JP7349124B2 - Method and program for determining the condition of a prime mover - Google Patents

Description

The present invention relates to a method and program for determining the state of a prime mover.

Internal combustion engines such as gas engines are often operated in conditions close to the occurrence of knocking in order to improve thermal efficiency. FIG. 14 is a state diagram showing a state in which knocking occurs, a state in which stable combustion occurs, and a state in which misfire occurs, with the horizontal axis representing the fuel-air ratio and the vertical axis representing the average effective pressure within the combustion chamber. In FIG. 14, near the boundary region from the stable combustion region to the knocking occurrence region, there is a possibility that pressure oscillations may occur as shown in FIG. 15 due to spontaneous combustion in the combustion chamber. Such pressure oscillations create the potential for damage to the internal combustion engine.

Therefore, a technique for analyzing knocking that occurs in the combustion chamber of an internal combustion engine has been disclosed. For example, a noise signal sensed by a knock sensor placed in or near the combustion chamber of a combustion engine is pre-conditioned to generate a pre-conditioned noise signal, and the pre-conditioned noise signal is subjected to wavelet processing etc. A technique for determining the position or time of peak combustion pressure in a combustion chamber of an engine is disclosed (Patent Document 1).

In addition, the system performs discrete wavelet transform on the original signal, calculates wavelet coefficients at two levels, identifies the necessity of noise removal based on the wavelet coefficients, and determines the need for noise removal based on the wavelet coefficients. A technique is disclosed in which the level of the coefficient is suppressed and the presence or absence of knocking is determined based on the suppressed wavelet coefficient (Patent Document 2).

In addition, the signal detected by the knock sensor that detects the vibration of the engine's cylinder block is stored in correspondence with the crankshaft angle, and the signal is divided into short time intervals and frequency analysis (FFT) is performed for each time interval. A technique for determining whether or not knocking has occurred based on the results has been disclosed (Patent Document 3).

In addition, the technology analyzes the frequency of the signal output from the knock sensor and determines the presence or absence of knocking based on the correspondence between the number of times fuel is injected during one combustion cycle and the frequency component correction amount. is disclosed (Patent Document 4).

Japanese Patent Application Publication No. 2016-169735 Japanese Patent Application Publication No. 2011-1867 Japanese Patent Application Publication No. 2005-188297 International Publication No. WO2016/125687 pamphlet

However, the conventional technology has been insufficient as a technology for early and appropriate detection of abnormalities such as knocking in a prime mover including an internal combustion engine. Therefore, it is difficult to detect an abnormality such as knocking while it is still mild and to issue an alarm or take measures in response to the alarm.

An object of the present invention is to provide a state determination method and a state determination program that enable early and appropriate detection of abnormalities such as knocking in a prime mover including an internal combustion engine.

A method for determining the condition of a prime mover according to claim 1 of the present invention includes acquiring at least one learning data of sound, vibration, and pressure fluctuation indicating the operating condition of the prime mover, and performing frequency analysis of the acquired learning data. Performing singular value decomposition (SVD) on the result of the frequency analysis, further performing machine learning using the feature space extracted by the singular value decomposition (SVD) as input data and outputting the operating state of the prime mover, A judgment prediction means is constructed that derives judgment information on the operating state of the prime mover based on the results of machine learning and clusters input data, and acquires judgment target data indicating the operating status of the prime mover. The frequency analysis of the judgment target data is performed, the singular value decomposition (SVD) is performed on the result of the frequency analysis, and the feature quantity extracted by the singular value decomposition (SVD) is used as the input data to predict the judgment. The method is characterized in that the judgment information is input to a means and the judgment information for the judgment target data of the prime mover is derived .

Here, it is preferable to perform an analysis using wavelet transform as the frequency analysis.

Further, as the machine learning, it is preferable to perform learning using a support vector machine (SVM) that performs pattern recognition using supervised learning on the feature space extracted by the singular value decomposition (SVD). be.

