Large pump station operation safety monitoring system and method based on multi-source data fusion

Abstract

The application discloses a large pump station operation safety monitoring system and method based on multi-source data fusion, and belongs to the field of water conservancy projects and industrial internet of things monitoring. The system comprises a multi-parameter acquisition module, a data transmission module, an edge computing module, a comprehensive evaluation module and a decision support module. The multi-parameter acquisition module comprises a sensor array arranged on a unit, a data acquisition unit connected with the sensor array and a signal preprocessing submodule connected with the data acquisition unit. The sensor array arranged on the unit and the data acquisition unit connected with the sensor array are used for collecting heterogeneous data of the unit and performing high-quality preprocessing on original signals. The signal preprocessing submodule is used for filtering the original signals according to an algorithm combining variational mode decomposition and wavelet threshold denoising. The application solves the problems of isolated traditional monitoring parameters and static evaluation, and improves the operation safety of the pump station and the intelligent level of operation and maintenance.

Description

Technical Field

[0001] This invention relates to the field of water conservancy engineering and industrial Internet of Things monitoring, specifically to a large-scale pumping station operation safety monitoring system and method based on multi-source data fusion. Background Technology

[0002] Large-scale pumping stations, as a core component of water conservancy infrastructure, are widely used in critical scenarios such as urban water supply, drainage and flood control, farmland irrigation, and inter-basin water transfer. Their operational safety is directly related to the safety of people's lives and property, the stability of industrial and agricultural production, and the sustainability of the ecological environment. As pumping station units develop towards larger scale and intelligence, higher requirements are placed on the comprehensiveness, real-time nature, and accuracy of operational status monitoring.

[0003] Existing large-scale pump station operation monitoring technologies have several shortcomings: First, monitoring parameters are limited, mostly focusing on a few indicators such as electrical quantities or water levels, which fails to comprehensively reflect the overall operating status of the unit. Second, data processing methods are simplistic, often involving threshold judgments for single parameters, lacking the ability to integrate and analyze multi-source heterogeneous data, leading to false alarms and missed alarms. Third, assessment methods are static, making it difficult to dynamically track changes in the unit's operating status and achieve proactive assessments of safety conditions. Fourth, fault response is passive, only capable of issuing fault alarms and unable to proactively generate accurate handling suggestions based on fault type and severity, resulting in delayed operation and maintenance decisions, increasing the risk of equipment damage and operation and maintenance costs.

[0004] Therefore, developing a large-scale pumping station operation safety monitoring system and method based on multi-source data fusion, capable of comprehensive acquisition of multiple parameters, multi-dimensional integrated analysis, dynamic safety assessment, and intelligent decision support, has become an urgent need to address the shortcomings of existing technologies. Summary of the Invention

[0005] The purpose of this invention is to provide a large-scale pumping station operation safety monitoring system and method based on multi-source data fusion, which can realize real-time monitoring, accurate evaluation and intelligent decision-making.

[0006] To solve the above-mentioned technical problems, the present invention provides the following technical solution, specifically including: A large-scale pumping station operation safety monitoring system based on multi-source data fusion, the system comprising: The system comprises a multi-parameter acquisition module, a data transmission module, an edge computing module, a comprehensive evaluation module, and a decision support module, forming a closed-loop operation for acquisition, transmission, processing, evaluation, and decision-making.

[0007] The multi-parameter acquisition module includes a sensor array deployed on the unit, a data acquisition unit connected to the sensor array, and a signal preprocessing submodule connected to the data acquisition unit. The signal preprocessing submodule is used to filter the original signal according to an algorithm combining variational mode decomposition and wavelet threshold denoising. This module adopts an advanced filtering algorithm combining variational mode decomposition (VMD) and wavelet threshold denoising, which can effectively separate and filter out power frequency interference and random noise in the strong noise environment of the pumping station, significantly improve the signal-to-noise ratio and quality of the original signal, and lay a reliable data foundation for subsequent accurate feature extraction and state analysis, fundamentally solving the problems of poor signal quality and the submersion of weak fault features in traditional monitoring.

