Comprehensive pipe gallery operation and maintenance prediction method based on long short-term memory recurrent neural network
By using the LSTM model in the pipeline system combined with the CNN and BP neural networks in the integrated pipeline operation and maintenance data prediction system, the shortcomings of traditional neural networks in dealing with long-term dependence problems are solved, and the stability and security of the system are improved. , improves the accuracy of predicting risk events and control efficiency.
Patent Information
- Application Number
- CN202011385524.9
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2020-11-30
- Publication Date
- 2024-11-05
- Estimated Expiration
- 2040-11-30
AI Technical Summary
Traditional neural networks are difficult to deal with long-term dependence problems in pipeline systems, resulting in poor system stability and autonomy, and lack of effective verification and error correction capabilities, which affects the accuracy and security of risk assessment.
Using a comprehensive pipeline operation and maintenance data prediction system based on the LSTM model, combined with CNN convolutional operation and BP neural network, time series data is processed through long and short-term memory recurrent neural network, short-term memory and long-term memory risk prediction results are generated, and the results are obtained through The gate structure controls the flow of information to achieve independent learning and error correction.
It improves the operating stability and safety of the pipeline system, enhances the short-term and long-term prediction capabilities of risk events, reduces labor intensity and cost, and improves the efficiency and accuracy of risk judgment.
Smart Images

Figure CN112598222B_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to a pipe gallery risk assessment system and method, and belongs to the technical field of information communication and artificial intelligence. Background Art
[0002] In the operation of the urban integrated pipe corridor system, in order to improve the comprehensive supervision ability of operation risk events and improve the safety of pipe corridor operation, a large number of artificial intelligence-based risk prediction system platforms are currently used in supervision work. However, in actual work, it is found that during the operation of the current artificial intelligence risk prediction platform, the traditional neural network architecture has not been able to solve some basic problems, such as interpreting the input sequence that depends on information and context. This information can be certain words in a sentence, which we can use to predict what the next word is; it can also be the time information of the sequence, and we can analyze the context of the sentence based on the time element. In short, the traditional neural network only uses an independent data vector each time, and it does not have a concept like "memory" to handle various tasks related to "memory".
[0003] In order to solve this problem, an early method was to add a loop to the network. After the output value is obtained, its input information will be "inherited" to the output through the loop, which is the last input context it sees. These networks are called recursive neural networks (RNNs). Although RNNs have solved the above problems to a certain extent, they still have considerable defects. For example, gradient disappearance is prone to occur when dealing with long-term dependency problems. Therefore, the stability and autonomy of the system operation are relatively poor. In addition, the current operation of such systems also lacks the corresponding verification and error correction capabilities for the evaluation results, which causes great inconvenience and safety hazards to the risk assessment of the pipeline corridor system.
[0004] Therefore, in response to this situation, there is an urgent need to develop a new urban pipeline corridor risk prediction system to meet the needs of the actual operation, maintenance and management of the pipeline corridor system. Summary of the invention
[0005] In view of the shortcomings of the prior art, the present invention provides a comprehensive pipe gallery operation and maintenance data prediction system and construction method based on the LSTM model to overcome the defects of traditional equipment in operation and improve the stability, safety and reliability of the pipe gallery system operation. In order to achieve the above purpose, the present invention is implemented through the following technical solutions:
[0006] The comprehensive pipe gallery operation and maintenance prediction and method based on long short-term memory recurrent neural network includes the following steps:
[0007] S1, system prefabrication, first build a main program system based on SOA system, a risk server with LSTM intelligent prediction system as the core, and the risk assessment server is equipped with CNN convolution operation system and BP neural network system, and the LSTM intelligent prediction system is connected with CNN convolution operation system and BP neural network system respectively, and the CNN convolution operation system is connected with BP neural network system, so as to form a risk operation and maintenance prediction system;
[0008] S2, data preprocessing. After completing step S1, the LSTM intelligent prediction system first establishes a data operation model based on time series, and contains at least three cell neural networks, and each cell neural network contains a hidden layer structure of four ordinary neural networks in series. The activation functions of the first, second and fourth are all sigmoid functions, and the activation function of the third is tanh function. At the same time, first, the input X at time t and the output h(t-1) at time t-1 are concatenated and then input into the cell. In fact, it can be understood that our input X(t) is fed into the hidden layer structure of four ordinary neural networks respectively. The operations performed in the hidden layer structure of each ordinary neural network are the same as those of normal neural networks (matrix multiplication), and the memory part is completely controlled by each A gate structure is used to control (that is, 0 and 1). At the same time, the input not only has the original data set, but also adds the output result of the previous data, that is, h(t-1). Thus, two parallel distributed data processing paths are generated in the cell at the same time, one of which is used to control short-term memory, and the other is used to control long-term memory. On the other hand, three types of logic judgment nodes, namely input gate, forget gate and output gate, are generated to complete the setting of LSTM intelligent prediction system. Then, the operation and maintenance data collected in the operation and maintenance of the pipe gallery system are input into the risk server through external devices, and then the operation and maintenance data are sent to the CNN convolution operation system. After the preliminary operation of the CNN convolution operation system, the calculated data are sent to the LSTM intelligent prediction system and BP neural network system at the same time.
