JP2012048405A - Control device, electronic control unit and abnormality detection method - Google Patents

Abstract

PROBLEM TO BE SOLVED: To provide a control device capable of discriminating an abnormality a supply source of input data supplied to an ECU from a CPU incorporated in the ECU.SOLUTION: A control device 100 includes: data output means 22 for outputting first data and outputting second data after the first data; data prediction means 16 for predicting the second data from the first data and outputting it as predicted data; first arithmetic operation means A for performing arithmetic operation on the second data; second arithmetic operation means B for performing arithmetic operation on the predicted data; and comparison means 17 for comparing between a fist output value on which the first arithmetic operation means has performed an arithmetic operation and a second output value on which the second arithmetic operation means has performed an arithmetic operation on the predicted data.

Description

  The present invention relates to a control device that can detect an abnormality in a CPU or an input system, and in particular, a control device that can detect an abnormality in a CPU or an input system when the CPU has a plurality of cores, an electronic device, and the like. The present invention relates to a control unit and an abnormality detection method.

  When an abnormality occurs in an on-vehicle ECU (electronic control unit), the control of the vehicle is affected, so a function for detecting an abnormality of the ECU is often installed. For example, a WDT (Watch Dog Timer) for detecting an abnormality of one CPU and a technique for mutual monitoring between two ECUs are known. However, WDT detects processing delays such as the task processing not progressing for some reason. Mutual monitoring that detects a state where processing proceeds but correct calculation cannot be performed is more abnormal. There is a possibility that it can be detected.

  In the mutual monitoring of the two ECUs, two calculation results from physically different processing systems are compared, so that an abnormality can be detected as long as the same kind of abnormality does not occur at the same time.

  By the way, integration of ECUs has been promoted in order to reduce the number of ECUs mounted on the vehicle, and if the functions of two ECUs are conventionally mounted on one ECU, such mutual monitoring becomes impossible. Become. For example, as a method for detecting an abnormality, a technique for detecting an abnormality when a difference between a previous value and a current value of a calculation result exceeds an error has been devised (see, for example, Patent Document 1). When this technology is applied to one ECU, it is conceivable to perform the same calculation with a plurality of cores and determine whether or not the difference between the calculation results of both exceeds an error.

  Further, a system model construction method has been proposed in which parameters of a state space model modeled hypothetically from input / output data of the target system are obtained to perform system identification of the target system (see, for example, Patent Document 2). In Patent Document 2, the accuracy of the obtained state space model and the parameters is compared by comparing the estimated output data estimated by the model with the actually measured output data obtained by actually executing the input data on the target system. A technique for confirming the above is disclosed.

JP-A-61-175701 JP 2007-260504 A

  However, both Patent Documents 1 and 2 are based on the premise that the input data is correct. However, since the input data is supplied by a sensor or a circuit configured in hardware, the input data includes a large error. There is. In this case, Patent Documents 1 and 2 both provide input data to a plurality of computing means, and thus there is a problem that an abnormality such as a sensor that provides input data cannot be detected.

  In view of the above problems, an object of the present invention is to provide a CPU mounted on an ECU, a control device capable of determining an abnormality of a supply source of input data supplied to the ECU, an electronic control unit, and an abnormality detection method. .

  In view of the above problems, the present invention predicts the second data from the first data by outputting data of the first data and outputting the second data after the first data. Data predicting means for outputting as prediction data, first computing means for computing the second data, second computing means for computing the forecast data, and the first computing means Comparing means for comparing the first output value obtained by calculating the second data and the second output value calculated by the second calculating means on the prediction data. A control device is provided.

  It is possible to provide a CPU mounted on the ECU, a control device, an electronic control unit, and an abnormality detection method capable of determining an abnormality of a supply source of input data supplied to the ECU.

It is an example of the figure explaining the outline of ECU. It is an example of the hardware block diagram of ECU. It is an example of the figure which illustrates the functional block of ECU typically. It is an example of the figure which illustrates a hidden Markov model typically. It is an example of the figure which shows a learning result typically. It is an example of the figure explaining the reference vector used with a self-organizing map. It is an example of the figure which shows the relationship between a label and an input vector typically. It is an example of the flowchart figure which shows the operation | movement procedure of ECU.

