Multi-sensor early warning method, system and equipment based on industrial internet of things and medium

Abstract

The invention relates to the field of Internet of Things production, in particular to a multi-sensor early warning method, system and device based on industrial Internet of Things and a medium, and the method comprises the following steps: obtaining sensing data; performing feature extraction and feature fusion on the processed sensing data to obtain early warning feature vector data; obtaining early warning feature vector data to execute a decision function; training in combination with the optimal penalty factor to obtain a Lagrange multiplier and bias; mapping the category label to a security level corresponding to the industrial Internet of Things; the method comprises the following steps: analyzing original data acquired by a data acquisition network, eliminating abnormal values exceeding a historical fluctuation range, and eliminating dimensional differences by adopting standardization processing; in the training stage, an optimal kernel weight and a penalty factor are determined through an optimization algorithm, in the testing stage, a real-time feature vector is input into the model, the kernel function value weighted sum of the real-time feature vector and a support vector is calculated, and finally a level label reflecting the safety state of equipment is output.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of Internet of Things production, and in particular to a multi-sensor early warning method, system, device and medium based on industrial Internet of Things. BACKGROUND

[0002] With the popularization of industrial Internet of Things (IIoT) technology, multi-sensor-based early warning systems have become a core means to ensure industrial safety production. Traditional methods mainly rely on setting fixed thresholds for single sensors for alarm, or using simple mathematical statistical models for anomaly judgment. However, modern industrial systems are highly complex, and their operating states are the result of the combined action of multiple physical fields such as mechanics, electricity, and thermodynamics. It is difficult to achieve accurate and early fault warning based on data from a single or a few dimensions.

[0003] At the data processing level, industrial field sensors vary in model, protocol, and sampling rate, resulting in different scales and varying quality of raw data. Existing solutions lack an automated and standardized data analysis and cleaning mechanism, and are insufficient in handling noise and outliers, making it difficult to convert heterogeneous data into high-quality, same-scale analysis basis, which seriously affects the reliability of subsequent model input.

[0004] Secondly, at the feature fusion level, existing methods mostly stop at simple data splicing or shallow feature weighted averaging, failing to deeply mine the complex spatio-temporal correlation and nonlinear coupling relationship between different modal data. For example, the deep correlation between the frequency spectrum features of vibration signals and the temperature time series trend is often ignored. Due to the lack of effective multi-core learning and fusion mechanism based on class information, the discriminative power of the extracted fusion features is insufficient, resulting in limited sensitivity and accuracy of the early warning model in perceiving complex fault patterns. SUMMARY

[0005] The main purpose of the present application is to provide a multi-sensor early warning method based on industrial Internet of Things, aiming to solve the problem that the existing early warning network is difficult to associate the nonlinear coupling relationship between multi-sensor data.

[0006] To achieve the above-mentioned purpose, the present application provides a multi-sensor early warning method based on industrial Internet of Things, comprising the following steps: obtaining sensor data; analyzing and processing the obtained sensor data into sensor data of the same scale; performing feature extraction and feature fusion on the processed sensor data to obtain early warning feature vector data; The early warning feature vector data is input into a decision function to calculate kernel function values between the early warning feature vector data and all support vectors by using an optimal kernel function; and a Lagrange multiplier and a bias are trained by combining an optimal penalty factor to obtain a class label by weighted summation of the Lagrange multiplier and the bias. The class label is mapped to a corresponding security level of the industrial Internet of Things.

[0007] Optionally, the parsing of the obtained sensing data into sensing data in the same scale comprises the following steps: The raw data stream is obtained by the parser and is uniformly packaged into standard internal data; The average value of each sensor unit in the historical sensing data is calculated, and a stable range is set according to the average value, and the internal data that is not in the stable range is removed or smoothed; The original data value of the internal data point is subtracted from the average value of the data point in the historical sensing data, and then the difference value is divided by the standard deviation of the data point, so as to convert the internal data in the stable range into a distribution with a mean value of 0 and a standard deviation of 1.

[0008] Optionally, the feature extraction and feature fusion of the processed sensing data comprise the following steps: For the sensing data of each sensor unit, a fixed-length sliding time window is set, and at each time point, a plurality of data points in the current window are obtained, and statistical features are calculated; wherein the statistical features include the current value, the window mean value and the window standard deviation; A multi-kernel matrix set is constructed; wherein the multi-kernel matrix set includes a linear kernel, a polynomial kernel and a radial basis function kernel; A weighted combination kernel matrix is constructed; The spatial transformation fusion based on the category information is performed.

[0009] Optionally, the spatial transformation fusion based on the category information comprises the following steps: The constructed weighted combination kernel matrix is fused with the category structure information to obtain a category modulation matrix; The category modulation matrix is feature-decomposed; The final transformation matrix is calculated by using the result of the feature decomposition; A multi-kernel relationship vector is constructed; The multi-kernel relationship vector is left-multiplied by the final transformation matrix to obtain early warning feature vector data; The category structure information is obtained from a Laplacian matrix.

