A deep learning-based optical fiber temperature measurement signal compensation method and system

Abstract

The application provides a kind of based on deep learning's optical fiber temperature measuring signal compensation method and system, it is related to deep learning optical fiber sensing technical field, the application first acquires the original spectrum data of stokes and anti-stokes channel in distributed optical fiber temperature monitoring system, obtains original monitoring data;Again it is converted into electric signal, after being preprocessed by programmable gain analog front-end circuit, it is converted into digital signal, and digital spectrum data is obtained;Then input pre-training convolutional neural network, extract temperature-sensitive spectral deep features by alternating convolution layer and pooling layer;Subsequently input pre-training adversarial generative network, optimize feature distribution in latent space to suppress strong electromagnetic interference abnormal features;Finally, through the full connection neural network layer, the temperature feature vector is converted, and the continuous temperature data output is mapped by the temperature solution function, can effectively process optical fiber temperature measuring signal, suppress strong electromagnetic interference, accurately output continuous temperature data, realize optical fiber temperature measuring signal compensation.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of deep learning optical fiber sensing, and in particular to an optical fiber temperature measurement signal compensation method and system based on deep learning. BACKGROUND

[0002] In an industrial environment with strong electromagnetic interference (such as a power substation or a metallurgical electric arc furnace), the surface temperature of an industrial device needs to be monitored in real time and accurately to prevent overheating damage or safety accidents. In such a scenario, the photoelectric detector is affected by electromagnetic radiation, and the output signal is prone to nonlinear distortion, causing the original temperature measurement data to deviate from the true value. Therefore, there is an urgent need for an optical fiber temperature measurement signal compensation technology that can effectively suppress electromagnetic interference, correct signal distortion, and adapt to the needs of industrial dynamic monitoring, to ensure the accuracy and reliability of temperature monitoring data.

[0003] The mainstream existing scheme for this scenario mainly uses digital filtering and linear fitting technology. The digital spectrum data output by the photoelectric detector is denoised by Kalman filtering or wavelet filtering, and then the deviation caused by nonlinear distortion in the filtered data is fitted and corrected based on a predetermined linear mathematical model. Finally, the device surface temperature is output in combination with the temperature calculation formula, in an attempt to improve the temperature measurement accuracy in a strong electromagnetic interference environment.

[0004] This existing scheme has obvious technical limitations. On the one hand, traditional digital filtering can only filter out some high-frequency electromagnetic interference noise, and cannot identify and process the nonlinear distortion of the photoelectric detector caused by strong electromagnetic radiation, resulting in residual distorted signals and affecting the accuracy of subsequent temperature calculation. On the other hand, the linear fitting model relies on fixed mathematical assumptions and cannot adapt to the complex signal distortion patterns caused by fluctuations in electromagnetic interference intensity and dynamic changes in device surface temperature in industrial environments. When the interference intensity increases sharply or the temperature fluctuates sharply, the compensation effect decreases significantly, and the high-precision and high-stability temperature monitoring requirements of industrial devices cannot be met. SUMMARY

[0005] The present application aims to provide an optical fiber temperature measurement signal compensation method and system based on deep learning to solve the problem that the compensation scheme based on traditional digital filtering and linear fitting in the prior art cannot handle the nonlinear distortion of the photoelectric detector caused by strong electromagnetic interference, and cannot adapt to the dynamic changes in interference intensity and temperature, resulting in insufficient compensation accuracy.

[0006] To solve the above technical problems, in a first aspect, the present application provides an optical fiber temperature measurement signal compensation method based on deep learning, comprising:

[0007] Collecting the original spectrum data of the Stokes channel and the anti-Stokes channel in a distributed optical fiber temperature monitoring system to obtain original monitoring data containing intensity distance relationship and spectral morphological features;

[0008] convert the original monitoring data into an electrical signal, pre-process the electrical signal by using an analog front-end circuit with programmable gain, and convert the pre-processed electrical signal into a digital signal to obtain digitized spectral data containing Stokes channel and anti-Stokes channel intensity information;

[0009] input the digitized spectral data into a pre-trained convolutional neural network model, process through a plurality of alternately arranged convolutional layers and pooling layers in the convolutional neural network model, and extract temperature-sensitive spectral deep features as deep feature data;

[0010] input the deep feature data into a pre-trained generative adversarial network model, perform feature distribution optimization on the deep feature data in the latent space through the generative adversarial network model, suppress abnormal feature patterns caused by strong electromagnetic interference, and generate optimized feature representations;

[0011] convert the optimized feature representations into temperature feature vectors in the temperature domain through a fully connected neural network layer, and map the temperature feature vectors to continuous temperature data as temperature calculation results based on a temperature calculation function.

[0012] Optionally, the inputting the deep feature data into a pre-trained generative adversarial network model, performing feature distribution optimization on the deep feature data in the latent space through the generative adversarial network model, and suppressing abnormal feature patterns caused by strong electromagnetic interference to generate optimized feature representations, comprises:

[0013] input the deep feature data into an encoder part of a pre-trained generator network, map the input features to the latent space through a plurality of nonlinear transformations, and generate initial latent feature representations;

[0014] perform distribution optimization adjustment on the initial latent feature representations in the latent space, make the feature distribution close to the normal temperature feature distribution, and perform abnormal feature suppression processing to eliminate abnormal feature patterns that differ from the normal temperature feature distribution by more than a preset threshold;

[0015] perform feature reconstruction on the optimized and processed initial latent feature representations through a decoder part of the generator network, and adopt layer-by-layer upsampling and feature fusion operations to convert the optimized and processed initial latent feature representations into optimized feature representations with the same dimension as the original deep feature data.

[0016] Optionally, the performing distribution optimization adjustment on the initial latent feature representations in the latent space, making the feature distribution close to the normal temperature feature distribution, and performing abnormal feature suppression processing to eliminate abnormal feature patterns that differ from the normal temperature feature distribution by more than a preset threshold, comprises:

[0017] Calculate the statistical difference between the initial latent feature representation and the normal temperature feature distribution in the latent space to obtain the feature distribution deviation index;

[0018] Based on the characteristic distribution deviation index, the distribution parameters in the initial potential characteristic representation are adjusted so that the adjusted characteristic distribution approaches the normal temperature characteristic distribution.

[0019] In the adjusted initial latent feature representation, abnormal feature points that deviate from the mean of the normal temperature feature distribution by more than a preset threshold are identified.

[0020] The abnormal feature points are suppressed, and their feature value is adjusted to the normal temperature feature distribution range through feature value correction operation.

[0021] Optionally, the step of inputting the digitized spectral data into a pre-trained convolutional neural network model, and processing it through multiple alternating convolutional and pooling layers in the convolutional neural network model to extract temperature-sensitive deep spectral features as deep feature data includes:

[0022] The digitized spectral data is organized into an input tensor according to the channel dimension and then input into the pre-trained convolutional neural network model.

[0023] The input tensor is used to extract local features through the first convolutional layer. The convolutional kernel of a preset size is used to perform sliding calculations on the tensor to generate a first initial feature map containing primary features.

[0024] The first initial feature map is condensed by the first pooling layer, and the feature map dimension is reduced by the maximum value selection method to generate the first condensed feature map.

[0025] The first condensed feature map is input into the second convolutional layer for deep feature extraction. Convolutional kernels of different sizes are used to extract more complex feature patterns and generate a second initial feature map.

[0026] The second initial feature map is further condensed by a second pooling layer. The feature dimension is reduced by calculating the region average value to generate a second condensed feature map.

[0027] By repeatedly performing alternating convolution and pooling operations, higher-level feature representations are extracted step by step, and the output is deep feature data containing temperature-sensitive components in the spectral data.

[0028] Optionally, the step of converting the optimized feature representation into a temperature feature vector in the temperature domain through a fully connected neural network layer, and mapping the temperature feature vector to continuous temperature data based on a temperature calculation function as the output temperature calculation result, includes:

[0029] inputting the optimized feature representation into a first transformation unit of a fully connected neural network layer, generating an intermediate feature representation through linear combination operation of a weight matrix and a bias vector;

[0030] inputting the intermediate feature representation into a nonlinear activation unit, obtaining an activated feature representation with enhanced expression capability through feature transformation by a preset nonlinear function;

[0031] directly inputting the activated feature representation into a second transformation unit of the fully connected neural network layer, converting the high-dimensional feature into a temperature feature vector with temperature representation capability through dimension reduction mapping operation;

[0032] inputting the temperature feature vector into a temperature solving function, mapping each feature component in the feature vector to a temperature value of a corresponding fiber distance point through component weighted summation calculation;

[0033] arranging and combining the temperature values of the fiber distance points in ascending order of distance to form a complete continuous temperature data sequence distributed along the fiber, and outputting the continuous temperature data sequence as the final temperature solving result.

