Monitoring methods, computer program product, monitoring unit, gas analysis device, and use of artificial intelligence

Abstract

The invention relates to monitoring methods (200) for a system (10) having a plurality of devices (12) which are designed each to provide an associated measured value (15) and / or a control command (17). The monitoring methods (100, 200) are based on the use of a processing device (40) or of a neural network (80), with which measured values (15) and / or control commands (17) are evaluated. The invention also relates to a computer program product (70) with which the monitoring methods (100, 200) are able to be executed, and to a correspondingly equipped monitoring unit (75). The invention also relates to a gas analysis device (11) which is equipped with such a monitoring unit (75). The invention further relates to the use of artificial intelligence (99) for monitoring a gas analysis device (11).

Description

[0001] Description

[0002] Monitoring method, computer program product, monitoring unit, gas analysis device and use of artificial intelligence

[0003] The invention relates to a monitoring method for monitoring a system and a computer program product designed therefor. Furthermore, the invention relates to a monitoring unit having a corresponding computer program product and to a gas analysis device having such a monitoring unit. Furthermore, the invention relates to the use of artificial intelligence for monitoring a gas analysis device.

[0004] The publication KR 2021147318 A discloses a multisensor-based artificial intelligence designed to diagnose faults in a mechanical device. It collects sensor data and processes it using an autoencoder.

[0005] From the application US 2020 / 0310370 A1, a control system with an autoencoder is known in which sensor data is processed to produce derived sensor data. The autoencoder receives the derived sensor data and uses it to determine a prediction. Based on the prediction, a control intervention is triggered.

[0006] Increasingly complex systems with a growing number of individual devices are being used in various automation applications. Flexible operation for processing different input materials is also being sought. Due to the growing operational and system complexity, monitoring such systems is becoming more demanding. There is therefore a need for a method for monitoring complex systems that offers reliable, automatic error detection and can be quickly adapted to changing operating requirements. This applies in particular to gas analysis devices, which are becoming increasingly complex and sensitive due to increasing demands on measurement accuracy.

[0007] The problem is solved by a monitoring method according to the invention which is designed to monitor the operation of a system. The system comprises a plurality of devices which are each designed to provide a measured value and / or a control command. The devices can be, for example, sensors, control units, regulation units or combinations thereof. The measured values ​​can be recorded physical variables or status information about the respective device. The monitoring method comprises a first step in which an input signal array is provided which comprises a plurality of cells. One cell of the input signal array is suitable for receiving a predetermined measured value or control command and feeding it to a further processing unit. The input signal array can be designed as a one-, two- or higher-dimensional array, i.e. a data field.In the first step, a monitoring data array is provided that corresponds to the input signal array in terms of dimensionality and size (i.e., number of cells). Each cell of the input signal array corresponds to a cell of the monitoring data array.

[0008] The monitoring method also has a second step in which the system is made available in an active operating state. In the active operating state, measured values ​​and / or control commands are essentially generated continuously and made available to the monitoring method for further processing. The recorded measured values ​​and / or control commands are transferred to the input signal array and fed into the corresponding associated cells as content. These are processed by the further processing unit to form the contents of corresponding cells in the monitoring data array. In a third step, a system deviation parameter is determined based on the values ​​entered into the input signal array in the second step. The system deviation parameter is formed by a plurality of cells in the monitoring data array.The system deviation parameter can, for example, be a sum of the contents of cells in the monitoring data array or a sum of the squared contents of the cells in the monitoring data array. The contents of the cells in the monitoring data array can each be normalized individually or several of them together to a reference value. The contents are type-compatible with the contents of the corresponding cells in the input signal array. Weighting of individual cells in the monitoring data array is also possible. The weighting can be performed by an algorithm before or during the monitoring process. If the system deviation parameter exceeds an adjustable system threshold value in the third step, the presence of an improper state is detected.

[0009] The monitoring method also has a fourth step, which can be carried out if an improper operating state of the system is detected in the third step. In this fourth step, the contents of the cells of the monitoring data array combined in the system deviation parameter are processed separately. Pairwise cell deviation parameters are determined based on the contents of the cells of the monitoring data array and the contents of corresponding cells of the input signal array. This can be done, for example, by forming a difference. Furthermore, in the fourth step, a cell of the input signal array or the monitoring data array is detected as a defective cell if the amount of an associated cell deviation parameter exceeds an adjustable cell threshold. The cell threshold can be specified by a user, a table, an algorithm, or artificial intelligence.The defective cell can be a cell of the input signal array or a cell of the monitoring data array. An associated device can thus be identified as defective. In particular, a component of a device can also be identified as defective. Furthermore, in the monitoring method according to the invention, a warning is issued if a defective cell is detected. The warning can be in the form of a visual or acoustic warning to a user and / or as an electronic warning signal, for example to a control unit of the system. Based on an electronic control signal, the control unit can initiate a countermeasure during operation of the system.

