Machine learning program, machine learning method, inference program, inference method, and information processing apparatus

Abstract

To provide a machine learning program, a machine learning method, an inference program, an inference method, and an information processing apparatus that improve the performance of quality evaluation using ultrasonic signals.SOLUTION: The machine learning program causes a computer to execute processing of: acquiring a training data set including a plurality of pieces of training data in which a plurality of A-mode ultrasonic signals obtained for each of different positions of an object are associated with evaluation results for the object; for each training data included in the training data set, weighting a plurality of pieces of feature data acquired on the basis of the plurality of A-mode ultrasonic signals included in the training data by using a weighting model 152 that weights the feature data; acquiring inference results of evaluation for the object by inputting the plurality of pieces of weighted feature data to a classification model 154 that outputs inference results of evaluation in response to input of the plurality of pieces of feature data; and training a first machine learning model and a second machine learning model on the basis of the inference results and the evaluation results in the training data.SELECTED DRAWING: Figure 2

Description

Technical Field

[0001] The present invention relates to a machine learning program, a machine learning method, an inference program, an inference method, and an information processing apparatus.

Background Art

[0002] When nondestructively inspecting solids such as concrete structures and frozen substances, low-frequency ultrasonic waves are generally used. In inspection using low-frequency ultrasonic waves, an ultrasonic probe becomes large in order to output low-frequency ultrasonic waves. Therefore, inspection is often performed using an A (Amplitude) mode ultrasonic signal that displays amplitude information on the time axis obtained using a single ultrasonic probe.

[0003] Specifically, ultrasonic waves are transmitted from a single ultrasonic probe in a certain direction to an object, and the ultrasonic pulse reflected by the object and received again by the ultrasonic probe is detected to obtain an A mode ultrasonic signal. A graph showing the A mode ultrasonic signal represents time on the horizontal axis and the intensity of the reflected wave on the vertical axis. From the graph showing the A mode ultrasonic signal, the amplitude and intensity of the reflected wave are obtained, and the internal state of the object to be inspected is obtained using them.

[0004] As a method of handling an ultrasonic probe when acquiring an A mode ultrasonic signal, there are a method in which a person manually presses the ultrasonic probe against an object and a method in which the ultrasonic probe is mechanically pressed against the object. However, in any method, noise easily rides on the A mode ultrasonic signal, and the quality of data varies greatly depending on the state of the contact surface of the ultrasonic probe. Therefore, it is difficult to perform accurate analysis using the obtained A mode ultrasonic signal as it is.

[0005] Conventionally, a method has been proposed in which an ultrasonic probe is continuously applied to the same measurement site and the A-mode ultrasonic signals continuously acquired are averaged. Also, a method has been proposed in which signals acquired by applying an ultrasonic probe to different measurement sites of the same object are averaged. Additionally, a technique has been proposed in which a machine learning model trained using scores for single A-mode signals is used to infer scores for each of a plurality of A-mode ultrasonic signals obtained from different measurement sites of the same phantom, and the average thereof is calculated.

[0006] Note that as an inspection technique using ultrasonic waves, a technique has been proposed in which an ultrasonic inspection device is used to measure ultrasonic tomographic images of tuna, numerical values of volume elasticity, acoustic impedance, attenuation constant, and Doppler shift frequency of ultrasonic tissue characteristics, and quality evaluation is performed based on the measurement results.

Prior Art Documents

Patent Documents

[0007]

Patent Document 1

Summary of the Invention

Problems to be Solved by the Invention

[0008] However, some ultrasonic signals are not represented by images or the like, and in this case, it is often unclear whether an appropriate signal has been obtained. Therefore, the performance of quality evaluation using ultrasonic signals may decrease.

[0009] For example, in the technique of averaging A-mode ultrasonic signals continuously acquired by continuously applying an ultrasonic probe to the same measurement site, since the change in the signal when the probe is continuously applied is negligibly small, it is difficult to improve the performance of quality evaluation. Also, in the technique of averaging signals acquired by applying an ultrasonic probe to different measurement sites of the same object, since the information contained in the A-mode ultrasonic signals obtained varies depending on the measurement site, simply averaging is inappropriate and it is difficult to improve the performance of quality evaluation. Further, in the technique of calculating the average of scores for A-mode ultrasonic signals obtained from a machine learning model, the obtained A-mode ultrasonic signals are likely to contain noise, and there is a risk that the performance of quality evaluation will deteriorate due to being influenced by scores derived from noisy signals.

