Self-learning method for condition monitoring and predictive maintenance in drive systems

Abstract

Method for condition monitoring of a drive system consisting of at least one motor and at least one converter, wherein data from the drive, from the system in which the drive is used and / or from the environment of the system are collected and stored, characterized in that a learning machine (20) learns to predict future values ​​of quantities relevant for assessing the condition of the drive or the system from the stored data, wherein separate predictions for the thermal hazard of the converter and the motor are generated from the predicted values ​​of said quantities, and the switching frequency of the converter is influenced depending on these predictions.

Description

Field of the invention

[0001] The invention relates to a method for condition monitoring and predictive maintenance of an electric drive system and to a converter that uses this method. State of the art

[0002] The failure of an electrical drive in a machine or production facility can result in significant economic damage. Therefore, it is important to prevent such failures or at least issue a timely warning.

[0003] EP 1049 050 A2 describes a condition monitoring system for drive systems. A microsystem continuously records and stores certain operating parameters, which are used to evaluate the operating status. The operating status is classified into different classes, such as "normal," "pre-warning," and "alarm." If the classification changes, a corresponding electronic, optical, or acoustic signal is generated.

[0004] DE 10 2016 000 368 A1 describes a temperature estimation device for an electric motor. A temperature sensor is attached to the motor, and the detected temperature values ​​are stored in a memory device. The document discloses a simple method for estimating the motor temperature without having to consider the losses in the electric motor.

[0005] DE 10 2007 040 423 A1 describes a method for operating an inverter-fed electric motor with an electromagnetically actuated brake. In motor mode, motor currents and motor voltages are recorded. The current temperature is determined from these values, particularly taking the motor resistance into account. In generator mode, however, the temperature is determined from a mathematical model.

[0006] The failure of the motor or inverter in a drive system is often caused by an unacceptably high increase in temperature.

[0007] Often, the temperature in both the motor and the converter is monitored. Once the critical temperature limit is reached, the drive is shut down. This protects the drive itself from overheating. However, the system remains stationary in such a case, which can already cause significant damage. However, an early thermal failure of the drive cannot be completely ruled out, as both the still-tolerated overtemperature and the rapid temperature changes can cause faults in various locations, such as the solder joints, the winding insulation, or the power semiconductors.

[0008] To avoid system shutdowns, it is advisable to anticipate the drive's load. In the case of stationary or quasi-stationary operation, predicting the operating temperature using a thermal model is usually manageable with manageable effort. However, in other operating modes, predicting future temperature trends is considerably more difficult.

[0009] If the exact process sequence in a system in which the drive is used is known, a more accurate temperature forecast at critical points, such as in power semiconductors or motor insulation, would be feasible. However, this knowledge would have to be integrated into a complex temperature model. This work is not only time-consuming but also prone to errors. Testing such models is also very critical. And the most significant drawback: Even a simple change in the work plan or, for example, a format change in a production system could result in the calculation model having to be significantly revised, or in the quality of the forecast deteriorating significantly without a revision of the calculation model.

[0010] Therefore, it is proposed to use self-learning systems that recognize patterns in the process flow and use the detected patterns to predict the condition. This allows the drive to adapt to the current application area, even without prior knowledge of this application area. Subject of the invention

[0011] The aim of the invention is to use self-learning systems in the monitoring of drive systems that can predict critical states in the drives even during non-stationary operating modes. In particular, a method is proposed that can reduce the thermal hazard of the system. This is achieved by appropriately changing the switching frequency in the drive system's converter. Description of the invention

[0012] Learning machines are known to be capable of recognizing various, even complex, patterns in signals. If signals from many similar systems, some of which have failed, are evaluated, it may be possible to automatically identify the causes of the failures. This knowledge could then be used for prognosis and possibly for the prevention of future failures. However, this approach requires that the signals from the systems are available worldwide, which is often not the case. For reasons such as the protection of trade secrets (process information, economic data, warranty-relevant data, etc.), neither the machine or system manufacturers nor the operators of such systems are inclined to release their data. Therefore, a method is proposed that does not require global networking.

