METHOD FOR DETERMINING A FAULT FOR A SET OF SENSORS

Description

technical field

[0001] This presentation concerns the field of validating data acquired on a system by sensors, and more specifically a fault determination method for a set of sensors. Such a method finds application in numerous industrial fields, for example, the real-time or delayed monitoring of aeronautical or space systems. Previous technique

[0002] A physical system can be monitored by sensors to provide information about its operation, for example, to improve its reliability or detect system malfunctions. When the system in question is critical, such as the control of a rocket engine in flight, hardware redundancy of the sensors can be implemented: several sensors acquire the same or similar data, so that if one of them fails, it is still possible to monitor the system accurately and effectively without compromising its operation. Furthermore, US patent 2015 / 233730 and French patent 3 013 834 describe methods for fusing sensor data.

[0003] However, hardware redundancy is detrimental, particularly regarding the cost of sensors and their maintenance, for performing a given function. Conversely, reducing hardware redundancy raises the question of the reliability associated with the validity of the remaining measurements: if data is returned by only one sensor, how can we ensure that this data is reliable?

[0004] Therefore, there is a need for a new type of fault determination process for a set of sensors. Description of the invention

[0005] To that end, the present presentation relates to a method for determining failure for a set of sensors according to claim 1.

[0006] A sensor array comprises several sensors (at least two, and ideally at least three), and the method aims to determine if there is a fault in at least one of these sensors. The sensors provide measurements taken on the monitored physical system (hereinafter referred to as the system). Preferably, the sensors are mounted to monitor the same system.

[0007] In parallel, the system is also monitored by a plurality of estimators (at least two, and possibly at least three). An estimator determines (provides as output) an estimate of a physical quantity of the system under consideration. An estimator takes as input an electronic signal (for example, computer or logic), while a sensor takes as input a physical phenomenon directly. Each estimator can take as input one or more measurements returned by the sensors, directly or indirectly: for example, a first estimator can take as input three measurements; a second estimator can take as input one measurement and the estimate provided by the first estimator, which itself is based on three measurements: the second estimator is then considered to take as input, In short, a subset formed by said measure and said three measures (said measure possibly being confused with one of said three measures).

[0008] For the purposes of this discussion, and unless otherwise indicated, the mention of a "first" element, such as a first estimator, does not necessarily imply the existence of a "second" element, nor, if applicable, an order relationship between the first and second elements. Ordinal qualifiers are used in this context solely for clarity and identification, without prejudging any particular characteristics. Similarly, and conversely, the mention of a higher-ranking element (third, etc.) does not imply by no means that lower rank elements, such as a possible second element, exist and / or possess the characteristics that may have been presented elsewhere.

[0009] Some estimators can take the same subsets of measurements as input. For example, a third estimator might take as input the estimate provided by the second estimator. Ultimately, the basic information considered by the third estimator will therefore be the same as that considered by the second estimator. However, among the estimators considered, there are several that take different subsets of measurements as input: each estimator in this plurality takes as input, directly or indirectly, a subset of measurements that differs from the subset of measurements taken as input by the other estimators. Two subsets differ if their intersection is not equal to their union. The term "subset" is to be understood in a broad sense, meaning that an estimator can take as input the entire set of measurements, that is, all the measurements returned by all the sensors.

[0010] For the sake of brevity, we will refer to an estimator as determining one estimate. However, this does not preclude cases where an estimator provides multiple estimates, for example, of different but optionally related physical quantities.

[0011] At least one of the estimators includes an artificial neural network. An artificial neural network (ANN) is a computational model comprising one or more neurons, each neuron having a transfer function. The ANN thus possesses an overall transfer function that allows it to compute at least one output based on at least one input. The transfer functions of each neuron, or the relative weights of the neurons in the network, can be weighted by coefficients called synaptic weights (or simply weights) and biases. The weights can be modulated as the ANN learns. In principle, learning consists of providing the ANN with a set of situations in which the inputs and outputs (or target outputs) are known. During learning, the ANN adapts its synaptic weights and biases to conform to the learned situations, possibly with a certain degree of tolerance.RNAs are therefore an example of mathematical expressions whose coefficients are iteratively adapted based on learning. Weights and biases, once... determined Following the learning phase, these values ​​may or may not remain constant during the operational phase. Typically, learning can be carried out at least before, and possibly concurrently with, the implementation of the fault determination process.