[0008] The data transmission module uses a hybrid wired and wireless transmission architecture and employs the AES-256 algorithm to encrypt the transmitted data at the application layer, ensuring the continuity, reliability, and confidentiality of the monitoring data during transmission.

[0009] The edge computing module includes a data fusion submodule and a real-time early warning submodule. The data fusion submodule is configured to use an algorithm based on improved DS evidence theory to fuse heterogeneous data from multi-parameter acquisition modules and generate a comprehensive feature vector of unit operating status. The real-time early warning submodule is configured to perform threshold and trend analysis based on the comprehensive feature vector and trigger a local alarm, realizing early and accurate identification of abnormal unit status and millisecond-level local response, which greatly improves the system's real-time performance, reliability, and early warning capability for gradual faults.

[0010] The comprehensive evaluation module includes an evaluation model construction submodule, a status evaluation submodule, and a historical data analysis submodule. The evaluation model construction submodule builds a safety evaluation model based on an LSTM neural network. The input of the model is the comprehensive feature vector of the unit's operating status, and the output is the safety evaluation results of several levels set by the system. This improves the safety assessment from a simple static threshold alarm to a dynamic trend evaluation, realizing true predictive maintenance and significantly improving the foresight and scientific nature of the status assessment.

[0011] The decision support module includes a fault diagnosis submodule and a processing suggestion generation submodule. The fault diagnosis submodule is configured to diagnose abnormal states by combining a fault knowledge base with an algorithm that combines case reasoning and rule reasoning. The processing suggestion generation submodule generates processing suggestions with multiple dimensions based on the diagnosis results, extending the system function from alarms to diagnosis and decision-making, greatly reducing the reliance on the personal experience of operation and maintenance personnel, and improving the speed, standardization, and decision-making efficiency of operation and maintenance response.

[0012] Therefore, the aforementioned large-scale pumping station operation safety monitoring system based on multi-source data fusion constitutes a business closed loop with self-learning capabilities: First, the operation and maintenance feedback information generated by the decision support module triggers the multi-parameter acquisition module to perform targeted retesting or system parameter adjustments. Subsequently, the system re-acquires and implements the unit status data after processing the recommendations. This new data is ultimately used in two ways: first, to drive the safety evaluation model of the comprehensive evaluation module to perform self-optimization; and second, to update the fault knowledge base of the decision support module.

[0013] The sensor array includes a current sensor, a voltage sensor, a limit switch, a water level sensor, an electromagnetic flowmeter, a vibration acceleration sensor, an eddy current swing sensor, and a noise sensor, which are respectively deployed inside the unit's motor, pump body, coupling, inlet and outlet water channels, and pump house.

[0014] The signal preprocessing submodule is specifically configured to perform the following steps: set the number of modes K, perform variational mode decomposition on the original signal to obtain K intrinsic mode functions; calculate the correlation coefficient between each intrinsic mode function and the original signal, and select effective modes according to a preset correlation coefficient threshold; perform discrete wavelet transform on each of the selected effective modes, apply a soft threshold function to the wavelet detail coefficients obtained by decomposition for denoising, and reconstruct the denoised wavelet coefficients to obtain denoised mode components; sum all denoised effective mode components to obtain the final preprocessed signal.

[0015] The improved DS evidence theory algorithm used in the data fusion submodule constructs a basic probability allocation function for the same identification framework based on features extracted from data from different sensors; calculates the Jousselme distance between each pair of evidence to quantify the degree of evidence conflict; for highly conflicting evidence with a conflict level exceeding a preset threshold, calculates credibility weights based on the similarity matrix between all evidence, and uses the credibility weights to perform a weighted average correction on the basic probability allocation of highly conflicting evidence; and synthesizes all the corrected evidence using the Dempster combination rule to obtain the fused comprehensive credibility distribution.

[0016] The comprehensive evaluation module also has a model self-optimization function. When the amount of newly added valid fault case data exceeds the first preset threshold, or when the evaluation accuracy of the LSTM neural network security evaluation model on the recent validation set decreases by more than the second preset threshold, the model optimization process is automatically triggered. The newly added valid fault case data is used to incrementally learn or retrain the security evaluation model, and the LSTM neural network security evaluation model is updated after the validation is passed.

