Three-dimensional fluorescence spectrometer and pollution tracing method thereof

Abstract

The invention belongs to the technical field of analysis and detection, and particularly relates to a three-dimensional fluorescence spectrometer and a pollution tracing method thereof.The three-dimensional fluorescence spectrometer comprises a light source module, a light source monochromator, a fluorescence monochromator, a sample pool and a PMT detector, and the light source monochromator and the fluorescence monochromator are each provided with an entrance slit and an exit slit; the entrance slit and the exit slit are both of a slit continuous adjustable structure, the slit bandwidth is continuously adjustable, and compared with a traditional slit with fixed bandwidth or multi-gear discrete adjustment, the continuous slit has remarkable advantages in performance optimization and application flexibility; the pollution traceability method comprises the following steps: constructing a three-dimensional fluorescence spectrum database of a pollution source, carrying out data acquisition and pretreatment on an acquired water sample, and training to obtain an optimal BiA-GRU-LSTM traceability model; newly collected sample data are input into the traceability model, the type and characteristics of a pollution source can be rapidly identified, and the method is suitable for traceability of pollution in multiple scenes such as surface water and industrial wastewater.

Description

Technical Field

[0001] This invention belongs to the field of analytical detection technology, specifically relating to a three-dimensional fluorescence spectrometer and its pollution source tracing method. Background Technology

[0002] Three-dimensional fluorescence spectroscopy is a spectroscopic analysis technique based on the fluorescence properties of substances. It constructs a three-dimensional excitation-emission-intensity spectrum, also known as a fluorescence fingerprint, by simultaneously acquiring data on the excitation wavelength, emission wavelength, and corresponding fluorescence intensity of a fluorescent substance. This enables qualitative and quantitative analysis and source identification of fluorescent components in a sample. Due to its advantages of high sensitivity, non-destructive nature, and simultaneous analysis of multiple components, this technology is widely used in environmental monitoring, especially as a core technology in tracing the source of organic pollutants in water bodies. However, in practical applications, the slit width is either fixed or discretely adjustable in multiple ranges, making continuous adjustment impossible. Furthermore, existing technologies, through simple preprocessing, cannot distinguish characteristic peaks with similar positions; overlapping and masking of fluorescence peaks lead to large source-tracing deviations and low accuracy. Moreover, existing technologies rely on large amounts of labeled data, making them prone to failure when dealing with different process wastewaters. Summary of the Invention

[0003] Based on the aforementioned shortcomings and deficiencies in the prior art, one of the objectives of this invention is to at least solve one or more of the aforementioned problems in the prior art. In other words, one of the objectives of this invention is to provide a three-dimensional fluorescence spectrometer and its pollution source tracing method that meet one or more of the aforementioned requirements.

[0004] To achieve the above-mentioned objectives, the present invention adopts the following technical solution: A three-dimensional fluorescence spectrometer includes a light source module, a light source monochromator, a fluorescence monochromator, a sample cell, and a PMT detector. Both the light source monochromator and the fluorescence monochromator have an entrance slit and an exit slit, and both the entrance slit and the exit slit are continuously adjustable structures. The continuously adjustable slit structure includes a base frame and a slit opening disposed on the base frame, a first slit adjustment block, a second slit adjustment block, a slit adjustment gear, a transmission gear, and a stepper motor. The slit adjustment gear is coaxially distributed with the slit opening, the slit adjustment gear meshes with the transmission gear, and the transmission gear is driven and connected to the stepper motor. The first slit adjustment block and the second slit adjustment block are movably connected to two parallel guide columns, which are fixedly installed on the base frame. The first slit adjustment block and the second slit adjustment block are distributed on both sides of the slit opening. Each of the two guide columns is fitted with a return spring. The two ends of one return spring abut against the first slit adjustment block and the base frame, and the two ends of the other return spring abut against the second slit adjustment block and the base frame. The slit adjusting gear has a first limiting post and a second limiting post at both ends along its central axis. The first limiting post abuts against the second slit adjusting block, and the second limiting post abuts against the first slit adjusting block. By rotating the slit adjusting gear, the first slit adjusting block and the second slit adjusting block are adjusted to move closer or further away from each other relative to the slit opening, so as to adjust the slit width.

[0005] As a preferred embodiment, the first slit adjustment block includes a first movable connecting block and a first slit piece disposed on the first movable connecting block, the first slit piece being used to block the slit opening; The second slit adjustment block includes a second movable connecting block and a second slit plate disposed on the second movable connecting block, the second slit plate being used to block the slit opening; The first slit plate and the second slit plate are brought closer together or further apart to adjust the slit width.

