JPH08289613A - Plowing depth controller for tractor - Google Patents

Abstract

PURPOSE: To omit the previous analysis of a correlation of a value sensed with a sensor with the displacement of a rear cover or presetting of a control constant, etc., based on the analytical results in spite of the automatic plowing depth control performed based on a plowing depth sensing factor other than the displacement of the rear cover and thereby reduce the development cost and enable the automatic plowing depth control with a high accuracy adapted to an operating field. CONSTITUTION: This plowing depth controller for a tractor is obtained by using a neural network learning about values sensed with an arm angle sensor 23 and a pitching sensor 24 as an inputting teacher, on the other hand, a value detected with a rear cover sensor 11 as an outputting teacher, inputting the values detected with the arm angle sensor 23 and the pitching sensor 24 to the neural network, obtaining a virtual rear cover angle and automatically controlling a lifting cylinder so as to make the virtual rear cover angle coincide with the target rear cover angle.

Description

Detailed Description of the Invention

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a cultivating depth control device for a tractor that performs work such as plowing and plowing.

[0002]

BACKGROUND OF THE INVENTION Generally,
This kind of tractor has a rear cover, which is vertically moved in accordance with a change in the working depth, at a rear portion of a working unit which is vertically moved up and down in accordance with the operation of a lift cylinder, and detects a displacement of the rear cover by a rear cover sensor. There is one in which the lift cylinder is automatically controlled (automatic control of the working depth) so that the working depth detected by the rear cover sensor matches the set working depth of the working depth setting device. However, in the above, straw, grass,
When there are wheel marks, etc. in the field, rear cover displacement unrelated to the actual working depth occurs, and the lift cylinder is activated in response to the rear cover displacement, which reduces the accuracy of the automatic working depth control. On the other hand, when the rear cover is severely displaced, so-called hunting that repeats overshoot (cylinder operation exceeding the target position) occurs, which makes it difficult to perform stable work. Therefore, it is conceivable to provide a sensor for detecting a tilling depth detection factor or a tilling depth variation factor other than the rear cover displacement, and perform the tilling depth automatic control based on the detection value of the sensor. Since it is necessary to analyze the correlation in advance and set the control constants and the like in advance based on the analysis result, it takes time and effort for development, and the working environment is different for each field,
Depending on the field, there was a possibility that the previous setting would not be suitable.

[0003]

SUMMARY OF THE INVENTION The present invention was conceived in view of the above circumstances and was devised with the object of providing a tractor tilling depth control device capable of solving these problems. A working unit that moves up and down with the operation of the lift cylinder, a rear cover that rotates up and down with fluctuations in the working depth of the working unit, a rear cover sensor that detects the working depth based on the displacement of the rear cover, and a working unit A tractor having a tilling depth setting device for setting the tilling depth and an tilling depth automatic control means for automatically controlling a lift cylinder so that the detected tilling depth matches the set tilling depth, the rear cover being provided on the tractor. A non-rear cover sensor that detects a cultivation depth detection factor or a cultivation depth variation factor other than displacement, and a detection value of the non-rear cover sensor as an input teacher, and a detection value of the rear cover sensor as an output teacher A neural network means for performing learning, is characterized in that the output of the trained neural network is provided and a second tilling depth automatic control means for automatically controlling the lift cylinder to match the setting Kofuka. Further, the tractor is characterized in that it is provided with a control switching means that initially performs the first automatic plowing depth control and automatically switches to the second automatic plowing depth control when learning of the neural network has converged. is there. Further, the tractor is characterized by comprising learning result storage means capable of storing a plurality of learning results of the neural network, and learning result selection means capable of selecting an arbitrary learning result from the learning result storage means. is there. With this configuration, the present invention performs automatic plowing depth control based on a plowing depth detection factor other than the rear cover displacement, while pre-analyzing the correlation between the sensor value and the rear cover displacement, or based on the analysis result. It is possible to reduce the development cost by preliminarily setting the control constants and the like to reduce the development cost, and to execute highly accurate automatic plowing depth control suitable for the working field.

[0004]

Next, an embodiment of the present invention will be described with reference to the drawings. In the drawings, reference numeral 1 denotes a traveling machine body of a tractor, and the traveling machine body 1 includes an engine 2 for generating traveling power, PTO power, hydraulic power, etc., a transmission 3 for shifting and distributing engine power, and various operation tools. Is provided with the operation unit 4 and the like, and the basic configuration of each of them is the same as the conventional one.

