Abnormality determination device

Abstract

This abnormality determination device is an abnormality determination device to which n (n is a natural number of 3 or more) input signals are input, and which is provided with n automatic encoders configured so as to generate an input signal group of n-1 input signals from which one of the input signals has been removed, and to generate an input signal group of n-1 input signals from which one of the input signals has been removed. An output unit that outputs an output value group corresponding to the input signal group, and that has a comparison unit that calculates a difference value between the value of the input signal group and the output value group for each of the n autoencoders, and compares the difference value with a predetermined threshold value; and an abnormality determination unit that determines an abnormal portion of the plurality of input signals on the basis of a result of the comparison.

Description

TECHNICAL FIELD

[0001] The present application relates to an abnormality discriminating device that discriminates an abnormality of a signal input to a control device. BACKGROUND

[0002] In order to seek high-precision powertrain control and achieve development efficiency of powertrain control constituted by existing MAP control, control development using a multi-input calculation model such as a neural network model is being promoted.

[0003] PRIOR ART DOCUMENTS

[0004] PATENT DOCUMENTS

[0005] Patent Document 1: Japanese Patent Laid-Open No. 2020-045773 SUMMARY

[0006] PROBLEMS TO BE SOLVED BY THE INVENTION

[0007] Since a case where an abnormal input signal is input to the calculation model becomes a main cause of deterioration of calculation accuracy, it is necessary to detect the abnormal input signal. However, since the number of input signals in the case of multi-input is large, and depending on the model, it becomes complicated as various signal forms such as a sensor value, a control value from an ECU, a presumed value output from an upstream, and the like, there is a possibility that detection or determination of the abnormal input signal becomes difficult.

[0008] To such a problem, Patent Document 1 deals with an abnormal signal by excluding an abnormal input signal and performing relearning in a case where an abnormal input value is input to a neural network model. However, in the above-described Patent Document 1, a countermeasure in a case where an abnormal input value is input at the time of actuator or sensor failure is shown, detailed detection of the abnormal input signal is not described, various signal forms such as a sensor value, a control value from an ECU, a presumed value output from an upstream, and the like cannot be dealt with, and determination of an input signal that becomes a cause of error is not disclosed in Patent Document 1.

[0009] An object of the present application is to provide an abnormality discriminating device that can determine an input signal that becomes a cause of estimation error of a calculation model, maintain reliability of a calculation value of the model, and prevent deterioration of estimation accuracy due to abnormality of an input value of the model.

[0010] MEANS FOR SOLVING THE PROBLEMS

[0011] To achieve the above object, the abnormality discrimination device of the present application is an abnormality discrimination device to which n (n is a natural number of 3 or more) input signals are input, which includes n autoencoders configured to output an output value group corresponding to an input signal group when the input signal group of n-1 input signals excluding one input signal among the input signals is input, and a comparison section configured to calculate a difference value between the input signal group and the output value group for each of the n autoencoders, and compare the difference value with a predetermined threshold value, and an abnormality discrimination section configured to discriminate an abnormal signal included in the plurality of input signals based on a result of the comparison.

[0012] Effects of the Invention

[0013] According to the present application, it is possible to prevent deterioration of estimation accuracy caused by abnormality of input values of a model.

[0014] Other features related to the present application will become apparent from the description and drawings of the present application. In addition, the objects, configurations, and effects other than those described above will be clear from the following description of the embodiments. BRIEF DESCRIPTION OF DRAWINGS

[0015] Figure 1 is a block diagram showing a hardware configuration example of an ECU in which an abnormality discrimination device of an embodiment of the present application is mounted.

[0016] Figure 2 is a diagram showing an example of a neural network model applied in an embodiment of the present application.

[0017] Figure 3 is a diagram showing an example of an autoencoder applied in an embodiment of the present application.

[0018] Figure 4A is a block diagram showing a schematic configuration of an abnormality discrimination device of Embodiment 1 of the present application.

[0019] Figure 4B is a block diagram showing a detailed configuration of the abnormality discrimination device of Embodiment 1 of the present application.

[0020] Figure 5 is a flowchart showing a process performed by the abnormality discrimination device of Embodiment 1 of the present application.