[0006] As a preferred embodiment, the first movable connecting block has an extended first abutting arm, the first abutting arm and the first movable connecting block are respectively located on both sides of the slit opening, and the second limiting post abuts against the side of the first abutting arm away from the first movable connecting block; The second movable connecting block has an extended second abutment arm, which is located on both sides of the slit opening, and the first limiting post abuts against the side of the second abutment arm away from the second movable connecting block.

[0007] As a preferred embodiment, the first slit adjustment block and the second slit adjustment block are centrally symmetrical to each other.

[0008] As a preferred embodiment, both the first limiting post and the second limiting post are bearing structures.

[0009] As a preferred embodiment, the base frame is provided with a positioning optical coupler, and the first slit adjustment block or the second slit adjustment block is provided with an optical coupler baffle. The optical coupler baffle and the positioning optical coupler are used for positioning cooperation to adjust the slit width.

[0010] As a preferred embodiment, the base frame is further provided with a filter turntable and several bandpass filters arranged circumferentially on the filter turntable. The filter turntable is connected to a drive motor for driving, and the bandpass filter is switched to correspond with the slit opening by rotating the filter turntable.

[0011] As a preferred embodiment, the base frame is also equipped with a positioning sensor for aligning the center of the bandpass filter with the center of the slit opening.

[0012] This invention also provides a pollution source tracing method for a three-dimensional fluorescence spectrometer as described in any of the preceding embodiments, comprising the following steps: S1. Acquire a three-dimensional fluorescence spectral dataset using a three-dimensional fluorescence spectrometer and perform data preprocessing; S2. Perform parallel factor decomposition on the preprocessed three-dimensional fluorescence spectral data to decompose it into excitation load matrix, emission load matrix and relative concentration load matrix to form a feature dataset. S3. Train the BiA-GRU-LSTM model using the feature dataset to obtain the source tracing model. The BiA-GRU-LSTM model includes parallel Bi-GRU and Bi-LSTM. The outputs of the two are concatenated through the feature dimension and multi-head attention is added. The attention weights are weighted and aggregated to generate attention-weighted feature representations. Then, layer normalization and batch standardization are applied sequentially, followed by several repeated processing layers. Finally, the ReLU activation function is used to output the model result. The repeated processing layers include linear transformation, batch standardization, and dropout layers in sequence. S4. Collect three-dimensional fluorescence spectral data of the water sample to be identified using a three-dimensional fluorescence spectrometer, perform data preprocessing and parallel factor decomposition, input the data into the source tracing model for prediction, and output the source tracing results.

[0013] As a preferred option, the source tracing results include the name of the emission source with the highest matching degree, the matching degree value, and the confidence interval; When the predicted matching degree is greater than or equal to the preset threshold, it is determined that the water sample to be identified comes from the corresponding discharge source; When the matching degree is less than the preset threshold, it is determined to be a water sample from an unknown source or a mixed polluted water sample.

[0014] Compared with the prior art, the beneficial effects of this invention are: (1) The present invention designs the entrance slit and exit slit of the light source monochromator and the fluorescence monochromator as a continuously adjustable slit structure, so as to realize the slit bandwidth is continuously adjustable from 1 to 20 nm. The optimal slit width can be freely selected according to the sample characteristics. Compared with the traditional fixed bandwidth or multi-level discrete adjustment slit, the continuous slit has significant advantages in performance optimization and application flexibility. The slit width is adjusted by stepper motor, which has high positioning accuracy and controllability. When used with a bandpass filter, the stray light suppression effect can be further optimized, and the data reliability can be significantly improved in high-sensitivity detection. (2) This invention uses preprocessing to enhance features and then combines it with parallel factor decomposition PARAFAC, which can reduce the effects of fluorescence peak overlap and masking, and reduce source tracing bias. (3) The source tracing model of the present invention adopts the BiA-GRU-LSTM deep learning model, which has the advantage of being able to capture the temporal correlation of fluorescence signals (continuous features of excitation-emission wavelengths), and has a stronger ability to identify components with similar peak positions but different peak shape details, thereby improving the accuracy of source tracing. Attached Figure Description

[0015] Figure 1This is a schematic diagram of the three-dimensional fluorescence spectrometer architecture according to Embodiment 1 of the present invention; Figure 2 This is a schematic diagram of the continuously adjustable slit structure of Embodiment 1 of the present invention; Figure 3 This is a schematic diagram of the continuously adjustable slit structure of Embodiment 1 of the present invention from another perspective; Figure 4 This is a partial structural schematic diagram of the continuously adjustable slit structure of Embodiment 1 of the present invention; Figure 5 This is a schematic diagram of the structure of the first slit adjustment block in Embodiment 1 of the present invention; Figure 6 This is a flowchart of the pollution source tracing method of Embodiment 1 of the present invention; Figure 7 These are the original three-dimensional fluorescence spectrum and the corrected three-dimensional fluorescence spectrum of Embodiment 1 of the present invention; Figure 8 This is a model architecture diagram of BiA-GRU-LSTM in Embodiment 1 of the present invention. Detailed Implementation

[0016] To more clearly illustrate the embodiments of the present invention, specific implementation methods will be described below with reference to the accompanying drawings. Obviously, the drawings described below are merely some embodiments of the present invention. For those skilled in the art, other drawings and other implementation methods can be obtained based on these drawings without any creative effort.

