Method for generating a machine learning model for driver assistance to support the operation of an asphalt plant, driver assistance device for an asphalt plant, and asphalt plant

Abstract

This technology provides an asphalt plant that allows for the appropriate and simple reproduction of the manual driving techniques of skilled operators, even when the period for collecting operational data from skilled operators is limited, in accordance with specific conditions such as weather conditions. [Solution] A method for generating a machine learning model for assisting the operation of an asphalt plant, comprising a pre-training step S102 and a transfer learning step S105. Actual operation data to be used as training data is classified into two or more groups for specific conditions. In the pre-training step S102, pre-training is performed on an untrained machine learning model using actual operation data including all groups to obtain a basic model. In the transfer learning step S105, transfer learning is performed on the basic model using the actual operation data of a predetermined group to obtain a specific condition model. This makes it possible to appropriately and easily reproduce the manual operation techniques of a skilled asphalt plant operator according to specific conditions such as weather conditions.

Description

Technical Field

[0001] The present invention relates to a method for generating a machine learning model for operation support for an asphalt plant and an operation support device for assisting the operation of an asphalt plant.

Background Art

[0002] An asphalt plant has a dryer for heat-treating aggregates. Conventional asphalt plants are described, for example, in Patent Document 1. The asphalt plant described in Patent Document 1 has a fresh material dryer for heating and drying aggregates (fresh materials). Then, the aggregates heat-dried by these dryers are sieved by a vibrating screen provided at the upper part of the plant body and stored in an aggregate storage bin provided at the lower part of the vibrating screen for each particle size. Thereafter, the required amount of the aggregates is weighed from the aggregate storage bin and supplied to a mixer at the lower part of the plant body, and the mixer stirs and mixes them to produce an asphalt mixture.

[0003] Normally, the operation of the dryer of an asphalt plant can be switched between a manual operation mode and an automatic operation mode. In the automatic operation mode, the operator sets the target heating temperature of the aggregates on the operation panel, and the burner opening degree in the dryer is automatically operated. In the automatic operation mode, based on the sensor detection values such as the aggregate temperature at the dryer outlet, various control techniques such as PID control are used to automatically adjust the burner opening degree.

[0004] Such an automatic operation mode enables relatively stable operation when the temperature inside the drum of the dryer is kept at a predetermined temperature or higher. However, when starting the operation from a low state where the temperature inside the drum of the dryer is below the predetermined temperature, such as during nighttime operation suspension or long-term operation suspension, it is easily affected by conditions such as the weather, the water content of the input aggregates, and the supply amount of the input aggregates, and it is not always possible to perform stable operation.

[0005] When starting a dryer operation from a low drum temperature, the operator often adjusts the burner opening using manual operation mode. Experienced plant operators can appropriately control the burner opening based on conditions such as the moisture content of the aggregate being fed into the drum, the amount of aggregate supplied, and weather conditions, and heat the initially fed aggregate to the target aggregate temperature (160°C) before discharge.

[0006] On the other hand, if an inexperienced operator with relatively little experience operating the plant performs manual operation, it may be difficult to perform appropriate operations according to the conditions, as an experienced operator would. This can result in improper burner opening settings, leading to longer processing times for the aggregate discharged from the dryer outlet. Consequently, the time from the start of plant operation to the start of asphalt mixture production may be longer than in automatic operation mode.

[0007] To solve these problems, the applicant has filed the patent application shown in Patent Document 2. In the asphalt plant operation support device described in Patent Document 2, a trained model is created by learning the operation data of a skilled operator, and the burner opening predicted by the trained model is displayed on the control panel display, so that even an unskilled operator can appropriately and easily reproduce the manual operation techniques of a skilled operator. [Prior art documents] [Patent Documents]

[0008] [Patent Document 1] Japanese Patent Publication No. 2020-26647 [Patent Document 2] Japanese Patent Application No. 2023-165205 [Overview of the project] [Problems that the invention aims to solve]

[0009] In the conventional method described above, in order to create a highly predictive trained model that reproduces the proper operation of a skilled operator, it is necessary to train the model with operating data under conditions similar to those of the plant's actual operating conditions. This allows for the creation of a highly predictive trained model under operating conditions similar to those of the trained model. Plant operating conditions include, for example, ambient temperature, humidity, and aggregate moisture content. Specifically, examples include ambient temperature of 10-15°C, rainy weather, and wet aggregate.

[0010] However, attempting to collect operating data under similar conditions until sufficient data is available for machine learning can result in a long data collection period. For example, when the same collection period is set, the amount of operating data for the above-mentioned operating conditions of an outside temperature of 10-15°C, rainy weather, and wet aggregate tends to be less than the amount of operating data for the same outside temperature range but sunny weather and dry aggregate. Therefore, obtaining the necessary amount of data for these operating conditions tends to require a long data collection period. This means that it takes time to provide the trained model to users, and the challenge has been to create a trained model that can handle various operating conditions and has high predictive accuracy from operating data collected over a predetermined period.

[0011] The present invention has been made in view of the above points, and aims to provide a technology that can appropriately and easily reproduce the manual operation techniques of skilled operators in an asphalt plant, even when the period for collecting operational data from skilled operators in an asphalt plant is limited, in accordance with specific conditions such as weather conditions. [Means for solving the problem]

[0012] To solve the above problems, the first invention of the present application provides a method for generating a machine learning model for operation support for an asphalt plant, comprising: a) a pre-training step of performing pre-training on an untrained machine learning model to obtain a basic model which is the pre-trained machine learning model; and b) a transfer learning step of performing transfer learning on the basic model which is the transfer-trained machine learning model The process includes a transfer learning step to acquire a burner, and in steps a) and b), supervised learning is performed on the machine learning model so that when multiple input values ​​are input, it outputs a recommended burner opening, using multiple input values ​​of actual operating data as input variables and the burner opening operated by the operator as the output variable, and the actual operating data can be classified into two or more groups for a specific condition, in step a), pre-training is performed using the actual operating data including all of the groups for a specific condition, and in step b), transfer learning is performed using the actual operating data for one predetermined group for the specific condition.

[0013] The second invention of the present invention is a method for generating a machine learning model for driver assistance according to the first invention, wherein the machine learning model includes an input layer, a first LSTM layer, a second LSTM layer, a fully connected layer, and an output layer, wherein in step a), pre-training is performed on all layers of the untrained machine learning model, and in step b), transfer learning is performed on the fully connected layer and all output layers of the basic model.

