Prognostic system and method for thermal spray plasma gun nozzle-electrode degradation

Abstract

The present invention relates to a system and method for monitoring the condition of an electrode-nozzle pair in a thermal spray plasma gun. The system samples high-frequency voltage data and employs two distinct analytical models to estimate the health index (HI) and the remaining service life (RSL) of the electrode-nozzle pair. Firstly, a Deep Semi-Supervised Anomaly Detection Method (DeepSAD) generates a health index from a training dataset to determine the current health state of the electrode-nozzle pair. Secondly, the Remaining Service Life of the electrode-nozzle pair is predicted using a Mixture of Local Experts model. The system provides actionable alerts to the operator to exchange the electrode-nozzle pair before start of the next coating run or shutting down the thermal spray process at the end of the current run when coating is completed if the health index exceeds a pre-defined value.

Description

[0001] Prognostic System and Method for Thermal Spray Plasma Gun Nozzle-Electrode Degradation

[0002] The present invention relates to a system and method for monitoring the condition of an electrode-nozzle pair in a thermal spray plasma gun. The system samples high- frequency voltage data and employs two distinct analytical models to estimate the health index (HI) and the remaining service life (RSL) of the electrode-nozzle pair. Firstly, a Deep Semi-Supervised Anomaly Detection Method (DeepSAD) generates a health index from a training dataset to determine the current health state of the electrode-nozzle pair. Secondly, the Remaining Service Life of the electrode-nozzle pair is predicted using a Mixture of Local Experts model. The system provides actionable alerts to the operator to exchange the electrode-nozzle pair before start of the next coating run or shutting down the thermal spray process at the end of the current run when coating is completed if the health index exceeds a pre-defined value.

[0003] Background

[0004] Thermal spray methods, such as plasma spray, are coating methods using a heat source to melt a feedstock material and deposit the molten material on a substrate using a process gas. The material droplets accelerated towards the substrate solidify rapidly on the surface of the substrate or on a previously applied coating layer in the form of splats that overlap and build up the coating layer. As the molten particles and the heat source heat up the substrate, the heat source needs to travel at a certain speed relative to the substrate surface and thus only a limited coating thickness may be achieved during a single coating pass. Therefore, depending on the geometry of the substrate, the coating material to be applied and the final coating thickness to be achieved, several coating passes may be required to achieve the final coating thickness.

[0005] Thermally sprayed coatings are widely used in different industries to improve wear, friction, heat resistance or abradability of the parts being coated. Depending on the purpose to be achieved with the coating, a wide range of possible coating material feedstocks in the form of wire or powder can be used. Plasma spraying methods use a plasma jet to melt the coating material supplied and to accelerate the molten material towards the substrate to be coated. The plasma is generated by an arc established between the electrode (cathode) and the nozzle (anode) of the thermal spray plasma gun, which ionizes the plasma gas supplied to the thermal spray plasma gun. The electrode and nozzle paired on the same thermal spray plasma gun will be referred to as electrode-nozzle pair. In plasma spraying processes and atmospheric plasma spraying (APS) processes in particular, several parametric drifts and fluctuations of the process can be observed which lead to changes in the properties of the coating layer applied. These drifts and fluctuations stem from various sources, such as plasma jet instabilities, powder feed rate and / or powder carrier gas flow fluctuations and wear of the thermal spray plasma gun parts.

[0006] The condition of the electrode-nozzle pair is an important factor affecting the properties of the coating applied. Due to the high temperatures, the electrode and the nozzle wear. This wear leads to a decrease in the average thermal spray plasma gun voltage and / or an increased thermal spray plasma gun voltage fluctuation. While thermal spray plasma guns operate on direct current, the voltage fluctuates rapidly. A decreasing average or mean thermal spray plasma gun voltage and / or increased voltage fluctuations result in undesirable properties of the coating layer applied, such as increased porosity, lower hardness and / or decreased adhesion of the coating to the substrate or decreased cohesion within the coating and / or potentially in a decreased deposition efficiency as well, thus a higher percentage of the coating material fed into the process not being deposited on the substrate. Accurately determining the most appropriate time for replacement of a thermal spray plasma gun’s electrode-nozzle pair therefore ensures adherence to stringent coating quality standards, optimizes the utilization of the equipment and reduces cost for consumables used for the coating process. The service life of a specific electrode-nozzle pair for a specific type of thermal spray plasma gun varies from 10 to 60 hours depending on the arc energy the electrode-nozzle pair is subjected to, the plasma gas and the plasma gas flow used during the process, the assembly process of the electrode-nozzle pair, the geometric manufacturing tolerances and the tolerances of the material composition of the electrode and nozzle. According to a prior art approach, the electrode and / or nozzle are exchanged after a predefined operating time, indicated in the manuals of the thermal spray plasma gun, to ensure that the nozzle and the electrode are exchanged before the wear of these components leads to an unacceptable coating quality of the parts coated. However, the operating time that can be achieved with electrodes and nozzles of a thermal spray plasma gun until the properties of the coating layer applied are beyond the specifications may vary depending on several factors. One of these factors is the coating parameter set applied for coating parts. Depending on the thermal spray plasma gun voltage, the current, the plasma gas used and the plasma gas flow used, the electrode and / or nozzle are wearing faster or slower.