[0017] Example 1: like Figures 1 to 5 As shown, the three-dimensional fluorescence spectrometer in this embodiment adopts a concave grating dual monochromator design, including a light source module 1, a light source monochromator 2, a fluorescence monochromator 3, a sample cell 4, and a PMT detector. The coordination relationship of the above-mentioned devices is not described in detail here, but can be referred to the prior art.

[0018] In this embodiment, both the light source monochromator 2 and the fluorescent monochromator 3 have an entrance slit and an exit slit, and both the entrance slit and the exit slit are continuously adjustable structures. Specifically, the continuously adjustable slit structure includes a base frame 51 and a slit opening 52 disposed on the base frame 51, a first slit adjusting block 53, a second slit adjusting block 54, a slit adjusting gear 55, a transmission gear 56, and a stepper motor 57. The slit adjusting gear 55 is coaxially distributed with the slit opening 52, and the slit adjusting gear 55 meshes with the transmission gear 56. The transmission gear 56 is driven by the motor shaft of the stepper motor 57. The diameter of the transmission gear 56 is smaller than the diameter of the slit adjusting gear 55.

[0019] The first slit adjustment block 53 and the second slit adjustment block 54 are movably connected to two parallel guide posts 58, that is, the guide posts 58 pass through the first slit adjustment block 53 and the second slit adjustment block 54, and the guide posts 58 are fixedly installed on the base frame 51; the first slit adjustment block 53 and the second slit adjustment block 54 are distributed on both sides of the slit opening 52, and the two guide posts 58 are respectively fitted with return springs 59, the two ends of one return spring abutting between the first slit adjustment block 53 and the base frame 51, and the two ends of the other return spring abutting between the second slit adjustment block 54 and the base frame 51.

[0020] The slit adjusting gear 55 has a first limiting post 61 and a second limiting post 62 at both ends along its central axis. The first limiting post 61 abuts against the second slit adjusting block 54, and the second limiting post 62 abuts against the first slit adjusting block 53. By rotating the slit adjusting gear 55, the first slit adjusting block 53 and the second slit adjusting block 54 are adjusted to move closer or further away from each other relative to the slit opening 52, so as to adjust the slit width.

[0021] Specifically, the first slit adjustment block 53 includes a first movable connecting block 531 and a first slit piece 532 disposed on the first movable connecting block. The first slit piece 532 is used to block the slit opening 52. The first movable connecting block 531 has an extended first abutting arm 533. The first abutting arm 533 and the first movable connecting block 531 are respectively located on both sides of the slit opening 52. The second limiting post 62 abuts against the side of the first abutting arm away from the first movable connecting block. The structure of the second slit adjusting block is centrally symmetrical to the structure of the first slit adjusting block. The second slit adjusting block 54 includes a second movable connecting block and a second slit piece 542 disposed on the second movable connecting block. The second slit piece 542 is used to block the slit opening. The second movable connecting block has an extended second abutting arm. The second abutting arm and the second movable connecting block are respectively located on both sides of the slit opening. The first limiting post abuts against the side of the second abutting arm away from the second movable connecting block.

[0022] The first slit plate 532 and the second slit plate 542 are brought closer to each other or further apart, so that the slit width can be continuously adjusted from 1 to 20 nm.

[0023] In this embodiment, both the first limiting post 61 and the second limiting post 62 are bearing structures, which improves the stability of slit width adjustment.

[0024] In addition, the aforementioned base frame 51 is equipped with a positioning optical coupler 7, and the first slit adjustment block 53 is provided with an optical coupler baffle 8. The optical coupler baffle 8 and the positioning optical coupler 7 are used for positioning and cooperating to adjust the slit width.

[0025] In this embodiment, the base frame 51 is also equipped with a filter turntable 9 and several bandpass filters 90 circumferentially arranged on the filter turntable. The filter turntable 9 is driven by a drive motor 91, and the rotation of the filter turntable 9 switches the correspondence between the bandpass filter 90 and the slit opening 52. The base frame 51 is also equipped with a positioning sensor 10 (such as a conventional Hall element) for aligning the center of the bandpass filter with the center of the slit opening.

