Intelligent monitoring system for safe construction of spherical tank

Abstract

The invention relates to the field of industrial safety monitoring, and discloses an intelligent monitoring system for spherical tank safety construction, which comprises a multi-mode sensor unit, a data processing unit, a mode library unit, a self-adaptive threshold value unit, a grading alarm unit and a wireless transmission unit, the multi-mode sensor unit is integrated with sensing assemblies for multiple physical quantities such as vibration, temperature and stress; the data processing unit constructs a dynamic incidence matrix, quantifies nonlinear correlation among parameters, and extracts a maximum characteristic value to represent a system state; the mode library unit realizes self-evolution management of abnormal modes; the self-adaptive threshold value unit dynamically optimizes an alarm threshold value in combination with reinforcement learning and construction stage recognition; the grading alarm unit is used for implementing multi-dimensional composite criterion to trigger differential response; the wireless transmission unit guarantees the reliability of data transmission through a dual-mode communication architecture and an anti-interference protocol. According to the invention, the recognition precision and response efficiency of the composite risk in the construction process are obviously improved.

Description

Technical Field

[0001] The present invention relates to the technical field of industrial safety monitoring, and particularly to an intelligent monitoring system for the safe construction of spherical tanks. Background Art

[0002] In the field of pressure vessel construction, especially during the construction of large spherical tanks, the safety monitoring technology under complex working conditions faces severe challenges. Traditional monitoring systems mostly rely on a single physical quantity threshold alarm mechanism, and it is difficult to effectively capture the compound risks caused by the coupling effect of multiple parameters, such as the latent faults like the microcrack propagation caused by the synergistic effect of welding thermal stress and structural vibration.

[0003] Existing methods lack the adaptability to the dynamic characteristics of the construction stage at the data processing level. The fixed time window and static correlation model cannot accurately reflect the time-varying law of the parameter correlation relationship in different processes (civil engineering, welding, water filling test), resulting in distorted feature extraction and an increased misjudgment rate. In addition, the anomaly recognition module mostly adopts a predefined pattern matching strategy, and there are cognitive blind spots when facing new risk patterns emerging at the construction site, and the historical data experience is difficult to continuously accumulate through an online learning mechanism.

[0004] The above technical defects jointly restrict the reliability and intelligent level of the construction safety monitoring system, and there is an urgent need to construct a monitoring system that integrates multi-source perception, dynamic modeling, and autonomous evolution. Summary of the Invention

[0005] Aiming at the deficiencies of the existing technology, the present invention provides an intelligent monitoring system for the safe construction of spherical tanks, which solves the problems of the existing construction monitoring system in terms of compound risk identification, dynamic working condition adaptation, and communication reliability.

[0006] To achieve the above objectives, the present invention is realized through the following technical solutions: An intelligent monitoring system for the safe construction of spherical tanks, comprising:

[0007] A multi-modal sensor unit for real-time collecting vibration, temperature, stress, gas concentration, and corrosion data during the construction stage;

[0008] A data processing unit connected to the sensor unit for preprocessing the collected data and constructing a multi-parameter dynamic correlation matrix, and outputting matrix eigenvalues;

[0009] A pattern library unit connected to the data processing unit for storing historical abnormal feature patterns and providing real-time pattern matching;

[0010] An adaptive threshold unit connected to the data processing unit and the pattern library unit respectively for dynamically optimizing the alarm threshold according to the change trend of the eigenvalues and the pattern matching result;

[0011] The hierarchical alarm unit receives the output of the data processing unit and the threshold of the adaptive threshold unit, and triggers multi-level alarms based on the comprehensive determination of eigenvalue overrun and pattern matching;

[0012] The wireless transmission unit is connected to the hierarchical alarm unit and sends the alarm signal to the monitoring terminal.

[0013] Preferably, when the data processing unit constructs the multi-parameter dynamic correlation matrix:

[0014] Calculate the mutual information coupling degree of any two sensor data as the matrix element, and introduce the historical coupling degree attenuation factor to realize the dynamic update of the matrix element.

[0015] Preferably, when calculating the mutual information coupling degree:

[0016] Adopt the sliding window joint probability distribution estimation method, and the window length is dynamically adjusted according to the current construction stage.

[0017] Preferably, the adaptive threshold unit adopts a reinforcement learning algorithm, and generates a threshold adjustment action by constructing a state space including the eigenvalue change rate and the construction stage encoding.

[0018] Preferably, the state space includes:

[0019] The first derivative and the second derivative of the maximum eigenvalue, and the construction stage vector generated by one-hot encoding.

[0020] Preferably, the hierarchical alarm unit executes:

[0021] Trigger a level I alarm when the maximum eigenvalue exceeds the dynamic threshold and at least two sensor data exceed the statistical control limit;

[0022] Trigger a level II alarm when more than half of the eigenvalues reach the threshold critical area;

[0023] Trigger a level III early warning when the current mode matches the historical abnormal mode library successfully.

[0024] Preferably, the pattern library unit extracts the time-varying features of the dynamic correlation matrix through a convolutional neural network to generate a pattern fingerprint, and automatically expands the pattern library when the similarity between the new feature and the inventory pattern is lower than the preset threshold.