Reference numeral 5 denotes a rotary cultivating type working unit which is connected to the rear portion of the traveling machine body 1 via a lifting link mechanism 6 so as to be lifted and lowered. The working unit 5 is a lift associated with expansion and contraction of a lift cylinder 7. The arm 8 is moved up and down based on the vertical rotation of the arm 8, and is tilted to the left and right based on the expansion and contraction operation of the lift rod cylinder 9 provided between the lift arm 8 and the lower link 6a.

Reference numeral 10 denotes a rear cover provided on the rear portion of the working portion 5 so as to be vertically rotatable.
Functions as a flat plate that grounds the plowing soil in an oppressive manner, but the rear end of the rear cover 10 is an analog rear cover that detects the plowing depth (plowing depth) based on the vertical rotation angle of the rear cover 10. A sensor (potentiometer) 11 is connected, and a detection signal of the rear cover sensor 11 is input to a control unit 12 described later via a predetermined interface circuit (A / D conversion circuit).

Reference numeral 13 denotes an operation panel provided on the operation unit 4, which has a position lever 14 for moving the working unit 5 up and down, and an ON-OFF operation for automatic plowing depth control described later. Automatic working depth switch 15,
Tillage depth setting device 1 for setting the target tillage depth of automatic tillage depth control
6. Control method selection switch 17 for selecting the control method of automatic plowing control (rear cover control method, non-rear cover control method or automatic switching), the operation method of the non-rear cover control method (normal operation method or neuro operation method) The calculation method selection switch 18 for selecting, the below-mentioned embankment ratio (0.2, 0.5,
1.0 or automatic setting), the embankment ratio setter 19, a learning area selection switch 20a for selecting in which of a plurality of neuro learning storage areas the learning result is stored, 20b, 20c, learning result clear switches 21a, 21b, 21c for erasing learning results for each region, learning result execution switches 22a, 22b, 22c for executing learning results for each region are provided. The operation signal of is also input to the control unit 12 described later via a predetermined interface circuit.

The control unit 12 is constructed by using a so-called microcomputer (including CPU, ROM, RAM, etc.), which detects the working depth based on the rotation angle of the rear cover 10. The above-mentioned rear cover sensor 11, the arm angle sensor 23 for detecting the rotation angle of the lift arm 8, the pitching sensor 24 for detecting the longitudinal inclination angle of the traveling machine body 1, the engine rotation sensor for detecting the work load based on the engine speed. 25, sensors such as a lever angle sensor 26 that detects a lever angle of the position lever 14, a vehicle speed sensor 27 that detects a vehicle speed, and the above-described operation tools 15 to
Signals are input from 22 and, based on these input signals, the solenoids 28a for expansion and the solenoids 28b for contraction of the lift cylinder electromagnetic switching valve 28, the operating portion 4 are operated.
An operation signal is output to the embankment ratio display 29 or the like provided in the.

That is, the memory (RO
In M), the position control for automatically extending and contracting the lift cylinder 7 so that the detection value of the arm angle sensor 23 matches the detection value of the lever angle sensor 26, the rear cover sensor 1
A control routine such as automatic plowing depth automatic control for automatically extending and contracting the lift cylinder 7 so that the detected plowing depth 1 matches the plowing depth set by the plowing depth setting device 16 will be described below. The control routine of the automatic plowing depth control will be described in detail based on a flowchart.

In the automatic plowing depth control, after confirming the ON state of the plowing depth automatic switch 15, the pitching angle calculation described later is executed, and thereafter the state of the control method selection switch 17 is judged. Here, when the rear cover control method is selected, learning area selection defined as a subroutine, neuro learning, and rear cover control (first automatic plowing depth control) are executed. On the other hand, when the non-rear cover control method is selected, the state of the operation method selection switch 18 is further determined, and when the neuro operation method is selected, the learning area selection and the neuron defined as a subroutine are selected. When the non-rear cover control (second plowing depth automatic control) of the calculation method is executed, when the normal calculation method is selected, the state of the embankment ratio setter 19 is further judged, and the judgment is manual setting. After setting the set value to the embankment ratio variable,
While executing the non-rear cover control (third plowing depth automatic control) of the normal operation method defined as a subroutine, when the automatic setting is performed, the embankment ratio automatic setting defined as a subroutine and the non-operation of the normal operation method are performed. Rear cover control (third plowing depth automatic control) is executed. Furthermore, when automatic switching is selected by the control method selection switch 17, the learning area selection, neuro learning, and rear cover control (first plowing depth automatic control) are initially executed while learning of neuro learning is performed. When the values converge, the non-rear cover control (second plowing depth automatic control) of the neuro calculation method is automatically switched. The above-mentioned respective subroutines will be described below.

