JP6909670B2 - Anomaly detection method and anomaly detection system - Google Patents

Description

The present invention relates to an abnormality detection method and an abnormality detection system for detecting an abnormality in an air conditioner.

The domestic business of air conditioners is saturated, and there is an urgent need to commercialize a service that not only provides equipment to customers, but also collectively undertakes the operation and maintenance of equipment. For commercial air conditioners and refrigerating and refrigerating equipment, the Freon Emission Control Law, which came into effect in April 2015, requires users to perform simple inspections of refrigerant leaks. Abnormality detection for detecting an air conditioner failure or a sign thereof is extremely important for improving production efficiency and reducing inspection work and maintenance costs. Therefore, it is determined whether the target air conditioner is normal or abnormal based on the data associated with the plurality of IDs (Identifiers) output by the air conditioner.

In addition, in abnormality detection for air conditioners, normal sensor data for the past year is learned and evaluated for one air conditioner in order to suppress the decrease in abnormality detection sensitivity due to seasons and machine differences. .. For example, Patent Document 1 discloses a method of constructing learning data by grouping sensors according to the similarity of sensor signals. Further, a method of allocating to sensor groups having a high degree of similarity based on the sensor signals of the air conditioner to be evaluated and using the learning data is disclosed. Further, Patent Document 2 discloses a method of constructing learning data by grouping sensor signals by the k-means method. Further, a method of allocating to the nearest cluster based on the sensor signal of the air conditioner to be evaluated and using the learning data is disclosed.

JP-A-2010-191556 JP-A-2015-88078

As a method for grouping normal sensor data (learning data) based on the sensor signal described in Patent Document 1 and classifying the sensor data to be evaluated, normal learning sensor data having a high degree of similarity to the evaluation sensor may be used. Therefore, it is possible to detect an abnormality of the sensor with high sensitivity.

As a method for grouping normal sensor data (learning data) based on the sensor signal described in Patent Document 2 and classifying the sensor data to be evaluated, the learning sensor data of the cluster more suitable for the evaluation sensor by the k-mean method is used. Since it can be used, it is possible to detect an abnormality of the sensor at high speed and with high sensitivity.

However, since learning data is constructed by collecting sensor signals with high similarity of various air conditioners, it is not possible to consider changes in related sensors that are affected by each other in the air conditioner.

An object of the present invention is to provide an abnormality detection system and an abnormality detection method capable of detecting an abnormality in an air conditioner in consideration of the influence between sensors.

The abnormality detection system according to the present invention is the first air conditioner based on the similarity between the first sensor signals output from one or a plurality of sensors provided in each of the plurality of first air conditioners. Is constructed, and based on the similarity between the second sensor signals obtained from the second sensor signals output from one or more sensors provided in the second air conditioner. The second air conditioner is classified into a similar group in the training data, and an abnormality of the second air conditioner is determined based on the classified training data of the group. It is configured as an abnormality detection method.

The present invention is also understood as an abnormality detection system using the above abnormality detection method.

According to the present invention, it is possible to detect an abnormality in an air conditioner in consideration of the influence between sensors.

This is a configuration example of an abnormality detection system. This is the overall picture of the anomaly detection system. It is a flow chart which groups the normal air conditioners of a large number of sites. It is a flow chart which selects the learning group which has a high degree of similarity to the air conditioner to be evaluated. It is a creation flow chart of a two-dimensional distribution density graph in two types of sensors (learning data grouping). This is an example of a two-dimensional distribution density graph. It is a flow chart of the similarity calculation and the similarity determination between the two-dimensional distribution density graphs in the same sensor of each air conditioner. It is a figure explaining the similarity calculation method of the 2D distribution density graph in the same sensor of each air conditioner. This is an example of a similarity table of a two-dimensional distribution density graph for the same sensor of each air conditioner (classification into training data groups). It is a creation flow chart of a two-dimensional distribution density graph for two types of sensors during the evaluation period of the air conditioner to be evaluated (classification into a learning data group). It is a flow chart of the similarity calculation and the similarity determination between the two-dimensional distribution density graphs of the evaluation air conditioner and the same sensor of each learning data group. It is a figure which shows the similarity table between the two-dimensional distribution density graph in the evaluation air conditioner and the same sensor of each learning group. This is an example of a screen that displays the classification results of the evaluation air conditioner into the learning data group. It is a creation flow chart of the frequency distribution graph of the maximum value, the minimum value, and the average value every few hours in each sensor (learning data grouping). It is an example of the frequency distribution graph of the maximum value, the minimum value, and the average value every few hours in each sensor. It is a flow chart of the similarity calculation and the similarity determination between the frequency distribution graphs of the maximum value, the minimum value, and the average value in the same sensor of each air conditioner. It is a figure explaining the similarity calculation method of the frequency distribution graph of the maximum value, the minimum value, and the average value in the same sensor of each air conditioner. This is an example of the similarity table of the frequency distribution graph of the maximum value, minimum value, and average value in the same sensor of each air conditioner (classification into training data group). It is a creation flow chart of the frequency distribution graph of the maximum value, the minimum value, and the average value in the same sensor in the evaluation period of the air conditioner to be evaluated (classification into a learning data group). It is a flow chart of the similarity calculation and the similarity judgment between the frequency distribution graphs of the evaluation air conditioner and the same sensor of each learning data group.

FIG. 1A is a configuration example of the abnormality detection system 1000 in this embodiment. As shown in FIG. 1A, the abnormality detection system 1000 includes a sensor signal storage unit 103, a learning air conditioner grouping unit 104, an evaluation air conditioner classification unit 113, sensor signal input units 121 and 131, and a feature vector extraction unit. It includes 122 and 132, anomaly measurement calculation units 123 and 133, a threshold value calculation unit 124, an abnormality determination unit 134, and a related sensor identification unit 135. FIG. 1A shows an example in which two sensor signal input units, a feature vector extraction unit, and an anomaly measure calculation unit are provided, but one sensor signal input unit and a feature vector having the same functions, respectively. It may be configured as an extraction unit and an anomaly measure calculation unit. The anomaly detection system 1000 is composed of, for example, a general server device as hardware, and includes an external storage device such as a CPU (Central Processing Unit), a memory, and an HDD (Hard Disk Drive), a CD (Compact Disk), or a DVD. A reading device that reads and writes information to a storage medium that can be read by a portable computer such as (Digital Versatile Disk), an input device such as a keyboard and mouse, an output device such as a display, and a NIC for connecting to a communication network. It has a communication device such as (Network Interface Card) and an internal communication line such as a system bus that connects them. The processing executed by each of the above-mentioned parts of this system can be realized by loading a predetermined program stored in the external storage device into the memory and executing it in the CPU.

The predetermined program is stored (downloaded) in an external storage device from a storage medium via a reading device or from a network via a communication device, loaded on the memory, and executed by the CPU. May be good. Further, it may be loaded directly onto the memory from the storage medium via the reading device or from the network via the communication device and executed by the CPU. Further, the predetermined program stored in the external storage device may be recorded on the recording medium in a file in a format that can be installed in the information processing device or in an executable format, and provided or distributed. Further, the above-mentioned predetermined program may be provided or distributed by storing it on a computer connected to a communication network and downloading it via the network.

