Wellhead device fault prediction system based on narrowband Internet of Things

Abstract

The invention relates to the technical field of intelligent operation and maintenance and safety early warning of oil and gas equipment, in particular to a wellhead device fault prediction system based on the narrowband Internet of Things. Comprising a data acquisition module used for acquiring unstructured text data and structured time sequence data of a target wellhead device; the factor quantification module is used for determining a working condition abrupt change factor; the dynamic weight module is used for determining a dynamic weight factor; the fault prediction module is used for responding to the dynamic weight factor and determining a direct fault probability; the uncertainty quantification module is used for determining an uncertainty index based on multiple prediction results of the fault prediction module; the risk assessment module is used for determining a comprehensive risk index in combination with the direct fault probability and the uncertainty index; and the decision response module is used for generating a grading early warning instruction. According to the method, the conversion from passive response to active prediction is realized, the perspectiveness, the self-adaptability and the timeliness of fault prediction are remarkably enhanced, and earlier early warning is realized.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of intelligent operation and maintenance and safety early warning of oil and gas equipment, in particular to a wellhead device fault prediction system based on narrowband Internet of Things. BACKGROUND

[0002] The wellhead device is a key equipment for oil and gas exploitation, and its running state is directly related to production safety and efficiency; how to accurately predict its potential failure and realize the transition from passive response to proactive predictive maintenance is a core technical challenge to ensure the safe and efficient operation of oil and gas fields; Existing fault prediction techniques mostly rely on structured sensor time series data, and it is difficult to integrate expert experience and forward-looking risk information contained in unstructured text such as geological reports and operation instructions, resulting in incomplete information input of the prediction model; at the same time, traditional models are usually static and cannot dynamically adjust their prediction logic according to real-time changing working conditions, and their prediction accuracy and robustness decrease significantly when facing geological environment mutations or operation adjustments; in addition, existing methods often provide a single fault probability prediction value, lack of quantitative evaluation of the uncertainty of the prediction result itself, and are difficult to support reliable decision-making in high-risk scenarios; Therefore, there is an urgent need for a new fault prediction scheme that can integrate multi-source heterogeneous data, realize adaptive prediction and quantify uncertainty.

[0003] The above information disclosed in the above background section is only used to strengthen the understanding of the background of the present disclosure, and therefore it can include information that does not constitute prior art known to those of ordinary skill in the art. SUMMARY

[0004] To solve the above technical problems, the present application discloses a wellhead device fault prediction system based on narrowband Internet of Things, in particular, the technical scheme of the present application comprises: A data acquisition module for acquiring unstructured text data and structured time series data of a target wellhead device; A factor quantization module for determining a working condition mutation factor based on the unstructured text data; A dynamic weight module for determining a dynamic weight factor according to the working condition mutation factor; A fault prediction module for determining a direct failure probability in response to the dynamic weight factor and based on the structured time series data; An uncertainty quantization module for determining an uncertainty index based on multiple prediction results of the fault prediction module; A risk assessment module for determining a comprehensive risk index in combination with the direct failure probability and the uncertainty index; A decision response module for generating a graded early warning instruction according to the comprehensive risk index and a preset risk threshold.

[0005] Further, the factor quantization module is configured to determine a working condition mutation factor based on the unstructured text data, and the determining comprises: processing the unstructured text data by using a pre-trained language model to extract key semantic units; for the key semantic units, obtaining corresponding reference influence weights, confidence degrees, and intensity modifier scores; combining the reference influence weights, the confidence degrees, and the intensity modifier scores to calculate the working condition mutation factor.

[0006] Further, the fault prediction module adopts a physical information neural network model, and a loss function of the physical information neural network model comprises a data-driven loss and a physical information loss.

[0007] Further, the dynamic weight factor is used to adjust the weight of the physical information loss in the loss function. The dynamic weight module is specifically configured to: invoke the working condition mutation factor; calculate the dynamic weight factor by using a pre-set adaptive adjustment function.