[0017] The fault knowledge base supports dynamic updates. New fault cases verified on-site can be manually entered by maintenance personnel through their client. The latest fault feature database and handling specification documents can be automatically imported from industry standard databases or equipment manufacturer service platforms through a preset application programming interface.

[0018] The edge computing module is deployed on an industrial-grade edge server that supports multi-task parallel processing and is equipped with a local solid-state storage unit for continuously storing unit operation data and early warning logs during network interruptions of the data transmission module.

[0019] A method for monitoring the operational safety of large-scale pumping stations based on multi-source data fusion includes the following steps: S1: Synchronously collect multi-dimensional operating data of the unit and use an algorithm combining variational mode decomposition and wavelet threshold denoising for signal preprocessing; S2: Encrypt and transmit the preprocessed data to the edge computing module; S3: An algorithm based on improved DS evidence theory is used to fuse multi-source heterogeneous data, generate a comprehensive feature vector of unit operation, and provide real-time early warning based on the comprehensive feature vector of unit operation. S4: Input the comprehensive feature vector of the unit operation into the pre-trained LSTM neural network safety evaluation model to obtain the current safety level evaluation result of the unit; S5: When the evaluation result is abnormal, the fault diagnosis is performed by combining case reasoning and rule reasoning algorithms with the fault knowledge base, and multi-dimensional processing suggestions are generated. In this process, the feedback data from the unit after the processing suggestions are implemented is re-collected by the system and used for model self-optimization and knowledge base updates, forming a closed loop.

[0020] The real-time early warning in step S3 includes: threshold early warning by comparing key indicators in the integrated feature vector of unit operation with preset thresholds, and trend early warning by performing trend analysis on key indicators to identify deteriorating trends.

[0021] The training data for the LSTM neural network safety evaluation model described in step S4 is constructed from historical operating data of the unit, fault case data, and safety level labels marked according to industry standards.

[0022] Compared with the prior art, the beneficial effects achieved by the present invention are: Comprehensive monitoring parameters: This invention uses a multi-type sensor array to achieve simultaneous acquisition of multiple dimensions of parameters such as electrical quantities, switching quantities, water level, flow rate, vibration, swing, and noise. It covers the core operating indicators of the unit's electrical, mechanical, and hydraulic systems, and can comprehensively reflect the unit's overall operating status, thus solving the deficiency of existing technologies in monitoring only one parameter.

[0023] Data processing is real-time and efficient, employing an edge computing and cloud-based collaborative processing architecture. Localized edge computing enables real-time data fusion processing and rapid early warning, reducing cloud transmission pressure and response latency. The cloud platform enables in-depth comprehensive evaluation and historical data analysis, balancing real-time and in-depth analysis needs.

[0024] The dynamic safety evaluation model, based on LSTM neural network, can fully explore the temporal characteristics and correlations of multi-source data to achieve accurate evaluation and dynamic tracking of unit safety level. Compared with the traditional static threshold evaluation method, it significantly improves the scientific nature and foresight of the evaluation.

[0025] Intelligent decision support accurately locates the type and location of faults through fault diagnosis algorithms and automatically generates targeted handling suggestions. It realizes intelligent operation and maintenance decision-making throughout the entire process, from fault warning to operation and maintenance decision-making. It solves the problems of passive alarm and lack of decision support in existing technologies, effectively improves operation and maintenance efficiency, and reduces the risk of equipment damage and operation and maintenance costs. Attached Figure Description

[0026] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings: Figure 1 This is a schematic diagram of the module connection of the large-scale pumping station operation safety monitoring system based on multi-source data fusion according to the present invention; Figure 2 This is a flowchart of the steps of the method for monitoring the operational safety of large pumping stations based on multi-source data fusion, as proposed in this invention. Detailed Implementation

[0027] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0028] Please see Figures 1-2 The present invention provides the following technical solution: like Figure 1 As shown, the implementation of this system is based on a clear hardware architecture and modular deployment.