For the operation of the abnormality detection system 1000, "learning data grouping" 100 based on the sensor signal (first sensor signal) of the learning air conditioner (first air conditioner) which is a normal air conditioner at many sites. "Classification into learning data group" 110 based on the sensor signal (second sensor signal) of the evaluation air conditioner (second air conditioner) to be evaluated, and the sensor of the classified learning data group. There are four phases: "learning" 120 based on the signal and "evaluation" 130 that detects an abnormality of the air conditioner based on the sensor signal of the evaluation air conditioner. The air conditioner is provided with various sensors such as a temperature sensor, a current sensor, and a pressure sensor in order to control the air conditioning, and processing of each phase is performed based on the sensor signals output from these sensors. Is executed. In the following, the description is made on the premise that one sensor signal is output from the sensor, but the same can be considered when a plurality of sensor signals are output from one sensor.

In the learning data grouping 100, the sensor signal storage unit 103 receives the sensor signals 102 output from the sensors mounted on the learning air conditioners 101 at many sites at periodic predetermined time intervals and stores them in the storage device. .. The learning air conditioner grouping unit 104 groups the accumulated sensor signals 102.

In the classification 110 into the learning data group, the sensor signal storage unit 103 is output from the sensor signal 112 output from the sensor mounted on the evaluation air conditioner 111 and the sensor mounted on the learning air conditioner 101 at many sites. The sensor signal 102 is received at periodic predetermined time intervals and stored in the storage device. The learning air conditioner grouping unit 104 groups these accumulated sensor signals, and the evaluation air conditioner classification unit 113 classifies the evaluation air conditioner 111 into a learning data group.

In the learning 120, when the evaluation air conditioner classification unit 113 classifies the learning data group based on the sensor signal of the evaluation air conditioner 111, the sensor signal input unit 121 inputs the sensor signal classified by the evaluation air conditioner classification unit 113. Then, the feature vector extraction unit 122 extracts the feature vector based on the input sensor signal, and the anomaly measurement calculation unit 123 uses the feature vector of the learning period specified in advance every predetermined time (hereinafter, each). The anomaly measure is calculated for each feature vector (sometimes expressed as time). The threshold value calculation unit 124 calculates a threshold value according to the abnormality measure of the calculated learning data.

In the evaluation 130, when the sensor signal input unit 131 inputs the sensor signal 112 output from the sensor mounted on the evaluation air conditioner 111, the feature vector extraction unit 132 receives the feature vector based on the input sensor signal 112. Is extracted, and the abnormality measurement calculation unit 133 calculates the abnormality measurement for each feature vector at predetermined time intervals using the feature vector of the learning period specified in advance. The abnormality determination unit 134 determines an abnormality by comparing the calculated abnormality measure with the threshold value 124 obtained from the abnormality measure of the learning data, and the related sensor identification unit 135 identifies the related sensor of the abnormality.

Specifically, as will be described later, the learning air conditioner grouping unit 104 groups the learning air conditioners 101 at a large number of sites by the similarity of the sensor signal 102 based on the sensor signal 102 by the sensor signal storage unit 103, and learns. Build a data group group. The evaluation air conditioner classification unit 113 sets the evaluation air conditioner 111 and the learning data group group constructed by the learning air conditioner grouping unit 104 based on the sensor signal 112 of the evaluation air conditioner 111 received by the sensor signal storage unit 103. By comparison, the evaluation air conditioner 111 is classified into a learning data group composed of the learning air conditioners 101 having a high degree of similarity in the learning data group group.

FIG. 1B is a diagram showing the overall configuration of the abnormality detection system 1000 shown in FIG. 1A. The sensor signals 102, 102', 102'' output from the sensors mounted on the learning air conditioners 101, 101', 101'' at many sites are data centers provided with the server device constituting the abnormality detection system 1000. It is sent to 140 via radio or recording medium. Further, the sensor signal storage unit 103 accumulates these signals, inputs the signals to the learning air conditioner grouping unit 104, and groups the learning data.

Further, the sensor signal 112 output from the sensor mounted on the evaluation air conditioner 111 is sent to the data center 140 wirelessly or via a recording medium. Further, the sensor signal storage unit 103 accumulates these signals, inputs them to the evaluation air conditioner classification unit 113, and learn data having a high degree of similarity from a large number of learning data groups constructed by the learning air conditioner grouping unit 104. Select a group. The sensor signal input unit 121 receives the learning data of the learning data group, the feature vector extraction unit 122 extracts the feature vector, and the anomaly measurement calculation unit 123 uses the feature vector of the learning period specified in advance every predetermined time. The anomaly measure is calculated for each feature vector of. Through this process, the threshold value calculation unit 124 calculates the threshold value according to the abnormality measure of the learning data.

Further, the sensor signal 112 of the evaluation air conditioner 111 transmitted to the data center 140 is input to the sensor signal input unit 131, the feature vector extraction unit 132 extracts the feature vector, and the abnormality measure calculation unit 133 is designated in advance. The anomaly measure is calculated for each feature vector at predetermined time intervals using the feature vector of the learning period. Through the process, the abnormality determination unit 134 determines the abnormality, and the related sensor identification unit 135 identifies the sensor related to the abnormality by comparing with the threshold value 124 calculated above.

In the feature vector extraction units 122 and 132, the sensor signals are normalized so that the average is 0 and the variance is 1. Then, in order to extract short-term changes, a window of ± w (w = 0, 1, 2, ...) is set for a certain time t (t = 0, 1, 2, ...), and the following equation is used. , Extract one vector x (t) consisting of the elements of window width (2w + 1) x number of sensors (M) per time.

Here, x _m (t) (m = 1, 2, ..., M) is the m-th sensor signal at time t after normalization.

Further, the anomaly measure calculation units 123 and 133 employ a local subspace classifier (LSC). LSC creates a k-1 dimensional subspace using k neighborhood data xi (i = 1, ..., K) for unknown data. Then, it is a method of determining whether or not it is abnormal based on the unknown data and the projection distance of the k-1 dimensional subspace. Since the anomaly measure is represented by the projection distance between the unknown data and the k-1 dimensional subspace, it is sufficient to find the point on the subspace closest to the unknown data. Therefore, in order to calculate b from unknown data and k neighborhood data xi (i = 1, ..., k), a matrix Q and x _i (i = 1, ..., k) in which k unknown data are arranged are used. The correlation matrix C is obtained from the arranged matrix X by the following equation.

Then, b is calculated by the following equation.

Here, 1 _k is a k-dimensional vector whose elements are all 1. _{This equation means that the coefficient vector of the linear combination of x i} (i = 1, ..., K) that minimizes the square error with unknown data is b. The anomaly measurement by the anomaly measure calculation units 123 and 133 is performed by calculating the square root of the sum of squares of the difference between the unknown data and the points on the subspace closest to the unknown data. Further, in order to appropriately determine the threshold value for determining whether or not the abnormality is abnormal in the abnormality measurement, the threshold value calculation unit 124 is automatically set by cross-validation of learning data.

