Fault Inference Method for Energy Thermal System Diagram Embedding by Fusing Slow Signal Time-Frequency Characterization

Abstract

The invention discloses a method for fault inference of energy thermal system graph embedding by fusing time-frequency representation of slow signals. This method conducts joint time-domain and frequency-domain characterization of thermodynamic slow signals with low sampling frequency and slow signal changes in the thermal system through time-frequency analysis, constructs a knowledge graph among parameters, and uses the graph convolutional network (GCN) and spatio-temporal graph convolutional network (STGCN) to learn graph-structured data, realizing end-to-end fault inference for complex energy thermal systems. By adopting data interpolation encryption and sparse time-frequency representation methods, the invention can effectively extract the time-frequency domain features of slow signals, expanding the representation dimension of thermodynamic signals; integrating the correlation relationships among parameters by using a knowledge graph and combining CNN-GCN and STGCN for "time-space-frequency" cross-modal fusion decision-making, the fault inference ability of complex energy thermal systems is greatly improved, and the diagnostic accuracy and reliability of the system are enhanced.

Description

Technical Field

[0001] The present invention belongs to the technical field of intelligent fault reasoning for energy thermal systems, and particularly relates to a method for fault reasoning of energy thermal system graph embedding that integrates time-frequency characterization of slow signals. Background Technique

[0002] Efficient energy utilization is the main focus of current international research. The energy thermal system is the core basic system in the field of energy utilization and is the basic carrier for heat supply, cooling, and energy transmission. Faults in the thermal system will seriously affect the quality and efficiency of energy transfer and are technical difficulties that hinder the improvement of the quality and efficiency of the energy utilization process.

[0003] Energy thermal systems usually involve multiple energy flow coupling transmissions, involving complex heat-mass transfer and work-heat conversion, making it very difficult to diagnose faults in real time. Most traditional diagnostic techniques rely on means such as expert experience, rule reasoning, and single-variable analysis, and it is difficult to comprehensively capture the spatio-temporal dynamic characteristics in the system and the complex relationships between multiple variables, and the accuracy and real-time performance of fault diagnosis are limited.

[0004] With the rapid development of sensing technology, a large amount of operation data in the thermal system, such as pressure, temperature, flow rate, etc., are collected and saved in real time. Making high-quality use of the big data of the thermal system, learning, interpreting, and judging the high-order characterization information of the system state from the data, so as to more accurately identify the fault state of complex thermal systems, has become an important development direction in the future field of energy fault diagnosis.

[0005] Complex energy thermal systems, such as the primary loop and secondary loop of nuclear power plants and large central air-conditioning systems, have significant time-delay and time-varying characteristics, resulting in fault-related information being hidden in multi-dimensional and deep-level features, which puts higher requirements on data feature extraction. Through multi-dimensional spatio-temporal characteristic representation and cross-dimensional decision-making, the fault diagnosis of complex energy thermal systems can be effectively realized.

[0006] Fault diagnosis methods based on deep learning, as a decision-making method that comprehensively utilizes multi-dimensional deep information of data, have significantly improved the diagnosis accuracy compared with traditional means. Chinese invention patents that have been publicly disclosed (such as patent number: 202111114878.4, patent name: Adaptive Fault Diagnosis Method for Heat Pump Systems Based on Residual Data Scaling Strategy; patent number: 202210953909.3, patent name: An Intelligent Fault Diagnosis Method for Heat Pump Systems Based on Artificial Data Driving) record technical solutions for using convolutional neural networks for fault diagnosis of thermal systems. However, existing methods still have deficiencies, mainly reflected in the excessive dependence on the feature learning ability of deep learning models. With the increase in system complexity, deep learning models often have difficulty extracting sufficient rich and accurate feature information due to their own capacity limitations.

[0007] The time-frequency analysis algorithm enhances the accuracy of feature extraction by fusing the time domain and frequency domain information of the signal, and is often used in the field of mechanical fault diagnosis. The Chinese invention patents that have been published (such as patent number: 202210485072.4, patent name: A refrigeration equipment fault diagnosis method with multi-modal feature fusion neural network) record the technical solution of using both time domain and time-frequency domain modal information to diagnose screw refrigeration compressor faults and improve the diagnostic accuracy. However, time-frequency analysis has high requirements for the sampling frequency and duration of the signal. Generally, the higher the sampling frequency and the longer the sampling time, the higher the analysis accuracy. The patent uses the vibration signal of the compressor bearing as the analysis data. This signal is suitable for mechanical equipment fault diagnosis, and the vibration signal often has a very high sampling frequency. For the thermal system, it has its own thermal signals, such as temperature, pressure, etc. These signals change slowly, resulting in a low signal sampling frequency. Time-frequency processing is difficult to balance accuracy and real-time performance, making this method difficult to apply in thermal system fault diagnosis.

