Unattended method and system based on Internet of Things technology

Abstract

The invention provides an unattended operation method and system based on the Internet of Things technology, and the method comprises the steps: carrying out the multi-band scanning of a target substance through a tunable laser spectrum device of a distributed laser detection node, generating an original spectrum signal, carrying out the dynamic wavelength modulation and adaptive noise filtering processing, and carrying out the detection of the original spectrum signal. And matching and outputting a preliminary recognition result in combination with the dangerous substance feature library. And performing confidence evaluation and correction on the result by using a local federal learning model, generating a parameter update quantity through correlation analysis, and uploading the parameter update quantity to a central server. And dynamically updating a local model and laser scanning parameters based on the issued federal global model parameter weight distribution, and realizing accurate identification of the dangerous substances. According to the invention, the accuracy of dangerous substance identification and the self-adaptive capability of the system are improved.

Description

Technical Field

[0001] This application relates to the field of Internet of Things technology, and particularly to an unattended method and system based on Internet of Things technology. Background Art

[0002] In key places such as chemical industrial parks and border security inspections, there is an urgent need for an unattended detection system that can identify dangerous substances in real time and accurately, which requires the ability to detect multiple substances concurrently, anti-environmental interference, and a continuous self-optimization mechanism. At the same time, the security and privacy of detection data at each node need to be ensured.

[0003] The current existing solution uses a static spectral analysis system based on deep learning. It obtains substance spectral data by laser scanning at a fixed wavelength, uses a pre-trained convolutional neural network model for feature extraction and classification, and combines regular updates of the cloud model to improve the recognition accuracy.

[0004] This solution depends on preset scanning parameters and is difficult to dynamically adapt to spectral interference in complex environments; the centralized model update leads to the problem of data islands between nodes, and the generalization ability of the static network for new dangerous substances is limited, and frequent manual intervention is required for adjustment. Summary of the Invention

[0005] This application provides an unattended method and system based on Internet of Things technology to solve the problems of low accuracy in identifying dangerous substances and poor system adaptability in the prior art.

[0006] In a first aspect, this application provides an unattended method based on Internet of Things technology, including:

[0007] Using the tunable laser spectroscopy device in the distributed laser detection node to perform multi-band scanning on the target substance to generate an original spectral signal;

[0008] Performing dynamic wavelength modulation and adaptive noise filtering processing on the original spectral signal, and combining with a preset dangerous substance feature library to generate a preliminary recognition result;

[0009] Based on a preset local federated learning model, performing confidence evaluation on the preliminary recognition result to obtain a corrected recognition result;

[0010] Performing correlation analysis on the corrected recognition result and the current local parameters of the local federated learning model to generate updated local parameters;

[0011] Uploading the updated local parameters to a central server and receiving the parameter weight distribution of the federated global model;

[0012] Update the preset local federated learning model based on the parameter weight distribution of the federated global model, and dynamically adjust the scanning parameter configuration information of the tunable laser spectroscopy device, so as to identify the hazardous substances in the target substance through the updated local federated learning model. The federated global model is obtained by the central server aggregating the updated local parameters uploaded by multiple distributed laser detection nodes.

[0013] Optionally, the confidence evaluation of the preliminary identification result is performed based on the preset local federated learning model to obtain a corrected identification result, including:

[0014] Calculate the confidence score of the preliminary identification result in the feature space through the probability distribution model in the local federated learning model;

[0015] When the confidence score is lower than the preset confidence threshold, activate the error backpropagation path in the local federated learning model;

[0016] Perform error compensation calculation on the multi-spectral features of the preliminary identification result through the error backpropagation path to generate a compensated feature vector;

[0017] Perform phase synchronization superposition on the compensated feature vector and the preliminary identification result in the preset time-frequency joint domain to generate a corrected identification result.

[0018] Optionally, the performing phase synchronization superposition on the compensated feature vector and the preliminary identification result in the preset time-frequency joint domain to generate a corrected identification result includes:

[0019] Construct a multi-spectral phase distribution map of the preliminary identification result through the hidden layer activation state of the local federated learning model;

[0020] Based on the multi-spectral phase distribution map, perform time-frequency rasterization recombination on the compensated feature vector to generate a recombined feature vector with phase continuity with the preliminary identification result;

[0021] Establish a dynamic carrier modulation channel within the phase synchronization framework of the preset time-frequency joint domain, and in the dynamic carrier modulation channel, perform convolution processing on the recombined feature vector and the preliminary identification result;

[0022] Perform phase correction on the convolution processing result according to the phase synchronization parameters of the preset time-frequency joint domain to generate a corrected identification result.

[0023] Optionally, the performing convolution processing on the recombined feature vector and the preliminary identification result includes:

[0024] Generate a multi - band carrier group that matches the multi - spectral features of the preliminary recognition result in the dynamic carrier modulation channel;

[0025] Perform band - pass constraint modulation on the multi - band carrier group through the multi - spectral phase distribution map to generate a modulated carrier base;

[0026] Perform base - band loading on the recombined feature vector and the modulated carrier base in the preset time - frequency joint domain to generate a modulated feature signal;

[0027] Based on the multi - spectral phase distribution map, perform time - frequency window segmentation on the preliminary recognition result to generate a reference signal after multi - scale sub - band decomposition;

[0028] Establish a feedback loop within the dynamic carrier modulation channel. In the feedback loop, perform multi - scale sub - band convolution on the modulated feature signal and the reference signal.

[0029] Optionally, the associative analysis of the corrected recognition result and the current local parameters of the local federated learning model to generate updated local parameters includes:

[0030] Construct a feature space mapping function based on the hidden - layer activation state of the local federated learning model. Based on the feature space mapping function, calculate the feature correlation matrix between the corrected recognition result and the current local parameters of the local federated learning model;

[0031] Determine the set of basis vectors to be optimized in the parameter space of the local federated learning model according to the singular - value decomposition result of the feature correlation matrix;

[0032] Perform orthogonal projection on the set of basis vectors using the gradient direction of the corrected recognition result to generate a parameter update amount;

[0033] Perform geodesic interpolation on the parameter update amount and the current local parameters of the local federated learning model to generate updated local parameters.

[0034] Optionally, the processing of dynamically modulating the wavelength and adaptively filtering the noise of the original spectral signal, and combining with a pre - set hazardous substance feature library to generate a preliminary recognition result includes:

[0035] Construct a dynamic wavelength modulation channel based on the wavelength tuning characteristics of the tunable laser spectroscopy device, and generate a modulation wavelength sequence that matches the absorption spectral line of the target substance in the dynamic wavelength modulation channel;

[0036] Perform segmented wavelength scanning on the original spectral signal through the modulation wavelength sequence to generate a modulated spectral signal;

[0037] Analyze the noise distribution characteristics of the modulated spectral signal in the preset time-frequency joint domain, and generate a noise suppression template related to the characteristics of the target substance;

[0038] Perform selective filtering on the modulated spectral signal based on the noise suppression template to generate enhanced spectral data;

[0039] Establish a multi-dimensional matching space in the preset hazardous substance feature library, map the enhanced spectral data into the multi-dimensional matching space, and calculate the similarity between the enhanced spectral data and the characteristic spectra of each hazardous substance in the hazardous substance feature library;

[0040] Use the class label of the hazardous substance with the maximum similarity as the class information of the target substance to obtain a preliminary identification result including the class information of the target substance.

[0041] Optionally, updating the preset local federated learning model based on the parameter weight distribution of the federated global model and dynamically adjusting the scanning parameter configuration information of the tunable laser spectroscopy device includes:

[0042] Based on the structural constraints of the local federated learning model, perform parameter space mapping on the feature subspace components in the parameter weight distribution to generate a weight update amount compatible with the structure of the local federated learning model;

[0043] Calculate a weight fusion coefficient based on the feature correlation of the enhanced spectral data and the weight update amount;

[0044] Perform adaptive weighted fusion on the weight update amount according to the weight fusion coefficient to generate an updated local federated learning model;

[0045] Establish an association mapping function between the scanning parameters of the tunable laser spectroscopy device and the feature response of the updated local federated learning model;

[0046] Jointly optimize the wavelength tuning range, scanning step size, and integration time of the tunable laser spectroscopy device through the association mapping function, and dynamically adjust the scanning parameter configuration information according to the optimization results.

[0047] In a second aspect, the present application provides an unattended system based on Internet of Things technology, including:

[0048] A scanning module for performing multi-band scanning on a target substance using the tunable laser spectroscopy device in the distributed laser detection node to generate an original spectral signal;

[0049] A processing module for performing dynamic wavelength modulation and adaptive noise filtering on the original spectral signal, and generating a preliminary identification result in combination with a preset hazardous substance feature library;

[0050] An evaluation module, configured to evaluate the confidence of the preliminary recognition result based on a preset local federated learning model to obtain a corrected recognition result;

[0051] An analysis module, configured to perform correlation analysis on the corrected recognition result and the current local parameters of the local federated learning model to generate updated local parameters;

[0052] A receiving module, configured to upload the updated local parameters to a central server and receive the parameter weight distribution of the federated global model;

[0053] An adjustment module, configured to update the preset local federated learning model based on the parameter weight distribution of the federated global model, and dynamically adjust the scanning parameter configuration information of the tunable laser spectroscopy device, so as to identify hazardous substances in the target substance through the updated local federated learning model. The federated global model is obtained by the central server aggregating the updated local parameters uploaded by multiple distributed laser detection nodes.

[0054] In a third aspect, the present application provides a computing device, including a processor and a memory. A computer program is stored in the memory, and the processor is configured to run the computer program to execute any one of the methods for unattended operation based on Internet of Things technology in the first aspect.

[0055] In a fourth aspect, the present application provides a computer storage medium, on which computer program instructions are stored. When the computer program instructions are executed by a processor, they implement any one of the methods for unattended operation based on Internet of Things technology in the first aspect.