Furthermore, for the feature space extracted by the singular value decomposition (SVD), the degree of membership to clusters is determined using the Fuzzy C-Means method, and the degree of membership is calculated using the support vector machine (SVM). It is preferable to apply

Furthermore, as the machine learning, it is preferable to use self-organizing mapping (SOM), which is unsupervised learning, for the feature space extracted by the singular value decomposition (SVD).

Further, it is preferable that the operation of the prime mover is further controlled based on the judgment information of the operating state.

Further, it is preferable that the prime mover is an internal combustion engine and/or a movable part attached to the internal combustion engine.

Further, it is preferable that the determination information on the operating state is derived from the determination information on the knocking state of the internal combustion engine.

Further, it is preferable that a knocking index is derived as the determination information of the knocking state.

Further, it is preferable that control is performed to eliminate the knocking state of the internal combustion engine based on the determination information of the knocking state.

A prime mover state determination program according to an eleventh aspect of the present invention includes a step of causing a computer to acquire learning data of at least one of sound, vibration, and pressure fluctuation indicating the operating state of the prime mover; A frequency analysis step that performs frequency analysis of learning data, a singular value decomposition step that performs singular value decomposition (SVD) on the result of the frequency analysis, and inputs the feature space extracted by the singular value decomposition (SVD). a machine learning step of performing machine learning using the operating state of the prime mover as an output as data; and a judgment prediction means for deriving judgment information of the operating state of the prime mover based on the result of the machine learning and clustering the input data. a construction step of constructing a structure, acquiring judgment target data indicating the operating state of the prime mover, performing the frequency analysis of the acquired judgment target data, and performing the singular value decomposition (SVD) on the result of the frequency analysis. and further inputting the feature quantity extracted by the singular value decomposition (SVD) to the judgment prediction means as the input data, and executing an output step of outputting the judgment information for the judgment target data of the prime mover. Features.

Here, it is preferable to perform an analysis using wavelet transform as the frequency analysis in the frequency analysis step.

Further, as the machine learning in the machine learning step, the support vector machine (SVM) performs pattern recognition using supervised learning on the feature space extracted by the singular value decomposition (SVD) in the singular value decomposition step. ) is preferable.

Further, in the singular value decomposition step, the feature space extracted by the singular value decomposition (SVD) is applied to the support vector machine (SVM) using the Fuzzy C-Means method. It is preferable to obtain the degree of membership to a cluster for the purpose of determining the degree of membership in a cluster.

Further, as the machine learning in the machine learning step, it is preferable to perform learning using self-organizing mapping (SOM), which is unsupervised learning, on the feature space extracted by the singular value decomposition (SVD). It is.

Further, it is preferable that the engine further comprises an operation control step of controlling the operation of the prime mover based on the judgment information of the operation state in the output step.

Further, it is preferable that the prime mover is an internal combustion engine and/or a movable part attached to the internal combustion engine.

Further, in the output step, it is preferable that determination information on a knocking state of the internal combustion engine is outputted as the determination information on the operating state.

Further, in the outputting step, it is preferable that a knocking index is derived and outputted as the determination information of the knocking state.

Further, it is preferable that the operation of the internal combustion engine and/or the movable part attached to the internal combustion engine is controlled in the operation control step based on the determination information of the knocking state in the output step.

According to the state determination method according to claim 1 of the present invention, at least one learning data of sound, vibration, and pressure fluctuation indicating the operating state of the prime mover is acquired, and frequency analysis of the acquired learning data is performed. Performing singular value decomposition (SVD) on the result of the frequency analysis, further performing machine learning using the feature space extracted by the singular value decomposition (SVD) as input data and outputting the operating state of the prime mover, A judgment prediction means is constructed that derives judgment information on the operating state of the prime mover based on the results of machine learning and clusters input data, and acquires judgment target data indicating the operating status of the prime mover. The frequency analysis of the judgment target data is performed, the singular value decomposition (SVD) is performed on the result of the frequency analysis, and the feature quantity extracted by the singular value decomposition (SVD) is used as the input data to predict the judgment. By inputting the judgment information into the means and deriving the judgment information for the judgment target data of the prime mover , it is possible to appropriately judge the operating state of the prime mover.