[0029] Sensor arrays are deployed on the motor, pump body, coupling, and inlet / outlet water channels of the pump station unit. The sensor arrays are connected to the data acquisition unit, and the data acquisition unit is connected to the signal preprocessing submodule.

[0030] It adopts a hybrid wired and wireless architecture to provide an encrypted channel for data interaction between the multi-parameter acquisition module, edge computing module, comprehensive evaluation module and decision support module.

[0031] An industrial-grade edge server is deployed in the pump station control room as a carrier for edge computing modules. This server is connected to the data acquisition unit and local audible and visual alarms via an industrial network.

[0032] In the pump station's data center or cloud, server clusters are deployed to run the comprehensive evaluation module and decision support module. These servers communicate securely with local edge servers via a hybrid network consisting of industrial Ethernet and 4G / 5G wireless networks.

[0033] The maintenance personnel's computers and mobile terminals access the system via the network and interact with the comprehensive evaluation module and decision support module.

[0034] The modules are connected sequentially according to the signal flow of "multi-parameter acquisition module, data transmission module, edge computing module, comprehensive evaluation module, and decision support module", forming a physical closed loop from the physical site to the cloud and back to the operation and maintenance personnel.

[0035] like Figure 2 As shown, the implementation process is broken down as follows: S1. Specific operational steps for signal preprocessing: Taking a single vibration signal acquisition as an example, the specific implementation steps are as follows: Step S100: First, the multi-dimensional operating data of the unit is synchronously collected through the data acquisition unit of the multi-parameter acquisition module at a differentiated sampling rate. The sampling rate of the vibration signal is set to 10kHz, the current signal to 2kHz, and the water level signal to 1Hz.

[0036] Step S110: Define the signal length N = 10240 and the sampling frequency fs = 12800Hz. Perform a joint filtering algorithm combining variational mode decomposition and wavelet threshold denoising on the acquired original vibration signal x(t).

[0037] Variational mode decomposition was performed on the vibration signal x(t). The mode number K=5, penalty parameter α=2000, and noise tolerance τ=0 were set. The constrained variational problem was solved using the alternating direction multiplier method, iteratively updated until convergence, yielding five intrinsic mode function (IMF) components. and its center frequency ω k .

[0038] Subsequently, valid modes are selected, and a set of valid modes is defined as L. The eigenmode function components u of each mode are then calculated. k(t) The correlation coefficient ρ with the vibration signal x(t) k Set threshold ρth =0.1, retain all ρ k >ρ th The IMF components are used to form an effective mode set L, thereby eliminating noise-dominated spurious components.

[0039] Next, for each effective modal component L in L... i Wavelet thresholding denoising is performed sequentially. First, a 5-level discrete wavelet decomposition is performed using discrete wavelet basis functions to obtain the approximation coefficients ca5 and the set of detail coefficients. Then, the noise standard deviation σ of the first layer detail coefficient cd1 is estimated, and the denoising threshold is calculated using a general threshold formula: .

[0040] Assuming σ is estimated to be 0.15, then λ = 0.15 × (2ln(10240)). 1 / 2 )≈0.645; then for all detail coefficients The soft thresholding function is applied to obtain the denoised detail coefficients cD. j : ; Assuming a detail coefficient cd3 = 1.2 and a threshold λ = 0.645, then cD3 = sign(1.2)·max(|1.2|−0.645,0) = 0.555. Simultaneously, the approximation coefficient ca5 is retained. Finally, wavelet reconstruction is performed using the denoised coefficients to obtain the denoised modal component L. it After completing the denoising process for all valid modes, all denoised modal components L it Summing these values ​​yields the final pre-processed clean signal y(t), i.e.: y(t) ; Where M represents the total number of modal components in the effective modal component L.

[0041] S2. Specific operational steps for secure encrypted data transmission: Step S200: The clean signal y(t) preprocessed in step S1 is sent to the edge computing module through the data transmission module. First, application layer encryption is performed. Before the data is sent, the system uses the AES-256-GCM algorithm to encrypt the data packet throughout the process. During the encryption process, a 256-bit key and a randomly generated initialization vector (IV) are used to ensure the confidentiality and integrity of the data during transmission.