[0056] In the present application, a method for unattended operation based on Internet of Things technology is provided. The method includes: using a tunable laser spectroscopy device in the distributed laser detection node to perform multi-band scanning on a target substance to generate an original spectral signal;

[0057] Performing dynamic wavelength modulation and adaptive noise filtering processing on the original spectral signal, and combining a preset hazardous substance feature library to generate a preliminary recognition result;

[0058] Based on a preset local federated learning model, perform confidence evaluation on the preliminary recognition result to obtain a corrected recognition result; perform correlation analysis on the corrected recognition result and the current local parameters of the local federated learning model to generate updated local parameters; upload the updated local parameters to the central server and receive the parameter weight distribution of the federated global model; update the preset local federated learning model based on the parameter weight distribution of the federated global model, and dynamically adjust the scanning parameter configuration information of the tunable laser spectroscopy device, so as to realize the recognition of hazardous substances in the target substance through the updated local federated learning model. The federated global model is obtained by the central server aggregating the updated local parameters uploaded by multiple distributed laser detection nodes.

[0059] The technical solution of this application has the following beneficial effects:

[0060] This application uses tunable lasers to achieve high-resolution spectral acquisition, improving the ability to capture substance characteristics. It enhances the signal-to-noise ratio and suppresses the influence of environmental interference on spectral data. It quickly completes the preliminary classification of substances, reducing the computational complexity. It quantifies the recognition reliability through a probability model, reducing the misjudgment rate. It establishes a mapping relationship between the recognition result and the model parameters to achieve targeted optimization. It realizes knowledge sharing between distributed nodes while protecting data privacy. It adaptively optimizes the detection strategy to improve the subsequent scanning efficiency.

[0061] Furthermore, this application also evaluates the confidence of the preliminary recognition result through the probability distribution model of the local federated learning model. When the confidence is insufficient, it activates the error backpropagation path, compensates for the error of multi-spectral features to generate a compensated feature vector, and finally achieves precise correction of the recognition result through phase synchronization superposition in the time-frequency joint domain.

[0062] Moreover, this mechanism improves the robustness of hazardous substance recognition. It effectively corrects the recognition deviation caused by environmental interference or substance mixing through error compensation and phase synchronization technologies. At the same time, the federated learning framework ensures the continuous evolution ability of the correction strategy, enabling the system to maintain a high accuracy rate in complex scenarios.

[0063] These aspects or other aspects of this application will be more clearly understood in the following description of the embodiments. Description of the Drawings

[0064] In order to more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the following will briefly introduce the drawings required for the description of the embodiments or the prior art. Obviously, the following drawings are some embodiments of this application. For those of ordinary skill in the art, without creative efforts, other drawings can also be obtained based on these drawings.

[0065] Figure 1 Flowchart of an unattended method based on Internet of Things technology provided by an embodiment of the present application;

[0066] Figure 2 Schematic structural diagram of an unattended system based on Internet of Things technology provided by an embodiment of the present application;

[0067] Figure 3 Schematic structural diagram of a computing device provided by an embodiment of the present application. Detailed implementation manners

[0068] In order to enable those skilled in the art to better understand the solutions of the present application, the technical solutions in the embodiments of the present application will be clearly and completely described below in conjunction with the accompanying drawings in the embodiments of the present application.

[0069] In some processes described in the specification, claims and the above-mentioned drawings of the present application, there are multiple operations that appear in a specific order. However, it should be clearly understood that these operations may not be executed in the order in which they appear herein or may be executed in parallel. The serial numbers of the operations, such as 101, 102, etc., are only used to distinguish different operations, and the serial numbers themselves do not represent any execution order. In addition, these processes may include more or fewer operations, and these operations may be executed in sequence or in parallel. It should be noted that the descriptions such as "first" and "second" in this article are used to distinguish different messages, devices, modules, etc., and do not represent a sequence, nor do they limit that "first" and "second" are of different types.

[0070] Researchers found that existing hazardous substance detection systems have problems such as poor environmental adaptability, large interference on the recognition accuracy rate, and difficulty in coordinating and optimizing the data of each detection node in isolation. Based on this, an embodiment of the present application provides an unattended method based on Internet of Things technology. This method can obtain high-quality spectral data through tunable laser multi-band scanning, realize anti-interference recognition by combining dynamic wavelength modulation and federated learning framework, and continuously optimize the detection model by using the collaborative training of distributed nodes, and finally realize the accurate and adaptive recognition of hazardous substances. The technical solution of the present application is applicable to scenarios such as chemical industrial parks and border security inspections that require highly reliable monitoring of hazardous substances.

[0071] The technical solutions in the embodiments of the present application will be clearly and completely described below in conjunction with the accompanying drawings in the embodiments of the present application. Obviously, the described embodiments are only a part of the embodiments of the present application, rather than all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative efforts shall fall within the protection scope of the present application.

[0072] Figure 1The figure below is a flowchart of an unattended method provided by an embodiment of this application based on Internet of Things technology. As Figure 1 shown, the method includes:

[0073] Step 101: Use the tunable laser spectroscopy device in the distributed laser detection node to perform multi-band scanning on the target substance to generate an original spectral signal.

[0074] In this step, the tunable laser spectroscopy device refers to an optical device that performs multi-band scanning by adjusting the laser wavelength, and its output wavelength range covers the characteristic absorption spectral lines of the target substance. The target substance refers to the suspicious hazardous substances that need to be detected and identified, including chemicals such as organic solvents (such as acetone, benzene series), corrosive substances (such as strong acids and alkalis), explosives such as black powder, and toxic substances such as cyanides and organophosphorus pesticides. Its identification characteristics are reflected in the specific molecular vibration / rotation absorption spectrum. Multi-band scanning means continuous or discrete laser irradiation at multiple characteristic absorption wavelengths. The original spectral signal refers to the time-domain light intensity sequence obtained by laser scanning, which contains the absorption / scattering response of the target substance to lasers of different wavelengths.

[0075] In an embodiment of this application, when the distributed laser detection node is triggered by a motion sensor or a pressure sensing device, the tunable laser spectroscopy device in the distributed laser detection node is started. First, the scanning parameters of the tunable laser (including start wavelength, end wavelength, step size, etc.) are configured, and then the laser is controlled to irradiate the target substance according to the set parameters. At the same time, the transmitted / reflected light signal is collected by a photodetector and an original spectral signal containing the characteristic information of the substance is generated after analog-to-digital conversion. The key technologies include tunable laser control technology and spectral signal acquisition technology.

[0076] For example, in the safety monitoring system of a chemical plant, multiple distributed laser detection nodes are deployed, and each node is equipped with a tunable laser spectroscopy device. When it is necessary to detect whether there is a leakage of hazardous chemicals, the laser spectroscopy devices of each node are automatically started by the system to perform multi-band scanning on the surrounding environment and generate preliminary spectral signals.

[0077] Step 102: Perform dynamic wavelength modulation and adaptive noise filtering processing on the original spectral signal, and combine it with the pre-set characteristic library of hazardous substances to generate a preliminary identification result.

[0078] In this step, dynamic wavelength modulation refers to a wavelength fine-tuning technology that adjusts in real time according to the signal quality. Adaptive noise filtering refers to an intelligent noise reduction process based on signal characteristics. The characteristic library of hazardous substances refers to a pre-stored spectral characteristic database of standard substances. The preliminary identification result refers to the preliminary identification result of hazardous substance characteristics.

[0079] In the embodiments of the present application, first, the original spectrum is subjected to baseline correction and normalization processing, and then wavelet transform combined with Kalman filtering is used to achieve noise suppression. At the same time, the fine-tuning amount of the laser wavelength is dynamically adjusted according to the signal-to-noise ratio of the signal to enhance the characteristic peaks. The processed data is matched with the feature library for similarity, and the improved dynamic time warping algorithm is used to calculate the matching score, and the preliminary recognition result is output.

[0080] For example, continuing the above scenario, after receiving the original spectrum signals uploaded by each node, the system uses dynamic wavelength modulation and adaptive noise filtering processing technology for optimization, and matches them with the pre-stored characteristic library of hazardous chemicals to quickly screen out the possible hazardous chemicals.

[0081] Step 103: Based on a preset local federated learning model, perform a confidence evaluation on the preliminary recognition result to obtain a corrected recognition result.

[0082] In this step, the local federated learning model represents a lightweight deep learning model deployed on the node. Different nodes are located in different environments. The corrected recognition result includes: a substance category label (such as "acetone, concentration 500 ppm"), a confidence score (the corrected probability value, such as 88%), a characteristic band label (such as "main characteristic peak 3.4 μm"), an environmental interference evaluation (such as "water vapor interference degree: slight"), which is used to trigger a hierarchical alarm of the security system, guide the optimization of subsequent scanning parameters, and serve as a training sample for federated learning.

[0083] In the embodiments of the present application, the preliminary recognition result is input into the feature extraction network of the local model, and the probability distribution of each classification is calculated. When the maximum probability value is lower than the threshold, the attention mechanism of the model is activated to re-evaluate the key feature regions, and the corrected result is output in combination with the context information.

[0084] For example, based on the results of the previous step, the system uses a local federated learning model to evaluate the probability of the existence of each potential hazardous chemical, adjusts the initial recognition result, and improves the accuracy of the judgment.

[0085] Step 104: Perform an association analysis on the corrected recognition result and the current local parameters of the local federated learning model to generate updated local parameters.

[0086] In this step, the current local parameters represent trainable parameters such as model weights and biases. The association analysis means establishing a mapping relationship between the recognition result and the model parameters. The updated local parameters include convolution kernel weights (adjusted feature extraction parameters), attention mechanism parameters (optimized feature importance weights), classification layer biases (corrected decision boundary parameters), and temporal memory unit states, which are used to improve the recognition accuracy of subsequent similar substances, enhance the robustness of the model to current environmental interference, and serve as contribution parameters for federated learning.

[0087] In the embodiments of the present application, through the gradient backpropagation path of the model, the sensitivity matrix of the recognition result to each layer of parameters is calculated. The main feature directions are extracted by singular value decomposition, and the updated local parameters are generated based on the sensitivity distribution.

[0088] For example, the system further analyzes the relationship between the corrected recognition result and the existing model parameters, and adjusts the model parameters to better reflect the actual on-site situation.

[0089] Step 105: Upload the updated local parameters to the central server, and receive the parameter weight distribution of the federated global model.

