Information processing system, information processing device, information processing method, and program

Abstract

Provided are an information processing system, information processing device, information processing method, and program capable of reducing labor and costs required for analyzing physical quantity data. This information processing system includes: a requesting means which performs request processing for making a request, to a large language model, to generate interpretation information in which a physical phenomenon represented by a physical quantity is interpreted on the basis of physical quantity data related to a calculation result or a detection result of the physical quantity; and an output control means for acquiring the interpretation information generated by the large language model in accordance with the request processing, and causing a prescribed output means to output an interpretation result of the physical phenomenon on the basis of the interpretation information.

Description

Information processing system, information processing device, information processing method and program

[0001] The present invention relates to an information processing system, an information processing device, an information processing method, and a program.

[0002] In recent years, analysis of large-scale physical quantity data related to physical phenomena has become an important element in scientific investigations and research. Because analyzing such physical quantity data requires specialized skills and knowledge, attempts have been made to develop simpler analysis methods and improve analysis methods. One known method for analyzing physical quantity data is a method that uses a learning model trained by machine learning to output estimation results or analysis results for unknown physical quantity data (e.g., Patent Document 1).

[0003] Japanese Patent Application Laid-Open No. 2018-147280

[0004] However, the above-mentioned conventional techniques have the problem that a large amount of learning data and large-scale computational resources must be prepared for the design of a learning model and machine learning, and the amount of effort and cost required for analyzing physical quantity data is large.

[0005] An object of the present invention is to provide an information processing system, an information processing device, an information processing method, and a program that can reduce the effort and cost required to analyze physical quantity data.

[0006] In order to achieve the above object, the invention of an information processing system according to claim 1 comprises: a request means for performing request processing to request a large-scale language model to generate interpretation information that interprets a physical phenomenon represented by a physical quantity based on physical quantity data relating to a detection result or a calculation result of the physical quantity; and an output control means for acquiring the interpretation information generated by the large-scale language model in response to the request processing, and for causing a predetermined output means to output an interpretation result of the physical phenomenon based on the interpretation information.

[0007] The invention of claim 2 is the information processing system of claim 1, wherein the request means generates a prompt in a predetermined format that includes at least the physical quantity data, which is an input sentence for instructing the large-scale language model to perform an action in the request processing, and inputs the prompt to the large-scale language model to request the large-scale language model to generate the interpretation information.

[0008] The invention of claim 3 is the information processing system of claim 1, wherein the output control means acquires a response sentence from the large-scale language model as the interpretation information, converts the response sentence into result information according to a predetermined template related to the interpretation of the physical phenomenon, and outputs the result information to the output means as the interpretation result.

[0009] According to a fourth aspect of the present invention, in the information processing system of the second aspect, the requesting means generates the prompt, which further includes premise information for interpreting the physical phenomenon based on the physical quantity data, and inputs the generated prompt to the large-scale language model in the request processing.

[0010] In a fifth aspect of the present invention, in the information processing system according to the fourth aspect, the premise information includes information on preconditions related to detection or calculation of the physical quantity.

[0011] The invention described in claim 6 is the information processing system described in claim 4, wherein the premise information includes sample data consisting of example data that is an example of the physical quantity data and information resulting from interpreting the physical phenomenon based on the example data.

[0012] The invention of claim 7 is the information processing system of claim 6, wherein the request means repeatedly performs the request processing on the large-scale language model for certain physical quantity data, and in the second or subsequent request processing, inputs to the large-scale language model the prompt including sample data that has not been input to the large-scale language model since the initial request processing, out of the plurality of sample data that are different from each other.

[0013] The invention of claim 8 is the information processing system of claim 1, further comprising a re-learning means for re-learning the large-scale language model using training data including example data that is an example of the physical quantity data and information resulting from interpretation of the physical phenomenon based on the example data, before the requesting means performs the requested processing.

[0014] The invention of claim 9 is the information processing system of claim 1, wherein the request means performs the requested processing multiple times based on multiple pieces of physical quantity data that are different from each other, and the output control means performs statistical analysis on multiple pieces of interpretation information obtained from the large-scale language model in response to the multiple processing requests using a predetermined statistical method, and outputs the result of the statistical analysis to the output means as the interpretation result.

[0015] The invention of claim 10 is an information processing system of claim 6, wherein the request means identifies corresponding example data corresponding to the physical quantity data from among a plurality of mutually different example data, and in the request processing, generates the prompt including the identified corresponding example data and inputs it to the large-scale language model, and the corresponding example data is example data corresponding to vectorized example data, among a plurality of vectorized example data obtained by vectorizing the plurality of example data, that satisfies a predetermined similarity condition between the vectorized physical quantity data obtained by vectorizing the physical quantity data.

[0016] The invention described in claim 11 is the information processing system described in claim 4, wherein the physical quantity data is data relating to a detection result of the physical quantity, and the requesting means adjusts the premise information to be used for the next requested processing based on a comparison result between the interpretation result and the actual physical phenomenon for which the physical quantity data was detected.

[0017] The invention of claim 12 is the information processing system of claim 8, wherein the physical quantity data is data relating to a detection result of the physical quantity, and the re-learning means re-learns the large-scale language model using the training data that reflects a comparison result between the interpretation result and the actual physical phenomenon in which the physical quantity data was detected.

[0018] In order to achieve the above object, the invention of an information processing device according to claim 13 comprises: a request means for performing request processing to request a large-scale language model to generate interpretation information that interprets a physical phenomenon represented by a physical quantity based on physical quantity data relating to a detection result or a calculation result of the physical quantity; and an output control means for acquiring the interpretation information generated by the large-scale language model in response to the request processing, and for causing a predetermined output means to output an interpretation result of the physical phenomenon based on the interpretation information.

[0019] Furthermore, in order to achieve the above object, the invention of an information processing method according to claim 14 is an information processing method executed by a computer, comprising: a request step of performing request processing to request a large-scale language model to generate interpretation information that interprets a physical phenomenon represented by a physical quantity, based on physical quantity data relating to a detection result or a calculation result of the physical quantity; and an output control step of acquiring the interpretation information generated by the large-scale language model in response to the request processing, and causing a predetermined output means to output an interpretation result of the physical phenomenon based on the interpretation information.

[0020] Furthermore, in order to achieve the above object, the program invention described in claim 15 causes a computer to function as: a request means for performing request processing that requests a large-scale language model to generate interpretation information that interprets a physical phenomenon represented by a physical quantity based on physical quantity data related to a detection result or a calculation result of the physical quantity; and an output control means for acquiring the interpretation information generated by the large-scale language model in response to the request processing, and outputting the interpretation result of the physical phenomenon based on the interpretation information to a predetermined output means.

[0021] According to the present invention, it is possible to reduce the effort and cost required for analyzing physical quantity data.

[0022] 1 is a diagram showing the configuration of an information processing system. FIG. 1 is a block diagram showing the main functional configuration of a server. FIG. 1 is a block diagram showing the main functional configuration of an inspection device. FIG. 1 is a block diagram showing the main functional configuration of a user terminal. FIG. 1 is a block diagram showing the main functional configuration of an administrator terminal. FIG. 2 is a schematic diagram showing an example of detection results of magnetic flux density when there is a fracture in the steel material and when there is no fracture. FIG. 2 is a flowchart showing the control procedure for data analysis processing when Few-shot-learning is used. FIG. 3 is a diagram showing an example of measurement data. FIG. 4 is a diagram showing an example of a system prompt. FIG. 5 is a diagram showing an example of a user prompt. FIG. 6 is a diagram showing an example of a response sentence. FIG. 7 is a diagram showing example of result information. FIG. 8 is a flowchart showing the control procedure for result output processing. FIG. 9 is a flowchart showing the control procedure for data analysis processing when Chain-of-Though is used. FIG. 10 is a diagram showing an example of a prompt generated in step S306. FIG. 11 is a flowchart showing another control procedure for data analysis processing when Chain-of-Though is used. FIG. 12 is a flowchart showing the control procedure for data analysis processing when Fine-tuning is used. FIG. 13 is a flowchart showing the control procedure for data analysis processing when semantic search is used.