[0042] Subsequently, hybrid routing transmission is implemented. The system adopts a "wired as the primary and wireless as the secondary" architecture. Data is preferentially transmitted through a highly reliable industrial Ethernet wired link. When a failure is detected in the wired network in real time, the system automatically and seamlessly switches to the backup 4G / 5G wireless VPN channel, thereby ensuring the continuity of data transmission and the overall reliability of the system.

[0043] S3. Specific operational steps for data fusion and early warning in the edge computing module: Step S300: In the edge computing module deployed locally at the pump station, the system performs data fusion and real-time early warning. First, time-domain and frequency-domain features are extracted from the preprocessed heterogeneous data, and a basic probability allocation function is defined for each feature for the same identification framework T={T1: normal, T2: unbalanced, T3: misaligned, T4: bearing failure}.

[0044] For example, assuming the effective vibration characteristic value v = 7.5, the trapezoidal membership function μ is used. high(v) Construct its basic probability assignment (BPA). For the bearing failure proposition, its BPA is: m1(T4) = μ high(v) =0.8; For normal propositions, their BPA is: m1(T1)=1-μ high(v) =0.2, where μ high(v) It is the membership degree of the vibration value "high", and its value is between 0 and 1, increasing as v increases.

[0045] Next, the improved DS evidence theory fusion algorithm is executed: the Jousselme distance d between each pair of evidence is calculated. J (m i ,m j To quantify the degree of conflict, it is defined as: ; Where m i The BPA function representing the i-th piece of evidence, m j The BPA function representing the j-th piece of evidence, (m i - m j ) represents the difference between the basic probability assignment vectors of the two pieces of evidence, and the result is a column vector, (m i - m j )ᵀ represents the transpose of this difference vector, i.e., a row vector.

[0046] Where D is a matrix measuring the similarity of proposition sets; and the evidence body E i Average distance from the remaining evidence As indicated by the degree of conflict, where A represents the total number of pieces of evidence, an example is shown below: Assume there are 3 pieces of evidence, d j(m1,m2)=0.9,d j (m1,m3)=0.4,d j If (m2,m3)=0.6, then Conf2=(0.9+0.6) / 2=0.75. If the preset high conflict threshold is 0.7, then E2 is high conflict evidence; then the similarity matrix S between all evidence bodies is calculated, where... Based on this, the credibility weight of each piece of evidence Ei is calculated. Continuing from the previous example, Then w1 = (1 + 0.1 + 0.6) / 5.2 ≈ 0.327, where 5.2 is the sum of all elements of the matrix.

[0047] Then, using weight w i A weighted average correction is applied to the BPA of highly conflicting evidence. A simplified weighted average correction can be expressed as: R2 = w1·m1(T) i )+w2·m2(T i )+w3·m3(T i ).

[0048] For proposition T1: R2(T1) = 0.327 × 0.2 + 0.288 × 0.3 + 0.385 × 0.5 ≈ 0.344; For proposition T2: R2(T2) = 0.327×0 + 0.288×0.7 + 0.385×0 ≈ 0.202; For proposition T3: R2(T3) = 0.327×0 + 0.288×0 + 0.385×0 = 0; For proposition T4: R²(T4) = 0.327 × 0.8 + 0.288 × 0 + 0.385 × 0.3 ≈ 0.378; For uncertainty T: R²(T) = 0.327 × 0.8 + 0.288 × 0 + 0.385 × 0.2 ≈ 0.077; Verification: 0.344+0.202+0+0.378+0.077=1.001.

[0049] After the corrections were completed, Dempster's combination rule was used to synthesize all BPAs pairwise to obtain the final fused confidence distribution m. f Based on the final reliability distribution, the proposition with the highest reliability is selected as the current state judgment, and the reliability value m of each proposition is assigned. f (θ k The normalized values ​​of the key characteristic vibration effective value v are combined to generate a comprehensive feature vector F.