[0010] The monitoring method according to the invention is essentially structured in two stages. Firstly, the second and third steps determine whether an improper operating condition exists at all. Only when an improper operating condition is detected is the cell deviation parameters analyzed in detail by separately recording and evaluating them in order to identify a defective cell and thereby obtain a detailed picture of the existing improper operating condition. As long as the operating condition is normal, the third step simply compares the system deviation parameter with the system threshold value. Accordingly, the number of computationally intensive comparison operations to be performed in the normal operating condition is minimized.The system deviation parameter can, for example, be designed as a sum of cell deviation parameters, which can be determined with reduced computing effort. Only when an improper operating state is detected does an increased number of comparison operations take place using the cell deviation parameters and the associated cell threshold values. The monitoring method according to the invention can therefore be carried out with reduced computing power. The monitoring method according to the invention, in particular the second, third and fourth steps, consequently offer real-time capability for a wide range of systems. The term real-time capability is to be understood in the respective sense of the corresponding system. The invention is based, among other things, on the finding that the system deviation parameter in the third step can be reliably used as a decision criterion for carrying out the fourth step.The monitoring method according to the invention offers overall increased reliability with reduced computational effort.

[0011] In one embodiment of the claimed monitoring method, this comprises a fifth step in which a defective cell or a determined set of defective cells determined in the fourth step is used further. In the fifth step, the defective cell or the set of defective cells is recorded and evaluated by pattern matching. By means of the pattern matching, a cause of the defect is determined in the fifth step. For this purpose, a database with a large number of stored comparison patterns of different damage states, i.e. improper operating states of the system, can be used. Alternatively or additionally, the pattern matching can also be carried out using artificial intelligence, for example a machine learning algorithm that is trained before and / or during operation of the system.During pattern matching, it is also possible to take into account the extent to which, i.e. by what amount, one or more defective cells exceed the corresponding cell threshold value. The cause of the defect can be identified, for example, as a defective device in the system, a loss of functionality of such a device, or external influences on the system. Since the claimed monitoring method can be carried out with reduced computing power or can be carried out quickly with increased computing power, such artificial intelligence can be further developed and, in particular, further trained in parallel with the operation of the system. The claimed monitoring method therefore offers improved detection of improper operating states of the system.

[0012] Furthermore, in the claimed monitoring method, the first, second, third and / or fourth step can be carried out using an autoencoder. The system deviation parameter determined in this way using the autoencoder can be a so-called total reconstruction error. Autoencoders are available in different configurations and exhibit a high degree of reliability and robustness. Furthermore, autoencoders are powerful and allow the rapid processing of a large number of input data, i.e., extensive input signal arrays. The claimed monitoring method can therefore be implemented easily and adapted to different systems. Furthermore, the autoencoder allows a plurality of steps of the claimed monitoring method to be carried out. Furthermore, so-called unsupervised learning is possible with an autoencoder.A data set for training the autoencoder requires a simplified sorting process, consisting of reduced data labeling. Furthermore, in normal operation, the autoencoder utilizes the relationship between the available current data. Thus, the use of a comprehensive database is unnecessary for most of the operating time.

[0013] Furthermore, at least one of the adjustable cell threshold values ​​in the claimed monitoring method can be designed as a fixed threshold value or as a sliding threshold value. A fixed threshold value offers simple configurability of the monitoring method. In the case of cell threshold values ​​that are normalized to a reference value, the claimed monitoring method can be easily configured using at least one fixed threshold value, for example during initialization. A sliding threshold value, on the other hand, makes it possible to take into account, for example, a degradation effect on a device, and thus on a corresponding cell of the input signal array. For example, the cell threshold value can increase or decrease along a function depending on the operating hours of a device.This avoids inappropriate warnings and / or reliably detects premature degradation of devices. The diagnostic accuracy of the claimed monitoring method is thus further increased. For example, the threshold can be an adjustable multiple of a standard deviation of the corresponding data and / or an adjustable multiple of a mean total reconstruction error. This makes it possible to provide both a fixed and a sliding threshold. Whether the threshold is sliding or fixed depends on whether the underlying values, for example the standard deviation or the mean total reconstruction error, are maintained during the claimed monitoring method.

