Diagnostic methods for motor driven apparatus

Field of the invention

The present inventive concept relates to a method for diagnosing the operational status of and faults within mechanical apparatus driven by electrical motors.

Background to the invention

Electrical motors are well known. There are various types of motors, such as brushed DC motors and AC motors. Generators use similar technology to produce an electrical current harnessing the movement of a fluid such as steam or air.

A wide range of apparatus (often referred to as a load) may be driven by motors, and a wide range of apparatus may drive a generator (e.g. in a wind turbine or power station). For the purposes of this application we will also refer to apparatus which drives a generator as a load.

Given that the load is usually a mechanical arrangement, maintenance may be needed during the lifetime of the apparatus. Loads may be fairly complex items such as pumps, which may be difficult to monitor and maintain. Ideally maintenance is only commissioned when needed, but when needed the need for maintenance must be identified without undue delay.

Ascertaining conditions which contribute to the status of a load would normally require a sensor measuring a quantity associated with the load.

A widespread challenge is that motor driven apparatus may be implemented in locations and situations where access to the apparatus is limited, space is limited, the placing of sensors may be difficult or impossible and so on.

There is therefore a need for a method to ascertain the status of a load, without additional sensors or the like. Statuses inferred would include but not be limited to, for example, detecting whether a pump is operating normally, blocked, or running dry.

In recent years, brushless direct current (BLDC) motors have become a common type of electrical motor due to their lifespan. Brushless motors require a more complicated control arrangement than traditional mechanical commutator motors.

Summary of invention

Motors and generators are generally controlled by a control system comprising driving electronics controlled by a microcontroller. The microcontroller is in turn operated by embedded software, which in turn may have a human-machine interface and/or a machine-machine interface. The term control system thus encompasses the said driving electronics, microcontroller and embedded software.

The control system not only provides control signals to the motor or generator but can receive signals from the motor or generator.

The present inventive concept uses signals derived entirely from the motor driving the load or the generator being driven by the load. Multiple statuses or measurements may be inferred. This would include but not be limited to, for example, detecting whether a pump (load) driven by the motor is operating normally, blocked, or running dry, as well as more granular or continuous measurements such as an estimate of an output pressure, flow rate and the like.

The present inventive concept therefore provides a method of providing an estimate of a status of a load driven by a motor, comprising the steps of:

- providing a motor assembly having motor and a motor controller, the assembly being mechanically linked to the load;

- providing a pre-trained neural net connected to the motor controller;

- obtaining monitoring signals from the motor controller;

- passing the monitoring signals to the neural net;

- operating the neural net in an inference mode to provide an output of a probability rating of a pre-determined range of load states;

- outputting a signal indicating the most probable load status.

References to motor are intended to include the alternative of a generator for the purposes of the inventive concept.

The present inventive concept can thus provide a virtual sensor. This is useful when monitoring a status and/or measuring a quantity or property is not feasible because of space requirements, or reliability of traditional sensors, or the physical property to be measured.

The present inventive concept can thus provide inferred information about a system that is mechanically coupled to a motor, including for example: pressure or flow rate in a fluid path (for example where the motor drives a pump), bearing vibrations, wind speed (for example where the motor is driven by a wind turbine), tyre pressure (for example where the motor drives a vehicle), fill level of a container or tank (for example where the motor drives a vacuum pump such as in a vacuum cleaner). The said inferred information can be passed to a control system to adapt the operation of the system, or inform a user. Thus the inferred information can be fed back to the motor controller to provide variation in the motor drive, for example to compensate for an inferred property being out of line with the preferred degree, to provide for an emergency stop or to trigger an alarm or alert. Furthermore, the inferred information can be processed to trigger a request for pre-emptive maintenance or the like.

The method may further comprise a pre-processing step comprising processing data into a form that is better suited to neural networks. For example, by spectral analysis or Fourier transformation. The goal of this step is to make the information hidden in the raw data as easy to detect as possible to the neural net. However, pre-processing is not always helpful, so this step may be left out and instead the raw data then used as-is.

The concept of so-called digital twins is an established technique for building virtual sensors. These are typically hand-crafted mathematical models that take the same inputs and some readily available sensor data as the real system and then attempt to simulate what's going on in order to infer information that is otherwise hard to get to.

The pre-trained neural net may comprise a model comprising a digital twin adapted to simulate the load, the digital twin comprising a model of the said load derived by machine learning.