First, the anomaly measure calculation unit 123, 133 divides the learning data into N groups equally, and calculates the anomaly measure of each data by LSC using the learning data excluding the group for each group. Next, the threshold value calculation unit 124 sorts the anomaly measures calculated by the anomaly measure calculation units 123 and 133 in ascending / descending order, and the value of the anomaly measure reaches a parameter value of 0 to 1 in which the ratio to the total number of data reaches a predetermined parameter value. Is the threshold value. If we assume that the training data does not contain any anomalies, set the parameter value to 1. In this case, the threshold is the maximum value of the anomaly measure. Finally, the abnormality determination unit 134 determines whether or not the abnormality measure calculated by the abnormality measure calculation unit 123 and 133 exceeds the threshold value set by the threshold value calculation unit 124, and exceeds the threshold value. If it is determined that it is abnormal, it is determined that it is abnormal, and otherwise it is determined that it is normal.

The related sensor identification unit 135 obtains an abnormal section in which an abnormality is continuously detected, and extracts a related sensor for each section based on the two-dimensional distribution density. For that purpose, at the time of learning, the learning air conditioner grouping unit 104 calculates the two-dimensional feature distribution density of normal data by rounding out two types of sensor signals and saves it in an image format. This image is called a distribution density image. At the time of abnormality detection, the evaluation air conditioner classification unit 113 calculates the degree of isolation of the sensor according to where the detected abnormality data is plotted in the distribution density image, and extracts the sensor as an abnormality-related sensor in descending order of the degree of isolation. This means finding a sensor whose distribution of anomalous data deviates from the distribution of normal data.

Next, a method for grouping learning air conditioners and classifying evaluation air conditioners will be described with reference to FIGS. 2 and 3.

FIG. 2 is a flow diagram for grouping learning air conditioners, which are normal air conditioners at many sites. First, the sensor signal storage unit 103 stores the sensor signals 102 received from the sensors mounted on the learning air conditioners 101 at many sites in the storage device, and the learning air conditioner grouping unit 104 stores the stored sensor signals 102. Group (S201). Specifically, the learning air conditioner grouping unit 104 uses an arbitrary learning air conditioner 101 as a reference, and each learning data group (S202), each air conditioner (S203), each sensor (S204), and each learning period division (S204). The following processing is repeated for S205).

Here, the learning period division in step S205 can be a division in which one year is divided into a plurality of parts. For example, one year is divided into 12 and every month. The learning air conditioner grouping unit 104 extracts arbitrary two types of sensor data from all the sensor data of each learning air conditioner 101, and for all combinations of the extracted sensor data, two dimensions in the two types of sensors. A frequency distribution such as a distribution density graph or a maximum value, a minimum value, and an average value every few hours is created (S206). Further, the learning air conditioner grouping unit 104 calculates the similarity between the two-dimensional distribution density graph or the frequency distribution graph in the same two types of sensors (for example, a temperature sensor and a pressure sensor) for each learning air conditioner 101 (for example, a temperature sensor and a pressure sensor). S207), each learning air conditioner 101 is grouped based on the similarity (S208), and a learning data group of the learning air conditioner 101 is constructed (S209). The specific processing of S206 will be described later with reference to FIG. 4, and the specific processing of S207 and S208 will be described later with reference to FIG.

FIG. 3 is a flow chart for selecting a learning group having a high degree of similarity to the evaluation air conditioner 111.
First, the sensor signal storage unit 103 stores the sensor signal 112 received from the evaluation air conditioner 111 in the storage device, and the evaluation air conditioner classification unit 113 reads the sensor signal 112 stored in the storage device (S301). ). The evaluation air conditioner classification unit 113 sets an evaluation period (S302), and reads the sensor signal 102 for each learning data group in that period received from the learning air conditioner 101. The evaluation air conditioner classification unit 113 repeats the following processing for each sensor (S303) and each learning data group (S304). The evaluation air conditioner classification unit 113 refers to the sensor signal 112 of the evaluation air conditioner 111 and the sensor signal 102 for each learning data group in a two-dimensional distribution density graph for two types of sensors or every few hours, as in the case of S206. A frequency distribution such as the maximum value, the minimum value, and the average value of is created (S305). Further, the evaluation air conditioner classification unit 113 is a two-dimensional distribution density graph of the same two types of sensors (for example, a temperature sensor and a pressure sensor) between the evaluation air conditioner 111 and each learning air conditioner 101 of each learning data group. Alternatively, the similarity between the frequency distribution graphs is calculated (S306). Finally, the evaluation air conditioner classification unit 113 ranks the evaluation air conditioners 111 in the order of the learning data groups having the highest degree of similarity (S307), and assigns the evaluation air conditioner 111 to the learning data group having the highest similarity ranking. Classify (S308). Repeat S302 to S308 for each evaluation air conditioner. The specific processing of S305 will be described later with reference to FIG. 9, and the specific processing of S306 will be described later with reference to FIG.

As described above, the grouping and evaluation of the learning air conditioner 101 includes the two-dimensional distribution density of the two types of sensors, or the maximum value, the minimum value, and the average value of each sensor every few hours. Use the similarity of the frequency distribution. Grouping of these two types of learning air conditioners and classification of evaluation air conditioners are described below as Example 1 and Example 2.

In the first embodiment, a method of grouping the learning air conditioner 101 and classifying the evaluation air conditioner 111 using the two-dimensional distribution density graphs of the two types of sensors will be described.

As shown in FIG. 2, in order to group the learning air conditioners 101 of a large number of sites, step S206 for creating a two-dimensional distribution density graph for two types of sensors, for each air conditioner, 2 for the same two types of sensors. There is a step S207 for calculating the similarity between the two-dimensional distribution density graphs, and a step S208 for grouping each learning air conditioner 191 based on the similarity. Specific processing will be described below.

FIG. 4 is a flow chart for creating a two-dimensional distribution density graph for two types of sensors (learning data grouping).

The sensor signal storage unit 103 receives the sensor signals 102 mounted on the learning air conditioners 101 at many sites, stores them in the storage device, and inputs the stored sensor signals 102 to the learning air conditioner grouping unit 104 for grouping. (S401). The learning air conditioner grouping unit 104 executes the processes of steps S403 to S406 for the sensor signal 102 accumulated for each sensor (S402). The learning air conditioner grouping unit 104 performs a round-robin processing of two sensors in all learning air conditioners to obtain a maximum value (MAX) for each type of sensor signal so that the vertical and horizontal scales of the two-dimensional distribution density graph are the same. And the minimum value (MIN) are obtained (S403). Specifically, the detection range of the same type of sensor may differ depending on the environment in which the learning air conditioner 101 is installed. For example, in the case of a temperature sensor, one learning air conditioner 101 is set to −10 ° C. to + 40 ° C., while another learning air conditioner 101 is set to −5 ° C. to + 45 ° C. Therefore, the learning air conditioner grouping unit 104 first obtains the maximum value and the minimum value, and executes the following processing.

Subsequently, the learning air conditioner grouping unit 104 calculates the step size S when dividing the maximum value from the minimum value by a designated number N (S404). The step size S can be calculated by, for example, S = (MAX-MIN) / N. Further, the learning air conditioner grouping unit 104 calculates the processing range for calculating the distribution density by expanding the range outward from the minimum value and the maximum value in order to provide a spare processing range (S405). The range to be expanded is, for example, MIN is changed to MIN-S × M and MAX is changed to MAX + S × M. Here, M is a predetermined integer of 1 or more. Subsequently, the learning air conditioner grouping unit 104 calculates the bin number (BNO) from the feature value (F) for all the data in the learning period by the following equation (S406).