[0014] In a further embodiment of the claimed monitoring method, the second step can comprise interpolating measured values ​​and / or control commands that are passed to the input signal array. Synchronous measuring points are thereby provided to the input signal array, so that each input of the input signal array describes an operating state of the system in a specific, discrete time interval. This enables increased temporal resolution to be achieved in the claimed monitoring method. Ambiguities or inconsistencies in inputs to the input signal array during transient operating states of the system, and thus inappropriate warnings, are thus minimized. The interpolation can be implemented as so-called upsampling. Consequently, the claimed monitoring method is suitable for time-critical applications that require accelerated responses to the current operating state.The claimed monitoring method therefore has increased reliability and diagnostic accuracy. Alternatively, so-called downsampling can also take place in the second step. This means that only those measured values ​​and / or control commands are passed to the input signal array that already lie in a discrete time interval. This reduces the computing power required for the claimed monitoring method. Furthermore, the input signal array and / or the monitoring data array can comprise cells in which measured values ​​and / or control commands are stored during the claimed monitoring method that satisfy an adjustable relevance criterion. The relevance criterion can, for example, be an indication of how many known defects in the corresponding cell are expected to cause the cell threshold value to be exceeded.A cell that exceeds the cell threshold in terms of magnitude with a minimal number of known defects, in particular with no known defects, can be removed from the input signal array. The claimed monitoring method can therefore continue to optimize itself automatically as the operating time progresses, and thus accelerate. This allows, among other things, simple modification of the system without the modifications made having a negative impact on the performance of the claimed monitoring method. The relevance criterion can be used in particular for so-called pruning, for example of the input signal array.

[0015] The underlying problem is similarly solved by a further monitoring method according to the invention. The monitoring method serves to monitor a system that has a plurality of devices that interact to ensure the intended operation of the system. The devices can be, for example, sensors, control units, regulation units or combinations thereof. The devices are designed to form nodes in a neural network that maps the interaction of the devices in the system. The neural network includes zero correlations and / or operating correlations. The devices can be physically separate devices or have different functionalities in physically individual devices.The monitoring method comprises a first step in which a trained neural network is provided which has a plurality of edges which each correspond to a zero correlation or an operating correlation between two nodes of the neural network. The nodes are formed by devices in the system. A zero correlation is understood to be a correlation with a correlation value which lies below an adjustable correlation threshold. A so-called zero correlation describes a weak, negligible or non-existent connection between two nodes. The term zero correlation is to be understood in the sense of the previously unpublished German patent application with the official file number 10 2022 200 694 . 1 . The disclosure content of file 10 2022 200 694 . 1 is incorporated into the present application by reference.The zero correlations have correlation values ​​in the provided trained neural network that are below the adjustable correlation threshold. The corresponding correlation values ​​are determined when the neural network is trained. The zero correlations therefore describe pairs of devices between which no discernible connection is to be expected under normal operating conditions. Operating correlations are correlations in which there is a sufficiently strong connection between two nodes. An operating correlation is therefore the opposite or the counterpart to a zero correlation. The term operating correlation is to be understood as corresponding to the term zero correlation according to file 10 2022 200 694 . 1.

[0016] Furthermore, the monitoring method comprises a second step in which the system is provided in an active operating state, and measured values ​​and / or control commands that occur in the active operating state are recorded. The measured values ​​can be recorded physical quantities or status information about the respective device. The recorded measured values ​​and / or control commands each correspond to at least one node and / or edge of the neural network. The recorded measured values ​​and / or control commands are suitable for serving as input for the neural network.

[0017] In a third step, a network deviation parameter is determined which combines correlation values ​​from a number of edges of the neural network. For this purpose, correlation values ​​are determined for several edges based on the measured values ​​and / or control commands recorded in the second step, which are then summed up, for example in terms of their absolute value. The correlation values ​​can be recorded for edges of zero correlations and / or operating correlations in order to determine the network deviation parameter. The network deviation parameter thus represents a compressed value which maps the overall status of the system. This is followed by a fourth step in which an improper operating state of the system is detected if the network deviation parameter exceeds an adjustable network threshold in terms of absolute value.The network threshold can be specified by a user or an algorithm and represents a measure that reflects a deviation from a desired target state of the system. Exceeding or exceeding the network threshold can therefore be caused by correlation values ​​of edges that are intended to be zero correlations increasing to such an extent that they can no longer be accurately categorized as zero correlations, either individually or in combination. The monitoring method according to the invention thus detects the presence of spurious correlations where zero correlations are present or should be present in a normal operating state. Alternatively or additionally, operating correlations can also exhibit deviations which, when combined, lead to the network threshold being exceeded or exceeded.The monitoring method according to the invention is based, among other things, on the surprising discovery that improper operating conditions can be identified in a simple and reliable manner using zero correlations which otherwise have no further significance for operation. More detailed knowledge of the zero correlations is unnecessary since they are free of underlying physical relationships. Furthermore, edges with zero correlations can be generated essentially automatically in an increased number. Likewise, a deviation from operating correlations can be identified in a simple manner and an improper operating condition can be recognized. If the improper operating condition is detected, a warning is issued. The warning can be a visual or acoustic warning to a user and / or an electronic warning signal, for example to a control unit of the system.Using an electronic control signal, the control unit can initiate a countermeasure during the operation of the system.