Here, by performing analysis using wavelet transform as the frequency analysis, machine learning regarding the operating state of the prime mover can be appropriately performed.

In addition, as the machine learning, learning using a support vector machine (SVM) that performs pattern recognition using supervised learning is performed on the feature space extracted by the singular value decomposition (SVD). Appropriate machine learning can be performed to classify the operating state of the device into multiple clusters.

Furthermore, for the feature space extracted by the singular value decomposition (SVD), the degree of membership to clusters is determined using the Fuzzy C-Means method, and the degree of membership is calculated using the support vector machine (SVM). By applying this method, it is possible to appropriately perform supervised machine learning for classifying the operating state of a prime mover into multiple clusters.

In addition, as the machine learning, by using self-organizing mapping (SOM), which is unsupervised learning, on the feature space extracted by the singular value decomposition (SVD), the operating state of the prime mover is divided into multiple clusters. Machine learning for classification can be performed appropriately.

Further, by further controlling the operation of the prime mover based on the judgment information of the operating state, the prime mover can be appropriately controlled based on the judgment result of appropriately determining the operating state of the prime mover.

Further, since the prime mover is an internal combustion engine and/or a movable part attached to the internal combustion engine, the internal combustion engine And/or movable parts associated with the internal combustion engine can be appropriately controlled.

Furthermore, by deriving determination information on the knocking state of the internal combustion engine as the determination information on the operating state, it is possible to appropriately determine the knocking state of the internal combustion engine.

Furthermore, by deriving a knocking index as the determination information on the knocking state, a determination result indicating the knocking state of the internal combustion engine can be obtained as the knocking index.

Further, by performing control to eliminate the knocking state of the internal combustion engine based on the determination information of the knocking state, the internal combustion engine is appropriately controlled based on the determination result of appropriately determining the knocking state of the internal combustion engine. can do.

According to the prime mover state determination program according to claim 11 of the present invention, the learning data acquisition step causes the computer to acquire at least one learning data of sound, vibration, and pressure fluctuation indicating the operating state of the prime mover; a frequency analysis step of performing frequency analysis of the training data, a singular value decomposition step of performing singular value decomposition (SVD) on the result of the frequency analysis, and a feature space extracted by the singular value decomposition (SVD). a machine learning step in which machine learning is performed using input data as input data and the operating state of the prime mover as an output; and a judgment prediction step in which judgment information on the operating state of the prime mover is derived based on the result of the machine learning and the input data is clustered. a construction step of constructing a means, acquiring determination target data indicating the operating state of the prime mover, performing the frequency analysis of the acquired determination target data, and performing the singular value decomposition (SVD) on the result of the frequency analysis. ), and further inputs the feature quantity extracted by the singular value decomposition (SVD) to the judgment prediction means as the input data, and executes an output step of outputting the judgment information for the judgment target data of the prime mover. By this, the operating state of the prime mover can be appropriately determined.

Here, by performing analysis using wavelet transform as the frequency analysis in the frequency analysis step, machine learning regarding the operating state of the prime mover can be appropriately performed.

Further, as the machine learning in the machine learning step, the support vector machine (SVM) performs pattern recognition using supervised learning on the feature space extracted by the singular value decomposition (SVD) in the singular value decomposition step. ), it is possible to appropriately perform machine learning for classifying the operating state of the prime mover into multiple clusters.

Further, in the singular value decomposition step, the feature space extracted by the singular value decomposition (SVD) is applied to the support vector machine (SVM) using the Fuzzy C-Means method. By determining the degree of membership to a cluster for the purpose of classifying the operating state of the prime mover into a plurality of clusters, it is possible to appropriately perform supervised machine learning for classifying the operating state of the prime mover into a plurality of clusters.

Further, as the machine learning in the machine learning step, by performing learning using self-organizing mapping (SOM), which is unsupervised learning, on the feature space extracted by the singular value decomposition (SVD), Machine learning for classifying the operating state of a prime mover into a plurality of clusters can be appropriately performed.

Further, by further comprising an operation control step of controlling the operation of the prime mover based on the judgment information of the operating state in the output step, the prime mover is can be controlled appropriately.