[0012] An ANN (Automatic Numbering) system is therefore capable of intelligently modeling the relationship between physical quantities by learning from the actual behavior of these quantities within the system, without needing to know the theoretical laws that govern it. Consequently, the ANN is not necessarily constrained by the causal relationships that connect the different physical quantities to one another. Thus, for example, flow rate is the physical consequence of a difference between an inlet pressure and an outlet pressure. In a quasi-static system, an ANN will be able, unlike a classical model built on the laws of physics (hereafter "physical model" or "multiphysics model"), to determine, following a learning process, the upstream pressure as a function of the downstream pressure and the flow rate. In other words, an ANN is capable of swapping inputs and outputs compared to classical physical or multiphysics models. These are referred to as inverse relationships.

[0013] Furthermore, the ANN possesses a generalization capability (also called "inference"), meaning it can determine the output values ​​of a situation given input data, even if that situation was not trained during the learning phase, provided the inputs are contained within a domain of validity (in other words, it is ensured that the input values ​​of the situation to be calculated are between or close to the extreme input values ​​used for training). This is particularly useful for systems with a wide range of operating points, where explicitly parameterizing all possible situations is industrially impractical. For greater accuracy, it is necessary to verify that the training situations constitute a sufficiently refined mesh of the operating domain. In addition, the ANN is a readily modular system.These advantages also apply to other types of mathematical expressions with iteratively evaluated coefficients.

[0014] Furthermore, a trained RNA is generally faster at providing an estimate than a physical or multiphysics model, especially when the model is nonlinear. This is particularly advantageous for applications The real-time process allows for the construction and / or training of the ANN to be carried out beforehand. Finally, since an ANN has no internal restrictions dictated by physics, it is well suited to so-called "hybrid" systems, mixing continuous and discrete quantities.

[0015] Thanks to sensors and estimators, for at least one physical quantity of the system, there are at least two data points obtained differently: either by several sensors, or by several estimators, or by a combination of at least one sensor and at least one estimator. For at least some physical quantities, the data are obtained in whole or in part by estimators, which ensures software redundancy and allows for cost savings compared to purely hardware redundancy.

[0016] For a given physical quantity, for which at least two data points have been obtained, the procedure involves calculating residuals that represent the difference between data points representing that physical quantity. For example, data points representing this physical quantity are taken, and the differences between these data points taken two at a time are calculated as residuals. The difference can be a simple scalar difference or, more generally, a distance in an appropriate space.

[0017] Determining that at least one sensor has failed can be achieved using a transfer function that takes residuals as input, optionally all residuals, and returns a failure indicator as output for at least one sensor that provided a measurement on which the residuals are at least partially based. The failure is thus determined from residuals, which are themselves calculated from data, which are themselves measurements, or estimates determined from measurements, with the measurements themselves being returned by sensors; a potential failure is then identified for one of these sensors. Conversely, with this method, a failure cannot be identified for sensors that have not provided any measurements leading to the calculation of residuals.

[0018] The transfer function can be, for example, a matrix, notably determined analytically or numerically, one or more RNAs, or something else.

[0019] The fact that the estimators take different subsets of the measurements as input ensures that the residuals represent discrepancies between data from different measurements, so that examining the residuals allows us to trace the erroneous measurement back to the faulty sensor. Furthermore, the fact that at least one estimator includes a random variable recognition (RVR) provides greater robustness for determining the residuals. Thus, the present fault determination method allows for at least partial software redundancy, with the associated cost and weight savings, while offering a robust assessment of the validity of the obtained data.

[0020] Furthermore, at least some of the residuals are representative of the difference between a measurement and an estimate, or between two estimates. This ensures software redundancy.

[0021] In some embodiments, residuals are calculated between all pairs of data representing the same physical quantity. For example, all data representing the same physical quantity are taken and paired to form all possible pairs. Thus, with three data points representing the same quantity, there will be three pairs: (1, 2), (2, 3), and (1, 3). Residuals can then be calculated for each pair, for example, by subtracting the two data points from the pair.

[0022] In some embodiments, the sensor array comprises at most one sensor per physical quantity. This avoids hardware redundancy and further increases cost and weight savings.

[0023] In some embodiments, the plurality of estimators includes at least one estimator to provide an estimate of a physical quantity measured by one of the sensors. There is therefore software redundancy between the sensor and the estimator. Preferably, the estimator does not take as input the measurement returned by the sensor. More generally, the estimator can take as input measurements other than the measurement returned by the sensor. Fault detection is made more precise by exploiting the links between different measurements within the same monitored physical system.