[0090] In this step, the federated global model represents the aggregated model maintained by the central server. The parameter weight distribution represents the statistical characteristics of the parameters of each node.

[0091] In the embodiments of the present application, encrypting the local parameters means that after each node generates the local parameters, it will encrypt them to ensure data privacy and security, and prevent the parameters from being stolen or tampered with during transmission. The node uploads the updated encrypted parameters to the server, and the server uses a secure aggregation algorithm to calculate the global update. The weights of the updates of each node are allocated through an attention mechanism, and new global model parameters are generated and sent to each node.

[0092] For example, after all nodes complete the update of local parameters, they upload them to the central server. The central server synthesizes all the data, generates the latest global model parameters, and distributes them to each node.

[0093] Step 106: Update the preset local federated learning model based on the parameter weight distribution of the federated global model, and dynamically adjust the scanning parameter configuration information of the tunable laser spectroscopy device, so as to realize the identification of hazardous substances in the target substance through the updated local federated learning model. The federated global model is obtained by the central server aggregating the updated local parameters uploaded by multiple distributed laser detection nodes.

[0094] In this step, the scanning parameter configuration refers to the operating parameter settings of the laser, including the scanning band combination and the noise filtering threshold. The multiple distributed laser detection nodes refer to multiple detection devices deployed in an unattended environment. Each device operates independently and generates model parameters based on local data (such as laser spectroscopy data). The aggregation process of the central server means that after the central server receives the encrypted local parameters of all nodes, it merges these parameters through a specific aggregation algorithm (such as weighted average) to generate a global model.

[0095] In the embodiments of the present application, according to the feature importance analysis result of the global model, the laser scanning strategy of each node is dynamically adjusted. An reinforcement learning framework is adopted, with the recognition accuracy as the reward signal, to optimize parameters such as the scanning wavelength range and integration time. At the same time, the network structure and hyperparameters of the local model are updated.

[0096] For example, according to the update parameters sent by the central server, each node adjusts the scanning strategy of its laser spectroscopy device, improving the sensitivity and accuracy of hazardous substance detection. For the same target substance, the recognition results of different nodes in their respective corresponding environments may be inconsistent. Then, the present application can preferentially select the recognition result with the highest confidence as the final output.

[0097] This solution obtains high-quality spectral data through tunable laser scanning, combines intelligent signal processing and a federated learning framework, and realizes high-precision recognition of hazardous substances. The system has environmental adaptability. Each node continuously optimizes the detection performance through knowledge sharing, showing superior detection accuracy and robustness in scenarios such as chemical industry safety.

[0098] To solve the problem of misjudgment caused by environmental interference in hazardous substance recognition and further improve the recognition accuracy, in some embodiments, step 103: evaluating the confidence of the preliminary recognition result based on a preset local federated learning model to obtain a corrected recognition result, including:

[0099] Step 201: Calculate the confidence score of the preliminary recognition result in the feature space through the probability distribution model in the local federated learning model.

[0100] In step 201, the probability distribution model represents generating the probability distribution of each substance category based on the output layer of the classifier. The feature space refers to the multi-dimensional mathematical space spanned by the hidden layer activation states of the local federated learning model. Multi-dimensional includes the spectral feature characterization dimension, the model decision basis dimension, and the federated knowledge fusion dimension. The confidence score represents the prediction uncertainty measure calculated using the Bayesian algorithm, with a range of 0 - 100%.

[0101] In the embodiments of the present application, the spectral features corresponding to the preliminary recognition result are input into the local federated learning model, and the embedding vector in the feature space is obtained through forward propagation. Calculate the Mahalanobis distance from this vector to the centroids of each category, and combine the random sampling of the model to calculate the confidence score using the Bayesian approximation method. The final output includes: the category with the highest probability, the confidence score, and the analysis of the main interference factors.

[0102] Step 202: When the confidence score is lower than the preset confidence threshold, activate the error backpropagation path in the local federated learning model.

[0103] In step 202, the preset confidence threshold represents the quality threshold set according to the application scenario. The error backpropagation path represents the path of the automatic differentiation computational graph built into the model.

[0104] In the embodiment of the present application, the system monitors the confidence score in real time. When the detected score is lower than the threshold, the automatic differentiation computational graph of the model is activated. The intermediate variable gradient information of the current computational graph is retained, and a directional backpropagation path from the output layer to a specific convolutional layer (responsible for key feature extraction) is constructed. At the same time, the error compensation parameter matrix is initialized.

[0105] Step 203: Perform error compensation calculation on the multi-spectral features of the preliminary recognition result through the error backpropagation path to generate a compensated feature vector.

[0106] In step 203, the multi-spectral feature refers to the cross-band spectral characteristics contained in the preliminary recognition result. The generation path of the multi-spectral feature is as follows: the original spectral signal generated by multi-band scanning of the target substance by the tunable laser spectroscopy device, the enhanced spectral data obtained after dynamic wavelength modulation and adaptive noise filtering processing, and the feature components related to substance recognition are extracted during the multi-dimensional matching process in the hazardous substance feature library. These feature components retain the spectral resolution characteristics of the original multi-band scanning and constitute the multi-spectral features of the preliminary recognition result. The compensated feature vector represents the corrected 32-dimensional feature representation.

[0107] In the embodiment of the present application, the gradient of the preliminary recognition result in each feature dimension is calculated along the activated backpropagation path. The gated attention mechanism is used to perform weighted compensation on the feature dimensions with high gradient. The gradient information is converted into a feature compensation amount through a three-layer fully connected network to generate a compensation vector that matches the original feature dimension. The compensation calculation method is as follows: along the activated error backpropagation path, calculate the gradient value of the preliminary recognition result in each feature dimension, and quantify the contribution degree of the feature to the classification error. A gated mechanism (such as the Sigmoid (Logistic Sigmoid Function, Sigmoid) function) is used to generate an attention weight matrix, and higher compensation weights are assigned to the feature dimensions with high gradient values. For example, the formula is: w i = σ(g i ·θ); where, w i is the attention weight of the i-th dimensional feature, with a range of [0, 1]. g i is the gradient value of the i-th dimensional feature, reflecting the contribution degree of this feature to the classification error. θ is a learnable parameter used to adjust the scaling ratio of the gradient to the weight. σ is the Sigmoid function that maps the input to the interval [0, 1], and the formula is The gradient information is mapped to a compensation vector through a three-layer fully connected neural network (with the ReLU activation function). The network input is the gradient vector, and the output is a compensation feature vector with the same dimension as the original features. The compensation amplitude is dynamically scaled according to the real-time signal-to-noise ratio (SNR). For example:

[0108] Step 204: Phase-synchronously superimpose the compensation feature vector and the preliminary recognition result in a preset time-frequency joint domain to generate a corrected recognition result.

[0109] In step 204, the time-frequency joint domain refers to a composite signal space that simultaneously includes the spectral temporal dynamic characteristics and the frequency-domain resolution characteristics, which is specifically spanned by the scanning time axis (time dimension) and the multi-band spectral resolution axis (frequency dimension) of the tunable laser spectroscopy device. Its phase synchronization characteristics are quantified by the phase distribution map extracted from the hidden layer activation state of the local federated learning model, which is used to unify the signal representation consistency of the compensation feature vector and the preliminary recognition result in the time and frequency dimensions. The technology used for superimposition: The phase-synchronous superimposition is achieved through a dynamic carrier modulation channel, specifically using a multi-scale sub-band convolution technology. In the time-frequency joint domain, the recombined compensation feature vector and the preliminary recognition result are subjected to phase rotation compensation for each sub-band according to the phase synchronization parameter, and the signal fusion is completed through the coherent synthesis of the orthogonal carrier group. This technology ensures that the superimposition process simultaneously satisfies the constraints of time-domain waveform continuity and frequency-domain energy conservation.

[0110] In the embodiment of this application, the compensation vector is reconstructed into a time-domain signal through an inverse wavelet transform and superimposed with the feature representation of the original recognition result in the complex domain. The instantaneous phase information is extracted using the Hilbert transform, and phase synchronization is achieved through a phase-locked loop technology. Finally, the normalized corrected recognition result is output.

[0111] The following is a specific example:

[0112] In the pipeline leakage monitoring of a chemical plant, the distributed nodes collect the original spectral signal of the target substance, which is preliminarily recognized as "ethylene, confidence level 68%") after processing. The spectral matching similarity score with the feature library is 75 points (calculated based on the dynamic time warping distance). Since a water vapor interference peak is detected at 4.8 μm, a 7% environmental interference correction is deducted. The method for determining the environmental interference deduction ratio is as follows: Calculate the energy ratio of the interference peak at 4.8 μm through time-frequency analysis (such as wavelet transform): where, E 干扰It is the proportion of the interference peak energy. S(f) is the frequency-domain representation of the spectral signal. f is the frequency variable, corresponding to the wavelength of the spectrum (for example, the frequency at 4.8 μm is f = c / λ, where c is the speed of light). df is the frequency element, and the integration interval is the band where the interference peak is located and the full band. The standard energy attenuation coefficients of various interfering substances (such as water vapor) are pre-stored in the hazardous substance feature library. For example, the attenuation coefficient of water vapor at 4.8 μm is 0.07 (experimentally calibrated value). Adjust the deduction ratio according to real-time environmental parameters (such as humidity). For example: If 7% is deducted under the reference humidity and the current humidity is 1.2 times that of the reference, the adjusted deduction ratio is 8.4%. The sampling variance of the Monte Carlo method of the model itself brings an uncertainty of ±2%, so the final confidence level is 75% - 7% ± 2% = 68%. When the confidence level is lower than the preset threshold of 85%, the system activates the error backpropagation path. The specific operations include: calculating the gradient of the 64-dimensional feature vector (including ethylene characteristic peaks such as 2.7 μm and 3.3 μm) output by the model's third convolutional layer for the preliminary recognition result, and obtaining the weights of each dimension [0.12, 0.08,..., 0.15] (obtained by backpropagation derivation). The weight of the dimension corresponding to 4.8 μm affected by water vapor (weight 0.03) is reduced by 50%, and the 2.7 μm dimension (weight 0.12 becomes 0.18) is enhanced to generate a 32-dimensional compensation vector [0.18, 0.05,...,-0.11]. In the time-frequency joint domain (128×64 time-frequency matrix generated by the short-time Fourier transform), the compensation vector is reconstructed into a time-domain signal through the inverse wavelet transform, complex superimposed with the original feature, and the instantaneous phase of the original signal is extracted (such as the phase angle of 1.2 rad at 2.7 μm); the phase of the compensation signal is synchronized to within an error of ±0.1 rad through a phase-locked loop; weighted superposition is performed according to the signal-to-noise ratio (weight of the original signal is 0.6, weight of the compensation signal is 0.4), and finally the corrected recognition result "ethylene, confidence level 92%" is output. Exemplarily, the basic confidence level is adjusted to the result after interference deduction of the initial confidence level of 68% (75% similarity - 7% interference). After compensation, the 7% confidence level deducted due to interference is restored (68% becomes 75%). The feature optimization gain is that the enhancement of the 2.7 μm main characteristic peak means that the signal-to-noise ratio is increased (from 15 dB to 22 dB), bringing a +10% confidence gain. The calculation basis is that every 3 dB increase in the signal-to-noise ratio contributes approximately 4.3% to the confidence level (experimentally calibrated value); the interference suppression at 4.8 μm is that the weight of the interference peak is reduced by 60%, reducing the misjudgment probability by +5% confidence gain. The phase synchronization optimization means that the phase alignment in the time-frequency domain (±0.1 rad error) improves the feature consistency, bringing a +2% confidence gain. The comprehensive calculation is 75% (basic restoration) + 10% (main peak optimization) + 5% (interference suppression) + 2% (phase correction) = 92%.