Further, since the prime mover is an internal combustion engine and/or a movable part attached to the internal combustion engine, the internal combustion engine And/or movable parts associated with the internal combustion engine can be appropriately controlled.

Furthermore, in the output step, the knocking state of the internal combustion engine can be appropriately determined by outputting determination information on the knocking state of the internal combustion engine as the determination information on the operating state.

Furthermore, in the output step, by deriving and outputting a knocking index as the determination information on the knocking state, a determination result indicating the knocking state of the internal combustion engine can be obtained as the knocking index.

Further, the knocking state of the internal combustion engine is controlled by controlling the operation of the internal combustion engine and/or a movable part attached to the internal combustion engine in the operation control step based on the determination information of the knocking state in the output step. The internal combustion engine can be appropriately controlled based on the result of the appropriate judgment.

FIG. 1 is a functional block diagram showing the configuration of a state determining device according to an embodiment of the present invention. 1 is a diagram showing the configuration of a state determination device according to an embodiment of the present invention. It is a flow chart showing machine learning processing in an embodiment of the present invention. It is a figure showing an example of data for learning in an embodiment of the present invention. FIG. 3 is a diagram showing an example of learning data subjected to wavelet transform and singular value decomposition processing according to an embodiment of the present invention. It is a figure which shows the example of the data for learning which performed the conversion singular value decomposition process in embodiment of this invention. It is a figure which shows the example of the singular value (feature amount) extracted from the data for learning in embodiment of this invention. FIG. 3 is a diagram showing an example of a feature space for learning data in an embodiment of the present invention. FIG. 6 is a diagram showing an example of determining the degree of membership to a cluster by applying the fuzzy sea means method in the embodiment of the present invention. It is a figure which shows the result of machine learning by the support vector machine in embodiment of this invention. It is a figure which shows the result of machine learning by self-organizing mapping in embodiment of this invention. It is a flowchart which shows the judgment process of the operating state in embodiment of this invention. FIG. 3 is a diagram showing an example of a determination result regarding the occurrence of knocking in an internal combustion engine according to an embodiment of the present invention. FIG. 3 is a state diagram showing the operating state of the internal combustion engine. FIG. 2 is a diagram showing temporal changes in pressure within a combustion chamber of an internal combustion engine (crankshaft angle changes).

[Basic configuration]
As shown in FIG. 1, a prime mover state determination device 100 according to an embodiment of the present invention includes a learning data acquisition means 10, a frequency analysis means 12, a singular value decomposition means 14, a machine learning means 16, and a determination target data acquisition means. 18 and judgment prediction means 20.

The condition determining device 100 performs a process of determining the operating state of the prime mover (not shown) based on the data, by acquiring at least one data of sound, vibration, and pressure fluctuation that occurs with the operation of the prime mover (not shown). conduct.

In this embodiment, a prime mover means a machine that converts energy existing in the natural world into mechanical power, and a machine that uses the converted energy in general. For example, the prime mover includes a heat engine, a fluid machine, an electric motor, or the like, or a movable part such as a bearing or a propeller associated therewith.

The state determination device 100 can be realized by a computer including a processing section 30, a storage section 32, an input section 34, an output section 36, and a communication section 38, as shown in FIG.

The processing unit 30 includes a CPU and the like, and performs processing in the state determination device 100 in an integrated manner. The processing unit 30 makes it possible to implement the state determination method in this embodiment by executing the state determination program stored in the storage unit 32. That is, the computer functions as the learning data acquisition means 10, the frequency analysis means 12, the singular value decomposition means 14, the machine learning means 16, the judgment target data acquisition means 18, and the judgment prediction means 20. The storage unit 32 stores information used for processing, such as a state judgment program used in the state judgment processing in the state judgment device 100, learning data used for machine learning, and judgment target data to be subjected to state judgment. The storage unit 32 can be configured with, for example, a semiconductor memory, a hard disk, or the like. The input unit 34 includes means for inputting information to the state determining device 100. The input unit 34 includes, for example, a keyboard, a mouse, and the like. The output unit 36 includes means for displaying information processed by the state determining device 100. The output unit 36 can be, for example, a display. The communication unit 38 includes an interface for exchanging information with external devices. The communication unit 38 enables communication with external devices by being connected to an information communication network such as the Internet or a dedicated line, for example. Note that the combination of the state determination device 100 excluding the processing unit 30 is not limited to this configuration.

