NETWORK

FIELD OF THE INVENTION

The invention relates to a system for monitoring and controlling a dynamic network and more specifically, but not exclusively, to a system for monitoring and controlling a physical dynamic network such as an oil, gas, or water pipeline.

BACKGROUND TO THE INVENTION

Physical dynamic networks, and utility networks, pipelines and water reticulation networks specifically, are commonly monitored and controlled electronically. This is typically achieved in real time using a so-called supervisory control and data acquisition (SCADA) system which may include a variety of sensors, controllers, graphical user interfaces, and programmable devices. Such traditional systems rely on solving equations and absolute models to monitor the state of the network. These equations may be used to calculate energy or flow rates at specific points in the network.

SCADA systems are usually monitored by trained staff which interpret the feedback from the SCADA system and can control the system based on this. The training and experience of the staff is crucial to identify whether the system is operating normally or abnormally and to determine whether there are any fault conditions which may affect the network. As this relies so heavily on the abilities, training, and experience of staff, it is not possible to use the SCADA system to operate a dynamical system in an automated fashion. Prior art which discloses attempts to address aspects of this problem is described in further detail below.

United States patent number 6,155,292 in the name of Sakuji Kurata entitled “Method and a synthetic danger predicting and disaster preventing monitor system in which pipelines laid under the ground or through a structure are employed in their entirety for a synthetic danger predicting alarm sensor ¹ ’ discloses a method and a synthetic danger predicting and disaster preventing monitor system which includes a number of sensors associated with an underground pipeline which are employed to detect potential failures in order to carry out preventative maintenance on an affected portion before further damage might occur. A problem with the approach in this disclosure is that it relies heavily on solving equations in a deterministic fashion which is useful for a linear pipeline but cannot be applied to pipelines which have dynamic characteristics and changing boundary conditions.

United States patent number 8,447,706 in the name of Siemens Aktiengesellschaft entitled “Method for computer-aided control and/or regulation using two neural networks wherein the second neural network models a quality function and can be used to control a gas turbine" discloses a method which involves use of a cooperative learning method and artificial neural networks. Two feed-forward networks are linked to one another such that the architecture as a whole meets an optimality criterion and approximates the rewards observed to the expected rewards as an appraiser. In the network in this disclosure, the optimum action in respect of the quality function is modeled by a neural network which supplies the optimum action selection rule for a given control problem. A problem with this disclosure is that the method is specifically used to control a gas turbine which does not assist in monitoring and controlling a physical dynamic network.

United States 8,665,101 in the name of Aquarius Spectrum Ltd entitled “System method and device for leak detection and localization in a pipe network’ discloses a system for leak detection of a fluid in a pipe network which includes flow meters and vibration detectors adapted to be attached to a pipe at a location in the pipe network. A processor analyzes signals generated by the flow meters and vibration detectors to identify the presence of one or more leaks in the pipe network and detecting and localizing leaks in a pipeline network. The disclosure also provides for a device comprising a flow meter integral with a vibration detector for use in the system of the invention. Again, the system in this disclosure relies on solving equations from the flow and vibration data to detect the leaks in a deterministic fashion, which is not possible in a dynamic system.

United States patent number 7,705,747 in the name of Terahop Networks, Inc. entitled “Sensor networks for monitoring pipelines and power lines” discloses a sensor network for monitoring utility power lines which includes a sensor disposed for monitoring utility power lines capable of acquiring data related to the utility power lines and communicating sensor data. It further discloses a first remote sensor interface (RSI) which includes a data communications device capable of receiving the sensor data communicated from the sensor and transmitting data relating to the received sensor data. It further discloses a data communications device capable of receiving the data transmitted by the first RSI and transmitting data related to the sensor data directly or indirectly to a network external to the sensor network. A problem with this disclosure is, again, that it does not provide means of autonomously controlling aspects of the network and focuses specifically on the sensor networks.

None of the above disclosures, separately or collectively, disclose the invention as claimed and described herein.

OBJECT OF THE INVENTION

It is accordingly an object of the invention to provide a system for monitoring and controlling a dynamic network which, at least partially, alleviates at least some of the problems associated with the prior art.