BNO = INT ((F-MIN) / (MAX-MIN))
However, INT (X) represents the integer part of X.

The learning air conditioner grouping unit 104 executes the processes from steps S407 to S411 in order to create a two-dimensional distribution density graph by round-robin processing of two sensors for each learning period and each air conditioner. (S407).

The learning air conditioner grouping unit 104 repeats the following processing for each learning period division (S407), each air conditioner (S408), and each sensor (S409). Here, the learning period division in step S407 may be set in the same manner as in S205. Brute force includes the case where two sensors are of the same type and the same sensor. The learning air conditioner grouping unit 104 secures a two-dimensional array for calculating the distribution density, and sets 0 to all the elements. The size of the array is N + 2M. The learning air conditioner grouping unit 104 adds 1 to the elements of the array corresponding to the bin numbers of the two sensor values for all the data in the learning period. By this process, a two-dimensional frequency distribution (histogram) by two sensors is calculated. The learning air conditioner grouping unit 104 converts this frequency distribution into an image, and stores the converted image in association with an identification number (ID) for identifying the image (S410). The conversion method will be described later. Although not shown, the learning air conditioner grouping unit 104 records the size of the two-dimensional array and the minimum and maximum values of each sensor signal calculated in step S403.

An example of the image conversion method in step S410 will be described. The learning air conditioner grouping unit 104 obtains the maximum value of the array elements, that is, the maximum frequency. The image size is the same as the array size, and the pixel value of the corresponding coordinate from the value of each element is set to, for example, 255 × the element value of the array / maximum frequency. 255 is the maximum value when the pixel value is represented by 8 bits, and if this value is used, it can be saved as it is in the bitmap format.

The image obtained by the above processing is referred to as a distributed density image because the high density part in the two-dimensional feature space is represented by a high pixel value.

FIG. 5 is an example of an imaged two-dimensional distribution density graph. The pixel value of 0 is represented by white, the maximum is represented by black, and the space between them is represented by shades of gray. However, the method of creating an image is not limited to the above method. For example, instead of a simple frequency distribution, a Gaussian distribution or another weighted filter may be assigned to one data and superposed on the data. Alternatively, the image obtained by the above method may be filtered with a maximum value of a predetermined size, an average filter, or another weighted filter. Further, it is not always necessary to save in the image format, and the two-dimensional array may be saved in the text format. Further, the pixel value is not a shade of gray, and may be saved in a text format of a binarized two-dimensional array in which the pixel with distribution is 1 and the pixel without distribution is 0.

FIG. 5 shows, as an example, distribution density images of a sensor A (temperature sensor) and a sensor B (pressure sensor) provided in a learning air conditioner 101. With the minimum to maximum value of the scaled sensor A on the horizontal axis and the minimum to maximum value of the scaled sensor B on the horizontal axis, the pressure p1 at the temperature t1 is plotted as (X1, Y1), and the temperature t2. The pressure p2 in is plotted as (X2, Y2), and it can be seen that the distribution density image is associated with the identification information S0001A0001 (S0001: site number, A0001: air conditioner number). In this way, the learning air conditioner grouping unit 104 generates such an imaged two-dimensional distribution density graph for the combination of two sensors selected from all the sensors provided in all the learning air conditioners 101. And associate it with the identification number. In step S411, the learning air conditioner grouping unit 104 stores the two-dimensional distribution density graph calculated in the above process in the storage device and outputs the two-dimensional distribution density graph to the evaluation air conditioner classification unit 113. As described above, the two-dimensional distribution density graph can be said to be a graph showing the correlation between the output values of the two types of sensors.

FIG. 6 is a flow chart of similarity calculation and similarity determination between two-dimensional distribution density graphs in the same two types of sensors for each learning air conditioner 101. The two-dimensional distribution densities of all the learning air conditioners 101 created in FIG. 4 are input to the learning air conditioner grouping unit 104 (S601). First, the learning air conditioner grouping unit 104 uses an arbitrary learning air conditioner 101 as a reference, and each learning data group (S602), each sensor (S603), each learning period division (S604), and each learning air conditioner (S605). The following process is repeated. Here, the learning period division in step S604 may be set in the same manner as in S205.

The learning air conditioner grouping unit 104 has a similarity between the two-dimensional distribution density graphs of the same two types of sensors between the reference air conditioner selected from any learning air conditioner 101 and the remaining learning air conditioner 101. Is calculated (S606), and the degree of similarity thereof is determined (S607). That is, the determination of the degree of similarity means that the degree of similarity is high and it is determined whether or not the learning air conditioner 101 can be classified into the same group as the reference air conditioner. When the similarity between the two-dimensional distribution density graphs of the same two types of sensors between the reference air conditioner and the remaining learning air conditioner 101 is, for example, 80% or more, the learning air conditioner grouping unit 104 compares. It is determined that the learned air conditioner 101 belongs to the same group as the reference air conditioner. The similarity level is not fixed but adjustable. However, if the similarity level is too low, there is a high possibility that the anomaly detection accuracy will drop. The learning air conditioner grouping unit 104 groups the reference air conditioner and the learning air conditioner 101 having a high degree of similarity based on the similarity determination (S608).

The learning air conditioner grouping unit 104 sets the reference air conditioner again from the remaining learning air conditioners 101 excluding the reference air conditioner and the grouped learning air conditioner 101, and repeats steps S601 to S608. , A learning data group of the learning air conditioner 101 is constructed.

FIG. 7 is a diagram illustrating a method of calculating the similarity of the two-dimensional distribution density graph for each learning air conditioner. The learning air conditioner grouping unit 104 selects the reference air conditioner 701 from all the learning air conditioners 101 based on the two-dimensional distribution density graph for all the learning air conditioners 101 created in step S606 (S1). The similarity is calculated by comparing the two-dimensional distribution density graphs of the same two types of sensors with each of the learning air conditioners 101 selected from the remaining learning air conditioners 702 (S2). For example, the learning air conditioner grouping unit 104 reads a reference air conditioner superimposition distribution density image in which the distribution density image of the reference air conditioner 701 and the distribution density image of each learning air conditioner 101 are superimposed, and the superimposing portion of the distribution density image is read. The similarity is calculated based on the size of the area. This process is repeated for each sensor and each learning period division. For example, a two-dimensional distribution density graph for the same two types of sensors is represented by a binarized two-dimensional array in which pixels with a distribution are 1 and pixels without a distribution are 0. And the similarity is shown below.

Similarity (%) = 2D distribution density graph of air conditioner compared with standard air conditioner is "1"
Number of pixels showing / Number of pixels whose 2D distribution density graph of the reference air conditioner shows "1" FIG. 8 is an example of a similarity table of the 2D distribution density graphs of the same two types of sensors for each learning air conditioner. be. The similarity table is output by the learning air conditioner grouping unit 104 in step S607.

