Abnormality determination device

Abstract

An abnormality determination device according to the present invention, to which n (n is a natural number at least equal to 3) input signals are input, is provided with n autoencoders configured to output, when a group of n-1 input signals obtained by excluding one input signal from among the n input signals are input, a group of output values corresponding to the group of input signals. The abnormality determination device comprises: a comparison unit that calculates, for each of the n autoencoders, difference values between the values of the group of input signals and the group of output values, and compares the difference values with a prescribed threshold value; and an abnormality determination unit that determines abnormal portions of a plurality of input signals on the basis of the results of the comparison.

Description

Abnormality determination device

[0001] The present invention relates to an abnormality detection device that detects an abnormality in a signal input to a control device.

[0002] Highly accurate powertrain control is required, and in order to improve the development efficiency of powertrain control that has traditionally been based on MAP control, control development using multi-input calculation models such as neural network models is progressing.

[0003] Japanese Patent Application Laid-Open No. 2020-045773

[0004] When an abnormal input signal is input to a calculation model, it can cause a deterioration in calculation accuracy, so it is necessary to detect the abnormal input signal.However, in the case of multiple inputs, there are many input signals, and depending on the model, the signals can take various forms such as sensor values, control values ​​from the ECU, and estimated values ​​output from upstream, making it complicated, and there is a risk that detecting and identifying the abnormal input signal may become difficult.

[0005] To address this issue, Patent Document 1 addresses the problem by excluding the abnormal input signal and re-learning when an abnormal input value is input to a neural network model. However, Patent Document 1 above only describes a method for addressing the case where an abnormal input value is input due to an actuator or sensor failure, but does not describe detailed detection of abnormal input signals, and is unable to address various signal forms such as sensor values, control values ​​from an ECU, and estimated values ​​output from upstream, and does not disclose how to identify input signals that cause errors.

[0006] An object of the present invention is to provide an anomaly detection device that can identify input signals that cause estimation errors in a calculation model, maintain the reliability of the model's calculated values, and prevent a deterioration in estimation accuracy due to abnormalities in the model's input values.

[0007] In order to achieve the above object, the abnormality discrimination device of the present invention is an abnormality discrimination device that receives n (n is a natural number greater than or equal to 3) input signals, and is equipped with n autoencoders that are configured to output a group of output values ​​corresponding to a group of n-1 input signals in which one input signal has been removed from the input signals, and has a comparison unit that calculates a difference value between the group of input signals and the group of output values ​​for each of the n autoencoders and compares the difference value with a predetermined threshold value, and an abnormality discrimination unit that discriminates abnormal signals contained in the multiple input signals based on the results of the comparison.

[0008] According to the present invention, it is possible to prevent a deterioration in estimation accuracy due to an abnormality in the input value of the model. Further features related to the present invention will become apparent from the description of this specification and the accompanying drawings. In addition, problems, configurations, and effects other than those described above will become apparent from the following description of the embodiments.

[0009] FIG. 1 is a block diagram showing an example of the hardware configuration of an ECU equipped with an abnormality determination device according to an embodiment of the present invention. FIG. 2 is a diagram showing an example of a neural network model applied in an embodiment of the present invention. FIG. 3 is a diagram showing an example of an autoencoder applied in an embodiment of the present invention. FIG. 4 is a block diagram showing a general configuration of an abnormality determination device according to a first embodiment of the present invention. FIG. 5 is a block diagram showing a detailed configuration of an abnormality determination device according to a first embodiment of the present invention. FIG. 6 is a flowchart showing processing executed by an abnormality determination device according to a first embodiment of the present invention. FIG. 7 is a block diagram showing a detailed configuration of an abnormality determination device according to a second embodiment of the present invention.

[0010] Hereinafter, embodiments of the present invention will be described with reference to the accompanying drawings. In this specification and drawings, components having substantially the same functions or configurations are designated by the same reference numerals, and redundant description will be omitted.