[0007] Another factor influencing the wear of the electrode and nozzle is the torque applied to assemble the electrode-nozzle pair. The variation in torque applied for assembly leads to a different distance of electrode and nozzle and / or to loosening of these components after a certain number of ignitions or operating hours, which, in turn lead to a variation of the arc behaviour.

[0008] Other factors influencing the wear of the electrode and nozzle are manufacturing tolerances and material composition of the electrode and nozzle, as well as defects on the surface of the electrode and nozzle. All these factors lead to variation, and in most cases to a shortening of the operation time of the electrode-nozzle pair after which the components are worn such that the properties of the coating layer applied can no longer be accepted. Therefore, according to this prior art approach the operators of the thermal spray plasma gun tend to preliminarily exchange the electrode and / or nozzle to ensure that an acceptable coating quality is produced. This approach therefore leads to high cost and maintenance efforts, and reduced machine availability.

[0009] According to another prior art approach, the wear of the electrode-nozzle pair of the thermal spray plasma gun is determined using the average thermal spray plasma gun voltage. Thermal spray lasma controllers provide an approximation of the thermal spray plasma gun voltage by measuring the voltage, sampled at 1 Hz, within the junction and monitoring box (JAMBox). The JAMBox is connected to the thermal spray plasma gun using hoses. When the average thermal spray plasma gun voltage drops below a certain threshold, the coating process is stopped, and the electrode and nozzle are exchanged. This prior art approach is not reliable and requires continuous recalibration because the average thermal spray plasma gun voltage is influenced by other factors than the wear of the nozzle and electrode, such as variations of the hardware used, the operating conditions and environmental factors. Furthermore, the voltage signal acquired according to this prior art approach does not reflect the voltage within the thermal spray plasma gun where by an arc gas is converted to plasma. According to this prior art approach, the voltage signal is recorded within the JAMBox. Factors such as length, number and sharpness of bends and wear of the hoses connecting the thermal spray plasma gun to the JAMBox influence the voltage measured within the JAMBox. Therefore, the voltage measured within the JAMBox is only partially representative for the voltage being present at the thermal spray plasma gun. In addition to this, the sampling frequency of the voltage reduces the reliability of the voltage signal on one hand and on the other hand the above mentioned setup does not allow to capture all the data relevant for assessing the health state of the electrodenozzle pair.

[0010] In still another prior art approach, a camera is used to analyze the temperature and velocity of particles within the spray plume and the form of the spray plume generated by a thermal spray plasma gun. This approach allows to compare the parameters measured with a reference set for a given coating parameter setup before spraying parts and to detect anomalies of the spray plume. In contrast to the prior art approaches mentioned above, this prior art approach determines the condition of the spray plume rather than the condition of the thermal spray plasma gun components. However, due to its size and weight and, due to the fact, that a camera mounted to the thermal spray plasma gun manipulator would induce a danger of collision during the spray process, the camera used is normally mounted within the spray booth to monitor the spray plume during the startup and not during coating process.

[0011] According to yet another prior art approach, a flow induced acoustic signal recorded during the coating application process is used to analyze and determine the condition of the electrode-nozzle pair (See PhD Thesis of Blair, T.K. (2015), “Development of a Plasma Spray Process Monitoring System through Aeroacoustic Signal Analysis”, Available at: https: / / vtechworks.lib.vt.edu / server / api / core / bitstreams / 106ee589-8be9- 445e-8e81 -fa19f93da4ca / content, (accessed: 14.03.2024)). Offline measurements performed using this approach by generating a controlled gas flow through nozzles of a Oerlikon Metco 9MB thermal spray plasma gun, recording the acoustic signals and analyzing them showed that this approach is useful for identifying used nozzles as well as for identifying variations of the nozzle and powder port geometry. Using the same prior art approach, online measurements of acoustic signals were found to enable distinguishing between many different sources of coating process variations and to detect and identify causes of changes. However, this prior art is used to mitigate causes of coating process variations after they occurred, but it does not provide any information about the remaining service life of the electrode-nozzle pair and it does not allow to prevent anomalies of the coating process before they occur.

[0012] In CN110699683A, real-time measurement data from a pressure sensor, a temperature sensor and a powder flow monitoring device mounted to a cold spray gun are analyzed based on a convolutional neural network algorithm. Based on the realtime data gathered and analyzed, a microcomputer adjusts the pressure and powder flow rate to maintain pressure and temperature within the cold spray gun in order to achieve a stable coating layer quality. The application of this prior art approach is limited to thermal spraying by means of expansion of highly pressurized gas without combustion where there is not arc generated between an electrode and a nozzle. Like all other prior art approaches mentioned above, this prior art approach does not provide information about the remaining service life of thermal spray plasma gun components. Additionally, this prior art approach relies on sensors applied within, thus intrusive to the cold spray gun. The application of the sensors required for gathering process data influences the process and thus might lead to a discrepancy between the real condition of the process and the condition as measured.