[0013] Industrial drives are usually very reliable. Therefore, a large number of failures at a single location cannot be expected. This means that local data from one or a few plants is generally insufficient for analyzing failures and identifying regularities. Especially if a plant has never experienced a failure, the learning machine cannot learn from the local data how to predict a failure.

[0014] Therefore, it is proposed that the learning machine should not attempt to directly predict failures, but rather forecast future values ​​of variables relevant for assessing the drive's vulnerability, especially temperature values. More local data is available for learning this. These forecasts can detect many potential failures early on.

[0015] For a drive that neither operates at a steady-state operating point (constant speed and constant torque) nor at a quasi-steady-state operating point (e.g., strictly cyclical operation with short cycle times), temperature prediction is significantly more difficult. Here, it is helpful if the learning machine can also predict future loads (speed / torque). Similarly, it is helpful if the learning machine can predict future heat dissipation. However, these values ​​depend (along with other information such as location, time of day, season, weather, etc.) on the actual process in the system. The learning machine must therefore be able to independently recognize patterns in the process flow.

[0016] A further difficulty arises from the fact that the learned patterns are not universally valid indefinitely, but can quickly become obsolete. For example, if the shift schedule in the company changes, if different or modified products are produced with the system, or if the cooling conditions in the production hall are changed, the forecasts become worse or even completely worthless. Therefore, the learning machine must periodically or continuously review its forecast quality. If insufficient forecast quality is detected, the learning machine must repeat at least part of the learning process until the expected forecast quality is achieved again.

[0017] To obtain training data for the learning process and control data for monitoring the forecast quality, the learning machine is tasked with creating forecasts for temperature values ​​at locations in the drive or system. The current temperature values ​​at these locations must be detectable directly (via temperature sensors) or indirectly (through calculations from known variables) so that they can be used for the learning process. These values ​​and other detectable variables from the system and its environment are collected automatically and, according to the invention, serve as training and test data for the learning process. Suitable learning algorithms can be used for this purpose.

[0018] Since older forecasts can be compared with later actual values, algorithms from the group of methods known as "supervised learning" can be used advantageously. Both regression methods that predict temperature values ​​(e.g., in °C) and classification methods that divide future temperatures into temperature classes can be used.

[0019] If the learning machine can predict these temperature values ​​with sufficiently high quality, this prediction can be used to predict machine conditions, especially failures.

[0020] For a particularly reliable prediction of failures, additional thermal models can also be used so that temperature values ​​can be predicted even at points that cannot be measured directly or indirectly during operation, but which are of great importance for preventing failures.

[0021] For example, the learning machine can be trained to predict the temperature at the heat sink. The learning process requires the actual temperature values. However, these are usually available thanks to appropriate temperature sensors. The actual temperature of the semiconductor material of the power semiconductors ("junction temperature") or other electronic components, such as microcontrollers, is actually the decisive factor for a failure. These temperature values ​​can then be determined based on the learning machine's prediction of the heat sink temperature, incorporating a thermal model.

[0022] Furthermore, statistical, temperature-dependent failure models can be advantageously used to predict the probability of failure. For example, the recorded past temperature values ​​(preload) and the predicted future temperature values ​​for an insulation material, such as the stator insulation of a motor, can be used to determine the future failure probability of the part. However, a key element of the method according to the invention remains the process-dependent temperature prediction by a learning machine.

[0023] A key aspect of the invention is that the learning machine's forecast values ​​are stored for a specific period of time so that they can be compared with the actual temperature values ​​that later occur. This makes it possible to determine the quality of the forecast during operation. Without this control, the learning machine fails to recognize that changes have occurred in the system and that the modified processes must be relearned to produce a high-quality forecast.

[0024] Another option is the use of so-called online algorithms, such as the "stochastic" or "mini-batch" learning method. Here, training data continues to be collected during normal operation. The recorded actual temperature values ​​are used to label the data, which are then compared with the previously stored forecast values ​​during the learning process. Using such learning methods, the learning machine adapts its forecast virtually continuously ("stochastic") or periodically ("mini-batch") to potentially changing conditions. Another advantage of such methods is that they generally use the necessary resources more efficiently.

[0025] Monitoring forecast quality is also useful in this case, so that persistently poor forecast quality can be detected and reported. Otherwise, the plant operator may be under the illusion that the plant is being effectively monitored for future failures, even though this cannot be guaranteed due to the poor forecast quality.

