Learning device, route estimation system and learning method

Abstract

To estimate processes that a worker has performed, a sequence in which the worker has performed the processes, and a route in which the worker has moved to perform the processes.SOLUTION: A learning device comprises: a gathering part which gathers process operation information representing operation states of respective processes and reception level information including reception levels of signals that a first radio station receives from second radio stations installed for the respective processes; and a learning part which uses teacher data having the process operation information and the reception level information linked, process by process, so as to allow a generation model for estimating a state of movement of a third radio station to be learned.SELECTED DRAWING: Figure 1

Description

[Technical field]

[0001] The present disclosure relates to a learning device, a route estimation system, and a learning method. [Background technology]

[0002] One of the production methods for industrial products is the cell production method, which is highly adaptable to high-mix low-volume production and fluctuations in production volume. The cell production method is a method in which one or a small number of workers complete the product assembly process, with one worker taking charge of multiple processes. Understanding the processes performed by a worker, the order in which they performed the processes, and the route taken between processes to perform the processes provides information for productivity management, so there is a need to detect the relationship between the (work) position of a worker and the operating status of a process (process operation information).

[0003] There are various indoor position estimation technologies that are considered to be applicable to such a cell production system, but the demand for position estimation that applies wireless communication technology is increasing. For example, Patent Document 1 discloses a technology that collects in advance a list of radio wave strengths from many access points (called a wireless map or fingerprint), and estimates the position of a terminal based on the list of radio wave strengths. [Prior art documents] [Patent documents]

[0004] [Patent Document 1] Patent Publication No. 2021-56222 Summary of the Invention [Problem to be solved by the invention]

[0005] In the cell production system, a plurality of workers are present within a relatively small area, and the process operation information for each process often includes a mixture of process operation information from a plurality of workers.

[0006] High accuracy in location estimation is required to separate process operation information by multiple workers for each worker using location estimation technology. However, if the accuracy of location estimation using the fingerprint method is low, it is difficult to link the process operation information with the process layout of the cell, i.e., the (work) position of the worker, even if the technology disclosed in Patent Document 1 is applied to a cell production system. Therefore, it becomes difficult to grasp the processes performed by the worker, the order in which the processes were performed, and the route taken between processes to perform the processes.

[0007] Non-limiting examples of the present disclosure contribute to providing a learning device, a path estimation system, and a learning method that can estimate the steps performed by a worker, the order in which the steps were performed, and the path taken between steps to perform the steps, even using a fingerprint method. [Means for solving the problem]

[0008] A learning device according to one embodiment of the present disclosure includes a collection unit that collects process operation information indicating the operation status of each process and reception level information including the reception level of a signal received by a first radio station from a second radio station installed in each process, and a learning unit that learns a generative model for estimating a movement status of a third radio station using teacher data that links the process operation information and the reception level information for each process.

[0009] A path estimation system according to an embodiment of the present disclosure includes the learning device, an estimation unit configured to estimate a state of the movement based on the learning device, a process operation information sequence, a reception level information sequence including reception levels of signals received by the third radio station from a plurality of the second radio stations, and the generation model.

[0010] A learning method according to one embodiment of the present disclosure collects process operation information indicating the operation status of each process and reception level information including the reception level of a signal received by a first radio station from a second radio station installed in each process, and uses training data linking the process operation information and the reception level information for each process to learn a generative model for estimating the movement status of a third radio station.

[0011] These comprehensive or specific aspects may be realized by a system, an apparatus, a method, an integrated circuit, a computer program, or a recording medium, or may be realized by any combination of a system, an apparatus, a method, an integrated circuit, a computer program, and a recording medium. Effect of the Invention

[0012] According to the present disclosure, even when using the fingerprint method, it is possible to estimate the steps performed by a worker, the order in which the steps were performed, and the path taken between steps to perform the steps.

[0013] Further advantages and benefits of an embodiment of the present disclosure will become apparent from the specification and drawings. Such advantages and / or benefits may be provided by some of the embodiments and features described in the specification and drawings, respectively, but not necessarily all of them may be provided to obtain one or more identical features. [Brief description of the drawings]

[0014] [Figure 1] FIG. 1 is a configuration diagram illustrating an example of a route estimation system according to an embodiment of the present disclosure. [Diagram 2] FIG. 1 is a diagram showing an example of a process label sequence according to an embodiment of the present disclosure. [Diagram 3] FIG. 1 is a diagram showing examples of process labels and fingerprints according to an embodiment of the present disclosure. [Figure 4] FIG. 1 is a configuration diagram illustrating an example of a learning unit according to an embodiment of the present disclosure. [Diagram 5] 1 is a flowchart showing an example of a learning method according to an embodiment of the present disclosure. [Figure 6] FIG. 1 is a block diagram showing an example of a route estimation unit according to an embodiment of the present disclosure. [Figure 7] 1 is a flowchart showing an example of a route estimation method according to an embodiment of the present disclosure. [Figure 8] FIG. 1 shows examples of candidate sequences according to an embodiment of the present disclosure. [Figure 9]FIG. 13 is a diagram showing an example of an inter-process transfer probability table according to an embodiment of the present disclosure. [Figure 10] FIG. 1 is a diagram showing an example of an observed fingerprint sequence and a reconstructed fingerprint sequence according to an embodiment of the present disclosure. [Figure 11] FIG. 1 is a diagram showing an example of a display by a display unit according to an embodiment of the present disclosure. [Figure 12] FIG. 13 is a diagram showing another example of a display by the display unit according to the embodiment of the present disclosure. DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS

[0015] Hereinafter, the embodiments of the present disclosure will be described in detail with reference to the drawings as appropriate. However, more detailed explanation than necessary may be omitted. For example, detailed explanation of already well-known matters or duplicate explanation of substantially the same configuration may be omitted. This is to avoid the following explanation becoming unnecessarily redundant and to facilitate understanding by those skilled in the art.