[0023] Hereinafter, an embodiment of the present invention will be described with reference to the drawings.

[0024] (Configuration of Information Processing System) FIG. 1 is a diagram illustrating the configuration of an information processing system 1. The information processing system 1 includes a server 10 (information processing device), an inspection device 20, a user terminal 30, an administrator terminal 40, an analysis DB (database) 50, a vector DB (database) 60, and a large-scale language model 70. Hereinafter, the large-scale language model 70 will be abbreviated as "LLM (Large Language Models) 70." The devices constituting the information processing system 1 are communicatively connected via a network N. The network N may be, for example, the Internet, but is not limited thereto. The information processing system 1 analyzes measurement data (physical quantity data) obtained by detecting physical quantities using the inspection device 20 and outputs interpretation results of physical phenomena represented by the measurement data to a display unit 34 of the user terminal 30. The information processing system 1 uses the LLM 70 to interpret the physical phenomena represented by the measurement data. The server 10 controls the overall operation of each device in the information processing system 1 involved in analyzing the measurement data.

[0025] In this embodiment, the information processing system 1 is used for non-destructive inspection of a bridge. In this example, the inspection device 20 detects the magnetic flux density (physical quantity) of the magnetic field generated by steel materials in the bridge to obtain measurement data. Based on this measurement data, the information processing system 1 uses a magnetic leakage flux method to interpret whether or not fracture has occurred in the steel materials (physical phenomenon). The information processing system 1 displays the interpretation results on the display unit 34 of the user terminal 30.

[0026] 2 is a block diagram showing the main functional configuration of the server 10. The server 10 includes a central processing unit (CPU) 11, a random access memory (RAM) 42, a storage unit 13, and a communication unit 14. The components of the server 10 are connected via a communication path such as a bus. The server 10 may further include an operation unit, a display unit, etc. that are used by an administrator of the server 10.

[0027] The CPU 11 is a processor that controls the operation of each part of the server 10 by executing various processes in accordance with the program 131 stored in the storage unit 13. The CPU 11 functions as a request means and an output control means by executing various processes in accordance with the program 131. The RAM 12 provides a working memory space for the CPU 11 and stores temporary data.

[0028] The storage unit 13 is a non-transitory recording medium readable by the CPU 11 as a computer. The storage unit 13 is configured with a hard disk drive (HDD) or a solid state drive (SSD), etc. In addition to a program 131, the storage unit 13 stores or temporarily stores the following data: measurement data 132, premise information 133, prompt 134, response sentence 135, and result information 136. The premise information 133 includes sample data 1331 and prerequisite information 1332. The prompt 134 includes a system prompt 1341 and a user prompt 1342. The contents of each data will be described later.

[0029] The communication unit 14 performs communication operations in accordance with a predetermined communication standard. The communication unit 14 transmits and receives data to and from each device in the information processing system 1 via the network N through this communication operation.

[0030] 3 is a block diagram showing the main functional configuration of the inspection device 20. The inspection device 20 includes a CPU 21, a RAM 22, a storage unit 23, a detection unit 24, an operation unit 25, and a communication unit 26. The various units of the inspection device 20 are connected via a communication path such as a bus. The inspection device 20 is used by a user who performs non-destructive inspection of bridges. The inspection device 20 is a portable device. The user places the inspection device 20 against a position on the bridge where steel material is present inside and performs a predetermined operation on the operation unit 25, thereby causing the detection unit 24 to detect the magnetic flux density near the steel material.

[0031] The CPU 21 is a processor that controls the operation of each part of the inspection device 20 by executing various processes in accordance with a program 231 stored in the storage unit 23. The RAM 22 provides a working memory space for the CPU 21 and stores temporary data.

[0032] The storage unit 23 is configured by a nonvolatile memory such as a flash memory, etc. The storage unit 23 stores a program 231 and measurement data 232.

[0033] The detection unit 24 includes multiple magnetic sensors that detect magnetic flux density. Specifically, the detection unit 24 includes ten line sensors, each of which has 54 magnetic sensors linearly arranged at a predetermined interval. The predetermined interval may be, for example, 1 cm. The one-dimensional arrangement direction of the magnetic sensors in each line sensor corresponds to the X direction in an XYZ Cartesian coordinate system. The detection unit 24 includes five line sensors positioned at different positions in the Y direction. These five line sensors are arranged in two rows in the Z direction. That is, the detection unit 24 includes ten line sensors arranged in five rows and two columns when viewed from the X direction. Each device in the information processing system 1 identifies these ten line sensors as "sensor1" to "sensor10." The detection unit 24 acquires analog data of magnetic flux density at 540 locations using a total of 540 magnetic sensors included in the ten line sensors. The detection unit 24 converts the analog data into digital data using an ADC (analog-to-digital converter) (not shown) and outputs the digital data to the CPU 21. The CPU 21 stores the detection result of the magnetic flux density in the storage unit 23 as measurement data 232 .

[0034] The operation unit 25 includes operation buttons etc. that are operated by the user. When an operation is performed on an operation button, the operation unit 25 outputs an operation signal corresponding to the operation to the CPU 21.

[0035] The communication unit 26 performs communication operations in accordance with a predetermined communication standard. The communication unit 26 may be capable of data communication via, for example, mobile data communication or wireless LAN. Through this communication operation, the communication unit 26 transmits and receives data to and from the server 10 via the network N. The measurement data 232 is transmitted to the server 10 via the communication unit 26 and stored as measurement data 132 in the memory unit 13 of the server 10.

[0036] FIG. 4 is a block diagram showing the main functional configuration of the user terminal 30. The user terminal 30 includes a CPU 31, a RAM 32, a storage unit 33, a display unit 34 (output means), an operation unit 35, and a communication unit 36. The various units of the user terminal 30 are connected via a communication path such as a bus. The user terminal 30 is used by a user who checks the results of non-destructive inspection of bridges. This user may be the same as or different from the user who uses the inspection device 20. In this embodiment, the user terminal 30 is a tablet terminal. However, the present invention is not limited to this, and the user terminal 30 may also be a smartphone, a notebook PC, or the like.

[0037] The CPU 31 is a processor that controls the operation of each part of the user terminal 30 by executing various processes in accordance with a program 331 stored in the storage unit 33. The RAM 32 provides a working memory space for the CPU 31 and stores temporary data.

[0038] The storage unit 33 is configured by a nonvolatile memory such as a flash memory, etc. The storage unit 33 stores a program 331.

[0039] The display unit 34 includes a display device such as a liquid crystal display, etc. The display unit 34 displays various information display screens, operation screens, etc. in accordance with control signals and image data input from the CPU 31.

[0040] The operation unit 35 includes operation buttons operated by the user, and a touch panel superimposed on the screen of the display unit 34. When an operation is performed on the operation buttons or the touch panel, the operation unit 35 outputs an operation signal corresponding to the operation to the CPU 31.

[0041] The communication unit 36 ​​performs a communication operation in accordance with a predetermined communication standard. The communication unit 36 ​​transmits and receives data to and from the server 10 via the network N through this communication operation.

