A method for early warning of abnormal events in urban underground comprehensive pipe gallery

Abstract

A method for warning abnormal events in urban underground integrated pipe corridors, including: obtaining target attribute data through pipe corridor sensor equipment; using Kalman filtering to perform data-level fusion on homogeneous data sensor data; using the results of data-level fusion to predict future pipe corridor data through an improved and trained TS-Transformer model; using D-S evidence theory to perform decision-level fusion on the results of data-level fusion and predicted data; using the decision-level fusion results to realize real-time detection and warning analysis in parallel. The proposed secondary data fusion model can effectively reduce the redundancy and uncertainty of pipe corridor data; the proposed prediction model solves the problem that the traditional RNN-based time series prediction model cannot be trained in parallel and is difficult to handle long-term time dependence. The proposed abnormal event warning method can not only detect whether abnormal events occur in the pipe corridor in real time, but also warn of possible accidents in advance, effectively reducing the risk of major accidents in the pipe corridor and reducing economic losses.

Description

Technical Field

[0001] The present invention relates to the field of intelligent operation and maintenance of urban underground pipe corridors, and in particular to an abnormal event early warning method for urban underground integrated pipe corridors. Background Art

[0002] The urban underground comprehensive pipeline corridor is the underground nerve center of urban management. It integrates power, communication, gas, heating, water supply and drainage engineering pipelines. It is an important infrastructure and "lifeline" to ensure the normal operation of the city, and it is also a century-long plan for urban construction. Scholars and local governments have responded to the national call and put forward a series of guiding opinions and standardized measures for the information and intelligent construction of underground comprehensive pipeline corridors.

[0003] Although with the strong support of the state, the intelligent construction of underground comprehensive pipe corridors has attracted more and more attention, there are still some problems in the intelligent operation and maintenance of urban underground comprehensive pipe corridors: insufficient emergency response capabilities, and major hidden dangers in the construction process of operation and maintenance personnel; lack of intelligent operation and maintenance platform based on comprehensive perception, pipeline operation and maintenance requires a lot of manpower and material resources. Existing intelligent operation and maintenance solutions also have some defects: 1. The operation and maintenance system based on video surveillance still requires manual long-term monitoring of the screen, which is inevitably negligent and wrong; 2. In terms of abnormal event warning systems based on sensor equipment, real-time detection is usually performed based on the collected data, which has a certain lag, and only single sensor data is used in the final warning stage, which not only has high false alarm and missed alarm rates, but also makes it difficult to make efficient and accurate decisions, and also causes a large amount of sensor data to be wasted. Therefore, in view of the problems existing in the current operation and maintenance of pipe corridors, it is necessary to combine the characteristics of the Internet of Things to study the relevant technologies of the urban underground comprehensive pipe corridor warning system based on multi-source information fusion, so as to improve the ability to respond to emergencies, reduce the workload of operation and maintenance personnel, and ensure a safe working environment for operation and maintenance personnel. Summary of the invention

[0004] The present invention aims at the technical problems existing in the existing intelligent operation and maintenance solutions of the pipe gallery, and provides an abnormal event warning method for the urban underground comprehensive pipe gallery based on multi-source information fusion. A two-level data fusion model suitable for the underground comprehensive pipe gallery is proposed, and the Kalman filter (KF) algorithm is used for data-level fusion and the DS evidence theory is used for decision-level fusion. The data collected by a large number of sensors deployed in the pipe gallery can be comprehensively utilized to reduce the redundancy and uncertainty of the data, eliminate the ambiguity of the data, and effectively reduce the missed alarm rate and false alarm rate in the early warning stage. In addition, a Transformer-based deep learning model is used to predict the data in the future pipe gallery environment, which solves the shortcomings of the traditional RNN-based time series prediction model that cannot be trained in parallel and is difficult to handle long-term time dependence, and is embedded in the pipe gallery data fusion model. Finally, the results of data-level fusion and the predicted data are used, combined with the DS evidence theory for decision-level fusion, and the pipe gallery environment is monitored in real time and the early warning analysis of the future situation is performed, which can not only detect whether an abnormal event occurs in the pipe gallery in real time, but also take action before the accident occurs, effectively reducing the risk of major accidents in the pipe gallery and reducing economic losses.