[0024] In some embodiments, the artificial neural network is configured to determine an estimate of a first physical quantity occurring upstream of at least a second physical quantity whose measurement is taken as input by the artificial neural network.

[0025] As explained previously, there can be a physical cause-and-effect relationship between the first quantity (cause) and the second quantity (effect), in which case the first quantity is said to be upstream of the second quantity: the second physical quantity is normally a consequence of the first physical quantity. The RNA (Natural Reversal Network) is based on an inverse relationship, allowing it to take the effect (the second quantity) as input and determine, as output, an estimate of the cause (the first quantity). Such a determination, impossible with a multiphysics model, increases the flexibility for determining residuals and therefore, ultimately, the reliability of fault detection.

[0026] In some embodiments, in determining that there is a fault, the residuals are given as input to at least one artificial neural network configured to return, for at least one of the sensors, an indicator of the existence and optionally of the severity of a fault on said sensor.

[0027] In these embodiments, the aforementioned transfer function includes an RNA that takes residuals as input and returns at least one fault indicator as output. The RNA can return one indicator per sensor and / or one indicator per fault, or a single indicator for all sensors and all faults (for example, a number where each digit refers to a given sensor), or any intermediate situation. Similarly, the existence and severity of a fault can be reflected by the same indicator (for example, 0 for no fault and a positive number for a fault, increasing with severity) or by different indicators, or any intermediate situation.

[0028] The present presentation also relates to a fault determination system for a set of sensors according to claim 8.

[0029] This fault determination system may be suitable for implementing the fault detection process as described above.

[0030] In some embodiments, the different steps of the fault determination process are determined by computer program instructions.

[0031] Consequently, this presentation also concerns a program on an information medium, this program being likely to be implemented in a fault determination system or more generally in a computer, this program comprising instructions adapted to the implementation of the steps of a fault determination process as described above.

[0032] This program can use any programming language, and be in the form of source code, object code, or code somewhere between source code and object code, such as in a partially compiled form, or in any other desirable form.

[0033] This presentation also aims at an information medium readable by a computer or microprocessor, and containing instructions for a program as mentioned above.

[0034] The information medium can be any entity or device capable of storing the program. For example, the medium can include a storage means, such as a ROM, for example a CD ROM or a microelectronic circuit ROM, or a magnetic recording means, for example a floppy disk or a hard disk drive.

[0035] On the other hand, the information medium can be a transmissible medium such as an electrical or optical signal, which can be transmitted via an electrical or optical cable, by radio, or by other means. The program described herein can, in particular, be downloaded from a network such as the Internet.

[0036] This presentation also relates to a method for estimating a physical quantity from a set of sensors, comprising determining the failure of said set of sensors according to any of the embodiments described above, and producing an estimate of the physical quantity from the sensors measuring said physical quantity and the estimators determining an estimate of said physical quantity, ignoring the defective sensors for which a failure has been determined, and the estimators taking as input measurements returned by said defective sensors. By ignoring the defective sensors and the estimators based, directly or indirectly, on these sensors, the estimation of the physical quantity can be performed on the basis of reliable data.Redundancy can be structured so that reliable data always remains even if a number of sensors fail, this number being sized according to the needs and criticality of the monitored system.

[0037] The present exposition also relates to a method for learning an artificial neural network for a fault determination system for a set of sensors according to claim 9.

[0038] After training, the artificial neural network can be used for determining that there is a fault on at least one of the sensors in a process or the fault determination system according to any of the embodiments previously described.

[0039] The principle of the learning process is to generate a training dataset containing data into which known errors—the faults—have been introduced. The RNA is then trained using the residuals calculated from this data as input and the fault indicators established to correspond to the introduced known errors as output. From the same data representing multiple measurements, it is possible to create a training dataset comprising different instances of the data, generating faults that differ between instances in their intensity, the sensor(s) they affect, and so on. Thus, a fault can be characterized by at least one sensor and at least one amplitude. The sensors and amplitudes can optionally be selected randomly to easily generate a large training dataset.

[0040] The estimates are determined by the estimators after fault generation, so that they take faults into account in a way that is consistent with the measurements, as would happen on a real physical system in the event of a sensor failure.

