JPH07166939A - Diagnosing method of operation of exhaust-gas oxygen sensor and electronic engine controller - Google Patents

Abstract

PURPOSE: To provide a diagnostic method for diagnosing faults on-vehicle whereby exhaust gas oxygen sensors can be inspected for their operating conditions in their proper positions while an internal combustion engine is operating. CONSTITUTION: This diagnostic method comprises the step of storing signal parameters characterizing the operation of a plurality of reference exhaust gas oxygen sensors, the step 10 of gathering signal parameters characterizing the operation of the unidentified exhaust gas oxygen sensor for a certain time, the steps 11-18 of comparing the currently gathered signal parameters with previously stored signal parameters, and the step of determining the operating characteristic of the unidentified exhaust gas oxygen sensor on the basis of this comparison.

Description

Detailed Description of the Invention

[0001]

FIELD OF THE INVENTION The present invention relates to a method for diagnosing the operation of an exhaust gas oxygen sensor.

[0002]

BACKGROUND OF THE INVENTION It is known to use an electronic engine controller (engine or engine electronic controller) to control the amount of fuel introduced into the cylinder of an internal combustion engine. There is. In particular, it is known to use the output of an exhaust gas oxygen (EGO) sensor as part of a feedback control loop to control the air / fuel ratio. Improper operation of the exhaust gas oxygen sensor can adversely affect proper air-fuel ratio control, so it is desirable to know when such a sensor is not operating properly.

In particular, it may be desirable for vehicles to monitor their emissions controls to determine if the emissions of production vehicles are kept within certain limits. The EGO sensor is used to monitor the oxygen content of the engine exhaust gas and control the engine air-fuel ratio. Such a sensor can regulate the mixture to a control point close to the theoretical value so that the peak efficiency of the three-way catalyst is maintained. If the sensor is defective, vehicle emissions may exceed desired limits and the catalyst may be damaged by improper engine operation.

Known techniques for testing such sensors are:
Removing the sensor from the vehicle and performing various voltage and functionality tests in a test room. It is probably desirable to have diagnostic techniques that can be used to inspect the sensor in-situ when the sensor is operably installed in the vehicle. Performing such an on-board test is advantageous for detecting the exhaust gas oxygen sensor status during engine operation.

[0005]

The present invention observes its behavior over a relatively short period of time and compares current signal signatures with previously collected signal signatures of known defect sensors and normal sensors. This provides a reliable method for diagnosing a defective exhaust gas oxygen sensor. Therefore, fault detection and confirmation can be achieved on vehicles using this method.

A properly operating exhaust gas oxygen sensor is advantageous for proper operation of the vehicle transmission. For example, EGO
By detecting improper operation of the sensor, undesirably high vehicle emissions can be avoided and thus damage to the catalytic converter can be avoided.

[0007]

DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS In accordance with an embodiment of the present invention, a diagnostic strategy or main procedure or process determines whether an exhaust gas oxygen (EGO) sensor is nominally operating and if the sensor is defective. The EGO sensor can be monitored to determine the type of drift (rich or lean) in the air-fuel ratio control that has occurred (whether the air-fuel ratio deviates from fuel rich or fuel lean).
The method uses the measurement of several characteristics of sensor operation to determine the condition of the sensor, eliminate false alarms and provide accurate diagnostics. The system is efficient in that the sensor is advantageously diagnosed within a 25 limited cycle switch at selected operating points. This is typically about 1 operating time
Equivalent to less than a minute.

The diagnostic technique is based on trainable classification means or indicators or items trained, ie characterized, by reference sensors having known operating characteristics.
Reference sensors include normal sensors and lean or rich engine operation, as well as some sensors having anomalous properties that result in emission levels that are outside of specified limits. These trainable classifiers form the basis for decision rules used to analyze the signals from EGO sensors to be classified as rich, normal or lean. The data from these reference sensors are rich limit voltage, lean-rich switching time, lean limit voltage and rich-
Includes lean switching time. Additionally, the standard deviation of these distributions from each reference sensor characterized at the selected operating point is calculated.