[0021] Figure 6 is a block diagram showing a detailed configuration of the abnormality discrimination device of Embodiment 2 of the present application.

[0022] Figure 7 is a flowchart showing a process performed by the abnormality discrimination device of Embodiment 2 of the present application. DETAILED DESCRIPTION

[0023] Hereinafter, a mode for carrying out the present application will be described with reference to the drawings. In the present specification and the drawings, the same symbols are assigned to constituent elements having substantially the same functions or configurations, and overlapping descriptions will be omitted.

[0024] [Example 1]

[0025] [Example of hardware configuration of ECU]

[0026] Figure 1 is a block diagram showing an example of a hardware configuration of an ECU (Electric Control Unit) 100 including the abnormality discrimination device of Example 1.

[0027] The ECU 100 includes an input circuit 101, an input / output port 102, a RAM 103, a ROM 104, a CPU 105, a throttle valve drive circuit 106, a fuel injection valve drive circuit 107, and an ignition output circuit 108.

[0028] Output signals from various sensors are input to the input circuit 101. In Figure 1 , an example is shown in which output signals of a throttle valve sensor 1, a flow rate sensor 2, a supercharging pressure sensor 3, and the like are input to the input circuit 101 of the ECU. The signals input to the input circuit 101 are sent to the input / output port 102.

[0029] The signals sent to the input / output port 102 are stored in the RAM 103, and arithmetic processing is performed by the CPU 105. A control program describing the content of the arithmetic processing is written in advance in the ROM 104, and is executed by the CPU 105. The ROM 104 records programs, data, and the like necessary for the CPU 105 to operate, and functions as an example of a computer-readable non-transitory storage medium that stores the programs executed by the ECU.

[0030] The CPU 105 is the abnormality discrimination device in the present embodiment. Control signals calculated by the CPU 105 according to the control program are output to each device such as the throttle valve drive circuit 106, the fuel injection valve drive circuit 107, and the ignition output circuit 108. The throttle valve drive circuit 106 outputs a drive signal for controlling the opening / closing drive of the throttle valve to the throttle valve.

[0031] The fuel injection valve drive circuit 107 outputs a drive signal for controlling the opening / closing drive of the fuel injection valve at the fuel injection timing to the fuel injection valve. The ignition output circuit 108 outputs a drive signal for controlling the ignition of the spark plug to the ignition output circuit 108. An ECU is shown in the present embodiment, but other controllers such as a VCM can also be applied.

[0032] [Neural network model]

[0033] Figure 2 A summary of a neural network is shown. A neural network model is a mathematical model that simulates the structure of a human brain neural circuit, and weights and biases are set for each neuron constituting the model.

[0034] In addition, a function called an activation function is defined in the neuron. A logical function and a slope function, etc. are appropriately set in the activation function. One layer is formed by a plurality of neurons, and an intermediate layer is provided between the input layer and the output layer. By increasing the number of neurons and the number of layers of the intermediate layer, a more complex input-output relationship can be approximated. There is a trade-off relationship between the approximation accuracy and the model size, and a two-point that satisfies the requirements of both is selected. For example, by setting the frequency of the exhaust gas sensor signal on the input layer and setting the catalyst deterioration diagnosis on the output layer, and performing machine learning (supervised) on the weights and biases of each neuron, the input-output relationship can be approximated. In the machine learning algorithm, the error backpropagation method can be applied.

[0035] <Autoencoder>

[0036] Figure 3 A summary of an autoencoder (self-encoder) is shown. The autoencoder is one of the operators that learns by using a neural network unsupervised learning method. The autoencoder has a structure of an encoder that reduces the dimension of input data and a decoder that returns the original data again. As shown in Figure 3 , learning is performed so that the input and output of the autoencoder are the same. Since learning is performed by normal data, in the case where an abnormal input occurs, the input value and the output value are not equal, and abnormality detection can be performed. In addition, the input value referred to here means the value of the input signal.