Hereinafter, embodiments for carrying out the present invention will be described with reference to the drawings.
FIG. 1A is an example of a diagram illustrating an outline of the ECU according to the present embodiment. For comparison, a conventional ECU (electronic control unit) is also shown in FIG.

  As shown in the conventional diagram, conventionally, ECU_A and ECU_B are mounted on separate boards, and ECU_A executes application A and ECU_B executes application B, respectively. ECU_A monitors ECU_B for abnormality, and ECU_B monitors ECU_A for abnormality.

  Therefore, if ECU_A and ECU_B are integrated into one ECU_W as they are, ECU_W executes application A and application B on a common PF (platform) as shown in the lower diagram of FIG. In this case, an independent core is assigned to each of application A and application B. However, since the same input value (calculation target) is given to the two cores, it is detected that there is an abnormality in the input value simply by integrating them. Can not.

  On the other hand, in this embodiment, when integrating ECU_A and ECU_B, the probability model is connected to one core B. The probabilistic model may be a probabilistic model that is effective for estimating an input value, and is, for example, a hidden Markov model or a self-organizing map. The probabilistic model learns input values when the ECU_A and its input system and the ECU_B and its input system are in a normal state, and if the input value before one control cycle is normal, the “normal value most probable in probability” (This normal value is a “predicted input value” described later). In the following description, it is assumed that the hardware 22 provides an input value.

  The applications A and B of the core A are calculated based on input values output from the hardware 22, and the applications A and B of the core B are calculated based on predicted input values output from the probability model. Therefore, when the input value and the predicted input value are substantially the same, and the values output by the applications A and B of the core A and the applications A and B of the core B (hereinafter referred to as the current output value and the predicted output value, respectively) Some abnormality may occur in either or both of the cores A and B.

On the other hand, if the current output value output by the applications A and B of the core A and the predicted output value output by the applications A and B of the core B are different, the probability model outputs if the input value and the predicted input value are not the same. Since the predicted input value is a normal value, there is a possibility that an abnormality has occurred in the hardware 22. As described above, the ECU 100 of the present embodiment has a probability model that provides a normal input value. It is possible to detect which of the plurality of cores mounted on one ECU and the hardware 22 that provides input values to the plurality of cores has an abnormality.

〔Constitution〕
FIG. 2 shows an example of a hardware configuration diagram of the ECU 100. The ECU 100 includes a multi-core CPU 11, a reset circuit 12, an I / O bridge 13, an EEPROM 14, a comparison device 17, a probability model mounting unit 16, and a memory 15. The multi-core CPU 11 has a plurality of cores 18 represented by core A and core B. Although there are two in the figure, the number of cores 18 may be three or more.

  Various sensors, actuators, and switches are connected to the I / O bridge 13, and the I / O bridge 13 arbitrates connection requests from sensors and the like, and the actuator calculates data calculated by the multi-core CPU 11 in accordance with requests from the multi-core CPU 11. Send to etc.

  The reset circuit 12 is a circuit that outputs a reset signal to either the core A or the core B or both the core A and the core B. Since the reset core A and core B are restarted, they may be able to be recovered if any abnormality occurs.

  The EEPROM 14 is a non-volatile memory that stores applications A and B, an OS (Operating System), a driver, and the like. Software on the hardware side such as an OS and a driver corresponds to the PF 21 described above. The core A executes the applications A and B by using the memory 15 as a temporary storage area, and the core B executes the applications A and B in the same manner. The output value calculated by executing B is used (the output value calculated by the core B executing the applications A and B may be used for control). The memory 15 is a volatile memory such as a DIMM (Dual Inline Memory Module).

  The core A and the core B perform the same calculation on the input value based on the processing content of the application A or B. However, for example, when the sensor outputs an analog value, only the core A needs to be A / D converted.

  The comparison device 17 compares the current output value of the application A executed by the core A with the predicted output value of the application A executed by the core B, and the current output value of the application B executed by the core A and the execution of the core B. The predicted output value of application B is compared. If there is a difference greater than the degree that the absolute value of the difference between the two cannot be regarded as coincident, the comparison device 17 starts the abnormal avoidance action. The abnormal avoidance behavior will be described later.