[0009] Typically, EGO sensors can be divided into three types: normal, rich and lean. Training sample data is collected from a reference sensor operating in situ on a nominal vehicle of the type to which diagnostics are applied and statistics are calculated. The diagnostic decision rules are developed in the realistic environment of the vehicle. Using a reference sensor to generate all the data used to train a classifier is relatively time consuming and expensive. As a complementary technique to ensure robust classifier performance, additional data with the same statistics can be generated by Monte Carlo methods for training and testing purposes. The use of trainable classifiers and pattern recognition for diagnostics is entitled "Diagnostic System Using Pattern Recognition for Electronic Vehicle Controllers" assigned to the assignee of the present application. US Pat. Nos. 5,041,976 and 1990
Neural Network (n
K. K., presented at an international joint conference on oral networks). A. Marco, L.A. A. Feldcamp, and G. V. See Pascolius in "Vehicle Diagnostic Methods Using Trainable Classifiers, Statistical Testing and Paradigm Selection" for reference.

The data from the reference sensor is grouped into batches of data at similar operating points, 10 to 25.
Sample EGO switches are sampled batch by batch. The sample mean and standard deviation are calculated and stored as an 8-element "data vector". That is, each type of data (V _L , V _R , T shown in FIG. 3)
_LR , and T _RL ), one sample with multiple data points is taken and then the average value (av
g. ) And standard deviation (s.d.) are taken and the eight subvectors (ie, V _L (avg.), V _R (av).
g. ), T _LR (avg.), T _RL (avg.), L
_{L (s.d.), V R} (s.d.), T LR (s.
d. ), T _RL (s, d.)). The number of samples contained in a batch is determined by the averaging needed to provide both an early indication of defects and an acceptably low probability of false alarms. The sample size is also subject to the constraint that diagnostic evaluation must be available during a standard drive cycle.
This constraint established an upper bound of about 50-100 switches on the size of the sample window or sample window. These data vectors are associated with one of three classes, and with radial basis functions (radial basis).
It is used to train either a basis function classifier or a hyper plane classifier.

The data from the 16 sample sensors were evaluated in this way and this classification method was able to accurately identify all defect sensors and the nature of the defects. The tests were based on unseen data with real sensors or Monte Carlo data generated for testing purposes. Three-dimensional subspace of eight-dimensional space (subspa
(ce) is shown in FIG. 1 in which three classes of separation can be seen. In a perfect eight-dimensional space, the separation is even clearer.

Referring to FIG. 1, the sensor characteristics using the cluster diagnostic method are: rich-lean gradient, lean-rich gradient,
And with a tri-axial representation including the axes of the rich voltage. As an illustrative example, an elliptical decision surface is shown for classification of sparse sensors in a three-dimensional subspace. Advantageously, the real decision surface is determined in an eight-dimensional hyperspace that cannot be easily illustrated. The points shown in FIG. 1 within the oval surface and represented as "+" are sample averages from a single sensor. Each point in the cluster represents the average of a sample of data points from one sensor.

A flow chart for generating decision rules using training algorithms and diagnostics is shown in FIG. The procedure outlined in the flow chart uses the Monte Carlo method to determine the averaging window required to obtain a defined diagnostic accuracy based on a precise evaluation of diagnostic performance developed from a small sample of real data. These estimates make it possible to establish realistic limits on the calculations needed to obtain the desired precision. That is, one reference sensor may have measured data and data generated by the Monte Carlo method. The number of measurements of each type of data point (eg (ee), V _L ) to achieve the desired accuracy is the sample size that defines the window. The larger the sample size, the smaller the difference between the sample means.

For data analyzed from 14 sample sensors, a backpropagation algorithm with a three-layer network allows 100% diagnostic accuracy in an averaging window based on 16 samples per evaluation. Gave birth to a classification method to The classification categories in this case were rich, normal, and lean.

Referring to FIG. 2, the heated exhaust gas oxygen sensor (HEGO) diagnostic decision rule flow chart begins at block 10 where sample data is obtained from the steady state of the reference HEGO sensor. The logic flow then proceeds to block 11, where statistics on the sample distribution are calculated. The logic flow then proceeds to block 12, where the Monte Carlo method is advantageously used to generate an augmented data set with the same statistics. The logic flow then proceeds to block 13, where the sample window size is selected for the first trial by the classifier. The logic flow then proceeds to block 14, where the classifier is trained and performance is evaluated based on Monte Carlo data. The logic flow then proceeds to decision block 15, where the question "Performance is OK?" Is asked regarding classifier performance. If not (NO), logic flow returns to block 13. If yes (YES), logic flow proceeds to block 16 and a trainable classifier strategy is selected. Logic flow then proceeds to decision block 17
, Where the question “is the decision rule simple?” Is asked about the classifier. If not (NO), the logic flow goes to block 19 where the trainable classifier is deployed. If simple (YES), logic flow proceeds to block 18, where the decision plane equation is developed, then logic flow proceeds to block 20,
Therefore, conventional classification means are developed.