[0026] The learning machine can use a well-known learning algorithm, particularly from the group of supervised learning, such as decision tree learning, linear regression, logistic regression, Winnow, LASSO, ridge regression, ARIMA, perceptron, artificial neural networks, deep learning, Naive Bayes, Bayesian network, support vector machine, Markov chain, hidden Markov model, or boosting. Most algorithms also have so-called kernelized versions, which also allow the implementation of non-linear regressors and classifiers.

[0027] Since, in a widely used drive system, it is not known in advance which type of system the specific drive will be used in, and therefore how the processes will work, it is also impossible to know with certainty which learning algorithm is most suitable for a specific application. Therefore, it may be useful for the learning machine to independently test and compare several learning algorithms. In this comparison, both the achieved forecast quality and the resources required (required storage capacity and computing power) can be crucial. It may therefore be sensible to choose a resource-efficient method with sufficient forecast quality, rather than an algorithm that delivers a slightly better forecast but requires significantly more resources.

[0028] In addition to selecting the learning algorithm to be used, an important decision is which variables (reference variables, measurement signals, etc.) from the available signals are used in the learning process. With the increasing networking of components in automation technology, particularly due to trends such as "Industry 4.0," "Internet of Things," etc., the amount of accessible information is growing. However, too many variables lead to an increase in dimensionality in the learning process ("curse of dimensionality"). This leads to an excessive increase in the time required for the learning process and / or the time required to create a forecast, an unacceptable increase in the required storage capacity, or a failure of the learning algorithm to converge. Therefore, it is usually necessary to limit the number of variables used in the learning process.

[0029] A key problem is that the application of the drive is often unknown in advance. Therefore, the learning machine must be able to make this selection independently.

[0030] Sometimes, simple correlation calculations can be used to determine whether certain signals correlate with the values ​​to be predicted (future temperature values) and are thus suitable for forecasting. For example, the cross-correlation function or the cross-power spectrum must be calculated and evaluated, as is common in signal analysis.

[0031] However, it is often the case that signals do not directly correlate with future temperature values, but important information for the forecast can be obtained from the combination of several variables. For such cases, several well-known methods for reducing dimensionality are available, such as - Principal Components Analysis (PCA) - Independent Components Analysis (ICA), or - Random Components Analysis (RCA).

[0032] Another way to limit the number of variables used in the forecast is to use L1 regularization, which is well-known in the literature. This often results in some variables used in the learning process no longer being used in the forecast itself.

[0033] It can often be helpful to consider the variables used for learning and prediction in the frequency domain rather than the time domain. To do this, the time course of the variable can be transformed into the frequency domain using a Fourier transform, such as the fast Fourier transform (FFT). The spectral components can then be analyzed in the frequency domain. Another transformation that can be used for machine learning is the Hilbert-Huang transform, which has been proven to improve and simplify the learning process in some cases.

[0034] Since the application of the drive is often not known in advance, it is advantageous if the learning machine independently tries out several methods for reducing dimensionality and for signal transformation, compares them, and selects and later uses the most suitable method for the specific application.

[0035] Regarding the resources (storage capacity and computing power) required for selecting the parameters used, executing the learning algorithm, and calculating the forecast, there are several options for the physical implementation. The learning machine can be implemented entirely or partially in software using intelligent components already present in the drive, such as microcontrollers (µCs), programmable logic devices (FPGAs), or digital signal processors (DSPs). However, resources from controllers (e.g., PLCs or IPCs) found in drive subsystems can also be incorporated. Another option is the use of additional components, such as ASICs, FPGAs, CPUs, GPUs, and TPUs, to enable, improve, and accelerate the learning process and / or forecast calculation. Finally, the use of plant-internal, company-internal, or external cloud services can also be useful.

[0036] Using a cloud system offers another advantage: It enables the manufacturer to monitor the learning process of many different applications and optimize it manually or automatically.

[0037] Another important aspect of the invention is what should happen when a future threat to the drive system is detected. To prevent an imminent failure, an automatic shutdown is a possibility. However, if there is no acute, short-term threat, a warning may be sufficient. This can then lead to an inspection of the system, the drive, or the cooling system. Another option is manual or automatic load reduction, for example, by reducing the production speed in a production facility.