[0010] Optionally, the early warning method further comprises the following steps: Historical sensor data is acquired, and the data early warning model is optimized according to the set time period. Within the parameter range of the kernel function and penalty factor, an initial population containing several candidate solutions of kernel function and penalty factor is generated through tent chaos mapping and reverse learning strategy. For each candidate solution in the initial population, configure a temporary data warning model to train the temporary data warning model using historical sensor data; The artificial hummingbird algorithm incorporates Levy's flight guidance for foraging, allowing for migration to escape local optima. After a set number of iterations, a new kernel function and penalty factor are obtained and updated in the data early warning model to complete the iterative update of the data early warning model.

[0011] To achieve the above objectives, the present invention also provides a multi-sensor early warning system based on the Industrial Internet of Things (IIoT). The early warning system includes a management platform, a sensor network platform, and an object platform that are sequentially connected in communication. The management platform includes components that are sequentially connected in communication: A data acquisition network, comprising several sensor units, wherein the several sensor units respectively acquire different sensor data of the industrial Internet of Things; The data processing module is used to parse and process the sensing data acquired by different sensor units into sensing data at the same scale. The data fusion module is used to extract and fuse features from the processed sensor data to obtain early warning feature vector data. The data early warning module is equipped with a vector machine-based data early warning model. After acquiring early warning feature vector data, the data early warning model executes a decision function to calculate the kernel function value between the input early warning feature vector data and all support vectors using the optimal kernel function. Then, it combines the optimal penalty factor to train and obtain Lagrange multipliers and biases, and performs a weighted summation of the Lagrange multipliers and biases to obtain the category label of the data acquisition network. The data alarm module is used to map category labels to the corresponding security levels of the Industrial Internet of Things.

[0012] Optionally, the early warning system further includes an optimization module, which is used to acquire historical sensor data and optimize the data early warning model according to a set time period, so as to generate an initial population containing several candidate solutions of kernel function and penalty factor through tent chaos mapping and reverse learning strategy within the parameter range of kernel function and penalty factor.

[0013] Optionally, the optimization module includes: An evaluation unit is configured to configure a temporary data warning model for each candidate solution in the initial population, so as to train the temporary data warning model using historical sensor data. An iterative unit is used to introduce Levy's flight guidance for foraging into the artificial hummingbird algorithm, so as to escape the local optimum through migration foraging. After reaching a set number of iterations, the iterative unit obtains a new kernel function and penalty factor and updates it to the data early warning model to complete the iterative update of the data early warning model.

[0014] To achieve the above objectives, the present invention also provides a computer device including a memory and a processor, wherein the memory stores a computer program and the processor executes the computer program.

[0015] To achieve the above objectives, the present invention also provides a computer-readable storage medium storing a computer program, wherein a processor executes the computer program.

[0016] The beneficial effects that this invention can achieve are as follows: The multi-sensor early warning method based on the Industrial Internet of Things (IIoT) proposed in this invention includes the following steps: acquiring sensor data; parsing and processing the acquired sensor data into sensor data at the same scale; performing feature extraction and feature fusion on the processed sensor data to obtain early warning feature vector data; executing a decision function on the early warning feature vector data to calculate the kernel function value between the input early warning feature vector data and all support vectors using the optimal kernel function; training with the optimal penalty factor to obtain Lagrange multipliers and biases, and performing a weighted summation of the Lagrange multipliers and biases to obtain the category label; mapping the category label to the security level corresponding to the Industrial Internet of Things. By analyzing the raw data acquired by the data acquisition network and removing outliers that exceed the historical fluctuation range, standardization is used to eliminate dimensional differences. For sensor data, time-series features such as window mean and standard deviation are extracted using a sliding window. Subsequently, a multi-kernel matrix containing linear kernels and radial basis function kernels is constructed. Combined with the category structure information formed by historical sensor data, a discriminative fusion feature vector is generated. During the training phase, the optimal kernel weights and penalty factors are determined through optimization algorithms. During the testing phase, the real-time feature vectors are input into the model, and the weighted sum of the kernel function values ​​of the feature vectors and support vectors is calculated. Finally, a level label reflecting the safety status of the equipment is output. Attached Figure Description

[0017] To more clearly illustrate the specific embodiments of the present invention or the technical solutions in the prior art, the accompanying drawings used in the description of the specific embodiments or the prior art will be briefly introduced below. In all the drawings, similar elements or parts are generally identified by similar reference numerals. In the drawings, the elements or parts are not necessarily drawn to scale.

[0018] Figure 1 This is a flowchart of the early warning method in Embodiment 1 of the present invention; Figure 2 This is a structural block diagram of the early warning system in Embodiment 3 of the present invention.

[0019] The realization of the objective, functional features and advantages of the present invention will be further explained in conjunction with the embodiments and with reference to the accompanying drawings. Detailed Implementation

[0020] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of the present invention, and not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative effort are within the scope of protection of the present invention.