[0025] Preferably, when generating the pattern fingerprint:

[0026] Vertically splice the dynamic correlation matrices of three consecutive time windows, and use three-layer convolutional kernels with different scales to extract spatio-temporal features of the spliced matrix.

[0027] Preferably, the wireless transmission unit adopts the LoRa and Bluetooth Mesh dual-mode communication protocols, and automatically selects the transmission channel and transmission power according to the alarm level.

[0028] Preferably, the system further includes:

[0029] A construction stage identification unit that automatically identifies the current construction stage by analyzing the main frequency components of the vibration signal and the temperature change rate curve.

[0030] The present invention provides an intelligent monitoring system for the safe construction of spherical tanks. It has the following beneficial effects:

[0031] 1. Through the heterogeneous perception of the multi-modal sensor unit and the construction of the dynamic correlation matrix, the present invention realizes the collaborative monitoring of multiple physical quantities during the construction process. The calculation of the mutual information coupling degree of multi-source data such as vibration, temperature, and stress reveals the non-linear correlation characteristics between parameters. The introduction of KL divergence quantization and historical decay factor ensures the stability of the dynamic update of the matrix, overcomes the limitations of traditional single-parameter threshold monitoring, and significantly improves the recognition ability of composite risks under complex working conditions.

[0032] 2. Through the spatio-temporal feature fusion and self-evolution mechanism of the pattern library unit, the present invention enhances the adaptability of the system to unknown abnormal patterns. The spatio-temporal features of the dynamic correlation matrix extracted by the multi-scale convolutional neural network generate pattern fingerprints, and the incremental clustering algorithm realizes the autonomous expansion of the pattern library. Cooperating with the forgetting factor to eliminate stale data enables the system to continuously accumulate construction safety experience knowledge and effectively respond to the challenges of new safety hazards.

[0033] 3. The present invention adopts an adaptive threshold optimization strategy based on reinforcement learning to improve the intelligent level of alarm decision-making. By constructing a multi-dimensional state space including the derivative term of the eigenvalue and the construction stage code, and combining the TD3 algorithm to generate dynamic threshold adjustment actions, the alarm threshold can be autonomously optimized following the construction process and environmental changes, achieving a dynamic balance between false alarm suppression and missed alarm prevention.

[0034] 4. Through the multi-dimensional determination logic and priority scheduling mechanism of the hierarchical alarm unit, the present invention realizes the precise classification of risk responses. Integrating the composite determination conditions of eigenvalue overrun, statistical deviation degree of sensor parameters, and historical pattern matching degree, and cooperating with event merging and feedback parameter adjustment, a multi-granularity alarm trigger system is formed to ensure that risk events of different levels are appropriately handled and optimize the emergency resource scheduling efficiency. Description of the Drawings

[0035] Figure 1 It is a schematic diagram of the system architecture of the present invention. Detailed Embodiments

[0036] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings of the present invention. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all the embodiments. All other embodiments obtained by those of ordinary skill in the art based on the embodiments of the present invention without creative efforts shall fall within the protection scope of the present invention.

[0037] Please refer to the attached Figure 1 , the present invention provides an intelligent monitoring system for the safe construction of spherical tanks. By combining the construction of a multi-parameter dynamic correlation matrix with the optimization of reinforcement learning thresholds, it realizes the accurate identification and hierarchical early warning of compound abnormal risks during the construction process.

[0038] The system includes a multi-modal sensor unit, a data processing unit, a pattern library unit, an adaptive threshold unit, a hierarchical alarm unit, and a wireless transmission unit. The data interaction relationships among the modules are as Figure 1 shown. After receiving the sensor data, the data processing unit generates a dynamic correlation matrix. The matrix eigenvalues are input into the adaptive threshold unit for dynamic optimization. At the same time, the pattern library unit provides historical pattern matching results. Finally, the hierarchical alarm unit comprehensively determines the alarm level and sends it to the monitoring terminal through the wireless transmission unit.

[0039] The following is a detailed description of each component in the system of the present invention.

[0040] For the multi-modal sensor unit, in this embodiment, a heterogeneous sensing device collaborative deployment scheme is adopted. Through multi-dimensional physical quantity acquisition and data fusion processing, a monitoring network covering all elements of spherical tank construction is constructed. Specifically, this unit includes a vibration sensing component, a temperature sensing component, a stress sensing component, a gas sensing component, and a corrosion sensing component. The data acquisition time domain alignment among the components is realized through a unified clock synchronization module to ensure the timing consistency of the subsequent dynamic correlation matrix construction.

[0041] The vibration sensing component preferably adopts a triaxial accelerometer array, which is installed at the load-bearing support of the equatorial belt of the spherical tank. Its measurement range covers the frequency range from 0.1 Hz to 500 Hz, and it can synchronously capture mechanical excitation signals in different frequency bands such as construction machinery vibration and weld crack propagation. For the high-frequency impact interference in spherical tank construction, a band-pass filter is configured to filter out the noise in non-critical frequency bands above 50 Hz and retain the low-frequency vibration characteristics related to structural safety.