[Machine learning processing]
The machine learning process regarding the operating state of the prime mover in the state determining device 100 will be described below with reference to the flowchart of FIG. The state determination device 100 performs the following machine learning process by executing the state determination program stored in the storage unit 32.

In step S10, processing for acquiring learning data for performing machine learning regarding the operating state of the prime mover is performed. Through the processing in this step, the state determining device 100 functions as the learning data acquisition means 10. The processing unit 30 acquires data for performing machine learning via the input unit 34 or the communication unit 38. For example, time-series data indicating temporal changes in at least one of the sounds, vibrations, and pressure fluctuations generated in conjunction with the operation of the prime mover is acquired by sensors installed on or around the prime mover to be analyzed. The acquired data is stored in the storage unit 32.

FIG. 4 shows, as an example of learning data, changes in pressure within the combustion chamber with respect to the crankshaft angle of the internal combustion engine. However, the learning data is not limited to this, and may also be acoustic data or vibration data generated with the operation of the prime mover.

In step S12, frequency analysis processing is performed on the learning data. Through the processing in this step, the state determination device 100 functions as the frequency analysis means 12. The processing unit 30 performs frequency analysis processing to convert the time series data acquired in step S10 into frequency data in a frequency space. As the frequency analysis process, it is preferable to apply, for example, wavelet transformation. Wavelet transform is one of the methods of frequency analysis, and is frequency analysis using a wavelet function as a basis function. Basically, the waveform of the given learning data is expressed by scaling and translating small waves (wavelets) and adding them together. In wavelet transform, the basis functions are scaled up and down, so it is possible to analyze a wide frequency domain.

In step S14, singular value decomposition processing is performed on the learning data. Through the processing in this step, the state determination device 100 functions as a part of the singular value decomposition means 14. The processing unit 30 performs singular value decomposition processing (SVD) on the learning data subjected to frequency analysis in step S12. Singular value decomposition is a method of matrix decomposition for matrices whose components are complex numbers or real numbers in linear algebra. Note that the singular value decomposition (SVD) is preferably a transformed singular value decomposition (TSVD) that emphasizes singular values more when performing singular value decomposition of a matrix.

FIG. 5 shows an example of learning data that has been subjected to wavelet transformation and singular value decomposition processing. Further, FIG. 6 shows an example of the learning data that has been subjected to a transformed singular value decomposition process (TSVD) that performs spatial transformation of the learning data shown in FIG. 5 .

In step S16, feature amount extraction processing is performed from the learning data that has been subjected to singular value decomposition processing. Through the processing in this step, the state determination device 100 functions as a part of the singular value decomposition means 14. In the learning data subjected to the singular value decomposition process, a state having a signal energy equal to or higher than a predetermined reference value is extracted as a feature amount indicating the feature of the learning data.

FIG. 7 shows an example of the feature amount extracted from the example of the learning data shown in FIG. In FIG. 7, σi (tilt) indicates signal energy, ti (bar) indicates crankshaft angle (time), fi (bar) indicates frequency, and ti (hat) indicates crankshaft angle width (time span). FIG. 7 shows an example in which the reference value for signal energy is 0.65, and five states having energy of 0.65 or more are extracted as feature quantities.

Since machine learning requires extraction of feature amounts from various learning data, it is preferable to repeat the processes from step S10 to step S16 as necessary. FIG. 8 shows an example in which feature quantities extracted from a large amount of learning data are plotted in a feature space. In the example shown in FIG. 8, the feature quantity extracted from at least one learning data of sound, vibration, and pressure fluctuation generated with the operation of the prime mover is divided into two in the feature quantity space expressed by the crankshaft angle and frequency. It is divided into clusters (taxonomic groups).