[0013] Taking the above said into account, it becomes clear that the prior art approaches fail to offer solutions for accurate, real-time, non-disruptive and non-intrusive determination of the electrode-nozzle pair condition of a thermal spray plasma gun without need for frequent recalibration. Moreover, the prior art approaches mentioned above allow to determine the condition of a thermal spray process and to intervene if anomalies in the process are detected or they rely on a preliminary exchange of the electrode and / or nozzle but they lack the capability to predict the remaining service life of the electrode-nozzle pair and thus they fail to enable preventive maintenance of the thermal spray gun components whilst at the same time maximizing the service life of the electrode-nozzle pair.

[0014] Objective of the invention

[0015] The objective of the present invention therefore is to provide a system and a method to determine the health state of a thermal spray plasma gun electrode-nozzle pair in real-time, in a non-disruptive and non-intrusive manner in order to predict the remaining service life (RSL) to enable predictive maintenance of the electrode-nozzle pair.

[0016] Summary of the invention

[0017] According to the invention, this objective is met by equipping the thermal spray plasma gun of a thermal spray system with a high-frequency thermal spray plasma gun voltage measurement sensor, sampling on a computing device the thermal spray plasma gun voltage in real-time processing the data sampled. The Mel spectrograms are extracted and analyzed providing a health-index (HI) and the corresponding remaining service life for the thermal spray plasma gun electrode-nozzle pair. The alerts are produced and displayed on the human machine interface (HMI) of the thermal spray system controller, and new ignition of the thermal spray system is prevented if the remaining service life approaches a predefined threshold.

[0018] According to one aspect of the invention, a system for determining the health state and predicting the remaining service life of a thermal spray plasma gun’s electrode-nozzle pair is proposed. According to one embodiment of the present invention, the system for determining the health state and prediction of remaining service life of a thermal spray plasma gun’s electrode-nozzle pair comprises a high-frequency voltage sensor configured to generate at least one time-dependent thermal spray plasma gun voltage signal. In some embodiments, the high-frequency voltage sensor is configured to provide a sampling frequency of at least 20 kHz and up to 100 kHz, preferably a sampling frequency of at least 20 kHz and up to 50 kHz. Compared to the prior art sampling frequency of about 1 Hz, the increased sampling frequency allows to isolate and analyze specific frequency components indicative of the electrode-nozzle pair condition. The inventors have identified that specific thermal spray plasma gun voltage patterns within the frequency ranges of 3 kHz up to 10 kHz can be used to identify wear and tear of the electrode-nozzle pair. Therefore, to analyze amplitudes of the thermal spray gun voltage signal in a range up to 10 kHz it is essential that the sampling frequency for the thermal spray gun voltage signal is at least 20 kHz.

[0019] In one embodiment, the high-frequency high-frequency voltage sensor is mounted directly to the thermal spray plasma gun’s water in- and outlet without the high- frequency voltage sensor intruding into the thermal spray plasma gun. In this embodiment, in contrast to prior art approaches, hoses connected on one end to a junction and monitoring box (JAMBox) are connected at the other end to the high- frequency voltage sensor rather than directly to hose connection points at the thermal spray plasma gun. . Measuring the voltage directly at the water in- and outlet of the thermal spray plasma gun allows to measure the voltage as close as possible to the arc generated between the electrode and nozzle without specifically adapting the thermal spray plasma gun to the high-frequency voltage sensor. This non-intrusive and non-disruptive measurement of the thermal spray plasma gun voltage allows to measure the gun voltage without influencing the thermal spray plasma process.

[0020] The system further comprises a computing device. Primarily, the computing device is configured to convert the high-frequency voltage signal into Mel spectrograms, which are used as input data, for both methods. Based on these inputs, the computing device further determines the health index of the electrode-nozzle pair using the DeepSAD method. Furthermore, the computing device predicts the remaining service life using an adaptive mixture of local experts model, using the Mel spectrograms as input. In a preferred setup, upon determining that the remaining service life is approximately one hour, the computing device sends a signal to the thermal spray system controller for indicating the remaining service life and a signal for preventing ignition and / or for shutting down the thermal spray process. Concurrently, it triggers the controller to display a notification that the electrode-nozzle pair needs replacement. This notification remains active until an operator resets the system following the replacement of the electrode and / or nozzle.

[0021] According to another aspect of the invention a method for determining the health state of a thermal spray plasma gun’s electrode-nozzle pair by predicting the health index and the remaining service life is proposed. The method comprises the following steps. Thermal spray plasma gun voltage signals sampled at 20 kHz are sent to a computing device. The computing device converts the gun voltage signals into Mel spectrograms and provides a health index (HI) ranging from 0 to 1.0 to determine the electrodenozzle pair wear based on the DeepSAD with an incorporated diversity loss method. The method further comprises predicting the remaining service life of the thermal spray plasma gun’s electrode-nozzle pair by applying an Adaptive Mixture of Local Experts method and providing a signal to the HMI of the thermal spray controller indicating the remaining service life and / or providing a signal to the thermal spray controller to shut down the thermal spray process if the remaining service life is equal to or less than one hour.