[0042] 5 is a block diagram showing the main functional configuration of the administrator terminal 40. The administrator terminal 40 includes a CPU 41, a RAM 42, a storage unit 43, a display unit 44, an operation unit 45, and a communication unit 46. The components of the administrator terminal 40 are connected via a communication path such as a bus. The administrator terminal 40 is used by the administrator of the information processing system 1. The administrator terminal 40 is, for example, a notebook PC or a desktop PC, but is not limited to these and may also be a tablet terminal, a smartphone, or the like.

[0043] The CPU 41 is a processor that controls the operation of each part of the administrator terminal 40 by executing various processes in accordance with a program 431 stored in the storage unit 43. The RAM 42 provides the CPU 41 with a working memory space and stores temporary data.

[0044] The storage unit 43 is configured by an HDD, an SSD, etc. The storage unit 43 stores a program 431 and learning data 432.

[0045] The display unit 44 includes a display device such as a liquid crystal display, etc. The display unit 44 displays various information display screens, operation screens, etc. in accordance with control signals and image data input from the CPU 41.

[0046] The operation unit 45 includes input devices such as a keyboard and a mouse that are operated by a user. When an operation is performed on these input devices, the operation unit 45 outputs an operation signal corresponding to the operation to the CPU 41.

[0047] The communication unit 46 performs communication operations in accordance with a predetermined communication standard, and transmits and receives data to and from each device in the information processing system 1 via the network N through this communication operation.

[0048] 1, the analysis DB 50 is a database that stores, in a predetermined data format, the measurement data 132 and result information 136 transmitted from the server 10. As the analysis DB 50, for example, a relational database or NoSQL can be used.

[0049] The vector DB 60 is a database that saves and manages data in vector format. The vector DB 60 of this embodiment stores, in vector format, measurement data 132, which is a set of data on magnetic flux density detected by the inspection device 20, or example data, which is a sample of the measurement data 132. The vector DB 60 can derive the proximity of vector-format data to each other in vector space, i.e., the similarity.

[0050] The LLM70 (large-scale language model) is a language model pre-trained using massive amounts of data and deep learning technology to assign probabilities to word sequences. In the pre-training using deep learning technology, the neural network model parameters are adjusted so that appropriate probabilities are assigned to word sequences. When a prompt, which is an input sentence instructing the LLM70 to operate, is input, the LLM70 estimates and outputs the word sequence following the prompt, i.e., a response sentence. Specifically, the LLM70 divides the input prompt into the smallest units called tokens and extracts token features. The LLM70 constructs a response by repeatedly deriving the probability of the token following the prompt based on the extracted features. Through this operation, the LLM70 can execute (process) various tasks requested by the prompt. Tasks that the LLM70 can execute include, for example, chat, answering questions, information search, and essay writing.

[0051] The LLM 70 is a language model in which the number of model parameters in deep learning is 100 million or more, or the amount of data used for learning is 1 GB or more or 1 billion tokens or more. Examples of the LLM 70 include BERT, GLaM, LaMDA (all from Google), GPT-3 (registered trademark), and GPT-4 (registered trademark) (all from OpenAI).

[0052] (Principle of non-destructive inspection of bridges) In the information processing system 1 of this embodiment, the presence or absence of fractures in steel materials (presence or absence of fractures) is estimated and interpreted using a magnetic leakage flux method. The magnetic leakage flux method is a method in which a special magnet is applied to the steel materials inside the bridge concrete from the outside to magnetize the steel materials, and a change in magnetic flux density due to fractures in the steel materials is detected.

[0053] FIG. 6 is a schematic diagram showing examples of magnetic flux density detection results when there is a fracture in the steel material and when there is no fracture. The horizontal axis of the graph in FIG. 6 represents the position in a certain direction on the bridge, and the vertical axis represents the magnetic flux density measurement value by the inspection device 20 at each position. Line L1 represents an example of magnetic flux density distribution when there is no fracture in the steel material. Line L2 represents an example of magnetic flux density distribution when there is a fracture in the steel material at position p. When there is a fracture in the steel material, a change occurs in the magnetic flux density distribution before and after the fracture position p compared to when there is no fracture. The presence or absence of this change and the degree of this change can be used to estimate and interpret the presence or absence of a fracture in the steel material. Note that the shape of the magnetic flux density distribution shown by line L2 is just an example, and the shape will vary depending on the direction of the horizontal axis, etc. Note that other methods, such as the magnetic stream method, may be used instead of the leakage flux method.

[0054] (Operation of Information Processing System) Next, the operation of the information processing system 1 will be described. The operation of the information processing system 1 when estimating whether or not a steel material of a bridge has a fracture is outlined as follows. First, the CPU 11 of the server 10 acquires measurement data 132 related to the detection results of magnetic flux density (physical quantity) from the inspection device 20. The CPU 11, as a requesting unit, performs the following request processing to make a request to the LLM 70 based on the acquired measurement data 132. In the request processing, the CPU 11 generates a prompt 134 in a predetermined format that includes at least the contents of the measurement data 132. In the request processing, the CPU 11 inputs the generated prompt 134 to the LLM 70, thereby requesting the LLM 70 to generate interpretation information that interprets whether or not the steel material has a fracture based on the measurement data 132. Here, the interpretation information is a response statement 135 generated by the LLM 70 in response to the prompt 134. Furthermore, interpreting whether or not a steel material has a fracture based on the measurement data 132 corresponds to "interpreting a physical phenomenon represented by a physical quantity." The CPU 11 acquires a response statement 135 as interpretation information from the LLM 70. The CPU 11, which serves as output control means, generates result information 136 indicating whether or not there is a fracture in the steel material based on the acquired response statement 135, and displays this result information 136 on the display unit 34 of the user terminal 30.

[0055] In this way, the information processing system 1 of this embodiment uses the LLM 70 to interpret whether or not there is a fracture in the steel material represented by the measurement data 132. Methods for causing the LLM 70 to make such an interpretation include, for example, the following (1) to (4): (1) Few-shot-learning (2) Chain-of-though (3) Fine-tuning (4) Semantic search Each of the methods (1) to (4) will be described in detail below.

[0056] (1) Few-shot-learning Few-shot-learning is a technique for executing a task by adjusting the operation of the LLM 70 to suit the task by including premise information related to the task in the prompt 134. Few-shot-learning allows the task to be executed without updating the model parameters of the LLM 70, i.e., without relearning. Few-shot-learning uses a prompt 134 that includes a system prompt 1341 that includes premise information 133 and a user prompt 1342 that includes measurement data 132.

[0057] Fig. 7 is a flowchart showing the control procedure for data analysis processing when using few-shot learning. Fig. 7 shows the processes executed by the CPU 11 of the server 10, the CPU 21 of the inspection device 20, and the LLM 70. In Fig. 7, the process flow is depicted with solid lines, and data transmitted and received between devices is depicted with dashed lines.

[0058] When using few-shot learning, premise information 133 regarding the interpretation of whether or not a steel material has a fracture and a template of a system prompt 1341 are stored in advance in the storage unit 13 of the server 10. As shown in FIG. 2 , the premise information 133 may include sample data 1331 and prerequisite information 1332. The sample data 1331 includes example data D1 (see FIG. 9 ) of the measurement data 132 and interpretation information D2 (see FIG. 9 ) that is the result of interpreting whether or not a fracture exists based on the example data D1. The sample data 1331 may include, for example, example data D1 (sound data) for when there is no fracture in the steel material and interpretation information D2 based on the example data D1, i.e., information indicating "no fracture." The other sample data 1331 may include, for example, example data D1 for when there is a fracture in the steel material and interpretation information D2 based on the example data D1, i.e., information indicating "there is a fracture." The example data D1 may be the measurement results of magnetic flux density on a sample bridge that is different from the bridge to be analyzed. The prerequisite information 1332 of the premise information 133 includes information on the measurement method and measurement conditions used by the inspection device 20. In this embodiment, the prerequisite information 1332 includes information such as the fact that the steel material is magnetized and the magnetic flux density is measured, and that the magnetic flux density changes at the position where the steel material is broken. The prerequisite information 1332 may also include information on the number and arrangement of multiple magnetic sensors in the detection unit 24 of the inspection device 20. Furthermore, a template for the user prompt 1342 is stored in advance in the memory unit 13 of the server 10.