In the above-mentioned rear cover control, after the signals of the operating tools and the sensors are read, the detected rear cover angle (the detected value of the rear cover sensor 11) with respect to the target rear cover angle (calculated based on the set value of the working depth setting device 16). Of the lift cylinder 7 is calculated by differentiating the deviation to obtain the amount of change, and then using the first fuzzy inference using the deviation and the amount of change as input variables. It is designed to calculate. That is,
In the first fuzzy inference, the antecedent membership function with the deviation as the input variable, the antecedent membership function with the variation as the input variable, the antecedent membership function with the control flow as the output variable, and each variable And a first fuzzy rule indicating the relationship of (1) is set in advance. In this first fuzzy rule, for example, when the deviation is large on the shallow side (PB) and the variation is large on the shallow side (PB), When the lift arm 8 is lowered at a high speed (NB) and the deviation is medium (PM) on the shallow side and the change amount is medium (PM) on the shallow side, the lift arm 8 is lowered at the medium speed. (NM), the deviation is zero (ZO), and the amount of change is zero (ZO).
If, the correlation such as stopping the lift arm 8 is set. Then, in the first fuzzy inference, for each rule set in the first fuzzy rule (25 in this embodiment), the goodness of fit of the deviation and the change amount is calculated based on each antecedent membership function, whichever is lower. An area is obtained on the consequent membership function based on the value of the one, and the areas obtained for each rule are combined, and the center of gravity value in the horizontal axis direction is adopted as the control flow rate.
In the membership function and fuzzy rule, PB is positive and large, PM is positive and medium, PS is positive and small, ZO is zero, NB is negative and large, NM is negative and medium, NS is negative. Means small.

Further, in this embodiment, the antecedent membership function having the output value (control flow rate) of the first fuzzy inference as an input variable, the antecedent membership function having the engine speed as an input variable, and the final The consequent membership function that uses the dynamic control flow rate as an output variable and the second fuzzy rule indicating the relationship between the variables are set in advance, and the second fuzzy inference is performed. In other words, in the second fuzzy reasoning, when the engine speed is in the standard or rising state, the output flow rate of the first fuzzy reasoning is output as it is, while when the engine speed is in the falling state, the first fuzzy reasoning is performed. Of the lift arm 8 at a high speed (PB) regardless of the output flow rate, the final control flow rate is determined by the same calculation procedure as the first fuzzy reasoning described above. . For this reason, when an overload is applied to the working unit 5 due to a drop into the recess, the working unit 5 quickly rises to avoid the overload, and as a result, the engine is stopped due to the overload. It is possible to eliminate such inconveniences.

The neural learning executed simultaneously with the rear cover control is performed by forming a hierarchical neural network in which at least two input units, two intermediate units and one output unit are connected in a hierarchical manner. There is.
Then, each unit takes in the signal (detection value or calculation result) Oi from the sensor or the preceding unit into the unit through the coupling load Wi, and the sum x (x = ΣOi × Wi) of the taken-in values (Oi × Wi). )
Is calculated, and a predetermined threshold value H is subtracted from the total sum x, and this is transformed using a function such as a sigmoid function {1 / (1 + e ^−x )} to obtain an output y of the unit. In other words, the hierarchical neural network operates on a plurality of given input values in the order of the input unit, the intermediate unit and the output unit, and outputs a single operation result from the output unit. In the hierarchical neural network of the embodiment, the detected value of the arm angle sensor 23 and the detected value (calculated value) of the pitching sensor 24 are used as input teachers, while the detected value of the rear cover sensor 11 is used as output teacher, so-called reverse error propagation method. Each coupling weight Wi is learned based on (back propagation method, PB method). Therefore, after the learning of each coupling load Wi is completed (converged), the virtual rear cover angle (estimated tillage) can be obtained simply by inputting the detected value of the arm angle sensor 23 and the detected value (calculated value) of the pitching sensor 24 to the neural network. You can get the depth value). Incidentally, the back error propagation method is a supervised learning method that corrects the coupling weight Wi of each unit by back propagation so that the error between the output value of the output unit and the output teacher value is minimized. The connection load correction amount ΔWi of is calculated by the following formula. ΔWi = −ε · (∂E / ∂x) · Oi where ε is a learning coefficient and (∂E / ∂x) is an error of each unit.