[0021] If the embodiments of this invention involve descriptions such as "first" or "second," these descriptions are for descriptive purposes only and should not be construed as indicating or implying their relative importance or implicitly specifying the number of technical features indicated. Therefore, a feature defined with "first" or "second" may explicitly or implicitly include at least one of those features. Furthermore, the meaning of "and / or" throughout the text includes three parallel solutions; for example, "A and / or B" includes solution A, solution B, or a solution where both A and B are satisfied. Additionally, the technical solutions of the various embodiments can be combined with each other, but this must be based on the ability of those skilled in the art to implement them. When the combination of technical solutions is contradictory or impossible to implement, it should be considered that such a combination of technical solutions does not exist and is not within the scope of protection claimed by this invention.

[0022] Example 1: As attached Figure 1 As shown, this embodiment provides a multi-sensor early warning method based on the Industrial Internet of Things, including the following steps: Acquire sensor data; The acquired sensor data is parsed and processed into sensor data at the same scale; Feature extraction and feature fusion are performed on the processed sensor data to obtain early warning feature vector data; The decision function is executed to obtain the early warning feature vector data, and the kernel function value between the input early warning feature vector data and all support vectors is calculated using the optimal kernel function. Then, the optimal penalty factor is combined to train and obtain the Lagrange multipliers and biases. The class label is obtained by weighted summation of the Lagrange multipliers and biases. Map category labels to the corresponding security levels for the Industrial Internet of Things (IIoT).

[0023] In existing technologies, the data formats of various types of sensors deployed in industrial IoT environments, such as vibration, temperature, and pressure sensors, often become chaotic due to protocol differences. Traditional early warning systems use fixed threshold methods to independently judge data from a single sensor, which cannot effectively capture the nonlinear correlation between multiple data sources. For example, in a mechanical factory's production equipment monitoring scenario, the sampling frequencies of the temperature and pressure sensors are 10Hz and 50Hz, respectively. The raw data suffers from inconsistent dimensions and severe noise interference, causing the system to mistakenly identify instantaneous fluctuations under normal operating conditions as abnormal states.

[0024] Based on the above problems, this embodiment proposes a multi-sensor early warning method based on the Industrial Internet of Things (IIoT), including the following steps: acquiring sensor data; parsing and processing the acquired sensor data into sensor data at the same scale; performing feature extraction and feature fusion on the processed sensor data to obtain early warning feature vector data; executing a decision function on the early warning feature vector data to calculate the kernel function value between the input early warning feature vector data and all support vectors using the optimal kernel function; then training with the optimal penalty factor to obtain Lagrange multipliers and biases, and performing a weighted summation of the Lagrange multipliers and biases to obtain the category label; and mapping the category label to the security level corresponding to the Industrial Internet of Things.

[0025] Specifically, the raw data acquired by the data acquisition network is analyzed, outliers exceeding the historical fluctuation range are removed, and standardization is used to eliminate dimensional differences. For the sensor data, time-series features such as window mean and standard deviation are extracted using a sliding window. Subsequently, a multi-kernel matrix containing linear kernels and radial basis function kernels is constructed, and combined with the category structure information formed by historical sensor data, a discriminative fusion feature vector is generated. During the training phase, the optimal kernel weights and penalty factors are determined through optimization algorithms. During the testing phase, the real-time feature vectors are input into the model, and the weighted sum of the kernel function values ​​of the feature vectors and support vectors is calculated. Finally, a level label reflecting the safety status of the equipment is output.

[0026] Compared with existing technologies, traditional methods using a single kernel function to process multi-source data struggle to balance the contributions of different feature dimensions, leading to a high misjudgment rate for complex faults. This embodiment, however, through joint optimization of multi-kernel matrix combination and category modulation matrix, enables the model to adaptively adjust the fusion weights of vibration and temperature features. In an early warning case of bearing wear, it successfully identified a coupled fault mode characterized by slight vibration signal anomalies accompanied by a gradual temperature rise.

[0027] Through the above technical solution, this embodiment effectively solves the feature space mismatch problem caused by the heterogeneity of multi-sensor data. By standardizing the processing and multi-core fusion mechanism, the feature discrimination power is improved, enabling the early warning model to accurately capture the multi-physics coupling effect hidden in the equipment operation status. In production monitoring scenarios under multiple scenarios, this method reduces the false alarm rate from 12% of the traditional method to less than 5%, while shortening the detection response time of composite faults by 40%.

[0028] In some embodiments, the sensing data includes temperature sensing data, pressure sensing data, and vibration sensing data collected by temperature sensors, pressure sensors, and vibration sensors, respectively.

[0029] In some embodiments, the early warning method in this embodiment is implemented based on a vector machine data early warning model. The core lies in how the data early warning module utilizes the vector machine (SVM)-based data early warning model to output category labels through a decision function, which satisfies expression one: ; in, The output of the decision function; This is a sign function; it outputs +1 when the input is greater than 0 and -1 when the input is less than 0.

