value of the calculation results of the evaluation intermediate layer MLX and reducing the dimension corresponding to the coordinate information.

[0042] 5, the abnormality determination unit 55 determines whether the inter-data distance L between the first evaluation data and the second evaluation data is greater than the threshold value P stored in the threshold value storage unit 54. If it is determined in step S9 that the inter-data distance L is not greater than the threshold value P, the process proceeds to No in step S9. Then, in step S10, the abnormality determination unit 55 determines that the data acquisition conditions when the judgment data acquisition unit 41 acquired the judgment data were normal, and the flow ends.

[0043] On the other hand, if it is determined in step S9 that the inter-data distance L is greater than the threshold value P, the process proceeds to Yes in step S9. Then, in step S11, the abnormality determination unit 55 determines whether the state in which the inter-data distance L is greater than the threshold value P has continued for a predetermined period. If it is determined in step S11 that the state in which the inter-data distance L is greater than the threshold value P has not continued for a predetermined period, the process proceeds to No in step S11. Then, in step S10, the abnormality determination unit 55 determines that the data acquisition conditions when the judgment data acquisition unit 41 acquired the judgment data were normal, and the flow ends.

[0044] Furthermore, if it is determined in step S11 that the state in which the inter-data distance L is greater than the threshold value P has continued for a predetermined period of time, the process proceeds to Yes in step S11. Then, in step S12, the abnormality determination unit 55 determines that the data acquisition conditions under which the judgment data acquisition unit 41 acquired the judgment data are abnormal. Thereafter, in step S13, the determination result output unit 56 displays that the data acquisition conditions under which the judgment data acquisition unit 41 acquired the judgment data were abnormal, notifies the operator of the determination result, and ends this flow. Based on this notification, the operator can begin work to check the data acquisition conditions.

[0045] In the flow shown in FIG. 5, the flow is terminated after determining in step S10 that the data acquisition conditions are normal. Alternatively, the abnormality determination unit 55 may determine in step S10 that the data acquisition conditions were normal when the judgment data acquisition unit 41 acquired the judgment data, and then the judgment result output unit 56 may display that the data acquisition conditions were normal when the judgment data acquisition unit 41 acquired the judgment data, and then the flow may be terminated.

[0046] 3. Effects of Industrial Inspection System 1 Next, the effects of the industrial inspection system 1 of this embodiment will be described in detail. The industrial inspection system 1 of this embodiment acquires judgment data from an inspection object, and calculates the inter-data distance L between second evaluation data, which is judgment processing data in the middle of processing the judgment data to determine the quality of the product based on a pre-stored trained model A, and first evaluation data, which is intermediate data in the trained model A. Then, based on the result of comparing the inter-data distance L with a threshold value P, it is determined whether the data acquisition conditions when the judgment data acquisition unit 41 acquired the judgment data were abnormal.

[0047] As a result, if there is a discrepancy between the judgment data and the learned data used to create the trained model A, it can be detected as an anomaly in the data acquisition conditions when the judgment data acquisition unit 41 acquired the judgment data. This can suppress erroneous judgments, and also reduce the number of dimensions and the amount of calculations compared to when the judgment data is directly compared with the learned data (training data). Furthermore, since there is no need for a human to visually judge products judged as defective or unclassifiable by the system or to re-inspect them using a different quality judgment device, the first-pass rate of the production line can be improved. As a result, costs can be reduced and judgment accuracy can be improved.

[0048] In this embodiment, the trained model A constitutes a neural network. The first evaluation data is calculated based on the calculation results of an evaluation intermediate layer MLX, which is at least one of the intermediate layers ML between the input layer IL and the output layer OL of the neural network. The second evaluation data is calculated based on the calculation results of the evaluation intermediate layer MLX, which is the intermediate layer ML from which the first evaluation data was obtained by inputting judgment data into the neural network. Furthermore, the inter-data distance L is obtained as the distance between the first evaluation data and the second evaluation data in a feature space in which the features of the judgment data are expressed as dimensional elements. As a result, both the first evaluation data and the second evaluation data are based on the calculation results of the same evaluation intermediate layer MLX in the neural network, which reliably reduces the number of dimensions and reduces the calculation load for determining whether the data acquisition conditions of the judgment data are abnormal.