[0050] Specifically, the comprehensive feature vector F is a 10-dimensional vector, and its construction process is as follows: The fused confidence distribution m f The confidence values ​​for the four propositions {T1, T2, T3, T4} are sequentially filled into the first to fourth dimensions of the vector. Continuing the previous example, assume that the composite result is m. f (T1)=0.15,m f (T2)=0.10, m f (T3)=0.00,m f If (T4) = 0.70, then F[1:4] = [0.15, 0.10, 0.00, 0.70].

[0051] Calculate the reliability value, compare the reliability values ​​of dimensions 1 to 4 with the system's preset judgment threshold, calculate the difference as the reliability strength, and fill it into dimensions 5 to 8 of the vector. In the example above, with a threshold of 0.6, the reliability strengths are: 0.15 - 0.6 = -0.45, 0.10 - 0.6 = -0.50, 0.00 - 0.6 = -0.60, 0.70 - 0.6 = 0.10. Therefore, F[5:8] = [-0.45, -0.50, -0.60, 0.10].

[0052] Fill in the normalized value of key features and normalize the key features such as the effective value of vibration v. For example, the effective value of vibration v = 7.5 mm / s and the preset alarm value is 10 mm / s. Then its normalized value is 7.5 / 10 = 0.75. Fill this value into the 9th dimension, that is, F[9] = 0.75.

[0053] Other features are filled in, and the normalized value of another key feature is filled into the 10th dimension. Assuming the normalized value is 0.30, then F

[10] =0.30.

[0054] In summary, an example of the generated comprehensive feature vector F is shown below: F = [0.15, 0.10, 0.00, 0.70,-0.45,-0.50,-0.60,0.10, 0.75, 0.30] T ; Finally, real-time early warnings are implemented for key components in vector F: threshold warnings are implemented by comparing component values ​​with preset "attention values" and "alarm values," and an alarm is triggered if the values ​​exceed the limits; at the same time, trend warnings are implemented by taking the most recent 30 sampling points of the component and performing linear fitting to obtain the slope β. If β>0 and its significance test p-value<0.05, a "deteriorating trend" warning is triggered.

[0055] S4. Specific operational steps for model construction, evaluation, and self-optimization in the comprehensive evaluation module: Step S400: In the comprehensive evaluation module deployed in the cloud, the system first constructs an LSTM model based on the TensorFlow framework. Its input is a 10-dimensional comprehensive feature vector F. The network structure contains two layers of LSTM units, followed by a Dropout layer, a fully connected layer, and an output layer. The final output is a discrete probability distribution of safety levels such as excellent, good, average, and dangerous. Subsequently, using historical running data, a feature vector sequence is generated through the aforementioned signal preprocessing and feature fusion steps. Domain experts label the safety levels according to relevant standards, and the set is divided into training, validation, and test sets in an 8:1:1 ratio. The model is trained using the cross-entropy loss function and the Adam optimizer, and an early stopping strategy is introduced to prevent overfitting.

[0056] After training, the model is deployed as a RESTful API service for online use by the status assessment submodule, enabling real-time inference and security level determination of newly input feature vectors F. Furthermore, the system incorporates a model self-optimization mechanism, continuously monitoring the number of newly added valid fault cases and the model's performance on the validation set. When the number of new fault cases exceeds 50 or the recent validation accuracy drops by more than 5%, an incremental learning process is automatically triggered: the system locks the most recent six months of operational data, fine-tunes the training with a lower learning rate based on the existing model weights, and seamlessly updates the online service with the new model version after successful validation. This ensures that the security assessment capability possesses the characteristics of continuous self-evolution and optimization.

[0057] S5. Specific operational steps for fault diagnosis in the decision support module: Step S500: When the evaluation result is "moderate" or "dangerous", the decision support module is activated. The knowledge base representation module is responsible for building and maintaining the fault knowledge base, which includes a historical case library stored in the form of structured cases, and an expert rule library encoded in the form of production rules.

