Intelligent temperature detection method and Internet of Things device
The method addresses the challenges of complex wiring and low accuracy in traditional temperature detection by employing neural networks for data decomposition and prediction, enhancing remote monitoring accuracy in industrial settings.
Patent Information
- Application Number
- CN202311761663.0
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2023-12-20
- Publication Date
- 2024-12-20
AI Technical Summary
Traditional temperature detection methods in industrial and agricultural settings suffer from complex on-site wiring, inability to combine on-site and remote monitoring, and low accuracy due to strong coupling, interference, and nonlinear uncertainty in multi-point temperature measurements.
A temperature intelligent detection method utilizing empirical mode decomposition, TCN-BiLSTM neural networks, GCN-BiLSTM neural networks, and NARX-HRFNN recursive fuzzy neural networks to predict temperature, humidity, and wind speed data, with a parameter monitoring Internet of Things device for remote data transmission.
The method achieves high accuracy in temperature prediction by decomposing data into frequency bands, reducing interference, and using bidirectional LSTM networks for enhanced learning, resulting in improved prediction and remote monitoring capabilities.
Smart Images
Figure CN119168103A_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the technical field of automation equipment for intelligent temperature detection and parameter monitoring Internet of Things devices, and in particular to an intelligent temperature detection method and an Internet of Things device. Background Art
[0002] In many occasions of industrial and agricultural production, it is necessary to realize intelligent multi-point temperature intelligent detection and parameter monitoring. The traditional temperature detection method has the problems of complex on-site wiring, inability to combine on-site and remote monitoring, and low accuracy of multi-point temperature measurement. It does not take into account the low accuracy of multi-point temperature detection due to strong coupling, strong interference, large lag, and nonlinear uncertainty among the multi-point temperatures of the detected objects. Therefore, it is necessary to design an Internet of Things device for remote multi-point temperature detection and its parameter monitoring, which realizes multi-point temperature detection, parameter monitoring and remote data transmission functions.
[0003] When performing multi-point temperature detection and parameter monitoring, several parameter acquisition terminals can be deployed. After each parameter terminal collects data, it is sent to the central node through a wireless module, and the central node receives the data from the parameter acquisition terminal. Since the LoRa module has the advantages of low power consumption, long distance and high reliability, it is widely used. The LoRa module is used for multi-point parameter acquisition. The LoRa module of the central node and the LoRa module of each parameter acquisition terminal can adopt a transparent transmission mode. When the LoRa module of the central node receives the single-chip microcomputer, the single-chip microcomputer connects to the Alibaba Cloud Internet of Things platform through the WiFi module to realize remote access to data. Users can realize intelligent temperature detection methods and remote monitoring parameter acquisition and other early warning and other monitoring functions on the remote monitoring end and mobile phone APP end. A temperature intelligent detection method and parameter monitoring Internet of Things device realize multi-point parameter acquisition, intelligent temperature detection method and remote data transmission. The data transmission is complete and timely, and has good practicality. Summary of the invention
[0004] Purpose of the invention: In response to the problems pointed out in the background technology, the present invention discloses a temperature intelligent detection method and an Internet of Things device, which better predicts the temperature, humidity and wind speed detection data through empirical mode decomposition through multiple TCN-BiLSTM neural network models, and calibrates the prediction results of each component through the GCN-BiLSTM neural network model to predict the temperature value of the detected object, with high temperature prediction accuracy.
[0005] Technical solution: The present invention discloses a temperature intelligent detection method, comprising the following steps:
[0006] Step 1: Construct a temperature detection module. The temperature detection module includes a temperature empirical mode decomposition model, a humidity empirical mode decomposition model, a wind speed empirical mode decomposition model, TCN-BiLSTM neural network model 1 to TCN-BiLSTM neural network model n, GCN-BiLSTM neural network model 1 to GCN-BiLSTM neural network model 3 and a NARX neural network model-HRFNN recursive fuzzy neural network model. The structure and function of the temperature detection module are shown in Figure 1 shown.
[0007] The outputs of the temperature, humidity, and wind speed sensors are passed through the temperature empirical mode decomposition model, humidity empirical mode decomposition model, and wind speed empirical mode decomposition model to obtain decomposition signals. Each decomposition signal is passed through the corresponding TCN-BiLSTM neural network model and then input into the corresponding GCN-BiLSTM neural network model, and then calibrated and output by the NARX neural network model-HRFNN recursive fuzzy neural network model.
[0008] Step 2: Construct a temperature detection system, which includes a fuzzy C-means clusterer, the temperature detection module, a GCN-BiLSTM neural network model 4, an Elman neural network model, a TCN-BiLSTM neural network model A, a TCN-BiLSTM neural network model B, a TCN-BiLSTM neural network model C, a NARX neural network model-PID controller and a NARX neural network model-BiLSTM neural network model; the structure and function of the temperature detection system are shown in Figure 2 shown.