The following is a more detailed description of the schematic elements of FIG. Referring to block 10, data is obtained from various sensors and the emissions level control capabilities of these sensors are measured in separate tests. The transition of the sensor in response to engine exhaust gas flow is measured and the following parameters are: V _L , lean voltage level, V _R , rich voltage level, T _LR , rich lean transition time, T _RL , lean rich transition time, Is determined. Typically, these values are about 30-
It was measured at multiple speed-load points over the time 50 transitions were seen. To improve accuracy, this data is used to calculate sample statistics and generate a much larger data set for study.

Referring to block 11, the mean and standard deviation of the sample data are calculated. In addition, the covariance matrix is calculated. These statistics characterize the nature of the raw signals and their correlations and form the basis of the calculations performed in block 12.

Referring to block 12, block 11
More data is generated using the Monte Carlo method according to the statistics determined in. This process is desirable to avoid the cost of laboratory measurement time and to more thoroughly test the performance of diagnostic algorithms that operate on data. The data produced is calculated to have the same sample mean, sample variance and covariance as the original data sample. The covariance of the data is usually estimated to be zero, but it is often nonzero, and nonzero covariance has a significant impact on diagnostic accuracy. Therefore, classifier performance must be evaluated by proper data simulation.
Assuming that the augmented data generated by the Monte Carlo method is large, the logic flows to block 13, which extracts test samples from this dataset for evaluation.

Referring to blocks 13 and 14, a test sample from this data is derived from the average of the measurements from which the sample averages of the four variables and their variances are calculated. These eight variables form the input vector for the data classification task. In general, for longer averaging windows, there is less sample-to-sample change in the mean and diameter of clusters of points for a particular sensor. Within the limit of infinite time, the average of the samples within the window will probably reduce to a point for each type of sensor, as the averaged measurements characterize each sensor with little uncertainty. Let's do it.

Thus, in this diagnostic strategy, one
The average of information from transitions and information from many transitions is
Used together to reduce sound. That is, one
Information from one transition is V_L, V_R, T_LRAnd T_RLAnd
It The average of the information is that for each particular Thailand from a very large number of transitions.
Average of the data of the group, that is, V_L(Avg.), V _R
(Avg.), T_LR(Avg.), T_RL(Avg.)
is there. This process results in a relatively long wait for an answer
But the answers will be more reliable. Sump shown in Figure 1
The data cluster that represents the
Have a larger sample size, and
The equalization time will be longer. Suitable window size
And the process of determining the associated averaging time is
Choosing a reasonable time as a point, and a window
C. Evaluate the ability to classify data based on averaging time
Can be achieved by Usually a very short time
Produce unsatisfactory results. Increasing time is diagnostic
Reduce misclassification. Allowable averaging time once
If is established, the logic flow goes to block 15.
At block 15, the performance of the classifier is adjusted.
Be assessed. That is, the classifier gives the correct answer
Or does it conclude too generically? Do this
One method has not previously been used in connection with classifiers.
Applying a classification method to data that does not exist. Characteristics of data
Is known, and therefore the behavior of the classifier on the data
Can be evaluated.

Referring to block 16, it was determined that the appropriate choice of averaging time, consistent with the need for diagnostic accuracy and the requirement to complete the evaluation within a reasonable time (eg, about 30 seconds to 1 minute) was made. Sometimes the most efficient decision rule can be selected to accomplish the task. That is, a decision strategy is selected based on performance. This task is roughly equivalent to determining what surface and shape of the surfaces in the data space will be used to distinguish between good and bad sensors and which of the eight parameters should be retained. It can be used with the proviso that all the sensors on one side of the infinite plane due to the decision surface are good and all the sensors on the other side are bad. More complex surfaces may be used (paraboloids as an example of simple surfaces, or more complex convoluted surfaces). The closed surface may be used to enclose or contain some type of sensor. However, open surface classifiers classify all new data and closed surface classifiers are often chosen because they are more restrictive and predictable or predictable. The open surface classifier or hyperplane classifier classifies many data points that were not exposed. Because the condition that one point lies on one or the other side of the surface is very poorly descriptive. This type of classifier concludes generally, but always provides some answers. Closed surface or radial basis function classifiers are often preferred because they better limit their answers to the area trained there. Therefore,
These classifiers can be better trusted to classify good sensors, ie those sensors that are located within a certain range of what they were trained for. In this case, all cases of measurement that fall inside such a closed surface are of one kind, and all other cases that fall outside of said surface belong to some other kind. Needless to say, there are more complex permutations of these formulas such as multiple closed surfaces, mixed closed surfaces and hyperplanes. In addition, perhaps eight parameters are very necessary to achieve the required diagnostic accuracy. The ultimate goal is to determine the simplest set of decision rules or simplest classifiers that can accomplish the task. Therefore, several different classification paradigms can be tested and then the one that performs the task most accurately (with the fewest errors) with the least computational cost can be chosen. But,
This disclosed embodiment is based on a scheme that utilizes all information to achieve the highest possible accuracy.

