A temperature and humidity intelligent monitoring system and method

Abstract

The application provides a temperature and humidity sensor intelligent monitoring system and method, which simulates a high-temperature and severe environment test environment for specific products and materials (such as vehicle key components, electronic products and the like), configures a main system, a main electric system, a control system, a heating system, a temperature control system, a load detection and the like according to specific test requirements, and checks out defective products or defective parts through the system, and reproduces a test process through real-time data or stored data.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of experimental weather resistance detection, and in particular to a temperature and humidity intelligent monitoring system and method. BACKGROUND

[0002] With the development of science and technology, accurate measurement of temperature and humidity is becoming more and more important. The measured environmental temperature and humidity are taken as inputs, and through the calculation and output of the environmental control system, important protection can be provided for the normal operation of electronic equipment and instruments, the preservation of cold chain transportation, the safety of stored grain and medicine, etc. The artificial climate simulation accelerated test method is a test method that artificially controls one or several factors to simulate natural climate environment to accelerate the durability degradation of test pieces. This test method has the advantages of short test period, strict test condition control, good reproducibility, low cost and complexity, and high reliability of results. However, as the service period is prolonged, the durability performance degradation problem of the vehicle structure brings huge losses to the national economy. It is not enough to display or record the test environment conditions of the test sample only through the control of the test box. It is urgent to accurately reproduce or repeat a temperature / humidity test process.

[0003] Many scholars at home and abroad have carried out a lot of research on temperature monitoring. However, in the experimental process, the collected image signal data often has time-varying characteristics, which will cause signal data loss or inaccurate data. Therefore, some methods are needed to recover the wrong or lost data. The traditional data representation method cannot be applied to new network data sets with complex correlation characteristics, so a model graph network for representing correlated data in an irregular domain has emerged. However, how to better analyze and process these graph network represented data sets to more efficiently mine the depth information of the data set has become one of the hot issues of current research. In view of the problem that the single sensor of the artificial climate accelerated test box cannot accurately measure, this paper proposes a multi-sensor fusion temperature and humidity image signal processing. SUMMARY

[0004] The main purpose of the present application is to solve the above problems and deficiencies, and to provide a temperature and humidity intelligent monitoring system and method, which can realize multi-sensor fusion temperature and humidity signal processing and can reproduce or repeat a temperature and humidity test process.

[0005] To achieve the above purpose, the present application first provides a temperature and humidity intelligent monitoring system and method, which adopts the following technical scheme:

[0006] A temperature and humidity intelligent monitoring method, comprising the following steps:

[0007] S1, multiple temperature sensors and humidity sensors are arranged in the intelligent temperature and humidity monitoring system, the temperature and humidity are controlled according to the predetermined program, and the temperature and humidity data of each point are collected according to the test process;

[0008] S2, the collected data of each point is respectively denoised and reconstructed;

[0009] S3, output the denoised and reconstructed data, and reproduce the experiment process.

[0010] Further, in step S2, the data of each point is respectively denoised to meet the reconstruction condition, then different sampling rates are used to simulate the situation of partial data loss, and finally the time difference smoothing reconstruction method is used to reconstruct the temperature and humidity signals.

[0011] Further, the obtained data is denoised by db6 wavelet, the default threshold value of denoising is obtained, and then the denoising function is used for denoising, and the decomposition level is 5 levels.

[0012] Further, the basic function of wavelet denoising is,

[0013]

[0014] Wherein, a, b are continuous variables, wherein a is a scale factor, and a>0, b is a displacement factor, which changes with a, b unchanged, and a group of functions is obtained after shifting and stretching The smooth curve of the temperature data and humidity data about time is obtained.

[0015] Further, for the square integrable signal x(t), the wavelet transform is defined as,

[0016]

[0017] Further, the formula for data reconstruction is,

[0018]

[0019] Wherein, J is a sampling operator, X is a reconstructed signal, Y is an observation signal, D is a time difference operator, and λ is a regularization parameter.

[0020] Further, when the iterative algorithm is used to solve, the gradient is calculated as

[0021]

[0022] After obtaining the gradient, the gradient descent method is used to solve the optimal value, and the missing sampling points in the sampling signal are recovered.

[0023] Further, the reconstruction effect is evaluated by the square root error,

[0024]

[0025] When the calculated error threshold, when the calculated root mean square error is less than the error threshold, adjust the regularization parameter λ, make the calculated root mean square error greater than or equal to the error threshold, make the reconstructed data close to the real data.

[0026] Further, when a certain sensor fails, all or part of the signal is lost, according to the signal of the other sensors associated with it, the lost signal is obtained by using the graph signal smoothness index.

[0027] Further, the original data and the reconstructed data of the graph signal are input into the trained long short-term memory network model with increased attention mechanism as double signal, the two signals are respectively processed to obtain predicted data.

