JPH0895605A - High frequency circuit control method/device - Google Patents

Abstract

PURPOSE: To increase the adjusting speed of a high frequency circuit by generating various circuit output patterns by means of circuit simulator and then making a neural network learn these generated patterns. CONSTITUTION: The characteristic of a part that has the large variance of characteristic among parts included in a high frequency circuit 8 is set at the value that is selected at random within a fixed range against the nominal value. Then the value of the part for adjusting is varied so that the output of the circuit 8 has the most desirable characteristic. Under such condition, the value of the part for adjusting is set at r0. Then, the part for adjusting is set at its initial value r1 and the output of the circuit 8 is calculated. At the same time, an output pattern and the r0-r1 value are used as the teacher data that is learnt by a neural network 3. The preceding procedure is repeated and plural teacher data are obtained and then leant by the network 3. Then, the actual output pattern of the circuit 8 is inputted to the network 3, and the output of the network 3 is controlled as an adjusting variable of the part.

Description

Detailed Description of the Invention

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a method and a device for adjusting a high frequency circuit, and more particularly to a method and a device for adjusting a high frequency circuit for adjusting a circuit module designed by using a simulator at the time of design.

[0002]

2. Description of the Related Art Conventional TVs, communication devices and the like use high-frequency filters, high-frequency amplifiers, and other parts whose characteristics need to be adjusted after manufacture. In many cases, the adjustment work was performed manually while observing the display on the measuring instrument. Japanese Patent Laid-Open No. 4-281501 is an example of a conventional technique for automating and efficiently performing this. This device is an automatic frequency characteristic adjusting device for a TV tuner.

FIG. 6 is a block diagram showing a conventional device. The conventional device includes a high frequency signal generator 7, a fuzzy inference unit 10, an oscilloscope 11, a feature recognition unit 12, a coil 13, a tuner 14, and hands 15 and 16. In this device, the width of the coil is adjusted so that the measured waveform satisfies the standard. Measure the actual observation result with an oscilloscope. This is transferred to the feature recognition unit 1
In 2, the characteristic quantities such as bandwidth and tuning point are extracted by numerical processing. Then, the fuzzy inference unit 10 applies a fuzzy rule to the feature amount to calculate an adjustment amount for changing the width of the coil. Based on the adjustment amount, the hands 15 and 16 are moved to change the width of the coil 13 to perform the adjustment. The fuzzy rule and its membership function are determined based on the results of adjustments performed by a skilled person.

[0004]

The above-mentioned conventional apparatus uses a track record of a skilled person to determine the membership function of the fuzzy rule used for the adjustment, so that it takes a lot of man-hours to record the track record. There is a drawback that it costs. In addition, if each worker has a habit, it is difficult to derive a uniform fuzzy rule. In addition, in the case of adjusting a new product, there may be no expert in adjustment, and at this time, there is a disadvantage that fuzzy rules must be created by trial and error.

[0005]

A high-frequency circuit adjusting apparatus of the present invention comprises a simulator for simulating the characteristics of a high-frequency circuit, an output value of the high-frequency circuit calculated by the simulator, and an element constituting the circuit at that time. A neural network that learns element values as teacher data,
A configuration including a signal input unit that supplies a signal to the high-frequency circuit, a measurement unit that measures the output of the high-frequency circuit, and a control unit that inputs the output of the measurement unit to a neural network and instructs the adjustment value of the element. Sararu

The high-frequency circuit adjusting method of the present invention comprises: (A) a step of setting the value of an element on the circuit simulator having a large variation in characteristics to a certain value selected from a predetermined range; (B) After the above (A), in the circuit simulator, the values r0 to rn of the n adjusting elements are changed, and the circuit output is calculated each time, so that the circuit characteristics become the most desirable state. r0-r
Step of calculating the value of n, (C) element value r0 for adjustment
~ Setting rn to a value before adjustment and calculating the output pattern of the circuit with the circuit simulator, (D) the (C)
Using the output pattern calculated in step 1 as the input of the neural network, the optimum adjustment element values r0 to r calculated in (B) above
storing the output pattern and the difference between r0 to rn as a pair so that the difference between the value of n and the value of r0 to rn in (C) is used as the output teacher data of the neural network at that time, (E) The element value in (A) is set to a value different from the previous value, and the steps (A) to (D) are repeated a predetermined number of times, (F) the data stored in (D) is neural-coded. And (G) inputting an output pattern of an actual circuit before adjustment to the neural network learned in (F) and adjusting the adjustment element by the output of the neural network. Composed.