[0016] It should be noted that the accompanying drawings and the following description are provided to enable those skilled in the art to fully understand the present disclosure, and are not intended to limit the subject matter described in the claims.

[0017] (Embodiment) <Route Prediction System> FIG. 1 is a configuration diagram illustrating an example of a route estimation system 100 according to an embodiment of the present disclosure.

[0018] The path estimation system 100 estimates the steps performed by the worker 101 who moves between processes and performs work at the processes, the order of the steps performed, and the path taken between the processes to perform the steps (i.e., the state of movement between the processes of the portable sensor 102 linked to the worker 101), and presents the estimated steps, order, and path (hereinafter referred to as the estimation result) to a user (e.g., a process manager, a system manager, etc.). Note that movement in this disclosure includes not only movement between different processes, but also staying in the same process to perform the same work. In the following, estimating a path (or path estimation) means estimating the steps performed by the worker, the order of the steps performed, and the path taken between the processes to perform the processes.

[0019] The route estimation system 100 includes a portable sensor 102, process sensors 103-1 to 103-N (N is an integer greater than or equal to 2), a master station 104, a reception level information collection unit 105, a process information collection unit 106, a teacher data memory unit 107, a learning unit 108, a route estimation unit 109, and a display unit 110.

[0020] In this embodiment, the master station 104, reception level information collecting unit 105, process information collecting unit 106, teacher data storage unit 107, learning unit 108, route estimation unit 109 and display unit 110 are realized as one or more computer devices.

[0021] The route estimation system 100 is an example of a route estimation system according to the present disclosure. The reception level information collecting unit 105, the process information collecting unit 106, the teacher data storage unit 107, and the learning unit 108 are an example of a learning device according to the present disclosure. The reception level information collecting unit 105 and the process information collecting unit 106 are an example of a collecting unit according to the present disclosure. The learning unit 108 is an example of a learning unit according to the present disclosure. The route estimation unit 109 is an example of an estimating unit according to the present disclosure.

[0022] When there is no need to distinguish between the process sensors 103-1 to 103-N, they are referred to as process sensors 103. Furthermore, the portable sensor (wireless terminal device, third wireless station) 102, the process sensor (second wireless station) 103, and the master station (wireless base station device, first wireless station) 104 may be collectively referred to as wireless stations (wireless devices).

[0023] The mobile sensor 102, the process sensor 103, and the master station 104 are connected within the same wireless network, for example, by WiFi (registered trademark), Bluetooth (registered trademark), etc. For example, in FIG. 1, some of the wireless connections of these wireless stations are indicated by dotted lines.

[0024] Portable sensor 102 is carried by worker 101 performing work in a process. Here, the ID of portable sensor 102 and worker 101 (e.g., the ID of worker 101) are associated on a one-to-one basis. The associated portable sensor ID and, for example, the ID of worker 101 are stored in the form of a table in a storage unit (memory, etc.; not shown) of one or more of the computers. Portable sensor 102 periodically wirelessly transmits the ID of portable sensor 102 to master station 104.

[0025] When the portable sensor 102 receives or intercepts a wireless signal from another wireless station, the reception level such as the reception power and the received signal strength (RSSI: Receive Signal Strength Indicator) and the ID of the sender are linked to the time when the wireless signal was received, and stored as reception level information in the storage unit of the portable sensor 102. The portable sensor 102 periodically transmits the stored reception level information to the master station 104 by wireless.

[0026] The portable sensor 102 wirelessly transmits the ID of the portable sensor 102 and the reception level information at the same time to the master station 104. As a result, the ID of the portable sensor 102 and the reception level information are associated with each other.

[0027] The process sensors 103-1 to 103-N are installed at N work processes (also simply called processes). When a worker 101 works at a certain process, the process sensor 103 installed at that process wirelessly transmits to the master station 104 time-series process information including process operation information indicating that the process is operating (operational state of the process), i.e., that a certain worker 101-1 is present at the process, and process non-operation information indicating that the process is not operating (non-operational state of the process), i.e., that all workers 101 are not present at the process. Note that each process sensor 103 does not identify the ID of the worker 101 present at the process, and the process operation information is not linked to the worker 101 (e.g., the ID of the worker 101).

[0028] 1, for example, the arrow from the worker 101-1 to the process sensor 103-2 indicates the work in process "2" by the worker 101-1. At this time, the process sensor 103-2 wirelessly transmits the process operation information of the process "2" to the master station 104. The other process sensors 103 operate in the same manner.