Further, in the learning selection routine, a plurality of neuro learning storage areas are secured, and during the rear cover control, the learning area selection switches 20a, 2
Areas selected in 0b and 20c are saved as learning results while non-rear cover control is in progress (neuro calculation method)
, The learning result execution switches 22a, 22b, 2
The learning result of the area selected in 2c is set as the execution target, and the learning result clear switch 21 is set regardless of the control state.
When a, 21b, and 21c are operated, the learning result of the corresponding area is erased.

On the other hand, in the non-rear cover control of the neuro calculation method, after the signals of the operation tools and the sensors are read, the virtual rear cover angle is neuro calculated by using the learning result. That is, in the neuro operation, the step of loading the detection value of the arm angle sensor 23 into the first input unit via the coupling strength Wi and the threshold value Hi from the loaded value.
, A step of converting the value into an output value by a sigmoid function, a step of taking a detected value (calculated value) of the pitching sensor 24 into the second input unit via the coupling strength Wi, and a step of taking the taken value. The step of subtracting the threshold value Hi and the step of converting the value into the output value by the zigmoid function are executed, and the calculation of the input layer is completed.
Next, a step of sequentially taking in the output value On of each input unit via the coupling strength Wn, a step of obtaining a total sum of the taken values, a step of subtracting a threshold value Hi from the total sum, and outputting the value by a zigmoid function. The calculation of the intermediate unit including the step of converting into a value is executed for each intermediate unit to complete the calculation of the intermediate layer. Subsequently, a step of sequentially fetching the output value Om of each intermediate unit via the coupling strength Wm, a step of obtaining a total sum of the fetched values, a step of subtracting a threshold value Hi from the total sum, and outputting the value by a zigmoid function. And an operation of the output unit composed of a step of converting into a value. Then, after the virtual rear cover angle is obtained by the neuro calculation, the deviation of the virtual rear cover angle from the target rear cover angle is calculated, and the deviation is differentiated to obtain the change amount. Thereafter, the deviation and the change amount are used as input variables. The lift cylinder 7 is executed by executing the first fuzzy inference (the same as in the rear cover control) and the second fuzzy inference (the same as in the rear cover control) using the output of the first fuzzy inference and the engine speed as input variables.
The control flow rate of is calculated. That is, in the non-rear cover control, the virtual rear cover angle is calculated based on the lift arm angle and the machine body pitching angle, which are the cultivation depth fluctuation factors (rear cover fluctuation factors), without using the detection value of the rear cover sensor 11. Since the tilling depth automatic control is executed based on the above, even if there is straw, mowed grass, wheel marks, etc. in the provisional designated field, these effects are hardly affected, and as a result, good working accuracy is ensured. The occurrence of so-called hunting can be prevented.

In the normal operation type non-rear cover control, after the signals of the operation tools and sensors are read,
The virtual rear cover angle C is calculated using a preset function. That is, the normal calculation is the plowing depth set value (plowing depth setter 1
Setting value of 6) as a variable, the first function calculation step of calculating the target lift arm angle LM, and the lift arm target angle L
The second function calculation step for calculating the virtual rear cover target angle CM with M as a variable, the virtual rear cover target angle CM, the lift arm target angle LM, the lift arm detection angle L, and the aircraft pitching angle P (deviation of the pitching sensor value from the reference value ) As a variable, the third function calculation step {C = CM + (LM-L) × a + P ×
b, where a and b are constants}, the constant (slope of the linear function) used for the second function is the ratio m of the embankment to the plowing depth (m = embankment height / dip depth) It is designed to change according to. That is, since the amount of embankment, which is one of the determinants of the rear cover angle, changes depending on the soil quality and PTO rotation, the embankment ratio m is set manually or automatically, and the set embankment ratio m is calculated as the virtual rear cover angle. It is designed to be added to an element. Then, after the virtual rear cover angle is obtained by the above-described normal calculation, the deviation of the virtual rear cover angle from the target rear cover angle is calculated, and the deviation is differentiated to obtain the change amount. Thereafter, the deviation and the change amount are input variables. And the first fuzzy inference (the same as that of the rear cover control) and the second fuzzy inference (the same as that of the rear cover control) that uses the output of the first fuzzy inference and the engine speed as input variables to perform the lift cylinder. 7
Is calculated, and when the lift cylinder 7 is controlled based on this control flow rate, substantially the same operational effects as the above-described neuro calculation non-rear cover control are obtained.