[0018] The grid deviation parameter can be provided with reduced computational effort and checked for excess of the grid threshold using a minimum of comparison operations. Accordingly, the monitoring method according to the invention can be implemented with reduced computational effort and at increased speed. The first, second, third, and fourth steps thus increase the real-time capability of the monitoring method according to the invention. The term "real-time capability" should be understood in light of the specific application, i.e., the system being monitored by the monitoring method. At the same time, the monitoring method permits reliable diagnosis of an existing improper operating state.

[0019] According to one embodiment of the claimed monitoring method, the second, third and / or fourth step can be carried out separately for zero correlations and operating correlations. In particular, the second, third and / or fourth step for the zero correlations can be carried out with an adjustable first frequency. Furthermore, the second, third and / or fourth step for the operating correlations can be carried out with an adjustable second frequency. The first and second frequencies can be specified by a user or an algorithm. Furthermore, the first frequency can be lower than the second frequency. The claimed monitoring method is therefore adaptable to different applications in which an improper operating state is more likely to be detected via the operating correlations or the zero correlations. Overall, the claimed monitoring method has a broad range of applications.Alternatively, the second, third, and / or fourth steps can be performed in a first pass for the operating correlations. If an abnormal operating state is detected in the first pass, the second, third, and / or fourth steps can be performed for the zero correlations. Separate network deviation parameters are determined in the first and second passes and compared with the associated adjustable network threshold values. The claimed monitoring method can thus be implemented with reduced computing power.

[0020] In a further embodiment of the claimed monitoring method, this comprises a fifth step in which an edge of the neural network which corresponds to a zero correlation in a normal operating state of the system is detected as a defective edge. The edge is detected as a defective edge if its correlation value exceeds the associated correlation threshold in terms of amount. The fifth step is carried out if the improper operating state of the system is detected in the fourth step. The absolute exceeding of the correlation threshold is carried out by a separate comparison of the correlation values ​​with the corresponding correlation threshold.As a result, computationally intensive comparison operations are only performed when, based on the exceeded network threshold, it can be expected that at least one correlation threshold has been exceeded, and frequent unsuccessful runs of computationally intensive operations can be avoided. The monitoring method according to the invention therefore offers a rational use of the computing power employed. At the same time, the claimed monitoring method offers a targeted diagnosis.

[0021] In addition, the correlation threshold of at least one edge in the neural network can be designed as a fixed threshold or as a sliding threshold. A fixed threshold offers simple configurability of the monitoring method. With correlation thresholds that are normalized to a reference value, the claimed monitoring method can be easily configured using at least one fixed threshold, for example during initialization. A sliding threshold, on the other hand, makes it possible to take into account, for example, a degradation effect on a device, and thus on a corresponding edge of the neural network. The sliding threshold can be formed, for example, by a changing total reconstruction error. For example, the correlation threshold can increase or decrease along a function depending on the operating hours of a device.This prevents inaccurate warnings and / or reliably detects premature device degradation. The diagnostic accuracy of the monitoring method is thus further increased.

[0022] Furthermore, the claimed monitoring method can have a sixth step in which a pattern comparison is carried out based on the defect edge detected, for example, in the fifth step. In the sixth step, the defect edge is detected and evaluated by a pattern comparison. In the sixth step, the cause of the defect is determined using the pattern comparison. For this purpose, a database with a large number of stored comparison patterns of different damage states, i.e. improper operating states of the system, can be used. Alternatively or additionally, the pattern comparison can also be carried out using artificial intelligence, for example a machine learning algorithm that is trained before and / or during operation of the system.During pattern matching, it is also possible to take into account the extent to which, i.e. by what amount, one or more defect edges exceed the corresponding correlation threshold. The cause of the defect can be identified, for example, as a defective device in the system, a loss of functionality of such a device, or external influences on the system. Since the claimed monitoring method can be carried out with reduced computing power or can be carried out quickly with increased computing power, such artificial intelligence can be further developed and, in particular, further trained in parallel with the operation of the system. The claimed monitoring method therefore offers improved detection of improper operating states of the system.