[0028] The application further provides a temperature and humidity intelligent monitoring system, which adopts the following technical scheme:

[0029] A temperature and humidity intelligent monitoring system executes the monitoring method as described above.

[0030] Compared with the prior art, the temperature and humidity intelligent monitoring system and method have the following technical advantages:

[0031] 1. By adjusting the working state of the heating system and the humidifying system, a high-temperature and high-humidity harsh environment test environment is simulated, and the test process can be accelerated and the test cycle can be shortened by adjusting the heating and humidifying time;

[0032] 2. The test conditions can be strictly controlled;

[0033] 3. Multi-point data is used, the temperature and humidity in the test box can be fully reflected, and the test result has high reliability;

[0034] 4. The entire experiment process has good reproducibility through data denoising and reconstruction, and the lost data is reproduced through graph signal smoothness processing;

[0035] 5. Low cost and complexity. BRIEF DESCRIPTION OF DRAWINGS

[0036] Figure 1 The temperature and humidity intelligent monitoring system provided by the application is shown in the whole structure diagram;

[0037] Figure 2 The temperature and humidity intelligent monitoring system provided by the application is shown in the whole structure diagram;

[0038] Figure 3 : The sample rack schematic diagram of the temperature and humidity intelligent monitoring system is provided;

[0039] Figure 4 : The magnetic type temperature sensor distributor schematic diagram of the temperature and humidity intelligent monitoring system is provided;

[0040] Figure 5 : The humidity sensor installation schematic diagram of the temperature and humidity intelligent monitoring system is provided;

[0041] Figure 6 : The sensor topology diagram of the temperature and humidity intelligent monitoring system is provided;

[0042] Figure 7 : The comparison effect diagram before and after data wavelet denoising in the temperature and humidity intelligent monitoring method is provided;

[0043] Figure 8 : The comparison schematic diagram before and after data reconstruction in the temperature and humidity intelligent monitoring method is provided;

[0044] Figure 9 : The influence of the regularization parameter on the reconstruction precision in the temperature and humidity intelligent monitoring method is provided;

[0045] Figure 10 : The comparison schematic diagram of three construction methods in the temperature and humidity intelligent monitoring method is provided.

[0046] Figure 11 : The data prediction logic schematic diagram in the temperature and humidity intelligent monitoring method is provided;

[0047] Figure 12 : The LSTM network model in the temperature and humidity intelligent monitoring method is provided;

[0048] Wherein, the display 1, the upper cover plate 2, the cabinet body 3, the sensor joint 4, the wheel 5, the cabinet door 6, the keyboard 7, the NI acquisition card 8, the industrial computer 9, the temperature sensor support 10, the sample 11, the temperature sensor 12, the humidity sensor 13, the humidity sensor support 14. DETAILED DESCRIPTION

[0049] The application will be further described in detail in combination with the drawings and specific embodiments.

[0050] The application provides a temperature and humidity sensor intelligent monitoring system and method, which simulates a high-temperature and harsh environment test environment for specific products and materials (such as vehicle key components, electronic products, etc.) and sample pieces 11, configures a main system, a main power system, a control system, a heating system, a temperature control system, and a load detection subsystem according to specific test requirements, checks out defective products or defective pieces through the system, and reproduces the test process through real-time data or stored data.

[0051] To achieve the above object, the temperature and humidity sensor intelligent monitoring system provided by the application comprises an experiment cabinet, an acquisition system, a temperature sensor support 10, a humidity sensor support 14, various types of temperature sensors 12, humidity sensors 7 and a test system.

[0052] The experiment cabinet comprises a cabinet body 3, an upper cover plate 2, a keyboard support, a cabinet door 6 and a partition plate. The inside of the cabinet body 3 is divided into multiple chambers by the partition plate. The front end of the lower chamber is provided with the cabinet door 6. The acquisition system can be placed in the lower chamber closed by the cabinet door 6. The upper chamber of the cabinet body 3 is a test chamber. The sample pieces 11 are rotatable in the upper chamber. The upper chamber is provided with various types of temperature sensors 12 and humidity sensors 7, each type of which is multiple. The top of the upper chamber is provided with the upper cover plate 2 which is upwardly turned and opened. The middle chamber between the two partition plates of the upper chamber and the lower chamber is provided with the keyboard support for placing the keyboard and other articles, such as small-volume and small-quantity sample pieces 11, test reports, etc. The bottom of the cabinet body 3 is provided with wheels 5 for facilitating the movement of the entire experiment cabinet.