Further, in the non-rear cover control of the normal calculation method, the embankment ratio m set manually or automatically.
Is displayed on the embankment ratio display 29. That is, since the operator can always recognize the embankment ratio m, it is possible to perform an operation that makes use of the information, and determine the soil quality based on the displayed embankment ratio m (when the embankment ratio is high, the clay quality, If it is low, it is possible to judge the sand quality etc.).

Further, in the automatic setting of the embankment ratio, a hierarchical neural network which is substantially the same as the neuro calculation is used, and the rear cover sensor 11 is used in the neural network.
And the detection value of the arm angle sensor 23 is input and the embankment ratio m is neuro-calculated. That is, the detection value characteristics of the rear cover sensor 11 and the arm angle sensor 23 are set in advance for each embankment ratio m (0.0, 0.25, 0.5,
0.75, 1.0) and the detected value characteristics of the measured rear cover sensor 11 and arm angle sensor 23 are used as an input teacher, while each embankment ratio m is used as an output teacher to learn the coupling load Wi in advance. Since this is mounted on an actual machine, the embankment ratio m can be automatically set only by inputting the detection values of the rear cover sensor 11 and the arm angle sensor 23 to the neural network.

Further, in the pitching angle calculation executed prior to each of the above-mentioned controls, among the detected values of the pitching sensor 24 sequentially input, a step of averaging the detected values input within the past predetermined time T And a step of adopting the averaged value as a pitching reference value and adopting the deviation of the current value from the pitching reference value as the current pitching angle. That is, when performing automatic plowing depth using the machine body pitching angle displacement as in the non-rear cover control, the average of the pitching detection values input within the past predetermined time T (substantially equal to the entire tilt angle of the field) is used as a reference. Since the pitching angle is calculated in this manner, the pitching angle relative to the entire tilt angle of the field can be obtained even when the entire provisional field is tilted. Therefore, the horizontal value is always used as the reference. As in the case of the above, there is no unnecessary raising and lowering of the working part in response to the entire inclination of the field, and as a result,
It is possible to eliminate the inconvenience that the plowing depth is generally deep on the downhill and the plowing depth is generally shallow on the uphill.

Further, in the pitching calculation, the averaging time T is obtained by dividing the front-rear dimension L from the front wheel center position to the working portion position (rear end of the rear cover or the center of the rotary) by the vehicle speed V detected by the vehicle speed sensor 27. The step of sequentially calculating is included. The averaging time T calculated as described above is the running time from when the front wheels 1a of the traveling machine body 1 reach the inclination change point to when the working unit 5 (the rear end of the rear cover or the center of the rotary) reaches the same point. Therefore, the pitching reference value (average value) does not match the current value at least until the working unit 5 reaches the inclination change point. In other words, since it is necessary to operate the plowing depth correction by the automatic plowing depth control in the traveling section, a deviation of the current value from the pitching reference value is generated, but after the working unit 5 passes the slope change point, the pitching reference value is set. Rapidly matches the field inclination, so that the automatic plowing depth control is executed based on the relative pitching angle with reference to the overall inclination as described above.

In the embodiment of the present invention configured as described above, when the rear cover control system is selected by the control system selection switch 17, the automatic working depth control based on the rear cover displacement is executed. Further, when the non-rear cover control method is selected by the control method selection switch 17 and the neuro operation method is selected by the operation method selection switch 18, the detected value of the arm angle sensor 23 is added to the neural network learned during the rear cover control. And the detection value (calculated value) of the pitching sensor 24 is input to acquire the virtual rear cover angle, and the automatic plowing depth control is executed based on the virtual rear cover angle. Further, when the non-rear cover control method is selected by the control method selection switch 17 and the normal operation method is selected by the operation method selection switch 18, the detection value of the arm angle sensor 23 and the detection value of the pitching sensor 24 (operation The virtual rear cover angle is calculated based on the value) and the embankment ratio, and the plowing depth automatic control is executed based on the virtual rear cover angle. Further, when the automatic switching is selected by the control method selection switch 17, the neural network is learned while initially performing the rear cover control method of the automatic plowing depth control, and automatically when the learning value converges. Then, the non-rear cover control of the neuro calculation method is executed.