  The probability model mounting unit 16 learns an input value when the hardware 22 is in a normal state according to a predetermined probability model, calculates a predicted input value from a past actual input value using the learning result, and outputs the predicted input value to the core B. . How to obtain the predicted input value will be described later. The probability model mounting unit 16 may provide the core B with “normal value most likely to be probable = predicted input value” by calculating the predicted input value from the verified input value (current input value). it can. For example, the probability model implementation unit 16 draws a map obtained by learning based on the previous input value to extract the current predicted input value, or inputs the previous input value to the function obtained by learning to input the predicted value. Calculate the value.

  The probability model mounting unit 16 calculates a predicted input value for application A and a predicted input value for application B, respectively. What is necessary is just to use a model suitable for estimating the input value which the application A or the application B uses for a calculation stochastically. In the present embodiment, application A and application B will be described without distinction.

  The comparison device 17 and the probability model mounting unit 16 can be mounted both in hardware and software, but more preferably in hardware. By implementing the hardware, the comparison device 17 can compare the current output value with the predicted output value without being affected by the abnormalities of the core A, the core B, and the input system, and the probability model mounting unit 16 determines the predicted input value. It can be calculated. When implemented in software, it is preferable to provide a third core C, and the core C compares the current output value with the predicted output value and calculates the predicted input value based on the probability model.

FIG. 3 is an example of a diagram schematically illustrating functional blocks of the ECU 100. The hardware 22 is a sensor, an actuator, a hardware part of the ECU 100, and the like, and corresponds to hardware that supplies an input value to the ECU 100.
(1) As illustrated, an input value (same as the current input value) is input to the core A from the hardware 22, but not input to the core B.

The comparison device 17 can store the current input value, the previous value (current input value of the previous control cycle), and the predicted input value in a RAM or the like, and can store the current output value and the predicted output value in a RAM or the like. The comparison device 17 stores the input value from the hardware 22 as the current input value.
(2) The core A calculates the current output value of the application A or B based on the input value from the hardware 22 and inputs it to the comparison device 17. The comparison device 17 stores the current output value. In the comparison device 17, for example, the application A or B of the core A outputs the current output value without comparing the current output value with the predicted output value until the ECU 100 is stabilized (for example, several seconds to several minutes after activation). Repeat overwriting of the current output value every time.
(3) Each time the application A or B of the core A outputs the current output value (when the comparison result matches), the comparison device 17 copies the already stored current input value to the previous value. The value is output to the probability model mounting unit 16. Therefore, the input value input to the core A and the input value (previous value) input to the probability model mounting unit 16 are shifted by one control cycle.
(4) Thereby, the probability model mounting unit 16 calculates the predicted input value from the previous value. The predicted input value is a predicted current input value that is input to the core A in the next control cycle. The probability model mounting unit 16 writes the calculated predicted input value in the comparison device 17. Thereby, the comparison device 17 can compare the predicted input value and the current input value.
(5) Further, the probability model mounting unit 16 provides the calculated predicted input value to the core B. As a result, the core B can output the predicted output value. The predicted output value output by the core B and the current output value output by the core A can be regarded as the calculation result of the same control cycle.
(6) The core B outputs the predicted output value calculated based on the predicted input value to the comparison device 17.
(7) If the current output value and the predicted output value substantially match, the comparison device 17 has verified that the current output value is correct, and outputs the current output value to an actuator or the like. If it is verified that the output value is correct this time, it can be verified that the input value is correct this time. Therefore, as shown in (3), when the current output value matches the predicted output value, the comparison device 17 copies the current input value verified to be correct to the previous value, and uses the previous value as the probability model implementation unit. 16 is output.

  In this way, the core B can always use the predicted input value that is most likely to be probabilistically calculated from the previous value verified to be correct as a calculation target. By using this, it is possible to isolate an abnormality between the multi-core CPU 11 and the hardware 22.

[Hidden Markov Model]
The probability model mounted by the probability model mounting unit 16 will be described.
The hidden Markov model is based on a Markov chain in which a state St at an arbitrary time point t is determined only on k state sequences St-1, St-2, St-3, ... St-k. In addition, it is a probabilistic generation model of time series patterns. The observer cannot observe the state St from the outside, but can observe the observation symbol output according to the probability.