[0053] The collection system comprises an industrial computer 9, a NI collection box, a NI collection card 8 and a USB signal converter. The industrial computer 9 can control the working state of the heating device and the humidifying device (not shown in the figure) to make the test cavity reach or achieve the corresponding temperature and humidity within a specified time. The NI collection card 8 is placed in the NI collection box and in the lower cavity. The NI collection card 8 is taken out or placed in through the switch cabinet door 6. Each NI collection card is connected with a sensor and connected with the test system through the USB signal converter. This is a conventional technology of climate experiment and is not the focus of the present application, and thus is not described herein. In the present application, the test system can be connected with 24 channels of three-wire or four-wire platinum resistance at the same time, that is, in the present embodiment, 24 sensors can be arranged in the upper cavity and distributed at different positions of the upper cavity to detect the humidity and temperature in the cavity and transmit the detection signals to the test system. The test system collects and processes the data according to the preset method to record or reproduce the experiment process. In the present embodiment, the sensors are 24, including 20 temperature sensors 12 and 4 humidity sensors. The temperature sensors include but are not limited to patch type temperature sensors, magnetic type temperature sensors or any type of temperature sensors in the prior art to ensure that stable temperature data with different accuracies can be obtained during the experiment. The temperature sensors and the humidity sensors can be directly fixed on the surface of the sample 11, or the sensor supports for each sample can be designed for different types and positions of the sensors to face the front surface of the sample 11.

[0054] A temperature sensor support 10 can be designed to fix the sensors by using the patch type;

[0055] A humidity sensor support 13 can be designed to fix the humidity sensor 14 above the humidity sensor support 13. The humidity sensor support 13 can move in the test cavity (the upper cavity) to conveniently solve the layout problem of the humidity sensor in the test cavity.

[0056] Further, the patch type temperature sensor used in the present embodiment is a patch type sensor customized according to the actual use environment of the test cavity. The upper hole diameter of the sensor is 4 mm and can be used with the temperature sensor support 10.

[0057] The magnetic type temperature sensor used in the present embodiment is a temperature sensor made of PT100 material. The sensor uses a stainless steel probe and a 304 stainless steel tube. The measurement range is (-50-250) DEG C. The magnetic head of the sensor can be attracted to the center of the sample 11 to directly detect the surface temperature of the sample 11, so that the measurement data is more accurate.

[0058] In practical applications, other different types of temperature sensors can also be used, and various types of sensors are arranged at different positions of the test chamber, and if necessary, multiple different types of temperature sensors can be arranged at a key detection position at the same time to detect the temperature at the point to ensure the experimental effect.

[0059] In the embodiment, the humidity sensor 14 is a high-temperature humidity sensor of Ruzhonike, and the measurement range is (0-100) %RH, and is used in combination with the humidity sensor support 13.

[0060] The test system includes a computing service module, a data acquisition module, a storage module, a display and sharing module. The data acquisition module acquires real-time data of the temperature sensor 10 and the humidity sensor 14 through the preset acquisition software at regular intervals, and sends the computing service module, analyzes through modeling, reconstructs the intelligent monitoring algorithm by using multi-sensor fusion temperature and humidity graph signal processing (GSP), calculates, reconstructs and analyzes the data collected at each point (each sensor is a point), and then processes the data. The processed data, graph, real-time raw data can be stored through the storage module, and can also be displayed through the display of the display and sharing module, or downloaded through the USB and the like. The data processing and reconstruction process includes: first, collecting temperature and humidity experimental data from the test chamber, and denoising to meet the reconstruction condition; then, using different sampling rates to simulate the condition of partial data loss, using a reconstruction method based on time difference smoothness to reconstruct the temperature and humidity signals; finally, analyzing the reconstruction result of the data to effectively restore the original signal, so that the entire detection process has high reconstruction accuracy and strong robustness.

[0061] The application further provides the intelligent monitoring method of the temperature and humidity sensor, and specifically as follows:

[0062] S1, according to the number and placement position of the sample, the type and number of the temperature sensor 10 and the humidity sensor 14 are arranged around the sample in the test chamber according to the test regulation, to ensure that the temperature and humidity of the surface and / or surrounding of the sample can be truly and accurately detected, and the sensor connector 4 on the cabinet body 3 of each sensor is connected, the sensor connector 4 is an aviation plug, one end of which is connected with the sensor, and the other end is connected with the USB signal converter in the lower chamber, so that the temperature and humidity signals of the sensor are converted into a signal mode that can be processed by the test system.

[0063] In the embodiment, the upper limit of the number of sensors in the test cavity is 28, including 24 temperature sensors and 4 humidity sensors. The temperature sensors include multiple types of sensors, and the number of each type of sensor is also multiple. The number of humidity sensors is less, which can be a single type or two or more types. In actual application, the type and number of temperature and humidity sensors are determined according to the volume of the test cavity, the volume and number of test samples, and the placement position in the test cavity, without any requirements and restrictions.

[0064] The test sample is installed on the sample holder of the test cavity, each sensor is placed on the corresponding sensor support according to the regulation, and if it is a magnetic temperature sensor, it is directly adsorbed on the surface of the test sample. The positions of the temperature sensor support 10 and the humidity sensor support 14 are adjusted to facilitate the sensors to more truly detect the temperature and humidity of the sample surface and / or the periphery of the sample.

