Abnormal early warning method and system based on artificial intelligence

Abstract

The invention discloses an abnormity early warning method and system based on artificial intelligence, and relates to the technical field of abnormity early warning. The method comprises the following components: S1, brain wave data acquisition and preprocessing, S2, emotional state recognition and feature extraction, S3, specific business data acquisition and arrangement, S4, data fusion and abnormal feature construction, and S5, abnormal early warning decision and notification. According to the method, the brain wave emotion features and the specific business data are fused, abnormal feature construction is carried out by adopting a fusion algorithm of multi-scale feature matching and a support vector machine algorithm based on quantum annealing optimization, the abnormal feature rule can be mined more accurately through the comprehensive data processing and analysis mode, and the accuracy of abnormal feature extraction is improved. Therefore, in the real-time data monitoring and analysis process, the accuracy of abnormal early warning is improved, meanwhile, the system can dynamically adjust the monitoring frequency according to the fluctuation characteristics of the data, the timeliness of early warning information is ensured, and related personnel can take countermeasures in time.

Description

Technical Field

[0001] The present invention relates to the technical field of anomaly warning, and specifically provides an anomaly warning method and system based on artificial intelligence. Background Art

[0002] In modern complex working environments and business systems, whether in financial transactions, industrial production control, or medical surgery assistance, the timely detection and warning of anomalies are of crucial significance for ensuring the safe and stable operation of the system, avoiding major losses, and ensuring the smooth progress of operations.

[0003] Traditional technologies have deficiencies. Firstly, the collection and processing technologies of emotional brainwaves are relatively complex, requiring professional equipment and personnel for operation, which increases the application cost and technical difficulty. Secondly, traditional methods often ignore the correlation between emotional brainwaves and specific business data, and only analyze emotional brainwaves as a single data source, which limits the accuracy and comprehensiveness of warning. In addition, when processing emotional brainwave data, traditional methods usually adopt relatively simple algorithms and models, making it difficult to deeply mine and extract effective emotional features, thus affecting the warning effect.

[0004] In summary, there are many deficiencies in traditional anomaly warning methods when using emotional brainwaves for auxiliary warning. Therefore, it is particularly important to develop an anomaly warning method and system based on artificial intelligence. Summary of the Invention

[0005] The purpose of the present invention is to make up for the deficiencies of the existing technology, and provides an anomaly warning method and system based on artificial intelligence, which can fuse brainwave emotional features with specific business data, and use artificial intelligence technology to construct anomaly features and make warning decisions, aiming to achieve more accurate and real-time anomaly warning.

[0006] To solve the above technical problems, the present invention provides the following technical solution: An anomaly warning method based on artificial intelligence, and the specific steps of this method are as follows:

[0007] S1. Collection and preprocessing of brainwave data

[0008] Using a multi-channel flexible brainwave collection device, for the collected original brainwave signals, first remove the common-mode interference, and the algorithm formula is:

[0009]

[0010] where x(n) is the sampling value of the original brainwave signal at time n, y(n) is the filtered output, w i (n) is the i-th filter coefficient at time n, M is the filter order, w i(n) is initially estimated according to the prior spectral characteristics of the electroencephalogram signal, and then in the signal processing process, it is updated according to the least mean square error criterion through the iterative formula w i (n + 1) = w i (n) + μe(n)x(n - i), where μ is the step size factor, e(n) = d(n) - y(n), d(n) is the desired signal, and the signal after adaptive filtering is amplified. Let the initial intensity be I0, and the target intensity range is [I min , I max , and the amplification factor Then, a method combining empirical mode decomposition (EMD) and independent component analysis (ICA) is used to extract electroencephalogram signals in different frequency bands. First, the amplified signal is decomposed by EMD to obtain several intrinsic mode functions (IMFs), and then the IMFs are analyzed by ICA to separate the independent components corresponding to different frequency bands. Specifically, by constructing the objective function of ICA where W is the separation matrix, s i and s j are the separated signal components, corr represents the correlation function, N is the number of signal components, and W is solved by an optimization algorithm to achieve frequency band separation. The crossover probability P c of the genetic algorithm ranges from 0.6 to 0.9, and the mutation probability P m ranges from 0.01 to 0.1;

[0011] S2. Emotional state recognition and feature extraction

[0012] The preprocessed electroencephalogram signal is input into the emotional recognition neural network model. The model structure combines a bidirectional long short-term memory network (BiLSTM) and an attention mechanism. The number of hidden units in the BiLSTM layer is set to H. The weight matrices and bias vectors of its forget gate, input gate, and output gate are initialized using the Xavier initialization method. The attention score calculation function of the attention mechanism is e t = tanh(W e h t + b e ), where h t is the hidden state of the BiLSTM at time t, W e and b e are trainable parameters, and the attention weight T is the sequence length. The model training uses the Adam optimization algorithm with adaptive learning rate adjustment. Its initial learning rate is set to L r . During the training process, every E epochs, if the loss function has not decreased for B consecutive batches, the learning rate is multiplied by the decay factor D. The loss function uses the multi-label cross-entropy loss function where C is the number of emotional categories, M is the number of samples, and yij The true value of the i-th sentiment label of the j-th sample in the sixth is the predicted value. Through training with a large amount of EEG data labeled with different emotional states, emotional feature vectors are obtained. This vector contains the energy values of EEG waves in each frequency band weighted by attention at different time points and the information on the dynamic proportional relationship between the energy values of different frequency bands. An element v of a certain emotional feature vector ij represents the ratio of the energy value of the i-th frequency band weighted by attention at the j-th time point to the total energy value of all frequency bands at this time point, that is where E ij is the energy value of the i-th frequency band at the j-th time point, and K is the total number of frequency bands;

[0013] S3. Specific business data collection and collation

[0014] Deeply dock with the business system, and obtain specific business data with the help of specially developed data collection middleware. In the financial field, it covers the bid-ask spread sequence of stocks, block trade data, and changes in margin trading balances. In the industrial field, it includes the material flow fluctuations in each link of the production line, wear amount monitoring data of key components of equipment, and data on the impact of environmental humidity on the production process. In the medical field, it involves specific gene expression data of patients at different disease stages, intermediate metabolite concentration data during drug metabolism, and electromyogram data during rehabilitation training. Clean the collected business data, and use a data cleaning algorithm based on density clustering. By calculating the local density of data points where d i j is the distance between data points i and j, and σ is the density parameter and relative density Identify and remove the noise points in the data, and at the same time correct the outliers of the data. By comparing with the mean and standard deviation of the same type of data, if the data point exceeds the range of the mean plus or minus 3 times the standard deviation, it is corrected according to the local trend of the data, verify the accuracy of the data, and use a data integrity verification method based on the hash function. Convert the business data into a hash value of a fixed length and compare it with the pre-stored correct hash value, standardize the data, and according to the statistical characteristics of the business data and the subsequent analysis requirements, use a quantile-based standardization method. Let q p be the p-th quantile of the data, then the standardized data Organize the business data into a sequence set according to the time series;