[0049] In this embodiment, the evaluation hidden layer MLX includes a hidden layer ML located upstream of the center between the input layer IL and the output layer OL among the multiple hidden layers ML in the neural network. This makes it possible to more accurately evaluate whether the judgment data deviates from the training data based on the inter-data distance L between the first evaluation data and the second evaluation data. Furthermore, even if an unexpected abnormality occurs in the data acquisition conditions, this can be detected.

[0050] In this embodiment, the neural network is a convolutional neural network, and the evaluation intermediate layer MLX from which the first evaluation data is acquired includes a convolutional layer in the convolutional neural network, thereby enabling more accurate evaluation of whether the judgment data deviates from the training data based on the inter-data distance L.

[0051] In this embodiment, the distance calculation unit 53 calculates any one of Euclidean distance, Mahalanobis distance, Manhattan distance, and Chebyshev distance in the feature space as the inter-data distance L, or calculates the inter-data distance L based on a probability value calculated using a probability density function of the second evaluation data with respect to the first evaluation data. This makes it possible to calculate the inter-data distance L easily and with high accuracy.

[0052] In this embodiment, the judgment data acquisition unit 41 acquires image data of the inspection target as judgment data. This makes it possible to judge whether the inspection target is good or bad and whether the data acquisition conditions for the judgment data are abnormal based on the image data.

[0053] In this embodiment, the image data is image data of the appearance of the inspection object, which makes it possible to determine whether the inspection object is good or bad and whether the data acquisition conditions for the determination data are abnormal based on the image data of the appearance.

[0054] In this embodiment, the first evaluation data storage unit 51 stores the intermediate data with the number of dimensions related to coordinate information reduced as the first evaluation data, and the second evaluation data acquisition unit 52 acquires the intermediate data with the number of dimensions related to coordinate information reduced as the second evaluation data during the processing of the judgment data, thereby further reducing the calculation load.

[0055] Furthermore, in this embodiment, the abnormality determination unit 55 determines that the data acquisition conditions when the judgment data acquisition unit 41 acquired the judgment data are abnormal if the inter-data distance L is greater than the threshold value P, and determines that the data acquisition conditions when the judgment data acquisition unit 41 acquired the judgment data are normal if the inter-data distance L is not greater than the threshold value P. This allows the determination of whether the data acquisition conditions when the judgment data acquisition unit 41 acquired the judgment data are abnormal to be made with high accuracy.

[0056] Furthermore, in this embodiment, if the state in which the inter-data distance L is greater than the threshold value P continues for a predetermined period of time, or if the state in which the inter-data distance L is greater than the threshold value P is detected a predetermined number of times in succession, the abnormality determination unit 55 determines that the data acquisition conditions when the determination data acquisition unit 41 acquired the determination data are abnormal. This makes it possible to eliminate cases in which the inter-data distance L becomes greater than the threshold value P due to an unexpected factor not attributable to the data acquisition conditions, and improves the accuracy of detecting abnormalities in the data acquisition conditions.

[0057] As described above, according to this embodiment, it is possible to provide an industrial inspection system 1 that reduces costs and improves determination accuracy.