As described above, according to the present invention, the plowing depth automatic control is performed so that the plowing depth of the working portion 5 matches the set plowing depth. In the non-rear cover control, the rotation angle of the lift arm 8 and the traveling machine body are controlled. The virtual rear cover angle is calculated based on the pitching angle of 1, and the lift cylinder 7 is automatically controlled so that the virtual rear cover angle matches the target rear cover angle. In other words, since the automatic working depth control is performed regardless of the actual rear cover displacement, it is possible to perform the automatic working depth control that is hardly affected by straw, mowing grass, wheel marks, etc. Therefore, it is possible to secure good working accuracy, and prevent hunting from occurring to improve work stability.

Further, in the non-rear cover control of the normal calculation method, the ratio of tilling soil to plowing depth is set, and
Since the virtual rear cover angle is calculated in consideration of the embankment ratio, the plowing depth is automatically controlled based on the virtual rear cover angle, but the plowing depth based on the finished surface can be set similarly to the rear cover control. In addition, since the embankment ratio can be arbitrarily set based on the soil quality, PTO rotation speed, etc., which are ratio variation factors, it is possible to cope with various work environments.

When the automatic setting is selected by the embankment ratio setting device 19, the operator detects the embankment ratio automatically by inputting the detection values of the rear cover sensor 11 and the arm angle sensor 23 to the neural network learned in advance. In addition to being able to reduce the operation labor of, it is possible to set the embankment ratio that sequentially corresponds to changes in the work environment, and as a result, it is possible to contribute to improvement in operability and work accuracy.

Since the embankment ratio is displayed on the embankment ratio display 29 regardless of the manual setting or the automatic setting, the operator can always recognize the embankment ratio and perform the operation utilizing the information. It can be used to judge the soil quality.

In the non-rear cover control of the neuro calculation method, the detection value of the arm angle sensor 23 and the detection value (calculation value) of the pitching sensor 24 are used as input teachers, while the detection value of the rear cover sensor 11 is learned as an output teacher. The detected value of the arm angle sensor 23 and the detected value (calculated value) of the pitching sensor 24 are input to this neural network to obtain the virtual rear cover angle. You can analyze the correlation with the rear cover displacement in advance,
It is not necessary to preset control constants based on the analysis results, and as a result, it is possible to save development effort and contribute to cost reduction. It can be performed.

When the automatic switching is selected by the control method selection switch 17, the rear cover control is executed at the beginning while the learning of the neural network is performed, and the non-rear cover is automatically set at the stage when the learning of the neural network converges. Since the control is executed, it is possible to eliminate the need for the switching operation and reduce the operation labor, or it is possible to eliminate the inconvenience of lowering the work accuracy due to an erroneous switching operation.

Further, since the learning of the neural network is always executed during the rear cover control, no matter how the timing is changed to the non-rear cover control,
The non-rear cover control in which the learning result is reflected can be executed, and moreover, since the operator is not made aware of learning, it is possible to contribute to simplification of the operation.

Further, since a plurality of learning results of the neural network are stored and any learning result can be selected from the plurality of stored learning results, the operator himself intentionally controls the contents of the control according to the work environment and the like. It is not permissible to change, and there is an advantage that the learning process at the beginning of work can be omitted by using the past learning results.

Further, since the rear cover sensor 11 is not required during the non-rear cover control, the rear cover sensor 1
There is an advantage that the plowing depth automatic control (ground height control) can be executed even by various working machines that do not have the No.1.

Further, the respective control methods and calculation methods are as follows:
Since it can be arbitrarily switched, there is an advantage that the working accuracy can be improved by performing the switching according to the work.