The learning air conditioner grouping unit 104 calculates the similarity of the remaining learning air conditioners 802 to the reference air conditioner 801, creates a similarity table, and determines the similarity. That is, the determination of the degree of similarity means that the degree of similarity is high and it is determined whether or not it can be classified into the same group. If the similarity between the two-dimensional distribution density graphs of the reference air conditioner 801 and the remaining air conditioner 802 in the same sensor is, for example, 80% or more, which is a numerical value set as a threshold value, it is judged to be in the same group as the reference air conditioner 801. .. In FIG. 8, for example, a two-dimensional distribution density graph of the sensor A and the sensor B provided in the air conditioner 1 which is a reference air conditioner, and the sensor A and the sensor B provided in the air conditioner 2 which is a learning air conditioner. The similarity with the two-dimensional distribution density graph of is 85% in January and 90% in February. In this way, the learning air conditioner grouping unit 104 calculates the similarity between the reference air conditioner and the learning air conditioner for all combinations of sensors, and outputs the result as the similarity table. The above threshold is not fixed and can be adjusted to any value. However, if the value of the above threshold value is too low, there is a high possibility that the abnormality detection accuracy will drop. The learning air conditioner grouping unit 104 groups the learning air conditioner 802 having a high degree of similarity with the reference air conditioner 801 based on the similarity determination.

The learning air conditioner grouping unit 104 sets the reference air conditioner 803 again from the remaining air conditioners excluding the reference air conditioner 801 and the grouped learning air conditioners, and repeats steps S601 to S608 to learn. A learning data group for the air conditioner 101 is constructed.

As shown in FIG. 3, in order to select a learning group having a high degree of similarity to the evaluation air conditioner 111, step S305 for creating a two-dimensional distribution density graph for two types of sensors during the evaluation period of the evaluation air conditioner 111. Step S306 for calculating the similarity between the two-dimensional distribution density graphs of the same two types of sensors of the evaluation air conditioner 111 and the learning air conditioner 101 belonging to each learning data group, and ranking based on the similarity (S307). , There is step S308 to classify into the training data group having the highest similarity rank. Specific processing will be described below.

FIG. 9 is a flow chart for creating a two-dimensional distribution density graph for two types of sensors during the evaluation period of the evaluation air conditioner 111 (classification into a learning data group).

The sensor signal storage unit 103 inputs the sensor signal 112 mounted on the evaluation air conditioner 111 to the evaluation air conditioner classification unit 113. Further, the sensor signal storage unit 103 also inputs the sensor signals 102 of the learning air conditioner 101 classified for each learning data group during the evaluation period into the evaluation air conditioner classification unit 113 (S901).

The evaluation air conditioner classification unit 113 executes the processes of steps 903 to S906 for each sensor (S902). Similar to the case of step S403, the evaluation air conditioner classification unit 113 performs each sensor signal type so that the vertical and horizontal scales of the two-dimensional distribution density graph become the same by the round-robin processing of two sensors in all the learning air conditioners. The maximum value (MAX) and the minimum value (MIN) of are obtained (S903). As in the case of step S404, the evaluation air conditioner classification unit 113 calculates the step size S when dividing the maximum value from the minimum value by the specified number N (S904), and moves outward from the minimum value and the maximum value. The range is expanded to calculate the processing range for calculating the distribution density (S905). The evaluation air conditioner classification unit 113 calculates the bin number (BNO) from the feature value (F) for all the data in the evaluation period, as in the case of step S406 (S906).

The evaluation air conditioner classification unit 113 performs the processes from steps S907 to S911 for the evaluation period, the evaluation air conditioner 111, and each learning data group in order to create a two-dimensional distribution density graph by round-robin processing of two sensors. Execute. (S907).

Similar to the case of steps S407 to S409, the evaluation air conditioner classification unit 113 has the following for the evaluation period (S907), each sensor (S908), and all the learning air conditioners 101 (S909) belonging to each learning data group. Repeat the process. The brute force of two sensors includes the case where the two sensors are the same. The evaluation air conditioner classification unit 113 secures a two-dimensional array for calculating the two-dimensional distribution density, and sets 0 to all the elements. The size of the array is N + 2M. The evaluation air conditioner classification unit 113 adds 1 to the elements of the array corresponding to the bin numbers of the two sensor values for all the data in the learning period. By this process, a two-dimensional frequency distribution (histogram) by two sensors is calculated. The evaluation air conditioner classification unit 113 converts this frequency distribution into an image, and stores the converted image in association with an identification number (ID) for identifying the image (S910). The conversion method will be described later. Although not shown, the evaluation air conditioner classification unit 113 records the size of the two-dimensional array and the minimum and maximum values of each sensor signal calculated in step S903. Since the image conversion method in step S910 is the same as that in step S410, the description thereof will be omitted here.

FIG. 10 is a flow chart of similarity calculation and similarity determination between the two-dimensional distribution density graphs of the evaluation air conditioner and the learning air conditioner belonging to each learning data group in the same sensor. First, the learning air conditioner grouping unit 104 inputs the total two-dimensional distribution density of the learning air conditioner 101 and the two-dimensional distribution density of the evaluation air conditioner 111 created in FIG. 4 to the evaluation air conditioner classification unit 113 (S1001). The evaluation air conditioner classification unit 113 repeats the following processing for each evaluation air conditioner (S1002), each sensor (S1003), and each learning group (S1004). As shown in FIG. 11, the evaluation air conditioner classification unit 113 calculates the similarity between the two-dimensional distribution density graphs of the evaluation air conditioner 111 and all the learning air conditioners 101 belonging to each learning group in the same sensor (S1005). .. For example, the evaluation air conditioner classification unit 113 evaluates the learning air conditioner superimposed distribution density image in which the distribution density images of the combination of the two types of sensors provided in all the learning air conditioners 101 belonging to each learning group are superimposed. The superposed distribution density image superposed on the distributed density image of the air conditioner 111 is read, and the similarity is calculated from the size of the area of the superposed portion. In FIG. 11, for example, the distribution density image in the combination of the sensor A and the sensor B of the evaluation air conditioner and the learning in the combination of the sensor A and the sensor B of each learning air conditioner 101 classified in the learning data group 1 It shows that the similarity with the air conditioner superposition distribution is 89%.

Finally, the evaluation air conditioner classification unit 113 ranks the evaluation air conditioners 111 in the order of the learning data groups having the highest degree of similarity (S1006), and classifies them into the learning data groups having the highest degree of similarity (S1007). For example, in FIG. 11, when the evaluation air conditioner classification unit 113 calculates the similarity between the evaluation air conditioner 111 and each learning data group, the calculated similarity is compared, and the learning data group 1 is the most evaluated air conditioner. Since it is similar to the machine 111, the order of similarity is set to 1. In the above comparison, when the average value of the degree of similarity with the learning air conditioner superposition distribution in each combination of sensors is the highest value, the order may be set to 1. Then, the evaluation air conditioner classification unit 113 learns the sensor data of the classified learning group and detects an abnormality of the evaluation air conditioner.

FIG. 12 is an example of a screen for displaying the classification result of the evaluation air conditioner into the learning data group. The screen is output to, for example, a display connected to the server device constituting the abnormality detection system 1000. The input device of the server device receives from the user the designation of the evaluation air conditioner to be displayed and the evaluation period from the evaluation air conditioner classification screen 1201. For example, the input device receives the selection of the identification number (ID) of the evaluation air conditioner 111 from the user in the evaluation air conditioner ID selection window 1202. Further, the input device receives input of the start date and the end date of the evaluation period of the selected evaluation air conditioner 111 from the evaluation period window 1203. When the input device accepts the press of the start button 1205, the classification flow of the evaluation air conditioners shown in FIGS. 9 and 10 into the learning data group is started. After the classification is completed, the learning data group having the highest degree of similarity to the evaluation air conditioner 101 and the degree of similarity 1206 are displayed in the classification result column 1204. Further, in the column below it, the air conditioner list 1207 of each site included in the learning data group is displayed.