[0009] S3, risk prediction. After completing step S2, on the one hand, the BP neural network system performs a second run on the received data and caches the run results for backup; on the other hand, the LSTM intelligent prediction system runs the received data and divides the received data into short-term memory risk prediction results and long-term memory risk prediction results after running. Then, the BP neural network system performs a second comparison operation based on the received short-term memory risk prediction results and long-term memory risk prediction results with the previously cached run results, and then outputs the risk time and risk level corresponding to the current pipe gallery system operation and maintenance data, and outputs this as the final result;
[0010] Furthermore, in the step S1, the constructed server is data-connected to the control terminal of the pipe gallery system through an external data communication network.
[0011] Furthermore, in the step S2, when setting the LSTM intelligent prediction system, the LSTM intelligent prediction system is constructed using the RNN model structure.
[0012] Furthermore, in the S3 step, when the LSTM intelligent prediction system is running, the following steps are performed:
[0013] In the first step, the forget gate runs and chooses to forget some past information. The first step in LSTM is to decide what information to discard from the cell state;
[0014] Its operating function is: ft = σ(wf · [h t-1, xt] + bf);
[0015] Where ht-1 represents the output of the previous cell, xt represents the input of the current cell, and σ represents the sigmoid function.
[0016] In the second step, the input gate runs to decide how much new information to add to the cell state. First, a sigmoid layer, which is the input gate, determines which information needs to be updated. Another tanh layer generates a vector, which is the candidate content for update. Then these two parts are combined to update the cell state.
[0017] Its operating function is: it = σ (wi [h t-1, xt] + bi)
[0018] it=σ(wi·[h t-1,xt]+bi)
[0019] In the third step, the output gate runs, which needs to determine what value to output. First, a sigmoid layer is run to determine which part of the cell state will be output; then, we process the cell state through a tanh layer (to obtain a value between -1 and 1) and multiply it with the output of the sigmoid layer, and finally output the part of the data that determines the output.
[0020] Furthermore, in the S3 step, the short-term memory risk prediction results and the long-term memory risk prediction results sent to the BP neural network system are both retained in the risk server, wherein the short-term memory risk prediction results are directly deleted from the LSTM's intelligent prediction system when the next risk assessment is conducted, and the long-term memory risk prediction results are retained in the LSTM's intelligent prediction system for continuous operation. In addition, when the frequency of occurrence of the short-term memory risk prediction results accounts for more than 40% of the risk prediction operation times, the short-term memory risk prediction results are converted into long-term memory risk prediction results and applied; when the frequency of occurrence of the long-term memory risk prediction results is less than 2%, they are deleted from the LSTM's intelligent prediction system and stored in the risk server as original backup data for backup.
[0021] On the one hand, the system structure of the present invention is simple and has strong data communication processing capabilities, which can effectively meet the needs of short-term and long-term risk prediction of risk events in the operation and maintenance of the corridor system; on the other hand, it has good autonomous learning and operation capabilities and error correction capabilities, which can greatly reduce the labor intensity and cost of risk event prediction operations in the corridor system; at the same time, it greatly improves the efficiency and accuracy of risk event judgment, and greatly improves the work efficiency and accuracy of corridor operation risk management and control. BRIEF DESCRIPTION OF THE DRAWINGS
[0022] The present invention will be described in detail below with reference to the accompanying drawings and specific embodiments.
[0023] Figure 1 It is a flow chart of the operating method of the present invention;
[0024] Figure 2 This is a schematic diagram of the operation flow of the LSTM intelligent prediction system. DETAILED DESCRIPTION
[0025] In order to make the technical means, creative features, objectives and effects achieved by the present invention easy to understand, the present invention is further explained below in conjunction with specific implementation methods.