Further, in the rear cover control and the non-rear cover control, the control flow rate is calculated based on the first fuzzy inference using the rear cover deviation (or virtual rear cover deviation) and the amount of change with respect to the target as input variables. Output value of one fuzzy reasoning (control flow rate)
Since the final control flow rate is calculated based on the second fuzzy inference using the engine speed (work load) as an input variable, when the engine speed decreases, the lift is lifted regardless of the output flow rate of the first fuzzy inference. It is possible to perform flow rate correction such as raising the arm 8 at high speed. Therefore, when an overload is applied to the working unit 5 due to the depression into the recess,
It is possible to quickly raise the working unit 5 to avoid an overload, and as a result, it is possible to eliminate inconveniences such as engine stop due to overload.

Further, in utilizing the pitching angle displacement of the machine body in the automatic plowing depth control, among the detection values of the pitching sensor 24 sequentially input, the detection values input within the past predetermined time T are averaged. , The averaged value is used as the pitching reference value, and the deviation of the current value from the pitching reference value is adopted as the current pitching angle. Therefore, even if the entire provisional field is inclined, the entire inclination of the field is inclined. It is possible to obtain the pitching angle relative to the angle, so that there is no unnecessary lifting of the working part in response to the overall tilt of the field, as when the horizontal value is always used as a reference, and as a result, It is possible to eliminate the inconvenience that the plowing depth is deep on the slope and the plowing depth is generally shallow on the uphill.

Further, the averaging time T is successively calculated by dividing the front-rear dimension L from the center position of the front wheel to the working portion position by the vehicle speed V detected by the vehicle speed sensor 27, so that the front wheel 1a of the traveling machine body 1 is calculated. Corresponds to the traveling time from when the work portion 5 reaches the inclination change point to when the working unit 5 reaches the same point. That is, since the pitching angle deviation is generated until the working unit 5 reaches the inclination change point, proper plowing depth automatic control can be performed in the inclination change section, and after the working unit 5 passes the inclination change point. Since the pitching reference value quickly matches the field inclination, the automatic working depth control can be executed based on the relative pitching angle based on the overall inclination as described above.

It should be noted that the present invention is not limited to the above-mentioned embodiment. For example, in the above-mentioned neuro arithmetic non-rear cover control, the lift arm angle and the machine body pitching angle are input variables of the neural network. However, it is also possible to adopt and combine other tillage depth detection factors (or tillage depth fluctuation factors), such as using the aircraft pitching angle and engine speed (work load) as input variables. However, it is desirable to include at least one of the aircraft pitching angle and the engine speed (work load) among all factors. Further, in the above-described embodiment, the learning of the neural network is always performed during the rear cover control, but it is also possible to select the presence or absence of the learning with a changeover switch or the like. Then, in this case, the operator can perform learning at any place, and thus control that reflects the intention of the operator can be performed. Further, in the above-mentioned second fuzzy inference that corrects the control flow rate based on the engine speed (work load), the control flow rate is corrected so that the working unit 5 rises only when the engine speed rapidly decreases, but otherwise. Even in the case of, it is possible to correct the control flow rate by considering the engine speed (work load), and in this case, always consider the engine speed (work load) which is the cultivation depth detection factor. Since the plowing depth automatic control is performed, the control accuracy can be improved.
Furthermore, the correction based on the engine speed does not necessarily need to use fuzzy inference. For example, when the engine speed is lower than a preset speed, the plowing depth target value (target rear cover angle) is reduced by a predetermined amount. It may be corrected to the shallow plowing side, or if the correction amount of the plowing depth target value is made proportional to the decrease amount of the engine speed (deviation from the reference value), it may be corrected depending on the situation. it can.

[0036]

In summary, since the present invention is configured as described above, the automatic working depth control is performed based on the detection value of the non-rear cover sensor that detects the working depth detection factor or the working depth variation factor other than the rear cover displacement. In doing so, the detected value of the non-rear cover sensor is used as an input teacher, while the detected value of the rear cover sensor is used as an output teacher to learn the neural network, and the automatic working depth control is executed based on the output of the learned neural network. It is not necessary to analyze the correlation between the detection value of the non-rear cover sensor and the working depth in advance, or to preset the control constants based on the analysis result, and as a result, it is possible to save development labor and contribute to cost reduction. Perhaps, it is possible to perform highly accurate automatic plowing depth control based on a neural network adapted to the working field.

