Distributed Microsystems Laboratory: Developing Microsystems that Make Sense

1
Distributed Microsystems LaboratoryDeveloping
Microsystems that Make Sense
Sensor Validation Techniques Sponsoring Agency
Center for Process Analytical Chemistry PI D.
Wilson Research Assistant Garth Tan
2
Distributed Microsystems LaboratoryDeveloping
Microsystems that Make Sense

• Goals To use various pattern recognition and
  preprocessing tools to validate incoming sensor
  data for enhancing system task accuracy and
  minimizing downtime
• Sensor Response Validation using Hidden Markov
  Models
• Preliminary Analysis of sensor arrays to
  determine experiment health
• Follow-up Analysis of individual sensors to
  pinpoint sensor health
• Sensor Baseline Validation using Statistical
  Analysis
• Gaussian Analysis detects deviations from
  Gaussian calibration distribution for each
  individual sensor
• Flags sensors, who in the absence of a stimulus,
  have drifted or otherwise deviated from
  calibration state sufficiently to diminish the
  accuracy of their contribution to system tasks
• Analysis of Data Sets/Proof-of-Concept
  Applications
• Composite Polymer Film Chemical Sensor Arrays
  Caltech
• Honeywell Sensor Arrays for Monitoring Vacuum
  Drier Process

3
Distributed Microsystems LaboratorySensor
Validation using HMMs

• What is a Hidden Markov Model (HMM)?
• An HMM is a doubly embedded stochastic process
  with an underlying stochastic process that is not
  observable (it is hidden), but can only be
  observed through an other set of stochastic
  processes that produce the sequence of
  observations. wrote Lawrence R. Rabiner, 1989
  1
• In simpler terms, it is a collection of states
  connected by transitions and emitting an output
  during each transition. The model is named
  hidden since the state of the model at a time
  instant t is not observable directly. wrote L.
  Satish and B.I. Gururaj, 1993 2
• The particular HMM model that is used to
  characterize response in this application is the
  Left to Right model. This model is more
  computationally efficient the model is forced to
  start in state one and transition sequentially.
  This results in a bi-diagonal transition matrix
  and eliminates the need for a probability vector
  describing the the models initial state.

4
Distributed Microsystems LaboratorySensor
Validation using HMMs

• What is a Hidden Markov Model (HMM)?
• The HMM is an analog state machine
• What does this mean?
• The model is trained to understand important
  transitions in the input training data
• Time between important transitions are the states
• The transitions themselves are modeled
  probabilistically
• The rules for switching states are determined
  during the training phase
• During the testing phase, the response travels
  through the model via a series of states,
  transitions between states having been defined by
  rules acquired during training.

5
Distributed Microsystems LaboratorySensor
Validation using HMMs

• Outputs of the HMM
• From each state
• Observation sequence sensor response
• Probability that an observation (progression in
  sensor response) can occur in that particular
  state for that particular model.
• The states through which the sensor progresses
  can be tracked over time.
• Can be used to indicate where major transitions
  in the sensor response that would identify it
  with a particular HMM would occur.
• The final output of the model is a probability
  that This model produced the current sensor
  response

6
Distributed Microsystems LaboratorySensor
Validation using HMMs

• Recent Results Chemically Sensitive Polymer
  Films
• Array Composition
• 10 types of sensors
• 4 types of each sensor
• 40 total sensors
• HMM trained on each set of four redundant sensors
  (4 inputs to each model, 10 total models).
• Data sets are cross-validated against different
  models.
• Cross-validation of Training Data exposes classes
  of sensors which cannot be readily discriminated
  (for good reason).
• Cross-validation of Testing Data captures
  unlike/invalid responses for all models.
• Conclusion the HMM method can detect broken or
  incorrect sensors at the sensor level and
  invalid/extreme environmental conditions at the
  system level.

7
Sample Data Acetone
Distributed Microsystems LaboratorySensor
Validation using HMMs

• Each row represents four sensors of the same type
• The output is taken as a percentage change from
  the baseline resistance, which is then normalized
  across the entire sensor response.
• Each figure shows results from each of 35
  experiments.

8
Recent Results Model 1 for Sensor Type 1
(Acetone)
Distributed Microsystems LaboratorySensor
Validation using HMMs
Sensor Responses
Progression of States
9
Recent Results Model 1 for Sensor Type 4
(Acetone)
Distributed Microsystems LaboratorySensor
Validation using HMMs
Sensor Responses
Progression of States

• The responses look similar to that of Sensor Type
  1, but a closer look reveals differences in
  magnitudes, slopes, and the final settling
  points.
• Sensor Type 4 was not able to move from model 1s
  first state, thereby causing it to fail in the
  model and be invalidated as a potential sensor of
  Type 1.

10
Results Model 1 for Sensor Type 2 (Acetone)
Distributed Microsystems LaboratorySensor
Validation using HMMs
Sensor Responses
Progression of States

• The responses look similar to that of Sensor Type
  1
• However, the sensor response did not finish in
  the third state as Sensor 1 does, but progresses
  onto the fourth state, thereby invalidating it as
  a match for the Model for Sensor 1.

11
Distributed Microsystems LaboratorySensor
Validation using HMMs

• Recent Results Acetone Sensor Validation
  Confusion Matrix
• The matrix matches sensor model to correct
  recognition of sensor type (along the diagonal).
  For example, the HMM for sensor type 1 (Model 1,
  first row) recognized sensors of type 1 (first
  column of first row) 100 of the time.

12
Discussion and Conclusions
Distributed Microsystems LaboratorySensor
Validation using HMMs

• The HMM Model picks out responses that are not
  characteristic. Failures of right sensors in
  right models tend to be related to noisy
  experiments rather than failure of the model.
• Once sensors are invalidated, they can be removed
  from the decision making process to improve
  overall system accuracy.
• Other work being done in the DMS lab determines
  the optimal number and combination of sensors in
  an array so that the array can be rearranged or
  reconstructed when a sensor is invalidated to
  provide the next best system accuracy.
• If a model is trained with an uncharacteristic
  response, it will not affect the models ability
  to train and test correctly, unless a majority
  of the responses used to train the model have
  problems. Therefore, poor training data does not
  significantly impact the accuracy of the model.