[0026] The slit width adjustment process in this embodiment is as follows: When the stepper motor rotates in the forward direction, the transmission gear drives the slit adjustment gear to rotate, changing the position of the first and second limit posts. The first and second limit posts push the two slit adjustment blocks to move left and right respectively, compressing the reset spring, opening the slit. When the baffle moves to the optical coupler position, the slit opens to its maximum. When the stepper motor rotates in the reverse direction, the reset spring closes the slit adjustment blocks. The bandpass filter is used to reduce stray light. The bandpass filter divides the spectrum into five segments: 200-300nm, 300-400nm, 400-500nm, 500-600nm, and 600-900nm. During spectral scanning, the corresponding filter is switched according to the scanning band.

[0027] The three-dimensional fluorescence spectrometer in this embodiment has two monochromators: a light source monochromator and a fluorescence monochromator. The light source monochromator expands the composite spectrum of the xenon lamp light source and selects the target wavelength as the excitation light. The fluorescence monochromator expands the fluorescence spectrum and sends it to the PMT detector for acquisition. Each monochromator has two slits (incident slit and exit slit). The continuously adjustable structure of the slits is the same, and its function is to adjust the spectral bandwidth and light flux. The monochromator principle of this embodiment is as follows: After light passes through the entrance slit, it illuminates the concave grating. The concave grating separates the incident light according to wavelength and projects it to the exit slit. The stepper motor rotates, driving the grating to rotate and selecting the target wavelength to enter the exit slit. The light source monochromator illuminates the sample cell with the light passing through the exit slit as excitation light. The light from the fluorescence monochromator exit slit directly enters the PMT detector. The specific principle can be referred to in the prior art, and will not be elaborated here.

[0028] The workflow of the three-dimensional fluorescence spectrometer in this embodiment is as follows: ① The light source module converges the divergent light from the xenon lamp (focusing at the entrance slit of the light source monochromator); ② The light enters the light source monochromator through the light gate. A bidirectional self-holding electromagnet connects to the light gate baffle. When the electromagnet is energized in the forward direction, the baffle moves forward to block the light, closing the light gate. When the electromagnet is energized in the reverse direction, the baffle moves backward to leave the light path, opening the light gate; ③ The light source monochromator spreads the light into 200-900nm, and a stepper motor drives the grating to rotate, sequentially passing the spectrum from 200-900nm through the light source monochromator. The light exits through a monochromator exit slit (with a bandpass filter turntable); ④ Half of the emitted light is intercepted by a beam splitter and enters the light source reference detector as reference light; ⑤ After passing through the beam splitter, the light is collimated by two curved mirrors and enters the measurement cell to excite fluorescence. Compared with collimation by a convex lens, collimation by a concave mirror results in less energy loss and no chromatic aberration; ⑥ The fluorescence in the measurement cell is collected by a fluorescence acquisition lens and enters the fluorescence monochromator; ⑦ The fluorescence monochromator scans the fluorescence at 200-900 nm to obtain the fluorescence spectrum; for details, please refer to existing technologies, which will not be elaborated here.

[0029] In this embodiment, the slit bandwidth is continuously adjustable from 1-20 nm, allowing for the free selection of the optimal slit width based on sample characteristics. Compared to traditional fixed-bandwidth or multi-level discretely adjustable slits, continuous slits offer significant advantages in performance optimization and application flexibility. The slit width is adjusted using a stepper motor, ensuring high positioning accuracy and controllability. The slit module, when used with a specially designed bandpass filter, further optimizes stray light suppression, significantly improving data reliability in high-sensitivity detection. Furthermore, the compact dual monochromator design, with the sample cell positioned close to the fluorescence monochromator inlet to increase fluorescence acquisition efficiency and sensitivity, ensures that the excitation light is highly collimated by two curved mirrors before entering the sample cell, reducing stray light and enhancing sensitivity.

[0030] like Figure 6 As shown, the pollution source tracing method of the three-dimensional fluorescence spectrometer in this embodiment includes the following steps: S1. Collect three-dimensional fluorescence data of original pollution source water samples and three-dimensional fluorescence data of unknown pollution source water samples using a three-dimensional fluorescence spectrometer; Specifically, water samples discharged by various enterprises in the industrial park at different time periods are systematically collected. Three-dimensional fluorescence spectrometers are used to collect three-dimensional fluorescence spectral data of water samples from various sources to ensure the consistency and accuracy of data collection. Based on the distribution characteristics of enterprises in the industrial park and the layout of sewage pipe network, the coverage area to be traced is delineated. Target discharge enterprises in different industries within the coverage area are systematically sampled, and three-dimensional fluorescence spectrometers are used to scan the three-dimensional fluorescence spectra of each water sample to obtain the corresponding three-dimensional fluorescence raw spectral data of each water sample.

[0031] S2, Data Preprocessing; By subtracting the background fluorescence value of blank water samples to correct the spectral data, an original three-dimensional fluorescence spectral database of discharged water samples in the industrial park was constructed. The obtained raw three-dimensional fluorescence spectral data were preprocessed in a targeted manner: cubic interpolation was used to remove Rayleigh scattering interference, and Raman scattering interference was subtracted by characteristic band fitting algorithm; SG (Savitzky-Golay) smoothing was applied to denoise the data, preserving the core spectral features while eliminating noise interference to the maximum extent.