[0010] The disclosed technology has been made in view of the above, and an object thereof is to provide a machine learning program, a machine learning method, an inference program, an inference method, and an information processing apparatus that improve the performance of quality evaluation using ultrasonic signals.

Means for Solving the Problem

[0011] The machine learning program, machine learning method, inference program, inference method, and information processing apparatus disclosed in the present application, in one aspect, acquire a training data set including a plurality of pieces of training data in which a plurality of A-mode ultrasonic signals obtained for different positions of an object are associated with an evaluation result for the object, and for each piece of training data included in the training data set, perform weighting on a plurality of feature quantity data obtained based on the plurality of A-mode ultrasonic signals included in the training data using a first machine learning model that performs weighting on the feature quantity data, input the plurality of feature quantity data weighted by the first machine learning model into a second machine learning model that outputs an inference result of evaluation in response to an input of the plurality of feature quantity data to obtain an inference result of evaluation for the object, and based on the inference result and the evaluation result in the training data, cause a computer to execute a process of training the first machine learning model and the second machine learning model.

Effect of the Invention

[0012] According to one aspect of the machine learning program, machine learning method, inference program, inference method, and information processing apparatus disclosed in the present application, there is an effect that the performance of quality evaluation using ultrasonic signals can be improved.

Brief Description of the Drawings

[0013]

Figure 1

Figure 2

Figure 3

Figure 4

Figure 5

Figure 6

Figure 7

Modes for Carrying Out the Invention

[0014] Hereinafter, embodiments of the machine learning program, machine learning method, inference program, inference method, and information processing apparatus disclosed in the present application will be described in detail based on the drawings. Note that the machine learning program, machine learning method, inference program, inference method, and information processing apparatus disclosed in the present application are not limited by the following embodiments.

Examples

[0015] FIG. 1 is a block diagram of an information processing apparatus according to an embodiment. The information processing apparatus 1 is connected to ultrasonic probes 21 to 24. In the following description, when not distinguishing each of the ultrasonic probes 21 to 24, it is referred to as "ultrasonic probe 20". The information processing apparatus 1 is a device that measures the internal state of an object using the A-mode ultrasonic signals obtained from the object by the ultrasonic probes 21 to 24. In this embodiment, a frozen tuna is used as the object, and the case where measurements are performed on different measurement parts of the frozen tuna using four ultrasonic probes 21 to 24 will be described. Here, although four ultrasonic probes 21 to 24 are used in this embodiment for measurement, the number of ultrasonic probes 20 is not particularly limited. For example, it is also possible to use eight ultrasonic probes 20, and the accuracy of quality evaluation improves as the number of ultrasonic probes 20 used increases.

[0016] The information processing apparatus 1 includes a measurement control unit 11, a learning execution unit 12, a quality evaluation unit 13, an evaluation notification unit 14, and an inference unit 15. The inference unit 15 includes a feature extraction model 151, a weighting model 152, a data combination unit 153, and a classification model 154. The information processing apparatus 1 has two operation phases: a learning phase and an inference phase. In the learning phase, the information processing apparatus 1 trains and learns the feature extraction model 151, the weighting model 152, and the classification model 154, which are machine learning models used for quality evaluation inference. In the inference phase, the information processing apparatus 1 infers the quality evaluation of a frozen tuna whose correct answer for quality evaluation is unknown using the learned feature extraction model 151, weighting model 152, and classification model 154.

[0017] The measurement control unit 11 preliminarily receives and holds information on ultrasonic irradiation settings such as frequency, power, and direction, which are used for ultrasonic measurement of frozen tuna, from an input device (not shown). The measurement control unit 11 receives an instruction to execute measurement from the input device. Then, the measurement control unit 11 transmits the information on the irradiation settings to each of the ultrasonic probes 21 to 24 attached to the frozen tuna, and causes each of them to irradiate the frozen tuna with ultrasonic waves. Here, in the information processing apparatus 1 according to the present embodiment, since measurement is performed using an A-mode ultrasonic signal, the measurement control unit 11 outputs ultrasonic waves in a low-frequency band of several hundreds of kHz to the ultrasonic probes 21 to 24. For example, the measurement control unit 11 outputs ultrasonic waves of 500 kHz to the ultrasonic probes 21 to 24.