[0005] A method for early warning of abnormal events in an urban underground integrated pipe gallery comprises the following steps:

[0006] Step 1: Obtain sensor data of target attributes through various sensor devices and wireless communication devices deployed in the tunnel;

[0007] Step 2: Use Kalman filtering technology and sensor credibility to perform first-level data fusion on homogeneous sensor data, i.e., data-level fusion;

[0008] Step 3: Using the results of the first-level data fusion, the built and trained TS-Transformer model is used to predict future pipeline corridor data.

[0009] Step 4: Use DS evidence theory to perform secondary data fusion on the results of primary data fusion and the data predicted by the TS-Transformer model, i.e., decision-level fusion;

[0010] Step 5: Use the results of secondary data fusion to perform real-time detection and early warning analysis simultaneously.

[0011] Furthermore, in step 1, the sensor types include temperature sensors, humidity sensors, CO sensors, CH4 sensors, H2S sensors, smoke sensors, and light sensors, and the combination of sensor data of different types is used for the detection of different abnormal events.

[0012] Furthermore, in step 2, during the first-level data fusion, it is assumed that there are n homogeneous sensors that simultaneously monitor a target attribute in the tunnel, and the credibility of the n homogeneous sensors is calculated as a weight reference for the Kalman filter fusion stage; n sensor data within a fixed time period are selected, and their variances are calculated respectively. The smaller the variance, the lower the degree of data fluctuation, and the more stable the data collected by the sensor within the corresponding time, the higher the weight of the data collected by the sensor when using the Kalman filter for data fusion. The sensor weight calculation based on the variance credibility is performed by the following formula:

[0013]

[0014] The Kalman filter fusion stage is performed by the following formula:

[0015]

[0016] in is the prior estimate, W i is the weight obtained from the sensor confidence, K k is the Kalman gain, H is the transfer matrix, Z k is the actual observed value.

[0017] Finally, the fusion result of n homogeneous sensor data is obtained, which is used for secondary data fusion and TS-Transformer prediction.

[0018] Furthermore, in step 3, the improved Transformer deep learning model, namely the TS-Transformer model, is used to predict the pipeline corridor data. The specific construction process of the TS-Transformer model includes:

[0019] Step 3-1, based on the Transformer model, use the multi-head attention mechanism in the Transformer model to capture the long-term temporal dependency in the sequence;

[0020] Step 3-2: To avoid the impact of extreme weather events, causal convolution is used instead of linear transformation when calculating Q and K in the attention mechanism to capture local information in the time series, and a self-attention layer after causal convolution is added;

[0021] Step 3-3, discard the Decoder part of the original Transformer model and replace it with a fully connected layer to speed up model training and prediction; finally, use RELU as the activation function to train the model;

[0022] In step 3-4, multiple attention layers are stacked to maximize the prediction effect of the model.

[0023] Furthermore, in step 4, the secondary data fusion uses the result of the primary fusion as input and performs decision-level fusion through DS evidence theory for the final early warning alarm.

[0024] Furthermore, in step 4, DS evidence theory is used for secondary data fusion, and the process is as follows: for different target events, a corresponding recognition framework is established; the basic probability distribution BPA is obtained according to the measurement data of each sensor; BPA is used to perform orthogonal sum of probability distribution functions, that is, combination of evidence; based on the combination of various types of evidence, the decision result of the target event to be detected is finally obtained.

[0025] Furthermore, the TS-Transformer model is embedded into the secondary data fusion model to perform real-time detection and early warning analysis simultaneously.

[0026] The beneficial effects achieved by the present invention are as follows: the method for early warning of abnormal events in the urban underground comprehensive pipe corridor provided by the present invention first constructs a secondary data fusion model suitable for the underground comprehensive pipe corridor, which ensures the accuracy and credibility of the sensor data, and then predicts the future data of the pipe corridor through the TS-Transformer model and embeds it into the secondary data fusion model, which overcomes the shortcomings of the traditional RNN-based time series prediction model that cannot be trained in parallel and is difficult to handle long-term time dependence. It can not only realize the real-time abnormal detection of the pipe corridor sensor data, but also can more accurately warn of abnormal events that may occur in the future, improve the response capability of emergencies in the pipe corridor, reduce the workload of operation and maintenance personnel, ensure the personal safety of maintenance personnel, minimize the disaster losses of the pipe corridor, and realize the intelligent operation and maintenance of the urban underground comprehensive pipe corridor. In addition, the method has strong versatility and low cost, and is suitable for wide promotion and use. BRIEF DESCRIPTION OF THE DRAWINGS

[0027] Figure 1 It is a flow chart of the abnormal event early warning method in the implementation of the present invention.