[0015] S4. Data fusion and abnormal feature construction

[0016] Fuse the emotional feature vector with the specific business data sequence, and adopt a fusion algorithm based on multi-scale feature matching. First, perform multi-scale decomposition on the emotional feature vector and the business data sequence respectively. The emotional feature vector uses wavelet packet decomposition, and the business data sequence uses sliding window Fourier transform. For wavelet packet decomposition, the decomposition level is determined according to the dimension and complexity of the emotional feature vector. Let the dimension of the emotional feature vector be D, and the decomposition level Obtain sub-feature vectors at different scales through wavelet packet decomposition, and obtain business sub-data sequences in different frequency bands. Then calculate the similarity matrix S between the emotional sub-feature vectors and the business sub-data sequences at different scales. The similarity measurement uses a method that combines dynamic time warping DTW and mutual information. s ij =MI ij ×(1 - DTW ij ), where MI ij is the mutual information between the i-th emotional sub-feature vector and the j-th business sub-data sequence, and DTW ij is the DTW distance between them. The calculation of MI ij is carried out by constructing a joint probability distribution matrix P, and then calculated according to the formula . The calculation of DTW ij solves the minimum path cost through a dynamic programming algorithm. Determine the fusion weight vector W according to the similarity matrix. The weight where I is the number of emotional sub-feature vectors and J is the number of business sub-data sequences. Finally, obtain the fusion feature vector through weighted summation where e i is the i-th emotional sub-feature vector, and b j is the j-th business sub-data sequence. On this basis, use the support vector machine algorithm optimized by quantum annealing to construct abnormal features for the fusion feature vector. The kernel function of SVM uses the polynomial kernel function K(x,y)=(x T y + c) d , where c is a constant term and d is the polynomial degree. The Hamiltonian in the quantum annealing optimization algorithm is H = H p + H m , where H p is the problem Hamiltonian and H m is the transverse field Hamiltonian. Find the optimal parameter combination of SVM through the quantum annealing process to determine the abnormal feature model;

[0017] S5. Abnormal warning decision and notification

[0018] Apply the constructed abnormal feature model to real-time data monitoring and analysis. When new fused data is input into the model, the model makes a judgment based on the abnormal feature rules learned in advance. If it detects that the current data conforms to the abnormal feature pattern, it immediately triggers an abnormal warning decision. The warning decision system determines the warning level, warning method, and warning recipients according to the pre-set warning strategies, and conveys the warning information to relevant personnel in a timely and accurate manner through various communication channels and terminal devices.

[0019] Furthermore, in the brain wave data acquisition and preprocessing step, the electrode layout of the multi-channel flexible brain wave acquisition device adopts a non-uniform distribution method. According to the activity differences in brain functional areas, electrodes with a higher density are set in the frontal lobe and temporal lobe emotional-related brain areas, and the electrode spacing is between 1 mm and 3 mm. While in the parietal lobe and occipital lobe areas, the electrode spacing is appropriately increased to 5 mm - 10 mm. This layout method can focus on capturing the weak signal changes in the emotional-related brain areas, while taking into account the whole-brain signal acquisition, improving the pertinence and effectiveness of emotional brain wave signal acquisition, providing a more discriminative signal source for subsequent emotional state recognition. And in the adaptive filtering algorithm, the prediction model for common-mode interference signals is pre-trained using a deep learning-based signal generation model. This model uses a large number of known interference signals and interference-free brain wave signals as training data to learn the feature patterns of interference signals, so as to more accurately determine the initial filter coefficient w i (n), improving the filtering effect and reducing the loss of useful brain wave signals.

[0020] Even further, in the emotional state recognition and feature extraction step, in the model combining bidirectional long short-term memory network and attention mechanism, a feature compression layer is added after the BiLSTM layer. The principal component analysis algorithm is used to reduce the dimension of the output of BiLSTM, and the principal components with a cumulative contribution rate of more than 90% are retained to reduce the computational amount of the subsequent attention mechanism and highlight the main emotional feature information. At the same time, during the model training process, the early stopping method is used to prevent overfitting. Set the loss function of the validation set to stop training when it has not decreased for E s consecutive epochs, and save the model parameters at this time. In addition, to improve the adaptability of the model to the brain wave feature differences of different individuals, individual identification information is added to the training data and used as an additional feature input into the model, so that the model can learn the subtle differences in brain wave features of different individuals in the same emotional state, and thus more accurately identify the emotional state of individuals and extract effective emotional feature vectors.

[0021] Furthermore, in the specific business data collection and collation step, the data collection middleware has an intelligent data caching function. It dynamically adjusts the cache size according to the real-time traffic and processing speed of business data. When the business data traffic suddenly increases, the cache size automatically increases to avoid data loss. The adjustment of the cache size is calculated based on the traffic change rate and the current cache occupancy rate. Let the traffic change rate be r, the current cache occupancy rate be o, and the cache adjustment coefficient be k. Then the new cache size C new = C old (1 + k × r × (1 - o)), where C old is the original cache size. And in the data cleaning algorithm based on density clustering, for the calculation of the local density of data points, a time decay factor is introduced, that is where λ is the time decay coefficient, t i and t j are the timestamps of data points i and j, so that recent data points have higher weights in the clustering analysis, which is more in line with the dynamic change characteristics of business data, and improves the accuracy and timeliness of data cleaning.