[0037] <Configuration diagram>

[0038] Next, the configuration of the abnormality discrimination device (CPU 105) of the present embodiment will be described using Figure 4A and Figure 4B . Figure 4A A summary diagram of the abnormality discrimination device 105 is shown. The input signal is input from 101 and 102 of Figure 1 . The input signal can also be input from the RAM 103. The input signal is input to the abnormality determination section 440 constituted by the model 401 in the CPU 105, and the autoencoder 402 and the comparator 403 described in Figure 4B . In the case where it is determined that the input signal contains an abnormality, the input signal that becomes abnormal is determined by the abnormal signal determiner 400, and the final output is input to the RAM 103.

[0039] Figure 4BThis diagram shows the detailed configuration of the anomaly detection device. In this embodiment, it is assumed that only one of the multiple input signals is abnormal, or that no abnormal signal is present. As a first step, multiple input signals are input to model 401. These input signals include control values, sensor values, and estimated values ​​from other models. The multiple input signals are then input to a first-step autoencoder 402. This first-step autoencoder 402 is an autoencoder that has been learned using known inputs. Then, a comparator 403 compares the input and output values ​​of the first-step autoencoder 402. If the input and output values ​​are inconsistent or the difference exceeds a certain threshold, the anomaly signal determiner 400 is activated, proceeding to the second step. As a second step, the (n-1) input signals are input to an autoencoder 404 that excludes one of the n input signals. Furthermore, the n input signals input to the first-step autoencoder 402, which are used to determine whether an abnormal signal is present, are called the diagnostic input signal group, and the corresponding output is called the diagnostic output value. In addition, the (n-1) input signals that are input to the automatic encoders 404-406 in the second step, after removing one of the n input signals, are called the input signal group, and the corresponding output is called the output value.

[0040] exist Figure 4B In this process, two of the three input signals are input to the auto encoder 404. The auto encoder 404 outputs an output value based on the input signals. Then, comparator 407 compares the input signal value and the output value to determine whether the difference value is above a certain threshold. If it is above the threshold, the output is 1. Additionally, using the auto encoder 405 which excludes one other input signal, comparator 408 compares the input signal value and the output value to determine whether the difference value is above a certain threshold. If it is above the threshold, the output is 2. Finally, excluding the third input signal, the auto encoder 406 is used, and comparator 409 compares the input signal value and the output value to determine whether the difference value is above a certain threshold. If it is above the threshold, the output is 3.

[0041] In the case where the abnormal signal is excluded, since the input 2 signals are normal, the difference between the input value and the output value becomes below a certain threshold value. Therefore, the signal excluded at this time can be determined as an abnormal signal. Then, step 3 is entered. The adder 410 adds the output results of the comparators 407, 408, and 409. As an example, in the case where the input signal c is an abnormal signal, the comparator 407 outputs 1, the comparators 408 and 409 output 0, and therefore the adder 410 outputs 1. In the case where the input signal b is an abnormal signal, the comparator 408 outputs 2, the comparators 407 and 409 output 0, and therefore the adder 410 outputs 2. In the case where the input signal a is an abnormal signal, the comparator 409 outputs 3, the comparators 407 and 408 output 0, and therefore the adder 410 outputs 3.

[0042] Then, the output signal of the model group 411 is output by the plurality of switching sections 412, and the model group 411 changes the abnormal signal appropriate to the value output from the adder 410 to a substitute value. This model group functions as a substitute signal replacement section that replaces the abnormal signal with a substitute signal. As an example, in the case where the input signal c is abnormal, the output value of the model in which the input signal c' is output is changed. The substitute value uses a signal highly correlated with the abnormal signal or a signal physically close thereto. By confirming that the output value of the model changed to this substitute value is a correct value, the correctness of the model can be ensured. Finally, the final value is output by the switching section 413. In the case where all the input signals input to the autoencoder 402 in the first step are normal, 1 is output from the comparator 403, and therefore the output value of the model 401 is valid, and the abnormal signal determiner 400 does not operate.

[0043] On the other hand, in the case where the input signal input to the autoencoder 402 includes an abnormal signal, 0 is output from the comparator 403, and therefore the output value of the plurality of switching sections 412 is valid. The autoencoder is a manner in which learning is performed before shipment, but can also have a function of updating the learning value by connecting to a network after shipment. In addition, the respective threshold values are in the case where the number of input signals is 3, and can vary depending on the number of input signals.