[0030] x j This is the warning feature vector data of the current input.

[0031] x i For the first i There are 10 support vectors (from historical training data).

[0032] y i For the first i Each support vector corresponds to a category label (e.g., +1 or -1).

[0033] N This represents the total number of support vectors.

[0034] K ( x j, x i ) is a kernel function that calculates the similarity between the input vector and the support vector in a high-dimensional space.

[0035] a i For the first i The Lagrange multipliers corresponding to each support vector are obtained by training in conjunction with the optimal penalty factor, reflecting the importance of that support vector.

[0036] b The bias term for the classification hyperplane.

[0037] In summary, the above expression is the core decision function of this SVM model, used to classify the input warning feature vector data. Specifically, it calculates the similarity between the input vector and all support vectors through an optimal kernel function, then multiplies these similarities by their respective Lagrange multipliers and categories and sums them, finally adding a bias; the sign function determines the final category based on the sign of the summation result, and this category label will be mapped to the corresponding security level.

[0038] In this embodiment, the process of parsing and processing the acquired sensor data into sensor data at the same scale specifically includes the following steps: The original data stream is obtained through a parser and uniformly encapsulated into standard internal data; Calculate the average value of each sensor unit in the historical sensing data, and set a stable range based on the average value. Then, remove or smooth the internal data that is not within the stable range. The internal data is transformed into a distribution with a mean of 0 and a standard deviation of 1 by subtracting the original data value of the data point from the average value of the data point in the historical sensor data, and then dividing the difference by the standard deviation of the data point.

[0039] Understandably, the raw data stream from sensors is converted into a structured internal data format with a unified timestamp and data identifier after protocol parsing, resolving compatibility issues between data interfaces of different industrial equipment. During the data cleaning phase, the system automatically calculates the operating baseline values ​​for each sensor over a historical period. When real-time data deviates from the baseline value by more than three standard deviations, an abnormal data removal mechanism is triggered, or an exponentially weighted moving average method is used for smoothing and repair. The cleaned data then enters the scale normalization stage. Each data point is normalized by subtracting its historical mean and dividing by the standard deviation, eliminating dimensional differences between sensors for different physical quantities such as pressure, temperature, and vibration, forming a feature vector conforming to a Gaussian distribution.

[0040] It is also understandable that the above technical solutions have enabled automated protocol parsing and format unification of multi-source heterogeneous data in industrial settings, solving the data silo problem caused by differences in equipment interfaces; established a data cleaning mechanism based on dynamic statistical ranges, effectively eliminating sensor drift and transient interference noise; and adopted standardized processing to convert sensor data of different dimensions into a unified distribution space, providing a mathematical basis for multimodal feature fusion. In some embodiments, the raw sensor data is processed to a uniform scale, preferably using the Z-score normalization method, which satisfies expression two: ; These are the standardized values.

[0041] x The raw data values ​​collected by the sensor unit.

[0042] μ This is the arithmetic mean of the historical sensing data for this type of sensor unit.

[0043] σ This represents the standard deviation of this type of sensor unit in historical sensing data.

[0044] The above expression transforms a raw data point x Subtract the historical average of this type of data μ Divide by its historical standard deviation σ This operation transforms data with different units and dimensions (such as temperature and pressure) into a standard distribution with a mean of 0 and a standard deviation of 1, thereby eliminating scale differences.

[0045] In this embodiment, the feature extraction and feature fusion of the processed sensor data specifically includes the following steps: For the sensing data of each sensor unit, a fixed-length sliding time window is set. At each time point, multiple data points within the current window are acquired, and statistical characteristics are calculated. The statistical characteristics include the current value, the window mean, and the window standard deviation. Construct a multi-core matrix set; wherein the multi-core matrix set includes linear kernels, polynomial kernels, and radial basis function kernels; Construct a weighted combination kernel matrix; Perform spatial transformation fusion based on category information.

[0046] Understandably, in the temporal feature extraction stage, the continuous data stream of each sensor unit is divided into a fixed-length sliding window. At each time point, the current value, mean, and standard deviation within the window are extracted to form a feature vector containing both local dynamics and global statistical characteristics. In the multi-kernel transformation and fusion stage, a linear kernel is used to capture simple linear correlations between sensors, a polynomial kernel is used to model higher-order interaction relationships, and a radial basis function kernel is used to fit complex nonlinear patterns. The combination of these three kernels covers different levels of correlation characteristics between data. By dynamically adjusting the weight coefficients of each kernel function, the weighted combined kernel matrix can adaptively balance the contributions of different kernels. Subsequently, the combined kernel matrix is ​​spatially transformed using category structure information, and the similarity constraints between categories are encoded through the Laplacian matrix. Prior category knowledge is integrated into the feature fusion process, ultimately generating a highly discriminative warning feature vector.