[0022] Furthermore, in the data fusion and abnormal feature construction step, in the fusion algorithm based on multi-scale feature matching, when performing wavelet packet decomposition and sliding window Fourier transform, for different business types and emotional feature complexities, adaptive decomposition scales and window sliding step sizes are adopted. In the financial field, due to the rapid and complex data changes, the wavelet packet decomposition scale is set to L f -L g layers, and the window sliding step size of the sliding window Fourier transform is set to S1 - S2 data points. In the industrial field, the data is relatively stable, and the wavelet packet decomposition scale is L h ·L i layers, and the window sliding step size is S3 - S4 data points. In the medical field, it is adjusted according to the acute or chronic degree of the disease. For data related to acute diseases, a higher decomposition scale L j -L k layers and a smaller window sliding step size of S5 - S6 data points are adopted. For data related to chronic diseases, it is the opposite. This adaptive setting can better match the characteristics of data in different fields, improve the matching accuracy between the emotional feature vector and the business data sequence, so as to obtain more accurate fusion weight vectors and fusion feature vectors, and help to more accurately mine the abnormal feature rules.

[0023] Furthermore, in the data fusion and abnormal feature construction step, in the support vector machine algorithm based on quantum annealing optimization, when constructing the problem Hamiltonian H of quantum annealing pWhen considering this, in addition to the objective function of SVM, a measure of data uncertainty is introduced. Let the uncertainty of the fused feature vector be U, which is measured by calculating the entropy value of the feature vector, i.e., U = -∑ i p i logp i , where p i is the probability distribution of the i-th element in the feature vector. The uncertainty term is added to H p so that the optimization process can take into account the data uncertainty, improve the adaptability and robustness of the abnormal feature model to complex and changing data. At the same time, the annealing schedule in the quantum annealing process adopts an adaptive adjustment strategy, which is adjusted according to the quality of the currently searched solution and the search progress. If the quality of the solution has not improved in several consecutive iterations, the annealing speed is increased to improve the search efficiency and reduce the consumption of computing resources.

[0024] Furthermore, in the abnormal warning decision and notification step, the warning strategy is set using a collaborative decision-making method based on a multi-agent system. Multiple agents are established, including a data monitoring agent, an emotion analysis agent, a business risk assessment agent, and a warning level determination agent. The data monitoring agent is responsible for real-time monitoring of changes in the fused data, and its monitoring frequency is dynamically adjusted according to the fluctuation characteristics of the data. Let the average change rate of the data be V. When V ≥ V1, where V1 is a preset high-speed change threshold, the monitoring frequency is F5 times per second; when V1 > V ≥ V2, where V2 is a medium-speed change threshold, the monitoring frequency is F6 times per second; when V < V2, the monitoring frequency is F7 times per second. The emotion analysis agent conducts in-depth analysis of the emotion feature vector to judge the degree and trend of emotion abnormality. Its analysis model uses emotion transfer analysis based on a hidden Markov model. By constructing an emotion state transition matrix T, where the element t ij represents the probability of transferring from emotion state i to emotion state j, and T is obtained by training with a large amount of historical emotion data, and then the emotion change trend is predicted. The business risk assessment agent evaluates the degree of business risk based on the business data sequence and uses a risk scoring model. This model calculates the risk score S according to the deviation degree of the business data from the preset risk indicators. The warning level determination agent determines the warning level by comprehensively considering the analysis results of each agent and uses a fuzzy inference system. With the emotion abnormality degree score E s and the business risk score S as input variables and the warning level L as the output variable, by establishing a fuzzy rule base, multi-factor comprehensive judgment is realized, the accuracy and scientificity of warning level determination are improved, so that the warning information can more accurately reflect the actual abnormal situation, so that relevant personnel can take appropriate countermeasures. Each agent communicates through an information interaction protocol, and this protocol uses blockchain-based encrypted communication technology to ensure the security and integrity of information transmission and prevent information from being tampered with or stolen.

[0025] On the other hand, an anomaly warning system based on artificial intelligence, characterized in that the system includes an electroencephalogram data acquisition and preprocessing module, an emotional state recognition and feature extraction module, a specific business data acquisition and collation module, a data fusion and anomaly feature construction module, and an anomaly warning decision-making and notification module:

[0026] The electroencephalogram data acquisition and preprocessing module: Using a multi-channel flexible electroencephalogram acquisition device, for the collected original electroencephalogram signals, first remove the common-mode interference, and the algorithm formula is:

[0027]

[0028] where x(n) is the sampling value of the original electroencephalogram signal at time n, y(n) is the filtered output, w i (n) is the i-th filter coefficient at time n, M is the filter order, and the initial value of w i (n) is estimated according to the prior spectral characteristics of the electroencephalogram signal, and then in the signal processing process, according to the least mean square error criterion, through the iterative formula w i (n + 1) = w i (n) + μe(n)x(n - i) for update, where μ is the step factor, e(n) = d(n) - y(n), d(n) is the desired signal, and the signal after adaptive filtering is amplified. Let the initial intensity be I0, and the target intensity range is [I min , I max , and the amplification factor Then use a method combining empirical mode decomposition (EMD) and independent component analysis (ICA) to extract electroencephalogram signals in different frequency bands. First, perform EMD decomposition on the amplified signal to obtain several intrinsic mode functions (IMFs), and then perform ICA analysis on the IMFs to separate the independent components corresponding to different frequency bands. Specifically, by constructing the objective function of ICA where W is the separation matrix, s i and s j are the separated signal components, corr represents the correlation function, N is the number of signal components, and solve W through an optimization algorithm to achieve frequency band separation. The crossover probability P c of the genetic algorithm ranges from 0.6 to 0.9, and the mutation probability P m ranges from 0.01 to 0.1;

[0029] The emotional state recognition and feature extraction module: Input the preprocessed electroencephalogram (EEG) signals into the emotion recognition neural network model. The model structure combines a bidirectional long short-term memory network (BiLSTM) with an attention mechanism. The number of hidden units in the BiLSTM layer is set to H. The weight matrices and bias vectors of its forget gate, input gate, and output gate are initialized using the Xavier initialization method. The attention score calculation function of the attention mechanism is e t = tanh(W e h t + b e ), where h t is the hidden state of the BiLSTM at time t, W e and b e are trainable parameters, and the attention weight T is the sequence length. The model is trained using the Adam optimization algorithm with adaptive learning rate adjustment. The initial value of its learning rate is set to L r . During the training process, every E epochs, if the loss function has not decreased for consecutive B batches, the learning rate is multiplied by the decay factor D. The loss function uses the multi-label cross-entropy loss function where C is the number of emotion categories, M is the number of samples, y ij is the true value of the i-th emotion label of the j-th sample, is the predicted value. Through training with a large amount of EEG data labeled with different emotional states, an emotional feature vector is obtained. This vector contains the energy values of EEG waves in each frequency band weighted by attention at different time points and the dynamic proportional relationship information between the energy values of different frequency bands. An element v ij of a certain emotional feature vector represents the ratio of the energy value of the i-th frequency band weighted by attention at the j-th time point to the total energy value of all frequency bands at this time point, that is where E ij is the energy value of the i-th frequency band at the j-th time point, and K is the total number of frequency bands;