[0041] Learning strictly speaking of the ANN, given an input (a first set of data) and an output target (a second set of data), is an operation known in itself to the person skilled in the art, using for example a loss function whose purpose is to make the estimate (third set of data), provided by the ANN on the basis of the first set of data, converge to the second set of data.

[0042] In some embodiments, the generation of a fault indicator includes comparing the corresponding fault with a threshold associated with the sensor reporting the measurement in question, and issuing a fault severity level for said sensor based on said comparison. The corresponding fault is the fault introduced on the measurement in question and for which the fault indicator was generated.

[0043] For example, a failure can be considered more serious if the fault exceeds a certain threshold, such as the magnitude of the measured physical quantity. It is also possible to define several successive thresholds, with the failure being considered more serious the greater the number of thresholds crossed. Brief description of the drawings

[0044] Other features and advantages of the object of this presentation will emerge from the following description of embodiments, given by way of non-limiting examples, with reference to the attached figures. [ Fig. 1 ] There figure 1 schematically represents a fault detection system for a set of sensors, according to one embodiment. Fig. 2 ] There figure 2 schematically represents a learning process for an artificial neural network for a fault determination system for a set of sensors, according to one embodiment. Detailed description

[0045] A fault determination method for a set of sensors will be described with reference to the figure 1In this example, a sensor set 10 comprises three sensors A, B, and C. However, any number of sensors, preferably two or more, could be included in the sensor set 10. The sensor set 10 is intended to be mounted on a physical system to be monitored (not shown), for example, a rocket engine. The physical quantities to be monitored in the system and the appropriate types of sensors for this purpose can be chosen by a person skilled in the art who is familiar with the system, according to their monitoring requirements. In the case of a rocket engine, the quantities considered could be chosen from among pressure, flow rate, temperature (particularly of a fluid or solid component), rotational speed, valve position, vibration, and so on.

[0046] In this case, a first sensor A is configured to measure a first physical quantity Y1. The first sensor A returns a measurement XA, which is equal to the sum of the first quantity Y1 and a measurement bias ZA: XA=Y1+ZA.

[0047] A second sensor B is configured to measure a second physical quantity Y2. The second sensor B returns a measurement XB, which is equal to the sum of the second quantity Y2 and a measurement bias ZB: XB=Y2+ZB.

[0048] A third sensor C is configured to measure a third physical quantity Y3. The third sensor C returns a measurement XC, which is equal to the sum of the third quantity Y3 and a measurement bias ZC: XC = Y 3 + ZC .

[0049] In this example, the quantities Y1, Y2, Y3 are assumed to be different, but it is possible to plan for several sensors measuring the same quantity (possibly with a negligible offset).

[0050] Since the quantities Y1, Y2, and Y3 are different, the sensor array in this example comprises at most one sensor per physical quantity. Therefore, this embodiment is free of hardware redundancy.

[0051] In this example, the third quantity Y3 is considered a physical quantity that occurs upstream of the first quantity Y1: thus, given the physical laws governing the system, the first quantity Y1 is a consequence of the third quantity Y3. For example, the third quantity Y3 could be a pressure in a propellant tank, and the first quantity Y1 could be a flow rate in a pipe leading from this propellant tank.

[0052] The XA, XB, XC measurements can be one-off, possibly repeated over time, or extend over a certain period of time.

[0053] Furthermore, a plurality of estimators 20 can be provided, in addition to the sensor set 10, to monitor the system. In this case, the plurality of estimators 20 comprises two estimators D and E. However, any number of estimators, preferably two or more, could be included in the plurality of estimators 20. The physical quantities to be estimated in the system and the appropriate estimators for this purpose can be chosen by a person skilled in the art who is familiar with the system, according to their monitoring needs.

[0054] In this instance, a first estimator D is configured to determine an estimate XD of the first physical quantity Y1, which is equal to the sum of the first quantity Y1 and an estimation bias ZD: XD = Y1 + ZD. In this embodiment, the first estimator D can be a physical model, linear or nonlinear, or any other suitable estimator. To determine the estimate XD, the first estimator D takes as input a subset of the measurements returned by the sensors, in this case, the measurement XB returned by the second sensor B. However, the estimate XD could be determined based on more than one measurement.