[0018] Thereafter, the measurement control unit 11 acquires an A-mode ultrasonic signal, which is a reflected signal of the ultrasonic wave propagated through the frozen tuna, received by each of the ultrasonic probes 21 to 24. In the learning phase, the measurement control unit 11 outputs the acquired A-mode ultrasonic signals at each measurement site to the learning execution unit 12. Also, in the inference phase, the measurement control unit 11 outputs the acquired A-mode ultrasonic signals at each measurement site to the quality evaluation unit 13.

[0019] The inference unit 15 receives the input of the A-mode ultrasonic signal at each measurement site, and performs inference on the quality evaluation of the frozen tuna using the feature amount extraction model 151, the weighting model 152, the data coupling unit 153, and the classification model 154, and outputs the result of the inferred quality evaluation. The feature amount extraction model 151, the weighting model 152, and the classification model 154 can use, for example, a neural network. FIG. 2 is an overall image diagram of the quality inference process by the inference unit according to the first embodiment.

[0020] As shown in FIG. 2, the feature extraction model 151 receives the input of the A-mode ultrasonic signals at each measurement site of the frozen tuna. In this embodiment, the feature extraction model 151 receives the input of four A-mode ultrasonic signals corresponding to each of the ultrasonic probes 21 to 24. Then, the feature extraction model 151 extracts the features of each input A-mode ultrasonic signal and outputs feature data including the information of the extracted features. Conceptually, this feature is information that summarizes the amplitude information, frequency information, etc. of the A-mode ultrasonic signal. In this embodiment, the feature extraction model 151 outputs four pieces of feature data corresponding to four measurement sites.

[0021] This feature extraction model 151 is an example of the "third machine learning model". That is, for a plurality of A-mode ultrasonic signals included in the training data 200, the feature data of each of the plurality of A-mode ultrasonic signals is generated using the third machine learning model.

[0022] As shown in FIG. 2, the weighting model 152 receives the input of the feature data of the A-mode ultrasonic signals at each measurement site of the frozen tuna extracted by the feature extraction model 151. Then, the weighting model 152 outputs the weight for each feature indicated by the input feature data. Conceptually, the weighting model 152 determines the weight of the feature so that the better the quality of the measurement data, the heavier the weighting. This weighting model 152 is an example of the "first machine learning model".

[0023] The data combining unit 153 receives the input of the feature data from the feature extraction model 151. In addition, the data combining unit 153 receives the input of the weight for each feature indicated by each piece of feature data from the weighting model 152.

[0024] Next, the data combining unit 153 adds the weight for each feature amount determined by the weighting model 152 to each of the feature amounts of the A-mode ultrasonic signals at each measurement site of the frozen tuna included in the feature amount data. For example, the data combining unit 153 performs weighting by adding a weight to the feature amount of the A-mode ultrasonic signal. Then, the data combining unit 153 generates one input data by collecting the weighted feature amounts. For example, the data combining unit 153 calculates the weighted sum of the feature amounts to obtain one input data.

[0025] As shown in FIG. 2, the classification model 154 receives the input of one input data generated by the data combining unit 153. Then, the classification model 154 classifies the quality of the target frozen tuna as good or bad using the input data, makes an inference on the quality evaluation of the frozen tuna, and outputs the result of the quality evaluation.

[0026] This classification model 154 is an example of the "second machine learning model". That is, one input data obtained by collecting a plurality of weighted feature amount data generated by the data combining unit 153 is input to the second machine learning model, and an inference result of the evaluation of the target object is obtained.

[0027] In the learning phase, the learning execution unit 12 receives the input of the A-mode ultrasonic signal at each measurement site of the frozen tuna from the measurement control unit 11. Also, the learning execution unit 12 acquires the correct answer of the quality evaluation for the frozen tuna, which is the target object from which the acquired A-mode ultrasonic signal is obtained. Then, the learning execution unit 12 trains the feature extraction model 151, the weighting model 152, and the classification model 154 using the A-mode ultrasonic signal at each measurement site of the frozen tuna and the correct answer of the quality evaluation of the frozen tuna. The details of the operation of the learning execution unit 12 will be described below.

[0028] FIG. 3 is a diagram for explaining the generation of a training data set in the learning execution unit. The learning execution unit 12 assigns a classification label indicating the correct quality evaluation corresponding to the frozen tuna to each A-mode ultrasonic signal at each measurement site of the frozen tuna obtained by the ultrasonic probes 21 to 24, and generates the training data 200 shown in FIG. 3.

