Multi-modal sensor drift self-calibration method in complex industrial environment

Abstract

The invention provides a multi-mode sensor drift self-calibration method in a complex industrial environment, and relates to the technical field of sensor calibration. Comprising the following steps: acquiring temperature gradient, equipment operation duration and environment humidity change data in real time by deploying an environment reference sensor group, constructing a multivariable autoregression model to dynamically predict the temperature deviation of a thermal imaging sensor, and introducing a compensation coefficient to correct a temperature measurement value after standardization processing and recursive parameter optimization so as to obtain the temperature deviation of the thermal imaging sensor. Calculating high-frequency energy of the image blocks by adopting two-dimensional discrete cosine transform, and generating a definition score; and when the score is lower than a preset threshold value, searching an optimal focusing position in a preset focal length interval through a golden section method to realize self-adaptive calibration. Calibration is dynamically triggered through a self-supervised closed-loop mechanism, sliding window data updating and abnormal value filtering are combined, and it is ensured that the model adapts to dynamic changes of the environment. The method solves the problems of temperature drift of the thermal imaging sensor and image blurring of the visible light sensor in a high-temperature and high-dust scene.

Description

Technical Field

[0001] The present invention relates to the technical field of sensor calibration, and particularly to a multi-modal sensor drift self-calibration method in a complex industrial environment. Background Art

[0002] In complex industrial scenarios such as cotton and linen textile, chemical fiber processing, etc., where there are high temperatures, high dust, and continuous operation of equipment, multi-modal sensors (such as long-wave thermal imaging sensors, visible light cameras) have become the core components of fire warning systems by virtue of their collaborative monitoring capabilities for temperature fields and smoke characteristics. However, such sensors face multiple technical challenges during long-term operation: affected by high-temperature environments, the temperature measurement accuracy of thermal imaging sensors gradually decays over time, leading to a significant increase in the risk of fire missed alarms; at the same time, dust accumulation and equipment vibration can cause focal length deviation or image blurring of visible light lenses, resulting in a decrease in the detection accuracy of the fire and smoke detection model for low-contrast smoke targets, seriously affecting the effective identification of early fire signals.

[0003] In existing research, sensor calibration relies on manual periodic shutdown calibration, such as thermal imaging blackbody calibration and manual focusing of visible light lenses. Each calibration takes up to several hours, which not only fails to meet the requirements of 7×24-hour continuous operation of industrial production, but also the fixed calibration parameters set manually are difficult to respond in real time to dynamic interferences such as environmental temperature and humidity fluctuations and equipment layout changes, resulting in lag errors in the performance of the calibrated sensors and being unable to fundamentally solve the drift problem in complex environments. These defects cause the detection accuracy of multi-modal sensors to gradually decrease during long-term use, and the reliability of fire warning systems faces severe challenges. Summary of the Invention

[0004] In order to overcome the deficiencies of the prior art, the purpose of the present invention is to provide a multi-modal sensor drift self-calibration method in a complex industrial environment, and the present invention solves the problems of time-consuming and low accuracy in sensor calibration in the prior art.

[0005] To achieve the above purpose, the present invention provides the following solutions:

[0006] A multi-modal sensor drift self-calibration method in a complex industrial environment, comprising:

[0007] Deploying an environmental reference sensor group to collect temperature gradient data, equipment operation duration data, and environmental humidity change data in the industrial scenario in real time;

[0008] Using a multi-variable autoregressive model to construct an initial temperature deviation autoregressive model;

[0009] Optimizing the initial temperature deviation autoregressive model according to the temperature gradient data, equipment operation duration data, and environmental humidity change data to obtain a final temperature deviation autoregressive model;

[0010] Based on the final temperature deviation autoregressive model and introducing a compensation coefficient, the corrected temperature data is obtained;

[0011] Visible light image data is acquired and the image sharpness score is calculated based on the high-frequency energy of the two-dimensional discrete cosine transform. When the sharpness is lower than the preset threshold, the golden section method is used to adaptively adjust the lens focal length to obtain the calibrated image sharpness.