[0042] The temperature sensing component is composed of an infrared thermocouple array and a distributed optical fiber temperature measurement device. Among them, the infrared thermocouples cover the welding operation area in a checkerboard layout to monitor the temperature field distribution of the molten pool in real time; the distributed optical fiber is arranged along the circumferential weld of the spherical tank, and the axial temperature gradient data is obtained by using the Raman scattering principle. After the data of the two are fused, a temperature change rate curve is generated, which provides key input for the identification of the construction stage.

[0043] The stress sensing component is implemented based on fiber Bragg grating technology. An FBG sensor array is cross - arranged on the inner and outer wall surfaces of the spherical tank, and the circumferential stress and axial stress components of each monitoring point are measured in real - time through a wavelength demodulation device. Further, a dual - optical - path compensation mechanism is adopted to eliminate the cross - sensitivity effect of environmental temperature on stress measurement, ensuring the accuracy of stress data.

[0044] The gas sensing component is deployed in potential leakage areas such as the manhole and sewage outlet of the spherical tank. An integrated electro - chemical and infrared absorption composite sensor is used to synchronously detect the volume concentrations of gases such as H2S, O2, and CH4. A cyclone separation type pretreatment device is configured to solve the problem of dust interference in the construction environment. Particles with a particle size greater than 5μm are removed through the inertial separation principle to ensure the smoothness of the sensor gas path.

[0045] The corrosion sensing component adopts a cooperative working mode of a resistance probe and an ultrasonic thickness gauge. The resistance probe is embedded in the surface of the spherical tank base metal specimen, and the corrosion rate is inverted through the resistance change caused by metal loss. The ultrasonic thickness gauge periodically scans the wall thickness of key parts of the tank body. The data of both are fused through Kalman filtering and then the corrosion state evaluation result is output.

[0046] Preferably, the clock synchronization module adopts the IEEE 1588 precise time protocol to achieve micro - second - level synchronization accuracy of each sensor node through a master - slave clock architecture. For wireless transmission nodes, a forward error correction coding mechanism is additionally introduced to compensate for the uncertainty of wireless channel transmission delay.

[0047] For the data processing unit, in this embodiment, its fusion analysis and feature extraction of multi - source heterogeneous data are carried out through a dynamic association modeling and pattern evolution mechanism, converting the original sensor data into an interpretable safety - state characterization quantity. Specifically, this unit includes a data pre - processing module, a dynamic association matrix generation module, and a feature analysis module. A cascade processing pipeline is formed among the modules to realize the mining of the essential characteristics of the multi - physical - quantity coupling relationship during the construction process.

[0048] The data pre - processing module first performs a sliding - window normalization operation to eliminate the data scale difference caused by the construction stage transition. For the original data stream x i (t) of each sensor channel, the normalization window length W is dynamically selected according to the real - time identified construction stage s∈{civil engineering, welding, water filling experiment} s . Specifically, in implementation, by calculating the mean value μ i (W s ) and the standard deviation σ i (W s ) within the window, the original data is standardized into a dimensionless form:

[0049]

[0050] This operation effectively suppresses the baseline drift caused by environmental factors while retaining the relative amplitude characteristics of data fluctuations. Preferably, for the high-frequency transient interference during the welding stage, a limiting filtering link is added, and a dynamic threshold K clip =β·σ i (W s ) is set, where β is an adjustable proportionality coefficient, and saturation processing is performed on the abnormal points outside the threshold range.

[0051] After completing data standardization, the dynamic correlation matrix generation module starts the multi-parameter coupling degree calculation process. During the construction of the dynamic correlation matrix, the mutual information coupling degree between any two sensor data streams and is calculated using the sliding window joint probability distribution estimation method. Define the observation data set within the time window [t - ΔT, t], and obtain the joint distribution and the marginal distribution through histogram statistics, and then calculate the KL divergence to characterize the coupling strength:

[0052]

[0053] In the formula, α ∈ (0, 1) is the historical coupling degree attenuation factor, which is used to smooth the time series mutation of matrix elements. This calculation method can dynamically capture the non-linear interaction patterns between multiple physical quantities during the construction process and overcome the limitations of traditional linear correlation coefficients.

[0054] Based on the above coupling degree calculation results, a symmetric dynamic correlation matrix is constructed, where the matrix element m ij (t) = φ ij (t). Preferably, a negative definite correction term is introduced to ensure the positive semi-definiteness of the matrix:

[0055] M′(t) = M(t) + δ·I

[0056] where δ is the minimum eigenvalue compensation amount and I is the identity matrix. This correction process ensures the numerical stability of subsequent eigenvalue decomposition.

[0057] The eigenvalue analysis module performs spectral analysis on the dynamic correlation matrix and extracts the maximum eigenvalue λ max (t) as the system state sensitive index. Solve the eigenvalue equation through the Jacobi iteration method:

[0058] M'(t)v k =λ k v k , k = 1, 2,..., n

[0059] Select λ max (t) = max{λ1, λ2,..., λ n} Input to the threshold optimization module. Further, calculate the eigenvalue change rate parameter:

[0060]

[0061] This differential term is used to characterize the mutation trend of the system state and provides a dynamic basis for adaptive threshold adjustment.