FIG. 4 is an example of a diagram schematically illustrating the hidden Markov model. The value represented by “a” and the subscript ij is the transition probability from the state Si to Sj, and the value represented by “b” and the subscript is the output probability of the observation symbol output at the time of state transition. is there. Since “N” below is the number of states, N = 4.
a _ij (t) = P (S _{t + 1} = S _j | S _t = S _i ) (1 ≦ i, j ≦ N)

b _jk (t) = P (O _t | S _i = S _j ) (1 ≦ i, j ≦ N)
Note that bi1 + bi2 + bi3 = 1.

Here, O _t is an observation symbol, and has T symbols (three in the figure) of O _{1 to} O _T.
O _t = O ₁ , O ₂ ,… O _T
The hidden Markov model expresses the probability of initially taking the state St as the initial state π. In the present embodiment, π = {1, 0, 0}.

The probability of an event according to a hidden Markov model can be derived from a _ij (t), b _jk (t) and π. A set of a _ij (t), b _jk (t) and π is called a hidden Markov model λ.
As described above, since the state St cannot be observed from the outside in the hidden Markov model, the probability that the next observation symbol is observed from the observation symbol and the probability that the observation sequence of the observation symbol is generated are calculated. For example, the probability that O ₁ is observed in the initial state (state S ₁ ) is “a ₁₁ · b ₁₁ ”, the probability that O ₂ is observed is “a ₁₁ · b ₁₂ ”, and the probability that O ₃ is observed is “A ₁₁・ b ₁₃ ”. The probability that {O ₁ , O ₂ } is observed is as follows.
_{a 11 · b 11 · a 11} · b 12 + a 11 · b 11 · a 12 · b 22 + a 11 · b 11 · a 13 · b 32
+ A ₁₂ , b ₂₁ , a ₂₂ , b ₂₂ + a ₁₂ , b ₂₁ , a ₂₃ , b ₃₂
+ A ₁₃・ b ₃₁・ a ₃₃・ b ₃₂ + a ₁₃・ b ₃₁・ a ₃₂・ b ₂₂
Therefore, if a _ij (t), b _jk (t), and π are obtained by a method described later, an output symbol that is most likely to be observed next can be determined from a past observation sequence. For example, it is possible to calculate which observation symbol is most easily observed next to {O ₁ , O ₂ }.

The probability that “O ₁ ” is observed after the first term “a ₁₁・ b ₁₁・ a ₁₁・ b ₁₂ ” in the above equation is
a ₁₁ , b ₁₁ , a ₁₁ , b ₁₂ , a ₁₁ , b ₁₁
The probability that “O ₂ ” is observed is
a ₁₁ , b ₁₁ , a ₁₁ , b ₁₂ , a ₁₁ , b ₁₂
The probability that “O ₃ ” is observed is
a ₁₁ , b ₁₁ , a ₁₁ , b ₁₂ , a ₁₁ , b ₁₃
It becomes. Such a calculation is performed for all the terms of Equation (1), and the observation symbol with the highest numerical value becomes the observation symbol that is most easily observed next to {O ₁ , O ₂ }.

From the above, it is possible to generate a hidden Markov model by calculating a _ij (t), b _jk (t), and π, assuming the number of states St and possible transitions from the number of observed symbols. Although a complete solution for a _ij (t) and b _jk (t) has not been found, a forward algorithm can estimate a _ij (t) and a backward algorithm can estimate b _jk (t). The Baum-Welch algorithm is known as a method for obtaining the local optimum value of the parameter of the model λ to be maximized. Since these estimation methods are known, they will be briefly described. Details are described in, for example, “Pattern Recognition and Learning Algorithm” by Yoshinori Uesaka.

Prospective algorithm The probability of being in the state St = i at time t is defined as follows.
α _t (i) = P (O ₁ O _2... O _T, St = i | λ)
First, α _t (i) is initialized. For all α _t (i) (1 ≦ i ≦ N)
α ₁ (j) = π _j b _jk . Also, an appropriate initial value (for example, an initial value in which each element is uniform) is set in b _jk (t).
Next, the following calculation is repeated for all j (1 ≦ j ≦ N) at each time t (1 to T).