[0026] like Figure 1 and 2 As shown, the comprehensive pipe gallery operation and maintenance prediction and method based on long short-term memory recurrent neural network includes the following steps:
[0027] S1, system prefabrication, first build a main program system based on SOA system, a risk server with LSTM intelligent prediction system as the core, and the risk assessment server is equipped with CNN convolution operation system and BP neural network system, and the LSTM intelligent prediction system is connected with CNN convolution operation system and BP neural network system respectively, and the CNN convolution operation system is connected with BP neural network system, so as to form a risk operation and maintenance prediction system;
[0028] S2, data preprocessing. After completing step S1, the LSTM intelligent prediction system first establishes a data operation model based on time series, and contains at least three cell neural networks, and each cell neural network contains a hidden layer structure of four ordinary neural networks in series. The activation functions of the first, second and fourth are all sigmoid functions, and the activation function of the third is tanh function. At the same time, first, the input X at time t and the output h(t-1) at time t-1 are concatenated and then input into the cell. In fact, it can be understood that our input X(t) is fed into the hidden layer structure of four ordinary neural networks respectively. The operations performed in the hidden layer structure of each ordinary neural network are the same as those of normal neural networks (matrix multiplication), and the memory part is completely controlled by each A gate structure is used to control (that is, 0 and 1). At the same time, the input not only has the original data set, but also adds the output result of the previous data, that is, h(t-1). Thus, two parallel distributed data processing paths are generated in the cell at the same time, one of which is used to control short-term memory, and the other is used to control long-term memory. On the other hand, three types of logic judgment nodes, namely input gate, forget gate and output gate, are generated to complete the setting of LSTM intelligent prediction system. Then, the operation and maintenance data collected in the operation and maintenance of the pipe gallery system are input into the risk server through external devices, and then the operation and maintenance data are sent to the CNN convolution operation system. After the preliminary operation of the CNN convolution operation system, the calculated data are sent to the LSTM intelligent prediction system and BP neural network system at the same time.
[0029] S3, risk prediction. After completing step S2, on the one hand, the BP neural network system performs a second run on the received data and caches the run results for backup; on the other hand, the LSTM intelligent prediction system runs the received data and divides the received data into short-term memory risk prediction results and long-term memory risk prediction results after running. Then, the BP neural network system performs a second comparison operation based on the received short-term memory risk prediction results and long-term memory risk prediction results with the previously cached run results, and then outputs the risk time and risk level corresponding to the current pipe gallery system operation and maintenance data, and outputs this as the final result;
[0030] Furthermore, in the step S1, the constructed server is data-connected to the control terminal of the pipe gallery system through an external data communication network.
[0031] Furthermore, in the step S2, when setting the LSTM intelligent prediction system, the LSTM intelligent prediction system is constructed using the RNN model structure.
[0032] Furthermore, in the S3 step, when the LSTM intelligent prediction system is running, the following steps are performed:
[0033] In the first step, the forget gate runs and chooses to forget some past information. The first step in LSTM is to decide what information to discard from the cell state;
[0034] Its operating function is: ft = σ(wf · [h t-1, xt] + bf);
[0035] Where ht-1 represents the output of the previous cell, xt represents the input of the current cell, and σ represents the sigmoid function.
[0036] In the second step, the input gate runs to decide how much new information to add to the cell state. First, a sigmoid layer, which is the input gate, determines which information needs to be updated. Another tanh layer generates a vector, which is the candidate content for update. Then these two parts are combined to update the cell state.
[0037] Its operating function is: it = σ (wi [h t-1, xt] + bi)
[0038] it=σ(wi·[h t-1,xt]+bi)
[0039] In the third step, the output gate runs, which needs to determine what value to output. First, a sigmoid layer is run to determine which part of the cell state will be output; then, we process the cell state through a tanh layer (to obtain a value between -1 and 1) and multiply it with the output of the sigmoid layer, and finally output the part of the data that determines the output.
[0040] Furthermore, in the S3 step, the short-term memory risk prediction results and the long-term memory risk prediction results sent to the BP neural network system are both retained in the risk server, wherein the short-term memory risk prediction results are directly deleted from the LSTM's intelligent prediction system when the next risk assessment is conducted, and the long-term memory risk prediction results are retained in the LSTM's intelligent prediction system for continuous operation. In addition, when the frequency of occurrence of the short-term memory risk prediction results accounts for more than 40% of the risk prediction operation times, the short-term memory risk prediction results are converted into long-term memory risk prediction results and applied; when the frequency of occurrence of the long-term memory risk prediction results is less than 2%, they are deleted from the LSTM's intelligent prediction system and stored in the risk server as original backup data for backup.
[0041] On the one hand, the system structure of the present invention is simple and has strong data communication processing capabilities, which can effectively meet the needs of short-term and long-term risk prediction of risk events in the operation and maintenance of the corridor system; on the other hand, it has good autonomous learning and operation capabilities and error correction capabilities, which can greatly reduce the labor intensity and cost of risk event prediction operations in the corridor system; at the same time, it greatly improves the efficiency and accuracy of risk event judgment, and greatly improves the work efficiency and accuracy of corridor operation risk management and control.