[0032] Specifically, the raw spectral data should first undergo background correction. Pure water itself will produce a certain fluorescence signal, mainly from trace organic matter in the water or the inherent noise of the measurement system. All subsequent processing (such as scattering elimination, noise smoothing, etc.) should be based on the data after background removal; otherwise, these processing methods may incorrectly alter or enhance the background signal. Background correction uses the blank water sample fluorescence value interpolation method, and the calculation formula is as follows: ; in, Representing the One sample, Represents the excitation wavelength. Represents the emission wavelength; For the blank water sample that does not contain pollutants, at the first Each water sample corresponds to an excitation wavelength Emission wavelength The fluorescence intensity matrix below; For the original water sample in the first Each water sample corresponds to an excitation wavelength Emission wavelength The fluorescence intensity matrix below; To correct the water sample at the first Each water sample corresponds to an excitation wavelength Emission wavelength The fluorescence intensity matrix below; Next, the corrected original three-dimensional fluorescence spectral data undergoes targeted preprocessing to eliminate scattering and noise interference and enhance feature identification. First, Rayleigh scattering interference needs to be removed. Rayleigh scattering is a common interference in three-dimensional fluorescence spectroscopy, which can mask true fluorescence characteristics. Generally, at the diagonal and overtone positions where the excitation and emission wavelengths are equal, cubic interpolation is used to replace the fluorescence values ​​within this range, estimating and filling the values ​​in the scattering region while retaining valid data in the non-scattering region. Subsequently, a polynomial fitting subtraction method is used to remove Raman scattering interference. Based on the spectral characteristics of Raman scattering, which are generally characteristic emission peaks at specific excitation wavelengths, a Raman scattering baseline is fitted from the original data. Baseline values ​​are subtracted point by point to eliminate the masking of target fluorescence features by Raman scattering. Removal of Rayleigh and Raman scattering can eliminate non-target fluorescence interference and reduce the confusion of characteristic peaks by impurities, laying the foundation for subsequent separation. Then, the Savitzky-Golay (SG) smoothing method is applied for denoising. Local polynomial fitting is performed on the spectral data through a sliding window. While preserving the core spectral features such as the position and intensity of characteristic peaks, random noise is eliminated to the maximum extent, making the spectral lines smoother, reducing the interference of noise and scattering on fluorescence peaks, avoiding the "pseudo-overlap" phenomenon caused by the superposition of noise or scattering signals with target peaks, making the contours of characteristic peaks clearer, and indirectly improving the accuracy of characteristic peak separation. The calculation formulas for the above targeted preprocessing methods are as follows: ; ; ; in, For the first Each node The cubic interpolation function for wavelength position; This represents the number of interpolation nodes. Representing the non-Rayleigh scattering interference region, the first The fluorescence intensity matrix corresponding to each interpolation node; The fluorescence intensity matrix after Rayleigh scattering removal; Representative at Baseline intensity at wavelength position; , This represents the fluorescence intensity matrix after Raman scattering removal; These are the SG filter coefficients. The width is half the width of the sliding window, and the total size is... , The excitation wavelength within the representative window relative to the current calculation position The dimensional offset, with values ​​ranging from arrive ; This is the fluorescence intensity matrix after SG smoothing.

[0033] S3. Apply the Parallel Factorization (PARAFAC) algorithm to the preprocessed three-dimensional fluorescence data to accurately decompose the three-dimensional data into excitation load matrix, emission load matrix and relative concentration load matrix. Through this decomposition process, extract the specific fluorescence core features of each enterprise's effluent samples to form a feature dataset. Specifically, the parallel factor PARAFAC decomposition algorithm is applied to the preprocessed three-dimensional fluorescence spectral data for feature extraction. The optimal decomposition group number is determined based on the core consistency diagnostic method and residual analysis (generally, the group number range is set to 2-21 groups, and the group number corresponding to core consistency ≥80% and the smallest residual is selected as the optimal value). The PARAFAC decomposition precisely decomposes the three-dimensional fluorescence data into three loading matrices: an excitation loading matrix, an emission loading matrix, and a relative concentration loading matrix. Based on these three loading matrices, characteristic fluorescence component parameters (such as characteristic excitation wavelength, characteristic emission wavelength, and relative concentration percentage) specific to each emitting enterprise's water samples are selected, forming a core feature vector set for each sample. This achieves the transformation from high-dimensional three-dimensional fluorescence data to low-dimensional core features, accurately capturing the core fluorescence characteristics of each pollution source and highlighting the specific differences in water samples from different enterprises. Specifically, the PARAFAC decomposition formula is calculated as follows: ; in, The fluorescence intensity matrix after PARAFAC decomposition; The number of fluorescent components; Representing the The first water sample The relative concentration loading values ​​of each fluorescent component reflect the component contribution of each water sample; Representing the The excitation wavelength for the first... Excitation characteristic loading values ​​of each fluorescent component; Representing the The emission wavelength is for the first Emission characteristic loading values ​​of each fluorescent component; Let be the residual matrix.