[0038] Another option is available specifically for drive systems: influencing the converter's switching frequency. Solutions are known in which the converter's switching frequency is reduced when the temperature of the power electronics exceeds a certain limit. However, the process-dependent temperature forecast according to the invention also allows for predictive control of the switching frequency.

[0039] If the learning machine is able to predict temperature values ​​in both the motor and the inverter, even balancing different needs can be done automatically.

[0040] A reduced switching frequency reduces the load on the power electronic switches (e.g., MOSFETs or IGBTs) in the converter, as switching losses can be reduced proportionally with the switching frequency. This also reduces the risk of temperature-related failure of the power semiconductors.

[0041] However, the lower switching frequency means increased ripple in the motor current. This leads not only to increased acoustic noise but also to additional losses in the DC link capacitors, in the cabling or busbars, and even in the motor itself. In the latter case, the higher ripple currents can cause increased copper losses in the winding, significantly higher ohmic losses (exacerbated by current displacement) in the squirrel cage of an asynchronous motor, and generally increased magnetization and eddy current losses due to higher induction ripple in the soft magnetic parts of the motor, such as the stator and rotor laminations.

[0042] However, the time component must be considered, as the time constants of the motor and the power electronic switches are typically very different. The thermal time constant of the motor is usually several times that of the power switch. Therefore, the component most likely to fail must be considered first.

[0043] The invention therefore proposes automatically evaluating temperature forecasts in the converter and motor. If a temperature-related failure of circuit breakers is imminent, the switching frequency is reduced. Conversely, if a thermal failure of the motor is imminent, the switching frequency is increased. Description of the drawings

[0044] Fig. Figure 1 shows a simplified, exemplary external signal sequence of the learning machine according to the invention. X1, X2, and X3 are the input signals that can be used for the forecast. Y1 is the quantity to be forecast. The dashed line (5) represents the point in time at which the values ​​of all quantities for the elapsed time T1 are available in memory. The task of the learning machine is to use these quantities to predict the future value of the output quantity, in this case Y1, represented as point 10. Point 10 is located at a time distance of T2 from the end of the recordings for X1, X2, and X3. Since the quantity Y1 is fed to the learning machine, after the elapse of time T2, it can compare its forecast with the actual value that occurred at this later point in time. This enables the learning machine to learn the forecast and monitor the forecast quality.Depending on the learning algorithm used, it can continuously improve the forecast quality.

[0045] Fig. Figure 2 shows an exemplary combination of the learning machine (20) and a machine model (40), in particular a thermal model. The learning machine calculates the future values ​​(Y1*, Y2*, Y3*) of the directly or indirectly measurable variables (Y1...Y3). For this purpose, it can use the variables X1...Xn that are also measurable for the system.

[0046] For monitoring the power switches in the converter, for example, it may be advisable to predict the current and voltage of the power switches (e.g. IGBTs) as well as the temperature of the heat sink. Since both the input and output variables for the system are Fig. 1 are available for the learning process, the learning machine can learn the prediction using this training data.

[0047] However, for assessing the risk to electronics, non-directly measurable temperature values, such as the internal temperature in the semiconductor material of the circuit breaker (“junction temperature”), may be more suitable. One way to predict such values ​​is Fig. 2. From the forecasts for the converter currents and voltages (Y1* and Y2*) calculated by the learning machine (20), the calculation block (30) calculates the future power loss in the circuit breaker (Z1). From this power loss (Z1) and the forecasts for the heat sink temperature (Y3*) provided by the learning machine (20), the thermal model creates the forecast for the non-directly measurable quantity W1, in this example, the blocking temperature of the electronic circuit breaker. This value is then used in the drive system to assess the thermal hazard.

