JPH05257918A - Parts allocation determination device - Google Patents

Abstract

PURPOSE:To reduce the changing frequency of programs and to determine the allocation of parts to an assembling machine by collecting the sorts of printed boards to be assembled by the same program of the assembling machine as much as possible within a certain production period in the assembling of printed boards. CONSTITUTION:This parts allocation determining device is provided with a two-layer competition type neural circuit network having the same number of input nodes 1 to 5 and output nodes 6 to 10 and capable of fixing outputs nodes corresponding to ON input nodes to ON and suppressing the number of ON output nodes to a fixed value, the parts constitution of each printed board is inputted from the input nodes 1 to 5 and the output nodes 6 to 10 are allowed to correspond to the parts allocation of the assembling machine, so that the group of semiappropriate printed boards is outputted and parts allocation can be determined.

Description

Detailed Description of the Invention

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a component allocation determining device, and more particularly to a component allocation determining device for a printed circuit board assembly machine.

[0002]

2. Description of the Related Art In recent years, in the printed circuit board assembly industry, particularly in the communication equipment and exchanges called system products, and the medium-sized or larger information processing equipment industry, the production of a large number of products in small lots is progressing, and the production of small lot sizes and lots It became normal. Therefore, the types of parts used are enormous, but the number of parts that can be set in the printed circuit board assembling machine is defined by the number of parts feeders of the machine. Generally, the number of feeders that can be set at the same time is about tens to hundreds and tens. Therefore, the feeder must be replaced (setup change), but the frequent setup change of the assembly machine delays the production of printed circuit boards and increases inventory. Therefore, there is a need for a method of grouping similar printed circuit boards together and performing setup change for each group to reduce the frequency of setup change.

Conventionally, in the printed circuit board assembly industry,
Often, some empirical rules (heuristics) were used in determining the input groups to reduce the setup change of the assembly machine within a certain production period. This kind of heuristics is mainly adopted by people in the field leader class based on experience, and it is common to use different rules for each field. As an example, sort the component types in descending order of the number of printed circuit boards that use that component type, and then sort the printed circuit boards in the order of using the most commonly used component types. Is used.

[0004]

In the above-mentioned conventional heuristics, the input group can be manually determined while the number of types of printed circuit boards is small. But,
As the number of types increases, a program is inevitably required, and the creation of the program requires specialized knowledge of printed circuit boards and assembly machines. Human resources having such program creation ability and printed circuit board assembly expertise are rarely obtained easily, and in the end, a great deal of man-hours and time are required for program development.

On the other hand, in recent years, research on neural networks has progressed, and due to artificial changes in the state of the network, a kind of information processing device that can obtain a predetermined result without programming has been developed and put to practical use. It was

[0006]

According to a first aspect of the present invention, in a two-layer type neural network composed of an input layer and an output layer each having the same number of nodes, an output node corresponding to an input node in an ON state is in an ON state. An output node that behaves so as to be fixed to, and a penalty node that holds an arbitrary limit number and is connected to all output nodes so that the number of nodes in the ON state in the output layer can be kept below a certain limit number. The provided competitive neural network and the component allocation determining device according to the second aspect of the invention correspond to the component configurations of a plurality of types of printed circuit boards from the input node of the competitive neural network, and limit the number of assembly machines. The output node is made to correspond to the maximum number of attachable parts and the parts are assigned to the machine for assembling the printed circuit board.

[0007]

Embodiments of the present invention will now be described with reference to the drawings.

FIG. 1 is a block diagram showing an embodiment of the competitive neural network of the first invention.

Various methods can be considered as a neural network for realizing the present invention, but here, it is assumed that the method is realized by a Boltzmann machine. The connection between the nodes whose both ends are turned on is called an active connection. In short, the Boltzmann machine is a neural network that makes transitions so that the total sum of weights of active connections is maximized. .. The sum of weights of active connections is called consensus. That is, the Boltzmann machine is a neural network having behavior that maximizes consensus. The Boltzmann machine itself can be configured with existing technology, so it will not be described again here.

In FIG. 1, the neural network has input nodes 1, 2, 3, 4, 5 and output nodes 6, 7, 8, 9, 1.
0, a penalty node 30, and a line segment connecting the input and output nodes includes a connection 20 between the input and output nodes. The behavior of this neural network will be described as an example.
Now, suppose that the input pattern is [0,1,0,1,1]. Then, the input nodes 1, 2, 3, 4, 5 are fixed according to the input pattern, and also [0, 1, 0, 1,
1]. In FIG. 1, black circles represent nodes having a value “1”, that is, an on state, and white circles represent nodes having a value “0”, that is, an off state. This is why the input nodes 2, 4 and 5 are on.

On the other hand, the output nodes 6, 7, 8, 9,
10 exhibits the following behavior due to its own property.
That is, the output nodes 7, 9 and 10 corresponding to the input nodes 2, 4 and 5 having the value "1" are also fixed to the value "1", but the output nodes corresponding to the input nodes 1 and 3 having the value "0" are output. The values of the nodes 6 and 8 are still “0” at this point. But,
It may be turned on by a subsequent transition.

