Throughput-driven condition-based maintenance for frequently reconfigured mass production equipment

Abstract

Heavily utilised reconfigurable production equipment has uneven deterioration rate caused by frequently changing configurations and complex subsystem interaction. Thus, traditional maintenance approaches that rely on reliability modelling are ineffective when applied on this class of equipment. This paper introduces a throughput-driven condition-based maintenance framework intended to perform predictive maintenance for heavily utilised reconfigurable production equipments. The effectiveness of the proposed method is experimentally validated using actual production records of the semiconductor test handlers. The primary benefit of the proposed framework is its prediction efficacy while accounting for the complex subsystem interaction through Bayesian statistics. It is low cost to implement because it only uses existing resources already available in most manufacturing plants.

Appendix

1.1 A—Semiconductor production data records field summary

Field	Type	Description
Index time	Raw	Time between end of test of a unit and start of test of the next
Test time	Configuration	The test time depending on the test programme used
Jam time	Raw	The length of time that the equipment spends idle while jammed
Jam count	Raw	The number of jams that occur during the run
Lot run time	Raw	The time taken for the lot to complete its run
Lot size	Configuration	The size of the lot
Net units	Raw	Number of units that pass the test
Expected sites	Configuration	Number of test sites configured to be used during the run
Actual sites	Raw	Average number of test sites actually operating
MCBJ	Processed	Average number of units tested in between jams
MSE	Processed	Average percentage of working sites during individual runs
Yield	Processed	The percentage of units in a test lot that passes the test
PPH	Processed	The number of units tested within the window of 1 hour
Handler model	Configuration	The model of the test handler used
Tester model	Configuration	The model of the attached tester
Run type	Configuration	The run type of the lot run: Normal run, Retest rejects and Quality control run

Configuration parameters are the inputs. Raw parameters are outputs from each lot. Processed parameters are that which are calculated from raw values. This type of data is normally automatically logged by the semiconductor test equipment in the Standard Test Data Format (STDF), which is a format used by all major semiconductor test equipment companies such as Teradyne and Verigy. It should be noted that on a case by case basis, the records for a production plant that does not adopt STDF would likely have raw and post-processed data fields.

1.2 B—Artificial neural network setup for CBM

State-of-the-art neural network toolbox in MATLAB was used to build ANN for CBM. 10,000 ANNs were trained with randomly generated network architectures. For each handler, the best network was chosen for benchmarking trials according to the following 2 criteria:

• 1. Best correlation to corrective action

• 2. Lead time of more than 5 lots

These criteria are exactly the same as those in Table 3. A lead time of 5 lots advance is chosen as it provides just enough time for maintenance to be scheduled. The randomisation parameters for the network architecture are 1–2 hidden layers, 5–15 nodes per hidden layer and the transfer function used are positive linear/tangent sigmoid combination.

Each of the networks is trained by 500 production records that have known maintenance decisions. It is ensured that a ratio of 1:1 between records with ‘Yes’ and ‘No’ decisions is maintained in the training set to enable sufficient exposure to bad handler during training phase. Six inputs (IT, TT, Jam time, Site count, MCBJ, PPH) and one output (Yes/No maintenance action) are used for the training process. These variables were selected as they were understood to have influence the health of the handler earlier in Sections 3.2–3.3.

The neural networks are all trained with the Levenberg–Marquardt optimization (trainlm), back-propagation supervised-learning training algorithm [24]. By default, trainlm invokes MATLAB’s dividerand function, which segregates training data into 3 sets, with a ratio of 70 to 15 to 15 [25]. These sets are used in succession for training, validation and testing of the neural network.