[0044] <flowchart>

[0045] Figure 5 A flowchart showing the processing performed by the abnormality determination device 105 of Embodiment 1. In step 501, a plurality of all input signals identical to the input signals input to the model 401 are input to the autoencoder 402. In step 502, the input and output values of the autoencoder 402 are compared using the comparator 403, and in the case where the difference between the input value and the output value is above a certain threshold value, step 503 is entered. In step 503, the two signals excluding one of the three input signals are input to the autoencoder 404 as input.

[0046] Further, signals other than the excluded signal are excluded. In Embodiment 1, three autoencoders are used. In step 504, the comparator 407 determines whether the difference between the input value and the output value of the autoencoder 402 is below a threshold value. The processing of steps 503, 504 is performed in parallel for the other autoencoders 405, 406 and the comparators 408, 409. In step 505, the adder 410 calculates the conversion value of the output values of the comparators 407, 408, 409.

[0047] The comparators 407, 408, 409 output 0 when the input and output values are identical or the difference is below a certain threshold value, and thus by addition, output 1 when the abnormal signal is c, output 2 when the abnormal signal is b, and output 3 when the abnormal signal is a. In step 506, the output value of the model in which the abnormal value conforming to the added value is replaced by the alternative value is output.

[0048] As described above, according to the present embodiment, even when an abnormality is included in the input signal, the correctness of the model can be confirmed by determining the input signal of the abnormality, and thus the reliability of the model can be maintained without loss.

[0049] [Embodiment 2]

[0050] In Embodiment 2, the configuration when two or more abnormal signals are included in the input signal is described.

[0051] Figure 6 The detailed configuration of the abnormality determination device 105 of the present embodiment is shown. As a first step, a plurality of input signals are input to the model 601. The input signals include control values, sensor values, estimated values of other models, and the like. The plurality of input signals are input to the autoencoder 602. The autoencoder 602 is an autoencoder that has been learned using the input. Then, the input value and the output value of the autoencoder 602 are compared using the comparator 603, and when the difference between the input value and the output value is above a certain threshold value, the abnormal signal determiner 600 is activated, and a second step is entered.

[0052] As the second step, the autoencoder 604 that inputs (n-1) input signals other than one of the input signals is input with the (n-1) signals. In the autoencoder 604, the input signal is excluded from the input, and thus the input signal is not input to the autoencoder 604. Figure 6In the example shown in FIG. 6, two of the three input signals are input to the autoencoder 604. The autoencoder 604 outputs an output value from the input value. Then, the comparator 607 calculates a difference between the input value and the output value, and first makes a determination as to whether the difference is below a threshold value 1. In the case where the difference is below a certain threshold value 1, the comparator 607 outputs 1. Further, in the case where the signals a, b input to the autoencoder 604 are abnormal values, since the difference between the input value and the output value becomes larger, a comparison of the difference with a threshold value 2 is performed. When the difference is above the threshold value 2, the comparator 607 outputs 4. In the case where the threshold value 1 is above and less than the threshold value 2, 0 is output. The relationship between the threshold value 1 and the threshold value 2 is threshold value 2 > threshold value 1.

[0053] Further, the autoencoders 605, 606 excluding one other input signal are also input with the combination of the input signals. Also, the comparators 608, 609 perform a comparison of the difference between the input value and the output value of the autoencoder with the threshold values 1, 2, similarly to the comparator 607. In the case where the difference is above the threshold value 2, the comparator 608 outputs 5, in the case where the threshold value 1 is above and less than the threshold value 2, the comparator 608 outputs 0, and in the case where the threshold value 1 is below, the comparator 608 outputs 2. In the case where the difference is above the threshold value 2, the comparator 609 outputs 6, in the case where the threshold value 1 is above and less than the threshold value 2, the comparator 609 outputs 0, and in the case where the threshold value 1 is below, the comparator 609 outputs 3.