[0012] Preferably, it further includes:

[0013] Collect the corrected temperature data within a preset time and the image sharpness data of a preset number of frames;

[0014] Compare the corrected temperature data within the preset time with the environmental reference data to obtain a first comparison result;

[0015] Compare the image sharpness data of the preset number of frames with the image sharpness of the reference image to obtain a second comparison result;

[0016] Retrain the parameters of the final temperature deviation autoregressive model according to the first comparison result;

[0017] Readjust the lens focal length according to the second comparison result.

[0018] Preferably, optimizing the initial temperature deviation autoregressive model according to the temperature gradient data, equipment operation duration data, and environmental humidity change data to obtain the final temperature deviation autoregressive model, including:

[0019] Normalize the temperature gradient data, equipment operation duration data, and environmental humidity change data respectively to obtain normalized data;

[0020] Optimize the initial temperature deviation autoregressive model according to the normalized data to obtain the final temperature deviation autoregressive model, where the expression of the final temperature deviation autoregressive model is:

[0021]

[0022] Among them, ΔT t is the temperature deviation, p is the autoregressive order, α i , β, γ, δ are the first model parameter, the second model parameter, the third model parameter, and the fourth model parameter respectively, ΔH t,scaled , D t,norm , are the normalized environmental humidity change data, equipment operation duration data, and temperature gradient data respectively, ∈ t is zero-mean Gaussian noise, and i is a natural number.

[0023] Preferably, the expression of the corrected temperature data is:

[0024] T c,t =T sensor,t -K t ·ΔT t ;

[0025] Wherein, T c,t is the corrected temperature data, T sensor,t is the measured temperature data, and K t is the output compensation coefficient.

[0026] Preferably, obtaining visible light image data and calculating an image sharpness score based on the high-frequency energy of a two-dimensional discrete cosine transform includes:

[0027] Performing image block division on the visible light image data to obtain block data;

[0028] Performing two-dimensional DCT transformation on the block data to obtain a frequency domain coefficient matrix;

[0029] Determining the energy of the high-frequency components according to the frequency domain coefficient matrix;

[0030] Determining the image sharpness score according to the energy of the high-frequency components.

[0031] Preferably, the expression of the frequency domain coefficient matrix is:

[0032]

[0033] Wherein, C(u, v) is the frequency domain coefficient matrix, α(u) is the DC component coefficient, α(v) is the AC component coefficient; u is the frequency index of the image block in the horizontal direction; v is the frequency index of the image block in the vertical direction; I(x, y) is the pixel value of the image block at (x, y).

[0034] Preferably, the expression of the energy of the high-frequency components is:

[0035]

[0036] Wherein, E b is the energy of the high-frequency components.

[0037] Preferably, using the golden section method to adaptively adjust the lens focal length to obtain the calibrated image sharpness includes:

[0038] Setting an initial segmentation point within a preset focal length range and calculating the image sharpness at the corresponding position;

[0039] Narrowing the search interval according to the sharpness comparison result until the interval accuracy meets the preset threshold;

[0040] Select the endpoint with the highest clarity as the optimal focal length position, and verify the image clarity after calibration to obtain the image clarity after calibration.

[0041] The present invention discloses the following technical effects:

[0042] The present invention provides a multi-modal sensor drift self-calibration method in a complex industrial environment, including:

[0043] Deploy an environmental reference sensor group to collect temperature gradient data, equipment operation duration data, and environmental humidity change data in the industrial scenario in real time; construct an initial autoregressive model of temperature deviation using a multivariate autoregressive model; optimize the initial autoregressive model of temperature deviation according to the temperature gradient data, equipment operation duration data, and environmental humidity change data to obtain the final autoregressive model of temperature deviation; obtain the corrected temperature data according to the final autoregressive model of temperature deviation and introduce a compensation coefficient; acquire visible light image data and calculate the image clarity score based on the high-frequency energy of the two-dimensional discrete cosine transform. When the clarity is lower than the preset threshold, adaptively adjust the lens focal length using the golden section method to obtain the image clarity after calibration. The present invention realizes the real-time correction and adaptive optimization of sensor performance by constructing an autoregressive model of temperature deviation and an image clarity measurement method based on DCT high-frequency energy, combined with the golden section search and dynamic parameter update mechanism. This method adopts a self-supervised closed-loop design, can dynamically respond to environmental changes and equipment state fluctuations, and significantly improves the long-term stability and detection accuracy of the sensor. In addition, the system does not require manual intervention or shutdown operations, fully adapts to the continuous production requirements of the industrial scenario, and provides efficient and reliable technical support for the reliable operation of the fire warning system and multi-modal sensing monitoring in a complex industrial environment, with broad application potential and promotion value. Description of the Drawings

[0044] In order to more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the following will briefly introduce the drawings required in the embodiments. Obviously, the drawings in the following description are only some embodiments of the present invention. For those of ordinary skill in the art, without creative efforts, other drawings can be obtained based on these drawings.

[0045] Figure 1 A multi-modal sensor drift self-calibration method in a complex industrial environment provided by an embodiment of the present invention. Detailed Embodiments

[0046] The following will clearly and completely describe the technical solutions in the embodiments of the present invention with reference to the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all the embodiments. All other embodiments obtained by those of ordinary skill in the art based on the embodiments of the present invention without creative efforts shall fall within the protection scope of the present invention.

[0047] To make the above objects, features, and advantages of the present invention more obvious and understandable, the present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0048] As Figure 1 shown, the present invention provides a multi-modal sensor drift self-calibration method in a complex industrial environment, including:

[0049] Step 100: Real-time collect temperature gradient data, equipment operation duration data, and environmental humidity change data in the industrial scenario by deploying an environmental reference sensor group;

[0050] Specifically, the environmental reference sensor group: fixedly installed in an unobstructed area of the workshop, including high-precision thermocouples (accuracy ±0.5°C), temperature and humidity sensors (humidity accuracy ±2%RH), and the sampling frequency is 1Hz. Among them, the sensor group adopts a rectangular grid layout, with a horizontal (east-west) spacing of 2 meters and a vertical (north-south) spacing of 1.5 meters. A total of 3×3 high-precision thermocouples are deployed to form a temperature field monitoring network.

[0051] Step 200: Construct an initial autoregressive model of temperature deviation using a multi-variable autoregressive model;

[0052] Step 300: Optimize the initial autoregressive model of temperature deviation according to the temperature gradient data, equipment operation duration data, and environmental humidity change data to obtain a final autoregressive model of temperature deviation;

[0053] Step 400: Obtain the corrected temperature data according to the final autoregressive model of temperature deviation and introduce a compensation coefficient;

[0054] Step 500: Obtain visible light image data and calculate the image sharpness score based on the high-frequency energy of the two-dimensional discrete cosine transform. When the sharpness is lower than the preset threshold, use the golden section method to adaptively adjust the lens focal length to obtain the calibrated image sharpness.

[0055] Furthermore, it also includes:

[0056] Collect the corrected temperature data within a preset time and the image sharpness data of a preset number of frames;

[0057] Compare the corrected temperature data within the preset time with the environmental reference data to obtain a first comparison result;

[0058] Compare the image sharpness data of the preset number of frames with the image sharpness of the reference image to obtain a second comparison result;

[0059] Retrain the parameters of the final temperature deviation autoregressive model according to the first comparison result;

[0060] Readjust the lens focal length according to the second comparison result.