[0029] The learning execution unit 12 acquires a plurality of A-mode ultrasonic signals for each measurement site of the frozen tuna obtained by the ultrasonic probes 21 to 24, and similarly assigns a classification label to each of the plurality of A-mode ultrasonic signals for each measurement site to generate a plurality of training data 200. Then, the learning execution unit 12 collects the generated training data 200 and generates the training data set 210 shown in FIG. 3. In this way, the learning execution unit 12 acquires a training data set 210 including a plurality of training data 200 in which a plurality of A-mode ultrasonic signals obtained for different positions of the object are associated with the evaluation results for the object.

[0030] Next, the learning execution unit 12 executes the following processing for each training data 200 included in the training data set 210. The learning execution unit 12 inputs the A-mode ultrasonic signal included in the training data 200 to the feature extraction model 151. The input A-mode ultrasonic signal is sequentially processed by the feature extraction model 151, the weighting model 152, the data combination unit 153, and the classification model 154. Then, the learning execution unit 12 acquires the quality evaluation result output from the classification model 154 corresponding to the input A-mode ultrasonic signal. Next, the learning execution unit 12 compares the quality evaluation result output from the classification model 154 with the correct quality evaluation included in the training data 200, and adjusts and updates the parameters of the feature extraction model 151, the weighting model 152, and the classification model 154.

[0031] The learning execution unit 12 repeats the training of the above-described feature extraction model 151, weighting model 152, and classification model 154 until a predetermined learning end condition is satisfied. The learning end condition is specified using, for example, an upper limit value of the number of trials, an upper limit time, a convergence state of the inferred result, and the like. When the predetermined learning end condition is satisfied, the learning execution unit 12 ends the training of the feature extraction model 151, weighting model 152, and classification model 154. As a result, trained feature extraction model 151, weighting model 152, and classification model 154 are generated.

[0032] Here, in this embodiment, although the three models of the feature extraction model 151, the weighting model 152, and the classification model 154 are trained together, the training method is not limited to this. For example, a model for which learning has been completed may be used for the feature extraction model 151, and the learning execution unit 12 may train the weighting model 152 and the classification model 154 using the features obtained by inputting the A-mode ultrasonic signal into the feature extraction model 151 and the correct answer of the quality evaluation. In this case, the parameters of the weighting model 152 and the classification model 154 are appropriately adjusted according to the training data 200.

[0033] As described above, the learning execution unit 12 performs the following processing for each piece of training data 200 included in the training data set 210. The learning execution unit 12 performs weighting on a plurality of feature data obtained based on a plurality of A-mode ultrasonic signals included in the training data 200 using a first machine learning model that performs weighting on the feature data. Next, the learning execution unit 12 inputs the plurality of weighted feature data by the first machine learning model into a second machine learning model that outputs an inference result of evaluation according to the input of the plurality of feature data, and obtains an inference result of evaluation for the object. Further, the learning execution unit 12 trains the first machine learning model and the second machine learning model based on the inference result and the evaluation result in the training data 200.

[0034] Returning to FIG. 1, the description will be continued. In the inference phase, the quality evaluation unit 13 receives the input of the A-mode ultrasonic signal at each measurement site of the frozen tuna from the measurement control unit 11. Then, the quality evaluation unit 13 inputs the acquired A-mode ultrasonic signal at each measurement site to the feature extraction model 151. The input A-mode ultrasonic signal is sequentially processed by the feature extraction model 151, the weighting model 152, the data combining unit 153, and the classification model 154. Then, the quality evaluation unit 13 acquires the quality evaluation result corresponding to the input A-mode ultrasonic signal output from the classification model 154. After that, the quality evaluation unit 13 outputs the acquired quality evaluation result to the evaluation notification unit 14.

[0035] The evaluation notification unit 14 receives the input of the quality evaluation result for the frozen tuna of the object from the quality evaluation unit 13. Then, the evaluation notification unit 14 notifies the user of the quality evaluation result by, for example, displaying the quality evaluation result for the frozen tuna of the object on a display device (not shown).

[0036] FIG. 4 is a flowchart of the machine learning process in the learning phase by the information processing apparatus according to the first embodiment. Next, with reference to FIG. 4, the flow of the machine learning process in the learning phase by the information processing apparatus 1 according to the first embodiment will be described.