[0008] Further, the physical information loss is calculated by substituting a prediction result of the fault prediction module into a pre-set standard erosion model describing an erosion and wear process of the target wellhead device, to quantify a deviation degree of the prediction result from the standard erosion model.

[0009] Further, the uncertainty quantization module is configured to determine an uncertainty index based on multiple prediction results of the fault prediction module, and the determining comprises: in an inference stage, multiple times of invoking a Dropout layer in the fault prediction module and performing forward propagation to obtain a group of prediction results; based on sample variances of the prediction result set, performing normalization processing to generate the uncertainty index.

[0010] Further, the risk assessment module is configured to determine a comprehensive risk index by combining the direct fault probability and the uncertainty index, and the determining comprises: invoking the direct fault probability, the uncertainty index, and a pre-set uncertainty amplification coefficient; based on a product of the uncertainty index and the uncertainty amplification coefficient, weighting and amplifying the direct fault probability; outputting a result after the weighting and amplification as the comprehensive risk index.

[0011] Further, the decision response module is specifically configured to: compare the comprehensive risk index with pre-set first and second risk thresholds; output a first-level early warning instruction when the comprehensive risk index is lower than a first risk threshold value; output a second-level early warning instruction when the comprehensive risk index is between the first risk threshold value and a second risk threshold value; output a third-level early warning instruction when the comprehensive risk index is higher than the second risk threshold value.

[0012] Compared with the prior art, the present application has the following beneficial effects: 1. The present application innovatively integrates unstructured text containing expert experience and sensor time series data. The instructions or geological information in the text that predict future changes are quantified into working condition mutation factors by the factor quantization module, which are used for dynamic regulation and prediction model. This design enables the system to perceive potential risks that have not yet appeared in sensor data in advance, realizes the transition from passive response to active foresight, significantly enhances the foresight, adaptability and timeliness of fault prediction, and realizes earlier warning.

[0013] 2. The present application uses a physical information neural network as the prediction core, and introduces the text quantization factor into the loss function weight adjustment. When the text information predicts a smooth working condition, the model focuses on learning the historical data pattern; when it predicts that a dramatic change will occur, it automatically switches to rely more on universal physical laws for constraint and prediction. This dynamic balance mechanism ensures that the model can still make reasonable predictions in extreme working conditions with sparse training data, significantly improving the robustness and generalization ability of the model.

[0014] 3. The present application introduces an uncertainty quantization module, which evaluates the confidence of the model's own prediction results through the Monte Carlo Dropout method. It not only outputs a single fault probability, but also provides an uncertainty index to represent the reliability of the current prediction. When the system faces a working condition that has never been seen or is in a critical state, even if the output fault probability is not high, the abnormally high uncertainty index can effectively remind the operation and maintenance personnel, thereby avoiding misjudgment caused by excessive reliance on the model's single-point prediction, and improving the safety of the decision-making process.

[0015] 4. The present application constructs a comprehensive risk assessment model, which integrates the direct fault probability and the uncertainty index of the prediction result to generate a comprehensive risk index. This index can more comprehensively reflect the real risk, effectively distinguish different risk scenarios such as low probability and high uncertainty, and high probability and high certainty, and overcome the limitations of single probability index. Based on the early warning instruction generated by the index and the grading threshold, the automation and standardization of decision response are realized, making the operation and maintenance decision more accurate, reasonable and efficient. BRIEF DESCRIPTION OF DRAWINGS

[0016] The present application will be further explained below in conjunction with the drawings and examples: Figure 1is a system structure diagram of the present application. DETAILED DESCRIPTION

[0017] In order to make the purpose, technical solutions and advantages of the present application more clear, the present application is further described in detail below in combination with specific embodiments.