[0061] Furthermore, optimize the initial temperature deviation autoregressive model according to the temperature gradient data, device operation duration data, and environmental humidity change data to obtain the final temperature deviation autoregressive model, including:

[0062] Normalize the temperature gradient data, device operation duration data, and environmental humidity change data respectively to obtain normalized data;

[0063] Optimize the initial temperature deviation autoregressive model according to the normalized data to obtain the final temperature deviation autoregressive model;

[0064] Specifically, environmental reference data collection and preprocessing:

[0065] The temperature deviation of the thermal imaging sensor is defined as: the measured temperature data T sensor (t) and the reference temperature T provided by the environmental reference sensor ref The difference between them:

[0066] ΔT t = T sensor (t) - T ref (t);

[0067] Historical temperature deviation can be obtained according to the above formula: ΔT t-1 , ΔT t-2 , …

[0068] In order to statistically analyze the influence of environmental humidity fluctuations on sensor drift, calculate the environmental humidity change:

[0069] ΔH t = H t - H t-1 ;

[0070] Record the cumulative operation duration of the device: D t ;

[0071] The temperature distribution difference in different regions of the environment is described by the environmental reference temperature gradient:

[0072]

[0073] Among them, is the rate of change of temperature in the horizontal direction (east-west direction); is the rate of change of temperature in the vertical direction (north-south direction), and is calculated as follows:

[0074] For a certain sensor point (x i , y i ), select the horizontally adjacent sensors (x i-1 , y i ), (x i+1 , y i ); select the vertically adjacent sensors (x i , y i-1 ), (x i , y i+1 ), and calculate:

[0075]

[0076] Adopt the multivariate autoregressive model (AR), and the initial expression of the temperature deviation autoregressive model is:

[0077]

[0078] where p is the autoregressive order, determined by the Akaike information criterion (AIC); ∈ t is zero-mean Gaussian noise.

[0079] Variables with different dimensions may vary significantly in numerical range (for example, the time span is much larger than the temperature deviation value), resulting in the dominant optimization of the loss function by the variable with a larger value during the estimation process of the model parameters (α i , β, γ, δ), and the contributions of other variables are diluted. Therefore, it is necessary to ΔH t , D t be standardized to eliminate the influence of dimensions. The following operations are performed:

[0080] For , adopt z-score standardization:

[0081]

[0082] where μ x and σ x are the mean and standard deviation of respectively.

[0083] For ΔH t , adopt Min-Max normalization:

[0084]

[0085] where ΔH max and ΔHmin are the minimum and maximum values of the historical data respectively.

[0086] For D t Perform z-score normalization:

[0087]

[0088] where, μ D and σ d are the mean and standard deviation of D t respectively.

[0089] The expression of the final temperature deviation autoregressive model is:

[0090]

[0091] where, ΔT t is the temperature deviation, p is the autoregressive order, α i , β, γ, δ are the first model parameter, the second model parameter, the third model parameter and the fourth model parameter respectively, ΔH t,scaled , D t,norm , are the normalized ambient humidity change data, equipment operation duration data and temperature gradient data respectively, ∈ t is zero-mean Gaussian noise, and i is a natural number.

[0092] Furthermore, use the benchmark data of the past 30 days and the original sensor data to solve the model parameters by the least squares method, minimizing the sum of squared residuals:

[0093] R is the total number of days.

[0094] Collect new data every 5 minutes, and use the recursive least squares method to dynamically update the parameters to adapt to the fluctuations of environmental temperature and humidity and equipment aging.

[0095] The model outputs the compensation coefficient K t , and the corrected temperature value is:

[0096] T c,t = T sensor,t - K t ·ΔT t ;

[0097] where, K t is calculated from the deviation predicted by the AR model:

[0098]

[0099] Every 5 minutes, the corrected temperature T c,tCompare with the environmental reference data. If the absolute error exceeds ±2°C, retrain the model parameters.

[0100] Furthermore, obtain visible light image data and calculate the image sharpness score based on the high-frequency energy of the two-dimensional discrete cosine transform, including:

[0101] Perform image block division on the visible light image data to obtain block data;

[0102] Perform two-dimensional DCT transformation on the block data to obtain a frequency domain coefficient matrix;

[0103] Determine the energy of the high-frequency components according to the frequency domain coefficient matrix;

[0104] Determine the image sharpness score according to the energy of the high-frequency components.