[0008] In view of the above difficulties, the present invention comprehensively uses cutting-edge deep learning models with spatiotemporal feature learning capabilities, such as spatiotemporal graph convolutional networks (STGCN) and graph convolutional networks (GCN), and combines the characteristics of slow signals (low sampling frequency, slow changes) of thermal systems to propose a slow signal time-frequency characterization method based on Gaussian regression interpolation encryption. By integrating multi-dimensional and multi-level spatiotemporal features, a general framework for end-to-end reasoning of complex thermal system faults is established, which significantly improves the accuracy and reliability of complex energy thermal system fault reasoning, helps the energy industry transform into an intelligent one, and has broad application prospects. Summary of the invention

[0009] The present invention is applicable to fault inference of energy thermal systems. It can infer possible faults based on real-time monitoring data through a deep fusion model, and quickly and accurately troubleshoot system problems. The information that needs to be provided to the fault inference model includes: sensor information such as temperature / pressure / flow at different locations of the thermal system, and correlation structure information between parameters. Based on the above information, a fault inference model is established and trained, and then the model features are fused and the fault type is inferred. The specific steps of the present invention are as follows:

[0010] Step S1: Slice the time series data collected by various sensors of the thermal system (temperature, pressure, flow, power, etc.), and encrypt the sliced data using the adaptive kernel function Gaussian regression interpolation method to obtain a length-filled low-frequency response time series data set D1 and a dense high-frequency response time series data set D2.

[0011] Step S2: Using the sparse time-frequency analysis method, independently perform time-frequency analysis on each sensor parameter in the dense D2 data set to obtain a time-frequency diagram of each sensor parameter.

[0012] Step S3: The time-frequency graph is represented as a low-dimensional manifold through a convolutional neural network (CNN) to obtain a low-dimensional representation vector of each parameter.

[0013] Step S4: Construct a knowledge graph of parameter association relationships based on the association and causal relationship between the sensing parameters of the thermal system.

[0014] Step S5: Construct a graph convolutional neural network (GCN). The structure of the input graph data is based on the knowledge graph structure constructed in step S4. The features of each node in the graph data are the low-dimensional representation vectors of the parameters in step S3. The output layer of GCN is a one-dimensional fully connected neural network layer.

[0015] Step S6: Construct a spatiotemporal graph convolutional neural network (STGCN). The structure of the input graph data is based on the knowledge graph structure constructed in step S4. The feature of each node in the graph data is the time series value of each parameter in the length-padded D1 data in step S1. The output layer of STGCN is a one-dimensional fully connected neural network layer.

[0016] Step S7: Concatenate the output layers of GCN and STGCN, use the Attention layer to adaptively learn the weights of each parameter, and then add the softmax classification layer to realize fault classification decision.

[0017] Step S8: Use the existing fault labeling data to perform model supervision training, use cross entropy to calculate the loss value, and obtain an end-to-end fault inference model.

[0018] Wherein, in the step S1, each sensor data should be normalized first.

[0019] Among them, in the step S1, the slicing algorithm performs time series cutting according to the principle of unified duration, that is, considering that different sensors may have different sampling frequencies, the duration corresponding to each time series slice is ensured to be consistent, and the number of data points of each time series slice is allowed to be different.

[0020] Among them, in step S1, the adaptive kernel function Gaussian regression interpolation method linearly combines multiple kernel functions, sets a learnable weight coefficient for each kernel function, and uses the jellyfish search algorithm to achieve a global and efficient search for the optimal weight coefficient. The objective function of the search adopts the maximized marginal log-likelihood function. It should be noted that in order to ensure the uniformity of the model training and implementation stages, the weight coefficient obtained by the optimization will no longer change in this system.

[0021] Among them, in the step S1, for the length-filled D1 data set, other time series slices are interpolated and supplemented to the same length according to the data sequence length of the longest time series slice; for the dense high-frequency response D2 data set, the slice data length is greatly increased so that the length of each slice data sequence is not less than 512, while ensuring that each slice has the same length.

[0022] Among them, in the step S3, the parameters of the CNN should be shared among the sensing parameters, that is, no matter how many sensing parameters the thermal system has, the same CNN is used for low-dimensional representation.