[0037] The measurement control unit 11 transmits irradiation setting information to each of the ultrasonic probes 21 to 24 attached to the frozen tuna, and irradiates the frozen tuna with ultrasonic waves from each of them. Then, the measurement control unit 11 acquires the A-mode ultrasonic signal, which is the reflection signal of the ultrasonic wave propagated through the frozen tuna, received by each of the ultrasonic probes 21 to 24. Then, the measurement control unit 11 acquires the input of the A-mode ultrasonic signal, which is the reflected wave from the frozen tuna at each measurement site, from the ultrasonic probes 21 to 24 (step S1). The measurement control unit 11 outputs the acquired A-mode ultrasonic signal at each measurement site to the learning execution unit 12.

[0038] The learning execution unit 12 assigns a classification label indicating the correct answer of the quality evaluation corresponding to the frozen tuna to the A-mode ultrasonic signal at each measurement site of the frozen tuna obtained by the ultrasonic probes 21 to 24, and generates training data 200 shown in FIG. 3 (step S2).

[0039] Next, the learning execution unit 12 collects a plurality of pieces of training data 200 to which classification labels are assigned for the A-mode ultrasonic signals at each measurement site, and generates a training data set 210 (step S3).

[0040] Next, the learning execution unit 12 inputs the A-mode ultrasonic signal at each measurement site included in each piece of training data 200 into the feature extraction model 151. The feature extraction model 151 receives the input of the A-mode ultrasonic signal at each measurement site and extracts the feature amount of each A-mode ultrasonic signal at each measurement site (step S4).

[0041] The weighting model 152 receives the input of the feature amount of each A-mode ultrasonic signal at each measurement site extracted by the feature extraction model 151, and calculates the weight for each feature amount (step S5).

[0042] The data combining unit 153 assigns the weight for each feature amount calculated by the weighting model 152 to the feature amount of each A-mode ultrasonic signal at each measurement site extracted by the feature extraction model 151. Next, the data combining unit 153 combines the weighted feature amounts to generate one input data (step S6).

[0043] The classification model 154 receives the input of the input data generated by the data combining unit 153, performs classification, and infers the quality evaluation (step S7).

[0044] Next, the learning execution unit 12 compares the result of the inferred quality evaluation of the quality evaluation of the frozen tuna of the object output from the classification model 154 with the classification label, and adjusts the parameters of the feature extraction model 151, the weighting model 152, and the classification model 154 (step S8).

[0045] After that, the learning execution unit 12 determines whether to end the learning based on whether the learning end condition is reached (step S9). When continuing the learning (step S9: NO), the learning execution unit 12 returns to step S4.

[0046] On the other hand, when ending the learning (step S9: YES), the learning execution unit 12 ends the machine learning and ends the operation of the learning phase of the information processing apparatus 1.

[0047] FIG. 5 is a flowchart of the inference process in the inference phase by the information processing apparatus according to the first embodiment. Next, with reference to FIG. 5, the flow of the inference process in the inference phase by the information processing apparatus 1 according to the first embodiment will be described.

[0048] The measurement control unit 11 transmits irradiation setting information to each of the ultrasonic probes 21 to 24 attached to the frozen tuna, and irradiates the frozen tuna with ultrasonic waves from each of them. After that, the measurement control unit 11 acquires the A-mode ultrasonic signals, which are the reflection signals of the ultrasonic waves propagated through the frozen tuna received by each of the ultrasonic probes 21 to 24. Then, the measurement control unit 11 acquires the input of the A-mode ultrasonic signals, which are the reflected waves from the frozen tuna at each measurement site, from the ultrasonic probes 21 to 24 (step S11). The measurement control unit 11 outputs the acquired A-mode ultrasonic signals at each measurement site to the quality evaluation unit 13.

[0049] Next, the quality evaluation unit 13 inputs the A-mode ultrasonic signals at each measurement site to the feature extraction model 151. The feature extraction model 151 extracts the features of each of the A-mode ultrasonic signals at each measurement site in response to the input of the A-mode ultrasonic signals at each measurement site (step S12).

[0050] The weighting model 152 calculates the weight for each feature in response to the input of the features of each of the A-mode ultrasonic signals at each measurement site extracted by the feature extraction model 151 (step S13).

[0051] The data combining unit 153 assigns weights for each feature amount calculated by the weighting model 152 to the feature amounts of each A-mode ultrasonic signal at each measurement site extracted by the feature amount extraction model 151. Next, the data combining unit 153 combines the weighted feature amounts to generate one input data (step S14).