[0054] Next, by the abnormal signal number detection section 610, a determination is made as to whether all of the input signals are abnormal or whether at least one or more normal signals are input. In the case where the total of the output values of the comparators is 15, all of the input signals a, b, c are abnormal signals. In the case where it is not 15, it means that at least one normal signal is input. Figure 6

[0055] For example, in the case where the signal c is abnormal, the difference between the input and the output of the autoencoder 604 which is a combination not including the signal c becomes below the threshold value 1, and the comparator 607 outputs 1. On the other hand, since a part is an abnormal signal, the difference of the autoencoders 605, 606 becomes above the threshold value 1 and less than the threshold value 2, and 0 is output. Further, in the case where the abnormal signals are a and b, the difference between the input and the output of the autoencoder 604 becomes above the threshold value 2, and the comparator 607 outputs 4. Since a part is an abnormal signal, the difference of the autoencoders 605, 606 becomes above the threshold value 1 and less than the threshold value 2, and 0 is output. Further, in the case where all of the abnormal signals, the difference between the input and the output of the autoencoders 604, 605, 606 is above the threshold value 2, and the outputs of the comparators 607, 608, 609 are 4, 5, 6, respectively, and the total is 15. ​

[0056] In a case where at least one normal signal is input in the input signals, the maximum value of the output values of the comparators is output by the maximum value determination section 611. In a case where the maximum value of the output values of the maximum value determination section 611 is 4, 5, 6, it means that two abnormal signals are input. In contrast, in a case where the maximum value of the output values of the maximum value determination section 611 is 3, 2, 1, it means that one abnormal signal is input. Then, the model in which the abnormal signal is changed to the substitute value is changed by the switching sections 613, 614 according to the number of inputs, and the output value is output. By the comparator 612 and the switching section 615, in a case of 4, 5, 6, the output value of the switching section 613 is valid. This is also the same as in Embodiment 1, and is used to ensure the correctness of the model.

[0057] In a case where the maximum value of the output values of the maximum value determination section 611 is 3, 2, 1 or less, the output value of the switching section 614 is valid. Also, in the switching section 616, in a case where all of the input signals are abnormal, the sum of the output values of the comparators 607, 608, 609 is 15, the output of the model 7 is valid, and in a case where the sum is not 15 and a normal signal is included, the output value of the switching section 615 is valid. Finally, by the switching section 617, in a case where the input of the model 601 includes an abnormal signal, the output value of the switching section 616 is valid, and in a normal case, the output value of the model 601 is valid, and the action of the abnormal signal determiner 600 is not performed.

[0058] The autoencoder is a manner in which learning is performed before shipment, but can have a function of updating the learning value by connecting to a network after shipment. The threshold values are in a case where the number of input signals is 3, and can be changed according to the number of input signals.

[0059] <Flowchart>

[0060] Figure 7 A flowchart showing the processing performed by the abnormality determination device of Embodiment 2. In step 701, all of the inputs of the model 601 are input to the autoencoder 602. In step 702, the input and output values of the autoencoder 602 are compared using the comparator 603, and in a case where the difference between the input and output values is a certain threshold value or more, step 703 is entered.

[0061] In step 703, the two input signals excluding one of the three input signals are input to the auto-encoder 604 as inputs. Further, the signal other than the excluded signal is excluded. In Embodiment 2, three auto-encoders are used. In step 704, the comparator determines whether the difference between the input value and the output value of the auto-encoder is equal to or less than a threshold value 1. In the case where the difference is equal to or less than the threshold value 1, the signal indicating an anomaly is one, and the process proceeds to step 705, and the output of the model 1 to 3 in which the one abnormal signal is replaced with a substitute value becomes valid.

[0062] In the case where the difference is greater than the threshold value 1, it is indicated that there is a difference value equal to or greater than a threshold value 2. In step 706, the comparator determines whether the sum of the output values of the respective comparators is equal to a predetermined value. In Embodiment 2, 15 is set. In the case where they are equal, all of the signals are considered to be abnormal, and the process proceeds to step 707, and the output value of the model 7 in which all of the signals are replaced with substitute values becomes valid. If they are not equal, the process proceeds to step 708, and any one of the outputs of the models 4 to 6 in which two abnormal signals are replaced becomes valid.

[0063] According to the above-described embodiment of the present application, even in the case where two abnormal signals are included in the input signals, the abnormal input signal can be determined as in Embodiment 1, and the accuracy of the model is not lost.

[0064] According to the above-described embodiment of the present application, the following effects can be obtained.