[0105] Specifically, image sharpness is the core index to measure image quality, and its numerical change has a significant correlation with the subjective perception of human eyes. Low sharpness usually shows the phenomenon of image blurring. To achieve the automatic calibration of the visible light sensor, the system first calculates and obtains the sharpness value of the current image through calculation, and this value needs to be compared with the sharpness threshold established based on the reference image - when it is detected that the sharpness of the real-time image is lower than the preset threshold, the system will automatically trigger the self-calibration program. It should be emphasized that the establishment of this threshold comes from the sharpness value of the reference image obtained by manual focusing in advance. This setting method based on the standard reference object ensures the objectivity and reliability of the calibration system.

[0106] The image energy is mainly concentrated in the middle and low-frequency regions of the amplitude spectrum, while the richness of image details and the sharpness of the contour are determined by the high-frequency components. When the image is out-of-focus blurred, its high-frequency information will be significantly attenuated, so the degree of out-of-focus can be judged by the amount of high-frequency information. Although the image sharpness evaluation function based on the Fourier transform is effective, due to the involvement of imaginary coefficients in the transformation process, the computational complexity is relatively high.

[0107] The two-dimensional discrete cosine transform (Discrete Cosine Transform, DCT) is widely used in the field of image compression. This transform can realize the conversion from the spatial domain to the frequency domain, and decompose the image signal into the weighted sum of cosine function basis signals with different frequencies. Among them, the high-frequency components correspond to the image details and edge features, and the low-frequency components correspond to the smooth regions. When the image appears blurred, its high-frequency components will decrease accordingly. Therefore, the amplitude of the high-frequency part in the DCT coefficients can be used as a measure of image sharpness. In addition, the transformation coefficients of the DCT algorithm are all real numbers, which significantly reduces the computational complexity of the image sharpness evaluation algorithm compared with the Fourier transform.

[0108] Define the image I with a resolution of W×H, and the implementation steps of its two-dimensional DCT are as follows:

[0109] Image Blocking:

[0110] Divide I(x, y) into multiple non - overlapping N×N small blocks, where N = 8. Each block is denoted as I(x, y), where x, y ∈ [0, N - 1].

[0111] Two - Dimensional DCT Transformation:

[0112] Perform two - dimensional DCT transformation on each block to obtain the frequency - domain coefficient matrix C(u, v), which is calculated as follows:

[0113]

[0114] Where:

[0115]

[0116] In the above formula, when u = 0, the DC component coefficient When u > 0, the DC component coefficient The same applies to v and α(v).

[0117] Thus, each divided N×N matrix is transformed by DCT to obtain another N×N matrix, including the DC coefficient (DC component) and the AC coefficient (AC component). The DC component represents the low - frequency component of the image and contains the main information of the image brightness; the AC component represents the middle - and high - frequency components of the image and corresponds to the edges and details in the image. Therefore, only the DC component needs to be excluded, the high - frequency region is defined, and its energy is calculated. This energy is used as a measure of the sharpness of the divided N×N matrix.

[0118] N is the size of the divided image block (N×N), and in this paper, N = 8; u is the frequency index of the image block in the horizontal direction; v is the frequency index of the image block in the vertical direction; satisfying: u, v ∈ [0, N - 1]; I(x, y) is the pixel value of the image block at (x, y), where x, y ∈ [0, N - 1].

[0119] High - Frequency Energy Calculation:

[0120]

[0121] Where E b is the energy of the high - frequency component; C(0, 0) is the DC component.

[0122] Image Sharpness Scoring:

[0123] Average the high - frequency energies of all blocks (N×N) to obtain the sharpness of the entire picture:

[0124]

[0125] Where K is the total number of N×N small blocks into which the entire image is divided; The sharpness (high-frequency energy) corresponding to each N×N small block.

[0126] The larger the S value, the more significant the high-frequency energy, the richer the details, and the clearer the image. Conversely, the smaller the S value, the more the high-frequency components are suppressed and the blurrier the image.