[0052] The classification model 154 receives the input of the input data generated by the data combining unit 153, performs classification, and infers quality evaluation (step S15).

[0053] The evaluation notification unit 14 receives the input of the result of the quality evaluation of the frozen tuna as the object from the quality evaluation unit 13. Then, the evaluation notification unit 14 notifies the user of the result of the quality evaluation by causing the result of the quality evaluation of the frozen tuna as the object to be displayed on the display device or the like (step S16).

[0054] As described above, the information processing apparatus according to the present embodiment trains and learns a feature amount extraction model, a weighting model, and a classification model using A-mode ultrasonic signals obtained from different measurement sites and classification labels representing the correct answers of quality evaluation. Then, the information processing apparatus performs inference of quality evaluation on the A-mode ultrasonic signals obtained from different measurement sites using the learned feature amount extraction model, weighting model, and classification model.

[0055] When measuring using an ultrasonic probe, if the probe is not in good contact with the object, the quality of the data deteriorates. In particular, in the case of A-mode signals, unlike images obtained from B-mode ultrasonic signals or other audio data, it is difficult to determine the quality of the data, and there is a risk that the data quality of the data used for learning and inference is poor. Thus, if the data quality is poor, it is considered that sufficient information for quality evaluation has not been obtained, and the performance of quality evaluation deteriorates. In this regard, the information processing apparatus according to the present embodiment can perform quality evaluation by suppressing the influence of low-quality A-mode ultrasonic data including noise among A-mode ultrasonic data obtained from different measurement sites and emphasizing high-quality A-mode ultrasonic data. Therefore, the information processing apparatus according to the present embodiment can perform analysis that removes the influence of noise and low-quality signals when using A-mode ultrasonic signals, and can improve the performance of quality evaluation using A-mode ultrasonic signals.

[0056] Also, by using A-mode ultrasonic signals for each different measurement site, it is possible to obtain information suitable for inspections according to the measurement site. For example, in the case of tuna, it is possible to obtain information corresponding to each part with or without internal organs. Also in this regard, it is possible to improve the performance of quality evaluation using A-mode ultrasonic signals.

Example

[0057] Next, Example 2 will be described. The information processing apparatus 1 according to the present embodiment is also represented by the block diagram of FIG. 1. The information processing apparatus 1 according to Example 2 has a feature extraction model 151 for each measurement site. In the following description, the operations of each part similar to those in Example 1 may be omitted.

[0058] FIG. 6 is an overall image diagram of quality inference processing by the inference unit according to Example 2. The feature extraction model 151 according to the present embodiment has feature extraction models 151A to 151D corresponding to each measurement site. These feature extraction models 151A to 151D correspond to an example of "a plurality of individual models".

[0059] The feature extraction models 151A to 151D extract the feature quantities of the A-mode ultrasonic signals acquired at different measurement sites of the frozen tuna to generate feature quantity data. For example, the feature extraction model 151A extracts the feature quantities of the A-mode ultrasonic signals acquired by the ultrasonic probe 21. Also, the feature extraction model 151B extracts the feature quantities of the A-mode ultrasonic signals acquired by the ultrasonic probe 22. Also, the feature extraction model 151C extracts the feature quantities of the A-mode ultrasonic signals acquired by the ultrasonic probe 23. Also, the feature extraction model 151D extracts the feature quantities of the A-mode ultrasonic signals acquired by the ultrasonic probe 24. That is, the process of generating the feature quantity data includes a process of generating the feature quantity data for each of the plurality of A-mode ultrasonic signals using an individual model corresponding to the position where each A-mode ultrasonic signal was obtained.

[0060] The weighting model 152 receives the input of the feature quantities of the A-mode ultrasonic signals at each measurement site of the frozen tuna extracted by the feature extraction models 151A to 151D. Next, the weighting model 152 calculates and outputs the weights corresponding to the respective feature quantities.

[0061] The data combining unit 153 weights each of the feature quantities of the A-mode ultrasonic signals at each measurement site of the frozen tuna obtained by the feature extraction model 151, and combines the weighted feature quantities to generate one input data.

[0062] The classification model 154 uses the one input data generated by the data combining unit 153 to classify the quality of the target frozen tuna as good or bad, makes an inference on the quality evaluation of the frozen tuna, and outputs the result of the inferred quality evaluation.