[0047] It is also understandable that the above content can fully explore the dynamic changes of multi-sensor data in the time dimension, effectively integrate the linear, nonlinear and high-order interactive features captured by different kernel functions, and improve the discrimination ability of fused features through feature space optimization guided by category information, thereby significantly improving the early warning accuracy of multi-physics coupling faults in complex industrial scenarios.

[0048] In some embodiments, the linear kernel satisfies: ; The polynomial kernel satisfies: ; The radial basis function kernel satisfies: ; The weighted combination kernel matrix satisfies: ; in, , , These are the hyperparameters in each core; K m For the first m The kernel matrix is ​​calculated using a basic kernel function (such as a linear kernel); M The total number of basic kernel functions; β m Give the first m The weight of the seed kernel satisfies the condition that the weight of the seed kernel satisfies the condition. , ; This is the final weighted combination kernel matrix.

[0049] By defining a set of kernel functions that capture data relationships from different perspectives such as linear, polynomial, and nonlinear, and then by weighted summing of these individual kernel matrices, a composite kernel matrix is ​​obtained. This composite matrix integrates multiple data patterns and can describe the complex relationships between data more comprehensively than any single kernel matrix.

[0050] In this embodiment, the spatial transformation fusion based on category information specifically includes the following steps: A category modulation matrix is ​​obtained by fusing the constructed weighted combination kernel matrix with the category structure information; Perform eigenvalue decomposition on the category modulation matrix; The final transformation matrix is ​​calculated using the results of eigenvalue decomposition. Construct multi-core relation vectors; The multi-core relation vector is multiplied on the left by the final transformation matrix to obtain the early warning feature vector data; The category structure information is obtained from the Laplace matrix.

[0051] Understandably, embedding the topological relationships between sample categories into the multi-kernel feature space using the Laplacian matrix results in higher similarity among samples of the same category in the fused feature space. The construction process of the category modulation matrix jointly models the category structure information with multi-kernel features, strengthening the correlation between samples of the same category. After eigenvalue decomposition of the modulation matrix, the eigenvectors corresponding to the k largest eigenvalues ​​are selected to form the transformation matrix, realizing the projection of high-dimensional features onto the low-dimensional discriminative space. The multi-kernel relationship vector captures the complementary descriptions of sample relationships by different kernel functions by calculating the similarity between the test sample and historical support vectors under multiple kernel functions. Finally, a low-dimensional warning feature vector with class separability is generated through a linear combination of the transformation matrix and the multi-kernel relationship vector.

[0052] It is also understandable that the above content effectively solves the problem of insufficient feature discrimination caused by the lack of category information in the process of multi-sensor data fusion. By embedding category structure information into the multi-kernel feature space, the intrinsic correlation between samples of the same type in the fused features is enhanced, so that the subsequent classification model can more accurately distinguish the feature patterns corresponding to different security levels.

[0053] In some embodiments, The category modulation matrix satisfies: ; Eigenvalue decomposition satisfies: ; The final transformation matrix satisfies: ; Multi-core relation vectors satisfy: ; The final warning feature vector satisfies: ; Where L is the Laplacian matrix, used to encode category structure information; Q is the eigenvector matrix of G; Let G be the diagonal matrix of eigenvalues; This is the final transformation matrix; The multi-kernel relation vector of the sample; This is the final warning feature vector for the sample.

[0054] The above process involves fusing multi-core information with category structure information L (obtained from the Laplacian matrix) to generate a category modulation matrix G. Then, eigenvalue decomposition is performed on G to calculate the final transformation matrix. For any sample to be analyzed, its multi-core relationship vector with all historical samples is calculated first. Finally, the relationship vector is multiplied by the transformation matrix on the left to complete the projection and obtain the final warning feature vector data.

[0055] Understandably, category structure information refers to the correlation between different fault types. For example, faults like 'motor overheating' and 'overload' may have similar data features. The role of the Laplacian matrix here is to construct a topological graph of fault types, ensuring that similar fault modes are closer together in the feature space during feature projection, while normal and fault states are further apart. More specifically, assuming there are three states in the historical database: normal, slight wear, and severe fault, the adjacency matrix of these samples in the feature space is calculated, and then the Laplacian matrix is ​​obtained by subtracting the degree matrices. This matrix serves as a constraint to correct the projection direction.

[0056] Example 2: This embodiment only describes the parts that differ from Embodiment 1. Specifically, the early warning method further includes the following steps: Historical sensor data is acquired, and the data early warning model is optimized according to the set time period. Within the parameter range of the kernel function and penalty factor, an initial population containing several candidate solutions of kernel function and penalty factor is generated through tent chaos mapping and reverse learning strategy. For each candidate solution in the initial population, configure a temporary data warning model to train the temporary data warning model using historical sensor data; The artificial hummingbird algorithm incorporates Levy's flight guidance for foraging, allowing for migration to escape local optima. After a set number of iterations, a new kernel function and penalty factor are obtained and updated in the data early warning model to complete the iterative update of the data early warning model.