[0065] S2, open the cabinet door, start the industrial computer and the display, each sensor starts to collect temperature data and humidity data respectively, and transmits them to the calculation service module. Through modeling analysis, the intelligent monitoring algorithm is reconstructed by using multi-sensor fusion temperature and humidity graph signal processing (GSP). The collected data of each point (each sensor as a point) is calculated, reconstructed and analyzed respectively.

[0066] In the embodiment, the temperature data is measured by the temperature sensors distributed and arranged in the test cavity (including those adsorbed on the surface of the sample), and the humidity data is measured by the hygrometer placed at the center position of the air of the measurement object (the sample in the test cavity). When the test setting time is 5 hours, the environmental temperature range of the test cavity is-20℃-120℃, and the humidity range is 0-8.3×10 -3 (Kg / m 3 ), wherein the temperature rises from-20℃ to 120℃ in the first 100 minutes, and then remains at 120℃. The humidity rises to the maximum value in the first 100 minutes, and then decreases to 5.5×10 -3 (Kg / m 3 ) after 40 minutes, and then remains unchanged. This condition simulates the general device working environment, which is consistent with the actual situation. The topology structure between the sensors is described by using an undirected graph model. As shown in Figure 6 The topological graph formed by the relative positions of the sensors is shown in FIG. 1. Nodes 1 to 24 represent 24 temperature sensors, and nodes 25 to 28 represent 4 humidity sensors.

[0067] The sampling frequency of each sensor in the test box is 1 Hz, that is, data is collected once per second, and the experiment records 5 hours of data, so each sensor contains 18000 time points, and the recorded data X (including humidity data and temperature data) is 28x18000, wherein each row represents the data of one sensor. The collected data is processed by the calculation service module.

[0068] The temperature and humidity signals collected from the sensor are stored to obtain observation signals Y e RNXM. Due to the existence of a small measurement error of the actual sensor, the signal presents a fluctuating state, and the obtained signal is not smooth enough, as shown in Figure 7 (a) and Figure 7 (c), so it is necessary to denoise, so that the temperature-time or humidity-time curve is smooth and closer to the actual data. In the processing of the test data, wavelet denoising is used. The wavelet used is db6 wavelet, and the default threshold for denoising is obtained first, and then the denoising function is used for denoising, and the decomposition level is 5 levels. Figure 7 For comparison before and after signal denoising, the denoised image is shown in Figure 7 (b) and Figure 7 (d), which are the processed humidity-time and humidity-time, and from the figure it can be seen that the denoised data is obviously smoother, thereby satisfying the assumption of signal reconstruction.

[0069] Specifically, wavelet transform is an effective time-frequency joint analysis method, which can provide a "time-frequency" window that changes with frequency, and adaptively set the time and frequency resolution, and is commonly used for signal denoising and compression. In this embodiment, the given basic function of continuous wavelet transform is:

[0070]

[0071] With the continuous change of a and b, a set of functions can be obtained after shifting and stretching For a square-integrable signal x(t), the wavelet transform is defined as a function:

[0072]

[0073] In practical applications, the Mallat algorithm is often used to realize wavelet transform, and the algorithm flow is as follows:

[0074]

[0075] Wherein, a j (k), d jh0(k) are the discrete approximation coefficients in multiresolution analysis, and h0(k) are the weighting coefficients, which are a discrete sequence. h0(k) and h1(k) are two filters satisfying the two-scale difference equation, then a j (k), d j (k) has the following recurrence relation:

[0076]

[0077]

[0078] in

[0079] The Mallat algorithm divides the original data into frequency-based categories, distinguishing between high-frequency and low-frequency components to achieve noise reduction and obtain data such as... Figure 7 The curve shown.

[0080] During the sampling process, situations where data cannot be collected are inevitable, such as... Figure 8 As shown in Figures 8(a) and 8(c), the acquired signals are represented in gray, and the blank areas between the gray signals represent lost signals. For the lost data, the complete experimental process cannot be reproduced. Furthermore, the acquired signals are represented as discrete data points on the corresponding humidity-time or temperature-time graphs, forming discontinuous graphs. Directly connecting the data points will result in zigzag lines with large fluctuations and / or angled connections. Generally, for real-world data, signals are often smooth, rather than strictly band-limited relative to the underlying graph structure. Therefore, in this embodiment, for situations where signals cannot be completely sampled, a data reconstruction method is chosen to reconstruct and complete the signals involved in the entire experimental process (the complete signal from each sensor).