[0059] 7, first, the CPU 21 of the inspection device 20 generates measurement data 132 based on the detection result of the magnetic flux density output from the detection unit 24. The CPU 21 converts the measurement data 132 into a JSON format and transmits it to the server 10 (step S101).

[0060] 8 is a diagram showing an example of the measurement data 132. The measurement data 132 includes data on magnetic flux density detected by 54 magnetic sensors for each of the 10 line sensors "sensor1" to "sensor10." The name of the line sensor in the measurement data 132 and the data on the magnetic flux density measured by the corresponding sensor are arranged separated by commas. The measurement data 132 begins on a new line for each line sensor.

[0061] Returning to FIG. 7, the CPU 11 of the server 10 generates a system prompt 1341 including sample data 1331 based on the premise information 133 and template stored in advance in the storage unit 13 (step S102).

[0062] FIG. 9 is a diagram showing an example of a system prompt 1341. The system prompt 1341 in FIG. 9 begins with the following prerequisite information 1332: "As a prerequisite, the bridge is magnetized to determine whether it is broken or intact, and characteristic magnetic field changes occur at the break point. The magnetic field measured by multiple sensors at multiple positions on the same bridge is shown below. Please determine whether the bridge is broken." Following the prerequisite information 1332, the system prompt 1341 includes sample data 1331 related to the integrity data. This sample data 1331 includes interpretation information D2 indicating that there is no break and example data D1 for when there is no break. The example data D1 includes magnetic flux density data detected by 54 magnetic sensors for each of the ten line sensors "sensor1" to "sensor10." The name of the line sensor and the magnetic flux density data measured by the sensor are arranged in the example data D1, separated by commas. The example data D1 is separated into lines for each line sensor. The system prompt 1341 includes sample data 1331 for the above-mentioned healthy data, followed by sample data 1331 for when a break is present. This sample data 1331 includes interpretation information D2 indicating the presence of a break, and example data D1 for when a break is present. Note that either the sample data 1331 for when there is no break or the sample data 1331 for when there is a break may be omitted. The system prompt 1341 includes, following the above-mentioned sample data 1331 for when there is a break, prerequisite information 1332 with the following content regarding the number and arrangement of magnetic sensors: "The horizontal axis has 54 points from 0 cm to 53 cm, and there are 10 waveforms from sensor 1 to 10. Please perform a break determination from here."

[0063] 7 , the CPU 11 of the server 10 generates a user prompt 1342 including the measurement data 132 received in step S101 (step S103). Specifically, the CPU 11 parses the received measurement data 132 in JSON format, converts it into text data, and registers it in the analysis DB 50. The CPU 11 generates the user prompt 1342 based on the text data of the measurement data 132 and a template stored in advance in the storage unit 13.

[0064] 10 is a diagram showing an example of a user prompt 1342. The user prompt 1342 in FIG. 10 includes the text data of the measurement data 132 and the following wording: "Please determine whether this data is a fracture waveform or a healthy waveform."

[0065] Returning to FIG. 7 , the CPU 11 sends the prompt 134, which includes a system prompt 1341 and a user prompt 1342, to the LLM 70 and calls the API (Application Programming Interface) of the LLM 70 to generate a response sentence (step S104). Here, the CPU 11 links the system prompt 1341 and the user prompt 1342 using a predetermined library (e.g., language chain) to generate the prompt 134. The CPU 11 converts the prompt 134 into a format that can call the API of the LLM 70, in this case, the JSON format, and makes the API call. The system prompt 1341 included in the prompt 134 functions as input data for the Few-Shot-Learning. The above-described steps S102 to S104 correspond to the request processing executed by the CPU 11 as a requesting unit.

[0066] The LLM 70 generates a response sentence 135 in response to the input prompt 134 using the method described above, and transmits it to the server 10 (step S105). FIG. 11 is a diagram showing an example of the response sentence 135. In the example of FIG. 11, the response sentence 135 includes the following wording: "After checking the data, all sensors show characteristic peaks and drops at certain positions, which are likely to correspond to the break point. Specifically, all sensors show peak values ​​followed by a rapid drop. This may indicate that a part of the bridge broke, causing a large change in the magnetic field at that position. Furthermore, the fact that the waveform after the break is flat suggests that the structure after the break is magnetized normally without affecting the magnetic field. Therefore, it can be determined that this data is a break waveform."

[0067] Returning to FIG. 7 , the CPU 11 of the server 10 uses the library to convert the response statement 135 into result information 136 according to a predetermined template, and stores the result information in the analysis DB 50 (step S106). The template is defined according to the manner of interpretation of the physical phenomenon. In this embodiment, the template includes a Boolean value representing the presence or absence of a break, and text data of the reason for interpreting the presence or absence of a break. The CPU 11 identifies the information on the presence or absence of a break contained in the response statement 135 and acquires it as a Boolean value. The CPU 11 acquires the information on the reason for interpreting the presence or absence of a break contained in the response statement 135 as text data. The processing of step S106 is executed by the CPU 11 as output control means.

[0068] FIG. 12 is a diagram showing an example of result information 136. The description "Measurement result: no fracture" in the result information 136 in FIG. 12 is generated corresponding to a Boolean value indicating the presence or absence of a fracture. Instead of the description "Measurement result: no fracture", the Boolean value itself (for example, "0" or "1") may be used. The description of "Notes" in the result information 136 in FIG. 12 is text data regarding the reason for interpreting the presence or absence of a fracture, obtained by creating a template from the response sentence 135.

[0069] When step S106 in Fig. 7 is completed, each device ends the data analysis process. When performing fracture analysis on steel materials at multiple locations on a bridge, the data analysis process in Fig. 7 can be repeatedly executed. In this case, multiple pieces of measurement data 132 and corresponding pieces of result information 136 are stored in the analysis DB 50.

[0070] When the data analysis process is completed, a result output process is initiated to output the interpretation result of whether or not there is a fracture to the user terminal 30. FIG. 13 is a flowchart showing the control procedure of the result output process. Here, the result output process will be described focusing on the processing performed by the CPU 11 of the server 10. The result output process is executed by the CPU 11 as an output control means. When the result output process is initiated, the CPU 11 of the server 10 repeatedly determines whether or not a request for output of the interpretation result has been made from the user terminal 30 (step S201). If it determines that a request for output of the interpretation result has been made ("YES" in step S201), the CPU 11 determines whether or not multiple pieces of result information 136 are stored in the analysis DB 50 (step S202).

[0071] If it is determined that multiple pieces of result information 136 are stored in the analysis DB 50 ("YES" in step S202), the CPU 11 performs statistical analysis of the multiple pieces of result information 136 (step S203). Here, the CPU 11 outputs the multiple pieces of result information 136 to a table-format file and performs statistical analysis on this file. For example, the CPU 11 uses the Python libraries NumPy and / or pandas to perform average, variance, correlation, and regression analysis of the data. The CPU 11 saves the results of the statistical analysis in the memory unit 13 and converts them into a predetermined data format and stores them in the analysis DB 50. The CPU 11 transmits the results of the statistical analysis to the user terminal 30 and displays them on the display unit 34 as fracture interpretation results (step S204). In step S204, the display unit 34 may also display the multiple pieces of result information 136 used in the statistical analysis.