Further, when the first automatic plowing depth control based on the rear cover displacement is initially executed and the automatic plunging control is automatically switched to the second plowing depth automatic control when the learning of the neural network has converged, the switching operation is performed. It is possible to eliminate the need for the above and reduce the operation labor, or it is possible to eliminate the inconvenience of lowering the work accuracy due to an erroneous switching operation.

When a plurality of learning results of the neural network are stored and an arbitrary learning result can be selected from the plurality of stored learning results, the operator himself intentionally operates according to the work environment and the like. There is an advantage that the learning process at the beginning of the work can be omitted by using the past learning result, not allowing the control contents to be changed.

[Brief description of drawings]

FIG. 1 is a side view of a tractor.

FIG. 2 is a block diagram showing input / output of a control unit.

FIG. 3 is a flowchart of automatic plowing depth control.

FIG. 4 is a block diagram showing the concept of rear cover control.

FIG. 5 is a flowchart of rear cover control.

6A, 6B and 6C are membership functions used in the first fuzzy inference, and FIG. 6D is a fuzzy rule.

7 (A), (B) and (C) are membership functions used in the second fuzzy inference, and (D) is a fuzzy rule.

FIG. 8 is a side view of the tractor showing the operation during overload.

FIG. 9 is a timing chart showing an example in which the target value of plowing depth is corrected by a fixed amount based on the engine speed.

FIG. 10 is a timing chart showing an example of proportionally correcting the target value of plowing depth based on the engine speed.

FIG. 11 is a block diagram showing a control concept of non-rear cover control (neuro calculation method).

FIG. 12 is a flowchart of non-rear cover control (neuro calculation method).

FIG. 13 is a flowchart of a neuro calculation.

FIG. 14 is a model diagram showing a neural network.

FIG. 15 is a model diagram showing a unit.

FIG. 16 is a model diagram for explaining back error propagation.

FIG. 17 is a flowchart of learning selection.

FIG. 18 is a plan view of a learning selection panel.

FIG. 19 is a block diagram showing the concept of learning selection.

FIG. 20 is a block diagram showing a control concept of non-rear cover control (normal calculation method).

FIG. 21 is a flowchart of non-rear cover control (normal calculation method).

FIG. 22 is a flowchart of normal calculation.

23 (A), (B), (C), and (D) are side views showing the relationship between the working depth and the embankment.

FIG. 24 is a graph of a function used in normal calculation.

FIG. 25 is a detection value characteristic graph used in automatic setting of the embankment ratio.

FIG. 26 is a block diagram showing the concept of pitching angle calculation.

FIG. 27 is a flowchart of pitching angle calculation.

FIG. 28 is a timing chart showing a relationship between a pitching angle and a reference value.

FIG. 29 is a graph showing a pitching angle displacement in the case of cultivating a flat ground having unevenness.

FIG. 30 is a graph showing a pitching angle displacement when cultivating an inclined land.

FIG. 31 is a graph showing the pitching angle displacement when the slope change point is cultivated.

[Explanation of symbols]

 1 Traveling Vehicle 5 Working Section 7 Lift Cylinder 10 Rear Cover 11 Rear Cover Sensor 12 Control Section 23 Arm Angle Sensor 24 Pitching Sensor 25 Engine Rotation Sensor

Claims (3)

[Claims] 1. A working unit that moves up and down with the operation of a lift cylinder, a rear cover that rotates up and down with fluctuations in the working depth of the working unit, and a rear cover sensor that detects the working depth based on the displacement of the rear cover. A tractor having a working depth setting device for setting the working depth of the working unit, and a working depth automatic control means for automatically controlling the lift cylinder so that the detected working depth matches the set working depth, A non-rear cover sensor that detects a tilling depth detection factor or a tilling depth variation factor other than the rear cover displacement, and a neural network for learning by using the detected value of the non-rear cover sensor as an input teacher and the detected value of the rear cover sensor as an output teacher. The network means and the second automatic working depth control means for automatically controlling the lift cylinder so that the output of the learned neural network matches the set working depth. A tractor tilling depth control device characterized by being struck. 2. The tractor according to claim 1 is provided with a control switching means that initially executes the first automatic plowing depth control and automatically switches to the second automatic plowing depth control when learning of the neural network has converged. A tractor tilling depth control device. 3. The tractor according to claim 1, comprising learning result storage means capable of storing a plurality of learning results of the neural network, and learning result selection means capable of selecting an arbitrary learning result from the learning result storage means. A tractor tilling depth control device.