[0028] Figure 2 It is the overall structure diagram of the secondary data fusion model in the implementation of the present invention.

[0029] Figure 3 It is a data level fusion flow chart in the secondary data fusion model in the implementation of the present invention.

[0030] Figure 4 It is a decision-level fusion flow chart in the secondary data fusion model in the implementation of the present invention.

[0031] Figure 5 It is the basic structure diagram of the TS-Transformer model in the implementation of the present invention.

[0032] Figure 6This is an example diagram of the Attention layer after adding causal convolution in the TS-Transformer model in the implementation of the present invention. DETAILED DESCRIPTION

[0033] The technical solution of the present invention is further described in detail below in conjunction with the accompanying drawings.

[0034] The process of the embodiment of the present invention is as follows Figure 1 As shown, the described embodiments include:

[0035] S1, through various sensor devices deployed in the tunnel, sensor data of target attributes are obtained through 5G or GPRS;

[0036] S2, uses Kalman filtering technology and sensor credibility to perform first-level data fusion on homogeneous sensor data, i.e., data-level fusion;

[0037] S3, using the results of the first-level data fusion, predicts future pipeline corridor data through the built and trained TS-Transformer model;

[0038] S4, using DS evidence theory to perform secondary data fusion on the results of primary data fusion and the data predicted by TS-Transformer, i.e., decision-level fusion;

[0039] S5, using the results of secondary data fusion, conducts real-time detection and early warning analysis simultaneously.

[0040] Specifically, the proposed two-level data fusion model structure is as follows: Figure 2 As shown in the figure, the first-level data fusion is the data fusion between homogeneous sensors that monitor the same target attribute, also known as data-level fusion. The specific sensor types are: temperature sensor, humidity sensor, CO (carbon monoxide) sensor, CH4 (methane) sensor, H2S (hydrogen sulfide) sensor, smoke sensor, light sensor, etc. The combination of different types of sensor data can be used for the detection of different anomalies. Among them, the data set collected by [smoke sensor, CO sensor, temperature sensor, light sensor] is used for pipe gallery fire warning; the data set collected by [humidity sensor, temperature sensor, humidity sensor, CO sensor, CH4 sensor, H2S sensor] is used for pipe gallery environmental quality assessment; humidity sensor data can be used alone to prevent pipeline short circuit problems caused by excessive humidity in the pipe gallery. The second-level data fusion uses the result of the first-level fusion as input, and performs decision-level fusion through DS evidence theory for the final warning alarm.

[0041] The first-level data fusion process is as follows: Figure 3 As shown, including:

[0042] (1) Assuming that there are n homogeneous sensors monitoring a target attribute in the tunnel at the same time, the credibility of the n homogeneous sensors is calculated as a weight reference in the Kalman filter fusion stage. The method of the present invention is as follows: select n sensor data within a fixed time period, and calculate their variance D respectively. i , The smaller the variance, the lower the data fluctuation, and the more stable the data collected by the sensor in the corresponding time. When using Kalman filtering for data fusion, the weight of the data collected by the sensor should be higher. The sensor weight calculation based on variance credibility is performed by the following formula:

[0043]

[0044] Using the Kalman filter algorithm, combined with the obtained sensor credibility, the data of n homogeneous sensors are fused, and the collected data is fused at the first level using the Kalman filter technology, that is, data-level fusion. Due to the influence of the internal environment of the tunnel and the measurement error of the sensor itself, a single sensor is often difficult to accurately describe the properties of the target object being measured, which in turn affects the accuracy of subsequent data mining and analysis. By deploying multiple sensors to monitor the same target attributes, the accuracy and reliability of the tunnel data to be processed can be improved, and the missed alarm rate and false alarm rate in the later early warning alarm stage can be effectively reduced. The process is as follows:

[0045] ① Calculate the prior estimate using the following formula:

[0046]

[0047] Since the data monitored inside the tunnel are mostly relatively stable data such as gas and temperature, and the sensor sampling frequency is controlled to sample once every 30s-1min, the change between adjacent sampling points is small, so A is set to 1 here, and there is no external control u, so B is set to 0.