Referring to block 17, if the classifier (eg, neural network) is very small (eg, 8-input, 2-hidden node, 2-output category), then the network is a set of simple decision planes or small ones. It can be approached by a set of rules. This technique is routinely used in conventional diagnostics. According to an embodiment of the invention, simple rules are valid, but not optimal, and it is preferable to train the classifier and let it determine the optimal decision boundaries, rather than recursively determining the rules. However, the best solution depends on the characteristics of the data, the nature of the sensor (sensors from different manufacturers operate in very different ways) and diagnostic accuracy requirements.

At block 17, a question is asked as soon as a simple set of decision rules can be derived from the trained classifier. In some (rare) situations, the trained classifier may internally formulate a set of simple decision surfaces to classify the data it is exposed to. Moreover, product engineers often prefer a concise list of rules for decision making rather than relying on a trained system with an unknown internal representation of the problem. The present invention employs a simple process at block 17 to determine if such a simple set of rules is drawn from the trained classifier. The method determines whether the classifier is small enough to be trapped within a set of rules whose behavior is small (ie, just a few nodes in the hidden layer), and the coarse binning of the input vector values ( coarse b
is a simple decision to allow a small set of rules to be derived in order to reproduce the classifier output. To this end, the present invention uses an input function or feature space (input space) for each component of the input vector.
feature space), for example, when the voltage is 0
0.33V, 0.33-0.66V, 0.66-1.0
It is divided into small, medium, and large values corresponding to V, respectively. If these rough input value (e.g., V _L - small, and V _R - in, and T _LR - large mean that the sensor better) if simple relationship between the proper classification output is present, the Relationships can be recognized by induction or derivation, and as such are, in fact, captured within a set of rules that are a very crude expression of a classifier. The rule induction process yields an unambiguous representation of the classifier, but at the expense of less durable classifiers. Block 17 provides a means for the product technician to choose the right balance of simplicity and performance required for on-board diagnostic tasks.

In practical operation, it is advantageous that the sensor is preheated for about 1-2 minutes after engine start-up, before the sorter training is started. During the measurement, the sensor is observed under preheat conditions as well, maintaining the engine speed load characteristics in a stable state and advantageously as well as in the situation in which the classifier was trained.

Undoubtedly, various modifications and alterations will occur to those skilled in the art to which the present invention pertains. For example, the particular method of obtaining the sample data may differ from that disclosed herein. All of these and other modifications, through which the specification essentially depends on the teachings proposed in the art, are properly considered to be included within the scope of the present invention.

[Brief description of drawings]

FIG. 1 is a graphic diagram showing decision information according to an embodiment of the present invention.

FIG. 2 is a logic flow diagram for deploying a diagnostic technique for classifying exhaust gas oxygen sensors according to one embodiment of the present invention.

FIG. 3 is a graphic diagram showing the exhaust gas oxygen sensor voltage output and measured values used as a trainable classifier.

[Explanation of symbols]

 10 blocks 11 blocks 12 blocks 13 blocks 14 blocks 15 decision blocks 16 blocks 17 decision blocks 18 blocks 19 blocks 20 blocks

Claims (12)