Further, in the classification result column 1204, the similarity 1208 of each sensor of the learning data group classified as the evaluation air conditioner is displayed. These displays are made by, for example, the evaluation air conditioner classification unit 113.

According to the first embodiment, the air conditioners are grouped according to the similarity of the sensor signals of the learning air conditioner 101, and based on the sensor signals of the evaluation air conditioner 111, the air conditioners having a high degree of similarity are used by the correlation between the sensor signals. By learning the machine group, highly sensitive abnormality detection becomes possible by accumulating data in a short period of time. In addition, since existing air conditioners at many sites can be used, a large amount of diverse learning data can be secured, so that it is possible to increase the sensitivity of abnormality detection and suppress false reports. Further, in the above embodiment, the similarity between the air conditioners is determined by using the correlation between the two types of sensors, but the same can be considered when the correlation between the three or more types of sensors is used.

Next, a method of grouping learning air conditioners and classifying evaluation air conditioners using the frequency distribution of the maximum value, the minimum value, and the average value every few hours in each sensor will be described.

As shown in FIG. 2, in order to group the learning air conditioners 101 of a large number of sites, step S206 for creating a frequency distribution graph of the maximum value, the minimum value, and the average value every few hours in each sensor, each learning air conditioner. There is a step S207 for calculating the similarity between the frequency distributions of the maximum value, the minimum value, and the average value in the same sensor of the machine 101, and a step S208 for grouping each learning air conditioner 101 based on the similarity. Specific processing will be described below.

FIG. 13 is a flow chart for creating a frequency distribution graph of the maximum value, the minimum value, and the average value every several hours in each sensor (learning data grouping).

The sensor signal storage unit 103 receives the sensor signals 102 mounted on the learning air conditioners 101 at many sites, stores them in the storage device, and inputs the stored sensor signals 102 to the learning air conditioner grouping unit 104 for grouping. (S1301). The learning air conditioner grouping unit 104 executes the processes of steps S1303 to S1306 for the sensor signals 102 accumulated for each sensor (S1302). The learning air conditioner grouping unit 104 obtains the maximum value (MAX) and the minimum value (MIN) for each type of sensor signal so that the scales of the frequency distribution graphs in all the learning air conditioners are the same (S1303). The learning air conditioner grouping unit 104 calculates the step size S when dividing the maximum value from the minimum value by a designated number N (S1304). Since steps S1303 and S1304 are the same as steps S403 and S404 shown in FIG. 4, the description thereof will be omitted here. In order to provide a spare processing range, the learning air conditioner grouping unit 104 calculates the processing range for frequency distribution calculation by expanding the range outward from the minimum value and the maximum value (S1305). The range to be expanded is, for example, MIN is changed to MIN-S × M and MAX is changed to MAX + S × M. Here, M is a predetermined integer of 1 or more.

The learning air conditioner grouping unit 104 creates a frequency distribution graph of the maximum value, the minimum value, and the average value of each sensor every few hours for each learning period and each air conditioner, in order to create a frequency distribution graph from steps S1307 to S1311. (S1307).

The learning air conditioner grouping unit 104 repeats the following processing for each learning period division (S1307), each air conditioner (S1308), and each sensor (S1309). Here, the learning period division in step S1307 may be set in the same manner as in S205. The learning air conditioner grouping unit 104 secures a two-dimensional array for calculating the frequency distribution, and sets 0 to all the elements. The size of the array is N + 2M. The learning air conditioner grouping unit 104 adds 1 to the elements of the array corresponding to one sensor value for all the data in the learning period). By this process, a frequency distribution (histogram) is calculated. The learning air conditioner grouping unit 104 converts this frequency distribution into data or an image, and stores the converted image in association with an identification number (ID) for identifying the image (S1310). Although not shown, the learning air conditioner grouping unit 104 records the size of the two-dimensional array and the minimum and maximum values of each sensor signal calculated in step S1303.

FIG. 14 is an example of a frequency distribution graph of the maximum value, the minimum value, and the average value every several hours in each sensor. The learning air conditioner grouping unit 104 creates the frequency distribution graph using, for example, the hourly maximum value, minimum value, and average value of each sensor for one month.

Specifically, as shown in FIG. 14, the learning air conditioner grouping unit 104 sets the maximum value and the minimum value of the sensor signal 101 for the sensor signal 101 of the sensor A of a certain learning air conditioner 101 every week. Find the average value. In FIG. 14, for example, the maximum value, the minimum value, and the average value of the sensor A in the first week are MAX1, MIN1, and AVE1. The same applies to sensor B.

When the learning air conditioner grouping unit 104 obtains the maximum value, the minimum value, and the average value of the sensor signal 101 for each sensor, the learning air conditioner grouping unit 104 stacks them on the respective frequency distribution graphs. In FIG. 14, for example, when the sensor A is a temperature sensor, the learning air conditioner grouping unit 104 has a temperature on the horizontal axis, which is a processing range for calculating the frequency distribution, and a frequency number (number of occurrences) on the vertical axis. .. In the learning air conditioner grouping unit 104, since the maximum value MAX1 in the first week and the maximum value MAX2 in the second week in the sensor A are the same value t3, the number of times (in this case, twice) is set to the temperature t3 of the frequency distribution. ) Are stacked. On the other hand, since the maximum value MAX3 in the third week is a value t2 lower than the maximum value MAX1 in the first week and the maximum value MAX2 in the second week, the learning air conditioner grouping unit 104 has the temperature t2 of the frequency distribution. The number of times (in this case, once) is accumulated. The average value and the minimum value of the sensor signal 101 can be considered in the same manner. Further, the learning air conditioner grouping unit 104 also executes the same processing for the sensor B. The learning air conditioner grouping unit 104 generates such a frequency distribution graph for all the sensors of the learning air conditioner 101, and generates the above identification number (in FIG. 14, identification information S0001A0001 (S0001: site number, A0001: air conditioner). It is saved as an image in the storage device in association with the number). It may be saved in the text format of a binarized two-dimensional array in which the pixel with distribution is 1 and the pixel without distribution is 0. In this way, the frequency. It can be said that the distribution graph is a graph in which statistical values (maximum value, minimum value, average value) obtained from the output values of the sensor in a predetermined period are aggregated.

FIG. 15 is a flow chart of similarity calculation and similarity determination between frequency distribution graphs of maximum value, minimum value, and average value in the same sensor of each learning air conditioner 101. The frequency distribution graph of the maximum value, the minimum value, and the average value every several hours in each sensor of the learning air conditioner 101 created in FIG. 13 is input to the learning air conditioner grouping unit 104 (S1501). First, the learning air conditioner grouping unit 104 uses an arbitrary learning air conditioner 101 as a reference, and each learning data group (S1502), each sensor (S1503), each learning period division (S1504), and each learning air conditioner (S1505). The following process is repeated. Here, the learning period division in step S1504 may be set in the same manner as in S205.