[0022] The method further comprises training the neural network methods indicated above using a limited set of up to 7 different electrode-nozzle pair sets of one electrode-nozzle pair type. The electrode-nozzle sets are subjected to a thermal spray process expediting the wear and tear of the electrode-nozzle pair. During this process the electrode-nozzle pairs are closely monitored by an expert in the field of thermal spray plasma coating to identify the onset of a state of significant wear. The nozzles showing commencement of wear and tear are flagged as abnormal by the expert. Based on the high-frequency voltage data gathered during these test sequences and based on the expert labelling of the samples, the neural networks are trained.

[0023] According to an embodiment, a neural network method is applied to calculate the health index of the electrode-nozzle pair. In a preferred embodiment, the calculation is based on the DeepSAD method.

[0024] According to an embodiment, the calculation of the remaining service life is based on the method of Adaptive Mixture of Local Experts, a model that can leverage specialized models referred to as experts. The method generates a number between 0 and 1 , where 1 means that 0% of the service life of the electrode-nozzle pair is remaining and 0 means that 100% of the service life is remaining. This number is then compared with the run time of the electrode-nozzle pair accumulated to calculate the remaining service lifetime in hours of run time. The amount of remaining service life is displayed on the thermal spray process controller and a shutdown of the process is triggered if the remaining service lifetime is equal to or less than 1 hour.

[0025] In accordance with embodiments, it has been found that the inventive system and method for determining the health state by predicting the health index and the remaining service life of a thermal spray plasma gun’s electrode-nozzle pair reliably and consistently allows to determine the health state of the thermal spray plasma gun’s electrode-nozzle pair and to predict the remaining useful life. Experiments have shown that the end-of-life predicted by the method proposed is consistently before or maximum at the same time as the end-of-life label provided by an expert in the field. The method proposed thus allows to accurately determine the most appropriate time for replacement of a thermal spray plasma gun’s electrode-nozzle pair and therefore ensures adherence to stringent coating quality standards, optimizes the utilization of the equipment by reducing the likelihood of unexpected downtime and reduces cost for consumables used for the coating process.

[0026] It is preferred that the inventive solutions are further improved by equipping the nozzle and / or electrode with an RFID chip to ensure that the alert provided by thermal spray controller is reset only if the worn nozzle has been replaced.

[0027] The invention shall now be further exemplified with the help of figures.

[0028] Description of figures

[0029] Figure 1 depicts a cross section of a conventional thermal spray plasma gun.

[0030] Figure 2 illustrates a conventional thermal spray plasma system. Figure 3 shows a system for determining the health state and predicting the remaining service life of a thermal spray plasma gun’s electrode-nozzle pair according to the invention.

[0031] Figure 4 shows a high-frequency voltage signal recorded using an F4MB thermal spray plasma gun.

[0032] Figure 5 shows a Fast Fourier Transformation of a high-frequency voltage signal.

[0033] Figure 6A contains a Mels Spectrogram of a high-frequency voltage signal of a thermal spray plasma gun at the beginning of the lifecycle of an electrode-nozzle pair.

[0034] Figure 6B contains a Mels Spectrogram of a high-frequency voltage signal of a thermal spray plasma gun at the end of the lifecycle of an electrode-nozzle pair.

[0035] Figure 7 is a flow diagram illustrating an exemplary method for predicting the remaining service life of a thermal spray plasma gun electrode-nozzle pair.

[0036] Figure 8 shows the health index (HI) score derived using a DeepSAD method from one nozzle as an example and a comparison to the end of service life label provided by an expert in the field for this nozzle.

[0037] Figure 9 depicts an Adaptive Mixture of Local Experts model as applied for predicting the remaining service life of an electrode-nozzle pair.

[0038] Figure 10 compares the remaining service life determined by the Adaptive Mixture of Local Experts model with the end of service life label provided by an expert.

[0039] Detailed description

[0040] The particulars shown herein are by way of example and for purposes of illustrative discussion of the embodiments of the present invention only and are presented in the cause of providing what is believed to be the most useful and readily understood description of the principles and conceptual aspects of the present invention. In this regard, no attempt is made to show structural details of the present invention in more detail than is necessary for the fundamental understanding of the present invention, the description taken with the drawings making apparent to those skilled in the art how the several forms of the present invention may be embodied in practice.

[0041] Fig. 1 illustrates a cross-section of a conventional thermal spray plasma gun. Cooling water (101) is supplied through a hose from a junction and monitoring box (JAMBox) not shown in Fig. 1 to a water inlet at the thermal spray plasma gun, being the positive thermal spray plasma gun connection point. The cooling water flows through a predetermined path within the thermal spray plasma gun through the nozzle (105), acting as the anode, to the electrode (115), acting as the cathode of the thermal spray plasma gun and subsequently to a thermal spray plasma gun a water outlet (119), being the negative gun connection point to a hose back to the junction and monitoring box. In addition to transporting cooling water from the JAMBox to the thermal spray plasma gun, the hoses are configured to supply electrical power to the thermal spray plasma gun. This is usually achieved by including a flexible and electrically conductive grid tube with the hoses. An arc (113) is ignited between the nozzle (105) and the electrode (115), being electrically isolated by an insulator (103) by applying a voltage. Plasma gas (121) is ionized and forms a plasma jet (111) which melts the coating material supplied (109) through injectors (107) using a carrier gas and propels the molten particles towards a substrate not shown in Fig. 1 on which the molten particles solidify rapidly and form a coating material.