[0030] The specific business data collection and collation module: Deeply dock with the business system and obtain specific business data with the help of specially developed data collection middleware. In the financial field, it covers the bid-ask spread sequence of stocks, block trade data, and changes in margin trading balances. In the industrial field, it includes the material flow fluctuations in each link of the production line, wear amount monitoring data of key equipment components, and data on the impact of environmental humidity on the production process. In the medical field, it involves specific gene expression data of patients at different disease stages, intermediate metabolite concentration data during drug metabolism, and electromyogram data during rehabilitation training. Clean the collected business data using a data cleaning algorithm based on density clustering by calculating the local density of data points where d ij is the distance between data points i and j, and σ is the density parameter and relative density Identify and remove noise points in the data, and at the same time correct the outliers in the data. By comparing with the mean and standard deviation of the same type of data, if the data point exceeds the range of the mean plus or minus 3 times the standard deviation, it is corrected according to the local trend of the data, verify the accuracy of the data, adopt a data integrity verification method based on the hash function, convert the business data into a hash value of a fixed length, and compare it with the pre-stored correct hash value, standardize the data, according to the statistical characteristics of the business data and the subsequent analysis requirements, adopt a quantile-based standardization method, let q p be the p-th quantile of the data, then the standardized data Organize the business data into a sequence set according to the time series;

[0031] The data fusion and anomaly feature construction module: fuse the sentiment feature vector with a specific business data sequence, adopt a fusion algorithm based on multi-scale feature matching. First, perform multi-scale decomposition on the sentiment feature vector and the business data sequence respectively. The sentiment feature vector adopts wavelet packet decomposition, and the business data sequence adopts sliding window Fourier transform. For wavelet packet decomposition, the decomposition layer is determined according to the dimension and complexity of the sentiment feature vector. Let the dimension of the sentiment feature vector be D, and the decomposition layer Obtain sub-feature vectors at different scales through wavelet packet decomposition, obtain business sub-data sequences in different frequency bands, and then calculate the similarity matrix S between the sentiment sub-feature vectors and the business sub-data sequences at different scales. The similarity metric adopts a method combining dynamic time warping (DTW) and mutual information. s ij =MI ij ×(1 - DTW ij ), where MI ij is the mutual information between the i-th sentiment sub-feature vector and the j-th business sub-data sequence, and DTW ij is the DTW distance between them. The calculation of MI ij is carried out by constructing a joint probability distribution matrix P, and then calculated according to the formula The calculation of DTW ij is solved by the dynamic programming algorithm to find the minimum path cost. Determine the fusion weight vector W according to the similarity matrix. The weight where I is the number of sentiment sub-feature vectors, J is the number of business sub-data sequences, and finally obtain the fusion feature vector through weighted summation where e i is the i-th sentiment sub-feature vector, and b j is the j-th business sub-data sequence. On this basis, use the support vector machine algorithm based on quantum annealing optimization to construct anomaly features for the fusion feature vector. The kernel function of SVM adopts the polynomial kernel function K(x,y)=(xT y+c) d , where c is a constant term, d is the polynomial degree, and the Hamiltonian H in the quantum annealing optimization algorithm is H p +H m , where H p is the Hamiltonian of the problem, H m is the transverse field Hamiltonian, and the optimal parameter combination of SVM is found through quantum annealing process to determine the abnormal feature model;

[0032] The abnormal warning decision and notification module applies the constructed abnormal feature model to real-time data monitoring and analysis. When new fused data is input into the model, the model makes a judgment based on the abnormal feature rules learned in advance. If it is detected that the current data meets the abnormal feature pattern, the abnormal warning decision is triggered immediately. The warning decision system determines the warning level, warning method and warning recipients based on the pre-set warning strategy, and uses a variety of communication channels and terminal devices to convey the warning information to relevant personnel in a timely and accurate manner.

[0033] Compared with the prior art, this abnormal warning method and system based on artificial intelligence has the following beneficial effects:

[0034] 1. The present invention integrates brain wave emotional features with specific business data, adopts a multi-scale feature matching fusion algorithm and a support vector machine algorithm based on quantum annealing optimization to construct abnormal features. This comprehensive data processing and analysis method can more accurately mine the laws of abnormal features, thereby improving the accuracy of abnormal warnings in real-time data monitoring and analysis. At the same time, the system can dynamically adjust the monitoring frequency according to the fluctuation characteristics of the data to ensure the timeliness of the warning information, so that relevant personnel can take timely countermeasures.

[0035] 2. The present invention adopts an adaptive decomposition scale and window sliding step, and introduces a quantum annealing optimization algorithm for measuring data uncertainty. This adaptive and robust design enables the system to better adapt to the characteristics of data in different fields and improve the matching accuracy between sentiment feature vectors and business data sequences. At the same time, the blockchain-based encryption communication technology ensures the security and integrity of information transmission, further enhancing the stability and reliability of the system.

[0036] Other advantages, objectives and features of the present invention will be set forth in part in the following description and, in part, will be apparent to those skilled in the art based on an examination of the following or may be taught from the practice of the present invention. BRIEF DESCRIPTION OF THE DRAWINGS

[0037] To more clearly illustrate the technical solutions in the embodiments of the present invention or in the prior art, the following will briefly introduce the drawings required for the description of the embodiments or the prior art. Obviously, the drawings in the following description are only some embodiments of the present invention. For those of ordinary skill in the art, without creative efforts, other drawings can also be obtained based on these drawings.

[0038] Figure 1 It is a flow operation diagram of an anomaly warning method based on artificial intelligence;

[0039] Figure 2 It is a flow operation diagram of an anomaly warning system based on artificial intelligence. Detailed implementation manners

[0040] To further elaborate on the technical means and effects adopted by the present invention to achieve the predetermined invention purpose, the following will, in combination with the accompanying drawings and preferred embodiments, detail the specific implementation manners, structures, features, and their effects of the present invention as follows.