[0023] The described problem is equally solved by a computer program product according to the invention. The computer program product is designed to receive measured values ​​and / or control signals from devices in a system, to store them at least temporarily and to process them. The computer program product is also designed to issue a warning to a user and / or an electronic warning signal. The computer program product serves to monitor the system to which the devices belong and for whose operation they interact. The measured values ​​and / or control commands are processed as part of a monitoring method which can run concurrently, i.e. essentially in parallel, in particular in real time, to the operation of the system. According to the invention, the computer program product is designed to carry out a monitoring method according to at least one of the embodiments outlined above.The computer program product can be monolithic, i.e., designed for operation on a single hardware platform. Alternatively, the computer program product can be modular, i.e., comprise subprograms that can be executed on separate hardware platforms, for example, a computer cloud, and interact via a communicative data connection to execute the respective monitoring method. The computer program product according to the invention allows the monitoring method according to the invention to be implemented in a simple manner.

[0024] The underlying object is also achieved by a monitoring unit for a system according to the invention. The monitoring unit is designed to monitor operation of the system, to which it is at least functionally coupled, and to issue a warning. The system to be monitored comprises a plurality of devices which interact to operate the system and for which measured values ​​and / or control commands can be recorded. The recording of the measured values ​​and / or control commands, which also includes receiving, is carried out by the monitoring unit. The monitoring unit is also designed to process the recorded measured values ​​and / or control commands. According to the invention, the monitoring unit is designed to process the measured values ​​and / or control commands by means of a computer program product with which the monitoring unit is equipped.According to the invention, the computer program product is designed according to one of the embodiments described above.

[0025] The object set out at the outset is further achieved by a gas analysis device according to the invention. The gas analysis device comprises a plurality of devices which interact to prepare and measure a substance sample. In particular, the gas analysis device can be designed to determine the composition of the substance sample at least qualitatively. The gas analysis device is provided with a monitoring unit which is coupled to a plurality of devices. The monitoring unit is designed to detect an improper operating state of the gas analysis device. According to the invention, the monitoring unit is designed according to one of the embodiments set out above.

[0026] Furthermore, the outlined object is achieved by an inventive use of artificial intelligence which is designed to monitor a gas analysis device. The artificial intelligence is suitable for directly or indirectly receiving and processing measured values ​​and / or control commands from devices of the gas analysis device. The use of the artificial intelligence comprises at least issuing a warning, which can be, for example, a visual or acoustic warning to a user and / or an electronic warning signal. In order to monitor the gas analysis device, the artificial intelligence is designed to carry out a monitoring method. According to the invention, the monitoring method can be designed according to one of the embodiments described above.The use of artificial intelligence may in particular comprise the running of an appropriately designed computer program product, as set out above, for example.

[0027] The invention is explained in more detail below with reference to individual embodiments in figures. The figures are to be read as complementary to one another in that identical reference numerals in different figures have the same technical meaning. The features of the individual embodiments can also be combined with one another. Furthermore, the embodiments shown in the figures can be combined with the features outlined above. They show in detail:

[0028] FIG 1 schematically shows a sequence of a first embodiment of the claimed monitoring method;

[0029] FIG 2 schematically shows a sequence of a second embodiment of the claimed monitoring method.

[0030] A first embodiment of the claimed monitoring method 100 is shown schematically in FIG 1. The monitoring method 100 serves to monitor a system 10 with which a system process 20 is carried out. The system 10 is designed as a gas analysis device 11 and the system process 20 carried out thereon comprises determining a composition 25 of a substance sample 22. For this purpose, the system 10 is equipped with a plurality of devices 12 which are designed as sensors 14 or as a control unit 14. The control unit 16 is designed to influence the system process 20 by means of a running control program 18. The sensors 14 record physical variables relevant to the system process 20. The sensors 14 are designed to output measured values ​​15 and the control unit 14 to output control commands 17.The first embodiment shown in FIG. 1 assumes that a first step 110 has been completed, in which an input signal array 30 having a plurality of cells 32 is provided. Likewise, in the first step 110, a monitoring data array 50 having a plurality of cells 52 is provided. The cells 52 of the monitoring data array 50 and the cells 32 of the input signal array 32 correspond to one another, so that the input data array 32 and the monitoring data array 52 are identical in terms of structure.