[0062] In the mode feature extraction stage, the data processing unit vertically concatenates the current matrix M(t) and the historical matrix M(t - Δt) to generate a spatio-temporal extended matrix M ext (t) = [M(t); M(t - Δt)]. Deep features are extracted through a convolutional neural network:

[0063] F(T) = ReLU(W2 * ReLU(W1 * M ext (t) + b1) + b2)

[0064] where W1, W2 are convolutional kernel weight matrices, and b1, b2 are bias terms. The obtained feature vector F(t) is input to the pattern library unit for similarity matching and pattern update.

[0065] Preferably, the convolutional neural network adopts a multi-scale fusion architecture. The size of the first-layer convolutional kernel covers short-term correlation patterns, and the second-layer convolutional kernel captures medium- and long-term change trends. During the training process, an unsupervised autoencoder structure is adopted, and the loss function is defined as:

[0066]

[0067] where E(·) is the encoder, D(·) is the decoder, and ||·|| F represents the Frobenius norm. This training strategy enables the network to autonomously learn the essential feature representation of matrix data.

[0068] For the pattern library unit, in this embodiment, it realizes the representation learning and dynamic update of abnormal patterns through spatio-temporal feature fusion and self-evolution mechanism, providing a historical experience knowledge base for composite abnormal recognition. Specifically, this unit includes a pattern feature extraction module, a similarity matching module, and a pattern library self-expansion module, forming a closed-loop learning architecture from feature generation to knowledge accumulation, ensuring that the system can adapt to changes in the construction environment and improve the abnormal recognition ability.

[0069] The pattern feature extraction module receives the dynamic correlation matrix sequence output by the data processing unit where K is the number of time windows. A spatio-temporal extended matrix M is constructed through vertical concatenation extI(t) = [M(t); M(t - Δt); M(t - 2Δt)], this operation encodes the temporal correlation information into spatial dimension features, providing a structured input for subsequent convolutional processing. Preferably, the splicing dimension is dynamically adjusted according to the construction stage. Two-window splicing is adopted in the welding stage, and three-window splicing is adopted in the water filling test stage to adapt to the monitoring requirements of different processes.

[0070] On this basis, a multi-layer convolutional neural network is used to abstract the features of the spatio-temporal expansion matrix. The first-layer convolutional kernel slides along the time dimension to extract short-term correlation patterns; the second-layer convolutional kernel captures medium-term change trends; the third-layer convolutional kernel focuses on long-term stable features. After each layer of convolution, a ReLU activation function and a max-pooling operation are connected, and finally a dimensionality-reduced feature vector is output as the pattern fingerprint of the current state. The specific calculation process is expressed as:

[0071] F(t) = MP(ReLU(W3 * MP(ReLU(W2 * MP(ReLU(W1 * M ext (t)))))))

[0072] where * represents the convolution operation, and MP(·) is the max-pooling function. This hierarchical feature extraction scheme can effectively decouple the essential differences between normal working conditions and abnormal states.

[0073] The similarity matching module calculates the cosine similarity between the current pattern fingerprint F(t) and the historical feature vectors in the pattern library:

[0074]

[0075] Select the maximum similarity Sim max (t) = max{Sim1(t),..., Sim M (t)} as the pattern matching index. When Sim max (t) is lower than the preset threshold θ sim , it is determined that there is a significant difference between the current state and the historical pattern library, and the pattern library self-expansion process is triggered.

[0076] The pattern library self-expansion module updates the feature library using an incremental learning strategy. For the unmatched new pattern F(t), first calculate its Euclidean distance from the center of the existing pattern clusters:

[0077] d k = ||F(t) - C k ||2, k = 1, 2,..., K

[0078] If the minimum distance d min > θ cluster , a new cluster is created and F(t)F(t) is used as its center; otherwise, F(t) is assigned to the nearest neighbor cluster and the center is updated:

[0079]

[0080] where N k is the number of members of the original cluster. This mechanism realizes the adaptive expansion of the pattern library capacity while avoiding the excessive accumulation of redundant features.

[0081] Preferably, the pattern library unit introduces a forgetting factor mechanism to optimize the storage efficiency. For each cluster center C k maintains an access frequency counter c k , and when c k has not increased for T obs cycles, this cluster is determined to be a low-frequency invalid pattern and is removed. This design ensures that the pattern library dynamically reflects the latest state characteristics of the construction environment and eliminates the interference of historical stale data.

[0082] In the model training stage, an unsupervised autoencoder is used to pre-train the convolutional neural network. Define the reconstruction loss function:

[0083]

[0084] where E(·) is the encoder, D(·) is the decoder, and ||·|| F represents the Frobenius norm. By minimizing the reconstruction error, the network is forced to learn the essential low-dimensional representation of the spatio-temporal matrix. Further introduce the contrast loss:

[0085]

[0086] where τ is the temperature coefficient, F i and F j are the feature of the same kind of samples, and this loss function promotes the close aggregation of the same kind of patterns in the feature space.