[0034] Optionally, the original monitoring data is converted into an electrical signal, the electrical signal is preprocessed by using an analog front-end circuit with programmable gain, and the preprocessed electrical signal is converted into a digital signal to obtain digital spectrum data containing Stokes channel and anti-Stokes channel intensity information, including:

[0035] the Stokes light intensity distribution sequence and the anti-Stokes light intensity distribution sequence in the original monitoring data are input into independent photoelectric converters respectively to generate analog electrical signals corresponding to the Stokes channel and the anti-Stokes channel;

[0036] The analog electrical signals of each channel are input into a programmable gain amplifier in the gain programmable analog front-end circuit, and the amplification multiple is dynamically adjusted according to the monitored signal amplitude value, so that the output signal amplitude is stabilized within a preset working range;

[0037] The gain-adjusted analog electrical signals of each channel are input into a programmable filter group in the gain programmable analog front-end circuit, and the filter parameters are adaptively adjusted according to the signal frequency band characteristics to retain the effective signal frequency band components and suppress the out-of-band noise;

[0038] The filtered analog electrical signals of each channel are sent into a high-precision analog-to-digital converter in the gain programmable analog front-end circuit, and are synchronously sampled at a sampling rate not less than twice the highest frequency of the signal, and the sampling values are quantized into digital signals;

[0039] Adding corresponding channel identification codes to the digital signals of the Stokes channel and the anti-Stokes channel, and aligning and combining them according to the same sampling time stamp, to generate digital spectrum data containing two-channel intensity information.

[0040] Optionally, the original spectrum data of the Stokes channel and the anti-Stokes channel in the distributed optical fiber temperature monitoring system is collected to obtain original monitoring data containing intensity-distance relationship and spectral morphological features, including:

[0041] The original spectrum data of the Stokes channel and the anti-Stokes channel is synchronously collected by a signal collection unit of the distributed optical fiber temperature monitoring system;

[0042] The original spectrum data of the Stokes channel is sequentially extracted according to the fiber distance points to form a Stokes light intensity distribution sequence;

[0043] The original spectrum data of the anti-Stokes channel is sequentially extracted according to the same fiber distance points to form an anti-Stokes light intensity distribution sequence;

[0044] The Stokes light intensity distribution sequence and the anti-Stokes light intensity distribution sequence are correspondingly combined according to the fiber distance points to form original monitoring data containing both two-channel intensity information and distance corresponding relationship.

[0045] In a second aspect, the present application provides a fiber temperature measurement signal compensation system based on deep learning, comprising:

[0046] A collection module is configured to collect original spectrum data of a Stokes channel and an anti-Stokes channel in a distributed optical fiber temperature monitoring system, and obtain original monitoring data containing intensity-distance relationship and spectral morphological features;

[0047] A processing module is configured to convert the original monitoring data into an electrical signal, pre-process the electrical signal by using an analog front-end circuit with programmable gain, and convert the pre-processed electrical signal into a digital signal to obtain digital spectrum data containing Stokes channel and anti-Stokes channel intensity information;

[0048] An extraction module is configured to input the digital spectrum data into a pre-trained convolutional neural network model, process the digital spectrum data by a plurality of alternately arranged convolutional layers and pooling layers in the convolutional neural network model, and extract temperature-sensitive spectral deep features as deep feature data;

[0049] A generation module is configured to input the deep feature data into a pre-trained generative adversarial network model, perform feature distribution optimization on the deep feature data in a latent space by the generative adversarial network model, suppress abnormal feature patterns caused by strong electromagnetic interference, and generate an optimized feature representation.

[0050] an output module, configured to convert the optimized feature representation into a temperature feature vector in a temperature domain through a fully connected neural network layer, and map the temperature feature vector into continuous temperature data as a temperature solving result based on a temperature solving function.

[0051] In a third aspect, the present application provides an electronic device, comprising:

[0052] a memory, configured to store a computer program;

[0053] a processor, configured to implement the steps of the deep learning-based optical fiber temperature measurement signal compensation method according to the first aspect when the computer program is executed.

[0054] In a fourth aspect, the present application provides a computer readable storage medium, wherein the computer readable storage medium stores a computer program, and the computer program can implement the steps of the deep learning-based optical fiber temperature measurement signal compensation method according to the first aspect when the computer program is executed by a processor.

[0055] The deep learning-based optical fiber temperature measurement signal compensation method provided by the present application acquires basic data containing intensity distance relationship and spectral morphology by collecting Stokes and anti-Stokes channel original spectrum data in a distributed optical fiber temperature monitoring system, ensuring the integrity of the temperature measurement data; converts the original data into electrical signals, pre-processes and digitizes the electrical signals through a programmable gain analog front-end circuit, and completes the preliminary optimization of the signals; extracts temperature-sensitive spectrum deep features by using a pre-trained convolutional neural network, and then suppresses abnormal patterns caused by electromagnetic interference through a pre-trained generative adversarial network; finally, converts the temperature feature vector through a fully connected neural network layer, and outputs continuous temperature data through a temperature solving function, thereby realizing accurate and stable monitoring of the surface temperature of industrial equipment in a strong electromagnetic interference environment, and improving the reliability and precision of the optical fiber temperature measurement system.

[0056] Further, the deep feature data is input into a pre-trained adversarial generation network model to generate optimized feature representation: first, the data is mapped to the initial latent feature in the latent space through the multi-layer nonlinear transformation of the generator encoder; then, the distribution in the latent space is adjusted, and the abnormality is suppressed to make the features close to the normal temperature distribution and eliminate abnormal patterns with too large differences; finally, the optimized features with the same dimension as the original data are reconstructed through upsampling and feature fusion of the decoder. The process realizes effective mapping of deep features to the latent space through nonlinear transformation of the generator encoder, provides a reasonable spatial basis for feature optimization; with the help of distribution adjustment and abnormality suppression in the latent space, abnormal features caused by strong electromagnetic interference are accurately eliminated to ensure that the feature distribution conforms to the normal temperature data rule; and the feature reconstruction is completed through upsampling and feature fusion of the decoder, which not only ensures that the optimized features have the same dimension as the original deep features and can be adapted to subsequent processing, but also further improves the purity and temperature correlation of the features, providing high-quality feature support for subsequent accurate conversion of temperature feature vectors and output of reliable temperature data. BRIEF DESCRIPTION OF DRAWINGS

[0057] In order to more clearly illustrate the technical solutions of the embodiments of the present application or the prior art, the drawings needed to be used in the embodiments or the prior art description will be briefly introduced. Obviously, the drawings in the following description are only some embodiments of the present application, and those skilled in the art can obtain other drawings according to these drawings without creative labor.

[0058] Figure 1 A flowchart of a deep learning-based optical fiber temperature measurement signal compensation method provided by an embodiment of the present application;

[0059] Figure 2 A specific implementation flowchart of a deep learning-based optical fiber temperature measurement signal compensation method provided by an embodiment of the present application;

[0060] Figure 3 A specific implementation structure diagram of a deep learning-based optical fiber temperature measurement signal compensation method provided by an embodiment of the present application;

[0061] Figure 4 A structure diagram of a deep learning-based optical fiber temperature measurement signal compensation system provided by an embodiment of the present application. DETAILED DESCRIPTION

[0062] In the surface temperature monitoring scene of industrial equipment under strong electromagnetic interference, the existing optical fiber temperature measurement signal compensation scheme based on traditional digital filtering and linear fitting has significant limitations: on the one hand, traditional digital filtering can only filter out part of the high-frequency electromagnetic interference noise, and cannot identify and process the nonlinear distortion of the photodetector caused by strong electromagnetic radiation, resulting in residual distortion signals and affecting the temperature calculation accuracy; on the other hand, the linear fitting model relies on fixed mathematical assumptions and cannot adapt to the complex signal distortion patterns caused by the fluctuation of electromagnetic interference intensity and the dynamic change of equipment surface temperature in industrial environment. When the interference intensity increases sharply or the temperature fluctuates sharply, the compensation effect decreases significantly, which cannot meet the high-precision and high-stability temperature monitoring requirements of industrial equipment, and may even cause safety hazards due to temperature measurement errors.

[0063] To solve the above problems, the present application provides a fiber temperature measurement signal compensation method based on deep learning. The method first collects the original spectrum data of the Stokes and anti-Stokes channels in the distributed optical fiber temperature monitoring system, and after preprocessing and digitizing by the programmable gain analog front-end circuit, inputs the pre-trained convolutional neural network to extract the temperature-sensitive spectrum deep features, and then optimizes the feature distribution in the latent space through the pre-trained generative adversarial network to suppress the electromagnetic interference abnormal features. Finally, the continuous temperature data is output through the fully connected neural network and the temperature calculation function. This scheme accurately captures temperature-related deep features through convolutional neural networks, breaking through the limitations of traditional filtering that cannot handle nonlinear distortion; with the help of adversarial network to dynamically optimize the feature distribution, it solves the problem that linear fitting cannot adapt to complex interference and temperature changes, fundamentally eliminates the influence of strong electromagnetic interference on the temperature measurement signal, and significantly improves the accuracy and stability of industrial equipment surface temperature monitoring, meeting the monitoring requirements in complex industrial environments.

[0064] In order to enable personnel in the technical field to better understand the present application scheme, the present application will be further described in detail below in combination with the drawings and specific embodiments. Obviously, the described embodiments are only part of the embodiments of the present application, not all. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative labor are within the scope of protection of the present application.