[0055] A second estimator E is configured to determine an estimate XE of the third physical quantity Y3, which is equal to the sum of the third quantity Y3 and an estimation bias ZE: XE = Y3 + ZE. In this embodiment, the second estimator E includes an ANN. For example, the ANN can include one or more layers, notably of a structure called a Feedforward Network (FF or FFNN) or Multi-Layer Perceptron (MLP). The ANN can be trained on data from physical models and / or experimental data. To determine the estimate XE, the second estimator E takes as input a subset of the measurements returned by the sensors, in this case, the measurement XA returned by the first sensor A. Like the estimate XD, the estimate XE could be determined based on more than one measurement. Conversely, the second estimator E, and particularly its ANN, could return more than one estimate, that is, have multiple outputs.

[0056] Thus, estimators D and E take as input different subsets of the measurements XA, XB, XC returned by sensors A, B, C.

[0057] Furthermore, the plurality of estimators 20 therefore includes at least one estimator, here two estimators D and E, to provide an estimate of a physical quantity measured by one of the sensors, in this case of the first quantity Y1 measured by the first sensor A and the third quantity Y3 measured by the third sensor C, respectively.

[0058] This example is purely illustrative; the types of estimators can be adapted to suit the needs of a person skilled in the art, taking into account the present discussion. For example, the estimators may all include RNAs, or only some of them, provided that at least one of the estimators includes an RNA.

[0059] Note that the estimator E, which takes the measurement XA as input and determines the estimate XE, is such that the ANN it comprises is configured to determine the estimate XE of a physical quantity Y3 that occurs upstream of the physical quantity Y1, whose measurement XA is taken as input by the ANN. Thus, the estimator E, and more specifically its ANN, is configured to determine an estimate according to an inverse relationship.

[0060] This can be particularly useful for determining estimates of quantities for which exact redundancy is not possible. For example, a rotational speed (resultant of forces) is the result of the presence of torques (composed of pressure, temperature, and flow rate) applied to the turbine and the pump (inlet forces). It is possible to determine the pressure at the turbine inlet by knowing the speed and pressure at the pump outlet.

[0061] The data consists of the measurements XA, XB, XC and the estimates XD, XE. In this embodiment, the data is redundant for the first quantity Y1, evaluated using measurement XA and estimation XD, and for the third quantity Y3, evaluated using measurement XC and estimation XE. However, the second quantity Y2 is not redundant, as it is only measured by the second sensor B.

[0062] Based on this data, residuals are calculated (functional block 30). Residuals represent the difference between data points representing the same physical quantity. For example, the first quantity Y1 is represented by the data points XA (measurement) and XD (estimate). A residual R1 for the first quantity Y1 can be calculated as the difference between XA and XD: R 1 A D = XA − XD = Y 1 + ZA − Y 1 + ZD = ZA − ZD

[0063] Thus, in this example, the residual is equal to the difference in measurement or estimation biases. Under normal operating conditions, this difference should be very small, especially compared to Y1. If the residual R1 exceeds a certain threshold, then there is a fault in the first sensor A and / or the first estimator D, whose bias is too large.

[0064] Rather than a simple scalar difference, another type of deviation can be calculated as a residual, depending on the nature of the measurement XA and the estimation XD. For example, in the case of time-domain signals, the deviation can be represented by the integral of the absolute value or the square of the difference between the two signals. Other suitable deviations, including a distance in an appropriate space, can be developed by a person skilled in the art.

[0065] The residual R1(A,D) represents the difference between a measurement XA and an estimate XD. If other sensors F and / or other estimators G returned a measurement or an estimate of the quantity Y1, it would be possible to calculate other residuals, for example, by taking the data pairs: R1(A,F), R1(A,G), R1(D,F), R1(D,G), R1(F,G). Thus, we would obtain residuals representing the difference between two measurements (XA and XF), between a measurement and an estimate (XA and XD, XA and XG, XF and XD, XF and XG), or between two estimates (XD and XG). Optionally, the calculation of residuals can therefore be performed between all pairs of data representing the same physical quantity, in this case Y1.

[0066] If fault detection only targets the sensor(s) directly or indirectly related to quantity Y1, it is unnecessary to calculate residuals for other quantities. However, in this case, a more comprehensive detection is performed, so the residual R3 for the third quantity Y3 is also calculated. Thus, analogously: R3(C,E) = ZC - ZE.