[0029] When the process sensor 103 receives or intercepts a wireless signal from another wireless station, the process sensor 103 associates the reception level such as the reception power and RSSI and the ID of the sender with the time when the wireless signal was received, and stores the reception level information in the storage unit of the process sensor 103. The process sensor 103 periodically wirelessly transmits the stored reception level information to the master station 104.

[0030] The master station 104 receives the ID of the portable sensor 102 and reception level information of the portable sensor 102 wirelessly transmitted from the portable sensor 102, and receives the process information of the process sensor 103 and reception level information of the process sensor 103 wirelessly transmitted from the process sensor 103.

[0031] When master station 104 receives wireless signals from portable sensor 102 and process sensor 103, it links the reception levels such as the reception power and RSSI and the ID of the sender with the time when the wireless signals were received, and outputs them to reception level information collection unit 105 as reception level information of master station 104. Furthermore, master station 104 outputs the received ID of portable sensor 102, the reception level information of portable sensor 102, and the reception level information of process sensor 103 to reception level information collection unit 105. Furthermore, master station 104 outputs the received process information of process sensor 103 to process information collection unit 106.

[0032] The reception level information collecting unit 105 receives reception level information of all or some of the wireless stations in the wireless network, input from the master station 104. The reception level information collecting unit 105 generates and collects reception level information (also called Fingerprint (FP) or FP vector) in which the input reception levels are arranged in a predetermined order for each process (each process label) and vectorized, for example, as shown in FIG. 3 described below. The FP and FP sequence generated in this way are also called observed FP and observed FP sequence, or observed FP and observed FP sequence, respectively. Note that, although not shown in FIG. 3, information related to time may also be included.

[0033] When the learning unit 108 of the route estimation system 100 performs learning, the reception level information collecting unit 105 stores the generated FP in the teacher data storage unit 107 as input data for training.

[0034] During operation of the process (when it is desired to estimate the route of the worker 101 ), the reception level information collecting unit 105 outputs the generated FP sequence to the route estimating unit 109 .

[0035] The reception level information collecting unit 105 can capture a change in FP according to the position of the worker 101 by collecting FP at a speed that is sufficiently faster than the moving speed or working speed of the worker 101 .

[0036] The process information collecting unit 106 collects and consolidates the process operation information input from the master station 104 and sent from the process sensor 103 .

[0037] When the learning unit 108 of the path estimation system 100 performs learning, the process information collection unit 106 stores, in the teacher data storage unit 107, a process label indicating the process in which the worker 101 was present, as a training label (a process label as teacher data).

[0038] During operation of the process (when it is desired to estimate the path of the worker 101), the process information collection unit 106 outputs the aggregated process operation information as a process label sequence to the path estimation unit 109. An example of the process label sequence will be described later with reference to FIG.

[0039] The process information collection unit 106 outputs the collected process operation information to the display unit 110 as a process label sequence.

[0040] The teacher data storage unit 107 stores teacher data for learning an FP model for estimating the steps performed by the worker 101, the order in which the steps were performed, and the route traveled between steps to perform the steps (the state of movement between steps of the portable sensor 102 linked to the worker 101). The teacher data and FP model link the position of the worker (i.e., the step in which the worker is present and working, the step label) with the FP at that time. Therefore, the teacher data storage unit 107 stores teacher data that pairs the step label when the worker is in a certain step with the FP observed at that time.

[0041] The teacher data is collected in advance as a training model, for example, by moving the portable sensor 102 between a plurality of processes as the worker 101 carrying the portable sensor 102 performs the work in a predetermined process order, as a pair of a known process label and an FP observed when the worker 101 is in that process, and is stored in the teacher data storage unit 107. An example of the teacher data will be described later with reference to FIG.

[0042] The learning unit 108 learns the FP model using the process labels and FPs stored in the teacher data storage unit 107. The learning unit 108 outputs parameters of the learned FP model to the path estimation unit 109. Note that the FP model can be a generative model using various machine learning algorithms. In the following, an example will be described in which a conditional variational autoencoder (CVAE) is used as the FP model (generative model).

[0043] The path estimation unit 109 uses parameters of the FP model learned by the learning unit 108 to estimate a relationship between an FP sequence observed during operation of the process and a process label sequence, and estimates a path of the worker 101. The path estimation unit 109 outputs an estimation result including the estimated path of the worker 101 to the display unit 110.

[0044] The display unit 110 displays to the user a process label sequence during operation of the process, routes separated for each worker ((ID of) the portable sensor 102), and the like.

[0045] FIG. 2 is a diagram showing an example of a process label sequence according to the embodiment of the present disclosure.

[0046] 2, process labels "1" to "9" are assigned to the nine process sensors 103-1 to 103-9, respectively. The process information from each of the process sensors 103-1 to 103-9 is treated as time-series data. In other words, the process operation information, which is time-series data (also called a process label series (or a process operation information series)), is linked to time.

[0047] For example, process information from process sensor 103-1 is represented as time series data 201-1. The process information from process sensor 103-1 is represented as process non-operation information at L level when not in operation, and is represented as process operation information at H level when in operation. Furthermore, worker 101-1 and worker 101-2 work simultaneously at process 6 and process 9, respectively, and during the period when they work simultaneously at process 6 and process 9, time series data 201-6 and 201-9 are simultaneously at H level.