[0081] If the signal x(t) at each time step is relatively smooth compared to the underlying graph, meaning the signal loss is small, then X can be recovered column by column using the reconstruction method for smoothed signals. During the reconstruction process, smoothness is used as a regularization term, and the optimization objective can be written as:

[0082]

[0083] However, when the signal x(t) is not smooth enough, i.e. the signal loss is large, the reconstruction quality of the above method is poor. The above method only considers the change amount of all vertices at the same time, while the actual data also has certain correlation between adjacent time. The non-smoothness of the signal is mainly caused by the sampling process, such as data fluctuation caused by measurement error, or the absence of some sampling points. For incomplete signals, certain methods are needed to restore them. Considering the time correlation and spatial correlation comprehensively, using more complete constraint conditions can improve the reconstruction quality.

[0084] The signal smoothness of the actual data can be found that even if the signal at each time is not smooth enough, the difference signal x t -x t-1 still has good smoothness. Define D as the time difference operator, then:

[0085]

[0086] The time difference signal can be expressed using the time difference operator as X' = XD = [x2-x1, x3-x2,..., xn-xn-1]. M -x M-1 ].

[0087] Let J be the sampling operator, and:

[0088]

[0089] where S t is the point sampled at t, let the observed signal be Y = J0X * + V, where X * is the original signal, and V represents noise, then the reconstruction can be described as the following optimization objective function:

[0090]

[0091] In the objective function, the first term minimizes the reconstruction error, and the second term minimizes the smoothness of the difference time signal, i.e. ensures the smoothness of the time-varying signal. λ is the regularization parameter. Write equation (7) in matrix form:

[0092]

[0093] The solution of this function has a closed-form expression, but it involves the inversion of a large matrix, which has a large computational complexity. To simplify the program, an iterative algorithm can be used to solve it, and the gradient is calculated as:

[0094]

[0095] After obtaining the gradient, the gradient descent method or other methods can be used to solve the optimal value.

[0096] The signal reconstruction method based on differential smoothness described above can ensure good reconstruction quality even when signal smoothness is insufficient. This method can effectively recover lost sampling points in the sampled signal and improve the quality of the output signal. A comparison of the signals before and after reconstruction is shown in Figure 8.

[0097] In the experiment, sampling operator J was used to simulate the situation of partial signal loss. The sampling algorithm was random sampling; the higher the sampling rate, the better the realism of the reconstruction. Extensive experimental verification showed that when the sampling rate was 40%, meaning 40% of the vertices were randomly sampled at each moment, the data obtained by reconstructing the entire signal based on the sampled data met the realism requirements for reconstruction. The reconstruction effect was evaluated using the root mean square error.

[0098]

[0099] Where x is the reconstructed signal, x * It is the original signal, N x It is the signal length; in this embodiment, the signal length N x The data can be arbitrarily selected within the entire 5-hour timeframe, either a specific period or the entire test length. If the signal is a matrix, it should be vectorized before calculating the error. For example... Figure 8 The image shown is a comparison between the reconstructed signal and the sampled signal. Figure 8 As can be seen, the sampled signal contains many missing values, but the reconstructed signal can fill in these missing values, and the overall signal is relatively smooth. The reconstruction results confirm the smoothness of the differential signal. The good smoothness of the differential signal is consistent with the physical meaning of simultaneous measurement by multiple sensors, that is, the temperature changes at different locations of the object are synchronous.

[0100] Furthermore, during the verification of the reconstruction effect, it was found that the value of the regularization parameter λ had a significant impact on the reconstruction results. Figure 9 The figure shows the relationship between the regularization parameter and the reconstruction error. As can be seen from the figure, the reconstruction error first decreases as λ increases, then increases after reaching a minimum value (i.e., the optimal λ value). However, if λ is too large, the algorithm may not converge. Within the range of λ values ​​that allow the algorithm to converge, the reconstruction error is relatively small. In this embodiment, λ is set to 0-0.6, with a step size of 0.05.

[0101] As mentioned earlier, even if signals are occasionally uncollectible, the entire process signal of each sensor can be obtained through data reconstruction, realistically reproducing the test process of each sensor. However, in practice, if a sensor malfunctions, it will be impossible to obtain all signals after the sensor malfunctions, and the test process of that sensor cannot be reproduced. Therefore, this invention further utilizes a signal processing method to simulate and obtain all data after the sensor malfunctions based on signals from other sensors associated with the malfunctioning sensor. Specifically, as follows:

[0102] The structure of the graph signal can be represented as G. G＝{V,E,W} Where V = {v1,...,v} N Let} represent the set of nodes, and N be the number of nodes. E represents the set of edges, and W represents the weighted adjacency matrix between nodes, where the weights represent the degree of association between nodes. If node i and node j are related, then W... i,j >0 indicates a higher correlation, resulting in a greater weight. If the network is unweighted, then A∈R is used. N×N This represents an unweighted adjacency matrix, where each element takes the value 0 or 1. In a typical graph structure, each vertex has a signal value; the graph signal is denoted as x∈R. N If the signal is time-varying, then X = [x1, x2, ..., x...]. M ]∈R N×M Where M is the time length, x t ∈R N ,t∈{1,2,...,M} represents the graphical signal at time t.