以上の計算から
P(O|λ)は次のように表すことができる、

From the above calculation
P (O | λ) can be expressed as:

・後ろ向きアルゴリズム
時間ｔにおいて状態Siにいて、O_t+1，O_t+2…,O_Tという観測系列を出力する確率Pを次のように定義する。
β_t(i)＝Ｐ（O_t+1，O_t+2…,O_T，S_t＝i｜λ）
まず、β_ｔ(i)を初期化する。すべてのB_ｔ(i)（１≦i≦N）に対し、
β_T（i）＝１とする。
次に、各時間ｔ（T-1、T-2、…０）毎に、全てのi（１≦ｊ≦N）に対し以下の計算を繰り返す。す

Backward algorithm The probability P of outputting an observation sequence O _{t + 1} , O _{t + 2} ..., O _T in the state Si at time t is defined as follows.
β _t (i) = P (O _{t + 1} , O _{t + 2} ..., O _T, S _t = i | λ)
First, β _t (i) is initialized. For all B _t (i) (1 ≦ i ≦ N)
Let β _T (i) = 1.
Next, the following calculation is repeated for all i (1 ≦ j ≦ N) at each time t (T−1, T−2,... 0). You

以上の計算から
P(O|λ)は次のように表すことができる、

From the above calculation
P (O | λ) can be expressed as:

・Baum-Welch アルゴリズム
与えられた観測系列の生成確率を最大にするモデルλを推定する。まず、以下の式により、π_i、a_ij、b_ijを計算する。π_i、a_ij、b_ijの上のバー中間の計算値を意味する。

・ Baum-Welch algorithm Estimates the model λ that maximizes the probability of generating a given observation sequence. First, π _i , a _ij , and b _ij are calculated by the following equations. It means the calculated value in the middle of the bar on π _i , a _ij and b _ij .

これにより、モデルλが推定されるので、モデルλに対し与えられた観測系列Ｏ₁Ｏ_2...Ｏ_T，が観測される確率P（O｜λ）を計算する。計算された確率Pを記憶しておき、再度、上式によりP（O｜λ）を計算し、記憶しておいた確率Pと比較して、両者の差が閾値を超えなくなったら、演算を終了する。言い換えると、隠れマルコフモデルが与えられた観測系列による学習が終了したことになる。

Thus, since the model λ is estimated, the probability P (O | λ) that the observation sequence O ₁ O _2... O _T given to the model λ is observed is calculated. Store the calculated probability P, calculate P (O | λ) by the above formula again, and compare it with the stored probability P. If the difference between the two does not exceed the threshold, calculate finish. In other words, the learning by the observation sequence given the hidden Markov model has been completed.

[Application example]
The probability model mounting unit 16 based on the hidden Markov model obtained as described above outputs a normal value (input predicted value) that is most likely to occur with respect to the previous value given from the comparison device 17. For example, the probability model mounting unit 16 outputs the maximum likelihood value as the input predicted value based on the past m previous values.

An example will be described in which the input value provided by the hardware 22 is the duty of the PWM signal, and 0%, 100%, or an indefinite value is output as the duty. This duty becomes an observation symbol that can be observed from the outside in the hidden Markov model. Therefore, when the probability model mounting unit 16 learns observation symbols, 0%, 100%, or indefinite values are assigned to the observation symbols O ₁ , O ₂ , and O ₃ , respectively. In learning, the duty actually output by the hardware 22 is recorded in time series and assigned to the observation symbol.

  For example, the hardware 22 has a high probability of outputting 100% duty after 100% duty, outputs 0% duty after an indefinite value, and does not output 0% twice consecutively. Suppose there are many. The probability model mounting unit 16 learns the input value output from the hardware 22 from the observation sequence.