[0048] As described above, since the process operation information is not linked to the worker 101 (e.g., the ID of the worker 101), it is difficult to identify the worker 101-1 who was working in process 6 from the time-series data 201-6 alone.

[0049] FIG. 3 is a diagram showing an example of a process label and an FP according to the embodiment of the present disclosure.

[0050] Process step labels 301-1 to 301-N ("1" to "N" respectively assigned to process sensors 103-1 to 103-N) shown in FIG. 3 indicate the position (process) where worker 101 was.

[0051] A column vector (e.g., (-30 -50 . . . -90 -70)) of the reception levels of wireless signals received by the mobile sensor 102 from the process sensors 103-1 to 103-N and the master station 104 shown in FIG. T ) (unit: dBm) represents FP. The FP used as input data for training may be collected using one portable sensor 102 carried by a representative of the workers, or may be collected using multiple portable sensors 102 carried by multiple workers.

[0052] <Details of the study section> 4 is an (internal) configuration diagram showing an example of the learning unit 108 in the embodiment of the present disclosure. In this embodiment, CVAE is used as the generative model. The learning unit 108 performs learning of an encoder that compresses an input vector (FP) corresponding to a label into a low-dimensional latent variable, and a decoder that reconstructs the input vector (FP) using the input latent variable. The reconstructed FP and the reconstructed FP sequence are also called a reconstructed FP and a reconstructed FP sequence, respectively.

[0053] The learning unit 108 includes a generative model (FP model) 1084 including an encoder 1081, a latent variable 1082, and a decoder 1083, and an error learning unit 1085.

[0054] The encoder 1081 is configured by, for example, a neural network. The encoder 1081 receives as input a training FP vector and a training label (a process label as training data) stored in the training data storage unit 107, and obtains a low-dimensional mean vector and a variance vector based on these inputs.

[0055] The encoder 1081 samples a latent variable 1082 from the multivariate Gaussian distribution based on the obtained low-dimensional mean vector and variance vector. In this way, the encoder 1081 performs dimensional compression by imparting a certain degree of randomness to the mean vector, and obtains a latent variable 1082 after dimensional compression. The encoder 1081 outputs the latent variable 1082 to the decoder 1083.

[0056] The decoder 1083 is composed of, for example, a neural network. The decoder 1083 receives inputs of latent variables 1082 compressed to a low dimension and training labels (process labels as teacher data) stored in the teacher data storage unit 107, which are the same as those input to the encoder 1081, and reconstructs (the original high-dimensional) FP vectors based on these inputs. The decoder 1083 outputs the reconstructed FP vectors to the error learning unit 1085. Note that the randomness means that at least a part of the FP vectors reconstructed from the same process label varies. The randomness causes the vectors to be reconstructed into vectors that are slightly distant from each other in the vector space. This makes it possible to increase the variety of the reconstructed FP series.

[0057] The error learning unit 1085 adjusts parameters (e.g., weights, mean vectors, variance vectors, etc.) of the encoder 1081, latent variables 1082, and decoder 1083 so as to minimize the error between the reconstructed FP vector input from the decoder 1083 and the training FP vector input to the encoder 1081, and learns these parameters of the generative model 1084. Note that various optimization methods such as the stochastic steepest descent method can be used to learn these parameters.

[0058] The learning unit 108 outputs the parameters learned in this manner to the route estimation unit 109 for setting a generative model in the route estimation unit 109.

[0059] FIG. 5 is a flowchart illustrating an example of a training method 500 according to an embodiment of the present disclosure.

[0060] In step S501, pairs of N types of process labels and FP vectors observed in each process are prepared in advance as training teacher data, as shown in Fig. 3. Specifically, the reception level information collecting unit 105 and the process information collecting unit 106 store pairs of process labels and FP vectors as teacher data in the teacher data storage unit 107. The learning unit 108 uses the teacher data to learn the CVAE.

[0061] In step S502, the encoder 1081 obtains a low-dimensional mean vector and variance vector from the input FP vector and process label, and obtains a latent variable 1082 after dimensional reduction by sampling a latent variable 1082 from a multivariate Gaussian distribution based on the low-dimensional mean vector and variance vector.

[0062] In step S503, the decoder 1083 reconstructs an FP vector from the input process label and the latent variables 1082 compressed to a low dimension, which are the same as those input to the encoder 1081.

[0063] In step S504, the error learning unit 1085 learns parameters of the encoder 1081, the latent variable 1082, and the decoder 1083 so as to minimize the error between the FP vector input to the encoder 1081 and the reconstructed FP vector input from the decoder 1083.

[0064] If the error in step S504 is within a predetermined small value, that is, if a reconstructed FP sequence with a predetermined sufficiently small error has been generated for the training FP sequence, learning unit 108 ends learning (YES in step S505). On the other hand, if the error in step S504 is not within the predetermined small value, steps (S501,) S502, S503, and S504 are repeated (NO in step S505).