A penalty node 30 is connected to the output nodes 6, 7, 8, 9, and 10. The penalty node 30 can have an arbitrary limit number, and when the limit number is exceeded, the penalty node 30 has a property of turning on. Normal penalty node 30 and output nodes 6, 7,
A large negative weight is given to the connection between 8, 9 and 10. If the number of output nodes that are on exceeds the limit number and the penalty node 30 is turned on, the connection between the penalty node 30 and the output nodes that are turned on becomes active, and a large penalty is added to the consensus. Therefore, the number of output nodes that are turned on does not exceed the limit. In this example, if the limit number is 4, the number of nodes turned on at the output node does not exceed 4.

However, since the connection connecting the input node and the output node has a positive weight, the number of output nodes that are turned on as a result of the transition is limited (4 in this example).
Is equal to In this example, since the output nodes 7, 9 and 10 are turned on corresponding to the input nodes, any of the remaining 6 or 8 is turned on. In this example, the output nodes 6 and 8 are "competing" when they are turned on, and are therefore called competitive neural networks. Competitive neural networks tend to give common outputs to similar inputs in principle, and have a grouping ability. The present invention takes advantage of this grouping capability.

FIG. 2 is a flowchart showing the algorithm of the learning phase of this neural network. Step 10
In step 1, set one input pattern in the input layer, then step 1
In 02, the output node corresponding to the ON input node is fixed to ON. In step 103, as many output nodes as the limit number allows are turned on from the output nodes that are not fixed. In step 104, the weight of the connection that connects to the output node that has been turned on, that is, the winner, is updated so that the winner is more likely to win. In step 105, if there is another input pattern, another input pattern is selected and the same processing is repeated. In step 106,
The above processing is repeated until learning is firmly established.

FIG. 3 shows an algorithm of the association phase of this neural network. For the neural network whose weight is updated in the learning phase described above, each input pattern is set in the input layer in step 201 as in the learning phase, and the output node corresponding to the input node in the ON state is set in step 202. Lock on. In step 203, a winner (a node to be turned on) is determined based on the weight of the connection obtained in the learning phase. Of course, the weight is not updated in the associative phase.
If the input pattern still exists at 04, the same processing is repeated to obtain the output pattern corresponding to the input pattern.

An example of the output of this neural network will be described with reference to the input / output pattern correspondence diagram of FIG. FIG. 4 shows the output pattern obtained in the associative phase after the four input patterns shown in the left column are given and the weight of the connection is updated in the learning phase on the right side. The limit number is assumed to be 4. As is clear from this, the input patterns 1 and 4 output the same output pattern, and the input pattern 2
And 3 output another same output pattern. That is,
It is considered that the input patterns 1 and 4 are classified into the same group, and the input patterns 2 and 3 are classified into another same group.

FIG. 5 is an operation example diagram showing an embodiment of the parts allocation determination device of the second invention. The left column shows the component configuration of each printed circuit board.
It means that a few will be implemented. The printed circuit board assembling machine considered here has four component feeders. Then, the component configuration of each printed circuit board is considered as an input pattern and input to the neural network of the first invention. Then, since this pattern is similar to that given in FIG. 4, the same result as the output pattern of FIG. 4 is output in the association phase after the learning phase. However, since this output pattern corresponds to the component allocation of the assembly machine, it is better to assemble the printed circuit boards 1 and 4 with the same component allocation, that is, with the components 1, 2, 3, 4 set. become. Similarly, it is concluded that the printed boards 2 and 3 should be assembled with the components 1, 2, 4, and 5 set. Therefore, the right column of FIG. 5 is obtained. Now, consider a group of printed circuit boards assembled by the same component allocation, and consider that the smaller the number of groups resulting from component allocation, the better the component allocation. In this sense, component allocation by the competitive neural network is not necessarily optimal, but as a property of the competitive neural network, similar input patterns tend to be grouped into the same output pattern as much as possible. Suboptimal results can be obtained.

[0018]

As described above, by using the device allocation determining device by the competitive neural network of the present invention,
It is possible to easily obtain sub-optimal component allocation, which was difficult with the conventional heuristic method based on human experience. In addition, since it is not necessary to create a computer program, it can be easily used by a person who does not have knowledge of the program, and the cost and time for developing a special program are low.

[Brief description of drawings]

FIG. 1 is a configuration diagram showing an embodiment of a competitive neural network of a first invention.

FIG. 2 is a flowchart showing an algorithm of a learning phase of the competitive neural network of the first invention.

FIG. 3 is a flowchart showing an algorithm of an associative phase of the competitive neural network of the first invention.

FIG. 4 is an input / output pattern correspondence diagram of the competitive neural network of the first invention.

FIG. 5 is an operation example diagram showing an embodiment of a parts allocation determination device of the second invention.

[Explanation of symbols]

 1,2,3,4,5 Input node 6,7,8,9,10 Output node 20 Connection 30 Penalty node

Claims (2)

[Claims] 1. In a two-layer type neural network consisting of an input layer and an output layer each having the same number of nodes, an output node that behaves such that an output node corresponding to an input node in the ON state is fixed in the ON state. , Competitive neural network characterized by providing a penalty node that holds an arbitrary limit number and is connected to all output nodes so that the number of nodes in the ON state in the output layer can be kept below a certain limit number network. 2. The input node of the competitive neural network according to claim 1 is made to correspond to the component configuration of a plurality of types of printed circuit boards, the limit number is made to correspond to the maximum attachable component type number of the assembly machine, and the output node is set to A parts allocation determining device, which is adapted to parts allocation of a machine for assembling the printed circuit board.