[0042] In Fig. 2 a conventional thermal spray plasma system is shown. Dashed lines in Fig. 2 represent media supply lines, dotted lines represent signal cables. A plasma jet (205) is generated by a thermal spray plasma gun (205) mounted on a thermal spray plasma gun handling device such as a robot or a CNC axis not shown in Fig. 2. Coating material in the form of particles is transported from one or several powder feeders (221 ) through powder-gas hoses (219) by means of a carrier gas. The coating material is injected into the plasma jet (205) and subsequently deposited on a substrate (207) mounted on a part handling device (209). In order protect the operator and to remove dust from the spray process, the thermal spray plasma gun, the part, the part handling device and the thermal spray plasma gun handling device are enclosed in a spray booth (201). Cooling water and electrical power are supplied to the thermal spray plasma guns through hoses (213) connected to a junction and monitoring box (211 , JAMBox). The cooling water is supplied to the JAMBox from a chiller or heat exchanger not shown in Fig. 2 and flows back to the chiller or heat exchanger from the JAMBox. The electrical power needed is supplied to the JAMBox from a power supply not shown in Fig. 2. The plasma gas is supplied from a gas-management center (217) through a gas hose (215) to the thermal spray plasma gun (203). On the spray controller comprising a HMI (223), the operator sets the spray parameters and starts and stops the spray process. According to one prior art approach, the voltage and the current supplied to the thermal spray plasma gun is recorded within the JAMBox at a frequency of 1 Hz. The voltage data recorded is transmitted to the thermal spray controller (223) using a signal cable (225) from the JAMBox. If, according to this prior art approach, the voltage drops below a preset value, a warning message is displayed on the HMI and / or the thermal spray controller stops the process. This allows to stop the process in case of a short circuit created for example by a collision between the thermal spray plasma gun and the substrate and / or if the insulator between the electrode and the nozzle breaks. However, as according to this prior art approach the voltage is sampled at a frequency of 1 Hz and as the voltage measured within the JAMBox is biased by the hoses (213), the wear and tear of the electrode-nozzle pair cannot be determined reliably and a prediction of the remaining service life of the electrode-nozzle pair is not possible. The magnitude of the bias between the true voltage between the anode and the cathode within the thermal spray plasma gun and the voltage as measured within the JAMBox depends on factors such as length, number of bends and sharpness of bends of the hoses as well as on the wear of the hoses.

[0043] The thermal spray plasma spray system shown in Fig. 3 is equipped with an inventive system for determining the health state and predicting the remaining service life of a thermal spray plasma gun’s electrode-nozzle pair. A high-frequency voltage sensor (327) which samples the voltage at a frequency of at least 20 kHz and up to 100 kHz, preferably at least 20 kHz and up to 50 kHz is attached to the water in- and outlet connection points of the thermal spray plasma gun (305). The hoses (313) supplying cooling water and electrical power to the thermal spray plasma gun (305) are attached to the high-frequency voltage sensor (327). Sampling thermal spray plasma gun voltage data at a frequency considerably higher than according to the prior art approach allows to identify voltage patterns being specific for wear patterns of the electrode-nozzle pair and / or indicating impending failure. Mounting the sensor (327) directly to the water in- and outlet connection points of the thermal spray plasma gun allows to monitor the thermal spray plasma gun voltage in real-time without disrupting thermal spray plasma process and without changing the thermal spray plasma gun hardware which would lead to a bias in the thermal spray plasma gun voltage measured. The data generated by the sensor (327) is transmitted using a signal cable (329) or wirelessly to a computing device (331) comprising a high-frequency Data Acquisition System (DAQ) for digital conversion either located within the thermal spray controller or as a separate component of the thermal spray system. The voltage signals collected in real-time are stored and converted into a machine-readable format on the computing device.

[0044] Fig. 4 shows voltage signals sampled at a frequency of 50 kHz on a thermal spray system according to Fig. 3 employing an Oerlikon Metco F4 thermal spray plasma gun equipped with a nozzle with diameter 6 mm subjected to thermal spray plasma parameters of plasma gas flows of 45 nlpm Ar and 10.5 nlpm H2 utilizing a current of 650A. The sampling frequency of 50 kHz applied, corresponds to 50’000 samples of the thermal spray gun voltage collected in one second.