[Claims] 1. A method of diagnosing the operation of an unclassified exhaust gas oxygen sensor and an electronic engine controller, the process of storing signal parameters characterizing the operation of a plurality of reference exhaust gas oxygen sensors, the unclassified exhaust gas oxygen sensor for a certain time. Collecting the signal parameters that characterize the operation of the, and comparing the currently collected signal parameters that characterize the unclassified exhaust gas oxygen sensor with previously stored signal parameters that represent the operating parameters of multiple reference exhaust gas oxygen sensors, And a method of diagnosing the operation of an unclassified exhaust gas oxygen sensor and an electronic engine controller, the method comprising determining an operating characteristic of the unclassified exhaust gas oxygen sensor based on a comparison of the collected signal parameter with a previously stored signal parameter. 2. The method of claim 1, wherein the step of comparing the currently collected signal parameter with a stored signal parameter comprises selecting a plurality of basic classifiers. Calculating statistics characterizing the sample distribution of stored signal parameters from the sensor, a mathematical technique to generate an additional augmented data set with similar statistics to the stored signal parameters from the reference exhaust gas oxygen sensor , The process of selecting the size of the sample window for the initial application of the classifier to the sample distribution, the classifier to correctly evaluate the sample distribution of the signal parameters from the reference exhaust gas oxygen sensor. Training (trai
in the process of characterizing the signal parameters from the reference exhaust gas oxygen sensor, the performance of the classification means (performa).
, the process of establishing a predetermined level of performance of the classifier, the process of determining a satisfactory performance when the classifier evaluates the sample distribution at a level better than the predetermined level of its performance, multiple trained Using the process of selecting a trainable classifier strategy from the basic classifiers and the process of deploying the trainable classifier to evaluate the unclassified exhaust gas oxygen sensor,
A method comprising generating a classifier. 3. The method of claim 2, wherein the step of storing signals characterizing the operation of a plurality of reference exhaust gas oxygen sensors characterizes the operation of a plurality of reference exhaust gas oxygen sensors as pass or fail. , And the voltage representing lean operation,
A method comprising measuring the output voltage of each of a plurality of reference exhaust gas oxygen sensors to determine a voltage representative of rich operation, a rich voltage to lean voltage transition time, and a lean voltage to rich voltage transition time. 4. The method of claim 3, further comprising the step of measuring output voltage data from each reference exhaust gas oxygen sensor multiple times to produce a sample, calculating the mean and standard deviation for each sample distribution. A method, and a method of characterizing an output voltage of each of a plurality of reference exhaust gas oxygen sensors. 5. The method of claim 4, further comprising using a Monte Carlo technique to generate a data distribution that characterizes the additional reference exhaust gas oxygen sensor, and the output of such additional reference exhaust gas oxygen sensor. A method having a process of calculating a mean value and a standard deviation for each data distribution to be attached. 6. The method of claim 4, further comprising the step of reducing the coverage of each trainable classifier to better classify the sensor to be classified. 7. The method of claim 6, wherein the trainable classifying means defines a closed surface, the first class of measurements within the closed surface and the second class of measurements outside the closed surface. How to establish. 8. The method according to claim 4, wherein the number of signal parameters characterizing the operation of the unclassified exhaust gas oxygen sensor is at least 16. 9. The method of claim 4, wherein the step of collecting signal parameters characterizing the operation of the unclassified exhaust gas oxygen sensor for a period of time comprises: lean voltage magnitude, rich voltage magnitude, rich voltage magnitude. The process of measuring multiple output voltage waveform characteristics of the unclassified exhaust gas oxygen sensor voltage output including the lean voltage transition time and the lean rich voltage transition time, and the average value and standard deviation for each type of measurement data that characterizes the output voltage waveform. A method comprising the steps of calculating and forming a single vector characterizing the sample using the data mean and data standard deviation. 10. The method of claim 9, wherein the step of storing signal parameters characterizing the operation of a plurality of reference exhaust gas oxygen sensors includes: lean voltage magnitude, rich voltage magnitude, rich lean voltage transition. Process of measuring multiple output voltage waveform characteristics of the reference exhaust gas oxygen sensor voltage output, including time and lean rich voltage transition time, and the mean and standard deviation for each type of measurement data characterizing the output voltage waveform of the reference exhaust gas oxygen sensor A method including the step of calculating. 11. The method of claim 8 wherein the step of storing signal parameters characterizing the operation of a plurality of reference exhaust gas oxygen sensors further comprises: average lean voltage, average rich voltage, average rich lean transition time,
A method comprising forming a vector containing elements of average lean rich transition time, lean voltage standard deviation, lean voltage standard deviation, rich lean transition time standard deviation, and lean rich transition time standard deviation. 12. The method of claim 11, wherein the currently collected signal parameters characterizing the unclassified exhaust gas oxygen sensor are compared with previously stored signal parameters representative of operating parameters of a plurality of reference exhaust gas oxygen sensors. The method wherein the step comprises comparing the vector characterizing the unclassified exhaust gas oxygen sensor with a training classifier associated with each data point.