[0009] The parameter sensor group includes a temperature sensor, a humidity sensor and a wind speed sensor. The data collected by the multiple parameter sensor groups are used as the input of the fuzzy C-means clusterer. The multiple groups of parameter sensor group data output by the fuzzy C-means clusterer are used as the input of the corresponding temperature detection modules. The outputs of the multiple temperature detection modules are used as the corresponding inputs of the GCN-BiLSTM neural network model 4. The outputs of the GCN-BiLSTM neural network model 4 are used as the corresponding inputs of the Elman neural network model and the corresponding inputs of the NARX neural network model-BiLSTM neural network model. The output of the Elman neural network model is used as the input of the TCN-BiLSTM neural network model A and the corresponding input of the NARX neural network model-BiLSTM neural network model. The error and error variation between the output of the Elman neural network model and the output of the NARX neural network model-BiLSTM neural network model The quantization rate is used as the input of TCN-BiLSTM neural network model B, the input of NARX neural network model-PID controller and the corresponding input of NARX neural network model-BiLSTM neural network model respectively. The output of NARX neural network model-PID controller is used as the input of TCN-BiLSTM neural network model C and the corresponding input of NARX neural network model-BiLSTM neural network model respectively. The output of NARX neural network model-BiLSTM neural network model is used as the corresponding input of Elman neural network model. The outputs of TCN-BiLSTM neural network model A, TCN-BiLSTM neural network model B and TCN-BiLSTM neural network model C are used as the corresponding input of GCN-BiLSTM neural network model 4 respectively. The NARX neural network model-BiLSTM neural network model outputs the predicted temperature value of the object.
[0010] Furthermore, the structure of the temperature detection module is as follows:
[0011] The temperature sensor output is used as the input of the temperature empirical mode decomposition model. The low-frequency trend part and multiple high-frequency fluctuation parts of the temperature empirical mode decomposition model output are used as the input of the corresponding TCN-BiLSTM neural network model. The multiple TCN-BiLSTM neural network models are used as the input of the GCN-BiLSTM neural network model 1.
[0012] The humidity sensor output is used as the input of the humidity empirical mode decomposition model. The low-frequency trend part and multiple high-frequency fluctuation parts of the humidity empirical mode decomposition model output are used as the corresponding TCN-BiLSTM neural network model inputs respectively. Multiple TCN-BiLSTM neural network models are used as the input of GCN-BiLSTM neural network model 2 respectively.
[0013] The wind speed sensor output is used as the input of the wind speed empirical mode decomposition model. The low-frequency trend part and multiple high-frequency fluctuation parts of the wind speed empirical mode decomposition model output are respectively used as the corresponding TCN-BiLSTM neural network model inputs. Multiple TCN-BiLSTM neural network models are respectively used as the input of the GCN-BiLSTM neural network model 3.
[0014] GCN-BiLSTM neural network model 1, GCN-BiLSTM neural network model 2 and GCN-BiLSTM neural network model 3 are respectively used as the corresponding inputs of the NARX neural network model-HRFNN recursive fuzzy neural network model, and the outputs of the NARX neural network model-HRFNN recursive fuzzy neural network model are respectively used as the corresponding inputs of the GCN-BiLSTM neural network model 1, GCN-BiLSTM neural network model 2 and GCN-BiLSTM neural network model 3, and the NARX neural network model-HRFNN recursive fuzzy neural network model outputs the temperature value of the detected point as the output of the temperature detection module.
[0015] Furthermore, the temperature empirical mode decomposition model, the humidity empirical mode decomposition model and the wind speed empirical mode decomposition model will filter out trends of different characteristics existing in the temperature, humidity and wind speed historical data information step by step, and decompose the temperature, humidity and wind speed historical data information into a low-frequency trend part and multiple high-frequency fluctuation parts.
[0016] Furthermore, the TCN-BiLSTM neural network model is a TCN time series convolutional neural network model and a BiLSTM neural network model connected in series. The TCN time series neural network model learns the time series characteristics and spatial characteristics of temperature, humidity, and wind speed data, and the BiLSTM neural network model deepens the learning of the time series characteristics and spatial characteristics of temperature, humidity, and wind speed data.
[0017] Furthermore, the GCN-BiLSTM neural network model is a GCN graph neural network model connected in series with a BiLSTM neural network model. The GCN graph neural network model extracts feature data of output information of multiple TCN-BiLSTM neural network models, performs a series of convolution operations on the extracted data, and for a single node, extracts information from neighbors connected to the node, and aggregates node feature information with neighbor feature information using an aggregation function to obtain new information.
[0018] Further, the NARX neural network model-HRFNN recursive fuzzy neural network model is a NARX neural network model and an HRFNN recursive fuzzy neural network model connected in series;
[0019] The NARX neural network model is a dynamic network composed of interconnected nodes. It accepts one or more inputs and obtains the optimal weights between neurons and input parameters. The NARX neural network model adds a delay and feedback mechanism to store the historical data of input factors, and delays the output vector of the network and introduces it into the input vector through external feedback.
[0020] The HRFNN recursive fuzzy neural network model uses a gradient descent algorithm with cross-validation to train the weights of the neural network. The HRFNN recursive fuzzy neural network model introduces internal variables in the feedback link, weights the output of the rule layer, and then defuzzifies the output as the feedback, and uses the feedback together with the output of the membership function layer as the input of the rule layer at the next moment.
[0021] Furthermore, the NARX neural network model-PID controller uses the three values output by the NARX neural network model as Kp, Ki and Kd of the PID controller, thereby realizing dynamic adjustment of the proportional factor, integral factor and differential factor of the PID by the NARX neural network model.
[0022] The present invention also discloses a parameter monitoring Internet of Things device, which is composed of a parameter acquisition terminal, a parameter control terminal, a central node, a remote monitoring terminal, a cloud platform and a mobile phone APP. The parameter acquisition terminal, the parameter control terminal and the central node communicate through a self-organizing network through a LoRa communication module. The communication between the parameter acquisition terminal and the parameter control terminal and the remote monitoring terminal, the cloud platform and the mobile phone APP is realized through the central node. The remote monitoring terminal and the cloud platform perform big data monitoring, storage and processing. The mobile phone APP performs two-way communication with the central node through the cloud platform. The remote monitoring terminal and the mobile phone APP enable the management personnel to remotely monitor the operating status of the parameter acquisition terminal and the parameter control terminal. The computer program steps of the above-mentioned intelligent temperature detection method are deployed in the parameter acquisition terminal and the parameter control terminal. The structure and function of a parameter monitoring Internet of Things device are shown in FIG. Figure 3 shown.