[0113] In the embodiments of the present application, this solution reduces the false alarm rate caused by spectral interference through a dynamic error compensation mechanism. At the same time, the confidence evaluation result can be used as a system self-diagnosis index, providing a reliable basis for subsequent parameter optimization.

[0114] To solve the problem of recognition errors caused by phase mismatch during the spectral feature compensation process and further improve the accuracy of hazardous substance recognition, in some embodiments, step 204: The phase synchronization superposition of the compensation feature vector and the preliminary recognition result in the preset time-frequency joint domain to generate a corrected recognition result includes:

[0115] Step 301: Construct a multi-spectral phase distribution map of the preliminary recognition result through the hidden layer activation state of the local federated learning model.

[0116] In step 301, the hidden layer activation state refers to the set of non-linear responses of the neurons in the hidden layer of the local federated learning model to the input enhanced spectral data. Its numerical distribution encodes both the spectral feature abstract representation of the target substance and the phase correlation characteristics in the time-frequency joint domain, and is formed through the feature transformation and information integration of each hidden layer during the forward propagation process, providing a deep feature representation containing substance discrimination knowledge and signal phase structure for constructing the multi-spectral phase distribution map. The multi-spectral phase distribution map represents the instantaneous phase angle matrix (128×64) of each band extracted by the Hilbert transform.

[0117] In the embodiments of the present application, the spectral features corresponding to the preliminary recognition result are input into the local federated learning model, and the complex feature map output by the fourth convolutional layer is extracted. The Hilbert transform is performed on each feature channel to calculate its instantaneous phase angle, generating a phase distribution matrix corresponding one-to-one to the spectral bands.

[0118] Step 302: Based on the multi-spectral phase distribution map, perform time-frequency grid reorganization on the compensation feature vector to generate a reorganized feature vector with phase continuity with the preliminary recognition result.

[0119] In step 302, the time-frequency grid reorganization means rearranging the compensation feature vector according to the phase distribution. The phase continuity means that the phase difference between the reorganized vector and the original feature is <0.1 rad.

[0120] In the embodiments of the present application, first, the short-time Fourier transform is performed on the compensation feature vector to obtain the time-frequency representation, and then the phase rotation adjustment is performed on each frequency band component according to the phase distribution map. Specifically, the phase difference matrix between the compensation vector and the original feature is calculated, a group of all-pass filters is designed for phase compensation, and the time-domain reorganized vector is generated through the inverse transform.

[0121] Step 303: Establish a dynamic carrier modulation channel within the phase synchronization framework of the preset time-frequency joint domain. In the dynamic carrier modulation channel, perform convolution processing on the recombined feature vector and the preliminary recognition result.

[0122] In step 303, the dynamic carrier modulation channel represents a configurable complex convolution operation path. Convolution processing means operating on the time-frequency matrix using a 3×3 complex convolution kernel.

[0123] In the embodiment of the present application, a dynamic convolution layer including 32 complex convolution kernels is constructed, and its center frequency is aligned with the feature band. After the real - imaginary decomposition of the recombined vector and the time-frequency matrix of the original features, convolution operations are respectively performed and then combined to output enhanced time-frequency features.

[0124] Step 304: According to the phase synchronization parameters of the preset time-frequency joint domain, perform phase correction on the convolution processing result to generate a corrected recognition result.

[0125] In step 304, the phase synchronization parameters include a group delay compensation factor (0 - 1) and a phase offset (rad). Phase correction represents a closed-loop adjustment based on a phase-locked loop.

[0126] In the embodiment of the present application, by monitoring the instantaneous frequency deviation of the convolution output signal, the group delay compensation factor (step size 0.01) and the phase rotation amount (step size 0.05 rad) are dynamically adjusted. A second-order phase-locked loop is used to achieve fast convergence, and finally a corrected recognition result with phase alignment is output.

[0127] The following is a specific example:

[0128] In the continuous scenario of chemical plant pipeline leakage monitoring, the system has generated a 32-dimensional compensation vector [0.18, 0.05,..., -0.11] and a preliminary identification result with a 68% confidence level. First, the 128-channel hidden layer activation state (including data such as the phase of 1.2 rad at the 2.7 μm band and 0.8 rad at the 3.3 μm band) is extracted through the fourth convolutional layer to construct a multi-spectral phase distribution map (a 128×64 matrix with an accuracy of 0.01 rad). Then, the compensation vector is reorganized by time-frequency rasterization, converted into a time-frequency representation through the short-time Fourier transform, and a rotation of -0.15 rad is applied to the 2.7 μm component according to the phase distribution map (compensating for the difference between the original 1.2 rad and the compensated signal of 1.35 rad). After reorganization, the phase difference of the vector in the 2.7 μm band is reduced to 0.05 rad. Subsequently, in the dynamic carrier modulation channel, a complex convolutional kernel with a central frequency of 3.3 THz (real part [0.12, -0.05; 0.08, 0.03], imaginary part [0.07, 0.02; -0.04, 0.09]) is configured to perform a convolution operation (stride 1) on the time-frequency matrices of the reorganized vector and the original features, and a fused feature with a 6 dB improvement in signal-to-noise ratio is output. Finally, the phase of the 3.3 μm band is stabilized at 0.8±0.03 rad through a second-order phase-locked loop (bandwidth 50 Hz, damping coefficient 0.7), further improving the ethylene identification confidence level from 92% to 96%. The key improvement lies in that the phase consistency error in the 2.7 μm band is reduced from ±0.1 rad to ±0.03 rad, and the residual energy of the 4.8 μm interference peak is reduced by 40%. This process enables the system to reduce the false alarm rate from 5% to 2% while maintaining a 98% recall rate.

[0129] In the embodiments of the present application, this phase synchronization and superposition scheme improves the accuracy of spectral feature fusion, making the substance identification result more reliable and stable. Through precise phase matching and dynamic carrier modulation, the system effectively eliminates the feature distortion caused by environmental interference and significantly reduces the occurrence of false alarms. At the same time, this scheme maintains excellent substance detection ability, demonstrates strong adaptability and robustness in complex industrial environments, and provides more accurate technical support for hazardous substance monitoring.

[0130] To solve the problem of band mismatch in multi-spectral feature fusion and further improve the accuracy of feature compensation, in some embodiments, step 303: the convolution processing of the reorganized feature vector and the preliminary identification result includes:

[0131] Step 401: Generate a multi-band carrier group that matches the multi-spectral features of the preliminary identification result in the dynamic carrier modulation channel.

[0132] In step 401, the multi-band carrier group refers to a set of carriers composed of complex sine waves, and the center frequency of each carrier is aligned with the characteristic band. The multi-band includes a core characteristic band, an auxiliary reference band, a guard interval band, and a dynamic expansion band.

[0133] In the embodiment of the present application, according to the frequency components corresponding to each characteristic band (such as 2.7μm, 3.3μm, etc.) in the preliminary recognition result, complex carriers with corresponding center frequencies are generated. Specifically, a baseband carrier is generated through a direct digital frequency synthesizer, the carrier power is allocated according to the spectral feature importance, and a carrier frequency-band mapping table is established.

[0134] Step 402: Perform band-pass constraint modulation on the multi-band carrier group through the multi-spectral phase distribution map to generate a modulated carrier base.

[0135] In step 402, the band-pass constraint modulation means restricting the effective frequency band range of the carrier. This modulation process constrains the frequency domain distribution of the carrier group through the spectral feature correlation of the multi-spectral phase distribution map. Specifically: according to the absorption characteristics of the target substance in each band, effective bands matching the substance characteristics are dynamically selected in the multi-band carrier group, and the carrier components of the irrelevant bands are suppressed, so that the generated modulated carrier base only retains the spectral channel information related to the detection of dangerous substances. The modulation process adopts a phase-frequency band joint selection technology, and maps the phase distribution map into a frequency domain constraint weight function through the complex representation form of the carrier group. The specific modulation process: perform complex dot multiplication on each sub-carrier of the multi-band carrier group with the corresponding band of the multi-spectral phase distribution map to synchronize the carrier phase with the substance characteristic phase. Based on the correlation intensity of each band in the phase distribution map, a frequency domain mask matrix is generated, and an element-wise product operation is performed on the carrier group to filter out non-characteristic bands. Perform frequency-time domain transformation on the filtered carrier group to generate a modulated carrier base whose time domain waveform is strictly aligned with the substance characteristic phase, and its bandwidth is adaptively controlled by the spectral line sharpness of the phase distribution map. The modulated carrier base represents a set of carrier signals after band shaping.