[0011] 1 is a block diagram showing an example of the hardware configuration of an ECU (Electronic Control Unit) 100 including an abnormality determination device according to Example 1. 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.

[0012] Output signals from various sensors are input to the input circuit 101. In Fig. 1, an example is shown in which output signals from a throttle sensor 1, a flow rate sensor 2, a boost pressure sensor 3, etc. are input to the input circuit 101 of the ECU. The signals input to the input circuit 101 are sent to an input / output port 102.

[0013] Signals transmitted to the input / output port 102 are stored in the RAM 103 and are processed by the CPU 105. A control program describing the details of the processing is written in advance in the ROM 104 and executed by the CPU 105. The ROM 104 stores programs, data, etc. necessary for the operation of the CPU 105, and is used as an example of a computer-readable non-transitory storage medium that stores a program executed by the ECU.

[0014] The CPU 105 is an abnormality detection device in this embodiment. Control signals calculated by the CPU 105 in accordance with a control program are output to various devices such as a throttle valve drive circuit 106, a fuel injection valve drive circuit 107, and an ignition output circuit 108. The throttle valve drive circuit 106 outputs a drive signal to the throttle valve to control the opening and closing of the throttle valve.

[0015] The fuel injection valve drive circuit 107 outputs a drive signal to the fuel injection valve to control the opening and closing of the fuel injection valve at the fuel injection timing. The ignition output circuit 108 outputs a drive signal to the ignition output circuit 108 to control the ignition of the spark plug at the ignition timing. Although an ECU is shown in this embodiment, the present invention can also be applied to other controllers such as a VCM.

[0016] <Neural network model> An overview of a neural network is shown in Figure 2. A neural network model is a mathematical model that mimics the structure of the neural circuits in the human brain, and each neuron that makes up the model has a weight and bias set.

[0017] In addition, a function called an activation function is defined for neurons. Activation functions such as logistic functions and ramp functions are set as appropriate. A layer is formed from multiple neurons, and a hidden layer is set between the input layer and the output layer. By increasing the number of neurons and the number of hidden layers, it is possible to approximate more complex input-output relationships. There is a trade-off between approximation accuracy and model size, and a compromise that satisfies both requirements is selected. For example, the frequency of the exhaust gas sensor signal is set in the input layer and catalyst deterioration diagnosis is set in the output layer, and the input-output relationship can be approximated by machine learning (supervised) the weights and biases of each neuron. Error backpropagation can be applied to the machine learning algorithm.

[0018] <Autoencoder> Figure 3 shows an overview of an autoencoder (self-encoder). An autoencoder is a type of computing device that is trained using an unsupervised learning method that uses a neural network. An autoencoder is structured as an encoder that reduces the dimension of the input data and a decoder that restores the original data. As shown in Figure 3, learning is performed so that the input and output to the autoencoder are the same. Since learning is performed using normal data, if an abnormal input occurs, the input value and output value will not be the same, making it possible to detect the abnormality. Note that the input value here refers to the value of the input signal.

[0019] <Configuration Diagram> Next, the configuration of the anomaly detection device (CPU 105) according to this embodiment will be described with reference to Figures 4A and 4B. Figure 4A shows a schematic diagram of the anomaly detection device 105. An input signal is input from 101 and 102 in Figure 1. An input signal may also be input from RAM 103. The input signal is input to an anomaly detection unit 440 configured by a model 401 in the CPU 105 and an autoencoder 402 and comparator 403 shown in Figure 4B. If it is determined that the input signal contains an anomaly, an anomaly signal identifier 400 identifies the input signal that is abnormal, and the final output is input to RAM 103.