[0034] like Figure 7As shown, the left side is the original three-dimensional fluorescence spectrum, where the fluorescence peaks overlap and are obscured, resulting in unclear features. The right side is the corrected three-dimensional fluorescence spectrum. After using water sample blank correction, SG smoothing, Rayleigh scattering correction, and PARAFAC parallel factor decomposition, not only are the contours of the fluorescence characteristic peaks more prominent and the signals more concentrated, but the overlap and obscuration of peaks are also effectively reduced. This can reduce the deviation in subsequent fluorescence source tracing analysis and is more conducive to accurately identifying the fluorescent component information in the water sample.

[0035] S4. Divide the core feature dataset into a preset proportion for training and prediction. Select models such as PLS-DA (Partial Least Squares Discriminant Analysis), SVM (Support Vector Machine), RF (Random Forest), ANN (Artificial Neural Network), ResNet50 (Deep Residual Network), and BiA-GRU-LSTM for training. Evaluate the performance of the trained models from multiple dimensions using metrics such as confusion matrix, source tracing accuracy, recall, precision, and F1 score to select the optimal source tracing model. In this embodiment, the feature dataset is divided into a preset ratio of 8:2, using stratified sampling to create a training set and a prediction set. This ensures that the distribution ratio of different enterprise samples within each dataset is consistent with the original data, avoiding data partitioning bias. Regarding model construction, existing technologies primarily rely on traditional machine learning models such as SVM and RF, lacking a comparative approach combining deep learning with traditional models. This embodiment selects multiple machine learning and deep learning models for training, including Partial Least Squares Discriminant Analysis (PLS-DA), Support Vector Machine (SVM), Random Forest (RF), Artificial Neural Network (ANN), ResNet50 (Deep Residual Network), and BiA-GRU-LSTM. Optimal hyperparameter combinations for each model are determined through 5-fold cross-validation. For evaluation, the trained models are primarily compared based on their prediction set source tracing accuracy and F1 score. The model with the best performance is selected as the optimal model, ensuring good generalization ability and source tracing reliability. The model's performance results are shown in Table 1.

[0036] Table 1 Model running results ; The optimal source tracing model is BiA-GRU-LSTM, with a prediction set source tracing accuracy of 96.97% and an F1 score of 96.67%. Figure 8 As shown, the BiA-GRU-LSTM model structure includes a parallel bidirectional gated recurrent unit (Bi-GRU) and a bidirectional long short-term memory network (Bi-LSTM). The outputs of the two are concatenated through the feature dimension and multi-head attention is added. The attention weights are weighted and aggregated to generate attention-weighted feature representations. Then, layer normalization and batch standardization are performed sequentially, followed by two layers of repeated processing. Finally, the ReLU activation function is used to output the model result.

[0037] The repetition processing layer includes sequential linear transformation, batch normalization, and dropout processing. Furthermore, the specific number of repetition processing layers can be adjusted according to actual application requirements.

[0038] Specifically, the model structure mainly consists of a parallel architecture of a bidirectional gated recurrent unit (Bi-GRU) and a bidirectional long short-term memory network (Bi-LSTM). The Bi-GRU structure efficiently learns key features with a simple gating mechanism, while the Bi-LSTM structure excels at capturing long-sequence dependencies to extract complex temporal or spatial correlation features from 3D fluorescence data. The outputs of both are concatenated to achieve feature complementarity, thereby capturing the temporal correlation of fluorescence signals (continuous features of excitation-emission wavelengths) and exhibiting stronger recognition capabilities for components with similar peak positions but different peak shape details. Simultaneously, multi-head attention is incorporated to perform attention-weighted aggregation on the concatenated fused features, generating weighted feature representations. Layer normalization is used to stabilize the feature distribution during this process, enabling it to adaptively focus on key feature regions in the 3D fluorescence data and enhance the specificity of feature representation. In the feature processing stage, the model stabilizes the training distribution and accelerates convergence through batch normalization, gradually compresses the feature dimension to the classification space using linear layers, enhances nonlinear expression capabilities with the ReLU activation function, and further suppresses overfitting through two dropout layers. During training, the model performance was optimized by fine-tuning the model parameters. Finally, the optimal model in 3D fluorescence data processing was selected through cross-validation, which has comprehensive feature capture, key information focusing ability and anti-overfitting ability. The results show that the model is suitable for the accurate classification and analysis of 3D fluorescence data. It can efficiently capture nonlinear fluorescence features in complex matrices, and its component identification ability far exceeds that of traditional single models. The accuracy and discrimination of source tracing are significantly improved.