SUMMARY OF THE INVENTION

In accordance with the invention there is provided a system for monitoring and controlling a dynamic network comprising:

- a plurality of sensors for measuring aspects of a state of the network;

- each sensor being associated with a segment of the network and connected to a virtual sensor which accumulates and preprocesses measurements from the sensors for each segment of the network;

- a network topology processor for storing the topology of the network and relating sensors and virtual sensors to segments of the network and neighboring sensors and virtual sensors in accordance with the topology;

- a reinforcement learning artificial neural network (ANN) based nonlinear state estimation and predictive control model which uses measurements from the sensors and virtual sensors to model the state of the network and estimate sequential states of the network; - a classifier for classifying conditions of the segments of the network such that the condition of each segment may be monitored;

- wherein preceding measurements from the sensors are temporarily stored and provided to the state estimation and predictive control model as inputs for modeling and estimating sequential states of the network; and

- wherein the ANN uses reinforcement learning to train the artificial neural network based on the difference between the estimated state and the measured state.

The system may include a central controller which receives the measurements from the virtual sensors, stores historical measurements and estimated states, and provides measurements to the classifier and the state estimation and predictive control model.

State of a system is the actual value of physical properties of the system at a time before present and present.

The virtual sensors observe, identify, and learn inter-relationships of the measurements from the sensors of the segment which collectively represent a dynamic state of the network segment and behavior of fluid flow across the relevant segment.

The virtual sensors receive and process temporal physical measurements from the sensors such as pressure, flow, temperature, and computes derived measurements such as signal frequency, wavelength, autocorrelation, and signal mode.

The network topology processor computes boundaries and divides the network into segments and determines parameters of the segment such as interconnections, length, number, and type of sensors which exhibit similar observable behavior.

The ANN may be a dynamic ANN with dynamically configurable number of layers and neurons per layer.

The sensors may include actuated sensors which measure the state of an actuator which acts upon the system such that the state estimation and predictive control model learns the effect of the state of the actuator on the state of the network. The actuators may be controlled by the system and control of the actuator may be based on the estimated states of the network by the state estimation and predictive control model and the desired state of the network.

The system may include a feature extraction element in the form of a recurrent neural network with long short-term memory which uses temporal signals to build abstract profiles and extract repeatable phenomenon from the temporal wave behavior, its frequency elements and ability to propagate.

The system may include one or more preprocessing elements for preprocessing the measurements from the sensors or virtual sensors for use by the state estimation and predictive control model.

The preprocessing elements may normalize the measurements of the sensors or virtual sensors, extract features from measurements and derived measurements, classify the extracted features which is provided to the state estimation model, and/or apply filters or operations to the measurements to provide preprocessed values to the state estimation and predictive control model.

The classifier may include an ANN in the form of a Bayesian neural network for classifying the condition of each portion of the network. The condition may be steady state, transitioning, or abnormal. The contributing factors of an underlying unique behavior that may cause the network to leave the steady state can be identified using the classifier.

The classifier may also validate measurements from adjacent sensors and identify faulty readings.

The system may include a selective activation element which selectively provides and or withholds signals during the learning process. The selective activation element may include synapse which encode winning and/or most successful short-term memory activations into long-term memory.

The selective activation element may be utilized during training of the model to provide or withhold signals such that effective selections, where the estimation error is smallest, are stored such that the effective selections may be made based on the stored values. The network topology processor stores the topology of the network as a graph. The graph may include vertices, which represent nodes of the network, and edges, representing sections of the network.

The topology processor may include a geographical relation element which relates the nodes and vertices of the network to a physical location thereof. The topology and geographical relation element correlate data structures that hold the spatial data, connectivity, real-time data, computed state space and model information of the network.

The state estimation model may include an anomaly detection element. The anomaly detection element detects measurements or derived measurements which fall outside of the estimated state of the network.

The system may include a fault detection element. The fault detection element detects observable measured events which are measured at multiple locations which do not fall within the estimated state of the network and classifies the event as a fault. The fault detection element may also detect measurements which correspond to a predefined fault condition. A predefined fault condition may be water hammer, leak, and/or implosion in a section of the network.

The fault detection element may include a feature extraction element and a classifier. The fault detection element may include a long short-term memory (LSTM) ANN which learns and extracts features across time steps in a window to detect faults which have a temporal aspect.

The fault detection element may include an input LSTM layer, multiple hidden layers, and output layer with softmax activation.