[0020] FIG. 4B shows a detailed configuration diagram of the anomaly detection device. Note that this embodiment assumes a case where only one of the multiple input signals is abnormal, or where no abnormal signal is included. In the first step, multiple input signals are input to a model 401. The input signals include control values, sensor values, estimated values ​​from other models, etc. The multiple input signals are input to a first-step autoencoder 402. The first-step autoencoder 402 is an autoencoder that has been trained using known inputs. The input and output values ​​of the first-step autoencoder 402 are compared using a comparator 403. If the input and output values ​​do not match or the difference is greater than or equal to a threshold, the abnormal signal identifier 400 is activated and the process proceeds to the second step. In the second step, the (n-1) input signals are input to an autoencoder 404, which has (n-1) inputs excluding one of the n input signals. Note that the n input signals input to the first-step autoencoder 402, which are to be determined for the presence or absence of an abnormal signal, are referred to as a group of diagnosed input signals, and the corresponding output is referred to as a group of diagnosed output signals. In addition, the (n-1) input signals excluding one of the n input signals input to the autoencoders 404 to 406 in the second step are called an input signal group, and the corresponding output is called an output value.

[0021] In FIG. 4B , two of the three input signals are input to an autoencoder 404. The autoencoder 404 outputs an output value based on the input signals. Then, a comparator 407 compares the input signal value with the output value and determines whether the difference value is equal to or greater than a certain threshold. If it is equal to or greater than the threshold, a 1 is output. Another input signal is excluded from an autoencoder 405, and a comparator 408 compares the input signal value with the output value and determines whether the difference value is equal to or greater than a certain threshold. If it is equal to or greater than the threshold, a 2 is output. Finally, a third input signal is excluded, and similarly, an autoencoder 406 is used, and a comparator 409 compares the input signal value with the output value and determines whether the difference value is equal to or greater than a certain threshold. If it is equal to or greater than the threshold, a 3 is output.

[0022] When the abnormal signal has been excluded, the two input signals are normal, and the difference between the input value and the output value is below a certain threshold. Therefore, the signal that was excluded at that time can be determined to be an abnormal signal. Next, proceed to step 3. The adder 410 adds the output results of the comparators 407, 408, and 409. For example, if the input signal c is an abnormal signal, the comparator 407 outputs 1 and the comparators 408 and 409 output 0, so the adder 410 outputs 1. If the input signal b is an abnormal signal, the comparator 408 outputs 2 and the comparators 407 and 409 output 0, so the adder 410 outputs 2. If the input signal a is an abnormal signal, the comparator 409 outputs 3 and the comparators 407 and 408 output 0, so the adder 410 outputs 3.

[0023] Then, the multiple switching unit 412 changes the abnormal signal to a substitute value that matches the numerical value output from the adder 410, and outputs an output signal from the model group 411. This model group functions as a substitute signal replacement unit that replaces the abnormal signal with a substitute signal. For example, if the input signal c is abnormal, the output value of the model changed to input signal c' is output. A signal that is highly correlated with or physically close to the abnormal signal is used as the substitute value. By verifying that the output value of the model changed to this substitute value is correct, it is possible to ensure the accuracy of the model. Finally, the switching unit 413 outputs a final value. If all input signals input to the autoencoder 402 in the first step are normal, the comparator 403 outputs 1, so the output value of the model 401 is valid and the abnormal signal identifier 400 is not activated.

[0024] On the other hand, if the input signal input to the autoencoder 402 contains an abnormal signal, the comparator 403 outputs 0, and the output value of the multiple switching unit 412 becomes valid. The autoencoder is a system that performs learning before shipping, but may have a function that allows the learning value to be updated by connecting to a network after shipping. Note that the above threshold values ​​are for the case where the number of input signals is three, and change depending on the number of input signals.

[0025] 5 shows a flowchart of the processing executed by the abnormality detection device 105 according to the first embodiment. In step 501, all of the multiple input signals that are the same as 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 if the difference between the input value and the output value is equal to or greater than a certain threshold, the process proceeds to step 503. In step 503, one of the three input signals is excluded, leaving two inputs, and the two signals are input to the autoencoder 404.