[0031] In the first embodiment of the monitoring method 100, a second step 120 takes place in which the system 10 is made available in an active operating state in which measured values ​​15 are provided by the sensors 14 and / or control commands 17 are provided for the input signal array 30 by the control unit 16. The measured values ​​15 or the control commands 17 are provided 31 by separately storing them in a cell 32 of the input data array 30. The input signal array 30 is then provided 41 to a further processing unit 40, which is designed as an autoencoder 42. A third step 130 of the monitoring method 100 is carried out by means of the further processing unit 40.

[0032] By means of the further processing unit 40, contents, i.e. values, for cells 52 of the monitoring data array 50 are determined based on the input signal array 30. On the basis of this, a comparison value 54 of the monitoring data array 50 is in turn determined, which combines the contents of its cells 52. Likewise, a comparison value 34 of the input signal array 30 is determined, which likewise combines the contents of its cells 32. The comparison values ​​34, 54 can each be formed as a sum, for example. Furthermore, by comparing the comparison values ​​34, 54 of the input signal array 30 and the monitoring data array 50, a system deviation parameter 55 is determined. The system deviation parameter 55 is a total reconstruction error and represents an overall state of the system 10 in compact form. The system deviation parameter 55 is further compared with a system threshold value 57 in the third step 130.If the system deviation parameter 55 exceeds the system threshold value 57, an improper operating state of the system 10 is detected. The improper operating state detected in the third step 130 is non-specific, i.e., is not suitable for indicating an underlying cause of the defect. If the improper operating state of the system 10 is detected in the third step 130, a fourth step 140 follows in the claimed monitoring method 100.

[0033] In the fourth step 140, the contents of cells 52 of the monitoring data array 50 are compared separately, i.e. individually, with the contents of cells 32 of the input signal array 30. In a normal operating state of the system 10, the further processing unit 40, designed as an autoencoder 42, maps the contents of the cells 32 of the input signal array 30 essentially identically as contents into the cells 52 of the monitoring data array 50. The mapping 43 takes place by storing contents line by line. Accordingly, in the fourth step 140, a cell deviation parameter 45 is determined separately for a plurality of cells 32 of the input signal array 30 and corresponding cells 52 of the monitoring data array 50. The cell deviation parameter 45 can be determined, for example, by a difference between the associated cells 32, 53.The cell deviation parameters 45 determined in this way are each compared with an adjustable cell threshold value 47. Such a comparison is symbolized in FIG. 1 by the arrows 48. If a cell deviation parameter 45 exceeds the associated adjustable cell threshold value 47 in terms of amount, the associated cell 32, 52 of the input signal array 30 or of the monitoring data array 50 is recognized as a defective cell 58. For this purpose, the defective cells 58 are marked accordingly, for example with a so-called signature bit. Furthermore, a fifth step 150 takes place in the monitoring method 100, for which the contents of the cells 52 of the monitoring data array 50 are provided for comparison with a database 60. The provision for the fifth step 150 is shown in FIG 1 by an arrow 41.In the fifth step 150, an excess width 61 is determined for at least one defective cell 58, by which the content of the defective cell 58 exceeds the associated cell deviation parameter 45 in terms of amount. Furthermore, it is recorded for which cells 32, 52 of the monitoring data array 50 or of the input signal array 30 the corresponding cell threshold value 47 is exceeded in terms of amount. This results in a pattern 62 that is characteristic of an underlying cause of the defect, for example in connection with cell deviation parameters 45 that fall below the associated cell threshold value 47 in terms of amount. The pattern 62 is recognized in the fifth step 150 by a pattern comparison 65 using the database 60. A plurality of patterns 62 that serve as comparison patterns are stored in the database 60. If a pattern 62 is recognized in the fifth step 150, a corresponding warning 59 is issued.The warning 59 can indicate the presence of an improper operating state of the system 10, an indication of a defective device 12 associated with at least one defective cell 58, and / or an indication of the detected cause of the defect. The warning 59 can be sent to a display unit 64, which is designed to output the warning 59 to a user, for example, as a visual or acoustic warning. Alternatively or additionally, the warning 59 can be designed as an electronic warning signal, which is output to a communication interface 66, which can be connected to a higher-level control unit (not shown in detail). The monitoring method 100 in FIG. 1 is implemented by a computer program product 70, which is executed on a monitoring unit 75. The computer program product 70 forms an artificial intelligence 99, which is used to monitor the system 10, which is designed as a gas chromatograph 11.A second embodiment of the claimed monitoring method 200 is shown schematically in FIG 2. The monitoring method 200 serves to monitor a system 10 with which a system process 20 is carried out. The system 10 is designed as a gas analysis device 11 and the system process 20 carried out thereon comprises determining a composition 25 of a substance sample 22. For this purpose, the system 10 is equipped with a plurality of devices 12 which are designed as sensors 14 or as a control unit 14. The control unit 16 is designed to influence the system process 20 by means of a running control program 18. The sensors 14 record physical variables relevant to the system process 20. The sensors 14 are designed to output measured values ​​15 and the control unit 14 to output control commands 17.