[0041] Embodiment 1

[0042] In an industrial production scenario, workers wear a multi-channel flexible electroencephalogram acquisition device. The electrode layout is set according to the difference in the activity of brain functional areas. The electrode spacing in the frontal and temporal emotional-related brain areas is maintained at 1 mm - 3 mm, and the electrode spacing in the parietal and occipital regions is 5 mm - 10 mm. The collected original electroencephalogram signals are processed by removing common-mode interference. The initial filter coefficients of the adaptive filtering algorithm are pre-trained and determined by a signal generation model based on deep learning. This model uses a large number of interference and non-interference electroencephalogram signals as training data. Subsequently, the signal is amplified, and the initial intensity is set to I0, and the target intensity range is [I min , I max , and the amplification factor Then, a method combining EMD and ICA is used to extract electroencephalogram signals in different frequency bands. First, the amplified signal is decomposed by EMD to obtain IMFs, and then ICA is used to analyze and separate independent components in different frequency bands. The ICA objective function The separation matrix W is solved by a genetic algorithm, and the crossover probability P c takes a value of 0.6 - 0.9, and the mutation probability P m takes a value of 0.01 - 0.1. For example, when a worker on a production line operates equipment, his electroencephalogram signal is collected and processed.

[0043] By deeply connecting with the production line business system through intelligent data acquisition middleware, data such as the fluctuation of material flow in each link of the production line, the wear amount monitoring data of key components of equipment, and the data of the impact of environmental humidity on the production process are obtained. The middleware dynamically adjusts the cache size according to the real-time flow and processing speed of business data. Let the flow change rate be r, the current cache occupancy rate be o, the cache adjustment coefficient be k, and the new cache size be C new = C old (1 + k × r × (1 - o)), and data cleaning uses a density-based clustering algorithm to calculate the local density of data points (introducing a time decay coefficient λ) and relative density δ i , identify and remove noise points and outliers, correct outliers by comparing with the mean and standard deviation of the same type of data, verify the data accuracy using a hash function, and finally standardize the data based on quantiles and organize it into a sequence set. For example, data from each link of the production line is collected every 5 minutes for processing.

[0044] The preprocessed electroencephalogram signal is input into the emotion recognition neural network model (combined with BiLSTM and attention mechanism). The number of hidden units in the BiLSTM layer is set to 256, and the Xavier method is used for initialization. The attention mechanism calculates the scoring function e t = tanh(W e h t + b e ), and the attention weight The model is trained using the Adam optimization algorithm with adaptive learning rate adjustment. The initial learning rate is 0.002. If the loss function does not decrease for 2 consecutive batches every 3 epochs, the learning rate is multiplied by a decay factor of 0.6. The loss function is the multi-label cross-entropy loss function. The training data is added with individual identification information. After the BiLSTM layer, the principal component analysis algorithm is used for dimensionality reduction, and the principal components with a cumulative contribution rate of more than 90% are retained. Early stopping is used for training, and training stops when the loss function on the validation set does not decrease for 4 consecutive epochs. Taking the workers on the production line as an example, the model judges the emotional state according to their electroencephalogram, extracts the emotional feature vector, which includes the energy values of electroencephalograms in different frequency bands weighted by attention at different time points and the information of the dynamic proportional relationship of energy values in different frequency bands.

[0045] The emotional feature vector and the production line business data sequence adopt a fusion algorithm based on multi-scale feature matching. The decomposition layer number of the wavelet packet decomposition of the emotional feature vector is determined according to its dimension and complexity. Let the dimension be D, and the decomposition layer number obtains sub-feature vectors at different scales; the business data sequence uses sliding window Fourier transform. In the industrial field, the wavelet packet decomposition scale is set to 3 - 5 layers, and the window sliding step size is set to 8 - 12 data points. Calculate the similarity matrix between the emotional sub-feature vectors and the business sub-data sequences at different scales. The similarity metric is based on the method combining DTW and mutual information, sij =MI ij ×(1 - DTW ij ), Mutual information is calculated by constructing a joint probability distribution matrix The DTW distance is solved for the minimum path cost using the dynamic programming algorithm. The fusion weight vector is determined according to the similarity matrix, and the weighted sum gives the fusion feature vector. Then, an abnormal feature model is constructed using the support vector machine algorithm optimized by quantum annealing. The SVM kernel function is the polynomial kernel function K(x,y) = (x T y + c) d , and the Hamiltonian of the quantum annealing optimization algorithm is H = H p +H m . When constructing the problem Hamiltonian, a data uncertainty measure (calculating the entropy value of the fusion feature vector) is introduced, and the annealing schedule is adaptively adjusted. For example, an abnormal feature model is constructed by fusing workers' emotional features and production line data

[0046] The abnormal feature model is used for real-time data monitoring and analysis of the production line. After the new fused data is input, the model determines whether it is abnormal. The early warning strategy adopts a collaborative decision-making method based on a multi-agent system. The data monitoring agent adjusts the monitoring frequency according to the fluctuation characteristics of the production line data. Let the average change rate be V. When V ≥ V1 (the preset high-speed change threshold), the monitoring frequency is 8 times per second. The sentiment analysis agent uses the hidden Markov model to analyze the degree and trend of sentiment abnormalities, constructs a sentiment state transition matrix T, and trains according to historical sentiment data. The business risk assessment agent evaluates the degree of business risk using a risk scoring model based on the business data sequence, calculates the risk score S, and the early warning level determination agent uses a fuzzy inference system to determine the early warning level based on the analysis results of each agent. Taking the sentiment abnormality degree score E s and the business risk score S as inputs and the early warning level L as the output, each agent communicates through an information interaction protocol based on blockchain encryption communication technology. For example, when the wear amount data of key components of production line equipment is abnormal and the workers' emotions are unstable, the corresponding early warning level is triggered, and relevant personnel are notified by text message and pop-up window on the production management system

[0047] The above is only a preferred embodiment of the present invention, and it is not a limitation of the present invention in any form. Although the present invention has been disclosed as above with a preferred embodiment, it is not intended to limit the present invention. Any person skilled in the art can make some changes or modifications within the scope of the technical solution of the present invention. These equivalent embodiments, as long as they do not depart from the technical solution content of the present invention, any simple modifications, equivalent changes and modifications made to the above embodiments based on the technical essence of the present invention still fall within the scope of the technical solution of the present invention