[0042] It should be understood by those skilled in the art that the present invention is not limited by the above-described embodiments. The above-described embodiments and descriptions are merely illustrative of the principles of the present invention. Various changes and improvements may be made to the present invention without departing from the spirit and scope of the present invention. These changes and improvements fall within the scope of the present invention to be protected. The scope of protection of the present invention is defined by the appended claims and their equivalents.
Claims
1. A comprehensive pipe gallery operation and maintenance prediction method based on long short-term memory recurrent neural network is characterized by: The comprehensive pipe gallery operation and maintenance prediction and method based on long short-term memory recurrent neural network includes the following steps: S1, system prefabrication, first build a main program system based on SOA system, a risk server with LSTM intelligent prediction system as the core, and the risk server is also provided with CNN convolution operation system and BP neural network system, and the LSTM intelligent prediction system is connected with CNN convolution operation system and BP neural network system respectively, and the CNN convolution operation system is also connected with BP neural network system, so as to form a risk operation and maintenance prediction system; S2, data preprocessing. After completing step S1, the LSTM intelligent prediction system first establishes a data operation model based on time series, and contains at least three cell neural networks, and each cell neural network contains a hidden layer structure of four ordinary neural networks in series, where the first, second and fourth activation functions are all sigmoid functions, and the third activation function is tanh function. At the same time, first set the input X at time t and the output h(t-1) at time t-1 to be spliced, and then input into the cell. Specifically, the input X(t) is fed into the hidden layer structure of the four ordinary neural networks respectively. The calculations performed in the hidden layer structure of each ordinary neural network are the same as those of the normal neural network. The memory-related part is completely composed of various gate structures. To control, at the same time, when inputting, there is not only the original data set, but also the output result of the previous data, that is, h(t-1), so that two parallel distributed data processing paths are generated in the cell at the same time, one of which is used to control short-term memory, and the other is used to control long-term memory; on the other hand, three types of logic judgment nodes, input gate, forget gate and output gate, are generated to complete the setting of LSTM intelligent prediction system, and then the operation and maintenance data collected in the operation and maintenance of the pipe gallery system are input into the risk server through external devices, and then the operation and maintenance data are sent to the CNN convolution operation system. After the preliminary operation of the CNN convolution operation system, the calculated data is sent to the LSTM intelligent prediction system and BP neural network system at the same time; S3, risk prediction. After completing step S2, on the one hand, the BP neural network system performs a second run on the received data and caches the run results for backup; on the other hand, the LSTM intelligent prediction system runs the received data and divides the received data into short-term memory risk prediction results and long-term memory risk prediction results after running. Then, the BP neural network system performs a second comparison operation based on the received short-term memory risk prediction results and long-term memory risk prediction results with the previously cached run results, and then outputs the risk time and risk level corresponding to the current tunnel system operation and maintenance data, and outputs this as the final result.
2. The method for predicting operation and maintenance of a comprehensive pipe gallery based on a long short-term memory recurrent neural network according to claim 1 is characterized in that: In the step S1, the constructed server is connected to the control terminal of the pipe gallery system through an external data communication network.
3. The method for predicting operation and maintenance of a comprehensive pipe gallery based on a long short-term memory recurrent neural network according to claim 1 is characterized in that: In the step S2, when setting the LSTM intelligent prediction system, the LSTM intelligent prediction system is constructed using the RNN model structure.
4. The method for predicting operation and maintenance of a comprehensive pipe gallery based on a long short-term memory recurrent neural network according to claim 1 is characterized in that: In the S3 step, the short-term memory risk prediction results and the long-term memory risk prediction results sent to the BP neural network system are both retained in the risk server, wherein the short-term memory risk prediction results are directly deleted from the LSTM intelligent prediction system when the next risk assessment is performed, and the long-term memory risk prediction results are retained in the LSTM intelligent prediction system for continuous operation. In addition, when the frequency of occurrence of the short-term memory risk prediction result accounts for more than 40% of the risk prediction operation times, the short-term memory risk prediction result is converted into a long-term memory risk prediction result and applied. When the frequency of occurrence of the long-term memory risk prediction result is less than 2%, it is deleted from the LSTM intelligent prediction system and saved in the risk server as original backup data for backup.
Citation Information
Patent Citations
Series-wound long short-term memory recurrent neural network-based heating load prediction method
CN107239859A
A building energy consumption prediction method based on a recurrent neural network and a multi-task learning model
CN109685252A
Enterprise risk early warning analysis method and system based on a memory network
CN109829631A
Short-term wind power prediction method, device and equipment
CN110363360A
Artificial intelligence CNN and LSTM neural network voice recognition system
CN110738984A