[0023] Compared with the prior art, the present invention has the following advantages:
[0024] 1. The present invention decomposes the temperature, humidity and wind speed historical data sequence into components of different frequency bands through the empirical mode decomposition model. Each component shows different characteristic information implicit in the original sequence to reduce the non-stationarity of the temperature, humidity and wind speed sequence. The high-frequency part of the data is not strongly correlated and has a relatively high frequency. It represents the fluctuation component of the original temperature, humidity and wind speed sequence and has a certain periodicity and randomness, which is consistent with the periodic changes of temperature, humidity and wind speed; the low-frequency component represents the change trend of the original sequence of temperature, humidity and wind speed. It can be seen that EMD can decompose the fluctuation component, periodic component and trend component of the temperature, humidity and wind speed changes step by step. Each decomposed component contains the same deformation information, which reduces the mutual interference between different characteristic information to a certain extent, and the decomposed component change curve is smoother than the original temperature, humidity and wind speed deformation sequence curve. It can be seen that the empirical decomposition model can effectively analyze the deformation data of temperature, humidity and wind speed under the joint action of multiple factors. The decomposed components are conducive to better prediction of multiple TCN-BiLSTM neural network models. In order to avoid the problems of random selection of input dimensions of extreme learning machines and loss of component information, the phase space of each component is reconstructed first, and finally the prediction results of each component are calibrated through the GCN-BiLSTM neural network model to obtain the temperature of the detected object. The case study shows that the proposed temperature calibration prediction result of the detected object has a high prediction accuracy.
[0025] 2. The BiLSTM neural network model is composed of the forward and backward propagation directions of the LSTM neural network. The forward propagation layer unit learns the past information of the current sequence, and the backward propagation unit learns the future information of the current sequence. The forward LSTM and the backward LSTM are combined into the BiLSTM neural network model. The BiLSTM neural network model can effectively learn the front and back information between the temperature time series features and spatial features, and use the forward and reverse direction information of the current position for bidirectional learning. The TCN time series convolutional neural network model and the BiLSTM neural network model are connected in series to form a TCN-BiLSTM neural network model. The TCN time series neural network model learns the time series features and spatial features of the temperature data.
[0026] 3. The TCN temporal convolutional neural network model integrates causal convolution and dilated convolution on the basis of the traditional one-dimensional convolutional neural network, and adds residual connection on its basis. The improved TCN temporal convolutional neural network model has three structural characteristics - causal convolution, dilated convolution and residual connection, which makes it have a flexible receptive field. It can obtain temperature time series of any length by changing the dilation factor and map it to the output sequence of the same length. It will not forget or leak future temperature data information, and there will be no gradient explosion or gradient disappearance due to the deepening of the network. Therefore, TCN maintains a strong memory capacity while maintaining the stability of the entire network structure.
[0027] Fourth, the NARX neural network model adds a delay and feedback mechanism, which can store the historical data of the input factors, delay and keep the output vector of the network, and then introduce it into the input vector through external feedback, which greatly enhances the memory ability of historical data. It is a dynamic neural network.
[0028] 5. The GCN graph neural network model extracts feature data from the output information of multiple TCN-BiLSTM neural network models, performs a series of convolution operations on the extracted data, extracts information from neighbors connected to the node for a single node, and aggregates node feature information with neighbor feature information using aggregation functions to obtain new information. The GCN graph neural network model includes spectral domain and spatial domain methods, which can cleverly extract some data such as nodes and edges in the input information. BRIEF DESCRIPTION OF THE DRAWINGS
[0029] Figure 1 The temperature detection module of the present invention;
[0030] Figure 2 The temperature detection system of the present invention;
[0031] Figure 3 The parameter monitoring Internet of Things device of the present invention;
[0032] Figure 4 It is a parameter collection terminal of the present invention;
[0033] Figure 5 It is a parameter control terminal of the present invention;
[0034] Figure 6 It is the central node of the present invention;
[0035] Figure 7 This is the remote monitoring software function of the present invention. DETAILED DESCRIPTION
[0036] In order to better explain the present invention, the following Figure 1-7The technical solution of the present invention is described in detail. The following embodiments are explanations of the invention, and the present invention is not limited to the following embodiments.
[0037] The present invention discloses a temperature intelligent detection method, comprising the following steps:
[0038] Step 1: Construct a temperature detection module, which includes a temperature empirical mode decomposition model, a humidity empirical mode decomposition model, a wind speed empirical mode decomposition model, TCN-BiLSTM neural network model 1 to TCN-BiLSTM neural network model n, GCN-BiLSTM neural network model 1 to GCN-BiLSTM neural network model 3 and a NARX neural network model-HRFNN recursive fuzzy neural network model.
[0039] The outputs of the temperature, humidity, and wind speed sensors are passed through the temperature empirical mode decomposition model, the humidity empirical mode decomposition model, and the wind speed empirical mode decomposition model to obtain decomposition signals. Each decomposition signal is passed through the corresponding TCN-BiLSTM neural network model and then input into the corresponding GCN-BiLSTM neural network model, and then calibrated and output by the NARX neural network model-HRFNN recursive fuzzy neural network model.
[0040] The temperature sensor output is used as the input of the temperature empirical mode decomposition model. The low-frequency trend part and multiple high-frequency fluctuation parts of the temperature empirical mode decomposition model output are used as the input of their corresponding TCN-BiLSTM neural network models respectively. Multiple TCN-BiLSTM neural network models are used as the input of GCN-BiLSTM neural network model 1 respectively.