[0136] In the embodiment of the present application, using the band energy information in the multi-spectral phase distribution map, a finite impulse response filter bank is designed to perform band-pass shaping on each carrier. By adjusting the filter cut-off frequency and roll-off coefficient, it is ensured that the carrier spectrum is accurately matched with the characteristic band.

[0137] Step 403: Perform baseband loading on the recombined feature vector and the modulated carrier base in the preset time-frequency joint domain to generate a modulated feature signal.

[0138] In step 403, the baseband loading represents the modulation process of mapping the feature vector to the carrier base. The modulated feature signal represents a modulated wave carrying compensation information.

[0139] In the embodiments of the present application, the quadrature amplitude modulation method is adopted to modulate each dimensional component of the reconstructed feature vector onto the corresponding carrier wave. Specifically, it includes decomposing the feature vector, realizing up-conversion through a mixer, and synthesizing modulated signals of each waveband.

[0140] Step 404: Based on the multi-spectral phase distribution map, perform time-frequency window segmentation on the preliminary recognition result to generate a reference signal after multi-scale sub-band decomposition.

[0141] In step 404, time-frequency window segmentation refers to signal framing based on the feature waveband distribution. The reference signal refers to the time-frequency decomposition representation of the preliminary recognition result.

[0142] In the embodiments of the present application, a non-uniform time-frequency analysis window is designed according to the feature waveband interval, and the preliminary recognition result is processed as follows: a 5-ms Hanning window is used for the 2.7-μm waveband, a 3-ms rectangular window is used for the 3.3-μm waveband, and the overlap rate is set to 50%. A set of sub-band signals with waveband characteristics is generated.

[0143] Step 405: Establish a feedback loop in the dynamic carrier modulation channel. In the feedback loop, perform multi-scale sub-band convolution on the modulation feature signal and the reference signal.

[0144] In step 405, the feedback loop refers to a closed-loop control channel for real-time adjustment of convolution parameters. Multi-scale sub-band convolution refers to a joint time-frequency analysis operation performed on the modulation feature signal and the reference signal in the dynamic carrier modulation channel through multiple parallel convolution kernels with different time-frequency resolutions. The method used for convolution: construct a set of complex analytic wavelet basis functions with time-frequency localization characteristics as convolution kernels, perform complex-domain convolution operations on the signals in the dynamic carrier modulation channel, where the real part extracts the time-domain waveform correlation, the imaginary part captures the frequency-domain phase information, and the scale parameter and central frequency of the wavelet basis function are adjusted in real time using the feedback loop to make the convolution operation always focus on the time-frequency region with the most current spectral features.

[0145] In the embodiments of the present application, a feedback system including a delay-locked loop is established, and the dynamic adjustment refers to the variable size of each sub-band convolution kernel (from 3×3 to 7×7) and the adjustable convolution step (from 1 to 3). By monitoring the signal-to-noise ratio and phase consistency of the output signal, the convolution parameters are optimized in real time.

[0146] The following is a specific example:

[0147] In the continuous scenario of chemical plant pipeline leakage monitoring, based on the established 128×64 multi-spectral phase distribution map (including 1.2 rad phase data in the 2.7 μm band), the system first generates a multi-band carrier group with center frequencies of 3.7 THz (2.7 μm) and 4.5 THz (3.3 μm). The bandwidth of the 3.7 THz carrier is set to ±185 GHz (corresponding to a ±0.05 μm wavelength range), and this value is derived from the full width at half maximum characteristic of the C-H bond vibration spectrum of ethylene molecules. The carrier is band-pass constrained by designing a 128th-order filter, and the filter coefficients are calculated based on the energy gradient of the phase distribution map (passband ripple <0.1 dB). The 2.7 μm component (amplitude 0.18) of the reconstructed eigenvector is loaded onto the 3.7 THz carrier using a modulation method, and the modulation index is set to 0.8π to match the feature intensity. When performing time-frequency window segmentation on the preliminary recognition result, a 5 ms Hanning window (window function coefficients [0.08, 0.25,..., 0.08]) is used for the 2.7 μm band, and the window length is determined by the phase change rate (0.15 rad / ms) of this band. In the feedback loop, the convolution kernel size is dynamically adjusted by monitoring the signal-to-noise ratio (initial value 22 dB) of the convolution output in the 3.3 μm band. When the signal-to-noise ratio is lower than 25 dB, the kernel size is automatically expanded from 3×3 to 5×5 (the kernel weights are generated according to a two-dimensional Gaussian distribution, σ = 0.8). The final output fusion features keep the phase consistency error of the 2.7 μm band stable within the range of ±0.02 rad, and the signal-to-noise ratio of the 3.3 μm band is increased to 28 dB. On the basis of maintaining the original detection ability, the system further eliminates the residual spectral leakage interference.

[0148] In the embodiments of this application, this solution realizes the high-precision fusion of spectral features, effectively improving the accuracy and reliability of hazardous substance identification. Through dynamic carrier modulation and multi-scale convolution processing, the system can adaptively compensate for spectral distortion caused by environmental interference and maintain stable detection performance in complex industrial environments.

[0149] To solve the problem of accurately associating model parameter updates with substance identification features and further improve the knowledge fusion efficiency of federated learning, in some embodiments, step 104: The associating the corrected recognition result with the current local parameters of the local federated learning model to generate updated local parameters includes:

[0150] Step 501: Construct a feature space mapping function based on the hidden layer activation state of the local federated learning model, and calculate the feature correlation matrix between the corrected recognition result and the current local parameters of the local federated learning model based on the feature space mapping function.

[0151] In step 501, the feature space mapping function represents a non - linear transformation matrix from the feature space to the parameter space. The feature correlation matrix represents a 128×256 matrix of feature - parameter correlations.

[0152] In the embodiments of the present application, the Jacobian matrix of the corrected recognition result with respect to each layer of parameters is calculated through automatic differentiation, and combined with the outer product of the hidden layer activation state, to construct a mapping function from the feature space to the parameter space. The kernel method (kernel width 0.5) is used to calculate the correlation strength between features and parameters, generating a correlation matrix with matrix element values in the range [-1, 1].

[0153] Step 502: Determine the set of basis vectors to be optimized in the parameter space of the local federated learning model according to the singular value decomposition result of the feature correlation matrix.

[0154] In step 502, singular value decomposition refers to the spectral decomposition method of a matrix. The singular value decomposition result specifically refers to three sets of key data obtained after matrix decomposition of the feature correlation matrix: 1. The left singular vector matrix (128×20), whose column vectors represent the main change directions of the feature space, obtained by solving the matrix eigenvalue equation; 2. The singular value diagonal matrix (20×20), with diagonal elements arranged in descending order, and the numerical size reflecting the importance of each direction, obtained by the convergence of an iterative algorithm; 3. The right singular vector matrix (256×20), whose column vectors correspond to the optimization directions in the parameter space, derived from matrix product operations. This decomposition is implemented using a truncation algorithm, retaining components with singular values greater than a preset threshold (0.1). The set of basis vectors to be optimized specifically refers to the set of feature directions that are screened out through the singular value decomposition of the feature correlation matrix and have an impact on model parameter updates. The essential meaning of a basis vector refers to the unit vector representing the principal component direction of the feature correlation matrix in the parameter space. Each basis vector corresponds to a feature subspace component in the singular value decomposition, and the subspace spanned by them reflects the key correlation pattern between the corrected recognition result and the model parameters. The determination conditions for optimization: the amplitude of the singular value exceeds the preset energy threshold (retaining the main information components), the projection angle between the corresponding feature direction and the current gradient direction is less than the critical value (ensuring the effectiveness of optimization), and it satisfies the local geometric constraints on the parameter manifold (maintaining the stability of the model structure).

[0155] In the embodiments of the present application, the correlation matrix is truncated and decomposed (retaining components with singular values > 0.1), and the left singular vectors corresponding to the first 20 singular values are selected as the optimization bases. The dimension of the basis vectors is consistent with the model parameter space (256 - dimensional), and each vector represents a set of parameter update patterns related to features.

[0156] Step 503: Use the gradient direction of the corrected recognition result to perform an orthogonal projection on the set of basis vectors to generate a parameter update amount.

[0157] In step 503, the orthogonalization process of the set of basis vectors orthogonalizes the selected set of basis vectors {v1, v2, …, vk} to ensure that the basis vectors are orthogonal and normalized to each other, obtaining an orthonormal basis {u1, u2, …, uk}. Calculate the projection of the gradient direction on the orthogonal basis. Exemplarily, let the gradient direction of the corrected recognition result be g, and calculate the projection coefficient of it on each orthogonal basis ui as α i = g · u i . Only retain the basis vectors whose projection coefficient αi exceeds the set threshold (to ensure the effectiveness of optimization). Generate a parameter update amount, project the gradient direction g onto the selected orthogonal basis, and obtain the parameter update amount Δw as This update amount only adjusts the parameters along the direction of the key basis vectors, avoiding interference from irrelevant directions.

[0158] In the embodiment of the present application, calculate the gradient of the cross-entropy loss of the corrected recognition result with respect to the parameters (a 256-dimensional vector), and project it onto the subspace spanned by the basis vectors. Retain the basis vectors with a projection coefficient > 0.05 (usually 8 - 12), and generate a sparse parameter update amount (the proportion of non-zero elements is about 5%).

[0159] Step 504: Perform geodesic interpolation on the parameter update amount and the current local parameters of the local federated learning model to generate updated local parameters.

[0160] In step 504, geodesic interpolation is used to smoothly fuse the parameter update amount and the current local parameters along the shortest path (geodesic) on the Riemannian manifold of the parameter space, ensuring that the optimization process maintains the geometric constraints of the model structure.

[0161] In the embodiment of the present application, the system first converts the parameter update amount into a tangent vector in the tangent space, and then constructs an exponential mapping operator based on Lie group theory. The specific process is to calculate the projection of the current parameter point onto the tangent space; construct a geodesic equation along the direction of the tangent vector; numerically solve the integral path on the manifold using the fourth-order Runge - Kutta method; ensure that the updated parameters still satisfy the orthogonal constraint through re-orthogonalization. Finally, generate the updated parameters.