[0072] If it is determined that there is only one result information 136 stored in the analysis DB 50 ("NO" in step S202), the CPU 11 sends the result information 136 to the user terminal 30 and displays it on the display unit 34 as the interpretation result of the fracture (step S205).

[0073] The CPU 11 determines whether feedback has been provided from the user terminal 30 regarding the actual condition of the steel, i.e., whether the steel is actually fractured (step S206). Here, the user of the user terminal 30 checks whether the steel is fractured using a method other than the magnetic flux leakage method, such as breaking up the concrete of a bridge and visually inspecting the condition of the steel. The user then inputs the check results into the user terminal 30, thereby transmitting feedback regarding the condition of the steel to the server 10. The user terminal 30 may also acquire the results of labeling by the user any of the multiple pieces of result information 136 that differ from the actual condition of the steel, and transmit the labeling results to the server 10 as feedback. If it is determined that feedback has been provided from the user terminal 30 ("YES" in step S206), the CPU 11 executes step S207. In step S207, the CPU 11 extracts the result information 136 for which the interpretation results by the LLM 70 differ from the actual condition of the steel. The CPU 11 stores the measurement data 132 corresponding to the extracted result information 136 as example data D1 in the storage unit 13 or stores it in the analysis DB 50 (step S207). For example, the CPU 11 reflects the stored example data D1 and its interpretation information D2 in the premise information 133 and uses it in subsequent data analysis processes. In other words, the CPU 11 adjusts the premise information 133 to be used in the next requested process based on the comparison result between the acquired result information 136 and the actual state of the bridge (actual physical phenomenon). This improves the accuracy of interpretation by the LLM 70.

[0074] When step S207 is completed, or when it is determined in step S206 that there is no feedback from the user terminal 30 (step S206), the CPU 11 ends the result output process.

[0075] The following statistical analysis process may be performed based on the feedback received in step S206. For example, the accuracy, recall, precision, and confusion matrix of the multiple pieces of result information 136 may be derived. The derived results may be displayed on the user terminal 30. The confusion matrix is ​​used to evaluate the performance of binary or multi-valued classification problems in machine learning. In particular, it can be used to understand and evaluate prediction results in classification problems such as the presence or absence of fractures. The confusion matrix is ​​a two-row, two-column matrix whose elements are "True Positive (TP, true positive)," "False Positive (FP, false positive)," "True Negative (TN, true negative)," and "False Negative (FN, false negative)." When interpreting fractures in steel, for example, the absence of fractures corresponds to "Positive," and the presence of fractures corresponds to "Negative." "TP" indicates that the result was actually positive and was correctly interpreted (predicted) as positive. "FP" indicates that the result was actually negative but was incorrectly interpreted as positive. "TN" indicates that the result was actually negative and was correctly predicted as negative. "FN" indicates that the result was actually positive but was incorrectly predicted as negative. Accuracy, recall, specificity, and precision can be calculated using the elements of the confusion matrix. Accuracy represents the proportion of correct interpretations out of the total interpretations. Accuracy can be calculated as (TP + TN) / (TP + TN + FP + FN). Recall represents the proportion of cases that are correctly interpreted as positive out of those that are actually positive. Recall can be calculated as TP / (TP+FN). Recall is also called the true positive rate (TPR) or sensitivity. Specificity represents the proportion of cases that are correctly interpreted as negative out of those that are actually negative. Specificity can be calculated as TN / (TN+FP). Specificity is also called the true negative rate (TNR).The precision rate represents the percentage of results that were actually positive among those interpreted as positive by LLM70. The precision rate can be calculated as TP / (TP+FP). The precision rate is also called the positive predictive value.

[0076] Using evaluation indices such as confusion matrices, accuracy, precision, and recall, it is possible to evaluate the performance, weaknesses, misinterpretation patterns, and thresholds for interpretation judgment of the LLM 70. By feeding back the evaluation results to the system prompt 1341, it is possible to improve the behavior of the LLM 70. In other words, even if the LLM 70 is not actually being developed, by changing the prompt, it is possible to treat it as if a new LLM 70 has been developed. Therefore, since an appropriate LLM 70 can be easily prepared, the effort and cost required for analyzing physical quantity data can be reduced.

[0077] (2) Chain-of-Though Chain-of-Though is a method for realizing complex reasoning through multiple intermediate reasoning steps by describing the thought process for a given question in the prompt 134.

[0078] 14 is a flowchart showing the control procedure for data analysis processing when Chain-of-Though is used. When Chain-of-Though is used, a plurality of mutually different sample data 1331 is stored in advance in the storage unit 13 of the server 10.

[0079] Steps S301 to S305 of the data analysis process in Fig. 14 are the same as steps S101 to S105 in Fig. 7, and therefore will not be described here. Steps S302 to S304 correspond to the first request process.

[0080] When the CPU 11 of the server 10 receives the response sentence 135 output from the LLM 70 in step S305, it executes step S306. In step S306, the CPU 11 links the next sample data 1331 with the previous chat content using a predetermined library (e.g., language chains). The CPU 11 then sends a prompt 134 including the sample data 1331 to the LLM 70 and calls the API (step S306). Step S306 corresponds to the second request process.

[0081] FIG. 15 is a diagram showing an example of the prompt 134 generated in step S306. This prompt 134 includes, of the multiple sample data 1331 previously stored in the storage unit 13, sample data 1331 that is different from the sample data 1331 included in the prompt 134 input to the LLM 70 in step S304. In the example shown in FIG. 15, this sample data 1331 includes example data D1 and two interpretation information D2. The first interpretation information D2 states, "The following is also a typical break waveform." The second interpretation information D2 states, "The break point of this break waveform is at 30 cm on sensor 3. The reason for the break is that, compared to the other sensors, sensor 3 has what is called an inverted S-shaped waveform; when viewed in the horizontal direction, it bulges before the break point and drops after the break point, and therefore it has been determined to be a break."

[0082] The LLM 70 generates a response sentence 135 in response to the input prompt 134 and transmits it to the server 10 (step S307).

[0083] The CPU 11 of the server 10 determines whether the next sample data 1331 is stored in the storage unit 13 (step S308). Here, the next sample data 1331 is the sample data 1331 that has not yet been input to the LLM 70 after the initial request processing in steps S302 to S304, among the multiple sample data 1331 previously stored in the storage unit 13. If it is determined that the next sample data 1331 is available ("YES" in step S308), the CPU 11 returns the process to step S306. Step S306 here is identical to the previous step S306, except for the sample data 1331 to be used. The CPU 11 repeatedly executes step S306 for the number of sample data 1331. Step S306 in this loop processing corresponds to the third or subsequent request processing. This loop processing also corresponds to the presentation of an intermediate thought process in Chain-of-Thought. In this way, in Chain-of-Though, the CPU 11 repeatedly performs request processing on the LLM 70 for certain measurement data 132. In the second and subsequent request processing, the CPU 11 inputs to the LLM 70 a prompt 134 including sample data 1331 that has not been input to the LLM 70 since the initial request processing, among a plurality of different sample data 1331.

[0084] If it is determined in step S308 that there is no next sample data 1331 ("NO" in step S308), the CPU 11 converts the last received response sentence 135 into result information 136 according to a predetermined template and stores it in the analysis DB 50 (step S309). Thereafter, the CPU 11 ends the data analysis process. Thereafter, the result output process of FIG. 13 is executed in the same manner as in the case of Few-shot-learning.

[0085] Fig. 16 is a flowchart showing another control procedure for data analysis processing when Chain-of-Though is used. The flowchart in Fig. 16 differs from Fig. 15 in that sample data 1331 used in repeated request processing is input from the user terminal 30 to the server 10. The differences from Fig. 15 will be explained below.