The learning air conditioner grouping unit 104 is a frequency distribution graph of the maximum value, the minimum value, and the average value in the sensor between the reference air conditioner selected from any learning air conditioner 101 and the remaining learning air conditioner 101. The degree of similarity between the two is calculated (S1506), and the degree of similarity is determined (S1507). That is, the determination of the similarity means that, as in the case of the first embodiment, the similarity is high and it is determined whether the learning air conditioner 101 can be classified into the same group as the reference air conditioner. When the similarity between the frequency distribution graphs of the maximum value, the minimum value, and the average value in the same sensor between the reference air conditioner and each of the remaining learning air conditioners 101 is, for example, 80% or more, the learning air conditioner group. The chemical unit 104 determines that the compared learning air conditioner 101 belongs to the same group as the reference air conditioner. The similarity level can be considered in the same manner as in the case of the first embodiment. The learning air conditioner grouping unit 104 groups the reference air conditioner and the learning air conditioner 101 having a high degree of similarity based on the similarity determination (S1508).

The learning air conditioner grouping unit 104 sets the reference air conditioner again from the remaining learning air conditioners 101 excluding the reference air conditioner and the grouped learning air conditioner 101, and repeats steps S1501 to S1508. , A learning data group of the learning air conditioner 101 is constructed.

FIG. 16 is a diagram illustrating a method of calculating the similarity of the frequency distribution graph of the maximum value, the minimum value, and the average value in the same sensor of each learning air conditioner. The learning air conditioner grouping unit 104 uses the frequency distribution graph of the maximum value, the minimum value, and the average value of each sensor of the all learning air conditioners 101 created in step S1606 as a reference from among all the learning air conditioners 101. The frequency distribution graphs of the maximum value, the minimum value, and the average value of the same sensor are compared with each of the learning air conditioners 101 selected from the remaining learning air conditioners 1602 by selecting the air conditioner 1601 (SS1). By 1603 (SS2), the similarity is calculated. For example, the learning air conditioner grouping unit 104 superimposes the image of the frequency distribution graph of the maximum value of the reference air conditioner 1601 and the image of the frequency distribution graph of the maximum value of each learning air conditioner 101. The frequency distribution image is read, and the similarity is calculated based on the size of the area of the superposed portion. This process is repeated for each sensor and each learning period division. For example, the frequency distribution graph of the maximum value, the minimum value, and the average value in the sensor is represented by a binarized two-dimensional array in which pixels with a distribution are 1 and pixels without a distribution are 0.
And the similarity is shown below.

Similarity (%) = Number of pixels in which the frequency distribution graph of the air conditioner compared with the standard air conditioner shows "1" / Number of pixels in which the frequency distribution graph of the standard air conditioner shows "1" Figure 17 shows each learning air conditioner. This is an example of the similarity table of the frequency distribution graph of the maximum value, the minimum value, and the average value in the same sensor of the machine. The similarity table is output by the learning air conditioner grouping unit 104 in step S607.

The learning air conditioner grouping unit 104 calculates the similarity of the remaining learning air conditioners 1702 with respect to the reference air conditioner 1701, creates a similarity table, and determines the similarity. That is, the determination of the degree of similarity means that the degree of similarity is high and it is determined whether or not it can be classified into the same group. When the similarity between the frequency distribution graphs of the maximum value, the minimum value, and the average value of the reference air conditioner 1701 and the remaining air conditioner 1702 in the same sensor is, for example, 80% or more, which is a numerical value set as a threshold value, the reference air conditioner It is judged to be in the same group as 1701. In FIG. 17, for example, a frequency distribution graph of the maximum value of a certain sensor provided in the air conditioner 1 which is a reference air conditioner and a frequency distribution graph of the maximum value of the sensor provided in the air conditioner 2 which is a learning air conditioner. The similarity with the distribution graph shows that it is 85% in January and 90% in February. In this way, the learning air conditioner grouping unit 104 calculates the similarity between the reference air conditioner and the learning air conditioner for all the same sensors, and outputs the result as the similarity table. The above threshold is not fixed and can be adjusted to any value. However, the value of the above threshold value may be set in the same manner as in the case of the first embodiment. Based on this similarity determination, the learning air conditioner grouping unit 104 groups the learning air conditioner 1702 having a high degree of similarity with the reference air conditioner 1701.

The learning air conditioner grouping unit 104 sets the reference air conditioner 1703 again from the remaining air conditioners excluding the reference air conditioner 1701 and the grouped learning air conditioner, and repeats steps S1501 to S1508 for learning. A learning data group for the air conditioner 101 is constructed.

As shown in FIG. 3, in order to select a learning group having a high degree of similarity to the evaluation air conditioner 111, a frequency distribution graph of the maximum value, the minimum value, and the average value in the same sensor during the evaluation period of the evaluation air conditioner 111. Step S305, step S306 for calculating the similarity between the frequency distribution graphs of the maximum value, the minimum value, and the average value in the same sensor of the evaluation air conditioner 111 and the learning air conditioner 101 belonging to each learning data group, the similarity degree. There is a step S308 that ranks based on (S307) and classifies the training data group into the training data group having the highest similarity rank. Specific processing will be described below.

FIG. 18 is a flow chart for creating a frequency distribution graph of the maximum value, the minimum value, and the average value in the evaluation air conditioner 111 during the evaluation period (classification into a training data group).

The sensor signal storage unit 103 inputs the sensor signal 112 mounted on the evaluation air conditioner 111 to the evaluation air conditioner classification unit 113. Further, the sensor signal storage unit 103 also inputs the sensor signals 102 of the learning air conditioner 101 classified for each learning data group during the evaluation period into the evaluation air conditioner classification unit 113 (S1801).

The evaluation air conditioner classification unit 113 executes the processes of steps 1803 to S1806 for each sensor (S1802). As in the case of step S1303, the evaluation air conditioner classification unit 113 sets the maximum value (MAX) and the minimum value (MIN) for each type of sensor signal so that the scales of the frequency distribution graphs in all the learning air conditioners are the same. (S1803). As in the case of step S1304, the evaluation air conditioner classification unit 113 calculates the step size S when dividing the maximum value from the minimum value by the specified number N (S1804), and moves outward from the minimum value and the maximum value. The range is expanded to calculate the processing range for calculating the distribution density (S1805).

The evaluation air conditioner classification unit 113 creates a frequency distribution graph of the maximum value, the minimum value, and the average value in each sensor for the evaluation period, the evaluation air conditioner 111, and each learning data group, from steps S1807 to S1811. The process is executed (S1807).

Similar to the case of steps S1307 to S1309, the evaluation air conditioner classification unit 113 has the following for the evaluation period (S1807), each sensor (S1808), and all the learning air conditioners 101 (S1809) belonging to each learning data group. Repeat the process. The evaluation air conditioner classification unit 113 secures a two-dimensional array for calculating the frequency distribution, and sets 0 to all the elements. The size of the array is N + 2M. The evaluation air conditioner classification unit 113 adds 1 to the elements of the array corresponding to one sensor value for all the data in the learning period. By this process, the frequency distribution (histogram) by one sensor is calculated. The evaluation air conditioner classification unit 113 converts this frequency distribution into an image, and stores the converted image in association with an identification number (ID) for identifying the image (S1810). Although not shown, the evaluation air conditioner classification unit 113 records the size of the two-dimensional array and the minimum and maximum values of each sensor signal calculated in step S1803.