[0026] Furthermore, signals other than the excluded signals are excluded. In the first embodiment, three types of autoencoders are used. In step 504, a comparator 407 determines whether the difference between the input value and the output value of the autoencoder 402 is equal to or less than a threshold value. The processes of steps 503 and 504 are performed in parallel for the other autoencoders 405 and 406 and comparators 408 and 409. In step 505, an adder 410 calculates the converted values ​​of the output values ​​of the comparators 407, 408, and 409.

[0027] Comparators 407, 408, and 409 output 0 when the input and output values ​​match or when the difference is below a certain threshold, so by adding them together, if the abnormal signal is c, 1 is output, if the abnormal signal is b, 2 is output, and if the abnormal signal is a, 3 is output. In step 506, the abnormal value that matches the added value is replaced with a substitute value, and the output value of the model is output.

[0028] As described above, according to this embodiment, even if an input signal contains an abnormality, the accuracy of the model can be confirmed by identifying the abnormal input signal, thereby making it possible to prevent the reliability of the model from being lost.

[0029] [Example 2] In Example 2, a configuration will be described for a case where two or more abnormal signals are included in an input signal. FIG. 6 shows a detailed configuration of the abnormality determination device 105 according to this example. In a first step, a plurality of input signals are input to a model 601. The input signals include control values, sensor values, estimated values ​​of other models, etc. The plurality of input signals are input to an autoencoder 602. The autoencoder 602 is an autoencoder that has trained using the inputs. Then, the input value and output value of the autoencoder 602 are compared using a comparator 603. If the difference between the input value and the output value is equal to or greater than a certain threshold, the abnormal signal identifier 600 is activated and the process proceeds to a second step.

[0030] In the second step, the (n-1) signals are input to an autoencoder 604, which has (n-1) inputs excluding one of the input signals. In FIG. 6, two of the three input signals are input to the autoencoder 604. The autoencoder 604 outputs an output value based on the input value. Then, a comparator 607 calculates the difference between the input value and the output value and first determines whether the difference is equal to or less than threshold 1. If the difference is equal to or less than threshold 1, the comparator 607 outputs 1. Furthermore, if signals a and b input to the autoencoder 604 are abnormal values, the difference between the input value and the output value becomes larger, so the comparator 607 compares the difference with threshold 2. If the difference is equal to or greater than threshold 2, the comparator 607 outputs 4. If the difference is equal to or greater than threshold 1 and less than threshold 2, the comparator 607 outputs 0. The relationship between threshold 1 and threshold 2 is threshold 2 > threshold 1.

[0031] Similarly, autoencoders 605 and 606, excluding one other input signal, also receive inputs in the form of a combination of input signals. Similar to comparator 607, comparators 608 and 609 compare the difference between the input and output values ​​of the autoencoder with thresholds 1 and 2. If the difference is greater than or equal to threshold 2, comparator 608 outputs 5; if the difference is greater than or equal to threshold 1 and less than threshold 2, comparator 608 outputs 0; if the difference is less than or equal to threshold 1, comparator 609 outputs 6; if the difference is greater than or equal to threshold 1 and less than threshold 2, comparator 609 outputs 0; if the difference is less than or equal to threshold 1, comparator 609 outputs 3.

[0032] Next, the abnormal signal number detector 610 determines whether all input signals are abnormal or whether at least one normal signal is being input. In Fig. 6, if the sum of the output values ​​of each comparator is 15, all input signals a, b, and c are abnormal. If the sum is not 15, it means that at least one normal signal is being input.