[0063] The learning execution unit 12 assigns classification labels to the A-mode ultrasonic signals obtained from each measurement site to create training data 200, collects the created training data 200, and creates a training data set 210. Then, the learning execution unit 12 uses the training data set 210 to train the feature extraction models 151A to 151D. Here, as the configuration of the feature extraction models 151A to 151D, two configurations are conceivable.

[0064] One is a configuration in which different independent feature extraction models 151A to 151D are used for feature extraction of the A-mode ultrasonic signals for each measurement site. In this case, the learning execution unit 12 performs training for each of the feature extraction models 151A to 151D using the A-mode ultrasonic signals obtained at the corresponding measurement sites. Thereby, the learning execution unit 12 can perform learning for extracting features specialized for the measurement sites corresponding to each of the feature extraction models 151A to 151D.

[0065] This configuration is effective when the differences in the A-mode ultrasonic signals obtained for each measurement site are large. However, as the training data 200 used for the learning of each of the feature extraction models 151A to 151D, those including the A-mode ultrasonic signals obtained from the corresponding measurement sites are used.

[0066] Another configuration of the feature extraction models 151A to 151D is a configuration using a common model for all. In this case, the learning execution unit 12 may train each of the feature extraction models 151A to 151D using all of the A-mode ultrasonic signals obtained for each measurement site. As another method, the learning execution unit 12 may repeat sharing parameters among the feature extraction models 151A to 151D after training any one of the feature extraction models 151A to 151D to perform the training of each of the feature extraction models 151A to 151D. In any training method, the learning execution unit 12 can generate the feature extraction models 151A to 151D as a common model.

[0067] This configuration is effective when the differences in the A-mode ultrasonic signals obtained for each measurement site are small. In this case, the learning execution unit 12 can train the feature extraction models 151A to 151D using more training data 200 than in the case of the configuration described above, and overfitting can be suppressed.

[0068] As described above, the information processing apparatus according to the present embodiment provides separate feature extraction models for each measurement site and extracts features. Further, for the feature extraction models corresponding to each measurement site, machine learning models with independent and different configurations can be used, or machine learning models having a common configuration can be used.

[0069] By using independent feature extraction models for extracting the features of the A-mode ultrasonic signals for each measurement site, it is possible to extract features specialized for each measurement site. Further, by using a common feature extraction model for extracting the features of the A-mode ultrasonic signals for each measurement site, it is possible to suppress overfitting. Thus, the information processing apparatus according to the present embodiment can realize flexible feature extraction, perform inference according to the nature of the obtained A-mode ultrasonic signals, and improve the performance of quality evaluation using the A-mode ultrasonic signals.

[0070] (Hardware Configuration) FIG. 7 is a hardware configuration diagram of the information processing apparatus. Next, with reference to FIG. 7, an example of the hardware configuration for realizing each function of the information processing apparatus 1 will be described.

[0071] As shown in FIG. 7, the information processing apparatus 1 includes, for example, a CPU (Central Processing Unit) 91, a memory 92, a hard disk 93, and a network interface 94. The CPU 91 is connected to the memory 92, the hard disk 93, and the network interface 94 via a bus.

[0072] The network interface 94 is an interface for communication between the information processing apparatus 1 and an external apparatus. The network interface 94 relays communication, for example, between an external user terminal apparatus and the CPU 91.

[0073] The hard disk 93 is an auxiliary storage device. The hard disk 93 stores the feature amount extraction model 151, the weighting model 152, and the classification model 154 illustrated in FIG. 1. Further, the hard disk 93 stores programs for realizing the functions of the measurement control unit 11, the learning execution unit 12, the quality evaluation unit 13, the evaluation notification unit 14, and the inference unit 15 illustrated in FIG. 1.

[0074] The memory 92 is a main storage device. The memory 92 can use, for example, a DRAM (Dynamic Random Access Memory).

[0075] The CPU 91 reads out various programs from the hard disk 93, expands them in the memory 92, and executes them. Thereby, the CPU 91 realizes the functions of the measurement control unit 11, the learning execution unit 12, the quality evaluation unit 13, the evaluation notification unit 14, and the inference unit 15 illustrated in FIG. 1.