[0058] Based on this knowledge base, the system performs hybrid diagnostics. First, the rule-based reasoning engine matches the current anomaly feature vector F with the rule base, quickly generating a list H containing preliminary fault hypotheses and their confidence levels. r Simultaneously, the case reasoning engine calculates the current feature vector F and compares it with the feature vectors F of all historical cases in the case library. arr The Euclidean distance is used to retrieve the top-5 most similar historical cases. Finally, the diagnostic decision engine merges the hypothesis list output by RBR with the similar case information retrieved by CBR, and integrates the evidence from both through strategies such as weighted calculation. Finally, it outputs an accurate diagnostic result such as "pitting corrosion in the inner ring of the bearing" and its quantitative confidence level, and automatically associates and pushes the solutions recorded in the most similar historical cases.

[0059] S6. The specific process of closed-loop feedback: Step S600: The system forms a self-learning closed loop of human-machine collaboration. Maintenance personnel perform maintenance operations based on the multi-dimensional processing suggestions generated by the system, and the unit's status returns to normal. The system re-collects the unit's operating data after the processing suggestions are implemented. After the comprehensive evaluation module determines that the safety level remains "good," it automatically associates and labels the complete data stream from early warning triggering, fault diagnosis, processing suggestion generation to maintenance verification according to time sequence as positive cases and stores them in the historical database. These positive cases are then included in the candidate dataset for model self-optimization and knowledge base updates, and are used for incremental training of the LSTM model and expansion of the fault knowledge base after the triggering conditions are met. Thus, a complete intelligent closed loop is completed.

[0060] It should be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, method, article, or apparatus.

[0061] Finally, it should be noted that the above descriptions are merely preferred embodiments of the present invention and are not intended to limit the present invention. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art can still modify the technical solutions described in the foregoing embodiments or make equivalent substitutions for some of the technical features. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A large-scale pumping station operation safety monitoring system based on multi-source data fusion, characterized in that, It includes a multi-parameter acquisition module, a data transmission module, an edge computing module, a comprehensive evaluation module, and a decision support module; The multi-parameter acquisition module includes a sensor array deployed on the unit, a data acquisition unit connected to the sensor array, and a signal preprocessing submodule connected to the data acquisition unit; The sensor array deployed on the unit and the data acquisition unit connected to the sensor array are used to collect heterogeneous data from the unit and perform high-quality preprocessing on the raw signals. The signal preprocessing submodule is used to filter the original signal according to an algorithm that combines variational mode decomposition and wavelet threshold denoising. The data transmission module uses a hybrid wired and wireless transmission architecture and performs application-layer encryption on the transmitted data. The edge computing module includes a data fusion submodule and a real-time early warning submodule; The data fusion submodule fuses heterogeneous data from the multi-parameter acquisition module to generate a comprehensive feature vector of the unit's operating status.

2. The large-scale pumping station operation safety monitoring system based on multi-source data fusion according to claim 1, characterized in that, The real-time early warning submodule performs threshold and trend analysis based on the comprehensive feature vector of the unit's operating status and triggers a local alarm; the comprehensive evaluation module includes an evaluation model construction submodule, a status evaluation submodule, and a historical data analysis submodule; the evaluation model construction submodule configures and constructs a safety evaluation model; the input of the safety evaluation model is the comprehensive feature vector of the unit's operating status, and the output is several levels of safety evaluation results set by the system; The decision support module includes a fault diagnosis submodule and a processing suggestion generation submodule; The fault diagnosis submodule is configured to diagnose abnormal states by combining a fault knowledge base; the processing suggestion generation submodule generates processing suggestions containing multiple dimensions based on the diagnosis results.

3. The large-scale pumping station operation safety monitoring system based on multi-source data fusion according to claim 1, characterized in that, The sensor array deployed on the unit includes a current sensor, a voltage sensor, a limit switch, a water level sensor, an electromagnetic flowmeter, a vibration acceleration sensor, an eddy current slew rate sensor, and a noise sensor, which are respectively deployed inside the unit's motor, pump body, coupling, inlet and outlet water channels, and pump house.