[0034] The second embodiment shown in FIG. 2 assumes that a first step 210 has been completed, in which a neural network 80 is provided which is trained for use in the monitoring method 200. The neural network 80 comprises a plurality of nodes 82 which are connected to one another via edges 81. Each edge 81 is assigned a correlation value 83 which indicates how strongly a change at a node 82, that is to say a measured value 15 or control command 17 present there, follows a node 82 connected via the corresponding edge 81. The correlation threshold values ​​83 can be set, for example, by a user input, a value table or an algorithm, in particular an artificial intelligence. The edges 81 in the neural network 80 according to FIG. 2 are designed as zero correlations 85 in a normal operating state of the system 10.Accordingly, the magnitude of the correlation values ​​83 of the corresponding edges 81 is below an adjustable correlation threshold 86, which essentially describes a minimum of interaction between the corresponding nodes 82. The zero correlations 85 in the neural network 80 thus describe measured values ​​15 or control commands 17, between which, in the intended state, not even an apparent connection exists. The edges 81 corresponding to zero correlations 85 are determined in the first step 210 by training the neural network 80.

[0035] The monitoring method 200 comprises a second step 220 which builds on the completed first step 210. In the second step 220 the system 10 is made ready in an active operating state in which the system process 20 is carried out and the devices 12, that is to say the sensors 14 and the control unit 16, generate and provide measured values ​​15 and control commands 17. The providing 31 is symbolized by arrows in FIG. 2. The measured values ​​15 and the control commands 17 each correspond to a node 82 and / or an edge 81 of the neural network 80 and are fed into the neural network 80 as input values. This is followed by a third step 230 in which, starting from the measured values ​​15 and / or edge 81 provided in the second step 220, Control commands 17 are used to determine, i.e. calculate, the correlation values ​​83 present at the respective edges 81.The available correlation values ​​83 are stored, for example, in cells 92 of a table 90, and a grid deviation parameter 94 is determined based thereon. The grid deviation parameter 94 can, for example, be formed as the sum of the absolute values ​​of the correlation values ​​83. The grid deviation parameter 94 is a measure of the presence of an unspecified, improper operating state of the system 10. The monitoring method 200 further comprises a fourth step 240 being carried out. In the fourth step 240, the grid deviation parameter 94 determined in the third step 230 is compared with a grid threshold value 95. In a normal operating state of the system 10, the absolute value of the grid deviation parameter 94 falls below the grid threshold parameter 95.The network threshold value 95 can be designed as a fixed threshold value or as a sliding threshold value, which can increase, for example, taking into account the operating hours completed by the system 10. If an improper operating state is detected, a warning can be issued. Furthermore, a fifth step 250 belongs to the monitoring method 200 according to FIG. 2, in which the correlation values ​​83 of the individual edges 81 are compared with the associated correlation threshold values ​​86. If the correlation value 83 of an edge 81 exceeds the associated correlation threshold value 86 in terms of amount, the associated edge 81 is recognized as a defective edge 87. The defective edge 87 allows a conclusion to be drawn that the measured value 15 or Control command 17 at one of the nodes 82 connected by the edge 81 was caused by an improper operating condition of the system 10.Accordingly, it can be diagnosed that a device 12 belonging to the corresponding node 82 is defective or at least is expected to be defective.