[0067] Based on the residuals thus calculated, a determination is made whether there is a fault on at least one of the sensors A, B, C, according to the residuals R1(A,D) and R3(C,E) calculated from the data obtained from measurements returned by said sensor. Recall that the first estimator D takes as input the measurement XB; thus, the residual R1(A,D) is calculated from the data obtained from the measurements XA and XB returned by sensors A and B. Similarly, the second estimator E takes as input the measurement XA; thus, the residual R3(C,E) is calculated from the data obtained from the measurements XA and XC returned by sensors A and C.

[0068] It follows that if only residual R1 exceeds a threshold, while residual R3 does not, then it can reasonably be concluded that the second sensor B is defective. If only residual R3 exceeds a threshold, while residual R1 does not, then it can reasonably be concluded that the third sensor C is defective. If both residuals R1 and R3 exceed thresholds, then it can reasonably be concluded that the first sensor A, whose measurement XA is used in the calculation of both residuals, is defective.

[0069] This simple case clearly illustrates how estimators taking different subsets of measurements as input, particularly when evaluating the same physical quantity, can determine, based on the residuals, that at least one sensor has failed. Strictly speaking, high residuals R1 and R3 could be interpreted as indicating that sensors B and C, or even A, B, and C, are defective, although simultaneous failures are less likely than a single failure on the first sensor, A. However, increasing the number of sensors and estimators, compared to the present embodiment, allows for a more precise determination of the relationship between a residual exceeding a threshold and the location of a fault on one of the sensors.

[0070] Thus, in the real-world case of a rocket engine whose flight conditions are monitored by around fifty sensors measuring certain parameters from those listed previously, the present method reduces the number of hardware sensors to about thirty. The remaining sensors are replaced by one or more estimators without compromising fault detection, and therefore the overall reliability and accuracy of the system. This results in a significant reduction in mass and cost, without any loss of performance.

[0071] In the preceding discussion, the criterion described for identifying a sensor failure was a residual exceeding a threshold. However, other criteria are possible, such as a significant variation in the residual over time, etc. More generally, a determination transfer function 40 takes the residuals as input and determines that there is a failure on at least one of the sensors based on the residuals calculated from data obtained from measurements returned by that sensor.

[0072] As previously mentioned, the determination transfer function 40 can include at least one RNA that takes residuals as input and is configured to return, for at least one of the sensors, an indicator of the existence and optionally the severity of a fault on that sensor. If there are multiple RNAs, each RNA can be configured to identify a fault on a particular sensor. An example of RNA training will be described below.

[0073] Alternatively, the determination transfer function 40 can be a function determined analytically or numerically, for example a matrix taking as input a vector of residuals and returning a vector of indicators.

[0074] The determination 40 can provide the fault indication in the form of one or more indicators, or a vector of indicators 50. For example, with reference to the present embodiment, the indicator 50 can be a three-digit number, each digit corresponding to a sensor. For example, an indicator value of 010 indicates a minor fault on the second sensor B and no fault on the other two sensors, while an indicator value of 200 indicates a major fault on the first sensor A and no fault on the other two sensors. The indicators can be numeric, as in the present example, or take any other desired form (a signal, a letter, a shape, a color, etc.).

[0075] Optionally, the data XA to XE and the indicator 50 can be taken as inputs by a synthesizer 60 which performs the synthesis in order to take into account the faults translated by the indicator 50 to improve the estimation of the physical quantities Y1 to Y3.

[0076] In this embodiment, the synthesizer 60 is configured to produce an estimate of a physical quantity, for example Y1, from the first sensor A measuring said physical quantity and the first estimator D determining an estimate of said physical quantity, ignoring defective sensors for which a failure has been determined, and estimators taking as input measurements returned by said defective sensors. Thus, if the indicator 50 indicates that a failure has been detected on the first sensor A, then the synthesizer 60 can ignore the measurement XA and return the estimate XD as the estimate of the quantity Y1.

[0077] If indicator 50 indicates that a fault has been detected on the second sensor B, then it is likely that the estimate XD, determined by the first estimator D from the measurement XB, is bad; then, the synthesizer 60 can ignore the estimate XD and return the measurement XA as the estimate of the quantity Y1.

[0078] Finally, if indicator 50 shows that no fault has been detected, or that a fault has been detected on the third sensor C whose measurement XC does not affect the estimation of the first quantity Y1, then the synthesizer 60 can take into account both the measurement XA and the estimation XD to provide an estimate of the first quantity Y1. For example, the synthesizer 60 can average XA and XD. Other combinations are possible, depending on the choice of a person skilled in the art. Furthermore, the data can be weighted in the production of the estimate, for example, according to the degree of fault and / or the level of accuracy associated with each sensor / estimator.