[0048] ② Calculate the prior error covariance matrix:

[0049]

[0050] Here Q is the covariance matrix of the process noise.

[0051] ③Calculate the Kalman gain of each sensor:

[0052]

[0053] Here R is the covariance matrix of the measurement noise.

[0054] ④ Calculate the posterior estimation, that is, the final fusion result:

[0055]

[0056] In the formula is the prior estimate, W i is the weight obtained from the sensor confidence, K k is the Kalman gain, H is the transfer matrix, Z k is the actual observed value.

[0057] ⑤ Update the error covariance matrix of each sensor for the next iterative calculation:

[0058]

[0059] (2) Finally, the fusion result of n homogeneous sensor data is obtained, which is used for secondary data fusion and TS-Transformer prediction.

[0060] The secondary data fusion process is as follows: Figure 4 As shown, including:

[0061] (1) For different target events, establish the corresponding recognition framework H = {H1, H2, .., H k}.

[0062] Specifically, the identification framework of the tunnel fire warning system is H = {open flame, smoldering fire, no fire}, the humidity warning system in the tunnel divides the humidity into three intervals to establish a humidity level identification framework H = {dry, wet, relatively wet}, and the identification framework of the tunnel environment quality assessment system is H = {good, average, poor, severe}.

[0063] (2) Obtain a basic probability allocation (BPA) based on the measurement data of each sensor, where the BPA needs to satisfy:

[0064] m:2 H →[0,1],

[0065]

[0066] Where m is the basic probability distribution function of H and A is a subset of H.

[0067] (3) using BPA to perform orthogonal sum of probability assignment functions, i.e., combination of evidence;

[0068]

[0069] In the formula, A and B are both focal elements, is the basic probability distribution after synthesis.

[0070] (4) Based on the combination of various types of evidence, the decision result of the target event to be detected is finally obtained.

[0071] The structure of the Transformer-based deep learning model is as follows Figure 5 As shown in the figure, to avoid confusion with the original Transformer model, the modified Transformer model is named TS-Transformer (TimeSeries Transformer), including:

[0072] (1) Based on the Transformer model, the Multi-Head Self Attention mechanism in the Transformer is used to capture the long-term temporal dependencies in the sequence, which is described in detail as follows:

[0073] Assume that the target value z at each time step is t Obey the probability distribution l(z t θ t ), use the multi-head attention mechanism to calculate the hidden state of the current time step, and then calculate the parameter θ of the probability distribution t , and finally estimate by maximizing the log-likelihood To learn the network parameters. Because the attention mechanism is used, there is no need to perform a ground cabinet like the LSTM model, but to calculate the θ of all time steps at one time t ,like Figure 5 shown.

[0074] (2) To avoid the impact of events such as extreme weather, causal convolution is used so that the self-attention mechanism can capture local information.

[0075] There are usually some anomalies in time series. Whether an observation should be considered anomaly depends to a large extent on the context in which it is located. The original Transformer attention mechanism does not consider the local environment of each position when calculating the relationship between different positions in the sequence, which makes the prediction easily disturbed by outliers. If convolution is used instead of Q and K when calculating the self-attention mechanism, local information can be introduced. The schematic diagram is shown below. Figure 6 shown.

[0076] (3) Discard the Decoder part of the original Transformer to speed up model training and prediction; finally, use RELU as the activation function to train the model.

[0077] After inputting the target attribute feature vector into the TS-Transformer model and outputting corresponding prediction data, the method further includes:

[0078] Based on the SelfAttention layer in the existing TS-Transformer model, a new layer with the same structure as the first layer is added as the second layer, where the output of the first SelfAttention layer is used as input in the second layer;

[0079] Optimize the training of the TS-Transformer model after adding new layers;

[0080] Add new layers on top of the second layer and train until the performance of the TS-Transformer model no longer improves;

[0081] The predicted values ​​of the corresponding target attributes in the tunnel obtained by the final TS-Transformer model are taken as the final prediction results.