The fault detection element may include a geographical locating element, which uses the topological processor and the geographical relation element to locate the origin of the fault in the network. The geographical locating element may use the geographical relation element, wave parameters, and triangulation to determine the physical location of the origin of the fault. The system may include a graphical interface which shows the sections and nodes of the network schematically or to scale. The graphical interface may indicate the condition of each section of the network. The graphical interface may also indicate the type and location of a fault. The fault may be located based on the relative position of detection of specific features and the distance between detected features based on the network topology.

The network may be a pipeline network. The network may be a water reticulation network.

The sensors may include flow sensor, pressure sensors, temperature sensors, acoustic sensors, turbidity, pH, Oxidation Reduction Potential (ORP), and/or chlorine sensors.

An actuator may be a valve, wherein the sensor measures the operational position of the valve. An actuator may be a pump, wherein the sensor measure measures the operational level of the pump.

BRIEF DESCRIPTION OF THE DRAWINGS

An embodiment of the invention is described below, by way of a non-limiting example only, and with reference to the accompanying drawing in which:

Figure 1 is a schematic representation of a network; and

Figure 2 is a schematic representation of a system for monitoring and controlling a dynamic network. DETAILED DESCRIPTION OF THE DRAWINGS

With reference to the drawings, in which like features are indicated by like numerals, a system for monitoring and controlling a dynamic network is generally indicated by reference numeral 1 .

The system 1 described herein uses active temporal intelligence to learn the behavior of a physical dynamic network (such as an oil, gas, or water pipeline). The system includes a plurality of sensors 2 for measuring aspects of a state of the network 4. The measured aspect are physical properties and temporal signals of an inherently dynamical system and computes the relationship between signals from sensors 2 through deep learning and detects steady state changes and transients in the network 4. It should be noted that the network described in this embodiment is focused on a water reticulation network 4 and, as such, the description will refer to sensors 2 and actuators 5 which are typically used in such a network 4. Those skilled in the art will appreciate that the embodiment may be suitably modified, with the necessary changes to sensors, actuators, and layout, and be applied to a wide variety of dynamic physical networks. Figure 1 shows a portion of a very basic network which has been simplified to assist in the description of the example embodiment herein. However, the embodiment will provide sufficient context and information for the skilled person to apply it to much more complicated systems.

Each sensor 2 is associated with a segment 3 of the network and forms part of a virtual sensor which accumulates and preprocesses measurements from the sensors for each segment 3 of the network 4. In this example, the sensors shown in Figure 1 are virtual sensors 2. Each virtual sensor 2 represents the collective measurements from primary sensors or transducers of a section 3 of the network 4. The notation used herein associated the virtual sensors 2 sections 3 by its annotated value (for example, sensor 2 _a represents the collective measurements for section 3 _a , sensor 2b for section 3b and so on up to sensor 2x _+n representing section 3 _x+n ). These virtual sensors acquire measurements from several sensors or transducers (indicated collectively by reference numeral 2 _Z in Figure 2). For example, virtual sensor 2 _a (which is the virtual sensor for section 3 _a ) may receive measurements from transducers 2 _åi and 2ZN whilst virtual sensor 2b will only receive measurements from transducer 2 _Ziii . The transducers 2 _Z may include one or more of the following types of sensors 2 _Z placed at multiple locations along the relevant section 3:

- flow sensors,

- pressure sensors,

- temperature sensors,

- acoustic or sensors, turbidity sensors pH sensors

- Oxidation Reduction Potential (ORP) sensors, and/or

- chlorine sensors.

Such sensors are typically applicable where a water pipeline or reticulation network is monitored by the system 1 but may include many other types which might be useful in this application or other applications. In addition to receiving the raw measurements from the transducers 2 _Z , the virtual sensors 2 accumulates and preprocesses measurements such that the measurements from multiple sources with differing types of transducers 2 _Z and measurements may be interpreted in a meaningful way. The virtual sensors 2 receive and process temporal physical measurements from the sensors 2 _Z such as pressure, flow, temperature, and computes derived measurements such as signal frequency, wavelength, autocorrelation, and signal mode. All physical and derived measurements may be preprocessed and used by the system 1 , the virtual sensors 2 directly, or through the central controller 8 (described in further detail below). The virtual sensors 2 observe, identify, and learn inter-relationships of the measurements from the transducers 2 of a segment 3 which collectively represent a dynamic state of the network 4 segment 3 and behavior of fluid flow across the relevant segment 3.