[0048] Fig. Figure 3 shows the decision scheme for the switching frequency of the converter in a hazardous situation. 110: Is there a risk of thermal overload in the drive system? 120: Is a thermal hazard in the electronics to be expected first? 130: Leave switching frequency unchanged 140: Increase switching frequency 150: Reduce switching frequency

[0049] First, a check is made to determine whether a hazard can be predicted (110). If the answer is negative, the switching frequency is not changed (default operation 130). If a hazard is detected, the system checks whether the power electronics or the motor is at risk of failing sooner (120). If an impending thermal hazard to the power electronics is detected, the switching frequency is lowered (150) to reduce switching losses in the power switches and thus relieve their thermal load. However, if the power electronics still have reserve, the switching frequency is increased (140) to reduce the current ripple and thus relieve the motor's thermal load. Exemplary embodiment of the invention

[0050] In an exemplary implementation of the invention, the drive system is connected to the environment via at least one communication channel, such as OPC UA (Open Platform Communications Unified Architecture). The drive system receives information about command variables, state variables, and measured values ​​within and beyond the system via this channel. The drive system can thus receive information about other drives within the system (e.g., speed, torque), sensor signals (e.g., motion sensors, temperature sensors, etc.), and external information (e.g., weather reports and forecasts). The drive system can evaluate all of these variables for its learning process and for subsequent forecasts. These input variables are sampled by the drive system and stored for the learning phase.

[0051] In a supervised learning process, not only the input variables but also the corresponding output values ​​are available during the learning phase. For example, in order to predict a future temperature value, the temperature values ​​that actually occur later must be assigned to the stored input variables. While the input variables are sampled and stored, the future output variable is of course not yet known. A key feature of the process is the time to which the forecast lies in the future. After this time has elapsed following the end of the recording period, the variable to be predicted is already reality and can be recorded. According to the invention, this value is recorded by the system, stored and assigned to the already stored sequence of input variables.

[0052] In the classic "supervised learning" method, a collection of data ("training examples") is provided to the learning machine. These examples are previously labeled by the "trainer" ("knowledgeable external supervisor"). These labels can represent continuous values ​​("regression") or classifications into categories ("classifier"). In the first case, the learning machine learns, for example, to predict the temperature at a specific location. In the second case, the learning machine sorts the temperature values ​​into predefined groups.

[0053] A learning machine usually has the task of deriving as general rules as possible from the training data provided by the trainer so that it can calculate the desired output values ​​from future, as yet unseen data.

[0054] However, the inventive solution does not require an external trainer. The drive system must therefore collect the input data and the associated output values ​​itself as training data. However, this method requires that the drive system be able to acquire the output data to be predicted at a later point in time. The simplest approach is to use variables that can be directly measured by sensors. For example, in an inverter, the heat sink temperature is often monitored directly with a temperature sensor. This signal is therefore suitable for the data collection described above. Likewise, the stator temperature can be monitored with little effort using a temperature sensor, e.g., a thermistor.

[0055] There are also other variables that are typically measured in a drive and are therefore suitable for learning. Examples include voltage, current, and speed values. Since the heating in both the circuit breakers and the motor primarily depends on the current load, the motor current is particularly suitable for assessing the thermal hazard of the system.

[0056] Additionally, the modulation level of the modulator in the inverter can be calculated from the motor voltage. This value is also directly available in the drive control system, meaning it can be integrated directly into the learning process by the learning machine using the method described above.

[0057] If the duty cycle (i.e., PWM ratio) and load current are available, the power switch losses can be calculated. If both the future losses and the future heat sink temperature values ​​are predicted by the learning machine, the critical temperatures in the semiconductor material for failure can be calculated. A relatively simple thermal model is sufficient for this.

[0058] The future temperature values ​​in the semiconductor material calculated in this way could not otherwise be directly measured or otherwise recorded. Therefore, they are not suitable for the learning process. However, the combination of the learning machine and the thermal model described above enables their prediction and thus their use in hazard analysis.

[0059] In an electric motor, temperature values ​​can be measured directly in the slot. This temperature can be used directly to determine the failure probability of the motor insulation. However, sometimes the temperature sensor is placed in the winding head for design reasons. In such a case, the winding temperature in the slot can be calculated using a thermal motor model based on the predicted temperature and load values.

[0060] Once the collected training data and the corresponding values ​​to be predicted are available, the actual learning process can begin. Several learning methods known from the literature are available for this purpose. Ready-made software libraries such as scikit-learn or Tensorflow can also be used for implementation.