[0033] For example, if signal c is abnormal, the difference between the input and output of autoencoder 604, which is a combination that does not include signal c, is less than or equal to threshold 1, and comparator 607 outputs 1. On the other hand, the differences between autoencoders 605 and 606 are partly abnormal signals, so they are greater than threshold 1 and less than threshold 2, and output 0. Furthermore, if the abnormal signals are a and b, the difference between the input and output of autoencoder 604 is greater than threshold 2, and comparator 607 outputs 4. The differences between autoencoders 605 and 606 are partly abnormal signals, so they are greater than threshold 1 and less than threshold 2, and output 0. Furthermore, if all signals are abnormal, the differences between the input and output of autoencoders 604, 605, and 606 are greater than threshold 2, and the outputs of comparators 607, 608, and 609 are 4, 5, and 6, respectively, totaling 15.

[0034] When at least one normal signal is input among the input signals, the maximum value determination unit 611 outputs the maximum value of the output values ​​of each comparator. When the maximum value of the output value of the maximum value determination unit 611 is 4, 5, or 6, this means that two abnormal signals are input. On the other hand, when the maximum value of the output value of the maximum value determination unit 611 is 3, 2, or 1, this means that one abnormal signal is input. Then, depending on the number input by the switching units 613 and 614, the model is changed to one in which the abnormal signal is changed to an alternative value, and an output value is output. When the output value is 4, 5, or 6, the comparator 612 and the switching unit 615 make the output value of the switching unit 613 valid. This is also intended to ensure the accuracy of the model, as in the first embodiment.

[0035] If the maximum value of the output value of the maximum value determination unit 611 is equal to or less than 3, 2, or 1, the output value of the switching unit 614 becomes valid. Then, if the switching unit 616 determines that all input signals are abnormal and the sum of the output values ​​of the comparators 607, 608, and 609 is 15, the output of the model 7 becomes valid, and if the sum is not 15 and a normal signal is included, the output value of the switching unit 615 becomes valid. Finally, if the input of the model 601 includes an abnormal signal, the output value of the switching unit 616 becomes valid by the switching unit 617, and if it is normal, the output value of the model 601 becomes valid, and the abnormal signal identifier 600 does not operate.

[0036] The autoencoder is a system that performs learning before shipping, but may have a function that allows the learning value to be updated by connecting to a network after shipping. Each of the threshold values ​​is for the case where the number of input signals is three, and changes depending on the number of input signals.

[0037] 7 shows a flowchart of the process executed by the abnormality detection device according to the second embodiment. In step 701, all 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 if the difference between the input and output values ​​is equal to or greater than a certain threshold, the process proceeds to step 703.

[0038] In step 703, one of the three input signals is excluded, leaving two inputs to the autoencoder 604, and these two signals are input. Furthermore, the other signal from the excluded signal is excluded. In the second embodiment, three types of autoencoders are used. In step 704, a comparator determines whether the difference between the input value and the output value of the autoencoder is equal to or less than threshold 1. If it is equal to or less than threshold 1, this indicates that there is one abnormal signal, and the process proceeds to step 705, where the outputs of models 1 to 3 in which the one abnormal signal has been replaced with an alternative value are validated.

[0039] If it is equal to or greater than threshold 1, it indicates that there is a difference value that is equal to or greater than threshold 2. In step 706, a comparison is made to see if the sum of the output values ​​of each comparator is equal to a specified value. In the second embodiment, this is set to 15. If they are equal, all signals are considered to be abnormal, and the process proceeds to step 707, where the output value of model 7 in which all signals have been replaced with alternative values ​​becomes valid. If they are not equal, the process proceeds to step 708, where one of the outputs of models 4 to 6 in which two abnormal signals have been replaced becomes valid.

[0040] As described above, according to this embodiment, even if the input signal contains two abnormal signals, the abnormal input signals can be identified in the same manner as in the first embodiment, and the accuracy of the model can be maintained.

[0041] The above-described embodiment of the present invention provides the following advantageous effects.

[0042] (1) The abnormality discrimination device of the present invention is an abnormality discrimination device to which n (n is a natural number greater than or equal to 3) input signals are input, and includes n autoencoders configured to output an output value corresponding to a group of n-1 input signals when one of the input signals is removed, and a comparison unit that calculates a difference between the value of the input signal group and the output value for each of the n autoencoders and compares the difference value with a predetermined threshold value, and an abnormality discrimination unit that discriminates abnormal signals contained in the multiple input signals based on the comparison result.