[0023] Among them, in the step S4, the establishment of the association relationship is mainly based on correlation analysis algorithms such as Pearson correlation analysis and Spearman correlation analysis; the establishment of the causal relationship mainly relies on expert experience, considering the parameter space connection relationship and the thermophysical association relationship.

[0024] Compared with the prior art, the remarkable beneficial effects of the technical solution of the present invention include:

[0025] 1) The present invention encrypts the thermodynamic slow signal through the adaptive kernel function Gaussian regression interpolation method, and alleviates the defect that it is difficult to perform time-frequency processing on the ultra-short sequence of the slow signal without destroying the signal characteristics, and expands the representation dimension of the thermodynamic signal.

[0026] 2) The present invention makes full use of the association relationship between parameters to construct graph structure data, and at the same time uses CNN-GCN and STGCN for "time-space-frequency" cross-modal fusion decision-making, greatly improving the feature inference ability of complex time-delay time-varying systems. BRIEF DESCRIPTION OF THE DRAWINGS

[0027] Figure 1 is the logic flow chart of the method of the present invention;

[0028] Figure 2 is the schematic diagram of the overall model architecture of the fault inference of the present invention;

[0029] Figure 3 is the time-frequency analysis flow chart of the slow signal of the present invention;

[0030] Figure 4 is the schematic diagram of the system structure principle in the implementation case;

[0031] Figure 5 is the parameter association knowledge graph. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0032] In order to make the purpose, technical solution and advantages of the present invention clearer, the present invention will be further described in detail below with reference to the accompanying drawings of the specification and the embodiments. It should be understood that the specific embodiments described herein are only used to explain the present invention and are not used to limit the present invention.

[0033] In contrast, the present invention encompasses any alternatives, modifications, equivalent methods, and schemes defined by the claims that are within the spirit and scope of the present invention. Further, in order to enable the public to better understand the present invention, in the following detailed description of the present invention, some specific details are described in detail. Those skilled in the art can fully understand the present invention without the description of these details.

[0034] Example: In this implementation case, the central air-conditioning system in the ASHRAE RP-1312 open-source dataset is used as the target system. The structural principle of this system is shown in Figure 4 , where represents a temperature sensor, represents a flow sensor, represents a humidity sensor, represents a pressure sensor. This dataset is mainly used to evaluate the fault detection and diagnosis (FDD) methods of the central air-conditioning system and contains the operation data of typical fault states of the central air-conditioning system.

[0035] In this case, 19 parameters are selected as sensor data for input into the model of this case, which are: HWC-VLV: Heating coil opening; CHWC-VLV: Cooling coil opening; EA-DMPR: Exhaust valve position; RA-DMPR: Return air valve opening; OA-DMPR: Fresh air valve opening; SF-WAT: Supply fan power; RF-WAT: Return fan power; SA-CFM: Supply air flow; RA-CFM: Return air flow; OA-CFM: Outdoor air flow; SA-TEMP: Supply air temperature; MA-TEMP: Mixed air temperature; RA-TEMP: Return air temperature; SA-SP: Supply air static pressure; SF-SPD: Supply fan speed; RF-SPD: Return fan speed; SA-HUMD: Supply air relative humidity; E_hcoil: Heating coil energy consumption; E_ccoil: Cooling coil energy consumption.

[0036] This case considers 1 normal operating state and 11 fault operating states, which are: F1: Exhaust damper stuck (fully open); F2: Exhaust damper stuck (fully closed); F3: Return fan speed fixed at 30%; F4: Return fan completely fails; F5: Fresh air damper stuck (fully closed); F6: Cooling coil valve stuck (fully closed); F7: Cooling coil valve stuck (fully open); F8: Cooling coil valve stuck (partially open - 15%); F9: Cooling coil valve stuck (partially open - 65%); F10: Heating coil valve leak (2.0 GPM); F11: Fresh air damper leak (55% opening).

[0037] For each fault type, 1400 groups of samples are collected at a sampling interval of 1 minute. Using this dataset with reference to Figure 1The logical flow of the inference method shown and Figure 2 the model architecture shown are as follows in specific implementation steps:

[0038] S1: Process the 11 types of fault data in the dataset through normalization, time slicing, data filling, and encryption, etc., to obtain the low-frequency response time series dataset D1 with filled length and the intensive high-frequency response time series dataset D2.

[0039] S11: For the data of 11 fault types, respectively select the first 1000 groups of samples as the training dataset, and the parameters of this dataset are the 19 parameters described above.

[0040] S12: Perform normalization processing on the collected data. First, calculate the mean and variance of all samples. For each sample Y, use the formula to calculate its normalized value y, and finally obtain the normalized overall sample set.