[0087] For the adaptive threshold unit, in this embodiment, it realizes the dynamic optimization of the alarm threshold through a reinforcement learning framework, constructs a closed-loop control mechanism of state perception and policy iteration, and effectively balances the sensitivity of anomaly detection and the requirement of false alarm suppression. The core technical solution of this unit is to fuse the evolution law of the dynamic correlation matrix features with the context information of the construction stage to generate a threshold adjustment strategy with environmental adaptability.

[0088] The adaptive threshold unit first constructs a multi-dimensional state space to characterize the dynamic characteristics of the system. Define the state vector where:

[0089] λ max (t) is the maximum eigenvalue of the dynamic correlation matrix at the current moment;

[0090] Calculate the first derivative by the five-point central difference method;

[0091] Estimate the second derivative using the Savitzky-Golay filter;

[0092] e s (t) is the construction stage coding vector, and the one-hot coding is used to map the civil engineering, welding, and water supply experiment stages;

[0093] Sim(t) is the maximum historical pattern matching similarity output by the pattern library unit.

[0094] This state space design fully captures the time-domain evolution characteristics of the system's safe state and the context information of the working conditions, providing interpretable input features for the reinforcement learning strategy. Preferably, a time decay factor is introduced into the construction stage coding to perform a smooth transition process on the coding vector during the stage switching process.

[0095] On this basis, the Twin Delayed Deep Deterministic Policy Gradient (TD3) algorithm is used to generate threshold adjustment actions. An Actor-Critic network architecture is constructed, where the Actor network π φ (s t ) outputs a continuous action a t ∈[-1,1], and the Critic network Q θ (s t ,a t ) evaluates the state-action value function. The specific policy update process follows the following optimization objectives:

[0096]

[0097] Among them, is the experience replay buffer, which stores historical state transition samples (s t ,a t ,r t ,s t+1 ). The Critic network updates the parameters by minimizing the temporal difference error:

[0098]

[0099] In the formula, γ is the discount factor, θ' and φ' are the target network parameters, and the algorithm stability is improved through the delayed update mechanism.

[0100] The reward function r t is designed to comprehensively consider the requirements of false alarm suppression, missed alarm penalty, and policy smoothness:

[0101]

[0102] Wherein:

[0103] is a false alarm event indication function (false alarm triggered by manual review and confirmation);

[0104] is a missed alarm event indication function (failure to alarm in time during retrospective investigation of accidents after the event);

[0105] |a t |The penalty threshold fluctuates violently;

[0106] Sim(t) encourages a conservative strategy when matching historical abnormal patterns.

[0107] This reward mechanism guides the agent to learn a threshold adjustment strategy with safety priority through multi-objective trade-off. Preferably, the prioritized experience replay technique is introduced to assign a higher sampling probability to the samples containing to accelerate the convergence of the key strategy.

[0108] Based on the action a output by the policy network t , perform a threshold update operation:

[0109]

[0110] In the formula, γ is the learning rate parameter, and the tanh(·) function limits the action within a reasonable range. acts as an acceleration factor to enhance the response sensitivity to feature mutations. Preferably, when , freeze the threshold update to avoid over-adjustment when historical accident patterns reappear.

[0111] To ensure the rationality of the threshold during the system startup phase, the statistical process control method is used to set the benchmark threshold in the initialization phase:

[0112] Γ0 = μ λ + 3σ λ

[0113] Wherein, μ λ and σ λ are the mean and standard deviation of λ max (t) under historical normal operating conditions. This initial value is dynamically calibrated through the historical data stored in the pattern library unit to ensure the detection reliability during the cold start phase.

[0114] For the hierarchical alarm unit, in this embodiment, it realizes the hierarchical response to abnormal risks through a multi-dimensional criterion fusion and dynamic priority scheduling mechanism, and constructs a closed-loop decision-making link from feature analysis to emergency response. The core technical solution of this unit lies in comprehensively integrating the eigenvalues of the dynamic association matrix, the deviation degree of sensor parameters, and the historical pattern matching results, and implementing a multi-level alarm trigger logic to ensure that events of different risk levels are properly handled.

[0115] The hierarchical alarm unit receives the maximum eigenvalue λ max (t) from the data processing unit, the standardized data of each sensor parameter and the maximum similarity Sim max (t) output by the pattern library unit. Based on the preset trigger conditions, level-I alarm, level-II alarm, and level-III early warning signals are generated. The specific judgment logic is as follows:

[0116] Level-I alarm trigger condition:

[0117] When λ max (t)>Γ(t) and the over-limit count N of sensor parameters over ≥2, it is triggered. Among them, the over-limit of sensor parameters is defined as:

[0118]

[0119] In the formula, μ i and σ i are the mean and standard deviation under historical normal working conditions , and k is the statistical control coefficient. Preferably, k = 3 is taken for vibration and stress parameters, and k = 2.5 is taken for gas concentration parameters to reflect the risk sensitivity differences of different parameters.

[0120] Level-II alarm trigger condition:

[0121] When more than half of the eigenvalues of the dynamic association matrix satisfy λ j (t)>ηΓ(t), it is triggered, where η∈(0,1) is the critical zone ratio factor. This condition reflects the instability trend of the overall association structure of the system. Preferably, the instability state is confirmed by monitoring the change of the eigenvalue distribution entropy value:

[0122]

[0123] When H(t) is lower than the historical normal range, the confidence level of the level-II alarm is enhanced.