[0065] The core of the present application is to provide a fiber temperature measurement signal compensation method based on deep learning, and a specific implementation process of the method is shown in Figure 1 The method comprises the following steps:

[0066] S101, collect the original spectrum data of the Stokes channel and the anti-Stokes channel in the distributed optical fiber temperature monitoring system, and obtain the original monitoring data containing intensity distance relationship and spectrum morphological features;

[0067] Optionally, step S101 can specifically include the following steps:

[0068] S1011, synchronously collecting original spectrum data of the Stokes channel and the anti-Stokes channel by a signal acquisition unit of the distributed optical fiber temperature monitoring system;

[0069] S1012, sequentially extracting light intensity values according to fiber distance points from the original spectrum data of the Stokes channel to form a Stokes light intensity distribution sequence;

[0070] S1013, sequentially extracting light intensity values according to the same fiber distance points from the original spectrum data of the anti-Stokes channel to form an anti-Stokes light intensity distribution sequence;

[0071] S1014, corresponding combining the Stokes light intensity distribution sequence and the anti-Stokes light intensity distribution sequence according to fiber distance points to form original monitoring data containing both the intensity information of the two channels and the corresponding relationship with the distance.

[0072] In the above scheme, the original spectrum data refers to the basic light signal data directly obtained from the Stokes channel and the anti-Stokes channel in the distributed optical fiber temperature monitoring process without subsequent processing. The Stokes light intensity distribution sequence is an ordered set of light intensity values of each distance point of the Stokes channel arranged in order of fiber distance points. The anti-Stokes light intensity distribution sequence is an ordered set of light intensity values of each distance point of the anti-Stokes channel arranged in order of the same fiber distance points. The original monitoring data is a comprehensive data set containing both the intensity information of the Stokes channel and the anti-Stokes channel and the corresponding relationship with the fiber distance points after corresponding combination.

[0073] In the embodiment of the present application, first, the signal acquisition unit of the distributed optical fiber temperature monitoring system is connected with the optical fiber laid on the monitoring object through step S1011, to ensure that the signal acquisition unit can receive the optical signals transmitted by the optical fiber. Then, the synchronous acquisition function of the signal acquisition unit is started to receive the Stokes channel optical signals and the anti-Stokes channel optical signals from the optical fiber at the same time period. Then, the signal acquisition unit will convert the received two kinds of optical signals into corresponding original spectrum data, that is, record the intensity of each optical signal at different wavelengths. For example, when monitoring a device, the signal acquisition unit receives two kinds of optical signals at the same time within 1 minute, converts the Stokes channel optical signals into original spectrum data containing wavelengths of 800nm-1600nm and corresponding intensities, and also converts the anti-Stokes channel optical signals into original spectrum data of the same wavelength range and corresponding intensities, thereby completing the synchronous acquisition.

[0074] Secondly, the Stokes channel raw spectrum data synchronously collected in step S1011 is acquired through step S1012, and the data records the light intensity corresponding to different wavelengths; then the distance points on the optical fiber are determined, such as the four points of 1 meter, 2 meters, 3 meters and 4 meters; then a fixed wavelength reflecting temperature information is selected from the raw spectrum data, such as 1300 nm, and the light intensity value of each distance point at the wavelength is extracted in sequence, for example, the intensity value of the 1-meter point at the wavelength of 1300 nm is 110, that of the 2-meter point is 115, that of the 3-meter point is 120, and that of the 4-meter point is 118; finally, the extracted intensity values are arranged in the order of distance points from near to far to form a Stokes light intensity distribution sequence, such as [110, 115, 120, 118].

[0075] Then, the anti-Stokes channel raw spectrum data synchronously collected in step S1011 is acquired through step S1013, and the data also records the light intensity corresponding to different wavelengths; then the same distance points on the optical fiber as those in step S1012 are determined, such as the same 1-meter, 2-meter and 3-meter points; then the same reference wavelength as that in step S1012 is selected from the raw spectrum data, such as 1300 nm, and the light intensity value of each same distance point at the wavelength is extracted in sequence, for example, the intensity value of the 1-meter point at the wavelength of 1300 nm is 85, that of the 2-meter point is 90, and that of the 3-meter point is 95; finally, the extracted intensity values are arranged in the same order of distance points as in step S1012 to form an anti-Stokes light intensity distribution sequence, such as [85, 90, 95].

[0076] Then, through step S1014, the Stokes light intensity distribution sequence and the anti-Stokes light intensity distribution sequence are acquired first, and it is confirmed that the order of distance points on the optical fiber of the two sequences is completely consistent, and they are combined into a pair of data, for example, the 1-meter point in the Stokes sequence is 110, and the 1-meter point in the anti-Stokes sequence is 85, which are combined into (110, 85); then the intensity values corresponding to the second distance point and the third distance point are sequentially combined in the same way, and finally all the combined paired data are arranged in the order of distance points to constitute the original monitoring data containing the intensity information of the two channels and the distance corresponding relationship, for example, [(110, 85), (115, 90), (120, 95)].

[0077] In practical applications, when monitoring the surface temperature of a certain type of industrial equipment, first, the signal acquisition unit of the distributed optical fiber temperature monitoring system is started, and the original spectral data of the Stokes channel and the anti-Stokes channel on the optical fiber of the equipment are synchronously collected; then, the light intensity values of each point are extracted from the Stokes channel original spectral data according to the distance points of 1 meter, 2 meters, 3 meters, 4 meters and 5 meters on the optical fiber to form a Stokes light intensity distribution sequence of [120, 125, 130, 128, 126]; again, the light intensity values of each point are extracted from the anti-Stokes channel original spectral data using the same distance points of 1 meter to 5 meters to form an anti-Stokes light intensity distribution sequence of [100, 105, 108, 106, 104]; finally, the original monitoring data of [(120, 100), (125, 105), (130, 108), (128, 106), (126, 104)] is obtained by corresponding combination of the distance points, which is used for subsequent processing.

[0078] The overall scheme of S101 integrates the originally separated single-channel data into comprehensive data containing double-channel information by corresponding combination of the intensity sequences of the two channels according to the distance points, so that the temperature-related information of each monitoring position is more comprehensive; at the same time, the distance corresponding relationship is clearly reserved, so that the data can be accurately associated to the specific position of the equipment surface, avoiding data confusion; and finally, the original monitoring data formed provides a data basis with clear structure and complete information for subsequent electrical signal conversion, preprocessing and deep learning model processing, ensuring that each subsequent link can work based on accurate and comprehensive data, and guaranteeing the accuracy and reliability of temperature monitoring.

[0079] S102, converting the original monitoring data into electrical signals, pre-processing the electrical signals by using a programmable gain analog front-end circuit, and converting the pre-processed electrical signals into digital signals to obtain digitized spectral data containing Stokes channel and anti-Stokes channel intensity information;

[0080] Optionally, step S102 can specifically include the following steps:

[0081] S1021, inputting the Stokes light intensity distribution sequence and the anti-Stokes light intensity distribution sequence in the original monitoring data into independent photoelectric converters respectively to generate analog electrical signals corresponding to the Stokes channel and the anti-Stokes channel;

[0082] S1022, inputting the analog electrical signals of each channel into a programmable gain amplifier in a gain programmable analog front-end circuit, and dynamically adjusting the amplification multiple according to the signal amplitude value of the monitoring to make the output signal amplitude stable within a preset working range;

[0083] S1023, input the gain-adjusted analog electrical signals of each channel into a programmable filter set in the gain-programmable analog front-end circuit, adaptively adjust the filter parameters according to the signal frequency band characteristics, retain the effective signal frequency band components, and suppress the out-of-band noise;

[0084] S1024, send the filtered analog electrical signals of each channel into a high-precision analog-to-digital converter in the gain-programmable analog front-end circuit, perform synchronous sampling at a sampling rate not lower than twice the highest frequency of the signal, and quantize the sampling values into digital signals;

[0085] S1025, add corresponding channel identification codes to the digital signals of the Stokes channel and the anti-Stokes channel, and align and combine them according to the same sampling time stamp to generate digital spectral data containing two-channel intensity information.

[0086] In the above scheme, the analog electrical signal refers to a signal whose current or voltage changes continuously over time, which can reflect the change trend of the original intensity data. The programmable gain analog front-end circuit is a circuit that can flexibly adjust the signal amplification degree and preliminarily optimize the original electrical signal. The digital signal refers to a discrete digital coded signal corresponding to the Stokes channel and the anti-Stokes channel, which contains the digital characteristics of the two-channel analog electrical signals after sampling and quantization, including 0-1 coded information generated at a preset sampling rate, which can be used for subsequent addition of identification and combination into digital spectral data. The digital spectral data is spectral-related data containing two-channel intensity information and existing in digital form.

[0087] In the embodiments of the present application, first, the Stokes light intensity distribution sequence and the anti-Stokes light intensity distribution sequence are split from the original monitoring data in step S1021, ensuring that the distance point orders of the two sequences are consistent. Then, two independent photoelectric converters are prepared, the Stokes light intensity distribution sequence is input into the first photoelectric converter one by one, and the converter outputs corresponding continuous electrical signals according to the size of each intensity value; at the same time, the anti-Stokes light intensity distribution sequence is input into the second photoelectric converter one by one, and the same continuous electrical signals are output according to the size of the intensity value; finally, the analog electrical signals corresponding to the two channels are obtained. For example, when monitoring a certain type of equipment, the Stokes sequence [105, 110, 112] is input into the first converter, and the analog electrical signals [0.525V, 0.55V, 0.56V] are output, and the anti-Stokes sequence [92, 96, 94] is input into the second converter, and the analog electrical signals [0.46V, 0.48V, 0.47V] are output.