[0014] The third invention of this application is a method for generating a machine learning model for driver assistance according to the second invention, wherein the output layer is a second fully connected layer.

[0015] The fourth invention of this application is a method for generating a machine learning model for driver assistance according to the second invention, wherein the input layer comprises a masking layer and a convolutional layer.

[0016] The fifth invention of this application is a method for generating a machine learning model for driving assistance according to the first invention, wherein the input values ​​input to the machine learning model include at least a target temperature which is a target value of the temperature of the aggregate at the outlet of the dryer, a post-heated aggregate temperature which is the temperature of the aggregate at the outlet of the dryer detected, and an aggregate supply amount which is the amount of aggregate supplied to the dryer.

[0017] The sixth invention of this application is a method for generating a machine learning model for driving assistance according to the fifth invention, wherein the input value further includes at least one of the following: the filter inlet temperature, which is the temperature of the exhaust gas in the exhaust flue at the inlet of the dust collection filter; the pre-heating aggregate temperature, which is the temperature of the aggregate supplied to the dryer; the aggregate temperature change rate, which is based on the temperature difference of the post-heating aggregate temperature at predetermined time intervals; the flue temperature, which is the temperature inside the exhaust flue; the ambient temperature outside the asphalt plant; and the humidity outside the asphalt plant.

[0018] The seventh invention of this application is an operation support device for the asphalt plant, comprising: a recommended opening degree output unit that outputs a recommended opening degree for the burner; and a display unit that displays the recommended opening degree to the operator, wherein the recommended opening degree output unit has the specific condition model generated by any one of the first to sixth inventions.

[0019] The eighth invention of the present application is an operation support device for an asphalt plant, comprising: a recommended opening degree output unit that outputs a recommended opening degree for the burner; a display unit that displays the recommended opening degree to the operator; and a learning unit that performs machine learning on the machine learning model, wherein the learning unit receives input the basic model generated by step a) of the method for generating a machine learning model for operation support according to any one of claims 1 to 6, or the specific condition model generated by steps a) and b) of the method for generating a machine learning model for operation support according to any one of the first to sixth inventions, and the learning unit is capable of executing step b) on the basic model and is capable of executing step b) again on the specific condition model.

[0020] The ninth invention of the present application is an asphalt plant comprising a dryer having a burner at one end of a drum for heating and drying aggregates, an exhaust flue for discharging exhaust gas from the dryer to the outside, a dust collection filter inserted in the exhaust flue, an input section operable by an operator for the burner opening degree which is the opening degree of the burner, and the operation support device described in the seventh invention.

Advantages of the Invention

[0021] According to the first to ninth inventions of the present application, in an asphalt plant, even if the collection period of the operation data of a skilled operator is limited, the manual operation technique of a skilled operator of the asphalt plant can be appropriately and simply reproduced according to specific conditions such as weather conditions.

Brief Description of the Drawings

[0022] [Figure 1] It is a diagram showing the configuration of an asphalt plant. [Figure 2] It is a diagram conceptually showing the configuration of an operation section and a control section of an asphalt plant. [Figure 3] It is a flowchart showing the flow of a method for generating a machine learning model M used for operation support of an asphalt plant. [Figure 4] It is a diagram showing an example of the configuration of a machine learning model M used for operation support of an asphalt plant. [Figure 5] It is a diagram showing the configuration of a comparison model Mc obtained by deleting a fully connected layer from the machine learning model M. [Figure 6] It is a diagram showing the recommended opening degrees output by the machine learning model M and the comparison model Mc. [Figure 7] It is a diagram conceptually showing the configuration of an operation section and a control section of an asphalt plant according to a modification example.

Embodiments for Carrying Out the Invention

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

[0024] <1. Embodiments> <1-1. Configuration of an asphalt plant> Figure 1 shows the configuration of asphalt plant 1. This asphalt plant 1 produces asphalt mixture by mixing new material, recycled material, molten asphalt, and filler (stone powder). As shown in Figure 1, asphalt plant 1 comprises a new material hopper 10, a recycled material hopper 20, a new material dryer 30, a recycled material dryer 40, an exhaust path 50, and a plant body 90.

[0025] The new aggregate hopper 10 temporarily stores the new aggregate for the asphalt mixture (hereinafter referred to as "new aggregate") that has been brought into the asphalt plant 1. The new aggregate hopper 10 supplies the new aggregate to the drum 31 of the new aggregate dryer 30. Various types of natural coarse aggregate (natural gravel), crushed stone, natural fine aggregate (natural sand), crushed sand, etc., can be used as aggregate for the new aggregate.

[0026] New material of each particle size is stored in a new material hopper 10 according to the particle size, and after the required amount of new material according to the mix design of the asphalt mixture to be manufactured is cut out from each new material hopper 10, all the cut new material is supplied together to the drum 31 of the new material dryer 30.

[0027] The new material dryer 30 is a device that heats and dries new material supplied from the new material hopper 10. The new material dryer 30 comprises a drum 31, a machine base 32, and a burner 33. The drum 31 has a new material dryer inlet 301 at one end into which new material is fed from the new material hopper 10, and a new material dryer outlet 302 at the other end into which the heated and dried new material is discharged. The machine base 32 is provided to support the drum 31 from below so as to be rotatable. The burner 33 is located at the other end of the new material dryer 30.

[0028] The new material dryer 30 is connected to the plant body 90 by a vertical conveying device 101. Specifically, the new material dryer outlet 302 is connected to the upstream end of the vertical conveying device 101. In addition, the upper part of one end of the drum 31 is connected to the first exhaust flue 51 of the exhaust path 50, which will be described later.

[0029] The new material dryer 30 has scraping blades (not shown) on the inner circumference of the drum 31. The scraping blades rotate at a predetermined speed together with the drum 31 by a drive device (not shown), agitating the new material supplied to the drum 31. The burner 33 heats the new material supplied to the drum 31 to a predetermined temperature and performs a heat drying treatment.

[0030] The vertical conveying device 101 transports the new material, which has undergone heat drying treatment in the new material dryer 30, to the plant body 90. The exhaust gas generated by the heating of the new material is discharged into the first exhaust flue 51.

[0031] The recycled material hopper 20 temporarily stores the recycled aggregate (recycled material) that has been brought into the asphalt plant 1. The recycled material hopper 20 supplies the recycled material to the drum 41 of the recycled dryer 40.