[0079] A method for learning an artificial neural network for a fault determination system for a set of sensors is now described, according to one embodiment, with reference to the figure 2. Such an RNA can be used for determination 40 presented previously.

[0080] The learning process includes a step S70 for obtaining data representing a plurality of measurements returned respectively by the sensors. The sensors can be sensors A, B, and C described previously. It is preferable that the sensors be the same as those on which the RNA will be used to determine faults, in order to account for their inherent defects.

[0081] The data may directly include XA, XB, XC measurements returned by sensors A, B, C, for example during system tests or recordings from previous missions, as illustrated in the figure 2 , and / or simulated data reproducing measurements returned by the sensors. For brevity, we will refer to both cases as measurements, whether acquired, simulated, or obtained in another way.

[0082] This data is passed to step S72, which generates faults on the measurements. In this step, for example, the data obtained in step S70 is replicated, and on each copy, a known fault of a given amplitude is generated on one or more specific sensors. The choice of the fault amplitude and / or the sensors involved can be random, in order to quickly cover a large number of faults on a large number of sensors. The faults can target a single sensor or a combination of sensors, in order to simulate simultaneous failures.

[0083] Thus, from measurements XA, XB, XC, we can generate for example two copies with defects: XA', XB, XC, with a slight defect on the first sensor A, and XA, XB", XC', with a serious defect on the second sensor B and a slight defect on the third sensor C.

[0084] The data samples, after error generation, are passed to estimators D and E, which, for each data sample, determine estimates of the quantities based on the potentially flawed measurements. These estimates are then combined with representative measurement data from which they were derived to form complete datasets. Thus, based on the preceding example and the previously described estimators D and E, three datasets are generated: (XA, XB, XC, XD, XE), (XA', XB, XC, XD, XE'), and (XA, XB", XC', XD", XE). The first dataset represents normal system operation, where all sensors function normally; therefore, the estimators are expected to produce accurate estimates. In the second dataset, a slight error is expected to affect the XE' estimate due to the slight error introduced into the input measurement XA' by the second estimator, E.If the second estimator E is not linear, the defect can of course be amplified or diminished. Similarly, but independently, in the third dataset, a severe defect is expected to affect the estimate XD, due to the severe defect introduced into the input measure XB by the first estimator D. The minor defect in the measure XC does not affect any estimator in this example.

[0085] The data, including the measurements with defects generated in step S72 and the estimates with defects determined in step S74, are then used to calculate residuals in step S76. The residuals are calculated as described previously, for each dataset. This yields, in the previous example, three sets of residuals: (R1, R3), (R1', R3'), and (R1", R3").

[0086] In parallel, at step S78, independently of the estimates and residuals, fault indicators are generated on the sensors, corresponding to the faults generated on the measurements in step S72. Indeed, since the faults are known, it is possible to characterize the potential fault on each sensor with the format that we want the RNA to provide as output.

[0087] For example, with three-digit numeric indicators as described above, the first dataset would correspond to an indicator 000 (no fault on any sensor), the second dataset to an indicator 100 (slight fault on the first sensor A, none on the other sensors) and the third dataset to an indicator 021 (no fault on the first sensor A, serious fault on the second sensor B, slight fault on the third sensor C).

[0088] In particular, generating a fault indicator for a sensor (for example, a digit from the aforementioned indicators) may involve comparing the corresponding fault with a threshold associated with the sensor returning the measurement in question, and issuing a fault severity level for said sensor based on this comparison. In the example above, the severity level is 1 if the fault has an amplitude between 2% and 5%, and 2 if the fault has an amplitude greater than 5%. If the fault has an amplitude less than 2%, it is considered that there is no fault (the severity level is 0, which is also an indicator of the absence of a fault).

[0089] Of course, these thresholds are illustrative and can be modified, in number and value, depending on the accuracy and reliability of the sensors considered, and this for each sensor independently. Typically, instead of the two thresholds described above, one can use a single threshold (presence or absence of fault) or three or more thresholds.