[0018] Embodiment 1 Please refer to Figure 1 A wellhead device fault prediction system based on narrowband Internet of Things, comprising: A data acquisition module for acquiring unstructured text data and structured time series data of a target wellhead device; A factor quantification module for determining a working condition mutation factor based on the unstructured text data; A dynamic weight module for determining a dynamic weight factor according to the working condition mutation factor; A fault prediction module for determining a direct fault probability in response to the dynamic weight factor and based on the structured time series data; An uncertainty quantification module for determining an uncertainty index based on multiple prediction results of the fault prediction module; A risk assessment module for determining a comprehensive risk index in combination with the direct fault probability and the uncertainty index; A decision response module for generating a graded early warning instruction according to the comprehensive risk index and a preset risk threshold; The present application provides a wellhead device fault prediction system based on narrowband Internet of Things, aiming to integrate multi-source heterogeneous information and realize forward-looking and adaptive prediction and early warning of wellhead device performance degradation; the system builds a technical closed loop from data acquisition, risk quantification, dynamic prediction to decision response; In a specific implementation scenario, the system comprises: A data acquisition module which acquires unstructured text data and structured time series data of a target wellhead device in parallel through an industrial Internet of Things interface and a database interface; wherein the unstructured text data refers to text information such as real-time generated geological analysis reports and drilling operation daily reports, which are used to capture expert opinions and operation instructions that may indicate future working condition changes; the structured time series data refers to physical quantities such as pressure , temperature , flow rate v, etc. continuously collected through a wellhead sensor array, which are used to represent the current running state of the equipment; the module uses an NB-IoT module to remotely and low-power transmit the structured time series data (such as pressure, temperature, flow rate, etc.) of the wellhead device to the subsequent processing module through an NB-IoT network.

[0019] The factor quantization module receives text data acquired by the data acquisition module and determines a condition mutation factor based on this unstructured text data. This factor serves as the core driving input for subsequent dynamic adaptive correction of the model. The dynamic weighting module determines the operating condition mutation factors in real time based on the factor quantification module. Determine a dynamic weighting factor This is used to weigh the degree to which the predictive model depends on historical data experience and universal physical laws; The fault prediction module, which is a neural network model that integrates physical information, responds to the dynamic weight factors output by the dynamic weight module. Based on the structured time-series data acquired by the data acquisition module, a direct failure probability is determined. ; The uncertainty quantification module determines an uncertainty index based on the consistency of the multiple prediction results obtained by performing multiple random inferences on the fault prediction module. ; The risk assessment module uses a risk fusion model to convert the direct failure probability output by the failure prediction module. Uncertainty index output by the uncertainty quantification module Combined, a comprehensive risk index is determined. ; The decision response module is based on the comprehensive risk index determined by the risk assessment module. Compare with preset risk thresholds to generate and output tiered early warning instructions; This embodiment constructs an end-to-end intelligent fault prediction system through the collaborative work of the above modules. It not only utilizes easily quantifiable structured sensor data, but also innovatively quantifies the expert experience and early warning information contained in unstructured text, and uses them as key factors for dynamically adjusting the prediction model. This design enables the system to perceive potential risks brought about by changes in operating instructions or geological environment in advance, realizing the transformation from passive response to proactive prediction, significantly improving the accuracy, robustness and timeliness of fault prediction, and ultimately forming a closed-loop, forward-looking intelligent operation and maintenance decision-making system.