[0032] The regeneration dryer 40 is a device in the asphalt plant 1 that heat-treats the recycled material supplied from the recycled material hopper 20. The regeneration dryer 40 comprises a drum 41, a machine base 42, a burner 43, a recycled material storage bin 44, and a recycled material weighing tank 45.

[0033] The drum 41 has a regeneration dryer inlet 401 at one end through which recycled material is fed from the recycled material hopper 20, and a regeneration dryer outlet 402 at the other end through which the heated and dried recycled material is discharged. The machine base 42 is provided to support the drum 41 so as to be rotatable from below. The burner 43 is located at one end of the regeneration dryer 40.

[0034] The recycled material storage bin 44 is located at the other end of the regeneration dryer 40. The recycled material storage bin 44 is connected to the regeneration dryer outlet 402 of the drum 41. Therefore, the recycled material that has undergone heat drying treatment is supplied to the recycled material storage bin 44 from the drum 41.

[0035] The recycled material weighing tank 45 is located at the lower end of the recycled material storage bin 44. The recycled material weighing tank 45 weighs the recycled material stored inside the recycled material storage bin 44 and supplies it downwards.

[0036] The regeneration dryer 40 is connected to the plant body 90 by an input chute 102. Specifically, the upstream end of the input chute 102 is connected to the discharge port at the bottom of the recycled material weighing tank 45. The top of the recycled material storage bin 44 is connected to the second exhaust flue 52 of the exhaust path 50, which will be described later.

[0037] The regeneration dryer 40 has scraping blades (not shown) on the inner circumference of the drum 41. The scraping blades rotate at a predetermined speed together with the drum 41 by a drive device (not shown), agitating the regenerated material supplied to the drum 41. The burner 43 heats the regenerated material supplied to the drum 41 to a predetermined temperature and performs a heat drying treatment.

[0038] The recycled material storage bin 44 temporarily stores the heated recycled material. The recycled material weighing tank 45 weighs a predetermined amount of recycled material stored in the recycled material storage bin 44 and supplies it to the input chute 102. The input chute 102 transports the predetermined amount of recycled material to the plant body 90. The exhaust gas generated by heating the recycled material is discharged from the recycled material storage bin 44 to the second exhaust flue 52.

[0039] The exhaust path 50 is a gas discharge path consisting of piping and the like that leads exhaust gas from the new material dryer 30 and the regenerated dryer 40, which are the sources of exhaust gas, to the outside. The exhaust path 50 has a first exhaust flue 51, a second exhaust flue 52, and a common flue 53. The upstream end of the first exhaust flue 51 is connected to the top of the drum 31 of the new material dryer 30. The upstream end of the second exhaust flue 52 is connected to the top of the recycled material storage bin 44 of the regenerated dryer 40. The downstream end of the first exhaust flue 51 and the downstream end of the second exhaust flue 52 are connected to the upstream end of the common flue 53. As a result, the exhaust gas discharged from the new material dryer 30 and the exhaust gas discharged from the regenerated dryer 40 merge and flow through the common flue 53.

[0040] The common flue 53 has, in order from upstream, an inertial dust collector 61, a filter dust collector 62, and an exhaust fan 63. The downstream end of the common flue 53 is connected to the chimney 64. The exhaust fan 63 creates an exhaust gas flow within the common flue 53 from upstream to downstream. As a result, the exhaust gas flowing into the common flue 53 from the first exhaust flue 51 and the second exhaust flue 52 has its dust removed by the inertial dust collector 61 and the filter dust collector 62 before being discharged from the chimney 64. The inertial dust collector 61 and the filter dust collector 62 are collectively referred to as the "dust collection filter".

[0041] The inertial dust collector 61 separates dust particles with relatively large particle sizes from the exhaust gas. For example, a cyclone-type dust collector is used for the inertial dust collector 61. The relatively large particle size dust separated in the inertial dust collector 61 is transported to the vicinity of the new material dryer outlet 302 by a transport device (not shown in the figure), and together with the new material that has been heat-treated in the new material dryer 30, it is transported to the plant body 90 by a vertical transport device 101. The exhaust gas discharged from the inertial dust collector 61 is introduced into the downstream filtration dust collector 62.

[0042] The filter dust collector 62 separates fine dust (particulate matter) contained in the exhaust gas. The filter dust collector 62 captures and removes the fine dust using a filter cloth. The fine dust removed by the filter cloth is then transported via the pressurized piping 103 to the filler storage bin 94 of the plant body 90, which will be described later. The filter dust collector 62 discharges the exhaust gas from which the dust has been removed to the downstream side of the common flue 53. The exhaust gas from which the dust has been removed by the inertial dust collector 61 and the filter dust collector 62 is then discharged from the chimney 64.

[0043] The plant body 90 is equipment for producing an asphalt mixture by mixing heat-dried new material, heat-dried recycled material, filler, and molten asphalt. The plant body 90 includes a vibrating screen 91, an aggregate storage bin 92, an aggregate weighing tank 93, a filler storage bin 94, a filler weighing tank 95, a molten asphalt tank 96, a supply pump 97, an asphalt weighing tank 98, and a mixer 99.

[0044] The vibrating screen 91 receives the new material supplied from the vertical conveying device 101. The vibrating screen 91 sieves the received new material and supplies it to the aggregate storage bin 92 according to particle size. In other words, the aggregate storage bin 92 is an aggregate storage section that stores the aggregate (new material) dried by the new material dryer 30. The aggregate weighing tank 93 weighs the new material stored in the aggregate storage bin 92 according to particle size and supplies it to the mixer 99.

[0045] The filler storage bin 94 is a filler storage unit that temporarily stores the stone powder brought into the asphalt plant 1 and the powder transported from the filtration dust collector 62 via the pressurized piping 103. In the following description, the stone powder brought into the asphalt plant 1 and the powder transported from the pressurized piping 103 will be collectively referred to as "filler". The filler weighing tank 95 weighs a predetermined amount of filler stored in the filler storage bin 94 and supplies it to the mixer 99.

[0046] The molten asphalt tank 96 is a molten asphalt storage unit that temporarily stores molten asphalt, which is the material for the asphalt mixture. The supply pump 97 is an asphalt transport unit that transports the molten asphalt from the molten asphalt tank 96 to the asphalt weighing tank 98. The asphalt weighing tank 98 weighs the asphalt transported by the supply pump 97 and supplies it to the mixer 99.

[0047] Mixer 99 produces an asphalt mixture by mixing new aggregate supplied from aggregate weighing tank 93, recycled aggregate supplied from recycled aggregate weighing tank 45, filler supplied from filler weighing tank 95, and molten asphalt supplied from asphalt weighing tank 98 for a predetermined time.