[0090] Next, in step S80, the RNA is trained using the residuals calculated in step S76 as input and the fault indicators generated in step S78 as output target. Based on this, the training can be performed in a conventional manner to make the RNA capable, as illustrated in the... figure 1, to return indicators 50 having taken as input the residuals calculated in functional block 30: for each dataset, iteratively, the RNA takes as input the residuals, estimates at least one indicator and corrects its internal weights based on the comparison between the estimated indicator and the indicator provided as the output target.

[0091] Again, the example described above was based on a small number of datasets to clarify the illustration, but the RNA learning will be all the more refined as the datasets allow for a fine meshing of the system's operating domain and the faults that may occur.

[0092] Although the present description refers to specific embodiments, modifications may be made to these examples without departing from the general scope of the invention as defined by the claims.

Claims

1. A method for determining a fault for a set of sensors (10), comprising: - obtaining data, said data comprising a plurality of measurements returned respectively by the sensors (A, B, C) and a plurality of estimates determined respectively by a plurality of estimators (D, E) taking as input different subsets of said measurements, at least one of the estimators (E) comprising an artificial neural network; - calculating residuals (30) representative of the deviation between data representing a same physical quantity, wherein at least some of the residuals are representative of the deviation between a measurement and an estimate, or between two estimates, at least one of the estimators (E) that determines the estimates comprising an artificial neural network; - determining (40) that there is a fault on at least one of the sensors (A, B, C) as a function of the residuals calculated based on data obtained from measurements returned by said sensor.

2. The fault determination method according to claim 1, wherein the calculation of residuals is performed between all the pairs of data representing the same physical quantity.

3. The fault determination method according to any one of claims 1 to 2, wherein the set of sensors (10) comprises at most one sensor (A, B, C) per physical quantity.

4. The fault determination method according to any one of claims 1 to 3, wherein the plurality of estimators comprises at least one estimator (D, E) to provide an estimate of a physical quantity measured by one of the sensors (A, C).

5. The fault determination method according to any one of claims 1 to 4, wherein the artificial neural network is configured to determine an estimate of a first physical quantity occurring upstream of at least one second physical quantity whose measurement is taken as input by the artificial neural network.

6. The fault determination method according to any one of claims 1 to 5, wherein, in the determination (40) that a fault exists, the residuals are given as input to at least one artificial neural network configured to return, for at least one of the sensors (A, B, C), an indicator (50) of the existence and optionally of the severity of a fault on said sensor.

7. A method for estimating a physical quantity from a set of sensors, comprising the fault determination for said set of sensors according to any one of claims 1 to 6, and the production (60) of an estimate of the physical quantity from the sensors (A, B, C) measuring said physical quantity and the estimators (D, E) determining an estimate of said physical quantity, while ignoring the defective sensors for which the existence of a fault has been determined, and the estimators taking as input measurements returned by said defective sensors.

8. A fault determination system for a set (10) of sensors (A, B, C), comprising: - a data obtaining unit, configured to obtain data comprising a plurality of measurements returned respectively by the sensors (A, B, C) and a plurality of estimates determined respectively by a plurality of estimators (D, E) taking as input different subsets of said measurements, at least one of the estimators (E) comprising an artificial neural network; - a unit for calculating residuals (30) representative of the deviation between data representing a same physical quantity, wherein at least some of the residuals are representative of the deviation between a measurement and an estimate, or between two estimates, at least one of the estimators (E) that determines the estimates comprising an artificial neural network; - a determination unit (40) configured to determine that there is a fault on at least one of the sensors (A, B, C) as a function of the residuals calculated based on data obtained from measurements returned by said sensor.

9. A method for training an artificial neural network (40) for a fault determination system for a set of sensors (A, B, C), comprising: - obtaining (S70) data representing a plurality of measurements returned respectively by the sensors (A, B, C); - generating (S72) defects on the measurements; - generating fault indicators (S78) on the sensors, corresponding to the defects; - adding (S74), to the data, a plurality of estimates determined respectively by a plurality of estimators (D, E) taking as input different subsets of the measurements, after generation of the defects, wherein at least one of the estimators (E) comprises an artificial neural network; - calculating (S76) residuals representative of the deviation between data representing a same physical quantity; - training (S80) the artificial neural network (40) taking as input the residuals and as output target the fault indicators.

10. The training method according to claim 9, wherein the generation of a fault indicator (S78) comprises the comparison of the corresponding defect with a threshold associated with the sensor returning the measurement considered, and the emission of a fault severity level for said sensor based on said comparison.