[0020] Example 2: The factor quantification module is used to determine the factors that cause sudden changes in operating conditions based on unstructured text data, including: Unstructured text data is processed using pre-trained language models to extract key semantic units; For key semantic units, obtain the corresponding baseline influence weight, confidence level, and intensity modifier score; The working condition mutation factor is calculated by combining the baseline impact weight, confidence level, and intensity modifier score. This embodiment is a specific implementation of the factor quantification module based on Embodiment 1; its underlying logic lies in designing a method for mathematically quantifying risk descriptions in natural language text. The factor quantization module processes unstructured text data through a pre-trained language model to extract key semantic units. This pre-trained language model is a deep learning model that has been pre-trained on massive amounts of text in this field, enabling it to have contextual understanding and semantic feature extraction capabilities. The model receives texts such as geological analysis reports and work instructions, identifies and extracts key semantic units related to equipment risks, such as the detection of high-pressure gas layers or plans to carry out acidizing operations. For each extracted key semantic unit, obtain its corresponding baseline influence weight, confidence score, and intensity modifier score; baseline influence weight It refers to the first The inherent risk level of an event corresponding to a semantic unit; it is defined as a dimensionless normalized risk coefficient that is pre-calibrated by domain experts or obtained through statistical analysis of historical data, with a value range between [0,1], used to characterize the degree of objective impact of such events on the failure rate; Confidence This refers to the ability of a pre-trained language model to recognize the first... The degree of certainty for each semantic unit is the probability value directly output by the model when performing semantic recognition tasks, and is a dimensionless value in the interval [0,1]; intensity modifier score It refers to the quantitative scoring of words that modify the severity of semantic units. It is determined based on a modifier-score mapping table pre-built by domain experts based on an experience knowledge base, and is a dimensionless value. To integrate the above multi-dimensional information, this embodiment uses the following heuristic risk assessment model to calculate the operating condition mutation factor. , It is the total number of key semantic units identified in a single text analysis; the formula is as follows: The total risk in a single analysis of a text is the sum of the risks contributed by all identified key risk events; the risk of each event is determined by its inherent impact. The certainty of the model's recognition and the severity described in the text These three factors jointly determine the final operating condition mutation factor. It is a dimensionless scalar; It should be noted that the linear accumulation model used in this embodiment is... This is a simple and effective implementation method; in other implementation methods, nonlinear functions or attention mechanisms can also be used to fuse multiple risk events in order to capture the complex coupling effects that may exist between them. Compared to traditional methods that rely solely on structured data, this embodiment enables the system to capture potential or impending risks that are not yet reflected in sensor data by deeply mining information in unstructured text. This method enhances the system's foresight and adaptability to sudden operating conditions, achieving earlier warnings of risks.

[0021] Example 3: The fault prediction module adopts a physical information neural network model. The loss function of the physical information neural network model includes data-driven loss and physical information loss. The dynamic weighting factor is used to adjust the weight of physical information loss in the loss function; The dynamic weight module is specifically used for: Call the operating condition mutation factor; The dynamic weighting factor is calculated using a preset adaptive adjustment function; Physical information loss is calculated by substituting the prediction results of the fault prediction module into a preset standard erosion model that describes the erosion and wear process of the target wellhead device, so as to quantify the degree of deviation of the prediction results from the standard erosion model. This embodiment is based on Embodiment 1 and is a collaborative implementation of the fault prediction module, the dynamic weight module and their interaction. Its core lies in constructing a physical information neural network that is controlled in real time by external text information, thereby achieving a dynamic balance between the flexibility of data-driven approaches and the universality of physical laws. The fault prediction module employs a physical information neural network model. This model incorporates partial differential equations describing the physical process as regularization terms into the loss function, thereby constraining the model's solution space. In this embodiment, the model uses a long short-term memory network as its infrastructure to learn the wellhead pressure. ,temperature Dynamic patterns in structured time-series data directly predict instantaneous material loss rates. ; To obtain the final direct failure probability This module further analyzes the predicted instantaneous material loss rate. In the future, within a predetermined time window The total predicted material loss is obtained by integrating and accumulating the components. The total material loss is obtained. Material wear limit thresholds are preset based on equipment design specifications or industry safety standards. By using an S-shaped reliability function, such as a logic function, the reliability can be... Mapped to the probability of direct failure within a specific future time window For example ,in The constant is related to the slope of the function; the total loss function of this model. It consists of two parts: The optimization objective of the model is not only to minimize the data-driven loss. At the same time, it is also necessary to minimize the loss of physical information. ; Defined as the instantaneous material loss rate predicted by the model. The reference instantaneous loss rate is indirectly calculated or calibrated using high-frequency sensor data. The mean square error between them is expressed in units of mass / time (mass / time). This represents the model's pursuit of historical fidelity; The core task of the dynamic weighting module is to calculate the dynamic weighting factors. This factor is used to adjust the weight of physical information loss in the total loss function; this module calls the operating condition mutation factor calculated by the factor quantization module. The dynamic weighting factor is calculated using a preset adaptive adjustment function; this function is: It is the baseline physical loss weight, a dimensionless hyperparameter whose value is determined during the model development phase by performing grid search optimization on the data validation set under stable operating conditions. It is a dimensionless sensitivity coefficient, also a hyperparameter, used to control... Follow The degree of change; introducing the hyperbolic tangent function. This is to smooth out the adjustment process and... The effect is applied in a bounded manner. superior; Physical information loss The calculation aims to quantify the degree of deviation of the prediction results from physical laws; the loss of physical information is calculated by substituting the prediction results of the fault prediction module into a preset standard erosion model that describes the erosion and wear process of the target wellhead device; the specific formula is as follows: in, It is the predicted value directly output by the fault prediction module; It is a standard erosion model describing the rate of material loss and is the authoritative standard in the field of materials science. It is based on real-time fluid pressure obtained by a data acquisition module. Flow rate and the inherent material properties obtained by consulting material handbooks or experimental measurements. The function, for example, the model can take a specific form such as ,in and It is related to material properties Relevant empirical coefficients; The calculation results are also in units of mass / time, thus ensuring... The internal dimensions are consistent, making its final dimension mass-time (mass / time). ,and The dimensions are completely unified; This collaborative design introduces physical information loss. This ensures that the prediction model's results do not deviate from basic physical laws even in extreme operating conditions where training data is sparse, thus improving the model's robustness and generalization ability; by introducing a factor derived from sudden changes in operating conditions... Driven dynamic weighting factor This makes the model adaptive; when the text information indicates that the working conditions are stable, the model trusts historical data patterns more; when it indicates that drastic changes are about to occur, the model automatically switches to making predictions that rely more on universal physical laws.