[0048] <1-2. Regarding the control unit of the asphalt plant> The asphalt plant 1 has a control unit 80 and an operating unit 700 for controlling the above-mentioned parts. Figure 2 is a conceptual diagram showing the configuration of the control unit 80 and the operating unit 700 of the asphalt plant 1.

[0049] The control unit 700 is a device for the operator to perform operations. As shown in Figure 2, the control unit 700 includes an input unit 701 and a display unit 702.

[0050] The input unit 701 is equipped with numerous operation buttons, levers, or equivalent GUIs for operating various parts of the asphalt plant 1. The input unit 701 allows input of operating command values ​​and target values ​​for each part of the asphalt plant 1 to the control unit 80.

[0051] The following are some examples of the operating command values ​​that can be input to the input unit 701. • The first supply amount T11 is the amount of new material supplied from the new material hopper 10 to the new material dryer 30, for each particle size. • Second supply amount T12, which is the amount of recycled material supplied from the recycled material hopper 20 to the recycled material dryer 40. • The opening degree of the first burner T21 of the new material dryer 30, which is the opening degree of the burner 33. • The opening degree of the burner 43 of the regenerating dryer 40 is the second burner opening degree T22

[0052] Furthermore, the target values ​​that can be input to the input unit 701 are, for example, as follows: • The first target temperature T31 is the target temperature of the new material discharged from the new material dryer 30. • The second target temperature T32 is the target temperature for the recycled material discharged from the regeneration dryer 40.

[0053] The display unit 702 can display the operating status of each part of the asphalt plant 1, as well as the detected values ​​from sensors installed in each part of the asphalt plant 1. The display unit 702 can also display the recommended opening degree Do received from the recommended opening degree output unit 84, which will be described later.

[0054] As shown in Figure 1, the control unit 80 is composed of a computer consisting of, for example, a processor 801 such as a CPU or GPU, memory 802 such as RAM, and a storage medium 803 such as a hard disk drive or SSD. The control unit 80 controls the operation of each part within the asphalt plant 1 by having the processor 801 perform calculations based on computer programs P and data D stored in the storage medium 803.

[0055] Furthermore, as shown in Figure 2, the control unit 80 includes an operation control unit 81, an automatic driving control unit 82, a learning unit 83, and a recommended opening degree output unit 84, all of which are functionally implemented in software by such a computer. The control unit 80 also includes a storage unit 89.

[0056] The control unit 80 receives operating command values ​​and target values ​​input from the input unit 701. The control unit 80 also receives various detection values ​​detected by sensors installed in various parts of the asphalt plant 1. Based on these values, the control unit 80 controls the operation of each part of the asphalt plant 1.

[0057] The detected values ​​input to the control unit 80 are, for example, as follows: The temperature sensor S11 located at the outlet 302 of the new material dryer detects the temperature of the new material discharged from the new material dryer 30, which is the first aggregate temperature D11. The temperature sensor S12 located at the regeneration dryer outlet 402 detects the temperature of the recycled material discharged from the regeneration dryer 40, which is the second aggregate temperature D12. The temperature sensor S21 located at the connection point between the first exhaust flue 51 and the new material dryer 30 detects the temperature of the exhaust gas discharged from the new material dryer 30, which is the first flue temperature D21. The temperature sensor S22 located at the connection point between the second exhaust flue 52 and the regenerative dryer 40 detects the temperature of the exhaust gas discharged from the regenerative dryer 40, which is the second flue temperature D22. The temperature sensor S3 located at the inlet of the inertial dust collector 61 in the common flue 53 detects the filter inlet temperature D3. • The external temperature D4 is detected by the temperature sensor S4 installed outside the asphalt plant 1. • External humidity D5 detected by humidity sensor S5 located outside asphalt plant 1

[0058] The motion control unit 81 controls the operation of each part of the asphalt plant 1. The motion control unit 81 can be switched between automatic operation mode and manual operation mode. In manual operation mode, the motion control unit 81 controls the operation of each part of the asphalt plant 1 based on operation commands input by the operator from the input unit 701. In manual operation mode, the operator controls the burner opening of the burner 33 of the new material dryer 30 according to the first burner opening T21 input by the operator from the input unit 701, and controls the burner opening of the burner 43 of the regenerated dryer 40 according to the second burner opening T22 input by the operator from the input unit 701.

[0059] In automatic operation mode, the automatic operation control unit 82 calculates the first burner opening T21 and the second burner opening T22 based on the first target temperature T31 and the second target temperature T32 input from the input unit 701, and the temperatures D11, D12, D21, D22, and D3 input from the temperature sensors S11, S12, S21, S22, and S3 in the asphalt plant 1. Then, the operation control unit 81 controls the burner openings of the burner 33 of the new material dryer 30 and the burner 43 of the regenerated dryer 40 according to the first burner opening T21 and the second burner opening T22 calculated by the automatic operation control unit 82.

[0060] The automatic operation control unit 82 calculates the first burner opening T21 and the second burner opening T22 using feedback control such as PID control and feedforward control. Therefore, during startup, shutdown, or unsteady operation of the asphalt plant 1, if the plant is operated in automatic operation mode, depending on conditions such as weather, aggregate moisture content, and aggregate supply, the aggregate discharged from the outlets of dryers 30 and 40 may not reach the aggregate temperature (160°C) suitable for manufacturing asphalt mixtures. For this reason, during unsteady operation, it is preferable for the operator to manually operate the plant in manual operation mode, referring to the recommended opening Do output by the recommended opening output unit 84.

[0061] The learning unit 83 performs machine learning on the machine learning model M. In this embodiment, as shown in Figure 2, the control unit 80 receives a basic model M1 that has been pre-trained on the machine learning model M0 before training by the pre-training device 70. Here, the machine learning model M0 before training, the pre-trained basic model M1, and the specific condition models M21, M22, etc., which have undergone transfer learning and re-transfer learning described later, are collectively referred to as the machine learning model M.

[0062] The pre-learning device 70 includes, for example, a memory unit 71 and a pre-learning unit 72. The memory unit 71 stores a large number of actual operation data Ds0, which are time-series data including each sensing data and burner opening T21 from past operations of the asphalt plant 1.