[0057] It should be noted that during parameter optimization, a set of candidate solutions with ergodicity is first generated through tent chaotic mapping. Then, a back-learning strategy is used to generate a corresponding symmetric solution for each candidate solution, forming an initial population covering a wider parameter space. Each candidate solution is configured as a temporary model, trained using historical sensor data, and its early warning performance is evaluated. In the iterative optimization phase, individual hummingbirds in the artificial hummingbird algorithm perform territorial defense or migratory foraging behaviors based on their fitness values. The migratory foraging behavior utilizes the Levy flight mechanism to generate long-distance hopping search paths, effectively preventing the algorithm from prematurely converging to local optima. When the preset number of iterations is reached, the optimal combination of kernel function and penalty factor parameters is selected and updated to the early warning model, achieving dynamic adjustment of the model parameters.

[0058] In some embodiments, the tent chaotic mapping satisfies: ; The reverse learning strategy satisfies: ; Levi's flight-guided foraging candidate solution The best update will be selected from the standard update and the flight update from Levy. The standard update is as follows: ; Flight updates are as follows: ; in, X ( n ) is a chaotic variable in the interval [0,1]. The inverse solution to the candidate solution; , These are the upper and lower limits of the parameter search space, respectively; To satisfy the random step size of the Lévy distribution; This represents the hummingbird's target food source location (i.e., the current optimal solution).

[0059] Understandably, the process begins by generating a high-quality initial population, where each individual is a candidate solution for the (kernel function, penalty factor) combination. During the iteration process, the iterative unit employs an improved artificial hummingbird algorithm, the core of which is the described guided foraging behavior. This behavior incorporates Levy flight, enabling the algorithm to effectively escape local optima (achieved through migratory foraging) by using a combination of long and short walks when searching for the optimal parameter combination. After reaching the set number of iterations, the algorithm outputs the optimal kernel function and penalty factor combination and updates the data warning model.

[0060] It is also understandable that when searching for the optimal support vector machine parameters, conventional algorithms are prone to getting trapped in local optima (i.e., finding a parameter combination that seems to have high accuracy, but actually has poor generalization ability). The introduction of Levy flight simulates the behavior of organisms foraging over long distances. That is, when searching for parameters, it occasionally makes a 'long-distance jump', forcing the model to try vastly different parameter combinations, thereby escaping the local trap and finding the global optimal solution.

[0061] Example 3: As attached Figure 2 As shown, this embodiment provides a multi-sensor early warning system based on the Industrial Internet of Things. The early warning system includes a management platform, a sensor network platform, and an object platform that are sequentially connected in communication. The management platform includes the following components that are sequentially connected in communication: A data acquisition network, comprising several sensor units, wherein the several sensor units respectively acquire different sensor data of the industrial Internet of Things; The data processing module is used to parse and process the sensing data acquired by different sensor units into sensing data at the same scale. The data fusion module is used to extract and fuse features from the processed sensor data to obtain early warning feature vector data. The data early warning module is equipped with a vector machine-based data early warning model. After acquiring early warning feature vector data, the data early warning model executes a decision function to calculate the kernel function value between the input early warning feature vector data and all support vectors using the optimal kernel function. Then, it combines the optimal penalty factor to train and obtain Lagrange multipliers and biases, and performs a weighted summation of the Lagrange multipliers and biases to obtain the category label of the data acquisition network. The data alarm module is used to map category labels to the corresponding security levels of the Industrial Internet of Things.

[0062] It should be noted that the data acquisition network acquires multimodal data streams from the industrial site in real time through distributed sensor units, solving the problem of insufficient dimensionality in traditional single-sensor data. The data processing module performs protocol parsing and format unification on the raw data, establishes stable range thresholds by calculating the statistical characteristics of historical data, removes or smooths outliers exceeding the range, and eliminates dimensional differences through standardization and scaling to form a uniform-scale analysis dataset. The data fusion module uses a sliding time window to extract the temporal statistical features of each sensor, constructs a multi-kernel matrix set including linear kernels, polynomial kernels, and radial basis function kernels, and achieves deep fusion of cross-modal features through weighted combination and spatial transformation of the category modulation matrix. The data early warning module uses the optimal kernel function to calculate the similarity between the input features and support vectors, optimizes the classification boundary by combining a penalty factor, and generates category labels representing the health status of equipment through weighted summation of Lagrange multipliers and bias terms. The data alarm module converts the numerical classification results into corresponding early warning signals according to preset safety level mapping rules, forming a complete closed loop from data acquisition to decision output.