[0088] Secondly, by step S1022, the two channel analog signals obtained in step S1021 are respectively connected to the programmable gain amplifier in the programmable gain analog front-end circuit; then each amplifier will first monitor the amplitude value of the connected signal; then according to the comparison between the monitored amplitude value and the preset working range, the amplification multiple is adjusted: if the signal amplitude is lower than the lower limit of the preset range, the amplification multiple is increased so that the amplitude after amplification enters the range; if the signal amplitude is higher than the upper limit of the preset range, the amplification multiple is decreased so that the amplitude also enters the range; finally, the analog signal with the amplitude stabilized in the preset range is output. For example, the preset range is 0.5-2V, the Stokes channel original signal is 0.4V, the original signal amplitude is low, and after amplification by 1.3 times, it becomes 0.52V, and the anti-Stokes original signal is 2.1V, and after amplification by 0.95 times, it becomes 2.0V, the amplitudes are all stabilized in 0.5-2V.

[0089] Then, by step S1023, the two channel analog signals with the amplitudes stabilized are respectively input to the programmable filter group in the programmable gain analog front-end circuit; then each filter group will first analyze the frequency band characteristics of the connected signal to find out the frequency range of the useful signal set; then according to this frequency band characteristics, the filter parameter is adjusted, the allowed pass frequency range of the filter is set to be consistent with the useful signal frequency range; the out-of-band noise whose frequency exceeds the allowed pass frequency range is filtered out through the filtering operation of the filter. For example, when monitoring a certain device, the Stokes channel signal mixes 40Hz noise and 6000Hz high frequency interference, after adjusting the filter parameter to 120-900Hz, the 40Hz and 6000Hz noise is filtered out, only the 120-900Hz useful signal is left, and the anti-Stokes channel is the same, and a purer analog electrical signal is obtained.

[0090] Then, by step S1024, the two channel analog signals after filtering noise are respectively connected to the high-precision analog-to-digital converter in the programmable gain analog front-end circuit; the highest frequency of the useful signal of each channel analog signal is first detected; then the sampling rate is set to be not less than twice the highest frequency; then the synchronous sampling is started, and the analog signals of the two channels are collected, and the sampling value obtained each time is converted into a digital signal through quantization. For example, when monitoring a certain device, the highest frequency of the signal is 800Hz, the sampling rate is set to 1600 times / s, 1600 times are collected per second, and the voltage value collected each time is converted into an 8-bit digital signal, and finally the digital signal sequence of the two channels is obtained.

[0091] Finally, by step S1025, the Stokes channel digital signal and the anti-Stokes channel digital signal generated in step S1024 are obtained, and it is confirmed that the sampling time stamps of the two signal sequences are one-to-one corresponding; then a special channel identification code such as "ST-10100110" is added to each digital signal of the Stokes channel, and another special identification code such as "ANTI-ST-01100000" is added to each digital signal of the anti-Stokes channel; then the digital signals of the two channels are aligned and combined according to the same sampling time stamp; finally, all the aligned and combined signals are sorted in time stamp order to generate digital spectral data containing double-channel intensity information. For example, when monitoring a certain type of equipment, the combined data sequence is "time 1: (ST-10100110, ANTI-ST-01100000), time 2: (ST-10100111, ANTI-ST-01100001), …", which completely retains the information and time correspondence of the two channels.

[0092] In actual application, when monitoring the surface temperature of an industrial equipment, the Stokes light intensity distribution sequence [105, 112, 118, 116] and the anti-Stokes light intensity distribution sequence [82, 88, 92, 90] obtained previously are input into two independent photoelectric converters to obtain two-channel analog signals; then the two analog signals are input into a programmable gain amplifier, and it is found that the Stokes channel signal amplitude is 0.25V, which is lower than the preset lower limit of 0.4V, and the amplification factor is immediately adjusted from 10 times to 18 times, and the anti-Stokes channel signal amplitude is 2.3V, which is higher than the preset upper limit of 2.1V, and the amplification factor is immediately adjusted from 15 times to 13 times, so that the amplitudes of the two are stabilized at 0.4-2.1V; then the adjusted signals are input into a programmable filter group, and it is analyzed that the useful signals are concentrated in the range of 120-900Hz, so the noise outside this range is filtered out by adjusting the parameters; the highest frequency of the filtered signal is 850Hz, and a sampling rate of 1850 times / s is selected to convert the filtered signal into a digital signal; finally, the Stokes channel digital signal is marked as "ST02", the anti-Stokes channel is marked as "ANTI-ST02", and they are combined according to the same time stamp to obtain digital spectral data containing double-channel intensity information, which is used for subsequent deep learning processing.

[0093] The overall scheme of S102 converts the original monitoring data into electrical signals, completes the type conversion from optical correlation data to electrical signals, and prepares for circuit processing; the programmable gain amplifier is used to stabilize the signal amplitude and prevent the influence of subsequent processing caused by too weak or too strong; the filter set is used to filter out noise and improve signal purity; the analog-to-digital conversion is used to convert the analog signal into a digital signal to adapt to the processing requirements of the deep learning model; finally, the identification is added and the time alignment is performed to clarify the corresponding relationship of the double-channel data, generate digital spectral data with clear structure and reliable information, and provide high-quality data support for temperature feature extraction.

[0094] S103, input the digital spectral data into a pre-trained convolutional neural network model, process the digital spectral data through a plurality of alternately arranged convolutional layers and pooling layers in the convolutional neural network model, and extract temperature-sensitive spectral deep features as deep feature data;

[0095] Optionally, step S103 can specifically include the following steps:

[0096] S1031, organize the digital spectral data into an input tensor according to the channel dimension, and input the input tensor into the pre-trained convolutional neural network model;

[0097] S1032, perform local feature extraction on the input tensor through a first convolutional layer, perform sliding calculation on the tensor using a convolution kernel with a preset size, and generate a first initial feature map containing primary features;

[0098] S1033, perform feature condensation on the first initial feature map through a first pooling layer, reduce the dimension of the feature map using a maximum value selection method, and generate a first condensed feature map;

[0099] S1034, input the first condensed feature map into a second convolutional layer to perform deep feature extraction, use convolution kernels with different sizes to extract more complex feature patterns, and generate a second initial feature map;

[0100] S1035, further condense the second initial feature map through a second pooling layer, reduce the feature dimension using a regional average value calculation method, and generate a second condensed feature map;

[0101] S1036, repeatedly perform the alternating processing of the convolution operation and the pooling operation, extract higher-level feature representations step by step, and output deep feature data containing temperature-sensitive components in the spectral data.

[0102] In the above scheme, the pre-trained convolutional neural network model is a computer model trained in advance by a large amount of data and capable of extracting deep features of the data, wherein the convolutional layer is a component for extracting local features of the data, and the pooling layer is a component for compressing feature data and retaining key information, and the two are alternately arranged to gradually extract more important features. The first initial feature map is structured data containing primary features generated after the convolutional layer is processed. The first condensed feature map is data with reduced dimensions and retained key primary features generated after the pooling layer is processed. The second convolutional layer is a component for extracting more complex deep features from the condensed features. The second initial feature map is data containing deep complex features generated after the second convolutional layer is processed. The second condensed feature map is more concise data containing key deep features generated after the second pooling layer is processed. The temperature-sensitive component refers to the core information in the features directly related to the temperature change. The deep feature data is the final feature data integrating all temperature-sensitive components after multiple rounds of alternating processing.

[0103] In the embodiment of the present application, first, the digital spectrum data is arranged through step S1031: the number of channels contained in the data is determined, i.e. the Stokes and anti-Stokes, and the number of sampling points of each channel; then the data is organized in the order of channel dimension first and sampling point dimension last to form an input tensor, for example, the 80 sampling values of the Stokes channel are s1, s2…s80 in turn, and the 80 sampling values of the anti-Stokes channel are a1, a2…a80 in turn, the input tensor is organized into the structure of [[s1, s2…s80], [a1, a2…a80]], wherein the first sublist is the Stokes channel (first dimension), the second sublist is the anti-Stokes channel (second dimension), and the elements in each sublist are sampling points (second dimension); finally, the input tensor is directly input into the pre-trained convolutional neural network model to ensure that the model can identify the information corresponding to each dimension, for example, the model will automatically identify that the first dimension is the channel and the second dimension is the sampling point, so as to process the sampling data according to the channel.

[0104] Secondly, the convolution kernel parameters of the first convolution layer are determined through step S1032: the preset convolution kernel size is 3x1 (suitable for 1-dimensional spectral data, 3 is the window length, and 1 is the channel adaptation number), and the step is 1; then the convolution kernel is independently slid on each channel of the input tensor to calculate: taking the Stokes channel as an example, the convolution kernel first covers the first three sampling points s1, s2 and s3 to calculate "s1xk1+s2xk2+s3xk3", where k1, k2 and k3 are fixed values of the convolution kernel, which are trained in advance to obtain the first feature value; then the convolution kernel slides 1 sampling point to the right to cover s2, s3 and s4, and the second feature value is calculated in the same way, until the entire channel is slid; the anti-Stokes channel is processed in the same way; finally, all the feature values of the two channels are integrated to generate the first initial feature map, for example, the input tensor is "2 channels x 80 sampling points", and after sliding with a 3x1 convolution kernel, 78 feature values are generated in each channel, and the first initial feature map is a structure of "2 channels x 78 feature values".