[0045] Fig. 5 depicts the result of a Fast Fourier Transformation of a voltage signal sampled as shown in Fig. 4. As can be seen in Fig. 5, peak amplitudes of the voltage signal are present at around 3.9 kHz, approximately 6.4 kHz and at 9.1 kHz. To asses the influence of thermal spray plasma process parameters on the frequency ranges for peak amplitudes, the inventors created a Fast Fourier Transformation of high- frequency voltages sampled whilst applying different ranges of operating current and plasma gas flows. Using these representations, the inventors were able to identify that frequencies up to 10 kHz are relevant for analyzing the wear and health state of a thermal spray gun’s electrode / nozzle pair. Based on this relevant frequency range of up to 10 kHz, it is possible to determine that it is essential to sample voltage signals at at least 20 kHz. As shown in Fig. 6A and 6B, the voltage data sampled was subsequently converted into Mel spectrograms. A Mel spectrogram is a visual representation of a signal, where the frequency scale is transformed to the Mel scale, which is designed to mimic human perception of pitch. The process of converting voltage data sampled to Mel spectrograms comprises dividing the voltage data sampled into non-overlapping time segments with a duration of 1 .3 seconds each, each equivalent to 216voltage sample points, applying a Short-Time Fourier Transform (STFT) to each time segment and subsequently mapping the output of the STFT onto the Mel scale using 64 triangular filters. This process creates a matrix where rows represent Mel bands and columns represent time frames. The formula for converting frequency (f) to the Mel scale (m) is: m = 2595 * Iog10(1 + f / 700), with f being the frequency of the signal in Hertz. Compared to the result of the Fast Fourier Transformation shown in Fig. 6, the Mel spectrograms allow to visually represent variation of a signal’s amplitude over time at different frequencies. Based on Fig, 6A, showing a Mel spectrogram of a thermal spray plasma gun’s electrode-nozzle pair at the beginning of its lifecycle, it was identified that the Helmholtz frequency around 3 kHz and a frequency band ranging from 5 kHz to 6 kHz are characteristic for the electrode-nozzle pair. The highest amplitude at a frequency of 0 Hz represents the operating direct current. As shown in Fig. 6B, compared to Fig. 6A, the amplitudes of different frequency ranges changed during the lifetime of the electrode-nozzle pair.

[0046] Fig. 7 illustrates an exemplary method of predicting the remaining service life of a thermal spray plasma gun electrode-nozzle pair. In a first step (701), a predefined number of electrode-nozzle pairs is subjected to coating process parameters which are set based on the type of thermal spray plasma gun used. During the spray process, voltage is sampled at a frequency of 20 kHz in real-time (703) at a thermal spray plasma gun using a high-frequency voltage sensor. The data sampled is stored on a computing device and converted into Mel spectrograms (705). The process of converting voltage data sampled to Mel spectrograms further comprises dividing the voltage data sampled into non-overlapping time segments with a duration of 1.3 seconds each, each equivalent to 216voltage sample points, applying a Short-Time Fourier Transfrom (STFT) to each time segment and subsequently mapping the output of the STFT onto the Mel scale using 64 triangular filters. During the coating process, the health state of the electrode-nozzles pairs is assessed by an expert who possesses the ability to identify the commencement of wear and tear on the electrode-nozzle pair, and abnormal samples are identified based on this assessment (707). The datasets generated are then used to train the DeepSAD model to determine a health index and the Mixture of Local Experts model to determine a remaining service life index which is then used to determine the remaining service life in spray hours. The models trained on a specific thermal spray plasma system using a specific thermal spray plasma gun are then transferred to other thermal spray plasma systems where this specific thermal spray plasma gun type is used (711). During the spray process on these thermal spray plasma systems, high-frequency voltage signals provided by a sensor, preferably mounted to the thermal spray plasma gun, are recorded on a computing device, and converted to Mel spectrograms (713), the process of converting voltage data sampled to Mel spectrograms comprising dividing the voltage data sampled into nonoverlapping time segments with a duration of 1 .3 seconds each, each equivalent to 216voltage sample points, applying a Short-Time Fourier Transfrom (STFT) to each time segment and subsequently mapping the output of the STFT onto the Mel scale using 64 triangular filters. Subsequently, a health index ranging from 0 to 1 , with 0 meaning that the hardware is perfectly healthy and 1 meaning that the hardware is completely worn and has to be replaced, is determined based on the DeepSAD method (715). The dataset generated in step (713) is also used to the determine the remaining service life based on a Mixture of Local Experts method (717). The Mixture of Local Experts applied determines a remaining service life index ranging from 0 to 1 , 0 meaning the 100% of the service life remain and 1 meaning that 0% of the service life remain. Based on the remaining service life index and the spray hours accumulated for the electrodenozzle pair mounted to the thermal spray plasma gun, the remaining service life in spray hours is calculated in step (719). The remaining service life in spray hours is displayed on the HMI of the thermal spray controller and the thermal spray process. Once the remaining service life is approximately 1 spray hour, the computing device generates a signal that is used by the thermal spray controller to prevent ignition of the thermal spray process and / or to shut down the thermal spray process if the coating process for the part being coated is completed (721). Subsequently, the HMI displays a message that the electrode-nozzle pair must be replaced. This message is then reset by the operator after the electrode-nozzle pair has been replaced (723).