Claims

1. An abnormal warning method based on artificial intelligence, characterized in that: The specific steps of this method are: S1. EEG data collection and preprocessing Using a multi-channel flexible EEG collection device, the common mode interference of the collected original EEG signals is first removed. The algorithm formula is: Where x(n) is the sample value of the original EEG signal at time n, y(n) is the output after filtering, and w i (n) is the i-th filter coefficient at time n, M is the filter order, and w i The initial value of (n) is estimated based on the prior spectrum characteristics of the EEG signal, and then the minimum mean square error criterion is used in the signal processing process through the iterative formula w i (n+1)=w i (n)+μe(n)x(ni), where μ is the step size factor, e(n)=d(n)-y(n), d(n) is the desired signal, and the signal after adaptive filtering is amplified. The initial intensity is I0, and the target intensity range is [I min , I max ], magnification Then, a method based on empirical mode decomposition (EMD) and independent component analysis (ICA) is used to extract brain wave signals of different frequency bands. The amplified signal is first decomposed by EMD to obtain several intrinsic mode functions (IMFs). Then, the IMF is analyzed by ICA to separate the independent components corresponding to different frequency bands. Specifically, the objective function of ICA is constructed. Where W is the separation matrix, s j and j is the separated signal component, corr represents the correlation function, N is the number of signal components, W is solved by the optimization algorithm to achieve frequency band separation, and the crossover probability P of the genetic algorithm is c The value range is 0.6-0.9, and the mutation probability P m The value range is 0.01-0.1; S2. Emotional state recognition and feature extraction The preprocessed EEG signal is input into the emotion recognition neural network model. The model structure is a combination of a bidirectional long short-term memory network BiLSTM and an attention mechanism. The number of hidden units in the BiLSTM layer is set to H. The weight matrix and bias vector of the forget gate, input gate, and output gate are initialized using the Xavier initialization method. The attention score calculation function of the attention mechanism is: in is the hidden state of BiLSTM at time t, W e and b e is a trainable parameter, attention weight T is the sequence length. The model training adopts the Adam optimization algorithm with adaptive learning rate adjustment, and its initial learning rate is set to L r During the training process, after every E epochs, if the loss function does not decrease for B consecutive batches, the learning rate is multiplied by the decay factor D, and the loss function uses the multi-label cross entropy loss function Where C is the number of emotion categories, M is the number of samples, and yi j The true value of the i-th sentiment label of the j-th sample, To predict the value, a large amount of EEG data labeled with different emotional states is used for training to obtain the emotional feature vector, which contains the energy value of each frequency band of EEG at different time points after attention weighting and the dynamic proportional relationship between the energy values ​​of different frequency bands. ij It represents the ratio of the attention-weighted energy value of the i-th frequency band at the j-th time point to the sum of the energy values ​​of all frequency bands at that time point, that is, Where E ij is the energy value of the i-th frequency band at the j-th time point, and K is the total number of frequency bands; S3. Collection and collation of specific business data Deeply connect with the business system and use specially developed data collection middleware to obtain specific business data. In the financial field, it covers the bid-ask spread sequence of stocks, bulk transaction data, and changes in margin trading balances. In the industrial field, it includes material flow fluctuations in each link of the production line, wear monitoring data of key equipment components, and data on the impact of environmental humidity on production processes. In the medical field, it involves specific gene expression data of patients at different disease stages, intermediate product concentration data in drug metabolism, and muscle electrical signal data in rehabilitation training. The collected business data is cleaned using a data cleaning algorithm based on density clustering, which calculates the local density of data points. where d i j is the distance between data points i and j, σ is the density parameter and relative density Identify and remove noise points in the data, and correct the outliers of the data. By comparing with the mean and standard deviation of the same type of data, if the data point exceeds the range of the mean plus or minus 3 times the standard deviation, it will be corrected according to the local trend of the data. Verify the accuracy of the data. Use a data integrity verification method based on a hash function to convert the business data into a hash value of a fixed length, compare it with the pre-stored correct hash value, and standardize the data. According to the statistical characteristics of the business data and the needs of subsequent analysis, use a quantile-based standardization method. Set q p is the p quantile of the data, then the standardized data Organize business data into sequence sets according to time series; S4. Data fusion and abnormal feature construction The emotional feature vector is fused with the specific business data sequence. The fusion algorithm based on multi-scale feature matching is adopted. First, the emotional feature vector and the business data sequence are decomposed at multiple scales. The emotional feature vector is decomposed by wavelet packet, and the business data sequence is decomposed by sliding window Fourier transform. For wavelet packet decomposition, the number of decomposition layers is determined according to the dimension and complexity of the emotional feature vector. Assume that the dimension of the emotional feature vector is D and the number of decomposition layers is Sub-feature vectors at different scales are obtained by wavelet packet decomposition, and business sub-data sequences at different frequency bands are obtained. Then, the similarity matrix S between the emotion sub-feature vectors and the business sub-data sequences at different scales is calculated. The similarity measurement adopts a method based on the combination of dynamic time warping DTW and mutual information, s ij =MI ij ×(1-DTW ij ), where MI ij is the mutual information between the i-th sentiment sub-feature vector and the j-th business sub-data sequence, DTW ij is the DTW distance between them, MI ij The calculation is done by constructing the joint probability distribution matrix P, and then according to the formula Calculation, DTW ij The calculation of the minimum path cost is solved by the dynamic programming algorithm, and the fusion weight vector W is determined according to the similarity matrix. Where I is the number of emotional sub-feature vectors, J is the number of business sub-data sequences, and finally the fusion feature vector is obtained by weighted summation. , where e i is the i-th sentiment sub-feature vector, b j is the jth business sub-data sequence. On this basis, the support vector machine algorithm based on quantum annealing optimization is used to construct abnormal features for the fused feature vector. The kernel function of SVM adopts the polynomial kernel function K(x,y)=(x T y+c) d , where c is a constant term, d is the polynomial degree, and the Hamiltonian H in the quantum annealing optimization algorithm is H p +H m , where H p is the Hamiltonian of the problem, H m is the transverse field Hamiltonian, and the optimal parameter combination of SVM is found through quantum annealing process to determine the abnormal feature model; S5. Abnormal warning decision and notification The constructed abnormal feature model is applied to real-time data monitoring and analysis. When new fused data is input into the model, the model makes judgments based on the abnormal feature rules learned in advance. If it is detected that the current data meets the abnormal feature pattern, the abnormal warning decision is triggered immediately. The warning decision system determines the warning level, warning method and warning recipients based on the pre-set warning strategy, and uses a variety of communication channels and terminal devices to convey the warning information to relevant personnel in a timely and accurate manner.