The system 1 also receives information from actuated sensors 2A which measure the state of an actuator, shown in this example as valves 15 and pumps 16, which acts upon the network such that the state estimation and predictive control model 9 learns the effect of the state of the actuator 2A on the state of the network. This also allows the network to be controlled by the system 1 , based on the learned effects of the operational state of the actuators and estimated effect in order to reach a desired state of the network 4.

The system 1 includes a network topology processor 6 for storing the topology of the network 4 and relating virtual sensors 2, and the relevant positions of transducers 2 _Z within the segment, to segments 3 of the network 4 and neighboring sensors 2 in accordance with the topology. The topology processor 6 has the ability to change with the dynamics of the networked system 1 as the topology processor 6 computes the modified network 4 relationships in real time. The network topology processor 6 computes boundaries and divides the network 4 into segments 3 and determines parameters of the segment 3 such as interconnections, length, number, and type of sensors 2 which exhibit similar observable behavior. The division of network into segments are dynamically computed on the basis of network 4 topology, interconnections, and length of the specific section. This also allows changes to the network to be detected. The network topology processor 6 stores the topology of the network as a graph. The graph may include vertices, which represent nodes of the network, and edges, representing sections of the network. The graph may be stored by the processor in the form of an adjacency matrix. The topology processor may include a geographical relation element 7 which relates the nodes and vertices of the network 4, as stored by the topology processor 6 to a physical location thereof. The physical location will typically include geographic coordinates. The topology processor 6 and geographical relation element 7 correlate data structures that hold the spatial data, connectivity, real-time data, computed state space and model information of the network.

The system includes a central controller 8 which receives the measurements from the virtual sensors 2, stores historical measurements and estimated states, and provides values to a state estimation and predictive control model 9 and to a classifier 10. The controller 8 orchestrates messages and transfers data between the various components of the system. In this way, it acts as a message orchestrator for data transfer, integration and processing between the various components. The controller may also be responsible for further preprocessing before the values are provided to the model 9 or the classifier. The controller, along with the topology processor, may use the raw data from the sensors, along with their topological connections and physical locations, to derive preprocessed values which would be observable between neighboring sections. These values may, in addition to normalized measurements, be provided as inputs 11 to the components of the system 1. The model inputs 11 A may differ from the classifier inputs 11 B and the feature extraction inputs 11 C, and the central controller processes the measurements and information provided by the sensors 2 into a usable form for the components (shown in this example as the model 9 and the classifier 10 in Figure 2).

The state estimation and predictive control model 9 includes a reinforcement learning artificial neural network (ANN) which uses inputs 11 to model the state of the network 4 and estimate sequential states of the network 4. The ANN is a dynamic ANN with dynamically configurable number of layers and neurons per layer which optimized for a specific network 4 and relies on deep learning for training and automatic reconfiguration thereof. As such, the architecture of the network will vary considerably between different implementations. The error between the predicted sequential state and the actual measured state is used as feedback to train the network to reliably estimate the state of the network through reinforcement learning. The predicted values are also fed back into the central controller as inputs 13.

A selective activation element 14, which is a novel concept developed by the inventors, receives information from the central controller 8 and the model 9 and selectively activates and deactivates signals provided to the controller 8 during the learning process. The inventors refer to the selective activation element 14 as an agonist engine, these terms being used interchangeably in this description, because of its similarity to the way a human brain reacts to chemical agonists. The selective activation element 14 includes synapses which encode winning and/or most successful short-term memory activations into long-term memory. The selective activation element 14 is utilized during training of the model to provide or withhold signals such that effective selections, where the estimation error is smallest, are stored such that the effective selections may be made during monitoring and control based thereon.