In step S18, machine learning is performed to determine the operating state of the prime mover using the feature amount of the learning data obtained by singular value decomposition as input data. Through the processing in this step, the state determining device 100 functions as the machine learning means 16. Machine learning is not particularly limited as long as it is capable of cluster classification to classify the operating state of the prime mover into one of a plurality of states, but for example, support vector machine (SVM), It is possible to select and apply from Self-Organizing Maps (SOM) and the like.

When applying support vector machine (SVM) as machine learning, the degree of belonging to a cluster (taxonomic group) is determined in advance for the feature values that are the input data used for learning, and the degree of belonging to the cluster is associated with the feature values. and used as data for supervised learning. For example, by applying the Fuzzy C-Means method to the feature space obtained by singular value decomposition, the degree of membership of the feature to a cluster can be determined.

FIG. 9 shows an example in which the degree of membership of each feature to a cluster is determined by applying the Fuzzy C-Means method to the feature space shown in FIG. 8. The contour lines in FIG. 9 indicate the degree of belonging to cluster 1 (Cluster 1) and cluster 2 (Cluster 2) determined for each feature amount.

The degrees of membership to clusters obtained in this way are used as teacher data in combination with the feature quantities obtained by singular value decomposition, and a support vector machine is applied as data for supervised learning. Support vector machines utilize linear input elements to construct pattern discriminators for classes. Specifically, the parameters of the linear input element are learned based on the criterion (hyperplane separation theorem) of finding a margin-maximizing hyperplane that maximizes the distance to each feature based on supervised learning data. As a result, as shown in FIG. 10, it is possible to obtain a discriminator (judgment prediction means 20 in this embodiment) that clusters input data into cluster 1 and cluster 2.

Further, as machine learning, self-organizing mapping (SOM) may be applied instead of support vector machine (SVM). When applying self-organizing mapping (SOM) as machine learning, machine learning is performed using unsupervised learning data consisting of a feature space obtained by singular value decomposition as input data. Self-organizing mapping is a type of neural network, and is a discriminator (in this embodiment, a discriminator (judgment prediction) that maps input data to an arbitrary dimension by unsupervised learning and clusters the input data into cluster 1 and cluster 2. Means 20) can be obtained. FIG. 11 shows an example of a learning result when self-organizing mapping is applied to the feature space shown in FIG. 8. As shown in FIG. 11, by applying self-organizing mapping, it is possible to obtain a classifier (in this embodiment, the judgment prediction means 20) that clusters input data into a plurality of clusters.

As described above, the state determination device 100 uses at least one of sound, vibration, and pressure fluctuation that occurs with the operation of the prime mover as learning data through machine learning processing, and determines a plurality of operating states of the prime mover based on the data. A classifier (judgment prediction means 20) is generated that determines which cluster (classification group) belongs to.

[Operation status judgment processing]
Hereinafter, with reference to the flowchart of FIG. 12, a process of determining the operating state of the prime mover from input data using a classifier (judgment prediction means 20) obtained by machine learning processing will be described. The state determining device 100 executes the state determining program stored in the storage unit 32 to perform the following operation state determining process.

In step S20, a process is performed to obtain input data indicating the operating state of the prime mover whose operating state is to be determined. Through the processing in this step, the state determination device 100 functions as the determination target data acquisition means 18. The processing unit 30 acquires judgment target data indicating the operating state of the prime mover via the input unit 34 or the communication unit 38. For example, time-series data indicating temporal changes in at least one of sound, vibration, and pressure fluctuations generated with the operation of the prime mover is acquired by a sensor provided on the prime mover. The acquired data is stored in the storage unit 32. The judgment target data obtained by this process is similar to the learning data shown in FIG. 4.

In step S22, frequency analysis processing is performed on the judgment target data. Through the processing in this step, the state determination device 100 functions as the frequency analysis means 12. This process is similar to step S12 above, so the explanation will be omitted.

In step S24, singular value decomposition processing is performed on the judgment target data. Through the processing in this step, the state determination device 100 functions as a part of the singular value decomposition means 14. This process is similar to step S14 above, so the explanation will be omitted.