[0086] Steps S401 to S405 in FIG. 16 are the same as steps S301 to S305 in FIG. 15. When the CPU 11 of the server 10 receives the response sentence 135 output from the LLM 70 in step S405, it transmits the response sentence 135 to the user terminal 30 (step S406). The CPU 31 of the user terminal 30 displays the received response sentence 135 on the display unit 34. The CPU 31 accepts input of sample data 1331 from the user, for example, on a browser, and transmits the input sample data 1331 to the server 10 (step S407). The CPU 11 of the server 10 links the received sample data 1331 with past chat content. The CPU 11 then transmits a prompt 134 including the sample data 1331 to the LLM 70 and invokes the API (step S408). The subsequent steps S409 to S411 are the same as steps S307 to S309 in Fig. 15. In this way, the Chain-of-Though may be performed by inputting the sample data 1331 from the user terminal 30.

[0087] (3) Fine-tuning Fine-tuning is a technique in which the LLM 70 is retrained using training data to adjust the model parameters for a specific task. In fine-tuning retraining, the error in the output of the LLM 70 according to the training data is fed back in the neural network in the reverse direction. This adjusts the model parameters. Then, the retrained LLM 70 is made to interpret fractures in the same way as in few-shot learning.

[0088] 17 is a flowchart showing the control procedure for data analysis processing when fine-tuning is used. When fine-tuning is used, learning data 432 is stored in advance in the storage unit 43 of the administrator terminal 40. The learning data 432 includes example data D1 and interpretation information D2 based on the example data D1, similar to the sample data 1331. For relearning, it is preferable to use learning data 432 that includes, for example, tens to hundreds of different sets of example data D1 and the corresponding interpretation information D2.

[0089] When the data analysis process starts, the CPU 41 of the administrator terminal 40, which functions as a relearning means, makes a fine-tuning API call to the LLM 70 using the training data 432 (step S501). For example, the CPU 41 compiles the training data 432 into a CSV file, converts the CSV file into JSON format, and transmits it to the LLM 70 while making an API call. In response, the LLM 70 performs relearning using the input training data 432 and adjusts the model parameters (step S502). Thus, in steps S501 and S502, the CPU 41, which functions as a relearning means, causes the LLM 70 to re-train using the training data 432.

[0090] The subsequent steps S503 to S508 are the same as steps S101 to S106 in Figure 7, and therefore their explanation will be omitted. Note that by adjusting the model parameters through fine-tuning, it may be possible to generate an appropriate response sentence 135 even if the system prompt 1341 is omitted. In this case, the system prompt 1341 may be omitted from the prompt 134. In other words, step S504 may be omitted.

[0091] When the data analysis process is completed, the result output process of FIG. 13 is executed, as in the case of fine-tuning. When fine-tuning is used, in step S207 of FIG. 13, the CPU 11 transmits the measurement data 132 corresponding to the result information 136 extracted based on the user's feedback to the administrator terminal 40 as learning data 432. In other words, the measurement data 132 corresponding to the extracted result information 136 is used as the next learning data 432. This can improve the effectiveness of re-learning and the accuracy of interpretation by the LLM 70.

[0092] In the above example, the CPU 41 of the administrator terminal 40 functions as a relearning unit that causes the LLM 70 to relearn, but this is not limiting. For example, the CPU 11 of the server 10 or the CPU 31 of the user terminal 30 may function as the relearning unit. In this case, the learning data 432 is transmitted from the server 10 or the user terminal 30 to the LLM 70.

[0093] (4) Semantic Search Semantic search is a technology in which a search engine understands the user's search intent and purpose from input data and provides search results that match the user's intent. In the information processing system 1, semantic search can be applied to interpret the presence or absence of a fracture using the following method. First, multiple different sample data 1331 are prepared. When the inspection device 20 generates measurement data 132, one example data D1 that satisfies a predetermined similarity condition with the measurement data 132 is searched from multiple example data D1 included in the multiple sample data 1331. The similarity condition can be, for example, "being closest to the measurement data 132 in vector space." Then, the sample data 1331 containing the obtained one example data D1 is included in the system prompt 1341, and the LLM 70 interprets the presence or absence of a fracture using the few-shot learning method. The process of searching for one example data D1 that matches the measurement data 132 from among the plurality of example data D1 is performed by the vector DB 60.

[0094] 18 is a flowchart showing the control procedure for data analysis processing when using semantic search. When using semantic search, a plurality of sample data 1331 including a plurality of different example data D1 is stored in advance in the storage unit 43 of the administrator terminal 40.

[0095] When the data analysis process is started, the CPU 41 of the administrator terminal 40 acquires a plurality of vectorized example data obtained by vectorizing a plurality of example data D1 and transmits the acquired vectorized example data to the vector DB 60 (step S601). Here, the CPU 41 acquires the vectorized example data by calling an API of an embedding model for converting the example data D1 into a vector representation. Note that the process of step S601 may be executed by the server 10 instead of the administrator terminal 40. The vector DB 60 stores the plurality of vectorized example data received from the administrator terminal 40 (step S602).

[0096] When the inspection device 20 transmits the measurement data 132 to the server 10 (step S603), the CPU 11 of the server 10 requests the vector DB 60 to execute a similarity search with the measurement data 132 (step S604). Here, the CPU 11 first calls the API of the embedding model to obtain vectorized measurement data (vectorized physical quantity data) obtained by vectorizing the measurement data 132. Then, the CPU 11 transmits a similarity search request to the vector DB 60 together with the vectorized measurement data.

[0097] The vector DB 60 searches the plurality of vectorized example data for vector data that has the highest similarity to the received vectorized measurement data. That is, the vector DB 60 identifies the plurality of vectorized example data for data that has the shortest distance to the vectorized measurement data in vector space. The vector DB 60 transmits the search results to the server 10 (step S605).

[0098] The CPU 11 of the server 10 identifies the example data D1 corresponding to the vectorized example data having the highest similarity indicated by the received search results as one example data D1 (corresponding example data) corresponding to the measurement data 132. The CPU 11 generates a system prompt 1341 including the identified example data D1 (step S606). Specifically, the CPU 11 includes in the system prompt 1341 the sample data 1331 including the identified example data D1 and interpretation information D2 corresponding to the example data D1.

[0099] The subsequent steps S607 to S610 are the same as steps S103 to S106 in FIG. 7, and therefore a description thereof will be omitted.

[0100] (Effects) As described above, the information processing system 1 according to this embodiment includes a CPU 11. The CPU 11, as a requesting unit, performs a request process to request the LLM 70 to generate a response statement 135 (interpretation information) that interprets the presence or absence of fractures in the steel material (physical phenomenon) represented by the magnetic flux density based on measurement data 132 related to the detection result of the magnetic flux density (physical quantity). The CPU 11, as an output control unit, acquires the response statement 135 generated by the LLM 70 in response to the request process and displays the interpretation result of the presence or absence of fractures in the steel material based on the response statement 135 on the display unit 34. This allows the LLM 70 to interpret physical phenomena using its function of executing tasks through natural language processing. This enables more efficient and precise analysis of physical quantity data without requiring specialized skills or knowledge. In other words, it becomes possible to efficiently analyze complex patterns and correlations in physical quantity data related to physical phenomena using natural language processing techniques. This also increases the possibility of discovering unknown patterns and correlations in physical quantities related to physical phenomena. Furthermore, since the existing LLM 70 is used, there is no need to design a new learning model or perform machine learning for the analysis, which reduces the effort and cost required to analyze physical quantity data.

[0101] In the request process, the CPU 11 as a requesting means generates a prompt 134 in a predetermined format, which is an input statement for instructing the LLM 70 to perform an operation and includes at least the measurement data 132. The CPU 11 as a requesting means inputs the prompt 134 to the LLM 70, thereby requesting the LLM 70 to generate a response statement 135. This allows the physical quantity data to be analyzed by the simple process of inputting the prompt 134 to the LLM 70.