The inputs 11 may include historical measurements and values such that temporal aspects of the measurements may be trained into the model. The learning of temporal and dynamic aspects of the network 4 is crucial to modeling the behavior of dynamical networks based upon the acquisition of the knowledge of the network 4 and its operational states. A dynamical system transitions through different states and establishes relationships between the participating real-world variables (influencing factors) that are quantifiable and measurable. The relationships can be estimated using the model 9 such that the behavior of the dynamic network 4 may be learned over a period of time to establish the states. This allows the system 1 to extract relationships that are impossible to deduce through flow or energy equations and also to detect and control state changes in real-time without external stimuli or intervention. It also allows monitoring the inherent behavioral characteristics of the dynamic network 4, otherwise not observable or perceivable by humans. The model 9 learns states of the system 1 from inputs 11 and creates the model 9 of the network 4 around the behavior and inter-relationship of the inputs within a dynamic boundary computed by the system 1 . The model 9 differs from deterministic models such as hydraulic models as it does not require energy or flow equations to be solved deterministically. The deep neural network of the model 9 learns the relationships of the inputs presented and creates representation of the various states of operation.

The classifier 10 includes an ANN in the form of a Bayesian neural network for classifying the condition of each section 3 of the network 4. The condition may be steady state, transitioning, or abnormal. The contributing factors of an underlying unique behavior that may cause the network to leave the steady state can be identified using the classifier 10. The classifier may also validate measurements from adjacent sensors and identify faulty readings.

The system 1 also includes a feature extraction element 17. The feature extraction element 17 includes a recurrent neural network with long short-term memory (LSTM) which uses temporal signals to build abstract profiles and extract repeatable phenomenon from the temporal wave behavior, its frequency elements, and ability to propagate in relation to the network 4. This observed phenomena, typically attenuated over distance, may thus be used to automatically construct the topology of the network 4 as unique features are detected as it propagates through the network 4.

The feature extraction element may include a fault detection element which detects faults based on detected temporal features. The fault detection element may include an input LSTM layer, multiple hidden layers, and output layer with softmax activation and detects observable measured events, or features, which are measured at multiple locations which do not fall within the estimated state of the network and classifies the event as a fault. The fault detection element may also detect features which correspond to a predefined fault condition such as water hammer, leak, and/or implosion in a section of the network. The fault detection element learns and extracts features across time steps in a window to detect faults which have a temporal aspect. The fault detection element may include a geographical locating element, which uses the topological processor and the geographical relation element to locate the origin of a fault in the network 4. The geographical locating element may use the geographical relation element, wave parameters, and triangulation to determine the physical location of the origin of the fault.

Output of the various components of the system 1 may be correlated and displayed on one or more graphical user interfaces 18. The interface 18 shows the sections and nodes of the network 4 schematically or to scale (similar to the network 4 shown in Figure 1). The graphical interface 18, when displaying the network 4 for monitoring purposes, will indicate selected condition of each segment 3 of the network 4. The condition may be the state, it may indicate conditions such as steady state or transitional states, or detected anomalies, features and faults, all overlaid on the topography as displayed. These indications may be in the form of overlaid values in text, graphs, alerts, or colors associated with the various conditions detected. The graphical interface may also indicate the type and location of a fault which is located based on the relative position of detection of specific features and the distance between detected features based on the network topology 4 and is able to differentiate and visualize segments 3 of the network 4 that exhibit a unique behavior.

It is envisaged that the invention will provide a system for monitoring and controlling a dynamic network which enables operators thereof to achieve round-the-clock autonomous monitoring of the entire physical network in a supervised or autonomous fashion. The system improves efficacy of control loops and, through judging on a probable behavior patterns, improves operator visibility and response. The system can accurately determine the incidental operational state and is useful to monitor, assess, and control deviations and maintain stability of the system in an autonomous fashion.

The system further enables reading and interpretation of the information received through re-processing of the various measurements made by all the transducers in near real time so as to constantly control the network, defining a new and innovative methodology of continuous and constant control of fixed structures and, in the specific case of pipeline networks, enabling round-the-clock monitoring thereof. It provides operational intelligence on the state of the entire dynamical system, enabling detection of any alterations thereto and is able to identify possible fault conditions and representing geographically a map of the network with the indications of the fault conditions found. The invention is not limited to the precise details as described herein. For example, instead of using an LSTM ANN to process temporal data, another ANN with a suitable topology for processing temporal signals, such as a recurrent neural network with some form of temporal signal processing, may be employed. Further, instead of the system being used to monitor and control a pipeline, the system may be used to control any physical dynamic system such as a power grid or traffic network.