FIG. 19 is a flow chart of similarity calculation and similarity determination between the frequency distribution graphs of the evaluation air conditioner and the learning air conditioner belonging to each learning data group in the same sensor. First, the learning air conditioner grouping unit 104 inputs the total frequency distribution of the learning air conditioner 101 created in FIG. 13 and the frequency distribution of the air conditioner to be evaluated into the evaluation air conditioner classification unit 113 (S1901). The evaluation air conditioner classification unit 113 repeats the following processing for each evaluation air conditioner (S1902), each sensor (S1903), and each learning group (S1904). The evaluation air conditioner classification unit 113 calculates the similarity between the evaluation air conditioner and the frequency distribution graph in the same sensor of each learning group (S1905). For example, the evaluation air conditioner classification unit 113 superimposes the frequency distribution images of the sensors provided in all the learning air conditioners 101 belonging to each learning group in the same way as in the case shown in FIG. The superimposition frequency distribution image obtained by superimposing the machine superimposition frequency distribution image and the frequency distribution image of the evaluation air conditioner 111 is read, and the similarity is calculated based on the size of the area of the superimposition portion.

Finally, the evaluation air conditioner classification unit 113 ranks the evaluation air conditioners 111 in the order of the learning data groups having the highest degree of similarity (S1906), and classifies them into the learning data groups having the highest degree of similarity (S1907). The classification method may be carried out in the same manner as in FIG. Then, the evaluation air conditioner classification unit 113 learns the sensor data of the classified learning group and detects an abnormality of the evaluation air conditioner 111.

According to the second embodiment, the air conditioners are grouped according to the similarity of the sensor signals of the learning air conditioner 101, and the similarity is high by using the statistical values obtained from the sensor signals based on the sensor signals of the evaluation air conditioner 111. By learning the air conditioner group, highly sensitive abnormality detection becomes possible by accumulating data in a short period of time. In addition, since existing air conditioners at many sites can be used, a large amount of diverse learning data can be secured, so that it is possible to increase the sensitivity of abnormality detection and suppress false reports.

As described above, in the method of learning and evaluating the normal sensor data for the past one year for one air conditioner by the abnormality detection system to which the methods of Examples 1 and 2 are applied, the season and the machine of the air conditioner are used. It is possible to reduce the influence of the difference and detect abnormalities in the air conditioner with high sensitivity and at an early stage. That is, by grouping the air conditioners according to the similarity of the sensor signals of the normal air conditioners and classifying them into the air conditioner groups having a high degree of similarity based on the sensor signals of the air conditioners to be evaluated, for example, learning data is accumulated. Abnormalities can be detected with high sensitivity by accumulating data for a short period of time, such as a period (for example, one year).

101 ... Learning air conditioner, 102, 202 types of sensor signals, 103, 203 ... Sensor signal input unit, 104, 204 ... Feature vector extraction unit, 105, 205 ... Abnormal measurement calculation unit, 106 ... Threshold calculation unit, 120 ... Sensor signal storage unit, 201 ... Evaluation air conditioner, 206 ... Abnormality determination unit, 207 ... Related sensor identification unit.

Claims (4)

Learning data in which the first air conditioners are grouped is constructed based on the similarity between the first sensor signals output from one or a plurality of sensors provided in each of the plurality of first air conditioners. ,
Based on the second sensor signals output from one or more sensors provided in the second air conditioner, the second air conditioner is compared with the grouped learning data, and the second air conditioner is compared. And the second air conditioner are classified into similar groups in the training data based on the degree of similarity with the grouped training data.
Based on the training data classified the group, in abnormality detection method for determining an abnormality of the second air-conditioner,
In the grouping, a learning distribution density graph showing the correlation between the plurality of first sensor signals is generated, and the learning data is constructed based on the similarity of the generated learning distribution density graph.
In the classification, an evaluation distribution density graph showing the correlation between the plurality of second sensor signals is generated, and the second evaluation distribution density graph is based on the similarity between the generated evaluation distribution density graph and the learning distribution density graph. Classify the air conditioners into similar groups in the training data,
Anomaly detection method characterized by this. Learning data in which the first air conditioners are grouped is constructed based on the similarity between the first sensor signals output from one or a plurality of sensors provided in each of the plurality of first air conditioners. ,
Based on the second sensor signals output from one or more sensors provided in the second air conditioner, the second air conditioner is compared with the grouped learning data, and the second air conditioner is compared. And the second air conditioner are classified into similar groups in the training data based on the degree of similarity with the grouped training data.
In the abnormality detection method for determining the abnormality of the second air conditioner based on the learned data of the classified group.
In the grouping, a learning frequency distribution graph that aggregates the statistical values obtained from the first sensor signal in a predetermined period is generated, and the learning data is constructed based on the similarity of the generated learning frequency distribution graph.
In the above classification, an evaluation frequency distribution graph that aggregates statistical values obtained from the second sensor signal in a predetermined period is generated, and based on the degree of similarity between the generated evaluation frequency distribution graph and the learning frequency distribution graph, the above The second air conditioner is classified into a similar group in the training data.
Anomaly detection method characterized by this. Learning data in which the first air conditioners are grouped is constructed based on the similarity between the first sensor signals output from one or a plurality of sensors provided in each of the plurality of first air conditioners. With the grouping department
Based on the second sensor signals output from one or more sensors provided in the second air conditioner, the second air conditioner is compared with the grouped learning data, and the second air conditioner is compared. And a classification unit that classifies the second air conditioner into similar groups in the training data based on the degree of similarity with the grouped training data.
An abnormality determination unit that determines an abnormality of the second air conditioner based on the classified learning data of the group, and an abnormality determination unit.
In an abnormality detection system equipped with
The grouping unit generates a learning distribution density graph showing the correlation between the plurality of first sensor signals, and constructs the learning data based on the similarity of the generated learning distribution density graph.
The classification unit generates an evaluation distribution density graph showing the correlation between the plurality of second sensor signals, and based on the degree of similarity between the generated evaluation distribution density graph and the learning distribution density graph, the second Classify the air conditioners of the above into similar groups in the training data.
Anomaly detection system characterized by this. Learning data in which the first air conditioners are grouped is constructed based on the similarity between the first sensor signals output from one or a plurality of sensors provided in each of the plurality of first air conditioners. With the grouping department
Based on the second sensor signals output from one or more sensors provided in the second air conditioner, the second air conditioner is compared with the grouped learning data, and the second air conditioner is compared. And a classification unit that classifies the second air conditioner into similar groups in the training data based on the degree of similarity with the grouped training data.
An abnormality determination unit that determines an abnormality of the second air conditioner based on the classified learning data of the group, and an abnormality determination unit.
In an abnormality detection system equipped with
The grouping unit generates a learning frequency distribution graph that aggregates statistical values obtained from the first sensor signal in a predetermined period, and constructs the learning data based on the similarity of the generated learning frequency distribution graph. ,
The classification unit generates an evaluation frequency distribution graph that aggregates statistical values obtained from the second sensor signal in a predetermined period, and based on the degree of similarity between the generated evaluation frequency distribution graph and the learning frequency distribution graph, the classification unit generates an evaluation frequency distribution graph. The second air conditioner is classified into a similar group in the training data.
Anomaly detection system characterized by this.

Images