[0105] Then, the pooling window size of the first pooling layer is set to 2x1 (2 is the window length, and 1 is the channel adaptation number) through step S1033, and the step is 2; then the window is slid on each channel of the first initial feature map: taking the Stokes channel feature values f1-f78 as an example, the window first covers f1 and f2, and the maximum value of the two is selected as the condensed feature value f1'; then the window slides 2 feature values to the right to cover f3 and f4, and the maximum value is selected as f2', until the entire channel is slid; the anti-Stokes channel is processed in the same way; finally, the condensed feature values of the two channels are integrated to generate the first condensed feature map, for example, the first initial feature map is "2 channels x 78 feature values", and after sliding with a 2x1 window and a step of 2, 39 feature values are generated in each channel, and the first condensed feature map is a structure of "2 channels x 39 feature values".

[0106] Then, through step S1034, the convolution kernel of the second convolution layer is determined: two different sizes of convolution kernels, 2x1 and 4x1, are adopted, and the step is 1, which respectively captures the small-range and large-range feature association; then the first condensed feature map is input into the second convolution layer, and the two convolution kernels are respectively calculated by sliding on each channel: taking the first condensed feature map "2 channels x 39 feature values" as an example, the 2x1 convolution kernel slides on the Stokes channel, covers 2 adjacent feature values to calculate, and generates 39-2+1=38 feature values; the 4x1 convolution kernel also slides on the Stokes channel, covers 4 adjacent feature values to calculate, and generates 39-4+1=36 feature values; the anti-Stokes channel is processed in the same way, and the two convolution kernels each generate 38 and 36 feature values; finally, all the feature values are integrated to generate the second initial feature map, for example, after integration, it is "4 channels x (38+36)=4 channels x 74 feature values" structure, containing small-range and large-range feature association information.

[0107] Then, through step S1035, the size of the pooling window of the second pooling layer is set to 2x1, and the step is 2; then the window is made to slide on each channel of the second initial feature map, and the region average value calculation method is used for processing: taking the second initial feature map "4 channels x 74 feature values" as an example, taking the feature values m1-m74 of a channel, the window first covers m1, m2, and calculates (m1+m2)÷2 to obtain condensed feature value m1'; then the window slides 2 feature values, covers m3, m4, and calculates (m3+m4)÷2 to obtain m2', until the whole channel is slid; all 4 channels are processed in this way; finally, the condensed feature values of all channels are integrated to generate the second condensed feature map. For example, 37 feature values are generated for each channel (74÷2=37), and the second condensed feature map is "4 channels x 37 feature values" structure, which retains the average trend of every 2 adjacent deep features.

[0108] Finally, by step S1036, the number of rounds of alternating processing is determined, and the size of the convolution kernel and the pooling method are flexibly adjusted according to the dimension of the output feature map of the previous round to ensure that each round can extract higher-level features that are more closely related to temperature; then, taking the second condensed feature map as the starting data, the first round of alternating processing is performed: using the adjusted convolution kernel to extract more complex feature correlations, and then compressing the dimension through pooling to obtain an intermediate feature map with lower dimension and more focused features; then, taking the intermediate feature map as the input of the next round, repeating the above convolution feature extraction and pooling compression process, and further strengthening the temperature sensitivity of the features after each round of processing; finally, after completing the preset number of rounds of alternating processing, the operation is stopped, and the final obtained feature map is integrated into deep feature data containing temperature-sensitive components in the spectral data. For example, taking the second condensed feature map of "4 channels x 37 feature values" as the starting data, after 2 rounds of alternating processing, deep feature data of "16 channels x 8 feature values" can be obtained, each feature of which is closely related to temperature change.

[0109] In actual application, when monitoring the surface temperature of an industrial equipment, the digitized spectral data containing Stokes and anti-Stokes channels and 120 sampling points each are organized into an input tensor of "2 channels x 120 sampling points" according to the channel dimension, and input into the pre-trained convolutional neural network model; then, the first 3x1 convolution kernel is used for sliding processing to generate a first initial feature map of "2 channels x 118 feature values", and then 2x1 maximum value pooling is used to obtain a first condensed feature map of "2 channels x 59 feature values"; then, the 2x1 and 4x1 convolution kernels are used to process the condensed feature map to generate a second initial feature map of "4 channels x (58+56) = 4 channels x 114 feature values", and then 2x1 average value pooling is used to obtain a second condensed feature map of "4 channels x 57 feature values"; then, the feature extraction and pooling compression process of different size convolution kernels is repeated for 2 rounds, the first round uses 3x1 and 5x1 convolution kernels and 2x1 maximum value pooling to obtain a feature map of "8 channels x 27 feature values", and the second round uses 2x1 and 3x1 convolution kernels and 2x1 average value pooling, and finally outputs deep feature data of "8 channels x 13 feature values", which contains core information strongly related to the change of the surface temperature of the equipment.

[0110] The overall scheme of S103 above lays the foundation for subsequent feature extraction by organizing the input tensor in the channel dimension to adapt the input requirements of the convolutional neural network model to digital spectral data; gradually captures simple to complex spectral change information, especially temperature-related features, by using different size convolution kernels in the convolution layer; retains key information while reducing data volume and avoiding redundancy by using maximum or average value in the pooling layer; and finally generates deep feature data that accurately reflects the correlation between spectral data and temperature by alternating multiple rounds of processing, layer-by-layer filtering and strengthening temperature-sensitive features, providing high-quality feature support for subsequent adversarial generative network optimization and temperature calculation to ensure the accuracy of temperature monitoring.

[0111] S104, input the deep feature data into a pre-trained adversarial generative network model, and perform feature distribution optimization on the deep feature data in the latent space through the adversarial generative network model to suppress abnormal feature patterns caused by strong electromagnetic interference to generate optimized feature representation;

[0112] Optionally, step S104 can specifically include the following steps:

[0113] S1041, input the deep feature data into the encoder part of the pre-trained generator network, map the input features to the latent space through multiple layers of nonlinear transformation to generate initial latent feature representation;

[0114] S1042, perform distribution optimization adjustment on the initial latent feature representation in the latent space to make the feature distribution approach the normal temperature feature distribution, and perform abnormal feature suppression processing to eliminate abnormal feature patterns that differ from the normal temperature feature distribution by more than a preset threshold;

[0115] S1043, perform feature reconstruction on the optimized and processed initial latent feature representation through the decoder part of the generator network, and use layer-by-layer upsampling and feature fusion operations to convert the optimized and processed initial latent feature representation into optimized feature representation with the same dimension as the original deep feature data.

[0116] Wherein, step S1042 specifically includes the following process: calculate the statistical difference between the initial latent feature representation and the normal temperature feature distribution in the latent space to obtain a feature distribution deviation index; adjust the distribution parameters in the initial latent feature representation according to the feature distribution deviation index to make the adjusted feature distribution approach the normal temperature feature distribution; identify abnormal feature points in the adjusted initial latent feature representation that deviate from the mean value of the normal temperature feature distribution by more than a preset threshold; and perform suppression processing on the abnormal feature points by adjusting them to within the normal temperature feature distribution range through feature value correction operations.

[0117] In the above scheme, the initial latent feature representation is low-dimensional feature data in a low-dimensional latent space compressed by the adversarial generative network generator encoder. The feature distribution deviation index is a specific numerical value quantifying the statistical difference between the initial latent feature representation and the normal temperature feature distribution. The abnormal feature point refers to the feature value in the adjusted initial latent feature representation that deviates from the mean of the normal temperature feature distribution by more than a preset threshold. The original deep feature data dimension refers to the number of channels and feature values of the deep feature output by step S103, which is a key indicator of the size of the deep feature data structure. The optimized feature representation is a feature data that more accurately reflects the temperature information after feature distribution optimization and abnormality suppression in the latent space.

[0118] In the embodiments of the present application, as shown in Figure 2 As shown in the figure, first, the deep feature data obtained in step S103 is input into the generator encoder of the pre-trained adversarial generative network through step S1041; for example, the structure of "8 channels x 13 feature values" is input into the encoder of the pre-trained adversarial generative network, and the encoder performs multi-layer nonlinear transformation: the first layer uses the ReLU function to compress the 8x13 high-dimensional feature (a total of 104 numerical values) to 40 dimensions, filtering out part of the redundant information; the second layer uses the Tanh function to further compress the 40-dimensional feature to 8 dimensions, so that the numerical value of each dimension is controlled between -1 and 1; through the two layers of nonlinear transformation, the originally high-dimensional deep feature data is mapped to the low-dimensional latent space to generate an initial latent feature representation composed of 8 numerical values, for example, the final obtained initial latent feature is [0.6, -0.2, 0.4, 0.1, -0.3, 0.5, -0.1, 0.3], each numerical value represents a simplified core feature dimension, which is convenient for subsequent adjustment in the latent space.

[0119] Secondly, through step S1042, the statistical difference between the initial latent feature representation and the normal temperature feature distribution is calculated in the latent space. First, the statistical indicators of the normal temperature feature distribution are obtained, and then the mean and variance of each dimension of the initial latent feature are calculated. The statistical difference is calculated by the difference, and the difference is integrated into the feature distribution deviation index. Secondly, the distribution parameters of the initial latent feature are adjusted according to the deviation index: for the dimensions with large deviation, the mean value is adjusted so that the adjusted feature distribution is closer to the normal distribution. Then, the abnormal feature points are identified, and the preset threshold is set to 0.2, that is, the deviation from the normal mean value is greater than 0.2, which is abnormal. The abnormal feature points are screened by comparing the adjusted features with the normal mean value. Finally, the abnormal feature points are suppressed: the feature value correction operation is used to move the abnormal point value towards the normal mean value, and the optimized latent feature representation is finally obtained.