A key difference between previously-known digital twins and the digital twin of the present inventive concept is that previously-known digital twins have been based on bespoke algorithms derived from engineers' understanding of the system's physics. In contrast, the present inventive concept uses machine learning to achieve the digital twin. A key advantage is that a digital twin for complex system can be built, which a human would struggle to express in a comprehensible set of formulae.

The neural net may be trained by continuously monitoring signals from the motor controller for a fixed period of at least one load cycle at a full sample rate. The said signals may be stored in data files.

The said training stage records motor monitoring signals for a fixed period of at least one load cycle at the full sample rate so that all characteristics of the load status are captured at least once. The motor may be repeatedly run for each load state, varying control parameters such as speed, building a dataset that covering as much of the motor and load's operating envelope as possible.

The described steps which lead to the neural net being trained may be described as a data capture mode.

This dataset may be then used to build and train a neural net using current Artificial Intelligence (Al) techniques of which the skilled reader will be aware. A classification neural network is used for training and inferring load states; and a regression neural network is used for training and inferring measurement levels.

In the said data capture or inference mode, motor data may be captured using the same techniques as described above; instead of being stored in files the data may be fed directly into the neural network producing a probability rating against a given set of load states or measurement levels - for example "normal", "blocked", "running dry”, and so on as described above.

Further processing (which can be referred to as post-processing) may be effected, for example, to detect and remove anomalies or noise in the output. The most probable category or measurement level is then passed on to the device software to handle it as required.

The monitoring signals may be one or more of motor speed, phase currents, voltage vectors, shaft speed, shaft angle.

The sample rate of monitoring signals may be of the order of hundreds to thousands of hertz (Hz).

The skilled reader will appreciate that the term microcontroller could also include a microprocessor.

A general purpose machine learning framework may be provided and adapted (which for clarity may include relevant configuration) for the specific purpose according to the present inventive concept. The skilled reader will appreciate that adapting a general purpose machine learning framework is a key step because the adaptation affects the output from the framework based on a given set of input data. Such adaptation may include pre-processing input data. The pre-processing stage may include, for example, using discrete Fourier transformation (DFT/FFT), spectral analysis, or reducing the size of the input data set for optimal application of the framework.

Such adaptation may include a selection of the number of training iterations (sometimes referred to as "epochs") to apply, or the amount of data processed in a given training cycle - sometimes referred to as "batch size”. Such a selection can mitigate phenomena such as under- or over-fitting.

The framework may comprise a known framework such as the commercially available Tensorflow (RTM) offered by Google LLC. We have found that a Tensorflow (RTM) architecture with twelve layers, of which 1 "Input” type, 7 "1 D Convolutional", 1 "1 D Max Pooling", 1 "Flatten" and 2 "Dense" types performs well.

Furthermore, we have developed our own bespoke double cosine decay learning rate scheduler. This has been found to give a well-performing machine learning neural net in fewer epochs - i.e. with increased efficiency.

Thus, the machine learning neural net may be adapted to have an architecture with twelve layers, of which 1 "Input" type, 7 "1 D Convolutional", 1 "1 D Max Pooling", 1 "Flatten" and 2 "Dense" types and a double cosine decay learning rate scheduler.

To help evaluate model performance, we use two distinct datasets:

One large (several thousands of samples) set of training data, which is directly used by the training system to tweak the neural net; and

One smaller (-20% the size of the training set) set of validation data, which is recorded using the same techniques, but in a completely separate session. This is used to evaluate model performance after each training epoch, with the point being that the model will have never seen this data before. This avoids the risk that the model "gets used to” training data and just memorises individual samples it's already seen, rather than learning to recognise the information in the data, which leads to deteriorated performance in the field.

Preferably the motor is a brushless direct current motor (BLDC). The present inventive concept also provides an arrangement comprising a motor assembly having a motor and a motor controller, the assembly being mechanically linked to the load, the arrangement further comprising pre-trained neural net connected to the motor controller, and wherein the motor controller is adapted to obtain monitoring signals from the motor, to pass the monitoring signals to the neural net, and wherein the neural net is adapted to operate in an inference mode to provide an output of a probability rating of a pre-determined set of load states, the motor controller being further adapted to output a message indicating the most probable load status.

Preferably, the pre-trained neural net is obtained using the training stage as described above.