[0124] Level-III early warning trigger condition:

[0125] When Sim max (t)>θ simIt is triggered when the matching pattern is marked as the historical accident type. This condition is associated with the typical accident characteristics pre-stored in the pattern library. Preferably, a time continuity constraint is imposed on the similarity calculation result:

[0126]

[0127] to exclude instantaneous mis-matching interference, where T is the length of the observation window.

[0128] After the alarm signal is generated, a multi-level alarm encapsulation protocol is executed. The alarm data packet structure is defined to include:

[0129] Header field: synchronization code (0xAA55) and protocol version number;

[0130] Timestamp: encoding the absolute time in the GPS week second format;

[0131] Alarm level: 3-bit encoding (001: Level III, 010: Level II, 100: Level I);

[0132] Associated features: λ max (t), N over 、Sim max (t);

[0133] Location identifier: construction area grid code (based on Bluetooth beacon triangulation).

[0134] Preferably, forward error correction coding is enabled for Level I alarm data packets, and the Reed-Solomon (15,9) code is used to enhance the transmission reliability. Level II and Level III data packets use CRC-16 checksums.

[0135] At the alarm response level, a dynamic priority scheduling strategy is implemented. When multi-level alarms occur concurrently, they are processed in the order of Level I > Level II > Level III. However, an event merging mechanism is enabled for continuously triggered alarms of the same level:

[0136]

[0137] If the new alarm time t new < t merge , the original alarm timestamp is updated without adding a new record to avoid information overload.

[0138] The hierarchical alarm unit integrates a feedback adjustment mechanism to dynamically optimize the trigger condition parameters based on the manual confirmation results of the monitoring platform. Define the false alarm rate P false and the missed alarm rate P miss as:

[0139]

[0140] When P false > θ falseautomatically increase the conservative coefficient γ of Γ(t) when; when P miss > θ miss reduce the value of η. This adjustment process is implemented by a PID controller to avoid oscillations caused by sudden parameter changes.

[0141] For the wireless transmission unit, in this embodiment, it adopts a dual-mode heterogeneous communication architecture and dynamic parameter adaptation technology to construct a data link that takes into account both the coverage range and transmission reliability, and realizes the efficient and reliable transmission of alarm signals and associated monitoring data. This unit solves the communication reliability problem in the complex electromagnetic environment of the construction site through the coordinated implementation of protocol intelligent switching, channel optimization selection, and anti-interference strategies.

[0142] The wireless transmission unit integrates a LoRa wide-area communication module and a Bluetooth Mesh short-range networking module to form complementary communication capabilities. When the hierarchical alarm unit triggers a level-I alarm, the LoRa module is preferentially enabled, and the spread-spectrum modulation technology is used to transmit data packets in the Sub-1GHz frequency band. Its physical layer frame structure includes a preamble, a synchronization word, and payload data.

[0143] For level-II and level-III alarm signals, start the Bluetooth Mesh networking transmission mode. Construct a topology discovery protocol based on RSSI ranging to dynamically generate a multi-hop routing path.

[0144] At the data encapsulation level, define a unified transmission protocol frame structure:

[0145] Preamble field: 4-byte synchronization header (0xAA55AA55);

[0146] Control field: includes protocol version, encryption flag, priority marker;

[0147] Payload field: encapsulates data such as alarm level, timestamp, and eigenvalue in TLV (Type-Length-Value) format;

[0148] Checksum field: calculates the CRC-32 checksum for all data except the preamble field.

[0149] This embodiment constructs a reliable transmission system that adapts to the complex environment of spherical tank construction through the complementary advantages of heterogeneous communication protocols and intelligent parameter adaptation. The dual-mode dynamic switching mechanism balances the requirements of coverage range and networking flexibility, the anti-interference strategy effectively deals with on-site electromagnetic pollution, while the hierarchical encryption and positioning assistance functions strengthen the security and emergency response capabilities of data transmission, forming a complete wireless monitoring data transmission solution.

[0150] In a preferred embodiment of the present invention, the system further includes a construction stage identification unit.

[0151] In this embodiment, through multi-source sensing feature fusion and temporal pattern analysis, it realizes the autonomous identification of construction processes, providing key context information for the construction of a dynamic association matrix and threshold optimization. Based on the collaborative analysis of the vibration spectrum characteristics and the evolution law of the temperature field, this unit constructs a feature space mapping model to accurately distinguish different construction stages such as civil engineering, welding, and water supply experiments.

[0152] The construction stage recognition unit receives the original data streams from vibration sensors and temperature sensors and executes a parallel feature extraction process. For the vibration signal v(t), first, perform a windowed Fourier transform to obtain the power spectral density estimate:

[0153]

[0154] where w(n) is the Hamming window function and N is the window length. Extract the dominant frequency component as the vibration feature quantity. Preferably, for the high-frequency arc noise unique to the welding stage, a band-stop filter is added to suppress the interference frequency band of 100 - 150 Hz.