[0120] Finally, the optimized latent feature representation obtained in step S1042 is input into the decoder part of the generator in step S1043; for example, the optimized latent feature representation is an 8-dimensional feature, and the decoder first performs layer-by-layer upsampling: in the first step, the 8-dimensional feature is raised to 16 dimensions by copying and simply calculating supplementary values; in the second step, the 16-dimensional feature is raised to 32 dimensions to ensure that no information is lost; in the third step, the 32-dimensional feature is raised to the dimension of "8 channels x 13 feature values", which corresponds to the dimension of the original deep feature, and the 32-dimensional values are allocated by channel and feature point; at the same time, feature fusion is performed: after each round of upsampling, the features of the current layer are combined with the key features of the previous layer, for example, the important 3 dimensions in the 16-dimensional feature are merged with the corresponding dimensions in the 32-dimensional feature, and the core information after optimization is retained; through such upsampling and fusion, the optimized latent feature is finally converted into an optimized feature representation with the same dimension as the original deep feature data. For example, the original deep feature is "8 channels x 13 feature values", and the optimized feature output by the decoder is also "8 channels x 13 feature values", but each feature value has excluded the abnormality caused by electromagnetic interference.

[0121] In actual application, when monitoring the surface temperature of a transformer in a certain high-voltage substation, the output deep feature data of "16 channels x 17 feature values" is first input into the generator encoder of the pre-trained generative adversarial network; the encoder performs two layers of nonlinear transformation, where the first layer uses the ReLU function to compress the 16 x 17 features to 32 dimensions, and the second layer uses the Tanh function to compress to 10 dimensions, generating an initial latent feature representation of 10 dimensions, with values of 7.2, 4.8, 6.5, 3.9, 8.1, 5.3, 7.8, 4.2, 6.9, and 5.7; then the statistical difference degree of the initial latent feature representation and the normal temperature feature distribution is calculated: the mean of each dimension in the normal temperature feature distribution is 5.0, and the threshold is 2.0, the mean of the first dimension of the initial latent feature representation is 7.2, which deviates from 5.0 by 2.2, and the mean of the fifth dimension is 8.1, which deviates from 5.0 by 3.1, to obtain the deviation index; according to the index, the distribution parameters are adjusted, the mean of the first dimension is adjusted to 5.8, and the mean of the fifth dimension is adjusted to 5.9; then the abnormal points are identified: the original values of the first dimension and the fifth dimension are 7.2 and 8.1, which deviate from 5.0 by more than 2.0, and are corrected to 5.6 and 5.8; finally, the decoder restores the optimized 10-dimensional latent feature into an optimized feature representation of "16 channels x 17 feature values" through layer-by-layer upsampling (10 dimensions → 20 dimensions → 32 dimensions → 16 x 17 dimensions) and feature fusion, and completes feature optimization.

[0122] The overall scheme of S104 above provides a more operable low-dimensional environment for feature optimization by mapping deep features to a latent space through a generator encoder; reduces the distribution deviation caused by electromagnetic interference by calculating the statistical difference degree and adjusting the distribution; directly eliminates the false features caused by strong electromagnetic interference through abnormal feature point identification and correction; and finally restores the original dimension through the decoder to ensure that the optimized features can adapt to the subsequent processing flow. The overall process effectively suppresses the influence of strong electromagnetic interference on features, making the optimized feature representation more accurate in reflecting temperature information, and laying a reliable feature foundation for subsequent conversion of temperature feature vectors and accurate temperature data calculation.

[0123] S105, converting the optimized feature representation into a temperature feature vector in the temperature domain through a fully connected neural network layer, and mapping the temperature feature vector into continuous temperature data based on a temperature calculation function as a temperature calculation result output.

[0124] Optionally, step S105 can specifically include the following steps:

[0125] S1051, inputting the optimized feature representation into a first transformation unit of the fully connected neural network layer, and generating an intermediate feature representation through linear combination operation of a weight matrix and a bias vector;

[0126] S1052, inputting the intermediate feature representation into a nonlinear activation unit, and obtaining an activated feature representation with enhanced expression ability through feature transformation by a preset nonlinear function;

[0127] S1053, directly inputting the activated feature representation into a second transformation unit of the fully connected neural network layer, and converting the high-dimensional feature into a temperature feature vector with temperature representation ability through dimension reduction mapping operation;

[0128] S1054, inputting the temperature feature vector into a temperature calculation function, and mapping each feature component in the feature vector into a temperature value of the corresponding optical fiber distance point through component weighted summation calculation;

[0129] S1055, arranging and combining the temperature values of each optical fiber distance point in ascending order of distance to form a complete continuous temperature data sequence along the optical fiber distribution, and outputting the continuous temperature data sequence as the final temperature calculation result.

[0130] In the above scheme, the first transformation unit of the fully connected neural network layer is a module responsible for converting the optimization features into intermediate features. The weight matrix is a pre-trained numerical table that controls the importance of each optimization feature. The bias vector is a numerical set consistent with the dimension of the intermediate features, used to fine-tune the results of the intermediate features. The intermediate feature representation is a transition feature data with higher dimension than the optimization features after linear combination, used for subsequent activation processing. The activated feature representation is a feature data with consistent dimension with the intermediate features after nonlinear transformation, which eliminates invalid values and retains valid temperature correlation information. The temperature feature vector is a low-dimensional feature set with consistent dimension with the number of fiber distance points and only containing temperature representation information after dimension reduction, used for subsequent temperature calculation. The continuous temperature data sequence is a complete data set arranged in distance order, reflecting the temperature changes of each position on the device surface.

[0131] In the embodiments of the present application, first, the optimization feature representation is unfolded into a one-dimensional vector by channel-feature value through step S1051: if the optimization feature is "8 channels x 13 feature values", it is unfolded into a one-dimensional vector composed of 104 numerical values, such as [0.45, -0.15, 0.4,..., 0.3]; then the vector is input into the first transformation unit, and the unit calls the pre-trained weight matrix and bias vector; then the linear combination operation is performed: first, calculate "optimization feature vector x weight matrix", and then add the result to the bias vector; finally, generate a 64-dimensional intermediate feature representation, such as [2.2, 1.88,..., 2.32], which has a higher dimension than the optimization feature, facilitating subsequent activation processing.

[0132] Secondly, the intermediate feature representation generated by step S1051 is input into the nonlinear activation unit through step S1052, for example, the intermediate feature representation is a 64-dimensional vector [2.2, 1.88, -0.3, 0.5,..., -0.15]; then the unit calls the pre-set ReLU nonlinear function, which has the rule "if the feature value > 0, keep the value; if the value ≤ 0, set the value to 0"; then each value of the intermediate feature is transformed according to the rule: for example, 2.2 in the intermediate feature > 0, keep 2.2; 1.88 > 0, keep 1.88; -0.3 ≤ 0, set to 0; 0.5 > 0, keep 0.5; -0.15 ≤ 0, set to 0; finally, generate an activated feature representation consistent with the dimension of the intermediate feature, which eliminates invalid negative features and only retains valid features.

[0133] Then, through step S1053, the dimension of the temperature feature vector is determined: the dimension is equal to the number of distance points monitored by the optical fiber; then the activated feature representation, such as a 64-dimensional vector [2.2, 1.88, 0,..., 2.32], is input into the second transformation unit; the unit calls the pre-trained dimension reduction weight matrix, and the dimension reduction weight matrix has a dimension of 64x13, and each value controls the influence of the activated feature on the temperature feature, such as 0.05 in the first row and the first column, which means that the influence of the first value of the activated feature on the first value of the temperature feature is 0.05; then the dimension reduction mapping operation is performed: the 64-dimensional activated feature vector is multiplied by the 64x13 weight matrix; finally, a 13-dimensional temperature feature vector is generated, such as [5.2, 4.8, 5.3,..., 5.1], and each component corresponds to a temperature-related feature of a distance point of the optical fiber.

[0134] Then, through step S1054, the temperature feature vector generated in step S1053 and the pre-trained component weight and reference temperature are obtained; then each component of the temperature feature vector is input into the temperature calculation function, and the component weighting sum rule is used for calculation: the temperature value of each component = the component value x the corresponding weight + the reference temperature, for example, the component value of the first distance point is 5.2, the corresponding weight is 0.8, and the calculation is 5.2x0.8+25=4.16+25=29.16℃; the second component is 4.8x0.9+25=4.32+25=29.32℃; all 13 components are calculated according to this rule to obtain the temperature values [29.16℃, 29.32℃,..., 29.28℃] corresponding to the 13 distance points of the optical fiber, and each temperature value accurately corresponds to a monitoring position.

[0135] Finally, through step S1055, the temperature values of the distance points of the optical fiber and the corresponding distance information are obtained; then the temperature values are arranged and combined in the order of increasing distance: the temperature value corresponding to the smallest distance of 1 meter is found as the first element of the sequence; then the temperature value corresponding to 2 meters is found as the second element; in this way, the elements are arranged in order; finally, a complete continuous temperature data sequence along the optical fiber is formed, and the sequence is output as the final temperature calculation result.