[0061] Since it may be unclear in advance which learning methods are best suited for prediction, it is useful for the learning machine to autonomously try out and compare several learning methods for the prediction. Most methods have free parameters (so-called "hyperparameters") that can significantly influence the learning process. One such parameter, for example, determines the learning rate. A learning rate that is too slow increases the time required for the learning process. A learning rate that is too fast, on the other hand, can lead to the method not converging. Therefore, it is important to find the right learning rate. Some methods have several selectable parameters. Neural networks, for example, are characterized by the number of hidden layers and by the number of neurons in the individual layers. The learning machine can systematically try out these hyperparameters independently and determine the values ​​that will be used later.Several methods are known from the literature for this purpose, such as “grid search” and “random search”.

[0062] If the learning machine uses mathematical models (e.g., one or more thermal models) that themselves have adjustable parameters, the system can also autonomously determine the model parameters from the observed measured variables and also adjust them during operation. (For asynchronous three-phase motors, see, for example, G. Pfaff, H. Segerer, A. Lelkes: Resistance corrected and time-discrete calculation of rotor flux in induction motors, EPE 1989)

[0063] In an advantageous embodiment of the invention, the response to a detected thermal hazard in the system can be selected and configured from several alternatives. In addition to warnings via the communication channel or other means, such as visual or acoustic warnings, shutting down the drive or at least limiting its operating range (speed / torque limitation) may also be appropriate under certain circumstances. For highly dynamic drives, a temporary limitation of the dynamics (dn / dt) may also be considered to relieve the load on the drive system.

[0064] In some cases, the drive system is able to influence its internal cooling. For example, in air cooling, variable-speed fans are used, and the drive system can specify the fan speed. Another option for directly influencing cooling performance arises with liquid cooling (e.g., oil or water cooling), when the drive system can control the pump performance directly (via the pump speed) or indirectly (e.g., via the position of a throttle valve). In such cases, a forecast of a future thermal hazard can be used to avoid or at least reduce the potential hazard by increasing the cooling capacity in a timely manner. It is important that the cooling capacity is increased before the current temperature values ​​exceed specified limits, since the learning machine can forecast future temperature values.

[0065] If the cooling capacity is specified externally, the drive system can use its communication capabilities and request an increase in the cooling capacity via a communication channel.

[0066] Another possibility is to directly influence the switching frequency in the converter according to Fig. 3. Here, the switching frequency is independently controlled by the motor and inverter depending on the thermal hazard: increased or decreased. The switching frequency of the inverter can change the level of power loss in the motor and inverter. In this way, the drive system can mitigate the hazardous situation using predictions from the learning machine.

Claims

[1] Method for monitoring the condition of a drive system consisting of at least one motor and at least one converter, whereby data from the drive, from the system in which the drive is used and / or from the environment of the system are collected and stored, characterized by that a learning machine (20) learns to predict future values ​​of variables from the stored data which are relevant for assessing the condition of the drive or the system, wherein separate forecasts for the thermal hazard of the converter and the motor are created from the forecast values ​​of said variables, and the switching frequency of the converter is influenced depending on these forecasts. [2] Method according to claim 1, characterized by , that the switching frequency of the converter is reduced (150) if a thermal hazard in the converter is first expected, and that the switching frequency of the inverter is increased (140) if a thermal hazard to the motor is first expected. [3] Method according to claim 1 or 2, characterized by that the learning machine (20) is combined with a machine model (40), in particular with a thermal model, so that even variables not directly recorded can be predicted and used for the condition assessment, [4] Method according to one of the preceding claims, characterized by that a visual warning, an audible warning and / or a warning via a communication channel is issued if an impending thermal hazard to the system is detected. [5] Drive system consisting of at least one motor and at least one converter - with at least one communication channel with which data can be transmitted from the system in which the drive is used, - with at least one means for sampling and storing the data transmitted by the communication channel as well as data from the inverter and the motor, - with a learning machine (20) which learns from these data to predict variables which are relevant for assessing the hazard of the installation, in particular temperature values, and which creates separate forecasts from these predicted variables for assessing the hazard of the converter and the motor, - whereby the switching frequency of the converter is influenced depending on these forecasts.

Images