2. The abnormal warning method based on artificial intelligence according to claim 1 is characterized in that: In the step of collecting and preprocessing brain wave data, the electrode layout of the multi-channel flexible brain wave collection device adopts a non-uniform distribution mode. According to the difference in activity of brain functional areas, electrodes with higher density are set in the emotion-related brain areas of the frontal lobe and temporal lobe, and the electrode spacing is between 1mm and 3mm. In the parietal lobe and occipital lobe areas, the electrode spacing is appropriately increased to 5mm-10mm. In the adaptive filtering algorithm, the estimation model of the common mode interference signal is pre-trained using a signal generation model based on deep learning. The model uses a large number of known interference signals and non-interference brain wave signals as training data to learn the characteristic patterns of the interference signals, so as to more accurately determine the initial filter coefficient w i (n).

3. The abnormal warning method based on artificial intelligence according to claim 1 is characterized in that: In the emotion state recognition and feature extraction steps, in the model combining the bidirectional long short-term memory network and the attention mechanism, a feature compression layer is added after the BiLSTM layer, and the output of the BiLSTM is reduced in dimension using the principal component analysis algorithm, retaining the principal components with a cumulative contribution rate of more than 90%. At the same time, in the model training process, the early stopping method is used to prevent overfitting, and the loss function of the validation set is set to be continuous E s The training is stopped when the probability of the emotion decreases after epochs, and the model parameters are saved. In addition, in order to improve the adaptability of the model to the differences in brain wave characteristics of different individuals, individual identification information is added to the training data and input into the model as an additional feature, so that the model can learn the subtle differences in brain wave characteristics of different individuals under the same emotional state, thereby more accurately identifying the emotional state of the individual and extracting effective emotional feature vectors.

4. The abnormal warning method based on artificial intelligence according to claim 1 is characterized in that: In the specific business data collection and sorting step, the data collection middleware has an intelligent data caching function, which dynamically adjusts the cache size according to the real-time traffic and processing speed of the business data. When the business data traffic suddenly increases, the cache size is automatically increased to avoid data loss. The adjustment of the cache size is calculated based on the traffic change rate and the current cache occupancy rate. Assuming the traffic change rate is r, the current cache occupancy rate is o, and the cache adjustment coefficient is k, the new cache size C new =C old (1+k×r×(1-o)), where C old is the original cache size, and in the data cleaning algorithm based on density clustering, for the local density calculation of data points, a time decay factor is introduced, that is, Where λ is the time decay coefficient, t i and t j The timestamps of data points i and j are used to make recent data points have higher weights in cluster analysis, which is more in line with the dynamic characteristics of business data.

5. The abnormal warning method based on artificial intelligence according to claim 1 is characterized in that: In the data fusion and abnormal feature construction steps, the fusion algorithm based on multi-scale feature matching adopts adaptive decomposition scale and window sliding step size for different business types and emotional feature complexity when performing wavelet packet decomposition and sliding window Fourier transform. In the financial field, due to the rapid and complex changes in data, the wavelet packet decomposition scale is set to L f -L g The sliding step size of the sliding window Fourier transform is set to S1-S2 data points. In the industrial field, the data is relatively stable, and the wavelet packet decomposition scale is L □ ·L i Layer, the window sliding step is S3-S4 data points. In the medical field, it is adjusted according to the acuteness or chronicity of the disease. The data related to acute diseases adopts a higher decomposition scale L j -L k The sliding step size of the layer and smaller window is S5-S6 data points, while the opposite is true for chronic disease related data.

6. The abnormal warning method based on artificial intelligence according to claim 1 is characterized in that: In the data fusion and abnormal feature construction steps, in the support vector machine algorithm based on quantum annealing optimization, when constructing the problem Hamiltonian H of quantum annealing, p In addition to considering the objective function of SVM, the uncertainty measurement of data is also introduced. The uncertainty of the fused feature vector is set as U, which is measured by calculating the entropy value of the feature vector, that is, U = -∑ i p i logp i , where p i is the probability distribution of the i-th element in the feature vector, adding the uncertainty term to H p In this way, the optimization process can take into account the uncertainty of the data. At the same time, the annealing schedule in the quantum annealing process adopts an adaptive adjustment strategy, which is adjusted according to the quality of the currently searched solution and the search progress. If the quality of the solution does not improve in several consecutive iterations, the annealing speed is accelerated.

7. The abnormal warning method based on artificial intelligence according to claim 1 is characterized in that: In the abnormal warning decision-making and notification step, the setting of the warning strategy adopts a collaborative decision-making method based on a multi-agent system. Multiple agents are established, including a data monitoring agent, an emotion analysis agent, a business risk assessment agent, and a warning level determination agent. The data monitoring agent is responsible for real-time monitoring of changes in the fusion data. Its monitoring frequency is dynamically adjusted according to the fluctuation characteristics of the data. Let the average change rate of the data be V. When V≥V1, where V1 is a preset high-speed change threshold, the monitoring frequency is F5 times per second; when V1>V≥V2, where V2 is a medium-speed change threshold, the monitoring frequency is F6 times per second; when V<V2, the monitoring frequency is F7 times per second. The emotion analysis agent conducts in-depth analysis of the emotion feature vector to judge the degree and trend of emotional abnormality. Its analysis model adopts emotion transfer analysis based on a hidden Markov model. By constructing an emotion state transition matrix T, where the element t ij represents the probability of transitioning from emotion state i to emotion state j. T is obtained by training with a large amount of historical emotion data, and then the emotion change trend is predicted. The business risk assessment agent evaluates the degree of business risk based on the business data sequence and adopts a risk scoring model. This model calculates the risk score S according to the deviation degree between the business data and the preset risk indicators. The warning level determination agent determines the warning level by synthesizing the analysis results of each agent and adopts a fuzzy inference system. Using the emotional abnormality degree score E s and the business risk score S as input variables and the warning level L as the output variable, by establishing a fuzzy rule base, each agent communicates through an information interaction protocol, and this protocol adopts blockchain-based encrypted communication technology.