[0065] (1) The abnormality determination device of the present application is an abnormality determination device to which n (n is a natural number of 3 or more) input signals are input, and includes n auto-encoders configured to output an output value corresponding to an input signal group of n-1 input signals excluding one of the input signals when the input signal group is input, and a comparison section configured to calculate a difference value between the input signal group and the output value for each of the n auto-encoders, compare the difference value with a predetermined threshold value, and an abnormality determination section configured to determine an abnormal signal included in the plurality of input signals based on the comparison result.

[0066] With the above configuration, the input signal that is a cause of estimation error of the calculation model can be determined, the reliability of the calculation value of the model can be maintained, and deterioration of estimation accuracy caused by an abnormality of the input value of the model can be prevented.

[0067] (2) Further provided is a first step autoencoder that outputs a diagnosed output value corresponding to a diagnosed input signal group composed of n input signals when the input is the diagnosed input signal group, and performs comparison of the diagnosed input signal group and the diagnosed output value group, and in a case where a difference value between a value of the diagnosed input signal group and a value of the diagnosed output value is within a prescribed threshold value, it is presumed that one of the n input signals contains an abnormal signal, and in a case where it is presumed by the first step autoencoder that one of the n input signals contains an abnormal signal, it performs abnormality discrimination of the plurality of input signals based on the comparison results from the n autoencoders. The presence of such a first step autoencoder guarantees the effect of (1).

[0068] (3) The abnormality discrimination section discriminates the one input signal removed in the autoencoder in which the difference value is determined to be the smallest as an abnormal signal. This is the method of abnormality discrimination in Embodiment 1.

[0069] (4) The abnormality discrimination section discriminates the one input signal removed in each of the plurality of autoencoders in which the difference value is determined to be smaller than the prescribed threshold value as an abnormal signal. This is the method of abnormality discrimination in Embodiment 2.

[0070] (5) Further provided is a substitute signal replacement section that replaces the one input signal determined to be an abnormal signal with a substitute signal obtained by a method different from a prescribed calculation method for the one input signal. Thereby, by replacing the abnormal signal with another normal signal and outputting it and confirming that the output is correct, it is possible to confirm the correctness of the model.

[0071] Explanation of symbols

[0072] 105...CPU (abnormality discrimination device), 402...autoencoder, 403...comparator (comparison section), 404-406...autoencoder, 407-409...comparator (comparison section), 410...adder (abnormality discrimination section), 411...model group (substitute signal replacement section).

Claims

1. An anomaly detection device, which is an anomaly detection device that receives n input signals, where n is a natural number greater than or equal to 3, characterized in that, The system has n automatic encoders, each configured to output a value corresponding to one of the n-1 input signal groups (excluding one of the input signals) when the input signal is removed. It also includes: a comparison unit, which calculates the difference between the value of the input signal group and the output value for each of the n automatic encoders, and compares the difference with a predetermined threshold. And an anomaly detection unit, which, based on the comparison results, determines the abnormal signals contained in the plurality of input signals.

2. The anomaly detection device according to claim 1, characterized in that, It also includes a first-step automatic encoder. When a diagnostic input signal group consisting of the n input signals is input, the first-step automatic encoder outputs a diagnostic output value corresponding to the diagnostic input signal group, and performs a comparison between the diagnostic input signal group and the diagnostic output value. If the difference between the value of the diagnostic input signal group and the value of the diagnostic output value is within a predetermined threshold, it is presumed that one of the n input signals contains an abnormal signal. If the autoencoder in the first step determines that one of the n input signals contains an anomalous signal, anomaly detection of the plurality of input signals is performed based on the result of the comparison from the n autoencoders.

3. The anomaly detection device according to claim 1, characterized in that, The anomaly detection unit will identify the input signal that is removed from the automatic encoder that is determined to have the smallest differential value as the anomaly signal.

4. The anomaly detection device according to claim 1, characterized in that, The anomaly detection unit will identify the input signal that is removed from the plurality of automatic encoders whose difference value is determined to be less than a specified threshold as the anomaly signal.

5. The anomaly detection device according to claim 3, characterized in that, It also includes a substitution signal replacement unit, which calculates a substitution signal by a method different from the predetermined calculation method of the input signal that is judged as an abnormal signal, and replaces the input signal with the substitution signal.

Images