[0103] The most commonly used matrix in graph signal processing is the graph Laplacian matrix, which is defined as follows:

[0104] L = DW (11)

[0105] Where D is the degree matrix of the graph, and the diagonal elements d i This represents the sum of the i-th row in the adjacency matrix. The normalized graph Laplacian matrix is:

[0106] LN=D -1 / 2 LD -1 / 2 (12)

[0107] By performing eigenvalue decomposition on the normalized graph Laplacian matrix, we obtain the eigenvector matrix U and its corresponding eigenvalue matrix Λ.

[0108] Define the Fourier transform and inverse Fourier transform of the graph signal as follows:

[0109]

[0110]

[0111] The eigenvectors obtained from eigenvalue decomposition can be regarded as frequency components, and the physical meaning of the coefficients of the Fourier transform is the proportion of these frequency components relative to the signal.

[0112] To characterize the drastic changes in the signal on the graph (caused by missing faulty sensor data), this invention employs the indicator of graph signal smoothness. The smoothness of the graph signal can be determined by l p Norm representation:

[0113]

[0114] When p = 2, the above formula can be written as:

[0115]

[0116] For a time-varying signal, X∈R N×M Each row represents the time series of the corresponding vertex, and its smoothness is defined as the sum of the smoothness at each time step:

[0117]

[0118] Smoothness is a qualitative indicator that can only be used to compare different signals within the same graph structure. The smaller the value, the slower the signal changes relative to the underlying graph structure, meaning that the signal values ​​at adjacent vertices are closer, and low-frequency components account for a larger proportion. Missing data can be identified based on the sum of smoothness values.

[0119] In the embodiments provided by this invention, in addition to reproducing the experimental process, the data collected by each sensor during the experiment can be monitored in real time via a display to monitor and control the experimental process, predict the development trend of temperature and humidity data in the test chamber, and provide predictive data. Because the changes and development of temperature and humidity in the test chamber are quite complex, there is currently no mature and high-precision temperature and humidity prediction method in conventional technologies. Existing technologies typically use discrete single-point data for data fitting and prediction, which has low accuracy and results in significant deviations between the predicted data and the actual data later. Furthermore, the prediction process requires a certain amount of data within a certain time frame to complete the learning and prediction. However, as mentioned earlier, during intelligent temperature and humidity monitoring, there are problems such as significant data loss, failure to collect data normally, or malfunction of a sensor, leading to the inability to make predictions or inaccurate predictions.

[0120] To solve this problem, such as Figure 11As shown, this invention utilizes a graph signal method to address situations where data is lost, cannot be collected normally, or a sensor malfunctions. As described above, the graph signal method establishes a more accurate reconstructed signal based on the topological information between each sensor, thereby completing the data, including data missing during the acquisition process and all data that could not be collected due to sensor malfunction. The original sensor data and the graph signal reconstructed data are then input as dual signals into a pre-trained long short-term memory network model employing an enhanced attention mechanism. Simultaneously, feature signals acquired from both signals are also input to obtain predictive data. This dual-signal input method can more accurately predict changes in temperature and humidity within an artificial environment chamber and quickly reflect the true state of each sensor.

[0121] In this invention, the method for data prediction using the reconstructed signal and the original signal is as follows: Figure 11 As shown, it includes:

[0122] I. Data Acquisition: This includes raw signals collected by various sensors and reconstructed data obtained through the methods described above. The raw signals are n-dimensional discrete data, representing the actual data collected by the sensors. Predictive data obtained solely from discrete data lacks accuracy. The reconstructed signals, however, are n-dimensional continuous data obtained by wavelet denoising of the original signals and then reconstructing them based on smoothness characteristics. This data is more accurate and closely reflects the actual trend of the data, resulting in higher prediction accuracy. By coupling the raw and reconstructed signals, the obtained prediction data are mutually fitted, making the final prediction data closer to the actual trend of temperature and humidity signals, and thus more accurately predicting trend changes.

[0123] 2. Extract features from the input data.

[0124] Feature values ​​and target values ​​are extracted from the original and reconstructed signals respectively to address the issue of low output accuracy caused by large data spans. This standardizes the numerical data, which includes time-domain, frequency-domain, and matrix features. Three types of feature values ​​are extracted from the data, including:

[0125] 1. Extract time-domain feature data: For the one-dimensional signals in the time domain of the original data pairs and the reconstructed data, extract the necessary time-domain features such as mean, variance, root mean square, peak factor, kurtosis coefficient, waveform factor, margin factor, and impulse factor.