[0162] The following is a specific example;

[0163] In the continuation scenario of chemical plant pipeline leakage monitoring, the system first extracts the 128-dimensional activation features of the fourth convolutional layer to construct a feature space mapping function. The weight coefficient of the 2.7-μm feature channel is 0.15, which is determined by the performance of this wavelength band in correcting the recognition results. The 256×128 feature correlation matrix is calculated, and the matrix element values are calculated through a radial basis function kernel with a kernel width set to 0.3. After performing singular value decomposition on this matrix, the first 15 singular values are retained (the threshold is set to 0.12), and the corresponding right singular vectors form the parameter optimization basis set. The gradient vector [-0.13, 0.07,...] of the corrected result (ethylene recognition confidence of 96%) is projected onto the basis set, and the projection coefficients are calculated through orthogonalization, retaining 5 main directions with absolute values greater than 0.06. Finally, the parameters are updated along the geodesic on the orthogonal matrix manifold, with the step size set to 0.25, which is achieved through numerical integration by the fourth-order Runge-Kutta method, increasing the convolution kernel response value corresponding to the 2.7-μm feature from 0.42 to 0.58 while keeping the matrix orthogonality error less than 10^-6. This process improves the system's feature extraction ability for ethylene by 35% without introducing new false alarm samples.

[0164] To solve the problem of the decrease in the accuracy of substance recognition caused by noise interference in the spectral signal and further improve the reliability of hazardous substance detection, in some embodiments, this solution realizes the precise coupling of recognition features and model parameters, improving the knowledge transfer efficiency in the federated learning process. Through the parameter update mechanism of manifold optimization, the system can quickly adapt to the characteristic patterns of new hazardous substances while maintaining the model stability, providing an ability guarantee for continuous optimization of the distributed detection network.

[0165] Optionally, step 102: performing dynamic wavelength modulation and adaptive noise filtering on the original spectral signal, and combining with a pre-set hazardous substance feature library to generate a preliminary recognition result, including:

[0166] Step 601: constructing a dynamic wavelength modulation channel based on the wavelength tuning characteristics of the tunable laser spectroscopy device, and generating a modulation wavelength sequence matching the absorption spectral line of the target substance in the dynamic wavelength modulation channel.

[0167] In step 601, the wavelength tuning characteristics represent the tunable wavelength range (such as 2.5 - 10 μm) and resolution (0.01 nm) of the laser. The dynamic wavelength modulation channel refers to a signal processing path that adjusts the laser wavelength scanning strategy in real time according to the absorption characteristics of the target substance. The modulation wavelength sequence represents a list of scanning wavelengths designed according to the absorption characteristics of the substance.

[0168] In the embodiments of the present application, according to the absorption peak positions of the substances in the hazardous substance feature library, a modulation sequence including characteristic wavelengths and their adjacent bands is generated. Specifically, the characteristic absorption wavelengths of common hazardous substances in the target area are extracted; a scanning window of ±0.05 μm is set around each characteristic wavelength; and the scanning dwell time is allocated according to the absorption intensity.

[0169] Step 602: Perform segmented wavelength scanning on the original spectral signal through the modulation wavelength sequence to generate a modulated spectral signal.

[0170] In step 602, segmented wavelength scanning means dividing the full band into multiple characteristic intervals for scanning. The modulated spectral signal represents time-varying spectral data with wavelength modulation marks.

[0171] In the embodiments of the present application, the tunable laser is controlled to perform segmented scanning according to the modulation sequence, and the integration time of each wavelength point is dynamically adjusted according to the preset signal-to-noise ratio. The time-wavelength-intensity three-dimensional data is synchronously recorded to generate a modulated spectral signal with time marks.

[0172] Step 603: Analyze the noise distribution characteristics of the modulated spectral signal in the preset time-frequency joint domain to generate a noise suppression template related to the characteristics of the target substance.

[0173] In step 603, the time-frequency joint domain represents the time-frequency analysis space composed of short-time Fourier transform. The noise distribution characteristics refer to the spatial statistical characteristics of the interference components in the spectral signal in the time-frequency domain. The noise suppression template represents a binary mask matrix marking the noise distribution.

[0174] In the embodiments of the present application, a 128-point transform is performed on the modulation signal, and the statistical characteristics (mean, variance, etc.) of each time-frequency unit are calculated. The noise region is identified through adaptive threshold segmentation (threshold = 3σ) to generate a noise template complementary to the characteristic band of the substance.

[0175] Step 604: Perform selective filtering on the modulated spectral signal based on the noise suppression template to generate enhanced spectral data.

[0176] In step 604, selective filtering means a filtering method that retains the characteristic band and suppresses noise. The enhanced spectral data represents the spectral signal with improved signal-to-noise ratio.

[0177] In the embodiments of the present application, a filter bank based on the noise template is designed. A Chebyshev filter with 0.1 dB ripple is used for the characteristic band, and 40 dB stopband attenuation is applied to the noise region. The filtering is achieved through convolution operation, and the enhanced data with a signal-to-noise ratio improvement of more than 15 dB is output.

[0178] Step 605: Establish a multi-dimensional matching space in the pre-set hazardous substance feature library, map the enhanced spectral data into the multi-dimensional matching space, and calculate the similarity between the enhanced spectral data and the characteristic spectra of each hazardous substance in the hazardous substance feature library.

[0179] In step 605, the multi-dimensional matching space represents a 128-dimensional metric space composed of substance features. The similarity represents a quantitative index of the spectral shape matching degree.

[0180] In the embodiment of the present application, the enhanced data is projected into the feature space reduced by principal component analysis, and the similarity with the standard spectra of each substance is calculated. A band weight coefficient is introduced (characteristic peak band weight = 1.5, others = 0.8) to optimize the matching result.

[0181] Step 606: Use the category identifier of the hazardous substance with the maximum similarity as the category information of the target substance to obtain a preliminary recognition result including the category information of the target substance.

[0182] In step 606, the category identifier represents the unique classification code of the hazardous substance.

[0183] In the embodiment of the present application, the substance category with the highest similarity (threshold > 0.7) is selected, and the confidence level (range 0 - 100%) is calculated in combination with the matching metric and environmental parameters. A structured result including information such as the substance name, confidence level, and characteristic band is output.

[0184] The following is a specific example:

[0185] In the continuation scenario of chemical plant pipeline leakage monitoring, the system first constructs a modulation wavelength sequence that includes the ethylene characteristic wavelength band (2.7 μm ± 0.05 μm) and the benzene characteristic wavelength band (6.2 μm ± 0.08 μm). The number of scanning points in the 2.7-μm wavelength band is set to 50 points (obtained by converting the weight coefficient of this wavelength band, which is 0.15), and the adjacent wavelength interval is 0.001 μm (determined according to the response bandwidth of the model convolution kernel). When performing segmented scanning with a tunable laser, the integration time of the 2.7-μm wavelength band is set to 80 milliseconds (calculated from the signal-to-noise ratio requirement), generating a modulated spectral signal with time stamps. Through time-frequency analysis of the signal, it is found that there is a water vapor interference with an amplitude of 30% of the original signal at 4.8 μm (determined by wavelet transform energy detection), and based on this, a noise suppression template is generated (the attenuation coefficient at 4.8 μm is set to 0.3). The signal-to-noise ratio of the enhanced spectral data after filtering reaches 40 dB in the 2.7-μm wavelength band (the original signal is 25 dB). When matching with the hazardous substance feature library, the similarity score of ethylene is 0.91 (calculated using the improved dynamic time warping algorithm, and the wavelength band weights are updated synchronously with the model parameters), and the similarity score of benzene is 0.68. Finally, a preliminary recognition result of "ethylene, confidence level 95%" is output, and this confidence level is obtained by weighted averaging the similarity score (0.91) and the model output probability (0.96).

[0186] In the embodiments of the present application, this solution improves the recognition accuracy of hazardous substances in complex environments through dynamic modulation and adaptive filtering. The system can effectively suppress common environmental interferences and maintain the detection ability for weak characteristic signals, providing high-quality initial recognition results for subsequent federated learning optimization.

[0187] To solve the coordination problem of federated knowledge fusion and local device optimization and further improve the adaptive ability of hazardous substance recognition, in some embodiments, step 106: updating the preset local federated learning model based on the parameter weight distribution of the federated global model, and dynamically adjusting the scanning parameter configuration information of the tunable laser spectroscopy device, including:

[0188] Step 701: Based on the structural constraints of the local federated learning model, perform parameter space mapping on the feature subspace components in the parameter weight distribution to generate a weight update amount that is compatible with the structure of the local federated learning model.

[0189] In step 701, the structural constraint refers to the mathematical or physical limiting conditions that the parameter space of the local federated learning model must satisfy. Specifically, it includes the following aspects: Parameter topology constraint: The organization form of parameters (such as the local connectivity of convolution kernels, the temporal dependence of recurrent networks). Sparse / low-rank property of matrices (such as non-negative matrix factorization constraint in spectral analysis). Geometric constraint: Manifold structure (such as the manifold of orthogonal weight matrices, the manifold of covariance matrices). Boundary conditions (such as parameter value range constraints: non-negativity, unit norm, etc.). Federated knowledge consistency constraint: Isomorphism with the global model parameter space (ensuring that local updates can be aggregated). Compatibility with the dimension of the feature subspace (such as the matching of the number of principal components). The feature subspace components represent a subset of the parameters in the global model related to local features. Structural compatibility refers to mathematical compatibility, and the updated parameters still satisfy the structural constraints of the original model (such as remaining orthogonal after updating an orthogonal matrix), being consistent with the mathematical properties such as the parameter dimension and tensor rank of the local model. The weight update amount represents the parameter adjustment vector that conforms to the structure of the local model.

[0190] In the embodiments of this application, the manifold projection technology is used to map the global parameters to the parameter space of the local model. Specifically: Extract a 128-dimensional parameter subset related to ethylene recognition in the global model; Calculate the projection matrix through the Riemannian optimization algorithm; Generate a compatibility update amount with a dimension of 256, and the orthogonality error is controlled within 10^-6.