[0065] <Details of the route estimation section> FIG. 6 is an internal configuration diagram showing an example of the path estimation unit 109 in the embodiment of the present disclosure. The path estimation unit 109 generates multiple possible candidate sequences from process operation information (process label sequence) observed during the operation of the process. The path estimation unit 109 reconstructs an FP sequence based on the generated candidate sequence and a trained generative model. The path estimation unit 109 determines the similarity between the observed FP sequence linked to the ID of the portable sensor 102 and the reconstructed FP sequence. Then, the path estimation unit 109 selects a candidate sequence from the multiple candidate sequences based on the result of the similarity determination, and estimates the selected candidate sequence as the path of the worker 101.

[0066] The path estimation unit 109 includes a candidate sequence generation unit 1091, a trained generative model 1084a, and a similarity determination unit 1092.

[0067] The candidate sequence generation unit 1091 generates a plurality of possible candidate sequences from the observed process label sequence 602 input from the process information collection unit 106. The candidate sequence generation unit 1091 outputs the generated plurality of candidate sequences to the trained generative model 1084a and the similarity determination unit 1092. The observed process label sequence 602 is a sequence of process labels obtained from the process sensor 103.

[0068] The trained generative model 1084a includes a latent variable 1082a and a decoder 1083a having parameters trained by the learning unit 108 and input from the learning unit 108. Note that the latent variable 1082a and the decoder 1083a are the same as the latent variable 1082 and the decoder 1083 of the generative model 1084 of the learning unit 108, respectively, whereas the generative model 1084a differs from the generative model 1084 of the learning unit 108 in that it does not include an encoder.

[0069] For each candidate sequence, the generation model 1084a reconstructs an FP sequence using the process label of the candidate sequence input from the candidate sequence generation unit 1091 to the decoder 1083 and the random number input to the latent variable 1082a, and generates a reconstructed FP sequence 604.

[0070] In detail, the process label of the candidate sequence is input to the decoder 1083a, and a random number sampled from a multivariate Gaussian distribution according to the parameters of the latent variable 1082a is input to the latent variable 1082a, so that the generative model 1084a (decoder 1083a) reconstructs an FP sequence with a certain degree of randomness.

[0071] The generative model 1084 a (decoder 1083 a ) outputs the generated reconstructed FP sequence 604 to the similarity determining unit 1092 .

[0072] The similarity determination unit 1092 obtains the similarity between the observed FP sequence 601 input from the reception level information collection unit 105 and each of the multiple reconstructed FP sequences 604 input from the generation model 1084a (decoder 1083a), and determines which reconstructed FP sequence has the highest similarity. In this similarity determination, the similarity determination unit 1092 may use, for example, the Euclidean distance between vectors, and determine that a reconstructed FP sequence with a closer distance has a higher similarity. The similarity determination unit 1092 selects, from among the multiple candidate sequences, a candidate sequence that has resulted in a reconstructed FP sequence with the highest similarity as the prediction process performed by the worker, the prediction order in which the process was performed, and the prediction path moved between the processes, and estimates the selected prediction process, prediction order, and prediction path as the path of the worker 101. The similarity determination unit 1092 outputs the estimated result to the display unit 110.

[0073] FIG. 7 is a flowchart illustrating an example of a route estimation method 700 according to an embodiment of the present disclosure.

[0074] In step S701, the path estimation unit 109 sets the parameters learned by the learning unit 108 in the generative model 1084a (the latent variables 1082a and the decoder 1083a).

[0075] In step S702, the candidate sequence generation unit 1091 generates multiple candidate sequences from the observed process label sequence based on the inter-process movement probability table for a period for which path estimation is desired based on, for example, a user input to the display unit 110. Details of step S702 will be described later.

[0076] In step S703, the generative model 1084a reconstructs an FP sequence for each generated candidate sequence using the process label of the candidate sequence and a random number to generate a reconstructed FP sequence.

[0077] In step S704, the similarity determining unit 1092 finds the similarities between the generated multiple reconstructed FP sequences and the observed FP sequence.

[0078] In step S705, the similarity determining unit 1092 outputs the candidate sequence that yielded the reconstructed FP sequence with the highest similarity to the display unit 110 as the estimation result.

[0079] The route estimation unit 109 repeats steps S702 to S705 for each worker ((the ID of) the portable sensor 102) during the period for which route estimation is desired, thereby performing route estimation for each worker.

[0080] In step S706, the display unit 110 displays the observed process label sequence, the estimation results including the estimated route for each worker, and the like.

[0081] Details of step S702 will be described with reference to FIGS.

[0082] FIG. 8 is a diagram showing an example of candidate sequences 603-1 to 603-K in the embodiment of the present disclosure. For example, assume that a process label sequence 602 as shown in FIG. 8(a) is observed. When the labels of the processes that have moved along the time axis are extracted one by one in order from the process label sequence 602 in FIG. 8(a), a large number of combinations of candidate sequences 603-1 to 603-k are possible, as shown in FIG. 8(b). For example, the candidate sequence 603-1 represents the possibility of moving in the order of process 9 → process 8 → process 7 → process 5 → process 4 → . . . → process 3 → process 6. At this time, the movement (transition) probability between all processes is not necessarily equal. In view of the fact that the movement probabilities are not equal, the number of combinations of candidate sequences can be limited by considering that the probability of the movement considered in the production process design occurring is high.