[0007]

The present invention will be described in detail with reference to the drawings.

FIG. 1 is a block diagram showing an embodiment of the present invention. An apparatus for carrying out the adjusting method of the present invention is composed of a high frequency simulator 1, a control unit 2, an input signal generator 7, a measuring instrument 9, and a high frequency circuit 8 to be adjusted. The random element value generation unit 4, the teacher data storage unit 5, the neural network 3, which are included in the control unit 2,
The adjustment instruction unit 6 is realized as software. The high frequency circuit 8 is the object to be adjusted by the device of the present invention.

FIG. 2 is a diagram showing an example of the high frequency circuit 8. This circuit is a microwave band one-stage amplifier circuit, and has resistors, capacitors, inductors, and FETs 29 on the substrate.
Are mounted and are connected by the microstrip lines 23 to 26. As a specification of this amplifier circuit, the reflection coefficient (S11 of S parameter) and the transfer coefficient (S21 of S parameter) as viewed from the input port must satisfy a certain standard in a specific frequency range. , The open stubs 21 and 22 are provided for adjustment. The worker performs the work of satisfying the above-mentioned specifications by adjusting the number of island-shaped stubs connected to the circuit. When adjusting the amplifier circuit, the input signal generator 7 and the measuring instrument 9 are collectively a network analyzer. As the high frequency simulator 1, for example, M of HP
Consists of DS etc.

Next, the general operation of the adjusting method of the present invention will be described. The high-frequency simulator 1 simulates the amplifier circuit. At this time, the element values of the FET 29 and the like are randomly varied from the design value within a certain range, and the output value of the circuit is calculated each time. The correction value of the adjustment element at this time and the output value of the circuit are paired as teacher data, and when a certain amount of data is collected, the neural network 3
Let them learn. Next, the learned output of the high-frequency circuit 8 is input to the learned neural network 3, the corrected value of the adjustment element is output, and the adjustment element of the high-frequency circuit 8 is adjusted based on this value. To do.

Next, the operation of each section will be described in detail.

3A and 3B are flow charts showing a method of creating teacher data.

The high frequency simulator 1 simulates the above circuit example. Note that when the circuit such as the amplifier is usually designed, the simulation is often performed, and in this case, the simulation at the time of design is replaced with the simulation.

Teacher data to be learned by the neural network 3 is generated by simulation, which will be described with reference to the flow chart of FIG.

In process 101, the variation element value generation unit 4
At, the characteristic value of the FET 29 set on the simulator is set to a certain value generated within a certain range from the catalog value (standard value). This setting assumes variations in FET characteristics. In the present embodiment, it is assumed that the variation in the FET characteristics is the dominant factor in the variation of the entire circuit. However, when there is an element in which the variation greatly affects the characteristics of the circuit. Performs similar processing. All elements other than those set here are set to ideal values (design values).

In process 102, a simulation is performed based on each element value set in 101, and the adjustment stub 2
The lengths of 1 and 22 are changed to various values, and the length of the adjusting stub that maximizes the circuit characteristics is obtained. Let these lengths be L21 and L22, respectively.

Next, in process 103, when the stubs 21 and 22 are the shortest, that is, the output of the circuit in the state where no adjustment stub is connected is calculated by simulation.
The lengths of the stubs at this time are M21 and M22, respectively.
An output example of the circuit is as shown in FIG. 4, for example, when S21 of the S parameter is taken as an example. This figure shows the frequency characteristic of the amplitude of S21 in the defined frequency range.

In process 104, data O1 to On at n frequencies, which are predetermined in the output pattern of FIG. 4, are calculated (L21-M21) and (L22).
-M22) is stored in the teacher data storage unit 5 as one teacher data. Although only the amplitude component of S21 is dealt with in this figure, the phase component of S21 may also be included in the teacher data in the same manner. Other parameters such as S11 and S22 may be included depending on the standards to be satisfied.