[0127] Furthermore, the golden section method is used to adaptively adjust the lens focal length to obtain the calibrated image sharpness, including:

[0128] Set an initial segmentation point within the preset focal length range and calculate the image sharpness at the corresponding position;

[0129] Narrow the search interval according to the sharpness comparison result until the interval accuracy meets the preset threshold;

[0130] Select the endpoint with the highest sharpness as the optimal focal length position and verify the calibrated image sharpness to obtain the calibrated image sharpness.

[0131] Specifically, an adaptive focusing based on image sharpness measurement is designed. By adjusting the focal length position of the lens, the image sharpness at each position is calculated, and the focal length position corresponding to the maximum sharpness is found.

[0132] Preset an image with good sharpness as the reference image I r , and calculate its sharpness, denoted by S(I r ). When the sharpness measurement of the real-time image is lower than 80% of the reference image sharpness, i.e.: It is determined that there is a focal length offset or surface contamination in the lens, and the adaptive calibration process is triggered:

[0133] According to the historical preset focal length adjustment range Adjustment is implemented with the goal of maximizing the image sharpness, and the golden section method is used to search for the optimal focal length position.

[0134] Initial interval Take the initial segmentation point within Respectively satisfy:

[0135]

[0136] Calculate respectively and The image sharpness under and

[0137] Update the search interval according to the following rules:

[0138] When , then the maximum value is located in the interval Update the segmentation point:

[0139]

[0140] Calculate the image sharpness under and compare them.

[0141] Conversely, when the maximum value is located in the interval Update the segmentation point:

[0142]

[0143] Calculate the image sharpness under and compare them.

[0144] Execute in a loop until the stop condition is met: Among them,

[0145] When the termination condition is met, take the midpoint of the interval or the end point with higher image sharpness as the optimal focal length position:

[0146]

[0147] Among them,

[0148] After calibration, recalculate f best the image sharpness under, and compare it with the image sharpness of the reference image. If then start secondary focusing.

[0149] Furthermore, the calibration trigger condition:

[0150] Thermal imaging sensor calibration trigger condition:

[0151] When the temperature deviation |ΔT t |> 2°C for 3 consecutive sampling periods (15 minutes), or the cumulative operating duration D of the device t ≥ 720h, force the model to be retrained (using the data of the last 30 days) for calibration.

[0152] Visible light sensor calibration trigger condition:

[0153] When the sharpness drops rapidly, when the sharpness of 10 consecutive frames of images is less than 80% of the sharpness of the reference image, that is: S(I) < 0.8·S(I r ), trigger the self-calibration of the visible light sensor.

[0154] Historical data update:

[0155] Adopt a sliding window to store the calibration data of the most recent 30 days, and eliminate the data outside the window in chronological order to ensure that the model training data reflects the characteristics of the current environment;

[0156] Filter outliers from the environmental reference sensor data (such as thermocouple temperature, temperature and humidity). The IQR method is used for outlier filtering to avoid pulse noise contaminating the training set. Specifically, the interquartile range of the data is determined by calculating the difference between the first quartile (Q1, 25% quantile) and the third quartile (Q3, 75% quantile) of the data (IQR = Q3 - Q1), and the outlier boundary is defined based on this. Data points below Q1 - 1.5×IQR or above Q3 + 1.5×IQR are determined as outliers.

[0157] The various embodiments in this specification are described in a progressive manner. Each embodiment focuses on the differences from other embodiments. For the same or similar parts between the various embodiments, reference can be made to each other.

[0158] Specific examples are used in this article to elaborate on the principles and implementation manners of the present invention. The descriptions of the above embodiments are only used to help understand the method and its core idea of the present invention; at the same time, for those of ordinary skill in the art, there will be changes in the specific implementation manners and application scopes according to the idea of the present invention. In summary, the content of this specification should not be construed as a limitation to the present invention.