[0083] 9 is a diagram showing an example of an inter-step movement probability table 901 according to an embodiment of the present disclosure. The candidate sequence generating unit 1091 may generate a candidate sequence in consideration of the inter-step movement probability as shown in FIG.

[0084] For example, the inter-process movement probability table 901 is a table in which the current process label is vertically represented, the next process label is horizontally represented, and the element at the intersection is the probability of movement from the current process label to the next process label. Basically, the probability of continuing work at the same process is high, followed by the probability of moving to an adjacent process that is close in location or production procedure. On the other hand, the probability of moving to a distant process is low. In this way, the probabilities in the inter-process movement probability table 901 can be dynamically adjusted according to factors such as the procedure design and layout of the production line, and even the proficiency of the workers.

[0085] For example, it is considered that there is a high probability that worker 101-3 in charge of process labels 6 to 3 will repeatedly perform these processes. Taking such a probability into consideration, candidate sequence generation unit 1091 generates candidate sequence 603-2 as shown in FIG 8 according to inter-process movement probability table 901 as shown in FIG 9.

[0086] Fig. 10 is a diagram showing an example of an observed FP sequence 601 and a reconstructed FP sequence 604 according to the embodiment of the present disclosure. (a) of Fig. 10 shows the reception levels of transmission and reception (wireless links) between the portable sensor 102, the process sensors 103-1 to 103-N, and the master station 104 from time T_1 to T_k.

[0087] For example, in FIG. 10B, the reconstructed FP vectors for each process label (6 to 3) of the candidate sequence 603 are horizontally (in time series) combined to form a reconstructed FP sequence 604.

[0088] In step S703, the generation model 1084a generates such reconstructed FP sequences in accordance with the number of generated candidate sequences.

[0089] In step S704, the similarity determination unit 1092 determines the similarity between the observed FP sequence 601 as shown in FIG. 10(a) and the reconstructed FP sequence 604 as shown in FIG. 10(b), and in step S705, outputs the candidate sequence that resulted in the reconstructed FP sequence with the highest similarity as the estimation result.

[0090] <Display of estimated results, etc.> FIG. 11 is a diagram showing an example of a display by display unit 110 according to the embodiment of the present disclosure.

[0091] 11, the display unit 110 displays a process label sequence (shown as process information) 1101 observed by the process information collection unit 106 and a route (shown as an estimated route) 1102 for each worker 101 estimated by the route estimation unit 109. In FIG. 11, a route 802-2 for worker A 101a corresponds to the candidate sequence 603-2 in FIG. 8.

[0092] Here, the route 1102 may be displayed with different line colors or patterns for each worker 101 so that the routes of the multiple workers 101 can be easily distinguished. For example, as shown in Fig. 11, the route of worker A 101a may be displayed with a solid line, and the route of worker B 101b may be displayed with a dotted line.

[0093] In this way, through the display unit 110, the user can determine, for example, the process label sequence, the processes performed by the worker 101, the order of the processes performed, and the route taken between the processes, and can make decisions regarding improvements to work efficiency and processes.

[0094] FIG. 12 is a diagram showing another example of a display by the display unit 110 according to the embodiment of the present disclosure.

[0095] 12, a heat map graph 1103 of the observed FP series 601 and a heat map graph 1104 of the inter-process movement probability table 901 may be displayed on the display unit 110 simultaneously with a process label series 1101 and a path 1102 in order to show the relevance to the observed FP series. By displaying the heat map graphs 1103 and 1104 in this manner, part of the path determination criteria can be presented to the user.

[0096] A heat map graph 1103 of the observed FP series 601 indicates the change over time in the reception level between the portable sensor 102 and the process sensors 103-1 to 103-N (wireless link). The darker the color, the higher the reception level. A heat map graph 1104 of the inter-process movement probability table 901 indicates the probability of movement from the current process to the destination process. Since worker A 101a moves between processes 3 and 6, processes 3 to 6 are colored darker.

[0097] (Modification) Needless to say, the present disclosure is not limited to the embodiments described above, and various modifications can be made without departing from the spirit of the present disclosure.

[0098] Although an example has been described above in which the similarity determination unit 1092 outputs the candidate sequence that resulted in the reconstructed FP sequence with the highest similarity as the estimation result to the display unit 110, the present disclosure is not limited to this example.

[0099] For example, instead of outputting the candidate sequence that resulted in the most similar reconstructed FP sequence to the display unit 110, the similarity determination unit 1092 may output a predetermined number of candidate sequences in descending order of similarity for each worker 101 or for one worker 101 to the display unit 110. In response to this, the display unit 110 may display an estimation result showing a predetermined number of candidate sequences in descending order of similarity.

[0100] Further, in the above description, an example has been described in which the column vector of the reception levels of wireless signals received by the portable sensor 102 from the process sensors 103-1 to 103-N and the master station 104 is used for FP, but the present disclosure is not limited to this example.

[0101] For example, a column vector of the reception levels of wireless signals received by portable sensor 102 from process sensors 103-1 to 103-N may be used for FP, or a column vector of the reception levels of wireless signals received by process sensors 103-1 to 103-N from portable sensor 102 may be used for FP.