[0082] Abnormal events include fire inside the tunnel, poor environmental quality inside the tunnel, and high humidity inside the tunnel. The possible results are: major disasters and accidents in the tunnel, threats to the personal safety of maintenance personnel, and short circuits caused by excessive humidity. For the three types of abnormal events, the data sets that need to be selected are: [smoke sensor data, CO sensor data, temperature sensor data, light sensor data], [humidity sensor data, temperature sensor data, humidity sensor data, CO sensor data, CH4 sensor data, H2S sensor data], [humidity sensor data].

[0083] The above description is only a preferred embodiment of the present invention, and the protection scope of the present invention is not limited to the above embodiment. Any equivalent modifications or changes made by ordinary technicians in this field based on the contents disclosed by the present invention should be included in the protection scope recorded in the claims.

Claims

1. A method for early warning of abnormal events in an urban underground comprehensive pipe gallery, characterized in that: The method comprises the following steps: Step 1: Obtain sensor data of target attributes through various sensor devices and wireless communication devices deployed in the tunnel; Step 2: Use Kalman filtering technology and sensor credibility to perform first-level data fusion on homogeneous sensor data, i.e., data-level fusion; In step 2, during the first-level data fusion, it is assumed that there are n homogeneous sensors that simultaneously monitor a target attribute in the tunnel, and the credibility of the n homogeneous sensors is calculated as a weight reference for the Kalman filter fusion stage; Select n sensor data within a fixed time period and calculate their variances respectively. The smaller the variance, the lower the data fluctuation. The more stable the data collected by the sensor in the corresponding time, the higher the weight of the data collected by the sensor should be when using Kalman filtering for data fusion. The sensor weight calculation based on variance credibility is performed by the following formula: The Kalman filter fusion stage is performed by the following formula: in is the prior estimate, W i is the weight obtained from the sensor confidence, K k is the Kalman gain, H is the transfer matrix, Z k is the actual observed value; Finally, the fusion result of n homogeneous sensor data is obtained, which is used for secondary data fusion and TS-Transformer prediction; Step 3: Using the results of the first-level data fusion, the built and trained TS-Transformer model is used to predict future pipeline corridor data. Step 4: Use DS evidence theory to perform secondary data fusion on the results of primary data fusion and the data predicted by the TS-Transformer model, i.e., decision-level fusion; Step 5: Use the results of secondary data fusion to perform real-time detection and early warning analysis simultaneously.

2. The method for early warning of abnormal events in an urban underground comprehensive pipe gallery according to claim 1 is characterized in that: In step 1, the sensor types include temperature sensor, humidity sensor, CO sensor, CH4 sensor, H2S sensor, smoke sensor, and light sensor. The combination of different types of sensor data is used to detect different abnormal events.

3. The method for early warning of abnormal events in an urban underground comprehensive pipe gallery according to claim 1 is characterized in that: In step 3, the improved Transformer deep learning model, namely the TS-Transformer model, is used to predict the pipeline corridor data. The specific construction process of the TS-Transformer model includes: Step 3-1, based on the Transformer model, use the multi-head attention mechanism in the Transformer model to capture the long-term temporal dependency in the sequence; Step 3-2: To avoid the impact of extreme weather events, causal convolution is used instead of linear transformation when calculating Q and K in the attention mechanism to capture local information in the time series, and a self-attention layer after causal convolution is added; Step 3-3, discard the Decoder part of the original Transformer model and replace it with a fully connected layer to speed up model training and prediction; finally, use RELU as the activation function to train the model; In step 3-4, multiple attention layers are stacked to maximize the prediction effect of the model.

4. The method for early warning of abnormal events in an urban underground comprehensive pipe gallery according to claim 1 is characterized in that: In step 4, the second-level data fusion uses the result of the first-level fusion as input and performs decision-level fusion through DS evidence theory for the final early warning alarm.

5. The method for early warning of abnormal events in an urban underground comprehensive pipe gallery according to claim 1 is characterized in that: In step 4, DS evidence theory is used for secondary data fusion, and the process is as follows: establish a corresponding recognition framework for different target events; obtain the basic probability distribution BPA based on the measurement data of each sensor; use BPA to perform orthogonal sum of probability distribution functions, that is, combination of evidence; based on the combination of various types of evidence, finally obtain the decision result of the target event to be detected.

6. The method for early warning of abnormal events in an urban underground comprehensive pipe gallery according to claim 1 is characterized in that: The TS-Transformer model is embedded into the secondary data fusion model to perform real-time detection and early warning analysis simultaneously.

Images