[0043] With the above configuration, it is possible to identify input signals that cause estimation errors in a calculation model, maintain the reliability of the calculated values ​​of the model, and prevent deterioration of estimation accuracy due to abnormalities in the input values ​​of the model.

[0044] (2) The system further includes a first step autoencoder that, when a group of input signals to be diagnosed consisting of n input signals is input, outputs a diagnostic output value corresponding to the group of input signals to be diagnosed, compares the group of input signals to be diagnosed with a group of output values ​​to be diagnosed, and estimates that an abnormal signal is included in one of the n input signals if the difference between the value of the group of input signals to be diagnosed and the value of the output value to be diagnosed is within a predetermined threshold, and when the first step autoencoder estimates that an abnormal signal is included in one of the n input signals, performs an abnormality determination of the multiple input signals based on the comparison results from the n autoencoders. The effect of (1) is ensured by the presence of such a first step autoencoder.

[0045] (3) The abnormality determination unit determines, as an abnormal signal, the one input signal excluded from the autoencoder determined to have the smallest difference value. This is the method of abnormality determination in the first embodiment.

[0046] (4) The abnormality determination unit determines, as an abnormal signal, one input signal excluded from each of the plurality of autoencoders whose difference value is determined to be smaller than a predetermined threshold. This is an abnormality determination method in the second embodiment.

[0047] (5) The system further includes an alternative signal replacement unit that calculates an alternative signal using a method different from the predetermined calculation method for one input signal determined to be an abnormal signal, and replaces the one input signal with the alternative signal. This allows the abnormal signal to be replaced with another normal signal, output, and the accuracy of the model to be confirmed by confirming that the output is correct.

[0048] 105...CPU (abnormality determination device), 402...autoencoder, 403...comparator (comparison unit), 404 to 406...autoencoders, 407 to 409...comparators (comparison units), 410...adder (abnormality determination unit), 411...model group (alternative signal replacement unit)

Claims

1. An abnormality detection device that receives n (n is a natural number greater than or equal to 3) input signals, comprising n autoencoders configured to output an output value corresponding to a group of n-1 input signals obtained by excluding one of the input signals, for each of the n autoencoders, a comparison unit that calculates a difference between the value of the group of input signals and the output value and compares the difference with a predetermined threshold value, and an abnormality detection unit that detects abnormal signals contained in the plurality of input signals based on a result of the comparison.

2. An abnormality discrimination device as claimed in claim 1, further comprising a first step autoencoder which, when a group of diagnosed input signals consisting of the n input signals is input, outputs a diagnosed output value corresponding to the group of diagnosed input signals, and compares the group of diagnosed input signals with the diagnosed output value, and if a difference between the value of the group of diagnosed input signals and the value of the diagnosed output value is within a predetermined threshold, presumes that one of the n input signals contains an abnormal signal; and if it is presumed by the first step autoencoder that one of the n input signals contains an abnormal signal, performs abnormality discrimination of the multiple input signals based on the comparison result from the n autoencoders.

3. An abnormality discrimination device according to claim 1, characterized in that the abnormality discrimination unit discriminates, as the abnormal signal, the one input signal excluded in the autoencoder determined to have the smallest difference value.

4. An abnormality discrimination device as claimed in claim 1, characterized in that the abnormality discrimination unit discriminates, as the abnormal signal, each of the one input signals excluded in each of a plurality of autoencoders in which the difference value is determined to be smaller than a predetermined threshold value.

5. An abnormality discrimination device as claimed in claim 3, further comprising an alternative signal replacement section which determines an alternative signal by a method different from the predetermined calculation method of the one input signal determined to be an abnormal signal, and replaces the one input signal with the alternative signal.

Images