In process 105, it is determined whether or not the number of stored teacher data has reached a certain number, and if it is not more than a certain number, the process returns to step 101 and another FET and characteristic value different from the previous one are set. . When returning to the process 101, FET2
The characteristic value of 9 is different from the previous one.

Through these processes, a sufficient number of teacher data are generated. The above processing is implemented in the control unit 2 as a series of programs. The control unit 2 is composed of, for example, a personal computer.

Next, the method shown in FIG. 3B will be described.

In process 201, all the element values forming the circuit are set to ideal values (design values). At this time, the lengths of the adjusting stubs 21 and 22 are L21 and L22, respectively.

In step 202, the adjustment stubs 21 and 22 are used.
The length of is set to a predetermined value and the output of the circuit is calculated by simulation. The lengths of the stubs at this time are M21 and M22, respectively. The output example of the circuit is the same as the case of FIG.

In process 203, data O1 to O at n frequencies which are predetermined in the output pattern of FIG.
n and (L21-M21) and (L22-M) calculated previously
The group 22) is stored in the teacher data storage unit 5 as one piece of teacher data. Although only the amplitude component of S21 is dealt with in this figure, the phase component of S21 is similarly processed.
It may be included in the teacher data. Other parameters such as S11 and S22 may be included depending on the standards to be satisfied.

In the process 204, it is judged whether or not the number of stored teacher data has reached a certain number, and if it is not more than a certain number, the process returns to the process 202 and the lengths of the adjustment stubs 21 and 22 are the same as before. Is set to a different value.

Through these processes, a sufficient number of teacher data are generated. The above processing is implemented in the control unit 2 as a series of programs. The control unit 2 is composed of, for example, a personal computer.

Next, an embodiment of the neural network 3 will be described. The neural network is configured as shown in FIG. 5, for example. In the figure, the neural network includes an input layer 91 having input elements of a number corresponding to the number of data of O1 to On, which is a measured value of the circuit among the above-mentioned teaching data, and outputs of each of these input layers. Is provided with an intermediate layer 92 having a predetermined number of intermediate elements, and an output element 93 that receives the output of the intermediate layer 92 and outputs the adjustment amount of the adjustment stub.
The values of the input elements are sequentially assigned from O1 to On in FIG. The intermediate element weights the outputs from the input elements individually, and then calculates the sum of the weights. The intermediate element converts the value according to the sum by a sigmoid function and outputs the value. As an example of the sigmoid function, a known example such as Y = 1 / (1 + exp (−λX)) is used.

Similarly, for the output elements, the outputs of all the intermediate elements are respectively weighted in advance, and then the total sum is obtained, and the value corresponding to the total sum is converted and output by the sigmoid function. The value output from the output element has a real value between 0 and 1. This value corresponds to the adjustment amount of the stub. Note that the actual length of the stub is not necessarily a number between 0 and 1, so the maximum adjustment value of the teacher data is assigned to 1.0 and the minimum adjustment value is assigned to 0.0. , The intermediate value is scaled to a value between 0.0 and 1.0 by proportional calculation and then used for learning of the neural network. When the measured value of the circuit is input to the neural network to calculate the adjustment amount of the stub, the inverse conversion of the above scaling is performed to obtain the actual adjustment amount.

The operation of making the neural network learn the teacher data is specifically an operation of correcting the weights of the connection between the input layer and the intermediate layer of the neural network 3 and between the input layer and the output element to correct values. The correction algorithm uses, for example, the error back propagation method. The error back propagation method is a known method and will not be described in particular.
When the output error of the output element becomes sufficiently small for all the teacher data, the load correction is finished. The learning process is implemented in the control unit 2 as a series of programs.

Next, the actual adjustment of the circuit will be described.
First, the measurement is performed with the adjustment stub shortest. In addition,
Although not shown in FIG. 1, the measurement value is automatically transferred to the control unit by software installed in the control unit 2 in advance. Next, the measured value of the circuit is input to the neural network that has been learned by the above procedure. Input data is shown in Figure 4.
It is a value at several predetermined frequencies shown in.
Since the output value of the neural network 3 indicates the stub adjustment amount, after the inverse conversion of the above scaling, the adjustment instruction unit 6 compares it with a correspondence table stored in advance,
Find the connection point of the stub and display it on the screen. The above series of procedures is automatically performed by the program. Connect the stubs manually by following the instructions on the screen.