[0039] S5. Use the optimal source tracing model to process the newly collected source water samples using the same preprocessing procedure. Input the processed new water sample characteristic data into the optimal model for prediction and output the source tracing results. Based on the source tracing results, accurately trace the source to the corresponding discharge enterprise. If there are water samples from unknown sources that cannot be identified, update the database in a timely manner through transfer learning to improve the accuracy of source tracing. The above source tracing results include the name of the emission source with the highest matching degree, the matching degree value, and the confidence interval; When the predicted matching degree is greater than or equal to the preset threshold, it is determined that the water sample to be identified comes from the corresponding discharge source; When the matching degree is less than the preset threshold, it is determined to be a water sample from an unknown source or a mixed polluted water sample.

[0040] In the pollution source tracing method of this embodiment, existing technologies rely on a large amount of labeled data for new sample processing, which is prone to failure when faced with wastewater from different processes. The transfer learning method of this embodiment can reuse existing training results and can adapt to pollution samples from new processes and niche industries without the need to recollect massive amounts of data.

[0041] Specifically, the optimal source tracing model is loaded into the data processing system for pollution source tracing of water samples from unknown sources. Newly collected water samples from unknown sources undergo standardized processing according to a strict preprocessing procedure to ensure consistent processing conditions between the unknown samples and training samples. The preprocessed three-dimensional fluorescence data of the unknown water samples is decomposed using the PARAFAC decomposition algorithm to extract core feature vectors. These extracted core feature vectors are input into the optimal source tracing model, which calculates and outputs prediction results, including the name of the emitting company with the highest matching degree, the matching degree value, and the confidence interval. A matching degree threshold is set, such as ≥90%. When the predicted matching degree is ≥ the threshold, the unknown water sample is determined to originate from the corresponding emitting company; when the matching degree is < the threshold, it is indicated as an unknown source water sample or a mixed polluted water sample. Simultaneously, auxiliary verification methods such as cosine similarity and Euclidean distance can be used to further confirm the reliability of the prediction results. For newly emerging water samples from unknown sources, which may be contaminated by niche industries or new production processes, transfer learning can be used to reuse existing training results. The previously trained model can be transferred to contaminated samples from new processes or niche industries without collecting large amounts of new data, thus expanding the database's coverage of niche industries, special processes, or emerging pollution types. Through this efficient database update mechanism, the system's ability to identify unknown contaminated water samples is continuously improved, accurately outputting pollution source tracing results. This ensures the timeliness and comprehensiveness of water pollution tracing in industrial parks, ultimately achieving rapid location and accountability for enterprises discharging pollutants into industrial parks.

[0042] The pollution source tracing method in this embodiment has the following advantages: (1) By combining preprocessing enhancement features with PARAFAC, the influence of fluorescence peak overlap and masking can be reduced, thus reducing source tracing bias. SG smoothing can reduce the interference of noise and scattering on fluorescence peaks, avoid the "pseudo-overlap" phenomenon caused by the superposition of noise or scattering signals and target peaks, make the outline of overlapping peaks clearer, and indirectly improve the accuracy of feature peak separation; BiA-GRU-LSTM deep learning model has the advantage of being able to capture the temporal correlation of fluorescence signals (continuous features of excitation-emission wavelengths), and has a stronger ability to identify components with similar peak positions but different peak shape details; it can effectively solve the problem of source tracing bias caused by fluorescence peak overlap and masking. (2) For the same type of pollution source from different processes, the loading matrix corresponding to the core fluorescent component will have a high degree of consistency with the original value. The database does not need to cover all pollution samples from special processes. As long as the loading matrix features of the core fluorescent component are included, good accuracy can be obtained by tracing the source through model matching features. For pollution samples from niche industries, it is only necessary to cover pollution samples from all industries in the industrial park when establishing the database, or to collect pollution samples from newly emerging industries in a timely manner. The previously trained model can be transferred to pollution samples from new processes and niche industries through transfer learning. There is no need to collect a large amount of new data. Through this efficient database update mechanism, it is possible to accurately trace the source of emerging pollution types.

[0043] The above description is merely a detailed explanation of preferred embodiments and principles of the present invention. For those skilled in the art, there may be changes in specific implementation methods based on the ideas provided by the present invention, and these changes should also be considered within the scope of protection of the present invention.