[0047] In Fig. 8 a comparison of the end of service life labels provided on one hand by an expert and on the other hand based on a health index (HI) calculated based on the DeepSAD method for an electrode-nozzle pair of an Oerlikon Metco F4 thermal spray plasma gun equipped with a nozzle with diameter 6mm subjected to thermal spray plasma parameters of aplasma gas flow of 45 nlpm Ar and 10.5 nlpm H2 utilizing a current of 650A is shown. For semi-supervised methods of estimating a health index, a training dataset is required. To generate a training dataset for the DeepSAD method, the inventors mounted 14 electrode-nozzle pairs to an Oerlikon Metco F4 thermal spray plasma gun and subjected them to a plasma gas flow of 45 nlpm Ar and 10.5 Ipm H2 and a current of 650A. During the entire service life of the electrode-nozzle pairs used to generate a training dataset, the thermal spray plasma gun voltage was sampled at 50 kHz and converted to Mel spectrograms. During the generation of the training dataset, an expert assessed the wear and tear of the electrode-nozzle pairs and determinated abnormal labels for the electrode-nozzle pairs employed. The generation of the training dataset covered operation periods ranging from 15 to 45 spray hours until each electrode-nozzle pair employed attained a state of significant wear. This dataset was then used as training dataset for the DeepSAD method as indicated below:

[0048] X = {xv..., xN1, xN1+1, xN}

[0049] N = Nj + Nu

[0050] N represents the total number of samples, Ni the number of labeled samples and Nuthe number of unlabeled samples. The labels are denoted by

[0051] I £ {1. -1 } with a value of 1 assigned for samples that are a realization of a healthy system, and a value of -1 assigned when the sample represents a realization of a system with a severe fault. Samples in between are then unlabeled. The inventors used the dataset from the first 5 hours of operation of the seven training nozzles and of the test nozzle as benchmark for healthy labels. The time period flagged by the expert as the onset of abnormalities of the 7 training nozzles served as abnormal samples. To ensure the integrity and relevance of the training data the data corresponding to the end of service life of the current nozzle was excluded from the training. The method applied by the inventors extends the application of DeepSAD in the domain of estimating a health index (HI). Based on this training dataset, the DeepSAD method was applied to provide a health index ranging from 0 to 1.0 with 0 meaning that the electrode-nozzle pair is perfectly healthy, without wear and 1 meaning that the hardware is completely worn and has reached the end of service life.

[0052] The DeepSAD method aims to discover a transformation (|)e using a neural network with weights 9 to effectively separate healthy and unlabeled samples from the abnormal ones. The primary objective of the of the DeepSAD method is to minimize the volume of a hypersphere centered at a, encompassing healthy samples, while ensuring that abnormal samples lie outside this hypersphere. The DeepSAD loss function as LDS(X; 6), and the parameters 6 are determined by minimizing this loss function. It can be expressed as follows:

[0053] The parameter p serves as a hyperparameter that determines the extent to which unlabeled samples are incorporated within the hypersphere that encompasses healthy samples. In contrast, v is a crucial hyperparameter that regularizes the neural network’s weights, preventing overfitting.

[0054] Instead of directly using the embedding generated by the DeepSAD model as HI, the inventors propose to use the embedding generated by DeepSAD as a condition indicator that needs to be integrated to construct the HI. The condition indicator derived from the DeepSAD embedding is incorporated using a diversity loss function. With

[0055] C = (YTY) as the Gram matrix of the DeepSAD embeddings, the proposed diversity regularization can be expresses as:

[0056] LPDiversity(C) = -ln(det(C)) + trace(C)

[0057] The revised loss function incorporating a diversity regularization into the DeepSAD model can be expressed as: argmill £DS(X; 0) + A^Diversitv (C)

[0058] & with C being the Gram matrix of the DeepSAD embeddings Y and A a hyperparameter related to diversity regularization.

[0059] In addition to this, the inventors introduce a novel approach to HI estimation through feature fusion, employing an isotonic alternating projection algorithm. This concept involves projecting an index into both the input feature subspace and the space representing the ideal health index.

[0060] To predict the remaining useful life the electrode / nozzle pair of the thermal spray plasma gun an Adaptive Mixture of Local Experts model as shown in Fig. 9 is applied. This model can leverage specialized models referred to as experts. For this purpose, the complexities associated with domain generalization in regression tasks were investigated. The training phase involved handling a set of nslabeled instances from a specific source domain.

[0061] The mixture of experts can be formulated as: where n indicates the number of experts, and g(x)i is the i-th logit of the output of the gate g(x), which indicates the probability for the expert. More specifically, the gating network g produces a distribution over the n experts based on the input, and the final output is a weighted sum of the outputs of all experts. And both f and g are learnable neural networks. The output of the model is a number, referred to as remaining service life index, between 0 and 1 , where 1 means that 0% lifetime remained and 0 means 100% lifetime remained. Fig. 10 shows the prediction of the remaining service life for one of the nozzles as an example. There is a very good match between model prediction and the expert label. Subsequently, based on the number of spray hours accumulated at a specific remaining service life index, the remaining number of spray hours until the service life index is predicted to be 1 is calculated. The remaining number of spray hours is displayed on the HMI of the thermal spray controller. In addition to this, if the remaining service life in spray hours is equal to 1 , the computing device sends a signal to the thermal spray controller. This signal is used by the thermal spray controller such that startup of the thermal spray process is prevented and / or shut down the thermal spray process after completion of the spray process for a part is initiated. This signal can be reset by the operator once the electrode-nozzle pair is replaced.