[0041] S13: Perform time slicing on the normalized data. Set the slicing duration to 30 minutes to obtain time series samples with a data length of 30.

[0042] S14: The adaptive kernel function Gaussian regression interpolation method uses a linear combination of multiple kernel functions, including radial basis function, linear kernel function, Mahalanobis kernel function, exponential kernel function, etc.:

[0043] ;

[0044] where k1, k2, k n are different kernel functions, and a1, a2, a n are weight coefficients.

[0045] In this embodiment, three kernel functions, namely basis function, linear kernel function, and Mahalanobis kernel function, are selected, and a1, a2, and a3 are taken as 2, 1, and 2 respectively.

[0046] Using the maximum marginal log-likelihood function as the evaluation algorithm for interpolation quality, and using the jellyfish search algorithm for global parameter search, the combined kernel function for the dataset can be finally obtained.

[0047] S15: For the case where the sampling intervals of all parameters are the same, there is no need to perform sample filling processing. Use the original dataset slice as D1, and the sampling interval of this dataset is 1 minute.

[0048] S16: Continue to use the adaptive kernel function Gaussian regression interpolation algorithm to encrypt the time series data, increase the number of data points of each time slice to 600 points to make it adapt to the length required by the time-frequency processing algorithm, and obtain the intensive high-frequency response time series dataset D2, and the data interval of this dataset is 3 seconds.

[0049] Step S2: Use the sparse time-frequency analysis method to perform time-frequency analysis on each parameter of the D2 dataset to obtain the time-frequency diagrams of each sensing parameter. The slow signal time-frequency analysis process is as Figure 3 shown.

[0050] Among them, in this embodiment, the specific sparse time-frequency analysis method selects the sparse generalized S transform, constructs a sparse dictionary using the Gabor basis, and obtains the time-frequency characterization diagram using the S transform.

[0051] Step S3: Compress the time-frequency diagrams to obtain pictures of a unified size (such as 128×128 pixels). Construct a convolutional neural network (CNN) with an input size of 128×128, 3 layers can be set, the number of convolutional kernels is selected as 64, 128, 64, the activation function is selected as gelu, and the output is a 32-dimensional vector as the low-dimensional manifold representation of the parameters. This CNN is a shared model, that is, the low-dimensional manifold representations of all parameters share this model.

[0052] Step S4: Construct the topological structure of the association graph. The graph nodes represent different sensing parameters, and the edges represent the association relationships between different parameters, and 0 and 1 are used to indicate whether there is an association between two nodes.

[0053] Among them, the establishment of this spectral graph association should consider both the spatial position association between sensors and the thermodynamic association. First, use the Pearson correlation coefficient to calculate the correlation between parameters, establish connections for coefficients greater than 0.85, and supplement some physically associated connections according to experience. The knowledge graph constructed in this embodiment is as Figure 5 shown.

[0054] Step S5: Construct a graph convolutional neural network (GCN) with the input graph data topology being the spectral graph topology constructed in S4 and the graph node features being the low-dimensional representation vectors of each parameter in S3.

[0055] Among them, the graph convolutional layer of the GCN model is 2 layers, the feature dimension is set to 64, followed by a convolutional layer and a flattening layer, and finally a fully connected layer with a length of 64, and the activation function is selected as gelu.

[0056] Step S6: Construct a spatio-temporal graph convolutional neural network (STGCN) with the input graph data topology being the spectral graph topology constructed in S4 and the graph node features being the D1 data with its length filled in S1.

[0057] Among them, the spatio-temporal graph convolutional layer of the STGCN is 2 layers (each containing a spatial graph convolutional layer and a temporal convolutional layer), the feature dimension is set to 64, followed by a convolutional layer and a flattening layer, and finally a fully connected layer with a length of 64, and the activation function is selected as gelu.

[0058] Step S7: Concatenate the 64 fully connected layers output from S5 and S6 into a 128 fully connected layer as the cross-scale feature fusion layer, and finally perform classification through the Softmax layer.

[0059] Step S8: Use the existing fault annotation data for model supervised training. Calculate the prediction error using the cross entropy loss (CrossEntropyLoss), and use the Adam optimizer for training with a learning rate set to 0.001. Verify the accuracy on 300 groups of samples that do not participate in the training. To evaluate the performance of the diagnostic method proposed in the present invention, a support vector machine (SVM), a multi-layer perceptron (MLP), and a convolutional neural network (CNN) are selected for comparison. The comparison of the average accuracy with various comparison methods is shown in Table 1. It can be seen from the results that compared with these methods, the method (prod model) proposed in the present invention has achieved better results in the diagnostic accuracy of various faults, outperforming all other methods. For some fault types that are prone to misdiagnosis, such as F7: Cooling coil valve stuck (fully open); F8: Cooling coil valve stuck (partially open - 15%), etc., good results have also been achieved.