Although not shown in FIG. 1, it is also possible to add a bonder for bonding the adjustment stub based on the instruction of the position to be connected to the adjustment stub to completely automate the adjustment work. it can. The bonder is a known device and will not be described in particular.

Further, the above-mentioned embodiment shows an embodiment in which the present invention is applied to the adjustment of the FET amplifier, but in addition, the adjustment of the present invention can be applied to the adjustment of the characteristics that can be simulated by a simulator such as a high frequency filter. Adjustment method can be applied.

[0033]

The high-frequency circuit adjusting apparatus of the present invention uses a neural network for the adjustment, and since the learning data of the neural network is automatically generated from the circuit simulator, the performance of the expert is manually investigated. It is not necessary and has the effect of not requiring man-hours. In addition, since it does not depend on the know-how of the worker, it is not affected by the habit of each worker, and there is an effect that the adjustment system can be constructed even when there is no expert in adjustment.

[Brief description of drawings]

FIG. 1 is a block diagram showing an embodiment of a control unit of the present invention.

FIG. 2 is a circuit diagram showing an example of a circuit of a high frequency amplifier.

3A and 3B are flowcharts for generating teacher data.

FIG. 4 is a graph showing measured values of S parameters.

FIG. 5 is a diagram showing an example of a neural network.

FIG. 6 is a block diagram showing an example of a conventional device.

[Explanation of symbols] 1 high frequency circuit simulator 2 control unit 3 neural network 4 random element value generation unit 5 teacher data storage unit 6 adjustment instruction unit 7 input signal generator 8 high frequency circuit 9 measuring instrument 10 fuzzy inference unit 11 oscilloscope 12 feature extraction Part 13 Coil 14 Tuner 15,16 Hand 21,22 Adjustment stub 23-26 Microstrip line 27 Input terminal 28 Output terminal 29 FET

Claims (4)

[Claims] 1. A simulator for simulating characteristics of a high frequency circuit, a neural network for learning output values of the high frequency circuit calculated by the simulator and element values of elements constituting the circuit at that time as teacher data, A signal input section for supplying a signal to the high-frequency circuit,
A high frequency circuit adjusting device comprising: a measuring unit for measuring an output of the high frequency circuit; and a control unit for inputting an output of the measuring unit to a neural network to instruct an adjustment value of an element. 2. The high-frequency circuit adjustment device according to claim 1, further comprising a manipulator that rotates an adjustment knob of the element based on an adjustment value of the element by the neural network. 3. (A) Among the elements on the circuit simulator,
For those with large variations in characteristics, a step of setting the value to a certain value selected from a predetermined range,
(B) After the above (A), in the circuit simulator, the values r0 to rn of the n adjusting elements are changed,
Calculating the circuit output each time and calculating the values of r0 to rn such that the circuit characteristics are in the most desirable state,
(C) A step of setting the element values r0 to rn for adjustment to values before adjustment and calculating the output pattern of the circuit by the circuit simulator, (D) using the output pattern calculated in (C) as the input of the neural network , The optimum adjustment element values r0 to rn calculated in (B) above and r0 to (C) above
In order to use the difference in the value of rn as the output teaching data of the neural network at that time, the output pattern and r0 to r0
storing a difference of n as a pair, (E) setting the element value in (A) to a value different from the previous one, and (A)
To (D) are repeated a predetermined number of times, (F) the neural network learns the data stored in (D), (G) the neural net learned in (F) And a step of inputting an output pattern of the circuit before adjustment and adjusting the adjustment element by the output of the neural network. 4. (A) Among the elements on the circuit simulator,
Step to set the element value other than the adjustment element to the ideal value,
(B) after the step (A), in the circuit simulator, changing the values r0 to rn of the n adjusting elements to predetermined values, and calculating the circuit output,
(C) The output pattern calculated in (B) is used as an input of the neural network, and the difference between the adjustment element value r0 to rn set in (B) and the design value (ideal value) of the adjustment element value. 4. The method of adjusting a high frequency circuit according to claim 3, further comprising the step of storing the output pattern and the difference of r0 to rn as a pair so as to be the output teaching data of the neural network at that time.