[0041] The humidity sensor output is used as the input of the humidity empirical mode decomposition model. The low-frequency trend part and multiple high-frequency fluctuation parts of the humidity empirical mode decomposition model output are respectively used as the corresponding TCN-BiLSTM neural network model inputs. Multiple TCN-BiLSTM neural network models are respectively used as the input of GCN-BiLSTM neural network model 2.
[0042] The wind speed sensor output is used as the input of the wind speed empirical mode decomposition model. The low-frequency trend part and multiple high-frequency fluctuation parts of the wind speed empirical mode decomposition model output are respectively used as the corresponding TCN-BiLSTM neural network model inputs. Multiple TCN-BiLSTM neural network models are respectively used as the input of GCN-BiLSTM neural network model 3.
[0043] GCN-BiLSTM neural network model 1, GCN-BiLSTM neural network model 2 and GCN-BiLSTM neural network model 3 are respectively used as the corresponding inputs of the NARX neural network model-HRFNN recursive fuzzy neural network model, and the NARX neural network model-HRFNN recursive fuzzy neural network model outputs the temperature value of the detected point as the output of the temperature detection module.
[0044] Step 2: Build a temperature detection system, which includes a fuzzy C-means clusterer, n temperature detection modules, a GCN-BiLSTM neural network model 4, an Elman neural network model, a TCN-BiLSTM neural network model A, a TCN-BiLSTM neural network model B, a TCN-BiLSTM neural network model C, a NARX neural network model-PID controller and a NARX neural network model-BiLSTM neural network model.
[0045] The parameter sensor group includes a temperature sensor, a humidity sensor and a wind speed sensor. The data collected by the multiple parameter sensor groups are used as the input of the fuzzy C-means clusterer. The multiple groups of parameter sensor group data output by the fuzzy C-means clusterer are used as the input of the corresponding n temperature detection modules respectively. The outputs of the multiple temperature detection modules are used as the corresponding inputs of the GCN-BiLSTM neural network model 4. The outputs of the GCN-BiLSTM neural network model 4 are used as the corresponding inputs of the Elman neural network model and the corresponding inputs of the NARX neural network model-BiLSTM neural network model respectively. The output of the Elman neural network model is used as the input of the TCN-BiLSTM neural network model A and the corresponding input of the NARX neural network model-BiLSTM neural network model respectively. The error and error between the output of the Elman neural network model and the output of the NARX neural network model-BiLSTM neural network model are calculated. The change rate is used as the input of TCN-BiLSTM neural network model B, the input of NARX neural network model-PID controller and the corresponding input of NARX neural network model-BiLSTM neural network model respectively. The output of NARX neural network model-PID controller is used as the input of TCN-BiLSTM neural network model C and the corresponding input of NARX neural network model-BiLSTM neural network model respectively. The output of NARX neural network model-BiLSTM neural network model is used as the corresponding input of Elman neural network model. The outputs of TCN-BiLSTM neural network model A, TCN-BiLSTM neural network model B and TCN-BiLSTM neural network model C are used as the corresponding input of GCN-BiLSTM neural network model 4 respectively. The NARX neural network model-BiLSTM neural network model outputs the predicted temperature value of the object.
[0046] 1. Design of the above temperature detection module
[0047] 1. Temperature empirical mode decomposition model design
[0048] Temperature empirical mode decomposition (EMD) is an adaptive signal screening method with the characteristics of simple calculation, intuitiveness, experience-based and adaptive. It can filter out the trends of different characteristics existing in the temperature historical data information step by step to obtain multiple high-frequency fluctuation parts (IMF) and low-frequency trend parts. The IMF components decomposed by temperature empirical mode decomposition contain components of different frequency bands from high to low. The frequency resolution contained in each frequency band varies with the information itself, and has the characteristics of adaptive multi-resolution analysis. The purpose of using temperature empirical mode decomposition is to extract temperature historical data information more accurately. The steps of the "screening" process of temperature empirical mode decomposition method for temperature historical data are as follows:
[0049] ①. Determine all the local extreme points of the temperature history data information, and then use cubic splines to connect the local maximum points on the left and right to form an upper envelope.
[0050] ②. Connect the local minimum points of the temperature history data information with cubic spline to form the lower envelope line. The upper and lower envelope lines should envelop all the data points.
[0051] ③. The average value of the upper and lower envelopes is recorded as m1(t), and the solution is:
[0052] x(t)-m1(t)=h1(t) (1)
[0053] x(t) is the original signal of the temperature history data information. If h1(t) is an IMF, then h1(t) is the first IMF component of x(t). Let c1(t) = h 1k (t), then c1(t) is the first component of the signal x(t) that satisfies the IMF condition.
[0054] ④. Separate c1(t) from x(t) and get:
[0055] r1(t)=x(t)-c1(t) (2)
[0056] Repeat steps ① to ③ with r1(t) as the original data to obtain the second component c2 of x(t) that meets the IMF condition. Repeat the cycle n times to obtain n components of the signal x(t) that meet the IMF condition. In this way, the temperature empirical mode decomposition model decomposes the temperature historical data information into a low-frequency trend part and multiple high-frequency fluctuation parts. The design method of the temperature empirical mode decomposition model is similar to that of the humidity empirical mode decomposition model and the wind speed empirical mode decomposition model.