[0022] Example 4: The uncertainty quantification module is used to determine uncertainty indicators based on multiple prediction results from the fault prediction module, including: During the inference phase, the Dropout layer in the fault prediction module is used multiple times and forward propagation is performed to obtain a set of prediction results. Based on the sample variance of the prediction result set, normalization is performed to generate an uncertainty index. This embodiment is a specific implementation of the uncertainty quantification module based on embodiment 1. Its purpose is to provide the model with confidence in its own prediction results for the final decision of the system. The uncertainty quantification module employs the Monte Carlo Dropout method; during the inference phase, the Dropout layer in the fault prediction module is enabled multiple times and forward propagation is performed to obtain a set of prediction results; in conventional neural network inference, the Dropout layer is disabled; however, in this embodiment, even during the inference phase, it is forcibly and repeatedly enabled, for example... The Dropout layer in the network is enabled for the second time; for the input data in the same step, through... The forward propagation is performed on each different subnetwork to obtain a set containing A set of prediction results for different predicted values ,in It is the first The direct failure probability obtained from the first forward propagation; Based on the sample variance of the prediction result set, normalization is performed to generate an uncertainty index. To transform the original variance into a relative index with business meaning and easy comparison, this embodiment uses the following formula for calculation: in, yes The mean of the prediction results; It is the dimensionless uncertainty index of the final output; in the denominator It is the preset reference variance, which is the statistical average of the model output variance under a large number of historical stable operating conditions, used to scale the original variance to a relative scale with clear business meaning; This embodiment introduces uncertainty quantification to add self-evaluation capabilities to the fault prediction system; when the system faces an unseen or critical operating condition, even if it might output a seemingly normal low failure probability, the uncertainty index... It may rise abnormally; this can effectively remind operations and maintenance personnel that the current prediction results are unreliable, thereby avoiding misjudgments caused by over-reliance on single-point predictions of the model and improving the security of the entire operations and maintenance decision-making process.