[0036] In addition, the monitoring method 200 comprises a sixth step 260 in which at least the correlation values ​​83 are made available for further evaluation. The making available 41 is represented by an arrow in FIG 2. In the sixth step 260, based on at least one detected defect edge 81, a pattern 62 is determined which comprises a plurality of existing correlation values ​​83. For this purpose, the correlation values ​​83 of a plurality of edges 81 are compared jointly, i.e. essentially simultaneously, with the respectively associated correlation threshold value 86, which are zero correlations 85 in the intended operating state. From the absolute values ​​of the correlation values ​​83 and the extent to which these possibly exceed the associated correlation threshold value 86, characteristic patterns 62 arise for the respective defect case.The patterns 62 are recognized in the sixth step 260 by means of a database 60, with which a pattern comparison 65 is carried out. Accordingly, a cause of the defect is diagnosed by the pattern comparison 65. If a pattern 62 is recognized in the sixth step 260, a corresponding warning 59 is issued. The warning 59 can indicate the presence of the improper operating state of the system 10, an indication of a defective device 12 associated with the defect edge 81, and / or an indication of the detected cause of the defect. The warning 59 can be sent to a display unit 64, which is designed to output the warning 59 to a user, for example as a visual or acoustic warning. Alternatively or additionally, the warning 59 can be designed as an electronic warning signal, which is output to a communication interface 66, which can be connected to a higher-level control unit (not shown in detail).The monitoring method 200 in FIG 2 is implemented by a computer program product 70 which is executed on a monitoring unit 75.

[0037] The computer program product 70 forms an artificial intelligence 99 which is used to monitor the system 10 designed as a gas chromatograph 11.

Claims

Patent claims 1. Monitoring procedure (200) for a plant (10) with a plurality of devices (12) suitable for forming nodes (82) in a neural network (80) that represents an interaction of the devices (12) in the plant (10), comprising the steps of: a) providing a trained neural network (80) having a plurality of edges (81), each corresponding to a zero correlation (85) or an operating correlation, and in the case of the zero correlations (85), determining a correlation value (83) during training that is below an adjustable correlation threshold (86); b) providing the plant (10) in an active operating state and acquiring measured values ​​(15) and / or control commands (17), each corresponding to at least one node (82) and / or edge (81) of the neural network (80);c) Determining a network deviation parameter (94) that combines correlation values ​​(83) of a plurality of edges (81) of the neural network (80); wherein the correlation values ​​(83) are determined based on the measured values ​​(15) and / or control commands (17) acquired in step b); d) Detecting an unintended operating state of the system (10) when the network deviation parameter (94) exceeds an adjustable network threshold value (95) in magnitude; wherein a warning (59) is issued when the unintended operating state is detected, characterized in that the monitoring procedure (200) comprises a further step e) in which an edge (81) of the neural network (80) that corresponds to a zero correlation (85) in an intended operating state is detected as a defect edge (87) when its correlation value; (83) exceeds the corresponding correlation threshold (86) in absolute value.

2. Monitoring method (200) according to claim 1, characterized in that steps b) , c) and / or d) are performed for operating correlations with a first frequency and for zero correlations (85) with a second frequency .

3. Monitoring method (200) according to claim 1 or 2, characterized in that steps b) , c) and / or d) are carried out in a first pass for operational correlations and in a second pass for zero correlations (85) if an unintended condition is detected in the first pass.

4. Monitoring method (200) according to one of claims 1 to 3, characterized in that the correlation threshold (86) is designed as a fixed threshold or as a sliding threshold.

5. Monitoring method (200) according to one of claims 1 to 4, characterized in that the monitoring method (200) comprises a further step f) in which a pattern comparison is performed based on at least one detected defect edge (87). (65) is carried out and a cause of the defect is identified.

6. Computer program product (70) for monitoring a plant (10) which has a plurality of devices (12) which cooperate to operate a plant (10) and for which measured values ​​(15) and / or control commands (17) can be acquired and processed in a monitoring procedure (200), characterized in that the monitoring procedure (200) is designed according to one of claims 1 to 5.

7. Monitoring unit (75) for a plant (10) comprising a plurality of devices (12) that interact to operate the plant (10) and to which measured values ​​(15) are transmitted and / or control commands (17) are detectable, wherein the monitoring unit (75) is configured to receive and process the measured values ​​(15) and / or control commands (17) and is configured to issue a warning (59), characterized in that the monitoring unit (75) for processing the measured values ​​(15) and / or control commands (17) is equipped with a computer program product (70) according to claim 6.

8. Gas analysis device (11) comprising a plurality of devices (12) for preparing and measuring a substance sample (22) which is equipped with a monitoring unit (75) for detecting an unintended operating condition of the gas analysis device (11), characterized in that the monitoring unit (75) is configured according to claim 7.

9. Use of an artificial intelligence (99) for monitoring a gas analysis device (11), wherein the use includes at least issuing a warning (59) and the artificial intelligence (99) performs a monitoring procedure (200), characterized in that the monitoring procedure (100, 200) is configured according to one of claims 1 to 5.

Images