[0063] The pre-training unit 72 creates training data from these numerous actual operating data Ds0, using each of the above input values ​​as input variables and the first burner opening T21 as the output variable, and pre-trains the machine learning model M using this training data. The training completion condition in the learning unit 83 is, for example, when the difference between the recommended opening output by the machine learning model M and the first burner opening T21 in the training data is less than or equal to a predetermined threshold. The trained machine learning model M, after pre-training is complete, is handed over to the control unit 80 of the asphalt plant 1.

[0064] Furthermore, if the number of past operational data points for asphalt plant 1 is insufficient, such as when asphalt plant 1 is newly installed, the operational data Ds0 used for pre-training may include data from another asphalt plant of the same type as asphalt plant 1.

[0065] Then, the learning unit 83 performs transfer learning on the basic model M1 input from the pre-learning device 70 using specific condition data (aggregate drying data Ds1 and aggregate wetting data Ds2, described later) which are actual operating data for different conditions stored in the memory unit 89.

[0066] The memory unit 89 stores numerous sets of actual operating data, which are time-series data including sensing data and burner opening T21 from past operations of the asphalt plant 1. The learning unit 83 classifies these multiple sets of actual operating data into two or more groups based on specific conditions.

[0067] In this embodiment, these multiple operational data are divided into two groups: an aggregate drying group, which includes aggregate drying data Ds1, which is operational data of the aggregate in a dry state; and an aggregate wetting group, which includes aggregate wetting data Ds2, which is operational data of the aggregate in a wet state. In this embodiment, for ease of understanding the invention, there are two groups for conditional division, but the number of conditional groups may be three or more.

[0068] The learning unit 83 performs transfer learning on the basic model M1 for each group. The differences in the learning methods of the machine learning model M between pre-training and transfer learning will be described later.

[0069] In this embodiment, the learning unit 83 creates training data from multiple aggregate drying data Ds1, using each input value as an input variable and the first burner opening T21 as an output variable, and performs transfer learning of the basic model M1 using this training data. The learning completion condition in the learning unit 83 is, for example, when the difference between the recommended opening output by the machine learning model M and the first burner opening T21 in the training data is less than or equal to a predetermined threshold. The trained machine learning model M, which has completed transfer learning using the aggregate drying data Ds1, is passed to the recommended opening output unit 84 as the aggregate drying model M21, which is a specific condition model.

[0070] Furthermore, the learning unit 83 creates training data from multiple aggregate wet data sets Ds2, using each input value as an input variable and the first burner opening T21 as an output variable, and performs transfer learning of the basic model M1 using this training data. The trained machine learning model M, which has completed transfer learning using the aggregate wet data Ds2, is passed to the recommended opening output unit 84 as the aggregate wet model M22, which is a specific condition model.

[0071] During the operation of the asphalt plant 1, the operator may, as needed, operate the input unit 701 to cause the recommended opening degree output unit 84 to output a recommended opening degree Do1. At that time, the operator selects one of several specific condition models. In this embodiment, as shown in Figure 2, the recommended opening degree output unit 84 can select not only the aggregate drying model M21 and aggregate wetting model M22, which are specific condition models delivered from the learning unit 83, but also the basic model M1 delivered from the pre-learning device 70. This allows the use of the basic model M1 when it is unclear which group of specific condition models is appropriate or when the conditions are unknown.

[0072] The recommended opening degree output unit 84 outputs a recommended opening degree Do1 using one machine learning model M selected by the operator from among the specific condition models M21, M22 and the basic model M1. Specifically, it inputs values ​​to the selected machine learning model M from various values ​​input from the input unit 701 or detected values ​​input from various sensors.

[0073] The recommended opening degree output unit 84 displays the recommended opening degree Do1 output from the selected machine learning model M on the display unit 702. This allows the operator to manually operate the system using the input unit 701 while visually confirming the recommended opening degree Do1 displayed on the display unit 702. In this embodiment, the driving support device 100 is composed of the learning unit 83, the recommended opening degree output unit 84, the input unit 701, and the display unit 702.

[0074] For example, machine learning models M1, M21, and M22 output a recommended opening degree Do1 during the startup of asphalt plant 1, and their input variables are three values: a first target temperature T31 input from input unit 701, a first aggregate temperature D11 input from temperature sensor S11, and a first aggregate supply amount calculated from the first supply amount T11.

[0075] In this manner, the driver assistance device 100 outputs recommended opening degrees Do1 and Do2 using trained machine learning models M1, M21, and M22, which have been trained using actual driving data from skilled operators and actual driving data with good fuel efficiency, and displays them on the display unit 702.

[0076] When manually operating using the driving support device 100, the operator can adjust the burner opening T21 of the new material dryer 30 by referring to the recommended opening output by the trained machine learning model M. At this time, the operator first selects a machine learning model M that has undergone transfer learning in a group that matches the current operating conditions, if such a model exists. For example, if there has been no rainfall for three days or more, the operator selects the aggregate drying model M21, and otherwise selects the aggregate wetting model M22. If it is determined that there is no suitable model for specific conditions, the operator may select the basic model M1, which has not undergone transfer learning.

[0077] This allows even inexperienced operators with relatively little experience operating asphalt plant 1 to operate the plant manually with appropriate burner openings T21 and T22. As a result, the aggregate temperature can be efficiently raised to the desired temperature while suppressing burnout of the dust collection filters (inertial dust collector 61 and filtration dust collector 62). In other words, manual operation techniques can be appropriately and easily transmitted without the need for instruction from skilled operators.

[0078] In this embodiment, the specific condition for selecting actual operating data for transfer learning was the moisture content of the aggregate, but the present invention is not limited to this. The specific condition may be grouped by season, outside temperature, weather, etc. Furthermore, the specific condition may be, for example, actual operating data when a specific skilled operator is in charge of operation, actual operating data where fuel efficiency is better than a predetermined threshold, or actual operating data where the time until the start of asphalt mixture production is short. Thus, the specific condition for selecting actual operating data can be determined from various perspectives.

[0079] Furthermore, the actual operating data used for pre-training does not have to be limited to data from the actual asphalt plant 1 being used. For example, when introducing a new asphalt plant 1, a basic model M1 can be created using actual operating data from multiple existing asphalt plants of the same type, and as soon as actual operating data for the new asphalt plant 1 is obtained, transfer learning can be performed using that data. In this way, a machine learning model M suitable for the new asphalt plant 1, for which sufficient actual operating data has not yet been obtained, can be obtained.