[0063] It should also be noted that in this embodiment, the linear kernel is mainly used to capture linear trend features in sensor data (such as a steady rise in temperature); the radial basis function (RBF) kernel is mainly used to capture local nonlinear features in high-frequency abrupt signals (such as vibration spikes at the moment of bearing fracture). The physical significance of constructing a weighted combined kernel matrix is ​​that faults in industrial settings are often complex and cannot be described by a single feature. By combining them, both linear trends and nonlinear abrupt changes in equipment can be monitored simultaneously. Furthermore, the system does not directly perform abstract matrix operations, but first maps the time series data of different sensors to a high-dimensional feature space, and then performs a weighted summation of feature distances in different dimensions using preset weight coefficients to obtain a fusion distance that comprehensively reflects the health of the equipment.

[0064] Example 4: In this embodiment, the early warning system further includes an optimization module, which is used to acquire historical sensor data and optimize the data early warning model according to a set time period, so as to generate an initial population containing several candidate solutions of kernel function and penalty factor through tent chaos mapping and reverse learning strategy within the parameter range of kernel function and penalty factor.

[0065] It should be noted that the optimization module automatically uses historical sensor data as the training set when the set time period arrives. First, within the feasible region of the kernel function type and penalty factor value, a uniformly distributed sequence of initial parameters is generated using a tent chaotic map. This process breaks the correlation between parameters through the inherent randomness of the chaotic system. Then, a back-learning strategy is employed to compute the symmetric solution for each initial parameter solution, forming a candidate set that is twice the size, containing both the original and mirror solutions. By combining the ergodicity of the chaotic map with the symmetric expansion of back-learning, the candidate solution set can cover a wider parameter space. This mechanism enables subsequent optimization algorithms to have stronger global search capabilities during iteration, avoiding premature convergence caused by uneven distribution of initial solutions.

[0066] In this embodiment, the optimization module includes: An evaluation unit is configured to configure a temporary data warning model for each candidate solution in the initial population, so as to train the temporary data warning model using historical sensor data. An iterative unit is used to introduce Levy's flight guidance for foraging into the artificial hummingbird algorithm, so as to escape the local optimum through migration foraging. After reaching a set number of iterations, the iterative unit obtains a new kernel function and penalty factor and updates it to the data early warning model to complete the iterative update of the data early warning model.

[0067] It should be noted that after receiving a set of candidate solutions containing kernel function types, hyperparameters, and penalty factor values, the evaluation unit creates an independent data early warning model instance for each candidate solution. Model training and validation are performed by loading historical sensor datasets, recording performance metrics such as classification accuracy and false alarm rate for each candidate solution. The iterative unit initializes the population position based on the artificial hummingbird algorithm framework, dynamically adjusting the search strategy according to the fitness values ​​of individual hummingbirds during each iteration. When the population is detected to be trapped in a local optimum, the long-step hop generated by Levy's flight is used to update the hummingbird position, combined with the information sharing mechanism among individuals in the migration foraging strategy, guiding the population to migrate to a region with better parameters. When the preset iteration threshold is reached, the kernel function and penalty factor combination with the best fitness is selected to update the early warning model parameters.

[0068] It should be noted that the multi-sensor early warning system based on the Industrial Internet of Things in this embodiment also includes a user platform and a service platform that are interconnected. The service platform is interconnected with the management platform, thus forming a standard five-platform structure for the Internet of Things.

[0069] The physical entities of the user platform include various user terminals, such as mobile phones, computers, and dedicated terminals, which provide user services through integration with user information system software.

[0070] A service platform is a functional platform that enables service communication.

[0071] The management platform is a functional platform for realizing the operation and management of the Internet of Things system. In some embodiments, the management platform may include multiple object sub-platforms, which are respectively connected to the aforementioned sensor network platform. Each object sub-platform includes a data acquisition network, a data processing module, a data fusion module, a data early warning module, and a data alarm module.

[0072] A sensor network platform is a functional platform for realizing sensor communication.

[0073] The object platform is a functional platform for realizing perception and control.

[0074] Example 5: Based on the same inventive concept as the foregoing embodiments, this embodiment provides a computer device, which includes a memory and a processor. The memory stores a computer program, and the processor executes the computer program to implement the above-described method.

[0075] Example 6: Based on the same inventive concept as the foregoing embodiments, this embodiment provides a computer-readable storage medium storing a computer program, and a processor executes the computer program to implement the above-described method.

[0076] The above are merely preferred embodiments of the present invention and do not limit the scope of the patent. Any equivalent structural or procedural transformations made based on the description and drawings of the present invention, or direct or indirect applications in other related technical fields, are similarly included within the scope of patent protection of the present invention.

Claims

1. A multi-sensor early warning method based on the Industrial Internet of Things, characterized in that, Includes the following steps: Acquire sensor data; The acquired sensor data is parsed and processed into sensor data at the same scale; Feature extraction and feature fusion are performed on the processed sensor data to obtain early warning feature vector data; The decision function is executed to obtain the early warning feature vector data, and the kernel function value between the input early warning feature vector data and all support vectors is calculated using the optimal kernel function. Then, the optimal penalty factor is combined to train and obtain the Lagrange multipliers and biases. The class label is obtained by weighted summation of the Lagrange multipliers and biases. Map category labels to the corresponding security levels for the Industrial Internet of Things (IIoT).