[0136] In practical applications, when monitoring the surface temperature of the gearbox shell of a wind turbine, the 8-channel x 13-eigenvalue optimized feature representation [0.42, -0.18, 0.39, …, 0.29] is first input into the first variable unit of the fully connected neural network layer: after being expanded into a 104-dimensional vector, it is multiplied by a 104 x 64 weight matrix, and a 64-dimensional bias vector is added to generate a 64-dimensional intermediate feature representation [2.05, 1.78, …, 2.25]; then input into the ReLU activation unit, set the negative values -0.32, -0.15, etc. in it to 0, and get the activated features [2.05, 1.78, 0, …, 2.25]; then input into the second variable unit, calculate with a 64 x 13 weight matrix, and reduce to a 13-dimensional temperature feature vector [5.1, 4.7, 5.3, …, 5.0]; then input into the temperature solving function, and weighted sum according to the weight [0.82, 0.88, 0.91, …, 0.84]: 5.1 x 0.82 = 4.18℃, 4.7 x 0.88 = 4.14℃, 5.3 x 0.91 = 4.82℃, …, get 13 temperature values; finally, arrange them in ascending order of distance as [4.18℃, 4.14℃, 4.82℃, …, 4.20℃], output as the final temperature solving result.

[0137] The overall scheme of S105 above, through the linear combination of the first variable unit, makes the optimized features fully related in each dimension, laying a foundation for subsequent temperature feature extraction; through the nonlinear activation to eliminate invalid features, enhancing the expression ability of the features; through the second variable unit to reduce the dimension, accurately obtaining the feature vector suitable for temperature solving; through the weighted sum of the temperature solving function, converting the feature components into actual temperature values; finally, arranging the continuous data according to the distance to ensure that the temperature data accurately correspond to the equipment monitoring position. The whole process realizes the conversion from optimized features to available temperature data, effectively retains temperature information and eliminates interference, and the output continuous temperature data can directly reflect the surface temperature distribution of the equipment, meeting the needs of accurate monitoring in industrial scenarios.

[0138] The following is a complete example for steps 101-105, such as Figure 3As shown, when monitoring the surface temperature of the gearbox shell of a large wind farm, first start the distributed optical fiber temperature monitoring system, synchronously collect the original spectral data of the Stokes channel and the anti-Stokes channel through the signal acquisition unit, extract the light intensity values from the Stokes channel to form the sequence [118, 121, 125, 123, 128, 126, 130, 129] according to the 8 monitoring points of the optical fiber 1-8 meters, extract the light intensity values of the same distance points from the anti-Stokes channel to form the sequence [92, 95, 98, 97, 101, 100, 103, 102], and then combine the two sequences according to the corresponding distance points to obtain the original monitoring data [(118, 92), (121, 95), …, (129, 102)].

[0139] Then, the two intensity sequences are respectively input into independent photoelectric converters to generate analog electrical signals of the Stokes channel and the anti-Stokes channel; the signals are input into a programmable gain amplifier to amplify the Stokes channel signal by 1.35 times to 0.513V and reduce the anti-Stokes channel signal by 0.93 times to 1.9995V, so that it is stabilized in the range of 0.5-2.1V; then input into a programmable filter group to adjust the parameters according to the frequency band characteristics of the useful signal 140-880Hz to filter out out-of-band noise; then input into a high-precision analog-to-digital converter to sample synchronously at a sampling rate of 1780 times / s (more than twice the highest frequency 880Hz) and quantize into digital signals; finally, add “ST-B” and “ANTI-ST-B” labels to the two channel signals respectively, and combine them into digitized spectral data according to the time stamp alignment.

[0140] Then, the digitized spectral data is organized as an input tensor of 2 channels x 8 sampling points [[118, 121, …, 129], [92, 95, …, 102]] and input into a pre-trained convolutional neural network; a 2 channel x 3 feature value first condensed feature map is obtained by generating a 2 channel x 6 feature value first initial feature map through a 3 x 1 convolution kernel and performing 2 x 1 maximum value pooling; then a 4 channel x 2 feature value second condensed feature map is obtained by generating a 4 channel x 5 feature value second initial feature map through a 2 x 1 and 4 x 1 convolution kernel and performing 2 x 1 average value pooling, and after repeating the “convolution-pooling” for 1 round, a 6 channel x 4 feature value deep feature data is output; the deep feature data is input into the generative adversarial network, mapped to a 6-dimensional latent space through the encoder ReLU+Tanh transformation, and an initial latent feature [0.55, -0.18, 0.48, 0.12, -0.27, 0.65] is generated; the difference degree with the normal temperature feature distribution is calculated, the parameters are adjusted and the abnormal points are removed, and then the optimized feature representation is generated through the decoder upsampling (6 dimensions→12 dimensions→6 x 4 dimensions) and feature fusion.

[0141] Finally, the optimized features are unfolded into a 24-dimensional vector and input into the first transformation unit of the fully connected neural network, multiplied by a 24x28 weight matrix and added to a 28-dimensional bias vector to generate a 28-dimensional intermediate feature [1.95, 1.72, -0.25, …, 2.08]. After filtering negative values through the ReLU activation unit, the second transformation unit maps it to an 8-dimensional temperature feature vector [4.9, 4.7, 5.1, 4.8, 5.2, 5.0, 5.3, 5.2] through a 28x8 dimension reduction weight matrix. The temperature calculation function is input, calculated according to the weight [0.81, 0.79, …, 0.84] and the reference temperature 24℃, and the 8 distance point temperature values are arranged in order of 1-8 meters as a continuous temperature data sequence [27.97℃, 27.71℃, …, 28.17℃], output as the final result.

[0142] Figure 4 FIG. 1 is a structural schematic diagram of a specific embodiment of a deep learning-based optical fiber temperature measurement signal compensation system provided by the present application, Figure 4 The system can include:

[0143] The acquisition module 41 is configured to acquire original spectrum data of Stokes and anti-Stokes channels in a distributed optical fiber temperature monitoring system, and obtain original monitoring data containing intensity-distance relationship and spectral morphological features.

[0144] The processing module 42 is configured to convert the original monitoring data into an electrical signal, pre-process the electrical signal using a programmable gain analog front-end circuit, and convert the pre-processed electrical signal into a digital signal to obtain digitized spectrum data containing Stokes and anti-Stokes channel intensity information.

[0145] The extraction module 43 is configured to input the digitized spectrum data into a pre-trained convolutional neural network model, process it through a plurality of alternately arranged convolutional layers and pooling layers in the convolutional neural network model, and extract temperature-sensitive spectral deep features as deep feature data.

[0146] The generation module 44 is configured to input the deep feature data into a pre-trained generative adversarial network model, perform feature distribution optimization on the deep feature data in the latent space through the generative adversarial network model, suppress abnormal feature patterns caused by strong electromagnetic interference, and generate an optimized feature representation.

[0147] The output module 45 is configured to convert the optimized feature representation into a temperature feature vector in the temperature domain through a fully connected neural network layer, and map the temperature feature vector to continuous temperature data based on a temperature calculation function to output the temperature calculation result as a temperature calculation result.

[0148] The deep learning-based optical fiber temperature measurement signal compensation system of the embodiments of the present application is used to implement the aforementioned deep learning-based optical fiber temperature measurement signal compensation method, and therefore the specific embodiments of the deep learning-based optical fiber temperature measurement signal compensation system can be seen from the aforementioned embodiment part of the deep learning-based optical fiber temperature measurement signal compensation method, and the specific embodiments can be referred to the description of the corresponding embodiment part, which will not be repeated here.

[0149] The present application also provides an electronic device, comprising: a memory for storing a computer program; and a processor for implementing the steps of the deep learning-based optical fiber temperature measurement signal compensation method described above when executing the computer program.

[0150] The present application also provides a computer readable storage medium, which stores a computer program, and the computer program implements the steps of the deep learning-based optical fiber temperature measurement signal compensation method described above when executed by a processor.

[0151] In an exemplary embodiment, the computer readable storage medium described above can include, but is not limited to, a U disk, a read-only memory, a random access memory, a mobile hard disk, a magnetic disk or an optical disk, and various media that can store computer programs.

[0152] The embodiments of the present application also provide a computer program product, which comprises a computer program, and the computer program implements the steps in the deep learning-based optical fiber temperature measurement signal compensation method described above when executed by a processor.

[0153] The skilled person can further realize that the units and algorithm steps of the examples described in conjunction with the embodiments disclosed herein can be realized by electronic hardware, computer software or a combination of both. In order to clearly illustrate the interchangeability of hardware and software, the components and steps of the examples have been described in the above description in general terms. Whether the functions are performed in hardware or software depends on the specific application and design constraints of the technical solution. The skilled person can use different methods to implement the described functions for each specific application, but such implementation should not be considered beyond the scope of the present application.

[0154] The deep learning-based optical fiber temperature measurement signal compensation method and system provided by the present application are described in detail above. The principles and implementation of the present application are described by specific examples in this paper, and the above description of the examples is only used to help understand the method and its core idea of the present application. It should be pointed out that for ordinary skilled person in the technical field, without departing from the principles of the present application, some improvements and modifications can be made to the present application, and these improvements and modifications also fall within the protection scope of the present application.