FIG. 5 is an example of a diagram schematically showing a learning result. The three numerical values below the transition probability indicate the probability that the observation symbols O ₁ , O ₂ , and O ₃ are observed.
-When the observation series is {O ₂ , O ₂ } (100%, 100%), the observation symbols that can be observed next are O ₁ , O ₂ , and O ₃ , but the probability that each will be observed Is calculated as follows. Assume that π = {1, 0, 0} and the initial state is S ₁ .
O1: 0.9 x 0.5 x 0.7 x 1 x 0.2 x 0.5
O2: 0.9 x 0.5 x 0.7 x 1 x 0.7 x 1 + 0.9 x 0.5 x 0.7 x 1 x 0.2 x 0.5 + 0.9 x 0.5 x 0.7 x 1 x 0.1 x 0.5
O3: 0.9 x 0.5 x 0.7 x 1 x 0.1 x 0.5 + 0.9 x 0.5 x 0.2 x 0.5 x 0.1 x 15
Therefore, it can be seen that the observation symbol most likely to be observed next is O ₂ when the observation sequence is {O ₂ , O ₂ }.
When the observation series is {O ₃ , O ₁ } (indefinite value, 0%), the next observation symbols that can be observed are O ₁ , O ₂ , and O ₃ , but the probability that each will be observed Is calculated as follows.
O1: 0.9 × 0.1 × 0.2 × 0.5 × 0.9 × 0.4
O2: 0.9 × 0.1 × 0.2 × 0.5 × 0.9 × 0.5
O3: 0.9 × 0.1 × 0.2 × 0.5 × 0.9 × 0.1 + 0.9 × 0.1 × 0.2 × 0.5 × 0.1 × 1
Therefore, it can be seen that the observation symbol most likely to be observed next in the case of {O ₃ , O ₁ } is O ₂ .
As described above, the probability model mounting unit 16 can calculate a predicted input value output in a normal state of the hardware 22 as long as the probability model mounting unit 16 is normal. From the above example, it can be predicted as follows.

100% (t-2) → 100% (t-1) → 100% (t: present)
Undefined value (t-2) → 0% (t-1) → 100% (t: present)
In vehicles, PWM signals are often used to control motor torque. Actually, the duty can take a value of 1%, but it is not necessary to assign all the duty to different observation symbols, and if there are many types of input values, the observation symbols may be quantized.

  Further, not only the PWM signal but also a probability model can be generated when a voltage value or a current value is used as an input value. For example, when the voltage value ranges from 0 to 12 V, 0 ≦ voltage value <4 is observed symbol 1, 4 ≦ voltage value <8 is observed symbol 2, 8 ≦ voltage value ≦ 12 is observed symbol 3, Assign and determine a probability model.

[Self-organizing map]
An example in which a self-organizing map is used as a probability model will be described.
FIG. 6 is an example of a diagram for explaining reference vectors used in the self-organizing map. In the input layer, feature values x _{1 to} x _{n of} input values to be analyzed are arranged, and in the output layer, there are units (neurons) m _{1,1 to} m _5,5 that are more than the number of feature values. Any one unit m _{i, j} in the output layer has a reference vector of the same dimension as the feature quantities x _{1 to} x _n in the input layer. For example, unit m _5,1 has a reference vector of {m ^5,1 ₁ , m ^5,1 ₂ , m ^5,1 ₃ ,... ^{M 5,1} _n }.

  The self-organizing map having such a structure outputs the coordinates (i, j) of the unit having the reference vector closest to the input vector x (t). That is, the coordinates most relevant to the input vector x (t) are obtained. If meaning is given to these coordinates, the input vector x (t) can be converted into a meaningful value.

  In the self-organizing map, the reference vector is updated so that the reference vector of the unit is biased to be similar to the input vector x (t) by unsupervised learning when multiple input vectors x (t) are input. Go. This update of the reference vector is self-organizing map learning. Since the learning method is known, it will be described briefly.

(1) The reference vector {m ^{i, j} ₁ , m ^{i, j} ₂ , m ^{i, j} ₃ ,... M ^{i, j} _n } is determined at random.

  (2) An input vector x (t) for learning is prepared, and a unit most similar to x (t) is determined. A similar measure is, for example, the Euclidean distance.

(3) Update the reference vectors of the determined unit and its surrounding units. With this update, the reference vector gradually learns (closer to) the input vector. H _ci (t) in the following equation is a neighborhood coefficient, and becomes smaller as the distance between the determined unit and the surrounding unit is larger.

m ^ij (t + 1) = m ^ij (t) + h _ci (t) [x (t) -m ^ij (t)]
(4) By repeating (2) and (3), units similar to the input vector x (t) are gathered.

[Application example]
As a simplified example, for example, an example in which the input vector x (t) is classified into any coordinate will be described. {Previous value, previous value} is used as the input vector x (t). That is, there are two feature quantities, and each unit has a reference vector close to {previous value, previous value} by learning {previous value, previous value}.

  The input value provided by the hardware 22 is set as an input vector x (t), and the feature value is set as the duty of the PWM signal as described above. In other words, the input vector x (t) has the previous input value and the previous input value as feature amounts. In order to simplify the handling, the duty 0% is replaced with “0”, 100% is replaced with “1”, and the indefinite value is replaced with “2”.