[0057] 2. TCN-BiLSTM neural network model design
[0058] The TCN-BiLSTM neural network model is a series connection of the TCN temporal convolutional neural network model and the BiLSTM neural network model. The TCN temporal convolutional neural network model integrates causal convolution and dilated convolution on the basis of the traditional one-dimensional convolutional neural network, and adds residual connection on its basis. The three structural characteristics of the TCN temporal convolutional neural network model - causal convolution, dilated convolution and residual connection - give it a flexible receptive field. It can obtain temperature time series of any length by changing the dilation factor and map it to an output sequence of the same length. It will not forget or leak future temperature data information, and will not cause gradient explosion or gradient disappearance due to the deepening of the network. Therefore, TCN maintains a strong memory capacity while maintaining the stability of the entire network structure. The use of causal volumes can retain more original temperature time series feature information while learning the causal relationship between temperature time series change features and spatial features; by expanding the convolution to increase the receptive field of the convolution, a convolution can learn more features; the introduction of the residual module can deepen the number of TCN layers, and the convolution can more fully learn the implicit features of the temperature time series, and is less likely to have gradient problems. The performance of the TCN time series convolutional neural network model can be better improved.
[0059] Based on the above advantages, the TCN time series convolutional neural network model is very suitable for learning temperature time series features, with multiple temperature change features and coexistence of time series and space features. At the same time, in order to strengthen the learning of temperature time series features, the BiLSTM neural network model is introduced for learning. The LSTM neural network model consists of several LSTM units. The temperature time series feature sequence processed by the TCN time series convolutional neural network model is input into the LSTM neural network model, and finally expanded into the required dimensions by the fully connected layer. The BiLSTM neural network model is composed of the combination of the front and back propagation directions of the LSTM neural network. The forward propagation layer unit learns the past information of the current sequence, and the back propagation unit learns the future information of the current sequence. The forward LSTM and the backward LSTM are combined into a BiLSTM neural network model. The BiLSTM neural network model can effectively learn the forward and backward information between the temperature time series features and the spatial features, and use the forward and reverse direction information of the current position for bidirectional learning. The TCN time series convolutional neural network model and the BiLSTM neural network model are connected in series to form a TCN-BiLSTM neural network model. The TCN time series neural network model learns the time series features and spatial features of the temperature data, and the BiLSTM neural network model deepens the learning of the time series features and spatial features of the temperature data.
[0060] 3. GCN-BiLSTM neural network model design
[0061] The GCN-BiLSTM neural network model is a series connection of the GCN graph neural network model and the BiLSTM neural network model. The GCN graph neural network model extracts feature data of the output information of multiple TCN-BiLSTM neural network models, performs a series of convolution operations on the extracted data, and extracts information from neighbors connected to the node for a single node. The node feature information is aggregated with the neighbor feature information using the aggregation function to obtain new information. The GCN graph neural network model includes spectral domain and spatial domain methods, which can cleverly extract some data such as nodes and edges in the input information.
[0062] The GCN graph neural network model consists of two graph convolution layers, one pooling layer and one fully connected layer. The first two layers are feature learning layers, and the last two layers compress the data. The two feature learning layers perform graph convolution based on the node features in each graph to obtain the output node vector. By calling the graph message passing mechanism, the local information around each node is iteratively fused to obtain the node information of the current layer. All the output node vectors of each graph are then summed up and passed to the maximum pooling layer and the linear layer to obtain the graph structure table, thereby improving the robustness of the features.
[0063] 4. NARX neural network model-HRFNN recursive fuzzy neural network model design
[0064] NARX neural network model-HRFNN recursive fuzzy neural network model is a series connection of NARX neural network model and HRFNN recursive fuzzy neural network model. NARX neural network model is a dynamic network composed of interconnected nodes. It can accept one or more inputs to obtain the optimal weights of the connection between neurons and input parameters. Compared with the traditional static BP neural network, NARX neural network model adds delay and feedback mechanism. It can store the historical data of input factors, and delay and keep the output vector of the network, and then introduce it into the input vector through external feedback, which greatly enhances the memory ability of historical data. It is a dynamic neural network. The neural network includes input layer, input layer, input parameters of neural network, hidden layer, activation function of hidden layer is tansig (hyperbolic tangent sigmoid transfer function), output layer, activation function of output layer is purelin (linear transfer function), total 3 layers. The characteristic of NARX neural network model is that the result of output layer will be fed back to input layer at the same time. HRFNN recursive fuzzy neural network model is a multi-input single-output network topology structure. The network consists of 4 layers: input layer, member function layer, rule layer and output layer. The network contains n input nodes, each of which corresponds to m conditional nodes, m represents the number of rules, nm rule nodes, and 1 output node. The first layer introduces the input into the network; the second layer fuzzifies the input, and the membership function used is the Gaussian function; the third layer corresponds to fuzzy reasoning; and the fourth layer corresponds to defuzzification operations. Represent the input and output of the ith node of the kth layer respectively. The signal transmission process within the network and the input-output relationship between each layer can be described as follows. Layer I: Input layer. Each input node of this layer is directly connected to the input variable. The input and output of the network are expressed as:
[0065]
[0066] In the formula and is the input and output of the ith node in the network input layer, and N represents the number of iterations. Layer II: Membership function layer. The nodes in this layer fuzzify the input variables. Each node represents a membership function. The Gaussian basis function is used as the membership function. The input and output of the network are expressed as:
[0067]