Claims

1. A method for compensating fiber optic temperature measurement signals based on deep learning, characterized in that, include: Raw spectral data of Stokes and anti-Stokes channels in a distributed fiber optic temperature monitoring system are collected to obtain raw monitoring data containing intensity-distance relationships and spectral morphology characteristics. The raw monitoring data is converted into an electrical signal, the electrical signal is preprocessed using a programmable gain analog front-end circuit, and the preprocessed electrical signal is converted into a digital signal to obtain digital spectral data containing Stokes channel and anti-Stokes channel intensity information. The digitized spectral data is input into a pre-trained convolutional neural network model, and processed by multiple alternating convolutional and pooling layers in the convolutional neural network model to extract temperature-sensitive deep spectral features as deep feature data. The deep feature data is input into a pre-trained adversarial generative network model. The adversarial generative network model optimizes the feature distribution of the deep feature data in the latent space, suppresses abnormal feature patterns caused by strong electromagnetic interference, and generates optimized feature representations. The optimized feature representation is converted into a temperature feature vector in the temperature domain through a fully connected neural network layer, and the temperature feature vector is mapped to continuous temperature data based on the temperature calculation function as the output temperature calculation result.

2. The method according to claim 1, characterized in that, The step of inputting the deep feature data into a pre-trained Generative Adversarial Network (GAN) model, and optimizing the feature distribution of the deep feature data in the latent space through the GAN model to suppress abnormal feature patterns caused by strong electromagnetic interference, in order to generate optimized feature representations, includes: The deep feature data is input into the encoder part of the pre-trained generator network, and the input features are mapped to the latent space through multi-layer nonlinear transformation to generate an initial latent feature representation. In the latent space, the initial latent feature representation is optimized and adjusted to make the feature distribution closer to the normal temperature feature distribution. At the same time, abnormal feature suppression processing is performed to eliminate abnormal feature patterns that differ from the normal temperature feature distribution by more than a preset threshold. The decoder part of the generator network reconstructs the optimized and processed initial latent feature representation by employing layer-by-layer upsampling and feature fusion operations to transform the optimized and processed initial latent feature representation into an optimized feature representation with the same dimension as the original deep feature data.

3. The method according to claim 2, characterized in that, The step of optimizing the distribution of the initial latent feature representation in the latent space to make the feature distribution closer to the normal temperature feature distribution, and simultaneously performing abnormal feature suppression processing to eliminate abnormal feature patterns that differ from the normal temperature feature distribution by more than a preset threshold, includes: Calculate the statistical difference between the initial latent feature representation and the normal temperature feature distribution in the latent space to obtain the feature distribution deviation index; Based on the characteristic distribution deviation index, the distribution parameters in the initial potential characteristic representation are adjusted so that the adjusted characteristic distribution approaches the normal temperature characteristic distribution. In the adjusted initial latent feature representation, abnormal feature points that deviate from the mean of the normal temperature feature distribution by more than a preset threshold are identified. The abnormal feature points are suppressed, and their feature value is adjusted to the normal temperature feature distribution range through feature value correction operation.

4. The method according to claim 1, characterized in that, The process of inputting the digitized spectral data into a pre-trained convolutional neural network model, and processing it through multiple alternating convolutional and pooling layers in the convolutional neural network model to extract temperature-sensitive deep spectral features as deep feature data includes: The digitized spectral data is organized into an input tensor according to the channel dimension and then input into the pre-trained convolutional neural network model. The input tensor is used to extract local features through the first convolutional layer. The convolutional kernel of a preset size is used to perform sliding calculations on the tensor to generate a first initial feature map containing primary features. The first initial feature map is condensed by the first pooling layer, and the feature map dimension is reduced by the maximum value selection method to generate the first condensed feature map. The first condensed feature map is input into the second convolutional layer for deep feature extraction. Convolutional kernels of different sizes are used to extract more complex feature patterns and generate a second initial feature map. The second initial feature map is further condensed by a second pooling layer. The feature dimension is reduced by calculating the region average value to generate a second condensed feature map. By repeatedly performing alternating convolution and pooling operations, higher-level feature representations are extracted step by step, and the output is deep feature data containing temperature-sensitive components in the spectral data.

5. The method according to claim 1, characterized in that, The process of converting the optimized feature representation into a temperature feature vector in the temperature domain through a fully connected neural network layer, and mapping the temperature feature vector to continuous temperature data based on a temperature calculation function as the output temperature calculation result includes: The optimized feature representation is input into the first transformation unit of the fully connected neural network layer, and an intermediate feature representation is generated through a linear combination operation of the weight matrix and the bias vector. The intermediate feature representation is input into a nonlinear activation unit, and the feature is transformed through a preset nonlinear function to obtain an activated feature representation with enhanced expressive power. The activated feature representation is directly input into the second transformation unit of the fully connected neural network layer, and the high-dimensional features are converted into a temperature feature vector with temperature characterization capability through a dimensionality reduction mapping operation. The temperature feature vector is input into the temperature calculation function, and each feature component in the feature vector is mapped to the temperature value of the corresponding fiber distance point through a component weighted summation calculation method. The temperature values ​​at each fiber distance point are arranged and combined in ascending order of distance to form a complete continuous temperature data sequence distributed along the fiber. The continuous temperature data sequence is then output as the final temperature calculation result.

6. The method according to claim 1, characterized in that, The process of converting the raw monitoring data into an electrical signal, preprocessing the electrical signal using a programmable gain analog front-end circuit, and converting the preprocessed electrical signal into a digital signal to obtain digital spectral data containing Stokes channel and anti-Stokes channel intensity information includes: The Stokes light intensity distribution sequence and the anti-Stokes light intensity distribution sequence in the original monitoring data are respectively input into independent photoelectric converters to generate analog electrical signals for the corresponding Stokes channel and anti-Stokes channel; The analog electrical signals from each channel are input into the programmable gain amplifier in the programmable gain analog front-end circuit. The amplification factor is dynamically adjusted according to the monitored signal amplitude value to stabilize the output signal amplitude within the preset operating range. The analog electrical signals of each channel after gain adjustment are input into the programmable filter bank in the gain programmable analog front-end circuit. The filter parameters are adaptively adjusted according to the signal frequency band characteristics to retain the effective signal frequency band components and suppress out-of-band noise. The filtered analog electrical signals from each channel are sent to a high-precision analog-to-digital converter in a gain-programmable analog front-end circuit. The converter is then synchronously sampled at a sampling rate no less than twice the highest frequency of the signal, and the sampled values ​​are quantized into digital signals. Add corresponding channel identifier codes to the digital signals of the Stokes channel and the anti-Stokes channel, and align and combine them according to the same sampling timestamp to generate digital spectral data containing dual-channel intensity information.

7. The method according to claim 1, characterized in that, The raw spectral data of the Stokes and anti-Stokes channels in the distributed fiber optic temperature monitoring system are acquired to obtain raw monitoring data containing intensity-distance relationships and spectral morphology characteristics, including: The raw spectral data of the Stokes channel and the anti-Stokes channel are simultaneously acquired through the signal acquisition unit of the distributed optical fiber temperature monitoring system. The light intensity values ​​are extracted sequentially from the original spectral data of the Stokes channel according to the fiber distance points to form a Stokes light intensity distribution sequence; The light intensity values ​​of the original spectral data of the anti-Stokes channel are extracted sequentially at points with the same fiber distance to form an anti-Stokes light intensity distribution sequence. The Stokes light intensity distribution sequence and the anti-Stokes light intensity distribution sequence are combined according to the fiber distance points to form raw monitoring data that simultaneously contains dual-channel intensity information and distance correspondence.

8. A fiber optic temperature measurement signal compensation system based on deep learning, characterized in that, include: The acquisition module is used to acquire the raw spectral data of the Stokes channel and anti-Stokes channel in the distributed optical fiber temperature monitoring system, and obtain raw monitoring data containing intensity-distance relationship and spectral morphology characteristics. The processing module is used to convert the raw monitoring data into an electrical signal, preprocess the electrical signal using a programmable gain analog front-end circuit, and convert the preprocessed electrical signal into a digital signal to obtain digital spectral data containing Stokes channel and anti-Stokes channel intensity information. The extraction module is used to input the digitized spectral data into a pre-trained convolutional neural network model, and process it through multiple alternating convolutional and pooling layers in the convolutional neural network model to extract temperature-sensitive deep spectral features as deep feature data. The generation module is used to input the deep feature data into a pre-trained adversarial generative network model, and optimize the feature distribution of the deep feature data in the latent space through the adversarial generative network model to suppress abnormal feature patterns caused by strong electromagnetic interference, so as to generate an optimized feature representation. The output module is used to convert the optimized feature representation into a temperature feature vector in the temperature domain through a fully connected neural network layer, and to map the temperature feature vector into continuous temperature data based on the temperature calculation function as the output temperature calculation result.

9. An electronic device, characterized in that, include: Memory, used to store computer programs; A processor, configured to execute the computer program to implement the steps of the deep learning-based fiber optic temperature measurement signal compensation method as described in any one of claims 1 to 7.

10. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a computer program that, when executed by a processor, enables the implementation of the deep learning-based fiber optic temperature measurement signal compensation method as described in any one of claims 1 to 7.

Images