JPH07200702A - Abnormality diagnostic system for building facility - Google Patents

Abstract

PURPOSE:To easily perform state estimation such as the abnormality classification or abnormality diagnosis of installed equipment from the operation data of building installed equipment in a normal operating state without depending on any skilled person. CONSTITUTION:Any one of data at a certain time point in the normal state of equipment, all the data sampled for a fixed period and data to be time sequentially inputted are defined as the input data to a neural network concerning the building installed equipment in the normal operating state, and the abnormality classification or state estimation of installed equipment is enabled by outputs corresponding to the input data.

Description

Detailed Description of the Invention

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to an abnormality diagnosis system for building equipment that organically connects on-site work and management of building equipment management and supports work for smoothly performing the function.

[0002]

2. Description of the Related Art Conventionally, building facility management includes a building automation system that performs operations such as operation / monitoring, control, and measurement online, and a building maintenance system that performs operations and management offline, which are usually integrated. It is operated as a closed system that is centralized and centralized.

In the facility management of this building, it is essential to carry out rational maintenance by avoiding over-maintenance by properly performing facility diagnosis. In particular, the so-called predictive maintenance, which detects abnormal signs from situation observations such as online data of facility equipment and patrol monitoring data, predicts the causal relationship and notifies the maintenance time before a failure occurs, so-called predictive maintenance Currently, it is still at an early stage, which means that it has been judged.

By the way, the environment surrounding buildings has changed significantly in recent years due to the changes in the situation of the times.
As the number of highly functional building facilities that meet the needs of the times such as individualization is increasing, the shortage of specialists is becoming apparent, and it is strongly desired to reduce the maintenance cost at the same time as quickly responding to the black box. ing.

[0005] In recent years, several abnormality diagnosis expert systems have been announced for abnormality diagnosis of equipment, especially for plant equipment. These expert systems
This is a knowledge-based system based on writing down knowledge of an expert's abnormality in the form of "If ... TheN ...". This is the result that the knowledge base system that simplifies to derive the result from the cause by handing the "if A, then B" three-stage logic is considered to be a good system for abnormality diagnosis. is there. Certainly, if all situations can be symbolized and verbalized, it can be a preferred embodiment. However, it is very difficult to create a rule that interprets the situation from the numerical data and symbolizes it, and there is a limit to accurately performing equipment abnormality diagnosis with a conventional expert system or the like.

[0006]

[Problems to be Solved by the Invention] The above-described equipment abnormality diagnosis system has the following problems. (1) Acquiring knowledge about abnormalities In equipment where most of the equipment is in normal operation and rarely falls into an abnormal operating state, it is very common to document the abnormality as generalized knowledge and collect a large amount of data. This is a difficult task. Furthermore, it is difficult to build up the process leading to an abnormality from the numerical information obtained from each sensor. (2) Knowledge representation of unsteady behavior It is difficult to accurately symbolize and verbalize the operating status and behavior of equipment.

[0007]

Means for Solving the Problems A building equipment abnormality diagnosis system for solving the above-mentioned problem is a building equipment device in a normal operating state, in which a specific operation data of the device is input to a neural network. The output corresponding to the input data enables abnormality classification or state estimation of the equipment.

[0008]

The abnormality diagnosis system for building equipment according to the present invention, based on the operation data of the equipment in a normal operation state,
By using a neural network with an excellent pattern matching function, the situation can be grasped from a macro point of view, so that it is possible to recognize abnormalities not only in the steady operation state of equipment but also in the non-steady state. In addition, since it is possible to recognize anomalies in real time, it is possible to avoid serious damage due to early detection of a failure.

[0009]

Embodiment An embodiment of the present invention will be described below with reference to FIGS. FIG. 1 is a real-time non-stationary value estimation system using one data prediction type neural network (hereinafter referred to as NN) of the present invention.

1 is a supernet, subnets 2, 6,
It is composed of 7. Subnet 2 is input layer 3,
It is an NN consisting of a middle layer 4 and an output layer 5, and a subnet 6
Also, 7 and 7 have the same configuration as the subnet 2. Reference numerals 8 and 9 are the intermediate layer and the output layer of the supernet 1, respectively. As an example of the target system of FIG. 1, an example of vibration data of an air conditioner is shown in FIG. This vibration data is detected by an acceleration sensor as a change in acceleration, for example, and an abnormal state of equipment such as an air conditioner is predicted from the spectral distribution of the sensor output at this time. In FIG. 2, A is an observation section, t is the current time, and (t-1) to (t-
3) indicates past time, and (t + 1) indicates future time.

By the way, the vibration data of the equipment as shown in FIG. 2 is usually overwhelmingly normal. Therefore, the time-series data pattern in this normal state is learned and recognized by the neural network shown in FIG. 1, and the occurrence of an event different from the normal data pattern is detected as an abnormality in real time. explain.