[0068] Where mij and σ ij They represent the mean center and width value of the j-th Gaussian basis function of the i-th language variable in the second layer, respectively. m is the total number of language variables of the corresponding input node. The third layer: fuzzy reasoning layer, that is, the rule layer. Dynamic feedback is added to make the network have better learning efficiency. The internal variable h is introduced in the feedback link.k , the sigmoid function is selected as the activation function of the internal variables of the feedback link. The input and output of the network are expressed as:
[0069]
[0070] Where ω jk is the connection weight of the recursive part. The neurons in this layer represent the antecedent part of the fuzzy logic rule. The nodes in this layer perform π operation on the output of the second layer and the feedback of the third layer. is the output of the third layer, and m represents the number of rules when fully connected. The feedback link mainly calculates the value of the internal variable and the activation strength of the corresponding membership function of the internal variable. The activation strength is related to the matching degree of the rule nodes of the third layer. The internal variables introduced in the feedback link include two types of nodes: receiving nodes and feedback nodes. The receiving node uses weighted summation to calculate the internal variables to achieve the defuzzification function; the internal variables represent the results of fuzzy reasoning of the hidden rules. The feedback node uses the sigmoid function as the fuzzy membership function to realize the fuzzification of the internal variables. Layer IV: Defuzzification layer, that is, the output layer. The nodes in this layer perform a summation operation on the input. The input and output of the network are expressed as:
[0071]
[0072] In the formula, j is the connection weight of the output layer. The HRFNN recursive fuzzy neural network model has the performance of approximating highly nonlinear dynamic systems. The training error and test error of the recursive neural network with the addition of internal variables are significantly reduced. The HRFNN recursive fuzzy neural network model of this patent uses a gradient descent algorithm with cross-validation to train the weights of the neural network. The HRFNN recursive fuzzy neural network model introduces internal variables in the feedback link, takes the weighted sum of the output of the rule layer and then defuzzifies the output as the feedback, and uses the feedback together with the output of the membership function layer as the input of the rule layer at the next moment. The network output contains the historical information of the activation strength and output of the rule layer, which enhances the ability of the HRFNN recursive fuzzy neural network prediction model to adapt to nonlinear dynamic systems.
[0073] 2. Design of the above temperature detection system
[0074] The temperature detection system includes a fuzzy C-means clusterer, n temperature detection modules, a GCN-BiLSTM neural network model 4, an Elman neural network model, a TCN-BiLSTM neural network model A, a TCN-BiLSTM neural network model B, a TCN-BiLSTM neural network model C, a NARX neural network model-PID controller and a NARX neural network model-BiLSTM neural network model.
[0075] The design process of the GCN-BiLSTM neural network model, the temperature detection module and the TCN-BiLSTM neural network model 4 refers to the relevant design method of this patent, and the NARX neural network model-BiLSTM neural network model is a series connection of the NARX neural network model and the BiLSTM neural network model.
[0076] 1. Fuzzy C-means clusterer design
[0077] Suppose there is a finite set X = {x1,x2,…x n} is a set of n samples, which are the maximum, average and minimum values output by multiple parameter sensor groups, C is the number of predetermined categories, m i (i=1,2,…c) is the center of each cluster, μ j (x i ) is the membership of the i-th sample to the j-th class, and the clustering criterion function is defined by the membership function:
[0078]
[0079] Where:||x i -m j || is x i to m j The Euclidean distance between them; b is the fuzzy weighted power index, which is a parameter that can control the fuzziness of the clustering results; M is the fuzzy C partition matrix of X, and V is the cluster center set of X. The result of the fuzzy C-means clusterer is to obtain M and V that minimize the criterion function. In the fuzzy C-means clusterer method, the sum of the membership of the samples to each cluster is required to be 1, that is:
[0080]
[0081] The fuzzy C-means clusterer can be completed by following the following iterative steps:
[0082] A. Set the number of clusters c and parameter b, the algorithm termination threshold ε, the number of iterations t = 1, and the maximum number of iterations allowed is t max ;
[0083] B. Initialize each cluster center m i ;
[0084] C. Calculate the membership function using the current cluster center;
[0085] D. Update various cluster centers using the current membership function;
[0086] E. Select an appropriate matrix norm if ||V(t+1)-V(t)||≤ε or t≥t max, stop the calculation; otherwise, t=t+1, return to step C.
[0087] When the algorithm converges, the cluster centers of each class and the membership of each parameter sensor group output value sample to each class are obtained, and the fuzzy clustering division is completed. Finally, the fuzzy clustering results are defuzzified, and the fuzzy clustering is transformed into a deterministic classification to achieve the final classification of multiple parameter sensor group output values.
[0088] 2. Elman neural network model design
[0089] As a feedback neural network, the Elman neural network model mainly consists of four layers: input layer, hidden layer, receiving layer and output layer. The Elman neural network model adds a receiving layer as the feedback part, which receives the output information of the hidden layer as its own input signal. At the same time, as a one-step delayed operator, it acts on the hidden layer together with the input layer signal at the next moment, so that the Elman neural network model can process dynamic data information in real time and has the ability to adapt to time-varying characteristics. The input signal of the hidden layer of the Elman neural network model includes not only the input signal of the input layer at the current moment, but also the input signal of the receiving layer at the previous moment. In this way, the data of the input information at different moments can be processed, which enhances the ability of the Elman neural network model to process time-varying information. The dynamic memory function of the network can reflect the characteristics of the dynamic process system. The input information of the hidden layer of the Elman neural network model includes the feedback data of the receiving layer at different moments and the input data of the input layer. If the feedback data of the receiving layer at different moments is different, the output data of the output layer may also be different. However, through the self-learning process of the Elman neural network model itself, the input and output data information is dynamically predicted by adjusting the hidden layer and receiving layer structure inside the Elman neural network.
[0090] 3. NARX neural network model-PID controller design
[0091] The NARX neural network model-PID controller uses the three values output by the NARX neural network model as Kp, Ki and Kd of the PID controller, so as to realize the dynamic adjustment of the proportional factor, integral factor and differential factor of the PID by the NARX neural network model.