2. The multi-sensor early warning method based on the Industrial Internet of Things as described in claim 1, characterized in that, The process of parsing and processing the acquired sensor data into sensor data at the same scale specifically includes the following steps: The original data stream is obtained through a parser and uniformly encapsulated into standard internal data; Calculate the average value of each sensor unit in the historical sensing data, and set a stable range based on the average value. Then, remove or smooth the internal data that is not within the stable range. The internal data is transformed into a distribution with a mean of 0 and a standard deviation of 1 by subtracting the original data value of the data point from the average value of the data point in the historical sensor data, and then dividing the difference by the standard deviation of the data point.

3. The multi-sensor early warning method based on the Industrial Internet of Things as described in claim 1, characterized in that, The feature extraction and feature fusion of the processed sensor data specifically includes the following steps: For the sensing data of each sensor unit, a fixed-length sliding time window is set. At each time point, multiple data points within the current window are acquired, and statistical characteristics are calculated. The statistical characteristics include the current value, the window mean, and the window standard deviation. Construct a multi-core matrix set; wherein the multi-core matrix set includes linear kernels, polynomial kernels, and radial basis function kernels; Construct a weighted combination kernel matrix; Perform spatial transformation fusion based on category information.

4. The multi-sensor early warning method based on the Industrial Internet of Things as described in claim 3, characterized in that, The execution of spatial transformation fusion based on category information specifically includes the following steps: A category modulation matrix is ​​obtained by fusing the constructed weighted combination kernel matrix with the category structure information; Perform eigenvalue decomposition on the category modulation matrix; The final transformation matrix is ​​calculated using the results of eigenvalue decomposition. Construct multi-core relation vectors; The multi-core relation vector is left-multiplied by the final transformation matrix to obtain the early warning feature vector data; The category structure information is obtained from the Laplace matrix.

5. The multi-sensor early warning method based on the Industrial Internet of Things as described in claim 1, characterized in that, The early warning method also includes the following steps: Historical sensor data is acquired, and the data early warning model is optimized according to the set time period. Within the parameter range of the kernel function and penalty factor, an initial population containing several candidate solutions of kernel function and penalty factor is generated through tent chaos mapping and reverse learning strategy. For each candidate solution in the initial population, configure a temporary data warning model to train the temporary data warning model using historical sensor data; The artificial hummingbird algorithm incorporates Levy's flight guidance for foraging, allowing for migration to escape local optima. After a set number of iterations, a new kernel function and penalty factor are obtained and updated in the data early warning model to complete the iterative update of the data early warning model.

6. A multi-sensor early warning system based on the Industrial Internet of Things, characterized in that, The early warning system includes a management platform, a sensor network platform, and an object platform that are sequentially connected in communication. The management platform includes the following components that are sequentially connected in communication: A data acquisition network, comprising several sensor units, wherein the several sensor units respectively acquire different sensor data of the industrial Internet of Things; The data processing module is used to parse and process the sensing data acquired by different sensor units into sensing data at the same scale. The data fusion module is used to extract and fuse features from the processed sensor data to obtain early warning feature vector data. The data early warning module is equipped with a vector machine-based data early warning model. After acquiring early warning feature vector data, the data early warning model executes a decision function to calculate the kernel function value between the input early warning feature vector data and all support vectors using the optimal kernel function. Then, it combines the optimal penalty factor to train and obtain Lagrange multipliers and biases, and performs a weighted summation of the Lagrange multipliers and biases to obtain the category label of the data acquisition network. The data alarm module is used to map category labels to the corresponding security levels of the Industrial Internet of Things.

7. The multi-sensor early warning system based on the Industrial Internet of Things as described in claim 6, characterized in that, The early warning system also includes an optimization module, which is used to acquire historical sensor data and optimize the data early warning model according to a set time period. Within the parameter range of the kernel function and penalty factor, an initial population containing several candidate solutions of kernel function and penalty factor is generated through tent chaos mapping and reverse learning strategy.

8. The multi-sensor early warning system based on the Industrial Internet of Things as described in claim 7, characterized in that, The optimization module includes: An evaluation unit is configured to configure a temporary data warning model for each candidate solution in the initial population, so as to train the temporary data warning model using historical sensor data. An iterative unit is used to introduce Levy's flight guidance for foraging into the artificial hummingbird algorithm, so as to escape the local optimum through migration foraging. After reaching a set number of iterations, the iterative unit obtains a new kernel function and penalty factor and updates it to the data early warning model to complete the iterative update of the data early warning model.

9. A computer device, characterized in that, The computer device includes a memory and a processor, wherein the memory stores a computer program and the processor executes the computer program to implement the method as described in any one of claims 1-5.

10. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a computer program, and the processor executes the computer program to implement the method as described in any one of claims 1-5.

Images