[0080] <1-3. About the construction and training of machine learning models> Next, the configuration of the machine learning model M and the pre-training and transfer learning of the machine learning model M will be described. In this embodiment, the machine learning model M is a model for outputting recommended burner openings Do1 and Do2, which are estimated to be appropriate burner openings T21 and T22, during startup, shutdown, or non-steady-state operation such as sudden increases or decreases in aggregate supply of the asphalt plant 1. The machine learning model M, which is trained in the learning unit 83, outputs recommended burner openings Do1 and Do2 when multiple input values ​​are input.

[0081] Figure 3 shows an example of the configuration of a machine learning model M used to assist the operation of asphalt plant 1. In the example in Figure 3, this machine learning model M has, in order from the input side to the output side, an input layer L1, a first LSTM layer L2, a second LSTM layer L3, a fully connected layer L4, and an output layer L5. LSTM (Long Short Term Memory) is a type of RNN (Recurrent Neural Network) that is suitable for learning and predicting time series data and can learn both long-term and short-term features.

[0082] In this embodiment, the input layer L1 consists of a masking layer and a convolutional layer. However, the input layer L1 may take other forms. Also, in this embodiment, the output layer L5 is a fully connected layer. That is, the output layer L5 is a second fully connected layer.

[0083] Figure 4 is a flowchart showing the learning process of the machine learning model M in the pre-learning device 70 and the driving support device 100 of this embodiment.

[0084] As shown in Figure 4, first, the pre-training device 70 prepares the machine learning model M0 and a large dataset (step S101). Specifically, a large number of past actual operation data Ds0 are stored in the storage unit 71 of the pre-training device 70. This actual operation data Ds0 may include not only actual operation data of the asphalt plant 1 equipped with the driving support device 100, but also actual operation data of similar asphalt plants. Note that "a large number" means that the number of actual operation data Ds0 used for pre-training is greater than the number of specific condition data Ds1 and Ds2 used for the initial transfer learning. Note that, as will be described later, transfer learning may be performed multiple times if specific condition data Ds1 and Ds2 are added.

[0085] Next, the pre-training unit 72 of the pre-training device 70 performs supervised learning on all layers L1 to L5 of the machine learning model M0 before training (step S102). Specifically, the pre-training unit 72 creates training data from these numerous actual operating data Ds0, using each input value as an input variable and the first burner opening T21 as an output variable, and performs supervised learning of the machine learning model M0 using this training data. As a result, the pre-training unit 72 generates the basic model M1.

[0086] The driver assistance device 100 obtains the basic model M1 generated by the pre-learning in step S102 from the pre-learning device 70 (step S103).

[0087] Meanwhile, the driving support device 100 prepares specific condition data Ds1 and Ds2 from the actual operating data of the asphalt plant 1 that match specific conditions (step S104). In this embodiment, multiple aggregate drying data Ds1 and multiple aggregate wetting data Ds2 are selected from the actual operating data of the asphalt plant 1.

[0088] Then, the learning unit 83 performs transfer learning on some layers of the basic model M1 using aggregate drying data Ds1 (step S105). Specifically, the learning unit 83 creates training data from multiple aggregate drying data Ds1, using each input value as an input variable and the first burner opening T21 as an output variable. Then, the learning unit 83 performs transfer learning of the basic model M1 on two layers, the fully connected layer L4 and the output layer L5, using this training data. Similarly, the learning unit 83 also performs transfer learning of the basic model M1 on aggregate wet data Ds2.

[0089] In step S105, the learning unit 83 acquires specific condition models, namely the aggregate drying model M21 and the aggregate wetting model M22, and passes them on to the recommended opening output unit 84 (step S106). At the same time, the learning unit 83 also passes on the basic model M1 to the recommended opening output unit 84. As a result, when the asphalt plant 1 is in operation, the recommended opening output unit 84 can output the recommended opening Do1 using either the aggregate drying model M21, the aggregate wetting model M22, or the basic model M1.

[0090] In this way, by performing transfer learning, it is possible to generate highly accurate recommended opening degree output models tailored to each specific condition, even when there is insufficient actual operating data under certain conditions.

[0091] Generally, in deep learning of machine learning models, the input layers tend to learn general and universal features and knowledge, while the output layers tend to learn features and knowledge adapted to specific situations and tasks. Therefore, in this embodiment, transfer learning is performed on the two most output layers of the basic model M1, the fully connected layer and the output layer, using specific condition data Ds1 and Ds2, which have a small amount of data. This makes it possible to generate a highly accurate recommended opening degree output model even when the amount of data Ds1 and Ds2 for specific conditions is insufficient.

[0092] Here, we present empirical data on the effectiveness of transfer learning. In a certain asphalt plant, 285 operational data points were obtained as all operational data (unconditional data) without distinguishing between dry and wet aggregate conditions. Of these, 37 were conditional data points that met the specific condition that the previous day's rainfall was 2 mm or more and the aggregate was in a wet state.

[0093] Pre-training and transfer learning of the machine learning model M were performed on this actual operating data under the following three conditions. (Condition 1) Pre-training: 37 conditional data points, Transfer learning: None (Condition 2) Pre-training: 37 conditional data points, Transfer training: 15 conditional data points (Condition 3) Pre-training: 285 unconditional data points, Transfer training: 15 conditional data points The 15 specific condition data points used in transfer learning under conditions 2 and 3 were randomly selected from the 37 specific condition data points used in pre-training under conditions 1 and 2.

[0094] Table 1 below shows the evaluation of the machine learning model M created under the above conditions 1, 2, and 3, using the Mean Absolute Error (MAE) between the output recommended burner opening and the actual burner opening in the training data. [Table 1]

[0095] As shown in Table 1, in Condition 2, the MAE value is high despite further transfer learning being performed on the machine learning model generated in Condition 1. This may be because, with a small amount of data used in pre-training, general and universal features were not sufficiently learned, and transfer learning was performed using more than 40% of the overlapping data used in pre-training. As a result, despite insufficient training, skewed learning such as overfitting may have occurred.

[0096] In contrast, under condition 3, pre-training is performed on a large amount of unconditional data, and transfer learning is performed on a portion of that data (approximately 5.3%) that matches the specific conditions. As a result, it is thought that the model learns sufficiently general and universal features during pre-training, and then learns features under specific conditions through transfer learning, enabling it to output the recommended burner opening with high accuracy. The empirical data above also shows that by performing pre-training and transfer learning, the recommended burner opening can be estimated with high accuracy in asphalt plant 1.

[0097] <1-4. Regarding the layer for model selection and transfer learning> In this embodiment, the empirical data that led to the selection of the above-described model configuration for the machine learning model M is described below.