  FIG. 7 is an example of a diagram schematically illustrating a learning example of a self-organizing map. In the self-organizing map, the input values actually output by the hardware 22 are extracted as a set of three input values from the previous value to the current value.

  Then, the self-organizing map identifies a unit having a reference vector that is closest to the set of the previous value and the previous value, and learns the reference vector of the unit from the previous value and the previous value. Further, the current value is linked to the reference vector closest to the set of the previous time value and the previous value. By repeating this process, the reference vectors for each coordinate of the self-organizing map come to gather values that are close to the previous value and the previous value, and the current value is associated with each coordinate.

In the learning process, the label a associated with the unit m _ij at the time t may be associated with the label b with the same unit m _ij at the time t + 1. In order to learn the current value accurately, the self-organizing map stores the label of the learning process, and the value of the learning process label, such as the average, median, and most frequent values, as the unit label. To do.

In FIG. 7, unit m _2,2 has a reference vector of (2,0) by learning, unit m _4,2 has a reference vector of (1,1), and unit m _2,4 has (2 , 2), and the unit m _4,4 has a reference vector of (0, 1). Also, the label “100%” is _assigned to the unit m _2,2 , the label “100%” is _assigned to the unit m ₄ , _2, and the label “0%” is _assigned to the unit m ₂ , ₄ . Are labeled “100%”. These representative units are in close proximity to units having a reference vector close to the reference vector of each unit. The number of units can be adjusted according to the number of events to be identified.

The probability model mounting unit 16 calculates a predicted input value using the result learned in this way. For example, when the last time value is 100% and the previous value is 100%, the input vector x (t) is {1, 1}, and the probability model mounting unit 16 classifies the input vector x (t) into units m ₄ and ₂ . To do. Then, referring to the label, “100%” is provided to the core B as a predicted input value.

Similarly, when the last time value is an indefinite value and the previous value is 0%, the input vector x (t) is {2, 0}, and is classified into units m _2,2 . Therefore, the predicted input value is “100%”. When the previous value is 0% and the previous value is 100%, the input vector x (t) is {0, 1} and is classified into units m _4,2 . Therefore, the predicted input value is “100%”. When the last time value is an indefinite value and the previous value is an indefinite value, the input vector x (t) is {2, 2}, and is classified into units m _2,4 . Therefore, the predicted input value is “0%”.
In this way, the probability model mounting unit 16 can output an input predicted value from the previous time value and the previous input value.

[Operation procedure]
FIG. 8 is an example of a flowchart showing an operation procedure of the ECU 100. First, the current input value is input from the hardware 22 to the core A (S10). When the current input value is an analog value, the core A performs A / D conversion on the current input value (S20). In addition, the core A writes the input value this time in the comparison device 17.

  In addition, before at least one control cycle elapses, the probability model mounting unit 16 receives an input of a previous value from the comparison device 17 (S101). Then, the probability model mounting unit 16 calculates a predicted input value by the method described above (S201). In addition, the probability model mounting unit 16 writes the predicted input value in the comparison device 17.

  The subsequent processing is the same for the core A and the core B. Core A converts the current input value to a logical value (S30), and Core B converts the predicted input value to a logical value (S301). The logical value is obtained by converting the input value into a form handled by the application A or B. For example, in the case of duty, a percentage value of 0 to 100% is converted into a predetermined numerical value.

  Next, the core A converts the logical value into a control value (S40), and the core B converts the logical value into a control value (S401). That is, the logical value is converted into a control value according to the processing contents of the application A or B. For example, the control value can be obtained by calculation for feedback control according to the difference from the target control value or processing for reading the value according to the logical value from the map.

  Next, the core A converts the control value into a physical value (S50), and the core B converts the control value into a physical value (S501). The physical value is almost synonymous with a voltage value or a current value when controlling the actuator or the like, and is a duty in the description so far. Therefore, the physical value of the core A becomes the current output value and is written in the comparison device 17, and the physical value of the core B becomes the predicted output value and is written in the comparison device 17.

  When the current output value and the predicted output value are written, the comparison device 17 determines whether or not the two substantially match (S60). If the predicted input value calculated by the probability model mounting unit 16 does not completely match the current input value, the predicted output value may not completely match the current output value. For this reason, the comparison device 17 determines that the two values coincide when the absolute value of the difference between the current output value and the predicted output value is less than a predetermined threshold value.