4. The large-scale pumping station operation safety monitoring system based on multi-source data fusion according to claim 1, characterized in that, The signal preprocessing submodule is specifically configured to perform the following steps: set the number of modes K, perform variational mode decomposition on the original signal to obtain K intrinsic mode functions; calculate the correlation coefficient between each intrinsic mode function and the original signal, and select effective modes according to a preset correlation coefficient threshold; perform discrete wavelet transform on each of the selected effective modes, apply a soft threshold function to the wavelet detail coefficients obtained by decomposition for denoising, and reconstruct the denoised wavelet coefficients to obtain denoised mode components; sum all denoised effective mode components to obtain the final preprocessed signal.

5. The large-scale pumping station operation safety monitoring system based on multi-source data fusion according to claim 1, characterized in that, The data fusion submodule specifically includes: constructing a basic probability allocation function for the same identification framework based on features extracted from different sensor data; calculating the Jousselme distance between each pair of evidence to quantify the degree of evidence conflict; for highly conflicting evidence with a degree of evidence conflict exceeding a preset threshold, calculating the credibility weight of each evidence based on the similarity matrix between all evidence, and using the credibility weight to perform a weighted average correction on the basic probability allocation of highly conflicting evidence; and synthesizing all the corrected evidence using the Dempster combination rule to obtain the fused comprehensive credibility distribution.

6. The large-scale pumping station operation safety monitoring system based on multi-source data fusion according to claim 1, characterized in that, The comprehensive evaluation module also has a model self-optimization function. When the amount of newly added valid fault case data exceeds the first preset threshold, or when the evaluation accuracy of the LSTM neural network security evaluation model on the recent validation set decreases by more than the second preset threshold, the model optimization process is automatically triggered. The newly added valid fault case data is used to incrementally learn or retrain the security evaluation model, and the LSTM neural network security evaluation model is updated after the validation is passed.

7. A large-scale pumping station operation safety monitoring system based on multi-source data fusion according to claim 1, characterized in that, The fault knowledge base supports dynamic updates, which can be achieved through: manual input of complete information on new fault cases verified on-site by maintenance personnel via their client; and automatic import of the latest fault feature database and handling specification documents from industry standard databases or equipment manufacturer service platforms via a preset application programming interface.

8. The large-scale pumping station operation safety monitoring system based on multi-source data fusion according to claim 1, characterized in that, The edge computing module is deployed on an industrial-grade edge server that supports multi-task parallel processing and is equipped with a local solid-state storage unit for continuously storing unit operation data and early warning logs during network interruptions of the data transmission module.

9. A method for monitoring the operational safety of large pumping stations based on multi-source data fusion, applied to the operational safety monitoring system for large pumping stations based on multi-source data fusion as described in any one of claims 1-8, characterized in that, Includes the following steps: S1: By collecting multi-dimensional operating data of the unit, signal preprocessing is performed using an algorithm that combines variational mode decomposition and wavelet threshold denoising. S2: Encrypt the preprocessed data before transmission; S3: An algorithm based on improved DS evidence theory is used to fuse heterogeneous data, generate a comprehensive feature vector of unit operating status, and provide real-time early warning based on the comprehensive feature vector of unit operating status; S4: Input the comprehensive feature vector of the unit's operating status into the pre-trained LSTM neural network safety evaluation model to obtain the current safety level evaluation result of the unit; S5: When the evaluation result is abnormal, the fault diagnosis is performed by combining case reasoning and rule reasoning algorithms with the fault knowledge base, and multi-dimensional processing suggestions are generated. After the processing suggestions are executed, the feedback data of the unit is re-collected by the system. When the system detects that the unit status has returned to normal, the complete fault-diagnosis-processing-recovery data flow is automatically labeled and stored in the historical database to trigger model self-optimization and knowledge base updates, forming a human-machine collaborative business closed loop.

10. A method for monitoring the operational safety of large pumping stations based on multi-source data fusion according to claim 9, characterized in that, The real-time early warning in step S3 includes: threshold early warning, which compares key indicators in the comprehensive feature vector of the unit's operating status with preset thresholds, and trend early warning, which performs trend analysis on key indicators to identify deteriorating trends.

Images