[0098] As shown in Figure 3, the machine learning model M of this embodiment has two fully coupled layers: a fully coupled layer L4 and an output layer L5. By having two fully coupled layers in this way, when the burner opening is changed significantly to a specific value, such as when the dryer starts up, the estimated opening output from the machine learning model M can be made to closely resemble the actual burner opening performed by the operator.

[0099] Figure 5 shows the configuration of the comparison model Mc obtained by removing the fully connected layer L4 from the machine learning model M. As shown in Figure 5, this comparison model Mc has, in order from the input side to the output side, an input layer L1, a first LSTM layer L2, a second LSTM layer L3, and an output layer L5. Figure 6 shows the results after pre-training only the machine learning model M and the comparison model Mc, and then outputting a recommended opening degree using one actual operating data set from the machine learning model M and the comparison model Mc. Figure 6 shows the burner opening degree due to operator operation in the actual operating data and the recommended opening degree output by the machine learning model M and the comparison model Mc of this embodiment.

[0100] As shown in the upper part of Figure 6, in the comparative model Mc, which has only one fully connected layer, the recommended burner opening is divided into multiple stages when the operator increases the burner opening in steps. In this case, the operation becomes complicated when the operator changes the burner opening while referring to the recommended opening. In contrast, as shown in the lower part of Figure 6, in the machine learning model M of this embodiment, which has two fully connected layers, the recommended burner opening is changed in steps in the same way as the operator's operation when the operator increases the burner opening in steps.

[0101] Therefore, in this embodiment, a machine learning model M having two fully connected layers, a fully connected layer L4 and an output layer L5, is employed.

[0102] Next, based on the above results, we modified the layers used for transfer learning as follows, in the case where there are two fully connected layers and in the case where there are three fully connected layers, with one more layer added between the LSTM layer and the output layer. (Common conditions) Pre-training: 284 unconditional datasets, Transfer learning: 44 conditional datasets (Condition 1) Number of fully connected layers: 2 layers, Transfer learning target: 2 fully connected layers (Condition 2) Number of fully connected layers: 2 layers, Transfer learning target: 1 fully connected layer (output layer only) (Condition 3) Number of fully connected layers: 3 layers, Transfer learning target: 3 fully connected layers (Condition 4) Number of fully connected layers: 3 layers, Transfer learning target: 2 fully connected layers (the second fully connected layer and the output layer) (Condition 5) Number of fully connected layers: 3 layers, Transfer learning target: 1 fully connected layer (output layer only)

[0103] Table 2 below shows the evaluation of a machine learning model M with two fully connected layers and a machine learning model with three fully connected layers, using the Mean Absolute Error (MAE) between the output recommended opening and the actual burner opening in the training data, with the layers used for transfer learning changed as shown in conditions 1 to 5 above. [Table 2]

[0104] As shown in Table 2, among conditions 1 to 5, condition 1 has the lowest MAE value and the highest estimation accuracy. In other words, when there are two fully connected layers and the transfer learning target is these two fully connected layers, the MAE value is the lowest and the estimation accuracy is the highest.

[0105] Based on these evaluation results, in this embodiment, for a machine learning model M having two fully connected layers, a fully connected layer L4 and an output layer L5, the fully connected layer L4 and the output layer L5 are targeted for transfer learning.

[0106] <2. Variant> Although embodiments of the present invention have been described above, the present invention is not limited to the embodiments described above.

[0107] Figure 7 is a conceptual diagram showing the configuration of the control unit 80 and the operation unit 700 of the asphalt plant 1 according to one modified example. In the above embodiment, the pre-learning device 70, which is a device for pre-learning, and the control unit 80 of the asphalt plant 1, which is a device for transfer learning, were provided separately. However, both pre-learning and transfer learning may be performed within a single device.

[0108] In this modified configuration, the control unit 80 acts as the pre-training device 70. Therefore, the memory unit 89 stores both a large amount of actual operation data Ds0 and specific condition data Ds1 and Ds2. The learning unit 83 performs both pre-training of the machine learning model M0 using the large amount of actual operation data Ds0 and transfer learning of the basic model M1 using the specific condition data Ds1 and Ds2. Thus, both pre-training and transfer learning may be performed within a single asphalt plant 1.

[0109] Furthermore, in the above embodiment, the control unit 80 of the asphalt plant 1 had a learning unit 83 that performs transfer learning, but the present invention is not limited to this. The learning unit 83 that performs transfer learning may be implemented in a computer separate from the control unit 80. In that case, the control unit 80 transfers a plurality of specific condition data Ds1, Ds2 to the other computer, and the learning unit 83 performs transfer learning of the basic model M1. After that, the other computer transfers the specific condition models M21, M22 to the recommended opening degree output unit 84 of the control unit 80.

[0110] Furthermore, in the above embodiment, the machine learning model M only output the recommended opening degree Do1 for the burner 33 of the new material dryer 30. However, the machine learning model M may also output the recommended opening degree Do1 for the burner 33 of the new material dryer 30 and the recommended opening degree Do2 for the burner 43 of the regenerated dryer 40.

[0111] In that case, the input values ​​for the machine learning model M are at least, • The first target temperature T31 is the target temperature of the new material discharged from the new material dryer 30. • The second target temperature T32 is the target temperature for the recycled material discharged from the regeneration dryer 40. • The temperature sensor S11 located at the outlet 302 of the new material dryer detects the temperature of the new material discharged from the new material dryer 30, which is the first aggregate temperature D11 (aggregate temperature after first heating). The temperature sensor S12 located at the regeneration dryer outlet 402 detects the temperature of the recycled material discharged from the regeneration dryer 40, which is the second aggregate temperature D12 (second aggregate temperature after heating). • The first aggregate supply is the total amount of new material supplied to the new material dryer 30, calculated from the first supply amount T11. • Second aggregate supply amount, which is the total amount of new material supplied to the recycled dryer 40, calculated from the second supply amount T12. Includes.

[0112] Furthermore, the input values ​​for the machine learning model M are: • The rate of change of the first aggregate temperature based on the temperature difference at predetermined intervals of the first aggregate temperature D11 (aggregate temperature after first heating), which is the temperature of the new material discharged from the new material dryer as detected by the temperature sensor S11. • The rate of change of the second aggregate temperature based on the temperature difference at predetermined intervals of the second aggregate temperature D12 (aggregate temperature after second heating), which is the temperature of the recycled material discharged from the regeneration dryer as detected by the temperature sensor S12. It may also include

[0113] Furthermore, in the above embodiment, the asphalt plant 1 manufactured an asphalt mixture using both new and recycled materials as aggregate. However, the asphalt plant 1 may also manufacture an asphalt mixture using only either new or recycled materials as aggregate. For example, if only new materials are used as aggregate, the recycled material is not heat-treated, and therefore the recycled dryer 40 does not operate.