Detailed description of the invention

An exemplary embodiment of the present inventive concept will now be described with reference to a brushless direct current motor (BLDC). The skilled reader will appreciate that effecting the inventive concept would be substantially similar with another type of motor, or a generator, for example.

A typical BLDC motor installation consists of a Motor Controller, Motor and Load. The Motor Controller receives input signal(s), for example, requesting the motor to run at a given speed, or generate a given torque. To drive the motor, the Motor Controller requires accurate monitoring signals, such as phase currents, shaft speed and angle at a high sample rate (typically hundreds to thousands of Hertz). Changes to the load have an observable impact on those monitoring signals.

The invention uses the motor drive data that is already present in the system to infer information about the load that would conventionally require additional sensors such as a flow meter to detect pump blockages. While some characteristics of a load may be inferable through classic feature engineering (e.g. differentiating between no load and a constant load through a torque threshold), dynamic load situations are often hard to detect through a fixed set of rules. The present inventive concept works especially for dynamic loads that are cyclic in nature. These can either be continuous, like a rotating pump, or disjunct, such as a stamp press that performs a single action after being triggered.

A neural net would normally run either on the same microprocessor controlling the motor or on a separate microprocessor adjacent to the motor. A connection to a remote processing unit would not normally be required for operation of the system but could optionally be provided to allow transmission of status, measurement or control signals.

The neural net would be pre-trained to recognise statuses or events associated with a specific type of load.

The neural net inputs may include, but not be limited to, motor speed, shaft angle, current and voltage vectors, as well as configuration data. Data directly or indirectly derived from these may also be input to the neural net.

Neural net training

The inventive concept records motor monitoring signals for a fixed period of at least one load cycle at the full sample rate so that all characteristics of the load status are captured at least once.

This is first done in a "data capture only” mode, where the data is continuously recorded to files instead of passed through a neural net. The motor is repeatedly run for each load state, varying control parameters such as speed, building a dataset that covering as much of the motor and load's operating envelope as possible.

This dataset is then used to build and train a neural net using state-of-the-art Artificial Intelligence (Al) techniques.

Depending on the load's characteristics, this can include pre-processing data into a form that is better suited to neural networks, for example, by spectral analysis or Fourier transformation. The goal of this step is to make the information hidden in the raw data as easy to detect as possible to the neural net. Pre-processing is not always helpful, so this step may be left out and instead the raw data then used as-is. As set out above, the machine learning framework may be adapted to be based on a Tensorflow (RTM) neural net having an architecture with twelve layers, of which 1 "Input" type, 7 "1 D Convolutional", 1 "1 D Max Pooling", 1 "Flatten" and 2 "Dense" types; and a double cosine decay learning rate scheduler.

To help evaluate model performance, we use two distinct datasets:

One large (several thousands of samples) set of training data, which is directly used by the training system to adjust the parameters of the neural net; and

One smaller (-20% the size of the training set) set of validation data, which is recorded using the same techniques, but in a completely separate session. As set out above, this is used to evaluate model performance after each training epoch, with the point being that the model will have never seen this data before. This avoids the risk that the model "gets used to” training data and just memorises individual samples it's already seen, rather than learning to recognise the information in the data, which leads to deteriorated performance in the field.

Operation

A trained neural net is deployed and the device run in "inference" mode.

Motor data is captured using the same setup and pre-processing as before. Instead of being stored in files, it is fed directly into the neural network producing a probability rating against a given set of load states or measurement levels (e.g. "normal", "blocked", "running dry”, and so forth as described above).

Further processing may happen, for example, to detect and remove anomalies or noise in the output. The most probable category is then passed on to the device software to handle it as required.

The inventive concept will be further described with reference to the accompanying drawings.

Figure 1 shows a flow chart showing the interaction of the elements of the inventive concept as disclosed. The desired speed, torque and the like are provided to the motor control - by way of a human-machine interface for example. The motor and load are connected to the motor control, which both controls the motor and monitors signals received therefrom. Signals received from the motor and load are in turn provided to pre-processing, neural network and post-processing elements which lead to a load status being output.

Figure 2 shows a flow chart of the logical steps which may take place in the present inventive concept. From the Start point, data from the motor controller is received and optionally buffered before being passed to the pre-processing step. After pre-processing, the data is passed to the neural network, to be run. Subsequently data is passed for post- processing, after which an estimated load status or measurements can be updated.