[0102] Furthermore, for example, in addition to the reception levels of wireless signals received by the portable sensor 102 from the process sensors 103-1 to 103-N and the master station 104, a column vector of the reception levels of wireless signals received by the process sensor 103 from other process sensors 103 and the master station 104 may be used for FP.

[0103] (Effects of the embodiment) The learning device according to the embodiment of the present disclosure includes a collection unit that collects process operation information indicating the operation state of each process and reception level information including the reception level of a signal received by a first wireless station from a second wireless station installed in each process, and a learning unit that learns a generation model for estimating the movement state of a third wireless station using teacher data in which the process operation information and the reception level information are linked for each process. In this way, by learning a generation model linking the process operation information and the reception level information using teacher data in which the process operation information and the reception level information are linked for each process, it becomes possible to estimate the process performed by a worker who possesses the third wireless station and is linked to the third wireless station, the order in which the processes were performed, and the route taken between the processes, even when the reception level information is used. Similarly, by using a generative model that links process operation information and reception level information, the path estimation system in an embodiment of the present disclosure is able to estimate, even using reception level information, the processes performed by a worker who possesses a third wireless station and is linked to the third wireless station, the order in which the processes were performed, and the path taken between the processes.

[0104] In addition, the learning device and the path estimation system according to the embodiment of the present disclosure use a conditional variational autoencoder as a generative model. The variational autoencoder can respond to fluctuations in the wireless environment that are different from the time of learning by reconstructing a reception level information sequence that has a certain degree of randomness.

[0105] In addition, the path estimation system according to the embodiment of the present disclosure generates a plurality of candidate sequences from the process operation information sequence based on the movement probability between the processes, and estimates the movement state of the third wireless station based on the plurality of candidate sequences and the generation model. For example, the movement probability between the processes can be adjusted according to factors such as the procedure design and layout of the production line and the proficiency of the workers. This makes it possible to efficiently narrow down the number of combinations of the candidate sequences.

[0106] Moreover, the path estimation system according to the embodiment of the present disclosure includes a display unit that displays an estimation result of estimating the state of movement of the third wireless station. This allows a user to determine, for example, process operation information, the processes performed by a worker who owns the third wireless station and is linked to the third wireless station, the order in which the processes were performed, and the path taken between the processes, through the display unit, and thus allows the user to improve work efficiency and determine process improvement, etc.

[0107] In addition, the path estimation system according to the embodiment of the present disclosure displays an estimation result of estimating the movement state of a plurality of third wireless stations. For example, the path, etc., of a plurality of workers who each possess a plurality of third wireless stations and are associated with the plurality of third wireless stations and move between processes is displayed with different line colors or patterns for easy identification. This makes it easier to identify the path, etc., of the workers' movement between processes, so that the user can make decisions on improving work efficiency and processes.

[0108] In the above-described embodiments, the notation "... part" used for each component may be replaced with other notations such as "... circuit", "... assembly", "... device", "... unit", or "... module".

[0109] Although the embodiments have been described above with reference to the drawings, the present disclosure is not limited to such examples. It is clear that a person skilled in the art can come up with various modified or altered examples within the scope of the claims. It is understood that such modified or altered examples also belong to the technical scope of the present disclosure. In addition, the components in the embodiments may be arbitrarily combined within the scope of the present disclosure.

[0110] The present disclosure can be realized by software, hardware, or software in cooperation with hardware. Each functional block used in the description of the above embodiment may be realized partially or entirely as an LSI, which is an integrated circuit, and each process described in the above embodiment may be controlled partially or entirely by one LSI or a combination of LSIs. The LSI may be composed of individual chips, or may be composed of one chip that includes some or all of the functional blocks. The LSI may have input and output of data. Depending on the degree of integration, the LSI may be called an IC, a system LSI, a super LSI, or an ultra LSI.

[0111] The integrated circuit method is not limited to LSI, and may be realized by a dedicated circuit, a general-purpose processor, or a dedicated processor. Also, after LSI manufacturing, a programmable FPGA (Field Programmable Gate Array) or a reconfigurable processor that can reconfigure the connections and settings of circuit cells inside the LSI may be used. The present disclosure may be realized as digital processing or analog processing.

[0112] Furthermore, if a new integrated circuit technology that can replace LSI appears due to the progress of semiconductor technology or a derivative technology, it is possible to integrate the functional blocks using that technology. The application of biotechnology is also a possibility.

[0113] Summary of the Disclosure A learning device according to one embodiment of the present disclosure includes a collection unit that collects process operation information indicating the operation status of each process and reception level information including the reception level of a signal received by a first radio station from a second radio station installed in each process, and a learning unit that learns a generative model for estimating a movement status of a third radio station using teacher data that links the process operation information and the reception level information for each process.

[0114] In the above learning device, the generative model is a conditional variational autoencoder that includes an encoder that compresses the input reception level information into low-dimensional latent variables, and a decoder that reconstructs the reception level information using the latent variables.

[0115] In the above learning device, the learning unit learns parameters of the encoder, the latent variables, and the decoder so as to minimize an error between the reconstructed reception level information and the reception level information input to the encoder.