Time t, (t-1) to (t-3), (t +
1) is the same as that shown in FIG. In the input layer 3 at the time of the last three steps (t-3) at the present time t, (t-
The spectrum of the vibration data in 3) is input, which is performed as follows, for example.

Assuming that the interval (t-3) in FIG. 2 is 1 second, 1/60 second is 1 cycle when the frequency of the power source is 60 HZ, and the sensor output is FFT (Fast Fourier Transform) every cycle. To convert it to spectrum. Then, in order to reduce the influence of noise, one second, that is, 60 averages are added and averaged. Then, the amplitude is normalized so that the maximum value of the spectrum becomes 1, and this is input to the input layer 3 at the time (t-3) in FIG. The number of input layers 3 is, for example, 3
2 (or 64), the number of intermediate layers 4 is 25 (or 5)
0), and the number of output layers 5 is set to 1. Here, assuming that the output value of the output layer 5 is the rms value (effective value) of the spectrum at the past two step points (t-2) at the present time t, the number of output layers 5 is the same as that of the input layer 3 32 (or 64). Then, the rms value of the spectrum may be calculated, and the number of output layers 5 may be unified as shown in FIG. Subnets 6 and 7 are similar to subnet 2.

Subnet 2, obtained as described above,
As outputs of 6 and 7, 10 is output from the output layer 9 via the intermediate layer 8 of the supernet 1 as an estimated value one step (t + 1) ahead of the present time t. Here, the threshold function of the NN unit is a sigmoid function (not shown), and the well-known delta rule is used as a learning rule, and a network is constructed by recursive learning that modifies the weight of connection between units. . Therefore, the error that occurs during learning, that is, the maximum learning error with respect to the actual rms value of the spectrum is obtained, and if this value exceeds 5%, for example, it is determined to be abnormal. According to this method, it is possible to classify anomalies such as forecast, warning, and emergency according to the size of the maximum learning error.

Next, as specific operation data of the equipment,
The data (steady data) at a certain point in the steady state is N
An example of inputting N will be described with reference to FIG. Figure 3
Is an abnormality diagnosis system for air conditioning equipment.

The number of units in the input layer is 7, and the input elements include, for example, current, voltage power, cooling water inlet temperature, cooling water outlet temperature, room temperature, and room humidity. The numbers of units in the middle layer and the output layer are 4 and 1, respectively, and a steady value of vibration (amplitude) is output from the output layer. The learning of the NN in FIG. 3 is performed in the same manner as in the case of FIG. 1, but the difference is that the steady value estimation system is used because FIG. 3 handles steady data. In this case, the application is limited to the steady state, but advantages such as easy data collection, small calculation amount, and easy correspondence to a plurality of abnormal states are given.

As another example of the present invention, an unsteady value estimation system is shown in FIG. 4. In this case, the NN also has a three-layer structure with one input layer, one intermediate layer, and one output layer, and the input data is constant. All data (time series pattern) sampled during the period is targeted. In FIG. 4, P1 to P6 in the input layer are the input elements, and the input time series pattern P1 is in the output layer.
~ P6 corresponding to P6 (time series pattern) P'1
P'6 is output. In this case, it can be applied even in a non-steady state, and by treating the data sampled in a certain period as one case, it is easy to deal with a plurality of abnormal states.

[0018]

As described above, according to the present invention,
Since the specific operation data of the equipment in the normal operation state is learned and recognized as the input of the neural network, the data in the abnormal operation state, which is usually difficult to obtain, is not necessary and the steady operation of the equipment is possible. Not only the state,
Abnormality diagnosis in an unsteady state can be easily performed without relying on an expert. In addition, since it is possible to perform abnormality diagnosis in real time, there is an effect that predictive maintenance of equipment can be performed, and maintenance of building equipment can be performed with high reliability and at a low price.

[Brief description of drawings]

FIG. 1 is a network showing a real-time non-stationary value estimation system according to an embodiment of the present invention.

FIG. 2 is an example of vibration data of an air conditioner.

FIG. 3 is a network showing an abnormality diagnosis system for air conditioning equipment.

FIG. 4 is a network showing an unsteady value estimation system.

[Explanation of symbols]

 1 Supernet 2 Subnet 3 Input Layer 4 Middle Layer 5 Output Layer 6 Subnet 7 Subnet 8 Middle Layer 9 Output Layer 10 Output A Observation Section

Claims (2)

[Claims] 1. In a building facility device in a normal operating state, specific operation data of the device is used as input data to a neural network, and an output corresponding to the input data is used to estimate an abnormality or state of the facility device. An abnormality diagnosis system for building equipment. 2. The specific operation data is either data at a certain point in a steady state, all data sampled in a certain period, or data entered in a time series. Abnormality diagnosis system for building equipment.