[0092] The parameter monitoring Internet of Things device designed by the present invention is composed of a parameter acquisition terminal, a parameter control terminal, a central node, a remote monitoring terminal, a cloud platform and a mobile phone APP. The parameter acquisition terminal, the parameter control terminal and the central node communicate through a self-organizing network via a LoRa communication module. The communication between the parameter acquisition terminal and the parameter control terminal and the remote monitoring terminal, the cloud platform and the mobile phone APP is realized through the central node. The remote monitoring terminal and the cloud platform perform big data monitoring, storage and processing. The mobile phone APP performs two-way communication with the central node through the cloud platform. The remote monitoring terminal and the mobile phone APP enable the management personnel to remotely monitor the operating status of the parameter acquisition terminal and the parameter control terminal. A parameter monitoring Internet of Things device structure is shown in FIG. Figure 3 shown.
[0093] 1. Parameter collection terminal design
[0094] The present invention adopts a parameter acquisition terminal as a parameter sensing terminal. The parameter acquisition terminal, parameter control terminal and central node realize two-way communication through the LoRa communication module. The parameter acquisition terminal includes sensors for collecting temperature, humidity, wind speed and light intensity parameters and corresponding signal conditioning circuits and STM32 microprocessors. The software of the parameter acquisition terminal mainly realizes LoRa module communication and parameter acquisition and preprocessing. The software is designed with C language programming, which has a high degree of compatibility, greatly improves the work efficiency of software design and development, and enhances the reliability, readability and portability of the program code. Figure 4 shown.
[0095] 2. Parameter control terminal design
[0096] The parameter control terminal is designed with 4 D / A conversion circuits in the output path to realize the adjustment of temperature, humidity, wind speed and light intensity. Output control circuit, STM32 microprocessor and LoRa communication module interface are used to control the adjustment equipment. Figure 5 The STM32 microprocessor is provided with the above-mentioned intelligent temperature detection method to realize intelligent detection of temperature, humidity and wind speed.
[0097] 3. Central node design
[0098] The gateway node is composed of a LoRa communication module, an STM32 microprocessor, a Wifi module, and a USB interface. The LoRa communication module is used to build a self-organizing communication network with the parameter acquisition terminal and the parameter control terminal to realize data interaction between the central node, the parameter acquisition terminal, and the parameter control terminal. The central node realizes two-way data interaction between the central node and the cloud platform through the WiFi module, and realizes two-way data communication with the remote monitoring terminal through the USB interface. Figure 6 shown.
[0099] 4. Remote monitoring software design
[0100] The remote monitoring terminal is an industrial control computer. It mainly realizes parameter monitoring and intelligent temperature detection methods, and realizes information interaction with the central node, parameter acquisition terminal and parameter control terminal. The main functions of the remote monitoring terminal are communication parameter setting, data analysis and data management, and temperature detection system. Figure 7 shown.
[0101] The above embodiments are only for illustrating the technical concept and features of the present invention, and their purpose is to enable people familiar with the technology to understand the content of the present invention and implement it accordingly, and they cannot be used to limit the protection scope of the present invention. Any equivalent transformation or modification made according to the spirit of the present invention should be included in the protection scope of the present invention.
Claims
1. A temperature intelligent detection method, characterized in that: The steps include: Step 1: Construct a temperature detection module, which includes a temperature empirical mode decomposition model, a humidity empirical mode decomposition model, a wind speed empirical mode decomposition model, TCN-BiLSTM neural network model 1~ TCN-BiLSTM neural network model n, GCN-BiLSTM neural network model 1~ GCN-BiLSTM neural network model 3 and a NARX neural network model-HRFNN recursive fuzzy neural network model; The outputs of the temperature, humidity, and wind speed sensors are passed through the temperature empirical mode decomposition model, humidity empirical mode decomposition model, and wind speed empirical mode decomposition model to obtain decomposition signals. Each decomposition signal is passed through the corresponding TCN-BiLSTM neural network model and then input into the corresponding GCN-BiLSTM neural network model, and then calibrated and output by the NARX neural network model-HRFNN recursive fuzzy neural network model. Step 2: construct a temperature detection system, which includes a fuzzy C-means clusterer, the temperature detection module, a GCN-BiLSTM neural network model 4, an Elman neural network model, a TCN-BiLSTM neural network model A, a TCN-BiLSTM neural network model B, a TCN-BiLSTM neural network model C, a NARX neural network model-PID controller and a NARX neural network model-BiLSTM neural network model; The parameter sensor group includes a temperature sensor, a humidity sensor and a wind speed sensor. The data collected by the multiple parameter sensor groups are used as the input of the fuzzy C-means clusterer. The multiple groups of parameter sensor group data output by the fuzzy C-means clusterer are used as the input of the corresponding temperature detection modules. The outputs of the multiple temperature detection modules are used as the corresponding inputs of the GCN-BiLSTM neural network model 4. The outputs of the GCN-BiLSTM neural network model 4 are used as the corresponding inputs of the Elman neural network model and the corresponding inputs of the NARX neural network model-BiLSTM neural network model. The output of the Elman neural network model is used as the input of the TCN-BiLSTM neural network model A and the corresponding input of the NARX neural network model-BiLSTM neural network model. The error and error variation between the output of the Elman neural network model and the output of the NARX neural network model-BiLSTM neural network model The quantization rate is used as the input of TCN-BiLSTM neural network model B, the input of NARX neural network model-PID controller and the corresponding input of NARX neural network model-BiLSTM neural network model respectively. The output of NARX neural network model-PID controller is used as the input of TCN-BiLSTM neural network model C and the corresponding input of NARX neural network model-BiLSTM neural network model respectively. The output of NARX neural network model-BiLSTM neural network model is used as the corresponding input of Elman neural network model. The outputs of TCN-BiLSTM neural network model A, TCN-BiLSTM neural network model B and TCN-BiLSTM neural network model C are used as the corresponding input of GCN-BiLSTM neural network model 4 respectively. The NARX neural network model-BiLSTM neural network model outputs the predicted temperature value of the object.