[0023] Example 5: The risk assessment module is used to determine a comprehensive risk index by combining direct failure probability and uncertainty indicators, including: Invoke the direct failure probability, uncertainty index, and preset uncertainty amplification factor; The direct failure probability is weighted and amplified based on the product of the uncertainty index and the uncertainty amplification factor. The weighted and amplified result is output as a comprehensive risk index. This embodiment is a specific implementation of the risk assessment module based on embodiment 1. Its purpose is to build a comprehensive risk assessment model that is more comprehensive than a single failure probability and closer to the actual operation and maintenance decision-making logic. The risk assessment module combines direct failure probability and uncertainty indicators to determine a comprehensive risk index; it calls upon three key inputs: the direct failure probability output by the failure prediction module. Uncertainty index calculated by the uncertainty quantification module And a preset uncertainty amplification factor. Uncertainty amplification factor It is a dimensionless, adjustable parameter that is set by the operations and maintenance management department based on its risk tolerance. The direct failure probability is weighted and amplified based on the product of the uncertainty index and the uncertainty amplification factor; this is achieved through a risk fusion model, the calculation formula of which is as follows: Overall Risk Index Direct failure probability Based on this, and through an amplification factor driven by uncertainty. Adjust it; when the prediction is very certain. Approaching 0, the amplification factor is close to 1. Approximately equal to When the uncertainty of the forecast increases significantly Increasing the amplification factor will make it greater than 1, thus adjusting the original failure probability upwards; The weighted and amplified result is output as the comprehensive risk index; the calculated comprehensive risk index It is a dimensionless value that incorporates the probability of an event occurring. and the certainty of the perception of this possibility The linear scaling model here It is an intuitive and easy-to-explain risk fusion strategy; It should be noted that, although the direct failure probability The range of values ​​is However, the overall risk index Theoretically, there is no upper limit; when A value greater than 1 indicates that the model has extremely low certainty regarding the prediction result, even if the base probability is low. Even if the risk level is not high, its overall risk should still be considered extremely high and should immediately attract the highest level of attention from operations and maintenance personnel. This design ensures that highly uncertain prediction signals are not ignored. In other implementations, nonlinear fusion models such as exponential or piecewise functions can also be designed according to specific risk preferences. This embodiment achieves a deeper understanding of risk assessment; a low-probability but highly uncertain prediction may pose a greater potential risk than a moderate-probability but highly certain prediction; this embodiment calculates a comprehensive risk index. This allows for accurate quantification and differentiation of these differences, making subsequent early warning and operation and maintenance decisions more precise and reasonable.

[0024] Example 6: The decision response module is specifically used for: The comprehensive risk index is compared with the preset first risk threshold and second risk threshold. When the comprehensive risk index is lower than the first risk threshold, a first-level warning instruction is issued; When the comprehensive risk index is between the first risk threshold and the second risk threshold, a second-level early warning instruction is output. When the comprehensive risk index is higher than the second risk threshold, a third-level warning instruction is issued. This embodiment is a specific implementation of the decision response module based on Embodiment 1; its purpose is to convert the continuous and quantitative comprehensive risk index calculated by the upstream module into a more comprehensive and quantitative one. This is transformed into discrete, explicit, and directly operational hierarchical early warning instructions; The decision response module will output the comprehensive risk index from the risk assessment module. Compared with the preset first risk threshold Second risk threshold The two thresholds were determined through statistical methods such as receiver operation characteristic curve analysis on a large amount of historical data. The technical principle behind their setting is to achieve a balance between false alarm rate and false negative rate that meets specific operation and maintenance requirements. Based on the comparison results, the system automatically triggers different levels of response strategies: When the comprehensive risk index Below the first risk threshold When this occurs, a Level 1 warning instruction is issued; in practice, this instruction may correspond to maintaining routine monitoring. When the comprehensive risk index Between the first risk threshold With the second risk threshold During this period, a second-level warning instruction is issued; this instruction may correspond to increasing the data collection frequency or suggesting that attention be paid to this issue. When the comprehensive risk index Above the second risk threshold When this occurs, a Level 3 warning instruction will be issued; this instruction is the highest level of alarm, which may correspond to triggering a high-risk alarm and recommends a special inspection. This embodiment achieves automation, standardization, and differentiation of decision-making response by establishing clear, data-driven hierarchical thresholds. It can match the corresponding level of operation and maintenance resources according to the actual severity of the risk, which not only avoids the waste of resources caused by excessive alarms, but also ensures that the highest priority and most timely response can be given to truly high-risk signals.