[0126] 2. A Fast Fourier Transform (FFT) is applied to the input dual data to obtain frequency domain features such as characteristic frequency, mean square frequency, centroid frequency, and frequency difference. Common frequency domain feature parameter expressions are shown in the table below:

[0127] Common frequency domain characteristic parameter expression table

[0128]

[0129] 3. Obtain the matrix eigenvectors of the graph Laplacian matrix using mathematical tools, including the matrix's rank, eigenvalues, SVD-singular values, and other properties.

[0130] Extracting feature values ​​and target values ​​from data is not the focus of this invention. Any existing or future extraction methods can be used without limitation or requirement.

[0131] III. Constructing a Neural Network Model: LSTM (Long Short-Term Memory). LSTM is widely used for many sequence tasks and performs better than other sequence models (such as RNNs), especially with large amounts of data. LSTM is carefully designed to avoid the vanishing gradient problem of RNNs, storing more memories (hundreds of time steps), enabling the model to learn long-term dependencies. Compared to RNNs that maintain only a single hidden state, LSTM has more parameters, allowing for better control over which memories to save and discard at specific time steps; for example, the hidden state must be updated in each training step.

[0132] The acquired dual data, data feature reference values, and feature extraction methods are input into the LSTM base model to construct a model suitable for temperature and humidity data prediction. Figure 12 The neural network model shown.

[0133] Fourth, add an attention mechanism to the neural network model constructed in step three, select key temperature and humidity information for processing to improve the efficiency of the neural network, and determine the weights of the model.

[0134] Attention mechanisms enable neural networks to selectively focus on detailed information in a particular part of the input, allowing for the modeling of dependencies regardless of their distance in the input or output sequence. Self-attention (also known as internal attention) involves different positions in a single sequence to compute a representation of the sequence.

[0135] When neural networks process large amounts of temperature and humidity information, they can leverage the human brain's attention mechanism to select key temperature and humidity information for processing, thereby improving the efficiency of the neural network. In the neural network model constructed in step S3, max pooling and gating mechanisms can be approximated as bottom-up saliency-based attention mechanisms. Furthermore, top-down focused attention is also an effective information selection method.

[0136] Feeding a time series sequence into an LSTM model yields an output with dimensions (batch_size, time_steps, lstm_units), which is then used as the feature output for each time step. Figure 12 Model input;

[0137] After Permute flips the 2nd and 1st axes, its dimensions are transformed from (batch_size, time_steps, lstm_units) to (batch_size, lstm_units, time_steps).

[0138] After passing through a fully connected layer and Softmax, its dimensions are still (batch_size, lstm_units, time_steps). Its actual meaning is to use the fully connected layer to calculate the weight of each time_step.

[0139] After Permute flips the 2nd and 1st axes, its dimensions are transformed from (batch_size, lstm_units, time_steps) to (batch_size, time_steps, lstm_units). This represents the weight of each feature in each STEP.

[0140] V. Model Training.

[0141] The training data is input into the model to train the weights of the entire model in order to obtain the optimal model and weights.

[0142] It should be noted that this step is an offline training process to obtain the optimal model so as to obtain realistic prediction data in normal applications. In actual applications, this step is not required, and you can directly proceed to step S6.

[0143] During training, the training set includes a large amount of data from the entire process, corresponding reconstructed data, data feature values, and decision criteria. The training set is input into the model, and the obtained predicted data is compared with the real data in the training set. Based on the decision criteria, the degree of agreement between the predicted and real data, as well as the degree of agreement of the feature signals, is determined. The weights in the model are continuously adjusted until the predicted data obtained from the training set meets the decision requirements, ensuring that the predicted data matches the real data in the training set, or the error is within a threshold range. The obtained weights are then embedded into the model, resulting in a neural network model that can be used for temperature and humidity prediction.

[0144] VI. Prediction Results.

[0145] The test data is input into the model to obtain the temperature and humidity results predicted by the model, thereby quickly reflecting the true state of each sensor.

[0146] Of the six steps above, steps one through five are used for offline model training to obtain a mature, stable neural network model suitable for predicting temperature and humidity data. By checking whether the difference between the predicted value and the actual value meets the threshold requirement, the weights in the LSTM model are adjusted to obtain a suitable neural network model. Steps one through four and step six are used for actual temperature and humidity testing. The collected data and reconstructed data are input into the neural network model to obtain the predicted data, and step five is not required.

[0147] S3: The computing service module sends the processed data to the storage module and the display and sharing module. The denoised data, reconstructed data, and smoothness data obtained after processing are sent to the storage module for storage and future reference, and displayed on the screen of the display and sharing module.

[0148] Meanwhile, this invention also tested the performance of other methods at different sampling rates. In addition to the methods described above, the experiment also tested natural domain interpolation methods and graph regularization

[15] methods, and the results are as follows. Figure 10 As shown in the figure, the reconstruction error of various methods decreases with increasing sampling rate, and the method based on differential smoothness is always superior to other methods. The results indicate that the method of the present invention can effectively recover the sampled signal.

