We developed a synthetic flame detector based on a 55-parameter feed-forward neural network receiving as input seven physical quantities measured at diferent locations of a gas turbine. The model has been implemented on an edge device for the machine control (MarkVIe) and the synthetic flame detection has been performed in real time. The model has achieved full recall and full precision on an independent test set. Further improvements will be aimed at implementing the model on a MarkVIs device, so that the control loop can be closed.
The dataset used for training the model was acquired from a gas turbine prototype running on a test bench. Temperature, pressure, flow, speed, acceleration values were measured by probes installed on diferent sections of the machine and acquired with a sampling frequency of 1 Sample/40 ms. With the same sampling frequency, the continuous () and boolean () signal from a physical flame detector installed on the turbine were also acquired. The boolean , whose value is calculated by the acquisition system based on the value of , and has value 1 whenever there is flame, 0 whenever the flame is out. In the training dataset the proper behaviour of the physical flame detector installed has been assessed by subject matter experts. A total amount of about 760 kSamples, corresponding to 7 days of gas turbine operation, have been used as training and validation sets (with an exclusive splitting with ratio 70% vs 30%), while an independent dataset counting about 1.1 MSamples and acquired during 10 test days has been used as our test.
We leveraged a fully data-driven approach and trained a simple fully-connected neural network
(NN) [
The choice of the input features and the NN hyperparameters was led by the maximization of the recall and precision (see Sec. 2.3) on an independent validation dataset. Also, we preferred architectures which would keep the number of parameters as small as possible, so that this solution could be easily implemented on MarkVIe, which is our edge device.
Recall and precision are computed on an event basis. We have a true positive when these conditions are satisfied: 1. switches from 1 to 0 at a certain time instant 2. ˆ switches from 1 to 0 at time ˆ, such that | − ˆ| < ∆ The time tolerance ∆ is 1.2 s, corresponding to 30 samples, and derives from functional requirements. In fact, we do not want to have a late flame out detection, since this may cause explosion risks due to methane accumulation. We may thin to relax this constraint to 40-50 samples in case of an earlier detection, which instead is a desired behaviour. Analogously with the definition of a true positive, a false positive is given when: 1. ˆ switches from 1 to 0 at time ˆ 2. keeps on having value 1 in the time frame [ˆ − ∆ ; ˆ + ∆ ] Finally, we have a false negative when: 1. switches from 1 to 0 at time 2. ˆ keeps on having value 1 in the time frame [ − ∆ ; + ∆ ]
MarkVIe [
Many techniques were explored to deploy flame detector model onto MarkVIe Edge Controller: 1. Converting model weights into ONNX format and further generate xml of functional blocks that could run on MarkVIe 2. Converting model into a Dynamic Link Library 3. Developing the NN model using the coding language of MarkVIe We used the third approach. The necessary steps are detailed as follow: • Convert the trained model into equations whose coeficients are the model weights • Map the equations onto MarkVIe functional blocks • Create the functional block library to represent the activations of the neural layers • Deploy and test the implementation on a virtual controller • Implement it onto production controller • Check that the inference time is compatible with the requirements of real-time application. As shown in Fig. 1, the model deployed on MarkVIe consists of one block for each elemental operation. The architecture starts with the scaling of sensor data read from the gas turbine, followed by the multiplication by the weights of the first dense layer. The output is further added with the bias and a ReLu is applied to introduce non-linearity. The weights of the second layer are multiplied by the output from ReLu, then the bias is added and finally a sigmoid operation is performed to have the final output.
The model was producing real-time information about the flame status (on/of), given that the inference time is less than 40 ms, which is the sampling interval used by the control system. We tested the model on the same gas turbine from which training data were collected. The test set was chronologically after the training set.
We obtained full recall, meaning that all the flame-out events were correctly detected, and full precision, i.e. no false alert for spurious flame-out was given by the model, as shown in Fig. 2. The flame-out was always detected slightly in advance with respect to the physical flame detector (also see Fig. 3), allowing a larger safety margin.
The model performance were checked both real-time, while the gas turbine was running, and ofline, for a deeper analysis of results, by downloading data from MarkVIe.
This work shows a real-case application of a synthetic flame detector consisting of a small neural network implemented on a MarkVIe edge-device and producing real-time information on the control system of a gas turbine.
The synthetic flame detector is trained using temperature, pressure, speed, flow time series collected by sensors installed on the machine, and leveraging the time series acquired from a physical flame detector as our ground truth. Basically, a digital twin of the physical flame detector sensor is built.
The model allows a full recall and full precision in the flame out detection on a test set independent from training and validation sets. The deployment on MarkVIe allows to have the information about the flame status in real time, while the gas turbine is running, so that control engineers are able to suddenly shut down the machine in case of a flame out, thus preventing an explosive environment to be generated.
The next step will be to use this flame detector in the control closed loop. To achieve this, we will need to deploy the model to MarkVIs, which is an enhanced version of MarkVIe aimed at automatically manage security-related machine controls.