[0155] For the temperature signal T(t), calculate the first derivative within a sliding time window to characterize the temperature change rate:

[0156]

[0157] Preferably, the Savitzky-Golay differential filter is used to smooth the noise influence. Its convolution coefficient matrix S is obtained by least-squares fitting of polynomial derivatives, improving the anti-interference ability of the differential operation.

[0158] In the feature fusion stage, construct a two-dimensional feature vector F stage (t) = [f dom (t), r T (t)] T , and input it into a pre-trained classification model for stage discrimination. The classification model adopts the Support Vector Data Description (SVDD) algorithm and defines the decision function:

[0159] ||Φ(F stage ) - c|| 2 ≤R 2 + ξ

[0160] where Φ(·) is the kernel mapping function, c is the center of the feature space, R is the radius of the hypersphere, and ξ is the slack variable. When the feature vector exceeds the boundaries of the hyperspheres of the three categories of civil engineering, welding, and water supply experiments, a stage transition event is triggered.

[0161] In the model training stage, collect historical construction data to construct a labeled sample set where y i∈{1,2,3} corresponds to the construction stage label. By optimizing the objective function:

[0162]

[0163] Solve the optimal hypersphere parameters for each stage, where C is the penalty coefficient to control the fault tolerance of the model.

[0164] Preferably, use the radial basis kernel function k(x i ,x j )=exp(-γ||x i -x j || 2 ) to expand the linear separability of the feature space.

[0165] During the real-time recognition process, implement the state transition constraint logic. Define the stage transition matrix where the element a ij represents the allowable transition probability from stage i to j. When the classification result violates the transition constraint (for example, directly jumping from the welding stage to the civil engineering stage), start the time window integration verification:

[0166]

[0167] Take the mode result of the nearest K moments as the final output to eliminate the influence of instantaneous misjudgment. Preferably, set a state holding timer, and maintain the original stage determination when the stage duration is less than the threshold T hold .

[0168] In this embodiment, through the organic combination of vibration-temperature feature collaborative analysis and the dynamic classification model, accurate identification of the construction stage is achieved. The spectral features capture the mechanical excitation characteristics related to the process, the temperature change rate reflects the evolution law of the thermal process, and the data-driven classification model endows the system with the ability to adapt to complex working conditions, providing reliable support for the context awareness of the entire monitoring system.

[0169] Generally speaking, the working process of the system of the present invention can be described as follows:

[0170] After the system is started, first, the multi-modal sensor unit collects vibration, temperature, stress, gas concentration, and corrosion data during the construction process in real time. The vibration sensor array captures the mechanical vibration spectrum of the spherical tank structure, the temperature sensor synchronously monitors the thermal field distribution of the welding area, the stress sensor inversely calculates the structural stress state through the wavelength shift of the fiber grating, and the gas and corrosion sensors respectively detect the concentration of environmental hazardous substances and the material loss rate. When all the sensing data is transmitted through the RS-485 bus, the clock synchronization module ensures that the acquisition times of all parameters are aligned to eliminate the interference of time domain deviation on the subsequent correlation analysis.

[0171] After the sensing data enters the data processing unit, the construction stage identification unit immediately starts the analysis process: perform windowed Fourier transform on the vibration signal to extract the main frequency components, and at the same time calculate the temperature change rate curve. Input the feature vector after fusing the two into a pre-trained classification model to output the determination result of the current construction stage (civil engineering / welding / water filling experiment). The information in this stage dynamically controls the window length selection of the data processing unit - a short window (10-second level) is used in the welding stage to capture transient thermal shocks, a long window (60-second level) is used in the civil engineering stage to analyze low-frequency vibrations, and a medium window (30-second level) is used in the water filling experiment stage to track steady-state stress changes.

[0172] Based on the selected time window, the data processing unit performs sliding window normalization on the multi-source data. After eliminating the baseline drift, calculate the mutual information coupling degree of any two sensor data. Quantify the non-linear association strength between parameters through KL divergence, and introduce a historical decay factor to smooth the time series fluctuations to construct a dynamically updated symmetric association matrix. At this time, the maximum eigenvalue of the matrix and its derivative terms are extracted, and at the same time, the matrix sequence is input into the pattern library unit for spatio-temporal feature fusion: vertically splice the association matrices of three consecutive time windows, and use a multi-scale convolutional neural network to extract spatio-temporal feature vectors to generate the pattern fingerprint of the current system state.

[0173] The adaptive threshold unit receives the similarity data between the eigenvalue sequence from the data processing unit and the pattern library unit, and constructs a multi-dimensional state space in combination with the construction stage coding. Generate threshold adjustment actions through the reinforcement learning algorithm (TD3) to dynamically optimize the alarm threshold curve. When the eigenvalue exceeds the dynamic threshold, the hierarchical alarm unit starts multi-dimensional determination: if the maximum eigenvalue exceeds the limit and at least two sensor parameters exceed the statistical control limit, trigger a level I alarm; when more than half of the eigenvalues enter the threshold critical area, start a level II alarm; if the pattern fingerprint matches the historical accident database successfully, generate a level III early warning.