Claims

1. A multi-modal sensor drift self-calibration method in a complex industrial environment, characterized in that, Including: Real-time collecting temperature gradient data, equipment operation duration data, and environmental humidity change data in the industrial scenario through a deployed environmental benchmark sensor group; Constructing an initial autoregressive model of temperature deviation using a multivariate autoregressive model; Optimizing the initial autoregressive model of temperature deviation according to the temperature gradient data, equipment operation duration data, and environmental humidity change data to obtain a final autoregressive model of temperature deviation; Obtaining corrected temperature data according to the final autoregressive model of temperature deviation and introducing a compensation coefficient; Obtaining visible light image data and calculating an image sharpness score based on the high-frequency energy of a two-dimensional discrete cosine transform. When the sharpness is lower than a preset threshold, adaptively adjusting the lens focal length using the golden section method to obtain calibrated image sharpness.

2. A multi-modal sensor drift self-calibration method in a complex industrial environment according to claim 1, characterized in that, Also including: Collecting the corrected temperature data within a preset time and the image sharpness data of a preset number of frames; Comparing the corrected temperature data within a preset time with the environmental benchmark data to obtain a first comparison result; Comparing the image sharpness data of a preset number of frames with the image sharpness of a reference image to obtain a second comparison result; Re-training the parameters of the final autoregressive model of temperature deviation according to the first comparison result; Re-adjusting the lens focal length according to the second comparison result.

3. A multi-modal sensor drift self-calibration method in a complex industrial environment according to claim 1, characterized in that, Optimizing the initial autoregressive model of temperature deviation according to the temperature gradient data, equipment operation duration data, and environmental humidity change data to obtain a final autoregressive model of temperature deviation, including: Normalizing the temperature gradient data, equipment operation duration data, and environmental humidity change data respectively to obtain normalized data; Optimizing the initial autoregressive model of temperature deviation according to the normalized data to obtain a final autoregressive model of temperature deviation, where the expression of the final autoregressive model of temperature deviation is: Among them, ΔT t is the temperature deviation, p is the autoregressive order, α i , β, γ, δ are the first model parameter, the second model parameter, the third model parameter and the fourth model parameter respectively, ΔH t,scaled , D t,norm , are the normalized environmental humidity change data, the equipment operation duration data and the temperature gradient data respectively, ∈ t is zero-mean Gaussian noise, and i is a natural number.

4. A multi-modal sensor drift self-calibration method in a complex industrial environment according to claim 3, characterized in that The expression of the corrected temperature data is: T c,t = T sensor,t - K t ·ΔT t ; Among them, T c,t is the corrected temperature data, T sensor,t is the measured temperature data, K t is the output compensation coefficient.

5. A multi-modal sensor drift self-calibration method in a complex industrial environment according to claim 3, characterized in that, Obtaining visible light image data and calculating an image sharpness score based on the high-frequency energy of a two-dimensional discrete cosine transform, including: Performing image block division on the visible light image data to obtain block data; Performing a two-dimensional DCT transform on the block data to obtain a frequency domain coefficient matrix; Determining the energy of the high-frequency components according to the frequency domain coefficient matrix; Determining the image sharpness score according to the energy of the high-frequency components.

6. A multi-modal sensor drift self-calibration method in a complex industrial environment according to claim 3, characterized in that The expression of the frequency domain coefficient matrix is: Where C(u, v) is the frequency domain coefficient matrix, α(u) is the DC component coefficient, α(v) is the AC component coefficient; u is the frequency index of the image block in the horizontal direction; v is the frequency index of the image block in the vertical direction; I(x, y) is the pixel value of the image block at (x, y), and N is the number of side pixels of each sub-block.

7. A multi-modal sensor drift self-calibration method in a complex industrial environment according to claim 6, characterized in that The expression of the energy of the high-frequency components is: Among them, E b is the energy of the high-frequency component.

8. A multi-modal sensor drift self-calibration method in a complex industrial environment according to claim 6, characterized in that, Adapting to adjust the lens focal length using the golden section method to obtain calibrated image sharpness, including: Setting an initial segmentation point within a preset focal length range and calculating the image sharpness at the corresponding position; Narrowing the search interval according to the sharpness comparison result until the interval accuracy meets the preset threshold; Selecting the endpoint with the highest sharpness as the best focal length position and verifying the calibrated image sharpness to obtain the calibrated image sharpness.

Images