In step S26, feature amount extraction processing is performed from the judgment target data that has been subjected to singular value decomposition processing. Through the processing in this step, the state determination device 100 functions as a part of the singular value decomposition means 14. This process is similar to step S16 above, so the explanation will be omitted.

In step S28, a process is performed to determine the operating state of the prime mover using the feature amount of the determination target data obtained by singular value decomposition as input data. Through the processing in this step, the state judgment device 100 functions as the judgment prediction means 20. At this time, a determination result indicating the result of determining the operating state of the prime mover is output by inputting the feature amount of the determination target data to the discriminator (determination prediction means 20) obtained by machine learning in step S18.

The judgment result output from the judgment prediction means 20 can be used to control the operation of the prime mover. For example, when it is determined that an abnormality has occurred in the operating state of the prime mover, the operating conditions of the prime mover may be adjusted so that the abnormality is resolved. Furthermore, if a determination result is obtained that an abnormality has occurred in the operating state of the prime mover, an alarm may be output using the output unit 36 or the communication unit 38, and maintenance of the prime mover may be performed in response to the alarm. . A method may be adopted in which the operating status is always displayed and the display method is changed when an abnormality occurs.

FIG. 13 shows a discriminator (judgment/prediction means 20) that has been machine-learned using learning data to determine pressure changes in the combustion chamber with respect to the crankshaft angle of the internal combustion engine. This shows the results of inputting data to determine whether knocking has occurred. In FIG. 13, the horizontal axis indicates the number of judgments made by acquiring judgment target data (Analysed Instances: one plot indicates 10 combustion cycles), and the vertical axis indicates the index indicating that knocking has not occurred (Non- Knocking Index) and an index indicating that knocking has occurred (Knocking Index). In FIG. 13, as the number of combustion cycles increases, the determination result indicating the occurrence of knocking is output around the time when the number of determinations exceeds 500.

The state determining device 100 may perform control to eliminate knocking in the internal combustion engine based on the determination result indicating the occurrence of knocking. For example, the fuel-air ratio of the internal combustion engine may be adjusted in accordance with an index indicating that knocking is occurring to prevent knocking from occurring.

INDUSTRIAL APPLICATION This invention can be utilized for determining the operating state of a prime mover. In particular, gas engines are often operated in conditions close to the occurrence of knocking in order to increase thermal efficiency, and it is effective to use this method in gas engines used in ships where load fluctuations are likely to occur. Due to regulations on sulfur content, diesel engines use compliant heavy oil as marine fuel, but it tends to cause knocking under low load conditions, so it is necessary to switch to diesel oil. Therefore, by utilizing the present invention to determine the occurrence of knocking in a diesel engine, it is possible to appropriately control fuel switching.

10 learning data acquisition means, 12 frequency analysis means, 14 singular value decomposition means, 16 machine learning means, 18 judgment target data acquisition means, 20 judgment prediction means, 30 processing section, 32 storage section, 34 input section, 36 output section , 38 communication section, 100 status determination device.

Claims (20)

