Distributed fault detection

1 . A system, comprising:

at least one processor; and

at least one memory that stores executable instructions that, when executed by the at least one processor, facilitate performance of operations, comprising:

receiving, at a first computing node, first sensor data about a physical environment that is external to the system from at least one sensor;

analyzing, by the first computing node, the first sensor data with a restricted Boltzmann machine (RBM) neural network to determine whether there is a fault condition in the physical environment that is external to the system, wherein data used to train the RBM neural network comprises an identification of a safe state of the physical environment that is external to the system and omits an identification of the fault condition;

updating, by the first computing node, the RBM neural network based on the first sensor data to produce a first updated RBM neural network;

sending, by the first computing node, a first patch indicative of the first updated RBM neural network to a central server;

receiving, by the first computing node and from the central server, information indicative of a second updated RBM neural network, the second updated RBM neural network being based on an aggregation of the first patch and of a second patch generated by a second computing node, wherein the aggregation comprises combining neurons of the first patch that satisfy a similarity criterion with respect to neuron weight;

analyzing, by the first computing node, second sensor data with the second updated RBM neural network to determine whether there is the fault condition in the physical environment that is external to the system; and

in response to determining, by the first computing node and using the second updated neural network, that the second sensor data is indicative of the fault condition in the physical environment,

generating an alert, and

engaging an actuator to cause a physical change to the physical environment that mitigates the fault condition, wherein the first computing node is configured to engage the actuator independently of communicating with a centralized module, wherein a third computing node that stores a copy of the second updated neural network is configured to engage the actuator independently of the communicating with the centralized module and based on the third computing node identifying the fault condition, and wherein the engaging of the actuator is separate from the generating of the alert.

2 . The system of claim 1 , wherein the central server is unreachable by the first computing node when the first computing node performs the analyzing of the second sensor data.

3 . The system of claim 1 , wherein the fault condition is omitted from training data used to train the second updated RBM neural network.

4 . The system of claim 1 , wherein a first amount of time to analyze the second sensor data by the first computing node is less than a second amount of time to send the second sensor data to the central server and receive a result from the central server indicative of whether there is the fault condition in the physical environment.

5 . The system of claim 1 , wherein the operations further comprise:

iteratively updating the second updated RBM neural network.

6 . The system of claim 1 , wherein the second updated RBM neural network comprises an average of first neural network weights of the first patch and of second neural network weights of the second patch.

7 . The system of claim 1 , wherein the aggregation comprises adding a neuron from a layer of the first updated RBM neural network.

8 . A method, comprising:

analyzing, by a first computing node of a system comprising at least one processor, sensor data about a physical environment that is external to the system, the analyzing comprising using a neural network to determine whether there is a fault condition in the physical environment that is external to the system;

updating, by the first computing node, the neural network based on the sensor data to produce a first updated neural network;

sending, by the first computing node, a first patch indicative of the first updated neural network to a central server;

receiving, by the first computing node and from the central server, information indicative of a second updated neural network, the second updated neural network being based on an aggregation of the first patch and of a second patch generated by a second computing node, wherein the aggregation comprises combining neurons of the first patch that satisfy a similarity criterion with respect to neuron weight;

analyzing, by the first computing node, second sensor data with the second updated neural network to determine whether there is the fault condition in the physical environment that is external to the system; and

in response to determining, by the first computing node and using the second updated neural network, that the second sensor data is indicative of the fault condition in the physical environment,

rendering, by the system, an alert, and

engaging, by the first computing node, an actuator to cause a physical change to the physical environment that mitigates the fault condition, wherein the first computing node is configured to engage the actuator independently of communicating with a central module, wherein a third computing node that stores a copy of the second updated neural network is configured to engage the actuator independently of the communicating with the central module and based on the third computing node identifying the fault condition, and wherein the engaging of the actuator is separate from the rendering of the alert.

9 . The method of claim 8 , wherein the neural network comprises a restricted Boltzmann machine.

10 . The method of claim 8 , wherein the neural network comprises a stochastic recurrent neural network that omits intralayer connections between nodes.

11 . The method of claim 8 , wherein the physical environment is a first physical environment, and wherein the third computing node uses the second updated neural network to determine whether there is the fault condition in a second physical environment.

12 . A non-transitory computer-readable medium comprising instructions that, in response to execution, cause a system comprising at least one processor of a first fog node to perform operations, comprising:

analyzing sensor data about a physical environment that is external to the system comprising inputting the sensor data to a neural network and using an output of the neural network to determine whether there is a fault condition in the physical environment that is external to the system;

performing federated learning with respect to the neural network, comprising:

sending a first patch indicative of local learning of the federated learning to a server; and

receiving, from the server, information indicative of an updated neural network, the updated neural network being based on an aggregation of the first patch and of a second patch generated by a second fog node, wherein the aggregation comprises combining neurons of the first patch that satisfy a similarity criterion with respect to neuron weight;

analyzing, by the first fog node, second sensor data with the updated neural network to determine whether there is the fault condition in the physical environment that is external to the system; and

in response to determining, by the first fog node and using the second updated neural network, that the second sensor data is indicative of the fault condition in the physical environment,

raising an alert, and

engaging an actuator to cause a physical change to the physical environment that mitigates the fault condition, wherein the first fog node is configured to engage the actuator independently of communicating with a centralized component, wherein a third fog node that stores a copy of the second updated neural network is configured to engage the actuator independently of the communicating with the centralized component and based on the third fog node identifying the fault condition, and wherein the engaging of the actuator is separate from the raising of the alert.

13 . The non-transitory computer-readable medium of claim 12 , the performing the federated learning comprises:

determining a locally-trained updated neural network.

14 . The non-transitory computer-readable medium of claim 12 , wherein the first fog node is located at an edge of a network that comprises the server.

15 . The non-transitory computer-readable medium of claim 12 , wherein the first fog node is configured to analyze second sensor data with the updated neural network in response to the server being unreachable by the first fog node.

16 . The non-transitory computer-readable medium of claim 12 , wherein the fault condition is omitted from training data used to train the updated neural network.

17 . The non-transitory computer-readable medium of claim 12 , wherein a first amount of time associated with analyzing the sensor data by the first fog node is shorter than a second amount of time associated with analyzing the sensor data by the server.

18 . The non-transitory computer-readable medium of claim 12 , wherein a central server is unreachable by the first fog node when the first fog node performs the analyzing of the second sensor data.

19 . The non-transitory computer-readable medium of claim 12 , wherein the fault condition is omitted from training data used to train the second updated neural network.

20 . The non-transitory computer-readable medium of claim 12 , wherein a first amount of time to analyze the second sensor data by the first fog node is less than a second amount of time to send the second sensor data to a central server and receive a result from the central server indicative of whether there is the fault condition in the physical environment.