2. A temperature intelligent detection method according to claim 1, characterized in that: The structure of the temperature detection module is as follows: The temperature sensor output is used as the input of the temperature empirical mode decomposition model. The low-frequency trend part and multiple high-frequency fluctuation parts of the temperature empirical mode decomposition model output are used as the input of the corresponding TCN-BiLSTM neural network model. The multiple TCN-BiLSTM neural network models are used as the input of the GCN-BiLSTM neural network model 1. The humidity sensor output is used as the input of the humidity empirical mode decomposition model. The low-frequency trend part and multiple high-frequency fluctuation parts of the humidity empirical mode decomposition model output are used as the corresponding TCN-BiLSTM neural network model inputs respectively. Multiple TCN-BiLSTM neural network models are used as the input of GCN-BiLSTM neural network model 2 respectively. The wind speed sensor output is used as the input of the wind speed empirical mode decomposition model. The low-frequency trend part and multiple high-frequency fluctuation parts of the wind speed empirical mode decomposition model output are respectively used as the corresponding TCN-BiLSTM neural network model inputs. Multiple TCN-BiLSTM neural network models are respectively used as the input of the GCN-BiLSTM neural network model 3. GCN-BiLSTM neural network model 1, GCN-BiLSTM neural network model 2 and GCN-BiLSTM neural network model 3 are respectively used as the corresponding inputs of the NARX neural network model-HRFNN recursive fuzzy neural network model, and the outputs of the NARX neural network model-HRFNN recursive fuzzy neural network model are respectively used as the corresponding inputs of the GCN-BiLSTM neural network model 1, GCN-BiLSTM neural network model 2 and GCN-BiLSTM neural network model 3, and the output of the NARX neural network model-HRFNN recursive fuzzy neural network model is used as the temperature value of the detected point as the output of the temperature detection module.
3. The temperature intelligent detection method according to claim 1, characterized in that: The temperature empirical mode decomposition model, humidity empirical mode decomposition model and wind speed empirical mode decomposition model will filter out the trends of different characteristics existing in the temperature, humidity and wind speed historical data information step by step, and decompose the temperature, humidity and wind speed historical data information into a low-frequency trend part and multiple high-frequency fluctuation parts.
4. The intelligent temperature detection method according to claim 1, characterized in that: The TCN-BiLSTM neural network model is a series connection of a TCN time series convolutional neural network model and a BiLSTM neural network model. The TCN time series neural network model learns the time series characteristics and spatial characteristics of temperature, humidity, and wind speed data, and the BiLSTM neural network model deepens the learning of the time series characteristics and spatial characteristics of temperature, humidity, and wind speed data.
5. The intelligent temperature detection method according to claim 1, characterized in that: The GCN-BiLSTM neural network model is a series connection of a GCN graph neural network model and a BiLSTM neural network model. The GCN graph neural network model extracts feature data of output information of multiple TCN-BiLSTM neural network models, performs a series of convolution operations on the extracted data, and for a single node, extracts information from neighbors connected to the node, and uses an aggregation function to aggregate node feature information and neighbor feature information to obtain new information.
6. The temperature intelligent detection method according to claim 1, characterized in that: The NARX neural network model-HRFNN recursive fuzzy neural network model is a series connection of the NARX neural network model and the HRFNN recursive fuzzy neural network model; The NARX neural network model is a dynamic network composed of interconnected nodes. It accepts one or more inputs and obtains the optimal weights between neurons and input parameters. The NARX neural network model adds a delay and feedback mechanism to store the historical data of input factors, and delays the output vector of the network and introduces it into the input vector through external feedback. The HRFNN recursive fuzzy neural network model uses a gradient descent algorithm with cross-validation to train the weights of the neural network. The HRFNN recursive fuzzy neural network model introduces internal variables in the feedback link, weights the output of the rule layer, and then defuzzifies the output as the feedback, and uses the feedback together with the output of the membership function layer as the input of the rule layer at the next moment.
7. The intelligent temperature detection method according to claim 1, characterized in that: The NARX neural network model-PID controller uses the three values output by the NARX neural network model as Kp, Ki and Kd of the PID controller, so as to realize dynamic adjustment of the proportional factor, integral factor and differential factor of the PID by the NARX neural network model.
8. A parameter monitoring Internet of Things device, characterized in that: The method is composed of a parameter acquisition terminal, a parameter control terminal, a central node, a remote monitoring terminal, a cloud platform and a mobile phone APP. The parameter acquisition terminal, the parameter control terminal and the central node communicate with each other in a self-organizing network through a LoRa communication module. The communication between the parameter acquisition terminal and the parameter control terminal and the remote monitoring terminal, the cloud platform and the mobile phone APP is realized through the central node. The remote monitoring terminal and the cloud platform perform big data monitoring, storage and processing. The mobile phone APP performs two-way communication with the central node through the cloud platform. The remote monitoring terminal and the mobile phone APP enable the management personnel to remotely monitor the operating status of the parameter acquisition terminal and the parameter control terminal. The computer program steps of a temperature intelligent detection method described in any one of claims 1 to 7 are deployed in the parameter acquisition terminal and the parameter control terminal.