Acquire at least one learning data of sound, vibration, and pressure fluctuation indicating the operating state of the prime mover, perform frequency analysis of the acquired learning data, and perform singular value decomposition (SVD) on the result of the frequency analysis. Then, machine learning is performed using the feature space extracted by the singular value decomposition (SVD) as input data and the operating state of the prime mover as an output, and judgment information on the operating state of the prime mover is determined based on the result of the machine learning. construct a judgment prediction means that clusters the input data by deriving the above, obtains judgment target data indicating the operating state of the prime mover, performs the frequency analysis of the acquired judgment target data, and calculates the result of the frequency analysis. The singular value decomposition (SVD) is performed on the above-mentioned singular value decomposition (SVD), and the feature quantity extracted by the singular value decomposition (SVD) is inputted as the input data to the judgment prediction means, and the judgment on the judgment target data of the prime mover is performed. A method for determining the state of a prime mover characterized by deriving information. 2. The method for determining the state of a prime mover according to claim 1, wherein the frequency analysis is an analysis using wavelet transform. A claim characterized in that, as the machine learning, learning is performed using a support vector machine (SVM) that performs pattern recognition using supervised learning on the feature space extracted by the singular value decomposition (SVD). The method for determining the state of a prime mover according to claim 1 or claim 2. For the feature space extracted by the singular value decomposition (SVD), the degree of membership to a cluster is determined using the Fuzzy C-Means method and applied to the support vector machine (SVM). 4. The method for determining the state of a prime mover according to claim 3. According to claim 1 or claim 2, the machine learning uses self-organizing mapping (SOM), which is unsupervised learning, for the feature space extracted by the singular value decomposition (SVD). How to judge the condition of the prime mover. 6. The method for determining the condition of a prime mover according to claim 1, further comprising controlling the operation of the prime mover based on the determination information of the operating condition. The method for determining the state of a prime mover according to any one of claims 1 to 6, wherein the prime mover is an internal combustion engine and/or a movable part attached to the internal combustion engine. 8. The method for determining the state of a prime mover according to claim 7 , further comprising deriving determination information on a knocking state of the internal combustion engine as the determination information on the operating state. 9. The method for determining the condition of a prime mover according to claim 8, further comprising deriving a knocking index as the determination information of the knocking condition. The method for determining the state of a prime mover according to claim 8 or 9, wherein control is performed to eliminate the knocking state of the internal combustion engine based on the determination information of the knocking state. to the computer,
a learning data acquisition step of acquiring at least one learning data of sound, vibration, and pressure fluctuation indicating the operating state of the prime mover;
a frequency analysis step of performing frequency analysis of the acquired learning data;
a singular value decomposition step of performing singular value decomposition (SVD) on the result of the frequency analysis;
a machine learning step of performing machine learning using the feature space extracted by the singular value decomposition (SVD) as input data and outputting the operating state of the prime mover;
a construction step of constructing a judgment prediction means for clustering input data by deriving judgment information about the operating state of the prime mover based on the result of the machine learning;
Obtain judgment target data indicating the operating state of the prime mover, perform the frequency analysis of the obtained judgment target data, perform the singular value decomposition (SVD) on the result of the frequency analysis, and further perform the singular value decomposition (SVD) on the result of the frequency analysis. A state of the prime mover, characterized in that the feature quantity extracted by decomposition (SVD) is input to the judgment prediction means as the input data, and an output step is executed for outputting the judgment information for the judgment target data of the prime mover. judgment program. 12. The program for determining the state of a prime mover according to claim 11, wherein the frequency analysis in the frequency analysis step is an analysis using wavelet transform. The machine learning in the machine learning step includes a support vector machine (SVM) that performs pattern recognition using supervised learning on the feature space extracted by the singular value decomposition (SVD) in the singular value decomposition step. 13. The program for determining the state of a prime mover according to claim 11 or 12, wherein the program performs learning using the following methods. In the singular value decomposition step, the feature space extracted by the singular value decomposition (SVD) is applied to the support vector machine (SVM) using a fuzzy C-Means method. 14. The program for determining the state of a prime mover according to claim 13, wherein the program determines the degree of belonging to a cluster. The machine learning in the machine learning step is characterized in that learning using self-organizing mapping (SOM), which is unsupervised learning, is performed on the feature space extracted by the singular value decomposition (SVD). The program for determining the state of a prime mover according to claim 11 or 12. The prime mover according to any one of claims 11 to 15, further comprising an operation control step of controlling the operation of the prime mover based on the judgment information of the operating state in the output step. status determination program. The prime mover status determination program according to any one of claims 11 to 16, wherein the prime mover is an internal combustion engine and/or a movable part attached to the internal combustion engine. 18. The program for determining the state of a prime mover according to claim 17, wherein in the output step, determination information on a knocking state of the internal combustion engine is output as the determination information on the operating state. 19. The engine state determination program according to claim 18, wherein in the output step, a knocking index is derived and outputted as the determination information of the knocking state. Referring to claim 16, the operation control step controls the operation of the internal combustion engine and/or a movable part attached to the internal combustion engine based on the determination information of the knocking state in the output step. The program for determining the state of a prime mover according to claim 18 or 19 , which further cites claim 17 .