[0114] Furthermore, in the above embodiment, dust collectors 61 and 62 were provided in a common flue 53 connected to both the first exhaust flue 51 and the second exhaust flue 52. However, in the asphalt plant 1, dust collectors may also be provided in each of the first exhaust flue 51 and the second exhaust flue 52.

[0115] Furthermore, although the learning termination conditions in the learning unit 83 were not mentioned in the above embodiment, learning termination conditions can be set as appropriate. For example, the learning termination condition in the learning unit 83 may be set to a value of 5 or less for the MAE. In short, any method or numerical value can be adopted as long as the learned machine learning model M can actually imitate the driving of an expert that an inexperienced person is targeting.

[0116] Furthermore, the elements that appear in the above embodiments and modifications may be combined as appropriate, to the extent that no contradictions arise. [Industrial applicability]

[0117] The present invention can be used in a method for generating a machine learning model for driver assistance to support the operation of an asphalt plant, and in a driver assistance device for assisting the operation of an asphalt plant. [Explanation of symbols]

[0118] 1: Asphalt plant 10: New material hopper 20: Recycled material hopper 30: New material dryer 31: Drums 33: Burner 40: Regenerative hair dryer 41: Drums 43: Burner 70: Pre-training device 72: Pre-learning section 80: Control Unit 83: Learning Department 84: Recommended opening degree output section 100: Driving assistance system 701: Input section 702: Display section Do: Recommended opening Ds0: Actual operating data Ds1: Specific condition data (aggregate drying data) Ds2: Specific condition data (aggregate moisture data) L1: Input layer L2: First LSTM layer L3: Second LSTM layer L4: Fully connected layer L5: Output layer M: Machine learning model M0: Machine learning model before training M1: Basic Model M21: Specific Condition Model (Aggregate Drying Model) M22: Specific Condition Model (Aggregate Wetting Model) T21: First burner opening T22: Second burner opening

Claims

1. A dryer has a burner at one end of the drum and heats and dries the aggregate, An exhaust flue for discharging exhaust gas from the aforementioned dryer to the outside, A dust collection filter inserted into the aforementioned exhaust flue, An input unit that allows the operator to control the burner opening, which is the opening degree of the burner, A method for generating a machine learning model for driver assistance to provide driver assistance for an asphalt plant equipped with, a) A pre-training step in which an untrained machine learning model is pre-trained to obtain a basic model which is the pre-trained machine learning model, b) A transfer learning step in which transfer learning is performed on the basic model to obtain a specific condition model which is the machine learning model that has undergone transfer learning, It has, In steps a) and b) above, supervised learning is performed on the machine learning model so that when multiple input values ​​are input, it outputs a recommended burner opening, using multiple input values ​​of actual operating data as input variables and the burner opening operated by the operator as the output variable. The aforementioned operational data can be classified into two or more groups under specific conditions. In step a), pre-training is performed using the actual operating data including all the groups for specific conditions, A method for generating a machine learning model for driver assistance, wherein step b) involves performing transfer learning using the actual driving data for a predetermined group for the specific conditions.

2. A method for generating a machine learning model for driver assistance according to claim 1, The aforementioned machine learning model, The input layer, The first LSTM layer, The second LSTM layer, Fully connected layer, Output layer, Includes, In step a), the pre-training is performed on all layers of the untrained machine learning model. A method for generating a machine learning model for driver assistance, wherein in step b), the transfer learning is performed on the fully connected layer and all output layers of the basic model.

3. A method for generating a machine learning model for driver assistance according to claim 2, The output layer is a second fully connected layer, and the method for generating a machine learning model for driver assistance is described above.

4. A method for generating a machine learning model for driver assistance according to claim 2, The aforementioned input layer is Masking layer and Convolutional layers and A method for generating a machine learning model for driver assistance, consisting of the following.

5. A method for generating a machine learning model for driver assistance according to claim 1, The input values ​​input to the machine learning model are at least, The target temperature is the target value of the aggregate temperature at the outlet of the dryer, The temperature of the aggregate after heating, detected at the outlet of the dryer, The amount of aggregate supplied to the dryer is the aggregate supply amount, A method for generating machine learning models for driver assistance, including [the specified data].

6. A method for generating a machine learning model for driver assistance according to claim 5, The aforementioned input value is, The filter inlet temperature, which is the temperature of the exhaust gas in the exhaust flue at the inlet of the dust collection filter, The temperature of the aggregate supplied to the dryer is the temperature of the aggregate before heating, The rate of change in aggregate temperature based on the difference in the aggregate temperature after heating at predetermined time intervals, The flue temperature detected within the exhaust flue, The outside temperature of the aforementioned asphalt plant, The humidity outside the aforementioned asphalt plant, A method for generating a machine learning model for driver assistance, further comprising at least one of the following.

7. The aforementioned asphalt plant operation support device, A recommended opening degree output unit outputs a recommended opening degree for the aforementioned burner, A display unit that displays the recommended opening degree to the operator, Equipped with, The aforementioned recommended opening degree output unit is: A driving support device having the specific condition model generated according to any one of claims 1 to 6.

8. The aforementioned asphalt plant operation support device, A recommended opening degree output unit outputs a recommended opening degree for the aforementioned burner, A display unit that displays the recommended opening degree to the operator, A learning unit that performs machine learning on the aforementioned machine learning model. Equipped with, The learning unit receives the basic model generated by step a) of the method for generating a machine learning model for driver assistance according to any one of claims 1 to 6, or the specific condition model generated by steps a) and b) of the method for generating a machine learning model for driver assistance according to any one of claims 1 to 6, The aforementioned learning unit, The above basic model is capable of performing step b), A driving support device that enables the execution of step b) again for the aforementioned specific condition model.

9. A dryer has a burner at one end of the drum and heats and dries the aggregate, An exhaust flue for discharging exhaust gas from the aforementioned dryer to the outside, A dust collection filter inserted into the aforementioned exhaust flue, An input unit that allows the operator to control the burner opening, which is the opening degree of the burner, The driving support device according to claim 7, An asphalt plant equipped with [a specific feature / equipment].

Images