[0102] The CPU 11, which serves as output control means, converts the response sentence 135, which serves as interpretation information, into result information 136 according to a predetermined template relating to the interpretation of whether or not there is a fracture in the steel material. The CPU 11, which serves as output control means, displays the result information 136 as the interpretation result on the display unit 34. The response sentence 135 generated by the LLM 70 is prone to fluctuations in expression, but by using the above method, it is possible to display a clear interpretation result with little fluctuation in expression.

[0103] When using Few-shot-learning, the CPU 11 as requesting means generates a prompt 134 that further includes premise information 133 for interpreting the presence or absence of fracture in the steel material based on the measurement data 132, and inputs the prompt 134 to the LLM 70. This allows the LLM 70 to interpret the physical phenomenon without updating the model parameters of the LLM 70, i.e., without re-learning.

[0104] The premise information 133 includes precondition information 1332 related to the detection of magnetic flux density. This allows the LLM 70 to perform a proper interpretation based on the precondition information 1332 related to the detection.

[0105] The premise information 133 includes sample data 1331 made up of example data D1, which is an example of the measurement data 132, and interpretation information D2, which is the result of interpreting whether or not the steel material is fractured based on the example data D1. This allows the LLM 70 to perform a valid interpretation based on the sample data 1331, which is the result of a correct interpretation.

[0106] When Chain-of-Though is used, the CPU 11, as a requesting means, repeatedly performs a request process on the LLM 70 for certain measurement data 132. In the second and subsequent request processes, the CPU 11 inputs to the LLM 70 a prompt 134 including sample data 1331 that has not been input to the LLM 70 since the initial request process, among a plurality of different sample data 1331. This allows the LLM 70 to perform a valid interpretation of a physical phenomenon through intermediate inference steps.

[0107] When fine-tuning is performed, the CPU 41 of the administrator terminal 40, which functions as a relearning means, performs the following operation before the CPU 11, which functions as a requesting means, performs the request processing. That is, the CPU 41 re-learns the LLM 70 using the learning data 432, which includes example data D1, which is an example of the measurement data 132, and interpretation information D2, which is the result of interpreting whether or not the steel material is fractured based on the example data D1. This makes it possible to adjust the model parameters of the LLM 70 in accordance with the interpretation of the physical phenomenon. This allows the LLM 70 to perform a valid interpretation of the physical phenomenon. Furthermore, it is possible to omit input of the premise information 133 using the system prompt 1341.

[0108] The CPU 11, as a requesting means, performs multiple request processes based on multiple different measurement data 132. The CPU 11, as an output control means, performs statistical analysis using a predetermined statistical method on multiple response statements 135 obtained from the LLM 70 in response to multiple processing requests, and displays the results of the statistical analysis as interpretation results on the display unit 34. In this way, using the statistical method, it is possible to obtain appropriate interpretation results that suppress the influence of fluctuations in interpretation by the LLM 70.

[0109] When using semantic search, the CPU 11, acting as a requesting unit, identifies corresponding example data corresponding to the measurement data 132 from among a plurality of mutually different example data D1. Furthermore, in the request processing, the CPU 11 generates a prompt 134 including the identified corresponding example data and inputs it to the LLM 70. The corresponding example data is example data D1 corresponding to vectorized example data obtained by vectorizing the plurality of example data D1 and that satisfies a predetermined similarity condition with the vectorized measurement data obtained by vectorizing the measurement data 132. This allows interpretation of a physical phenomenon through few-shot learning using appropriate example data D1 that matches the measurement data 132. Therefore, more appropriate interpretation results can be obtained.

[0110] The CPU 11 as a requesting means adjusts the premise information 133 to be used in the next request processing based on the result of comparing the interpretation result with the presence or absence of fracture in the actual steel material detected in the measurement data 132. This makes it possible to improve the accuracy of interpretation by the LLM 70.

[0111] The CPU 41 of the administrator terminal, which serves as a relearning means, causes the LLM 70 to relearn using the learning data 432 that reflects the comparison result between the interpretation result and the presence or absence of fracture in the actual steel material for which the measurement data 132 was detected. This improves the effect of the relearning and the accuracy of interpretation by the LLM 70.

[0112] The server 10 according to this embodiment includes a CPU 11. The CPU 11, as a requesting means, performs a request process to request the LLM 70 to generate a response statement 135 interpreting whether or not the steel material is fractured, as indicated by the magnetic flux density, based on measurement data 132 related to the detection result of the magnetic flux density. The CPU 11, as an output control means, acquires the response statement 135 generated by the LLM 70 in response to the request process, and causes the display unit 34 to display the interpretation result of whether or not the steel material is fractured based on the response statement 135. This enables more efficient and precise analysis of physical quantity data without requiring specialized skills or knowledge. Furthermore, the effort and cost required for analyzing physical quantity data can be reduced.

[0113] The information processing method executed by the computer according to this embodiment includes a request step and an output control step. In the request step, a request process is performed to request the LLM 70 to generate a response statement 135 that interprets whether or not the steel material is fractured, as indicated by the magnetic flux density, based on measurement data 132 related to the detection result of the magnetic flux density. In the output control step, the LLM 70 acquires the response statement 135 generated in response to the request process, and displays the interpretation result of whether or not the steel material is fractured based on the response statement 135 on the display unit 34. This enables more efficient and precise analysis of physical quantity data without requiring specialized skills or knowledge. Furthermore, the effort and cost required for analyzing physical quantity data can be reduced.

[0114] The program 131 according to this embodiment causes the CPU 11 to function as a requesting means and an output controlling means. The requesting means performs a request process that requests the LLM 70 to generate a response statement 135 that interprets whether or not the steel material is fractured, as indicated by the magnetic flux density, based on measurement data 132 related to the detection result of the magnetic flux density. The output controlling means acquires the response statement 135 generated by the LLM 70 in response to the request process, and causes the display unit 34 to display the interpretation result of whether or not the steel material is fractured based on the response statement 135. This enables more efficient and precise analysis of physical quantity data without requiring specialized skills or knowledge. Furthermore, the effort and cost required for analyzing physical quantity data can be reduced.

[0115] (Other) The present invention is not limited to the above-described embodiment and various modifications are possible. For example, in the above-described embodiment, the information processing system 1 is used for non-destructive inspection of bridges, but the present invention is not limited to this. The information processing system 1 may be used for climate change data analysis. That is, the information processing system 1 may be used for the purpose of analyzing and predicting data related to climate change, such as temperature, precipitation, and sea surface temperature. The information processing system 1 may be used for earthquake data analysis. That is, the information processing system 1 may be used for the purpose of analyzing seismic activity and fault movement using seismic waveform data and crustal movement data. The information processing system 1 may be used for the prediction of chemical reactions. That is, the information processing system 1 may be used for the purpose of predicting the progress and products of chemical reactions using data on the molecular structure and state of chemical substances. The information processing system 1 may be used for the analysis of material properties. That is, the information processing system 1 may be used for the purpose of designing new materials using physicochemical property data of materials (thermal conductivity, electrical conductivity, etc.). The information processing system 1 may be used for analyzing biological data. That is, the information processing system 1 may be used for the purpose of analyzing the cause and treatment of disease from data on gene sequences and protein structures. The information processing system 1 may be used for space data analysis. That is, the information processing system 1 may be used for the purpose of analyzing the structure and evolution of the universe from observed space data (such as stellar spectra and images of black holes). The information processing system 1 may be used for analyzing environmental pollution. That is, the information processing system 1 may be used for the purpose of analyzing air and water quality data to evaluate the source and impact of pollution. The information processing system 1 may be used for ecosystem analysis. That is, the information processing system 1 may be used for the purpose of evaluating the health and resistance of ecosystems using biodiversity data and ecosystem interaction data. The information processing system 1 may be used for epidemiological data analysis. That is, the information processing system 1 may be used for the purpose of evaluating transmission patterns and infection risks using infectious disease data. The information processing system 1 may be used for energy consumption analysis.That is, the information processing system 1 may be used for the purpose of analyzing energy consumption data and proposing energy saving and sustainable development methods.

