Communication failure tolerant distributed kalman filter

A method comprising:

• receiving input vectors of process measurements at a computer implemented Kalman filter from neighboring filters in a distributed network of computer implemented Kalman filters corresponding to an interconnection map of subsystem of a process to be controlled, wherein the inputs correspond to a local state of the process;
• determining a communication failure;
• responsive to the communication failure, setting correlations of the estimated local state with the state of subsystems estimated by filters that are not accessible due to the communication failure to zero;
• increasing the local state covariance;
• and
• in further response to the communication failure, when a mean value prediction of a particular interconnection input cannot be obtained from an upstream neighbor, changing a model of this unknown input from one based on white noise correlated with the local state to a colored noise model using a previous sum of the mean estimated by the neighbor and a locally estimated deviation to set the state of the coloring filter.

A system comprising:

• a distributed network of filters having a topology corresponding to an interconnection map of a process to be controlled;
• each filter connected to receive a vector of process measurements corresponding to a local state of the process and to receive process variables;
• each filter being aware of the topology of the network up to the network radius r and of models of the sub-processes within the same radius;
• each filter modeling sub-process coupling inputs as sums of received mean values predicted by neighboring filters, and noise, correlated with the local state;
• and
• each filter operating in time slots that are fractions of the normal control period and communicating repeatedly, within these slots with its network neighbors;
• the filters receiving estimates of coupling variables made by neighboring filters fusing this information into its own state estimate in a stochastically optimal way;
• the filters receiving from network neighbors further information including Kalman gains and covariance matrices related to sub-processes that are within network radius r in a way of multi-hop communication, and using these data for statistically correct update of cross-covariance estimates, and
• performing compensation for limited communication radius by increasing local state covariance estimate onto the local covariance;
• after performing a number of data updates, performing, in the sub-sequent time slot, a one-step ahead prediction of the local state mean covariance and cross-covariance with other states within the network radius r;
• wherein the filters are adapted to determine a communication failure and responsive to the communication failure, correlations of the local state with state of subsystems estimated by filters that are not accessible due to the communication failure are set to zero and at the same time, the local state covariance is suitably increased;
• and
• in a further response of the filters to the communication failure, when a mean value prediction of a particular interconnection input cannot be obtained from an upstream neighbor, changing the model of this unknown input from one based on white noise correlated with the local state to a colored noise model using a previous sum of the mean estimated by the neighbor and the locally estimated deviation to set the state of the coloring filter.