  If the current output value and the predicted output value substantially match (Yes in S70), it is verified that the predicted input value is normal, so the comparison device 17 overwrites the previous input value with the predicted input value (S80). . Then, the core A outputs the current output value to the actuator or the like (S90).

  If the current output value and the predicted output value do not substantially match (No in S70), there is a possibility that an abnormality has occurred in the core A or the core B, so the comparison device 17 compares the current input value with the predicted input value. (S100). Similarly, since there is a possibility that the current input value and the predicted input value do not completely match, the comparison device 17 determines that the absolute value of the difference between the current input value and the predicted input value is less than a predetermined threshold value. Are determined to match.

  If the current input value and the predicted input value substantially match (Yes in S100), the hardware device 22 provides an appropriate current input value, so the comparison device 17 detects that there is an abnormality in the multi-core CPU 11. For this reason, the comparison device 17 requests the reset circuit 12 to reset the core A and the core B in order to avoid an abnormality (S110).

  If the current input value and the predicted input value do not substantially match (No in S100), the predicted input value is normal, so there is a high possibility that an abnormality has occurred in the hardware 22. For this reason, the comparison device 17 requests the hardware 22 for an abnormal avoidance action from the abnormality (S120). For example, when the hardware 22 is a sensor or actuator, the sensor 22 is switched to an alternative sensor or actuator, the sensor or actuator is restarted, or the sensor or actuator is stopped.

  When any abnormality is detected in steps S120 and S130, the comparison device 17 notifies the multi-core CPU 11. Thereby, the multi-core CPU 11 can record that an abnormality has occurred. The recorded contents are, for example, an estimated abnormality location (multi-core CPU 11 or hardware 22), abnormality detection date and time, vehicle position information, and other vehicle information (vehicle speed, accelerator opening, steering angle, etc.).

  As described above, the ECU 100 according to the present embodiment has a probability model for calculating a normal predicted input value, and thus an abnormality has occurred in either the multicore CPU 11 mounted in the ECU 100 or the input system hardware 22. Can be detected.

11 Multi-core CPU
12 Reset circuit 13 I / O bridge 14 EEPROM
DESCRIPTION OF SYMBOLS 15 Memory 16 Probability model mounting part 17 Comparison apparatus 22 Hardware 100 ECU

Claims (9)

Data output means for outputting first data and outputting second data after the first data;
Data prediction means for predicting the second data from the first data and outputting it as prediction data;
First computing means for computing the second data;
Second computing means for computing the predicted data;
A comparison for comparing the first output value obtained by computing the second data by the first computing means and the second output value obtained by computing the predicted data by the second computing means. Means,
A control device comprising:   The control device according to claim 1, wherein the data prediction unit predicts the prediction data based on a probability model.   The control apparatus according to claim 2, wherein the probability model is a hidden Markov model or a self-organizing map.   The comparison device determines whether or not there is an abnormality by comparing an absolute value of a difference between the first output value and the second output value with a threshold value. The control device according to any one of claims. When the absolute value of the difference between the first output value and the second output value is greater than or equal to a threshold, the comparison device
The control apparatus according to claim 4, wherein the second data and the predicted data are compared to determine whether there is an abnormality. When the absolute value of the difference between the second data and the prediction data is greater than or equal to a threshold, the comparison device determines that the data providing unit is abnormal,
If the absolute value of the difference between the second data and the prediction data is less than a threshold, it is determined that at least one of the first calculation means or the second calculation means is abnormal.
The control device according to claim 5. When it is determined that there is an abnormality in the calculation means, the first calculation means or the second calculation means has an abnormality elimination means for causing the abnormality elimination operation,
The control device according to claim 6.   The electronic control unit carrying the control apparatus of any one of Claims 1-7. An abnormality detection method for a control device including a first calculation means and a second calculation means,
A data output means for outputting the first data;
The data output means outputs second data;
A step of predicting the second data from the first data and outputting the second data as prediction data;
The first computing means computing the second data;
A second calculating means for calculating the prediction data;
The comparison means includes a first output value obtained by computing the second data by the first computing means, a second output value obtained by computing the predicted data by the second computing means, And a method of detecting an abnormality characterized by comprising:

Images