[0025] The above are merely preferred embodiments of the present invention and are not intended to limit the scope of protection of the present invention; any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A wellhead device fault prediction system based on narrowband Internet of Things, characterized in that, include: The data acquisition module is used to acquire unstructured text data and structured time-series data of the target wellhead device; The factor quantification module is used to determine the factors that cause sudden changes in operating conditions based on unstructured text data. The dynamic weighting module is used to determine the dynamic weighting factor based on the sudden change factor of the operating condition; The fault prediction module is used to determine the probability of direct failure in response to dynamic weighting factors and based on structured time series data. The uncertainty quantification module is used to determine uncertainty indicators based on the multiple prediction results of the fault prediction module. The risk assessment module is used to determine a comprehensive risk index by combining direct failure probability and uncertainty indicators; The decision response module is used to generate tiered early warning instructions based on the comprehensive risk index and preset risk thresholds.

2. The wellhead device fault prediction system based on narrowband Internet of Things according to claim 1, characterized in that, The factor quantification module is used to determine the factors of sudden changes in operating conditions based on unstructured text data, including: Unstructured text data is processed using pre-trained language models to extract key semantic units; For key semantic units, obtain the corresponding baseline influence weight, confidence level, and intensity modifier score; The operating condition mutation factor is calculated by combining the baseline impact weight, confidence level, and intensity modifier score.

3. The wellhead device fault prediction system based on narrowband Internet of Things according to claim 1, characterized in that, The fault prediction module adopts a physical information neural network model, and the loss function of the physical information neural network model includes data-driven loss and physical information loss.

4. The wellhead device fault prediction system based on narrowband Internet of Things according to claim 3, characterized in that, The dynamic weighting factor is used to adjust the weight of physical information loss in the loss function; The dynamic weight module is specifically used for: Call the operating condition mutation factor; The dynamic weighting factor is calculated using a preset adaptive adjustment function.

5. A wellhead device fault prediction system based on narrowband Internet of Things according to claim 3, characterized in that, The loss of physical information is calculated by substituting the prediction results of the fault prediction module into a preset standard erosion model that describes the erosion and wear process of the target wellhead device, so as to quantify the degree of deviation of the prediction results from the standard erosion model.

6. A wellhead device fault prediction system based on narrowband Internet of Things according to claim 1, characterized in that, The uncertainty quantification module is used to determine uncertainty indicators based on multiple prediction results from the fault prediction module, including: During the inference phase, the Dropout layer in the fault prediction module is used multiple times and forward propagation is performed to obtain a set of prediction results. Based on the sample variance of the prediction result set, normalization is performed to generate an uncertainty index.

7. A wellhead device fault prediction system based on narrowband Internet of Things according to claim 1, characterized in that, The risk assessment module is used to determine a comprehensive risk index by combining direct failure probability and uncertainty indicators, including: Invoke the direct failure probability, uncertainty index, and preset uncertainty amplification factor; The direct failure probability is weighted and amplified based on the product of the uncertainty index and the uncertainty amplification factor. The weighted and amplified result is output as a comprehensive risk index.

8. A wellhead device fault prediction system based on narrowband Internet of Things according to claim 1, characterized in that, The decision response module is specifically used for: The comprehensive risk index is compared with the preset first risk threshold and second risk threshold. When the comprehensive risk index is lower than the first risk threshold, a first-level warning instruction is issued; When the comprehensive risk index is between the first risk threshold and the second risk threshold, a second-level early warning instruction is output. When the comprehensive risk index is higher than the second risk threshold, a third-level warning instruction is issued.

Images