8. An abnormal warning system based on artificial intelligence, characterized in that: The system includes brain wave data collection and preprocessing module, emotional state recognition and feature extraction module, specific business data collection and sorting module, data fusion and abnormal feature construction module, and abnormal warning decision and notification module: The EEG data collection and preprocessing module uses a multi-channel flexible EEG collection device to remove common mode interference from the collected original EEG signals. The algorithm formula is: Where x(n) is the sample value of the original EEG signal at time n, y(n) is the output after filtering, and w i (n) is the i-th filter coefficient at time n, M is the filter order, and w i The initial value of (n) is estimated based on the prior spectrum characteristics of the EEG signal, and then the minimum mean square error criterion is used in the signal processing process through the iterative formula w i (n+1)=w i (n)+(μe(n)x(ni), where μ is the step size factor, e(n)=d(n)-y(n), d(n) is the desired signal, the signal after adaptive filtering is amplified, the initial intensity is I0, and the target intensity range is [I min , I max ], magnification Then, a method based on empirical mode decomposition (EMD) and independent component analysis (ICA) is used to extract brain wave signals of different frequency bands. The amplified signal is first decomposed by EMD to obtain several intrinsic mode functions (IMFs). Then, the IMF is analyzed by ICA to separate the independent components corresponding to different frequency bands. Specifically, the objective function of ICA is constructed. Where W is the separation matrix, s i and j is the separated signal component, corr represents the correlation function, N is the number of signal components, W is solved by the optimization algorithm to achieve frequency band separation, and the crossover probability P of the genetic algorithm is c The value range is 0.6-0.9, and the mutation probability P m The value range is 0.01-0.1; The emotional state recognition and feature extraction module: inputs the preprocessed brain wave signal into the emotion recognition neural network model. The model structure is a combination of a bidirectional long short-term memory network BiLSTM and an attention mechanism. The number of hidden units in the BiLSTM layer is set to H. The weight matrix and bias vector of the forget gate, input gate and output gate are initialized using the Xavier initialization method. The attention score calculation function of the attention mechanism is: in is the hidden state of BiLSTM at time t, W e and b e is a trainable parameter, attention weight T is the sequence length. The model training adopts the Adam optimization algorithm with adaptive learning rate adjustment, and its initial learning rate is set to L r During the training process, after every E epochs, if the loss function does not decrease for B consecutive batches, the learning rate is multiplied by the decay factor D, and the loss function uses the multi-label cross entropy loss function Where C is the number of emotion categories, M is the number of samples, and y ij The true value of the i-th sentiment label of the j-th sample, To predict the value, a large amount of EEG data labeled with different emotional states is used for training to obtain the emotional feature vector, which contains the energy value of each frequency band of EEG at different time points after attention weighting and the dynamic proportional relationship between the energy values ​​of different frequency bands. ij It represents the ratio of the attention-weighted energy value of the i-th frequency band at the j-th time point to the sum of the energy values ​​of all frequency bands at that time point, that is, Where E ij is the energy value of the i-th frequency band at the j-th time point, and K is the total number of frequency bands: The specific business data collection and sorting module is deeply connected with the business system, and obtains specific business data with the help of specially developed data collection middleware. In the financial field, it covers the stock bid-ask spread sequence, bulk transaction data, and changes in margin trading balances. In the industrial field, it includes material flow fluctuations in each link of the production line, wear monitoring data of key equipment components, and data on the impact of environmental humidity on production processes. In the medical field, it involves specific gene expression data of patients at different disease stages, intermediate product concentration data in drug metabolism, and muscle electrical signal data in rehabilitation training. The collected business data is cleaned by using a data cleaning algorithm based on density clustering to calculate the local density of data points. where d i j is the distance between data points i and j, σ is the density parameter and relative density Identify and remove noise points in the data, and correct the outliers of the data. By comparing with the mean and standard deviation of the same type of data, if the data point exceeds the range of the mean plus or minus 3 times the standard deviation, it will be corrected according to the local trend of the data. Verify the accuracy of the data. Use a data integrity verification method based on a hash function to convert the business data into a hash value of a fixed length, compare it with the pre-stored correct hash value, and standardize the data. According to the statistical characteristics of the business data and the needs of subsequent analysis, use a quantile-based standardization method. Set q p is the p quantile of the data, then the standardized data Organize business data into sequence sets according to time series; The data fusion and abnormal feature construction module: the emotional feature vector is fused with the specific business data sequence, and a fusion algorithm based on multi-scale feature matching is adopted. First, the emotional feature vector and the business data sequence are respectively decomposed at multiple scales. The emotional feature vector is decomposed by wavelet packet, and the business data sequence is decomposed by sliding window Fourier transform. For wavelet packet decomposition, the number of decomposition layers is determined according to the dimension and complexity of the emotional feature vector. Assume that the dimension of the emotional feature vector is D, and the number of decomposition layers is Sub-feature vectors at different scales are obtained by wavelet packet decomposition, and business sub-data sequences at different frequency bands are obtained. Then, the similarity matrix S between the emotion sub-feature vectors and the business sub-data sequences at different scales is calculated. The similarity measurement adopts a method based on the combination of dynamic time warping DTW and mutual information, s ij =MI ij ×(1-DTW ij ), where MI ij is the mutual information between the i-th sentiment sub-feature vector and the j-th business sub-data sequence, DTW ij is the DTW distance between them, MI ij The calculation is done by constructing the joint probability distribution matrix P, and then according to the formula Calculation, DTW ij The calculation of the minimum path cost is solved by the dynamic programming algorithm, and the fusion weight vector W is determined according to the similarity matrix. Where I is the number of emotional sub-feature vectors, J is the number of business sub-data sequences, and finally the fusion feature vector is obtained by weighted summation. where e i is the i-th sentiment sub-feature vector, b j is the jth business sub-data sequence. On this basis, the support vector machine algorithm based on quantum annealing optimization is used to construct abnormal features for the fused feature vector. The kernel function of SVM adopts the polynomial kernel function K(x, y) = (x T y+c) d , where c is a constant term, d is the polynomial degree, and the Hamiltonian H in the quantum annealing optimization algorithm is H p +H m , where H p is the Hamiltonian of the problem, H m is the transverse field Hamiltonian, and the optimal parameter combination of SVM is found through quantum annealing process to determine the abnormal feature model; The abnormal warning decision and notification module applies the constructed abnormal feature model to real-time data monitoring and analysis. When new fused data is input into the model, the model makes a judgment based on the abnormal feature rules learned in advance. If it is detected that the current data meets the abnormal feature pattern, the abnormal warning decision is triggered immediately. The warning decision system determines the warning level, warning method and warning recipients based on the pre-set warning strategy, and uses a variety of communication channels and terminal devices to convey the warning information to relevant personnel in a timely and accurate manner.

Images