[0191] Step 702: Calculate the weight fusion coefficient based on the feature correlation of the enhanced spectral data and the weight update amount.

[0192] In step 702, the feature correlation refers to the parameter change pattern whose correlation subject is the features of the enhanced spectral data and the weight update amount. Association process: Feature coding alignment means mapping the enhanced spectral features to the hidden layer activation space of the local federated learning model, and performing the same hidden layer space projection on the weight update amount. Cross-domain correlation analysis refers to calculating the canonical correlation analysis of the feature vector and the parameter update direction in the hidden layer space. Or quantifying the statistical dependence between the two through mutual information estimation. Dynamic fusion coefficient generation means allocating weights according to the correlation strength (such as a higher-correlation update amount obtaining a larger fusion coefficient). Fine-grained fusion is achieved through a gating mechanism (such as attention weights). The weight fusion coefficient represents the adaptive fusion weight of the update amount.

[0193] In the embodiments of this application, calculate the activation intensity of the enhanced data in each layer of the model (such as the activation value of the channel corresponding to the 2.7μm band is 0.58), and combine it with parameter sensitivity analysis (implemented through automatic differentiation) to generate a fusion coefficient in the range [0,1].

[0194] Step 703: Perform adaptive weighted fusion on the weight update amount according to the weight fusion coefficient to generate an updated local federated learning model.

[0195] In step 703, adaptive weighted fusion represents a parameter fusion method allocated according to importance.

[0196] In the embodiments of the present application, a layer-by-layer weighting strategy is adopted. The convolutional layer is weighted according to the activation intensity of channels (for example, the weight of the 2.7μm channel is 0.8); the fully connected layer is weighted according to the importance of features. Smooth update is achieved through parameter space interpolation. The method of adaptive weighted fusion is as follows: Calculate the activation intensity of the enhanced spectral data in each layer of the model as the feature correlation index: where α i is the fusion weight coefficient of the parameters in the i-th layer, satisfying ∑α i = 1. The activation value is the feature map output by the hidden layer of the model or the neuron activation intensity (such as the output after ReLU). N is the total number of layers of the model (for example, for a 128-dimensional feature space, N = 128). Calculate the gradient sensitivity of the parameter update amount to the loss through automatic differentiation to generate the sensitivity weight matrix. Allocate the fusion weights according to the model hierarchy. For example, the weight fusion coefficient of the convolutional layer is determined by the feature activation intensity, and the fully connected layer is determined by the gradient sensitivity: where ΔW 融合 is the fused parameter update amount (such as a 256-dimensional vector). ΔW k is the original parameter update amount of the k-th layer. β k is the fusion coefficient of the k-th layer, determined by the feature importance α i and the gradient sensitivity together. K is the number of layers participating in the fusion.

[0197] Step 704: Establish an association mapping function between the scanning parameters of the tunable laser spectroscopy device and the feature response of the updated local federated learning model.

[0198] In the embodiments of the present application, a polynomial regression model of the scanning parameters (wavelength, step size, etc.) and the model feature response is established. It is trained with 300 groups of historical data: the wavelength accuracy influence coefficient is 0.7, the integration time influence coefficient is 0.5, and the adjusted R-square reaches 0.92.

[0199] Step 705: Jointly optimize the wavelength tuning range, scanning step size, and integration time of the tunable laser spectroscopy device through the association mapping function, and dynamically adjust the scanning parameter configuration information according to the optimization results.

[0200] In the embodiments of the present application, the multi-objective genetic algorithm is used for optimization. The wavelength range is set to 2.6 - 2.8 μm (ethylene characteristic region); the step size is adjusted to 0.001 μm; the integration time is optimized to 75 ms. The signal-to-noise ratio of the 2.7 μm band is increased by 25%. Exemplarily, the baseline signal-to-noise ratio before optimization is 40 dB (measured from the enhanced spectral data). After obtaining the optimal parameter combination (wavelength range 2.6 - 2.8 μm, step size 0.001 μm, integration time 75 ms) through iterative search of the genetic algorithm, the measured signal-to-noise ratio is increased to 50 dB. This improvement is due to the synergistic effect of three aspects: 1. The wavelength range precisely matches the ethylene absorption peak (2.7 μm ± 0.1 μm), increasing the characteristic signal intensity by 15%; 2. A finer step size setting (0.001 μm) reduces the loss of spectral information, contributing a 6% increase; 3. Extending the integration time to 75 ms (original 50 ms) reduces random noise, bringing a 4% improvement. These improvements ultimately achieve a relative increase in the signal-to-noise ratio of 25% ((50 - 40) / 40 × 100%) through the synergistic effect between parameters, and the stability error is verified to be within ±2% after 100 experiments. The method of joint optimization is as follows: Establish a polynomial regression model between the scanning parameters and the model characteristic response: Characteristic response = a·λ + b·Δλ + c·t 积分 + d; where λ is the wavelength tuning range, in μm. Δλ is the scanning step size, which determines the spectral resolution. t 积分 is the integration time, which affects the signal-to-noise ratio. a, b, and c are regression coefficients. d is the bias term, which compensates for the baseline error. The genetic algorithm is used to optimize the parameter combination, and the objective function includes: maximizing the signal-to-noise ratio (SNR), minimizing the scanning time, and maximizing the feature matching degree. Configure the device in real time according to the optimization results: The wavelength range is focused on the range of ±0.1 μm around the characteristic peak (e.g., 2.7 μm → 2.6 - 2.8 μm). The scanning step size is dynamically adjusted according to the model resolution (e.g., a 0.001 μm step size corresponds to a 128-point scan). The integration time is inversely deduced according to the target SNR, and the formula is:

[0201] The following is a specific example:

[0202] In the continuous scenario of chemical plant pipeline leakage monitoring, the system first extracts a 128-dimensional parameter subset related to ethylene detection from the federal global model, maps it to the 256-dimensional parameter space of the local model through the Riemann projection algorithm, and controls the projection error within 10^-6. Calculate the weight fusion coefficient according to the characteristic intensity (activation value 0.58) of the 2.7μm band in the enhanced spectral data, and set the fusion coefficient of the third layer of the convolution kernel to 0.85 (calculated from the linear combination of the sensitivity of this layer to ethylene characteristics 0.7 and the band signal-to-noise ratio 40dB). Adopt a hierarchical weighted strategy to update the local model parameters, so that the convolution kernel weight corresponding to the 2.7μm feature is increased from 0.61 to 0.73. Subsequently, establish a scanning parameter optimization model, where the weight coefficient of the wavelength range is 0.7 (obtained by regression analysis of historical data), and optimize the scanning parameters according to the model output: the wavelength range is 2.65 - 2.75μm (centered on the 2.7μm ethylene characteristic peak, and the boundary is determined by the half-width of the characteristic peak 0.05μm), the scanning step is 0.0009μm (calculated according to the resolution ability of the updated model), and the integration time is 85 milliseconds (inferred from the target signal-to-noise ratio 45dB). After implementation, the measured signal-to-noise ratio of the 2.7μm band is further increased to 46dB, the confidence level of ethylene recognition reaches 97%, and the influence of water vapor interference at 4.8μm is reduced by 15% again.

[0203] In the embodiment of the present application, this solution realizes the accurate migration of global knowledge to the local model. Through the collaborative optimization of scanning parameters and model capabilities, the system maintains stable detection performance in complex environments, improving the reliability and adaptability of hazardous substance identification.

[0204] Figure 2 The following is a schematic structural diagram of an unattended system based on Internet of Things technology provided by the embodiment of the present application, as Figure 2 shown, this system includes:

[0205] A scanning module 21, configured to perform multi-band scanning on the target substance by using a tunable laser spectroscopy device in the distributed laser detection node to generate an original spectral signal.

[0206] A processing module 22, configured to perform dynamic wavelength modulation and adaptive noise filtering processing on the original spectral signal, and generate a preliminary recognition result in combination with a preset hazardous substance feature library.

[0207] An evaluation module 23, configured to perform confidence evaluation on the preliminary recognition result based on a preset local federated learning model to obtain a corrected recognition result.

[0208] An analysis module 24, configured to perform correlation analysis on the corrected recognition result and the current local parameters of the local federated learning model to generate updated local parameters.

[0209] A receiving module 25, configured to upload the updated local parameters to a central server and receive the parameter weight distribution of a federated global model.

[0210] An adjustment module 26, configured to update a preset local federated learning model based on the parameter weight distribution of the federated global model, and dynamically adjust the scanning parameter configuration information of the tunable laser spectroscopy device, so as to realize the identification of hazardous substances in the target substance through the updated local federated learning model. The federated global model is obtained by the central server aggregating the updated local parameters uploaded by multiple distributed laser detection nodes.

[0211] Figure 2 The unattended system based on the Internet of Things technology can execute Figure 1 The unattended method based on the Internet of Things technology described in the illustrated embodiment, and its implementation principle and technical effects will not be elaborated. For the unattended system based on the Internet of Things technology in the above embodiment, the specific manners in which each module and unit perform operations have been described in detail in the embodiment related to the method, and will not be elaborated here.

[0212] In a possible design, Figure 2 The unattended system based on the Internet of Things technology described in the illustrated embodiment can be implemented as a computing device, such as Figure 3 shown, the computing device may include a storage component 31 and a processing component 32;

[0213] The storage component 31 stores one or more computer instructions, wherein the one or more computer instructions are called and executed by the processing component 32.

[0214] The processing component 32 is the above Figure 1 The unattended method based on the Internet of Things technology described in the embodiment.

[0215] Among them, the processing component 32 may include one or more processors to execute computer instructions to complete all or part of the steps in the above methods. Of course, the processing component may also be implemented by one or more application-specific integrated circuits (ASICs), digital signal processors (DSPs), digital signal processing devices (DSPDs), programmable logic devices (PLDs), field programmable gate arrays (FPGAs), controllers, microcontrollers, microprocessors or other electronic components for executing the above methods.