[0174] After the alarm signal is generated, the wireless transmission unit starts a differential communication strategy according to the level: the level I alarm preferentially uses LoRa wide-area transmission to dynamically improve the signal coverage through the spreading factor; the level II and III alarms use Bluetooth Mesh networking to achieve multi-hop relay. Channel sensing and frequency hopping avoidance are implemented during the transmission process to combat electromagnetic interference at the construction site, and at the same time, positioning data is embedded to assist emergency response.

[0175] In the whole process, the pattern library unit continuously receives new spatio-temporal feature vectors. When an unknown pattern is detected, the feature library is expanded through an incremental clustering algorithm, and the forgetting factor mechanism is used to eliminate old patterns. The system dynamically adjusts the alarm threshold optimization direction and the pattern matching threshold through the manual feedback data of the monitoring terminal to form a closed-loop monitoring system with self-evolution ability.

[0176] Through the deep coupling of data flow and control flow among various units, the full process from raw data acquisition, feature extraction, dynamic analysis to intelligent decision-making is realized. The recognition results in the construction stage serve as the core variables of context awareness, running through key links such as window selection, matrix construction, and threshold optimization; the two-way interaction between the pattern library and the adaptive threshold endows the system with the ability to synergistically utilize historical experience and real-time status, ultimately achieving multi-level precise prevention and control of construction safety risks.

[0177] Although the embodiments of the present invention have been shown and described, those of ordinary skill in the art can understand that various changes, modifications, substitutions, and variations can be made to these embodiments without departing from the principles and spirit of the present invention. The scope of the present invention is defined by the appended claims and their equivalents.

Claims

1. An intelligent monitoring system for the safe construction of spherical tanks, characterized in that, Including: A multi-modal sensor unit for real-time collection of vibration, temperature, stress, gas concentration, and corrosion data during the construction phase; A data processing unit connected to the sensor unit for preprocessing the collected data and constructing a multi-parameter dynamic correlation matrix, and outputting matrix eigenvalues; A pattern library unit connected to the data processing unit, storing historical abnormal feature patterns and providing real-time pattern matching; An adaptive threshold unit connected to the data processing unit and the pattern library unit respectively, dynamically optimizing the alarm threshold according to the eigenvalue change trend and pattern matching result; A hierarchical alarm unit receiving the output of the data processing unit and the threshold of the adaptive threshold unit, triggering multi-level alarms based on the comprehensive determination of eigenvalue overrun and pattern matching; A wireless transmission unit connected to the hierarchical alarm unit, sending the alarm signal to the monitoring terminal.

2. The intelligent monitoring system for the safe construction of spherical tanks according to claim 1, wherein, When the data processing unit constructs the multi-parameter dynamic correlation matrix: Calculating the mutual information coupling degree of any two sensor data as matrix elements, and introducing a historical coupling degree attenuation factor to realize the dynamic update of matrix elements.

3. The intelligent monitoring system for the safe construction of spherical tanks according to claim 2, wherein When calculating the mutual information coupling degree: Adopting a sliding window joint probability distribution estimation method, and the window length is dynamically adjusted according to the current construction phase.

4. The intelligent monitoring system for the safe construction of spherical tanks according to claim 1, characterized in that, The adaptive threshold unit adopts a reinforcement learning algorithm, generates threshold adjustment actions by constructing a state space including the eigenvalue change rate and the construction phase encoding.

5. The intelligent monitoring system for the safe construction of spherical tanks according to claim 4, characterized in that, The state space includes: The first derivative and second derivative of the maximum eigenvalue, and the construction phase vector generated by one-hot encoding.

6. The intelligent monitoring system for the safe construction of spherical tanks according to claim 1, wherein The hierarchical alarm unit executes: Triggering a level-I alarm when the maximum eigenvalue exceeds the dynamic threshold and at least two sensor data exceed the statistical control limit; Triggering a level-II alarm when more than half of the eigenvalues reach the threshold critical area; Triggering a level-III early warning when the current pattern matches the historical abnormal pattern library successfully.

7. The intelligent monitoring system for the safe construction of spherical tanks according to claim 1, wherein The pattern library unit extracts the time-varying features of the dynamic correlation matrix through a convolutional neural network to generate a pattern fingerprint, and automatically expands the pattern library when the similarity between the new feature and the stored pattern is lower than the preset threshold.

8. The intelligent monitoring system for the safe construction of spherical tanks according to claim 7, wherein, When generating the pattern fingerprint: Vertically splicing the dynamic correlation matrices of three consecutive time windows, and using three-layer convolutional kernels with different scales to extract spatio-temporal features of the spliced matrix.

9. The intelligent monitoring system for the safe construction of spherical tanks according to claim 1, wherein The wireless transmission unit adopts a LoRa and Bluetooth Mesh dual-mode communication protocol, and automatically selects the transmission channel and transmission power according to the alarm level.

10. The intelligent monitoring system for the safe construction of spherical tanks according to claim 1, characterized in that, The system further includes: A construction phase identification unit for automatically identifying the current construction phase by analyzing the main frequency component of the vibration signal and the temperature change rate curve.

Images