[0149] In summary, the weld surface defect detection and classification method provided by this invention has the following technical advantages compared with the prior art:

[0150] 1. Weld contour data is obtained by 3D camera scanning based on line structured light. Various weld defects are inspected and classified through image curve analysis, realizing online monitoring of welding quality after welding.

[0151] 2. It has strong real-time performance. Data can be output and analyzed in real time during the camera scanning process, and the analysis results can be output as soon as the scan is completed. Compared with the current detection methods, it has high real-time performance, which speeds up the analysis process and makes it easier to find problems, adjust and solve them in a timely manner.

[0152] 3. The analysis results are more intuitive. During the scanning process, weld contour curves, width and height curves, depth images, etc. can be generated in real time, which can intuitively reflect and output the type, location and size of defects; offline analysis can also be performed on the saved data, and parameters can be modified to suit specific welding conditions.

[0153] 4. Reduced manual intervention: This method can automatically scan and analyze data, outputting the types of surface defects and various geometric parameters of the weld, greatly improving the degree of automation and making it more adaptable to the development of modern welding technology.

[0154] As described above, similar technical solutions can be derived from the given solutions. Any simple modifications, equivalent changes, and alterations made to the above embodiments based on the technical essence of this invention, without departing from the scope of the present invention, shall still fall within the scope of the present invention.

Claims

1. A method for intelligent monitoring of temperature and humidity, characterized in that: Includes the following steps, S1, multiple temperature and humidity sensors are deployed in the intelligent temperature and humidity monitoring system. According to a predetermined program, the temperature and humidity changes are controlled, and temperature and humidity data at each point are collected according to the test progress. S2, denoising and reconstruction are performed on the data collected from each point, and the lost data is reconstructed using a temporal difference smoothness reconstruction method. The formula used for data reconstruction is: Where J is the sampling operator, X is the reconstructed signal, Y is the observed signal, D is the time difference operator, and λ is the regularization parameter; When a sensor malfunctions and loses all or part of its signal, the lost signal can be obtained by using the graph signal smoothness index based on the signals from other related sensors. Using the graph signal method, a reconstructed signal is established based on the topological information between each sensor to complete the data. The original sensor data and the reconstructed graph signal data are input as dual signals into a pre-trained long short-term memory network model employing an enhanced attention mechanism. Simultaneously, feature signals acquired from both signals are input, and feature processing is performed on each signal to obtain predicted data. This includes... Acquire raw signals from each sensor and reconstructed data. Feature extraction is performed on the input data to obtain the feature values ​​and target values ​​of the data from the original signal and the reconstructed signal, respectively. The numerical data is standardized. The features of the data include time domain features, frequency domain features and matrix features. LSTM neural network model is constructed by inputting the acquired dual data, data feature reference values, and feature extraction methods into the LSTM base model to build a neural network model suitable for temperature and humidity data prediction. An attention mechanism is added to the constructed neural network model to select key temperature and humidity information for processing and determine the model weights. The test data is input into the model to obtain the temperature and humidity structure predicted by the model, so as to reflect the true state of each sensor. S3 outputs the data after noise reduction and reconstruction to reproduce the experimental process.

2. The intelligent temperature and humidity monitoring method as described in claim 1, characterized in that: In step S2, the data at each point are denoised to meet the reconstruction conditions. Then, different sampling rates are used to simulate the situation of partial data loss. Finally, the time difference smoothness reconstruction method is used to reconstruct the temperature and humidity signals.

3. The intelligent temperature and humidity monitoring method as described in claim 1, characterized in that: The acquired data was denoised using the db6 wavelet. The default denoising threshold was obtained, and then a denoising function was used to denoise the data. The decomposition level was 5.

4. The intelligent temperature and humidity monitoring method as described in claim 1, characterized in that: The basic function for wavelet denoising is, Where a and b are continuous variables, a is a scale factor (a > 0), and b is a displacement factor that varies invariantly with a and b. After displacement and scaling, a set of functions is obtained. A smooth curve of temperature and humidity data over time is obtained.

5. The intelligent temperature and humidity monitoring method as described in claim 4, characterized in that: For a square-integrable signal x(t), the wavelet transform is defined as follows:

6. The intelligent temperature and humidity monitoring method as described in claim 1, characterized in that: When solving using an iterative algorithm, its gradient is calculated as follows: After obtaining the gradient, the gradient descent method is used to solve for the optimal value and recover the lost sampling points in the sampled signal.

7. The intelligent temperature and humidity monitoring method as described in claim 1, characterized in that: The reconstruction effect is evaluated regardless of the root square error. When the calculated root mean square error is less than the error threshold, the regularization parameter λ is adjusted so that the calculated root mean square error is greater than or equal to the error threshold, making the reconstructed data closer to the real data.

8. A temperature and humidity intelligent monitoring system, characterized in that: Perform the monitoring method as described in any one of claims 1 to 7.

Images