Explanation of Signs

[0076] 1 Information processing apparatus 11 Measurement control unit 12 Learning execution unit 13 Quality evaluation unit 14 Evaluation notification unit 15 Inference unit 20 - 24 Ultrasonic probe 151, 151A - 151D Feature amount extraction model 152 Weighting model 153 Data combining unit 154 Classification model

Claims

1. Obtain a training data set including a plurality of pieces of training data in which a plurality of A-mode ultrasonic signals obtained for different positions of an object are associated with an evaluation result for the object, For each piece of the training data included in the training data set, Perform weighting on a plurality of feature data obtained based on the plurality of A-mode ultrasonic signals included in the training data using a first machine learning model that performs weighting on the feature data, Input the plurality of pieces of feature data weighted by the first machine learning model into a second machine learning model that outputs an inference result of evaluation according to the input of the plurality of pieces of feature data, and obtain an inference result of evaluation for the object, Train the first machine learning model and the second machine learning model based on the inference result and the evaluation result in the training data, A machine learning program characterized by causing a computer to execute the process.

2. The process of obtaining the inference result of evaluation includes a process of generating one input data by collecting the plurality of pieces of weighted feature data, inputting the input data into the second machine learning model, and obtaining an inference result of evaluation for the object. The machine learning program according to Claim 1.

3. Further cause the computer to execute a process of generating respective feature data of the plurality of A-mode ultrasonic signals using a third machine learning model that generates feature data according to the input of the A-mode ultrasonic signal for the plurality of A-mode ultrasonic signals included in the training data, The process of performing the weighting includes a process of performing weighting on the feature data obtained by the process of generating using the first machine learning model. The machine learning program according to Claim 1, characterized in that.

4. The third machine learning model includes a plurality of individual models corresponding to different positions of the object, The process of generating the feature amount data includes, for each of the plurality of A-mode ultrasonic signals, a process of generating the feature amount data using the individual model corresponding to the position where each A-mode ultrasonic signal is obtained. The machine learning program according to claim 3, characterized in that.

5. For a plurality of feature amount data obtained based on A-mode ultrasonic signals for different positions of an object, weighting is performed using a trained first machine learning model that weights the feature amount data, The weighted plurality of feature amount data is input to a second machine learning model that outputs an inference result of an evaluation for the object in response to an input of the plurality of feature amount data, and an evaluation for the object is inferred. An inference program characterized by causing a computer to execute the process.

6. An information processing apparatus, Obtains a training data set including a plurality of pieces of training data in which a plurality of A-mode ultrasonic signals obtained for different positions of an object are associated with an evaluation result for the object, For each piece of training data included in the training data set, For a plurality of feature amount data obtained based on the plurality of A-mode ultrasonic signals included in the training data, weighting is performed using a first machine learning model that weights the feature amount data, The plurality of feature amount data weighted by the first machine learning model is input to a second machine learning model that outputs an inference result of an evaluation in response to an input of the plurality of feature amount data, and an inference result of an evaluation for the object is obtained, Based on the inference result and the evaluation result in the training data, the first machine learning model and the second machine learning model are trained. A machine learning method characterized by executing the process.

7. An information processing apparatus, For a plurality of feature amount data obtained based on A-mode ultrasonic signals for different positions of an object, perform weighting using a learned first machine learning model that weights the feature amount data, Input the weighted plurality of feature amount data into a second machine learning model that outputs an inference result of an evaluation for the object in response to the input of the plurality of feature amount data, and infer the evaluation for the object An inference method characterized by executing the process.

8. A first machine learning model that weights feature amount data obtained based on A-mode ultrasonic signals, A second machine learning model that outputs an inference result of an evaluation in response to the input of a plurality of feature amount data, Obtain a training data set including a plurality of pieces of training data in which a plurality of A-mode ultrasonic signals obtained for different positions of an object are associated with an evaluation result for the object. For each piece of training data included in the training data set, perform weighting on a plurality of feature amount data obtained based on the plurality of A-mode ultrasonic signals included in the training data using the first machine learning model, input the plurality of feature amount data weighted by the first machine learning model into the second machine learning model to obtain an inference result of an evaluation for the object, and based on the inference result and the evaluation result in the training data, a learning execution unit that trains the first machine learning model and the second machine learning model An information processing apparatus characterized by comprising the above.

9. For a plurality of feature amount data obtained based on A-mode ultrasonic signals for different positions of an object, perform weighting using a learned first machine learning model that weights the feature amount data, and input the weighted plurality of feature amount data into a second machine learning model that outputs an inference result of an evaluation for the object in response to the input of the plurality of feature amount data, and an inference unit that infers the evaluation for the object An information processing apparatus characterized by comprising the above.

Images