[0216] The storage component 31 is configured to store various types of data to support the operation of the terminal. The storage component can be implemented by any type of volatile or non-volatile storage device or a combination thereof, such as static random access memory (SRAM), electrically erasable programmable read-only memory (EEPROM), erasable programmable read-only memory (EPROM), programmable read-only memory (PROM), read-only memory (ROM), magnetic memory, flash memory, magnetic disk or optical disk.

[0217] Of course, the computing device may also necessarily include other components, such as input / output interfaces, display components, communication components, etc.

[0218] The input / output interface provides an interface between the processing component and the peripheral interface module, and the above peripheral interface module may be an output device, an input device, etc.

[0219] The communication component is configured to facilitate communication between the computing device and other devices in a wired or wireless manner, etc.

[0220] Among them, the computing device may be a physical device or an elastic computing host provided by a cloud computing platform, etc. At this time, the computing device may refer to a cloud server, and the above processing component, storage component, etc. may be basic server resources leased or purchased from a cloud computing platform.

[0221] An embodiment of the present application further provides a computer storage medium storing a computer program, and when the computer program is executed by a computer, it can implement the above-mentioned Figure 1 unattended method based on Internet of Things technology shown in the embodiment.

[0222] Those skilled in the art can clearly understand that for the convenience and brevity of description, the specific working processes of the above-described systems, devices, and units can refer to the corresponding processes in the foregoing method embodiments, and will not be elaborated herein.

[0223] The device embodiments described above are merely illustrative. The units described as separate components may or may not be physically separated, and the components shown as units may or may not be physical units, that is, they may be located in one place, or may be distributed to multiple network units. Some or all of the modules can be selected according to actual needs to achieve the purpose of the solution of this embodiment. Those of ordinary skill in the art can understand and implement it without creative efforts.

[0224] Through the description of the above embodiments, those skilled in the art can clearly understand that each embodiment can be implemented by means of software plus a necessary general hardware platform, and of course, it can also be implemented by hardware. Based on this understanding, the essence of the above technical solution, or the part that contributes to the prior art, can be embodied in the form of a software product, and the computer software product can be stored in a computer-readable storage medium, such as ROM / RAM, magnetic disk, optical disk, etc., including several instructions for causing a computer device (which can be a personal computer, a server, or a network device, etc.) to execute the methods described in each embodiment or some parts of the embodiments.

[0225] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present application, and are not intended to limit them; although the present application has been described in detail with reference to the foregoing embodiments, those of ordinary skill in the art should understand that they can still modify the technical solutions recorded in the foregoing embodiments, or perform equivalent replacements for some of the technical features; and these modifications or replacements do not make the essence of the corresponding technical solutions deviate from the spirit and scope of the technical solutions of each embodiment of the present application.

Claims

1. An unattended method based on Internet of Things technology, characterized in that, Applied to distributed laser detection nodes, including: Using the tunable laser spectroscopy device in the distributed laser detection node to perform multi-band scanning on the target substance to generate an original spectral signal; Performing dynamic wavelength modulation and adaptive noise filtering on the original spectral signal, and combining with a preset hazardous substance feature library to generate a preliminary identification result; Performing confidence evaluation on the preliminary identification result based on a preset local federated learning model to obtain a corrected identification result; Performing correlation analysis on the corrected identification result and the current local parameters of the local federated learning model to generate updated local parameters; Uploading the updated local parameters to a central server and receiving the parameter weight distribution of the federated global model; Updating the preset local federated learning model based on the parameter weight distribution of the federated global model, dynamically adjusting the scanning parameter configuration information of the tunable laser spectroscopy device, so as to realize the identification of hazardous substances in the target substance through the updated local federated learning model, and the federated global model is obtained by the central server aggregating the updated local parameters uploaded by multiple distributed laser detection nodes.

2. The method according to claim 1, characterized in that, The performing confidence evaluation on the preliminary identification result based on a preset local federated learning model to obtain a corrected identification result includes: Calculating the confidence score of the preliminary identification result in the feature space through the probability distribution model in the local federated learning model; When the confidence score is lower than a preset confidence threshold, activating the error backpropagation path in the local federated learning model; Performing error compensation calculation on the multi-spectral features of the preliminary identification result through the error backpropagation path to generate a compensated feature vector; Performing phase synchronization superposition on the compensated feature vector and the preliminary identification result in a preset time-frequency joint domain to generate a corrected identification result.

3. The method according to claim 2, wherein The performing phase synchronization superposition on the compensated feature vector and the preliminary identification result in a preset time-frequency joint domain to generate a corrected identification result includes: Constructing a multi-spectral phase distribution map of the preliminary identification result through the hidden layer activation state of the local federated learning model; Based on the multi-spectral phase distribution map, performing time-frequency rasterization recombination on the compensated feature vector to generate a recombined feature vector with phase continuity with the preliminary identification result; Establishing a dynamic carrier modulation channel within the phase synchronization framework of the preset time-frequency joint domain, and in the dynamic carrier modulation channel, performing convolution processing on the recombined feature vector and the preliminary identification result; Performing phase correction on the convolution processing result according to the phase synchronization parameters of the preset time-frequency joint domain to generate a corrected identification result.

4. The method according to claim 3, wherein The performing convolution processing on the recombined feature vector and the preliminary identification result includes: Generating a multi-band carrier group matching the multi-spectral features of the preliminary identification result in the dynamic carrier modulation channel; Performing band-pass constraint modulation on the multi-band carrier group through the multi-spectral phase distribution map to generate a modulated carrier base; Perform baseband loading of the recombinant feature vector and the modulated carrier basis in the preset time-frequency joint domain to generate a modulated feature signal; Based on the multi-spectral phase distribution map, perform time-frequency window segmentation on the preliminary recognition result to generate a reference signal after multi-scale sub-band decomposition; Establish a feedback loop in the dynamic carrier modulation channel. In the feedback loop, perform multi-scale sub-band convolution on the modulated feature signal and the reference signal.

5. The method according to claim 1, characterized in that, The associative analysis of the corrected recognition result and the current local parameters of the local federated learning model to generate updated local parameters includes: Construct a feature space mapping function based on the hidden layer activation state of the local federated learning model. Based on the feature space mapping function, calculate the feature correlation matrix between the corrected recognition result and the current local parameters of the local federated learning model; According to the singular value decomposition result of the feature correlation matrix, determine the set of basis vectors to be optimized in the parameter space of the local federated learning model; Use the gradient direction of the corrected recognition result to perform orthogonal projection on the set of basis vectors to generate a parameter update amount; Perform geodesic interpolation on the parameter update amount and the current local parameters of the local federated learning model to generate updated local parameters.

6. The method according to claim 1, wherein The dynamic wavelength modulation and adaptive noise filtering processing of the original spectral signal, and combining with a preset hazardous substance feature library to generate a preliminary recognition result includes: Construct a dynamic wavelength modulation channel based on the wavelength tuning characteristics of the tunable laser spectroscopy device, and generate a modulation wavelength sequence matching the absorption spectral line of the target substance in the dynamic wavelength modulation channel; Perform segmented wavelength scanning on the original spectral signal through the modulation wavelength sequence to generate a modulated spectral signal; Analyze the noise distribution characteristics of the modulated spectral signal in the preset time-frequency joint domain to generate a noise suppression template related to the characteristics of the target substance; Perform selective filtering on the modulated spectral signal based on the noise suppression template to generate enhanced spectral data; Establish a multi-dimensional matching space in the preset hazardous substance feature library, map the enhanced spectral data into the multi-dimensional matching space, and calculate the similarity between the enhanced spectral data and the characteristic spectra of each hazardous substance in the hazardous substance feature library; Use the class label of the hazardous substance with the maximum similarity as the class information of the target substance to obtain a preliminary recognition result including the class information of the target substance.

7. The method according to claim 1, characterized in that Updating the preset local federated learning model based on the parameter weight distribution of the federated global model, and dynamically adjusting the scanning parameter configuration information of the tunable laser spectroscopy device includes: Based on the structural constraints of the local federated learning model, perform parameter space mapping on the feature subspace component in the parameter weight distribution to generate a weight update amount compatible with the structure of the local federated learning model; Calculate a weight fusion coefficient based on the feature correlation of the enhanced spectral data and the weight update amount; Adopt the weight fusion coefficient to adaptively weightedly fuse the weight update amount, and generate an updated local federated learning model; Establish an association mapping function between the scanning parameters of the tunable laser spectroscopy device and the feature response of the updated local federated learning model; Jointly optimize the wavelength tuning range, scanning step size, and integration time of the tunable laser spectroscopy device through the association mapping function, and dynamically adjust the scanning parameter configuration information according to the optimization results.

8. An unattended system based on Internet of Things technology, characterized in that, Comprising: A scanning module, configured to perform multi-band scanning on a target substance by using the tunable laser spectroscopy device in the distributed laser detection node, and generate an original spectral signal; A processing module, configured to perform dynamic wavelength modulation and adaptive noise filtering on the original spectral signal, and generate a preliminary identification result in combination with a preset hazardous substance feature library; An evaluation module, configured to evaluate the confidence level of the preliminary identification result based on a preset local federated learning model to obtain a corrected identification result; An analysis module, configured to perform correlation analysis on the corrected identification result and the current local parameters of the local federated learning model to generate updated local parameters; A receiving module, configured to upload the updated local parameters to a central server and receive the parameter weight distribution of the federated global model; An adjustment module, configured to update a preset local federated learning model based on the parameter weight distribution of the federated global model, and dynamically adjust the scanning parameter configuration information of the tunable laser spectroscopy device, so as to identify hazardous substances in the target substance through the updated local federated learning model, where the federated global model is obtained by the central server aggregating the updated local parameters uploaded by multiple distributed laser detection nodes.

9. A computing device, characterized in that, Comprising a processing component and a storage component; the storage component stores one or more computer instructions; the one or more computer instructions are used to be called and executed by the processing component to implement an unattended method based on Internet of Things technology according to any one of claims 1 to 7.

10. A computer storage medium, characterized in that, Stores a computer program, which when executed by a computer, implements an unattended method based on Internet of Things technology according to any one of claims 1 to 7.

Images