[0116] In the above embodiment, the interpretation of a physical phenomenon is exemplified by an example in which the interpretation result can be expressed as a Boolean value, such as the presence or absence of a fracture in a steel material, but the interpretation is not limited to this. For example, the interpretation of a physical phenomenon may be "clustering" in which physical phenomena are classified into several clusters based on predetermined rules.

[0117] In the above embodiment, the measurement data 132 obtained by detecting magnetic flux density as a physical quantity by the detection unit 24 of the inspection device 20 is exemplified as the physical quantity data related to the detection result of the physical quantity, but this is not limiting. The physical quantity data related to the detection result of the physical quantity may be observation data obtained by observing a physical phenomenon or experimental data obtained by an experiment. Furthermore, instead of the physical quantity data related to the detection result of the physical quantity, physical quantity data related to the calculation result of the physical quantity, for example, simulation data obtained by simulating a physical phenomenon, may be used.

[0118] The configuration of the information processing system 1 is not limited to that shown in FIG. 1 . For example, two or more of the inspection device 20, the user terminal 30, and the administrator terminal 40 may be integrated. Furthermore, the inspection device 20, the user terminal 30, or the administrator terminal 40 may be equipped with the functions executed by the server 10, and the server 10 may be omitted. Therefore, the CPU 21 of the inspection device 20, the CPU 31 of the user terminal 30, or the CPU 41 of the administrator terminal 40 may function as a request means and / or an output control means. Furthermore, the server 10 may be integrated with at least one function of the analysis DB 50, the vector DB 60, and the LLM 70.

[0119] Although the output of the interpretation result by the output means is exemplified as display on the display unit 34 of the user terminal 30, this is not limiting. The output of the interpretation result may be transmission of data related to the interpretation result by a communication unit serving as output means, or printing of the interpretation result by a printing device serving as output means, etc.

[0120] Although several embodiments of the present invention have been described, the scope of the present invention is not limited to the above-described embodiments, but includes the scope of the invention described in the claims and its equivalents.

[0121] The present invention can be used in an information processing system, an information processing device, an information processing method, and a program.

[0122] REFERENCE SIGNS LIST 1 Information processing system 10 Server (information processing device) 11 CPU (request means, output control means) 132 Measurement data (physical quantity data) 133 Prerequisite information 1331 Sample data 1332 Prerequisite information 134 Prompt 1341 System prompt 1342 User prompt 135 Response sentence (interpretation information) 136 Result information 20 Inspection device 30 User terminal 34 Display unit (output means) 40 Administrator terminal 41 CPU (relearning means) 432 Learning data 50 Analysis DB 60 Vector DB 70 LLM (large scale language model) D1 Example data D2 Interpretation information

Claims

1. An information processing system comprising: a request means for performing request processing to request a large-scale language model to generate interpretation information that interprets a physical phenomenon represented by a physical quantity based on physical quantity data related to the detection or calculation results of the physical quantity; and an output control means for acquiring the interpretation information generated by the large-scale language model in response to the request processing, and outputting an interpretation result of the physical phenomenon based on the interpretation information to a specified output means.

2. The information processing system of claim 1, wherein the request means generates a prompt in a predetermined format including at least the physical quantity data, which is an input sentence for instructing the large-scale language model to perform an action in the request processing, and inputs the prompt into the large-scale language model to request the large-scale language model to generate the interpretation information.

3. The information processing system according to claim 1, wherein the output control means acquires a response sentence from the large-scale language model as the interpretation information, converts the response sentence into result information according to a predetermined template relating to the interpretation of the physical phenomenon, and causes the output means to output the result information as the interpretation result.

4. The information processing system according to claim 2, wherein said request means, in said request processing, generates said prompt further including premise information for interpreting said physical phenomenon based on said physical quantity data and inputs the generated prompt to said large-scale language model.

5. The information processing system according to claim 4, wherein the premise information includes information on prerequisites related to detection or calculation of the physical quantity.

6. The information processing system according to claim 4, wherein the premise information includes sample data consisting of example data that is an example of the physical quantity data, and information resulting from an interpretation of the physical phenomenon based on the example data.

7. The information processing system according to claim 6, wherein the request means repeatedly performs the request process on the large-scale language model for certain physical quantity data, and in the second or subsequent request processes, inputs to the large-scale language model the prompt including sample data among a plurality of mutually different sample data that has not been input to the large-scale language model since the initial request process.

8. The information processing system according to claim 1, further comprising a re-learning means for re-learning the large-scale language model using learning data including example data which is an example of the physical quantity data and information resulting from an interpretation of the physical phenomenon based on the example data before the requesting means performs the request processing.

9. The information processing system according to claim 1, wherein the request means performs the requested processing multiple times based on multiple pieces of physical quantity data that are different from each other, and the output control means performs a statistical analysis on the multiple pieces of interpretation information obtained from the large-scale language model in response to the multiple processing requests using a predetermined statistical method, and causes the output means to output a result of the statistical analysis as the interpretation result.

10. The information processing system of claim 6, wherein the request means identifies corresponding example data corresponding to the physical quantity data from among a plurality of mutually different example data, and in the request processing, generates the prompt including the identified corresponding example data and inputs it to the large-scale language model, and the corresponding example data is example data corresponding to vectorized example data among a plurality of vectorized example data obtained by vectorizing the plurality of example data, which satisfies a predetermined similarity condition between the vectorized physical quantity data obtained by vectorizing the physical quantity data.

11. An information processing system as described in claim 4, wherein the physical quantity data is data relating to a detection result of the physical quantity, and the requesting means adjusts the premise information to be used for the next requested processing based on a comparison result between the interpretation result and the actual physical phenomenon in which the physical quantity data was detected.

12. The information processing system of claim 8, wherein the physical quantity data is data relating to a detection result of the physical quantity, and the re-learning means re-learns the large-scale language model using the learning data reflecting a comparison result between the interpretation result and the actual physical phenomenon in which the physical quantity data was detected.

13. An information processing device comprising: a request means for performing request processing to request a large-scale language model to generate interpretation information that interprets a physical phenomenon represented by a physical quantity based on physical quantity data related to a detection result or calculation result of the physical quantity; and an output control means for acquiring the interpretation information generated by the large-scale language model in response to the request processing, and outputting an interpretation result of the physical phenomenon based on the interpretation information to a predetermined output means.

14. An information processing method executed by a computer, comprising: a request step of performing request processing to request a large-scale language model to generate interpretation information that interprets a physical phenomenon represented by a physical quantity based on physical quantity data related to a detection result or calculation result of the physical quantity; and an output control step of acquiring the interpretation information generated by the large-scale language model in response to the request processing, and outputting an interpretation result of the physical phenomenon based on the interpretation information to a predetermined output means.

15. A program that causes a computer to function as: a request means that performs request processing to request a large-scale language model to generate interpretation information that interprets a physical phenomenon represented by a physical quantity based on physical quantity data related to the detection results or calculation results of the physical quantity; and an output control means that acquires the interpretation information generated by the large-scale language model in response to the request processing, and outputs the interpretation result of the physical phenomenon based on the interpretation information to a specified output means.

Images