[0058] The present invention is not limited to the above-described embodiment, and can be applied to various embodiments without departing from the spirit of the present invention. [Explanation of symbols]

[0059] 1. Industrial Inspection Systems 2. Detection unit 3 Learning processing device 4. Good / bad judgement device 5 Abnormality determination device 41 Judgment data acquisition unit 42 Trained model memory 43 Good / bad judgement section 44 Pass / fail judgment output section 51 First evaluation data storage unit 52 Second evaluation data acquisition unit 53 Distance calculation section 54 Threshold memory unit 55 Abnormality determination section 56 Judgment result output unit

Claims

1. An industrial inspection system including a quality determination device that determines the quality of industrially used products, and an abnormality determination device that determines an abnormality in a data acquisition condition of the determination data used for the quality determination, The quality determination device is a judgment data acquisition unit that acquires the judgment data from the product as an inspection target; a trained model storage unit that stores a trained model created in advance by machine learning the judgment data of the product and the pass / fail judgment result of the product; a quality determination unit that determines quality based on the determination data of the inspection target and the trained model, The abnormality determination device a first evaluation data storage unit in which first evaluation data based on intermediate data in the trained model is stored; a threshold value storage unit that stores a threshold value related to the inter-data distance from the first evaluation data; a second evaluation data acquisition unit that acquires second evaluation data based on data during processing of the judgment data in the pass / fail judgment unit; a distance calculation unit that calculates a data distance between the first evaluation data and the second evaluation data; an abnormality determination unit that determines whether or not the data acquisition conditions when the judgment data acquisition unit acquired the judgment data are abnormal based on a comparison result obtained by comparing the inter-data distance calculated by the distance calculation unit with the threshold value; a determination result output unit that outputs a determination result by the abnormality determination unit, Industrial inspection systems.

2. The above trained model constitutes a neural network, the first evaluation data is calculated based on at least one calculation result of an intermediate layer between an input layer and an output layer in the neural network, the second evaluation data is calculated based on a calculation result in an evaluation intermediate layer, which is the intermediate layer from which the first evaluation data is obtained by inputting the judgment data into the neural network, 2. The industrial inspection system according to claim 1, wherein the inter-data distance is acquired as a distance between the first evaluation data and the second evaluation data in a feature space in which the feature of the judgment data is represented as a dimensional element.

3. 3. The industrial inspection system of claim 2, wherein the evaluation intermediate layer includes an intermediate layer located upstream of a center position between the input layer and the output layer among the plurality of intermediate layers in the neural network.

4. The neural network is a convolutional neural network, 4. The industrial inspection system of claim 2, wherein the evaluation intermediate layer from which the first evaluation data is obtained comprises a convolution layer in the convolutional neural network.

5. 5. The industrial inspection system according to claim 2, wherein the distance calculation unit calculates, as the inter-data distance, one of Euclidean distance, Mahalanobis distance, Manhattan distance, and Chebyshev distance in the feature space, or calculates the inter-data distance based on a probability value calculated using a probability density function of the second evaluation data with respect to the first evaluation data.

6. 6. The industrial inspection system according to claim 1, wherein the determination data acquisition unit acquires image data of the inspection target as the determination data.

7. 7. The industrial inspection system of claim 6, wherein the image data is image data of the appearance of the inspection object.

8. the first evaluation data storage unit stores, as the first evaluation data, the intermediate data from which dimensions related to coordinate information have been reduced, The second evaluation data acquisition unit acquires, as the second evaluation data, data obtained by reducing dimensions related to coordinate information from data during processing of the judgment data.

8. An industrial inspection system according to claim 6 or 7.

9. The industrial inspection system of any one of claims 1 to 8, wherein the abnormality judgment unit judges that the data acquisition conditions when the judgment data acquisition unit acquired the judgment data are abnormal if the data distance is greater than the threshold value, and judges that the data acquisition conditions when the judgment data acquisition unit acquired the judgment data are normal if the data distance is not greater than the threshold value.

10. 10. The industrial inspection system of claim 9, wherein the abnormality determination unit determines that the data acquisition conditions when the judgment data acquisition unit acquired the judgment data are abnormal if the state in which the data distance is greater than the threshold value continues for a predetermined period of time, or if the state in which the data distance is greater than the threshold value is detected a predetermined number of times in succession.

Images