[0060] Table 1 Comparison of diagnostic effects of different diagnostic methods

[0061] 。

[0062] The above are only the preferred embodiments of the present invention and are not intended to limit the present invention. Any modifications, equivalent replacements, and improvements made within the spirit and principles of the present invention shall be included in the protection scope of the present invention.

Claims

1. An energy thermal system graph embedding fault inference method that fuses the time-frequency representation of slow signals, characterized in that, Based on the sensor parameter information at different locations of the thermal system and the correlation structure information between the parameters, a fault inference model is established and trained, the model features are integrated and the fault type is inferred; specifically, the following steps are included: Step S1: Slice the time series data collected by each sensor of the thermal system, and encrypt the sliced data using the adaptive kernel function Gaussian regression interpolation method to obtain a length-filled low-frequency response time series data set D1 and a dense high-frequency response time series data set D2; Step S2: using the sparse time-frequency analysis method, perform time-frequency analysis on each sensor parameter in the data set D2 to obtain a time-frequency diagram of each sensor parameter; Step S3: The time-frequency graph is represented in a low-dimensional manifold through a convolutional neural network (CNN) to obtain a low-dimensional representation vector of each parameter; Step S4: construct a knowledge graph of parameter association relationships based on the association relationships and causal relationships among the sensing parameters of the thermal system; Step S5: Construct a graph convolutional neural network (GCN) and a spatiotemporal graph convolutional neural network (STGCN), concatenate the output layers of GCN and STGCN, use the Attention layer to adaptively learn the weights of each parameter, and add a softmax classification layer to achieve fault classification decision; In the graph convolutional neural network GCN, the structure of the input graph data is in accordance with the knowledge graph structure constructed in step S4, the feature of each node of the graph data is a low-dimensional representation vector of each parameter in step S3, and the output layer of the graph convolutional neural network GCN is a one-dimensional fully connected neural network layer; In the spatiotemporal graph convolutional neural network STGCN, the structure of the input graph data is in accordance with the knowledge graph structure constructed in step S4, the feature of each node of the graph data is the time series value of each parameter in the length-padded D1 data in step S1, and the output layer of the spatiotemporal graph convolutional neural network STGCN is a one-dimensional fully connected neural network layer; Step S6: Use the existing fault annotation data to perform model supervision training, use cross entropy to calculate the loss value, and obtain an end-to-end fault inference model.

2. The method for fault inference by embedding an energy thermal system diagram integrating slow signal time-frequency characterization according to claim 1, characterized in that In the step S1, the slicing process performs time series cutting based on the principle of unified duration, ensuring that the duration corresponding to each time series slice is consistent, and allowing the number of data points of each time series slice to be different.

3. The method for fault reasoning by embedding an energy thermal system diagram with fused slow signal time-frequency characterization according to claim 1, wherein In step S1, a plurality of kernel functions are linearly combined by using an adaptive kernel function Gaussian regression interpolation method, and a learnable weight coefficient is set for each kernel function. The optimal weight coefficient is searched using a jellyfish search algorithm, and the objective function of the search is to maximize the marginal log-likelihood function; and the weight coefficient obtained by the optimization is a constant.

4. The energy thermal system diagram embedding fault inference method for fusing slow signal time-frequency characterization according to claim 1, characterized in that In step S1, for the data set D1, other time series slices are interpolated and supplemented to the same length according to the data sequence length of the longest time series slice; for the dense high-frequency response time series data set D2, the slice data length is increased so that the length of each slice data sequence is not less than 512, while ensuring that each slice has the same length.

5. The method for fault reasoning by embedding an energy thermal system diagram with fused slow signal time-frequency characteristics according to claim 1, characterized in that In step S3, the parameters of the convolutional neural network CNN are shared among the sensing parameters, that is, all parameters in the thermal system are represented in a low-dimensional manner using the same convolutional neural network CNN.

6. The method for inferring faults in the embedding of an energy thermal system diagram that fuses the time-frequency characterization of slow signals according to claim 1, characterized in that In the step S4, the establishment of the association relationship is implemented based on a correlation analysis algorithm; the establishment of the causal relationship is based on expert experience and combines the parameter space connection relationship and the thermophysical association relationship.

Images