[0062] It is preferred that the inventive solutions are further improved by equipping the nozzle and / or electrode with an RFID chip that enables resetting the process stop flag provided by thermal spray controller is reset only if the worn nozzle has been replaced.

[0063] It is noted that the foregoing examples have been provided merely for the purpose of explanation and are in no way to be construed as limiting of the present invention. While the present invention has been described with reference to an exemplary embodiment, it is understood that the words which have been used herein are words of description and illustration, rather than words of limitation. Changes may be made, within the purview of the appended claims, as presently stated and as amended, without departing from the scope and spirit of the present invention in its aspects. Although the present invention has been described herein with reference to particular means, materials and embodiments, the present invention is not intended to be limited to the particulars disclosed herein; rather, the present invention extends to all functionally equivalent structures, methods and uses, such as are within the scope of the appended claims.

[0064] List of reference signs

[0065] APS Atmospheric Plasma Spray

[0066] DeepSAD Deep Semi-supervised Anomaly Detection HF High Frequency

[0067] HMI Human Machine Interface

[0068] JAMBox Junction and Monitoring Box

[0069] RSL Remaining Service Life

[0070] STFT Short-Time Fourier Transformation

Claims

1. Claims1 . A system to determine the health state of a thermal spray plasma gun electrode-nozzle pair and to predict the remaining service life of the said electrode-nozzle pair, the system comprising- at least one high-frequency voltage sensor (327);- at least one signal transmitter configured to transmit the voltage measurements to a- a computing device (331 );- a thermal spray controller comprising an HMI (323) and- at least one thermal spray plasma gun (303) comprising;- at least one electrode-nozzle pair; characterized in that the sampling frequency of the at least one high- frequency voltage sensor (327) for the thermal spray plasma gun voltage is at least 20 kHz.

2. A system according to claim 1 characterized in that the high-frequency voltage sensor (327) is directly mounted to the thermal spray plasma gun (303) in between the thermal spray plasma gun’s water in- and outlet and the water hoses connected to the JAMBox.

3. A system according to claims 1 and 2 characterized in that the at least one high-frequency voltage sensor (327) does not intrude into the thermal spray plasma gun (303).

4. A system according to claim 1 characterized in that the computing device (331 ) is configured to convert the high-frequency voltage signals into Mel spectrograms.

5. A system according to claim 5 characterized in that the computing device (331 ) is configured to determine a health index of the at least one electrodenozzle pair using a DeepSAD method.

6. A system according to claim 6 characterized in that the computing device (331 ) is configured to determine the remaining service life of the at least one electrode-nozzle pair using an Adaptive Mixture of Local Experts method.

7. A system according to claim 1 characterized in that the thermal spray controller (323) is configured to display the remaining service life of the at least one electrode-nozzle pair of the least one thermal spray plasma gun.

8. A system according to claim 8 characterized in that the thermal spray controller (323) is configured to prevent ignition of the thermal spray process if the remaining service life of the at least one electrode-nozzle pair of the at least one thermal spray plasma gun is approximately 1 spray hour.

9. A system according to claim 9 characterized in that the thermal spray controller (323) is configured that the signal to prevent ignition of the thermal spray process can only be reset if the at least one electrode-nozzle pair of the at least one thermal spray gun has been replaced by a new electrode-nozzle pair.

10. Method to determine the health state and to predict the remaining service life of the at least one electrode-nozzle pair of a thermal spray plasma gun, the method comprising the steps of:- sampling the voltage of the thermal spray plasma gun;- determining a health index for the at least one electrode-nozzle pair;- determining the remaining service life of the at least one electrode-nozzle pair and- controlling a thermal spray controller using the remaining service life of the at least one electrode-nozzle pair characterized in that the voltage of the thermal spray plasma gun is sampledat a frequency of at least 20 kHz.11 . Method according to claim 11 characterized in that the voltage signal is converted to Mel spectrograms.

12. Method according to claim 12 characterized in that a health index is determined for the at least one electrode-nozzle pair using a DeepSAD method.

13. Method according to claim 11 characterized in that the remaining service life of the at least one electrode-nozzle pair is determined using an Adaptive Mixture of Local Experts method.

14. Method according to claim 11 characterized in that the thermal spray controller prevents ignition of the thermal spray plasma process if the remaining service life of the at least one electrode-nozzle pair is about one spray hour.

15. Method according to claim 15 characterized in that the signal preventing ignition of the thermal spray plasma process can only be reset on the thermal spray controller once the electrode-nozzle pair has been replaced.

16. Method according to claim 11 characterized in that a set of at least 7 distinct electrode-nozzle pairs are subjected to a thermal spray plasma process with predefined parameters for a specific setup of a specific thermal spray plasma gun and the thermal spray plasma gun voltage is sampled and recorded.

17. Method according to claim 17 characterized in that recordings of the thermal spray plasma gun voltage are used to generate a training dataset for the health index and remaining service life index determination.

Images