[0116] A path estimation system according to an embodiment of the present disclosure includes the learning device, an estimation unit configured to estimate a state of the movement based on the learning device, a process operation information sequence, a reception level information sequence including reception levels of signals received by the third radio station from a plurality of the second radio stations, and the generation model.

[0117] In the above-mentioned path estimation system, the estimation unit generates a plurality of candidate sequences from the process operation information sequence based on the movement probability between processes, reconstructs a plurality of reception level information sequences corresponding to the plurality of candidate sequences respectively based on the plurality of candidate sequences and the generation model, and selects a candidate sequence corresponding to the state of movement from among the plurality of candidate sequences based on the similarity between the reception level information sequence and each of the plurality of reconstructed reception level information sequences, thereby estimating the state of movement.

[0118] The above route estimation system further comprises a display unit that displays an estimation result of the movement state of the third wireless station.

[0119] In the above route estimation system, there are a plurality of the third wireless stations, and the display unit displays an estimation result obtained by estimating the movement states of the plurality of the third wireless stations.

[0120] A learning method according to one embodiment of the present disclosure collects process operation information indicating the operation status of each process and reception level information including the reception level of a signal received by a first radio station from a second radio station installed in each process, and uses training data linking the process operation information and the reception level information for each process to learn a generative model for estimating the movement status of a third radio station. [Industrial Applicability]

[0121] The present disclosure is useful for route estimation systems. [Explanation of symbols]

[0122] 100 Route Prediction System 101 Worker 102 Portable Sensor 103 Process Sensor 104 Master station 105 Reception level information collection unit 106 Process Information Collection Department 107 Teacher data storage unit 108 Learning Department 109 Route Estimation Unit 110 Display section 1081 Encoder 1082, 1082a Latent variables 1083, 1083a Decoder 1084, 1084a Generative Model 1085 Error Learning Unit

Claims

1. In wireless communication between a first wireless station, a second wireless station installed in each process, and a third wireless station carried by a worker who moves between processes, reception level information including at least a reception level value and transmission source information when a signal transmitted by another wireless station is received; process operation information indicating that each process is in operation when work is being performed in that process, and indicating that each process is in a non-operational state when work is not being performed in that process; Based on A learning device for estimating a moving route of the third wireless station, the third wireless station periodically transmits the reception level information of the reception signal; The second radio station periodically transmits the process operation information of the process in which the second radio station is installed and the reception level information of the reception signal; a collection unit that collects the reception level information received by the first wireless station from the third wireless station, and the process operation information and the reception level information received by the first wireless station from the second wireless station, and generates teacher data that associates the process operation information with the reception level information for each process; a learning unit that uses the teacher data to train a generation model to generate a reception level information sequence for estimating a movement state indicating that the third wireless station has moved or stopped at each of the steps; A learning device comprising:

2. The generative model is An encoder that compresses the input reception level information into a low-dimensional latent variable; a decoder that reconstructs the reception level information using the latent variables; is a conditional variational autoencoder including The learning device according to claim 1 .

3. The learning unit learns parameters of the encoder, the latent variables, and the decoder so as to minimize an error between the reconstructed reception level information and the reception level information input to the encoder. The learning device according to claim 2 .

4. A learning device according to any one of claims 1 to 3; an estimation unit that estimates a state of the movement based on a process operation information sequence, a reception level information sequence including reception levels of signals received by the third wireless station from a plurality of the second wireless stations, and the generation model; A path estimation system comprising:

5. the estimation unit generates a plurality of candidate sequences from the process operation information sequence based on a movement probability dynamically adjusted according to factors including a distance between processes and a process procedure, reconstructs a plurality of reception level information sequences corresponding to the plurality of candidate sequences based on the plurality of candidate sequences and the generation model, and selects a candidate sequence corresponding to the state of the movement from the plurality of candidate sequences based on a similarity between the reception level information sequence and each of the plurality of reconstructed reception level information sequences, thereby estimating the state of the movement. The route estimation system according to claim 4 .

6. a display unit for displaying an estimation result of the movement state of the third wireless station; Further comprising: The route estimation system according to claim 4 or 5.

7. a plurality of the third wireless stations are present, and the display unit displays an estimation result obtained by estimating the movement states of the plurality of the third wireless stations; The path estimation system of claim 6.

8. In wireless communication between a first wireless station, a second wireless station installed in each process, and a third wireless station carried by a worker who moves between processes, reception level information including at least a reception level value and transmission source information when a signal transmitted by another wireless station is received; process operation information indicating that each process is in operation when work is being performed in that process, and indicating that each process is in a non-operational state when work is not being performed in that process; Based on A learning method for a learning device for estimating a moving route of the third wireless station, comprising: the third wireless station periodically transmits the reception level information of the reception signal; The second radio station periodically transmits the process operation information of the process in which the second radio station is installed and the reception level information of the reception signal; The learning device includes: Collecting the reception level information received by the first wireless station from the third wireless station, and the process operation information and the reception level information received by the first wireless station from the second wireless station; generating teacher data for each process linking the process operation information with the reception level information; using the teacher data, a generation model is trained to generate a reception level information sequence for estimating a movement state indicating that the third wireless station has moved or stopped at each of the steps; How to learn.

Images