Claims

1. A three-dimensional fluorescence spectrometer, comprising a light source module, a light source monochromator, a fluorescence monochromator, a sample cell, and a PMT detector, wherein both the light source monochromator and the fluorescence monochromator have an entrance slit and an exit slit, characterized in that, Both the entrance slit and the exit slit are continuously adjustable structures. The continuously adjustable slit structure includes a base frame and a slit opening disposed on the base frame, a first slit adjustment block, a second slit adjustment block, a slit adjustment gear, a transmission gear, and a stepper motor. The slit adjustment gear is coaxially distributed with the slit opening, the slit adjustment gear meshes with the transmission gear, and the transmission gear is driven and connected to the stepper motor. The first slit adjustment block and the second slit adjustment block are movably connected to two parallel guide columns, which are fixedly installed on the base frame. The first slit adjustment block and the second slit adjustment block are distributed on both sides of the slit opening. Each of the two guide columns is fitted with a return spring. The two ends of one return spring abut against the first slit adjustment block and the base frame, and the two ends of the other return spring abut against the second slit adjustment block and the base frame. The slit adjusting gear has a first limiting post and a second limiting post at both ends along its central axis. The first limiting post abuts against the second slit adjusting block, and the second limiting post abuts against the first slit adjusting block. By rotating the slit adjusting gear, the first slit adjusting block and the second slit adjusting block are adjusted to move closer or further away from each other relative to the slit opening, so as to adjust the slit width.

2. The three-dimensional fluorescence spectrometer according to claim 1, characterized in that, The first slit adjustment block includes a first movable connecting block and a first slit piece disposed on the first movable connecting block, the first slit piece being used to block the slit opening; The second slit adjustment block includes a second movable connecting block and a second slit plate disposed on the second movable connecting block, the second slit plate being used to block the slit opening; The first slit plate and the second slit plate are brought closer together or further apart to adjust the slit width.

3. The three-dimensional fluorescence spectrometer according to claim 2, characterized in that, The first movable connecting block has an extended first abutting arm, the first abutting arm and the first movable connecting block are respectively located on both sides of the slit opening, and the second limiting post abuts against the side of the first abutting arm away from the first movable connecting block; The second movable connecting block has an extended second abutment arm, which is located on both sides of the slit opening, and the first limiting post abuts against the side of the second abutment arm away from the second movable connecting block.

4. The three-dimensional fluorescence spectrometer according to claim 3, characterized in that, The first slit adjustment block and the second slit adjustment block are centrally symmetrical to each other.

5. The three-dimensional fluorescence spectrometer according to any one of claims 1-4, characterized in that, Both the first limiting post and the second limiting post are bearing structures.

6. The three-dimensional fluorescence spectrometer according to any one of claims 1-4, characterized in that, The base frame is equipped with a positioning optical coupler, and the first slit adjustment block or the second slit adjustment block is equipped with an optical coupler baffle. The optical coupler baffle and the positioning optical coupler are used for positioning cooperation to adjust the slit width.

7. The three-dimensional fluorescence spectrometer according to any one of claims 1-4, characterized in that, The base frame is also equipped with a filter turntable and several bandpass filters arranged circumferentially on the filter turntable. The filter turntable is connected to a drive motor for driving. The rotation of the filter turntable switches the bandpass filter to correspond with the slit opening.

8. The three-dimensional fluorescence spectrometer according to claim 7, characterized in that, The base frame is also equipped with a positioning sensor for aligning the center of the bandpass filter with the center of the slit.

9. The pollution source tracing method of the three-dimensional fluorescence spectrometer as described in any one of claims 1-8, characterized in that, Includes the following steps: S1. Acquire a three-dimensional fluorescence spectral dataset using a three-dimensional fluorescence spectrometer and perform data preprocessing; S2. Perform parallel factor decomposition on the preprocessed three-dimensional fluorescence spectral data to decompose it into excitation load matrix, emission load matrix and relative concentration load matrix to form a feature dataset. S3. Train the BiA-GRU-LSTM model using the feature dataset to obtain the source tracing model. The BiA-GRU-LSTM model includes parallel Bi-GRU and Bi-LSTM. The outputs of the two are concatenated through the feature dimension and multi-head attention is added. The attention weights are weighted and aggregated to generate attention-weighted feature representations. Then, layer normalization and batch standardization are applied sequentially, followed by several repeated processing layers. Finally, the ReLU activation function is used to output the model result. The repeated processing layers include linear transformation, batch standardization, and dropout layers in sequence. S4. Collect three-dimensional fluorescence spectral data of the water sample to be identified using a three-dimensional fluorescence spectrometer, perform data preprocessing and parallel factor decomposition, input the data into the source tracing model for prediction, and output the source tracing results.

10. The data processing method according to claim 9, characterized in that, The source tracing results include the name of the emission source with the highest matching degree, the matching degree value, and the confidence interval; When the predicted matching degree is greater than or equal to the preset threshold, it is determined that the water sample to be identified comes from the corresponding discharge source; When the matching degree is less than the preset threshold, it is determined to be a water sample from an unknown source or a mixed polluted water sample.

Images