Oil and gas production data monitoring system and method

Abstract

The invention relates to the technical field of oil and gas production, in particular to an oil and gas production data monitoring system and method. The system comprises a data acquisition module used for acquiring multivariable time sequence data and preprocessing the multivariable time sequence data to generate a multivariable observation vector sequence; the reconstruction and anomaly scoring module is used for acquiring a reconstruction output vector and generating a Mama anomaly score and contribution variable sorting list; the quantile prediction module is used for constructing a historical context sequence and outputting a quantile prediction sequence; the alarm fusion judgment module is used for calculating a prediction deviation score and outputting an alarm level; and the result generation module is used for constructing an alarm evidence packet and generating a structured processing result. According to the invention, by introducing a dual-evidence fusion monitoring architecture of reconstruction residual error anomaly detection and quantile probability prediction deviation, real-time anomaly identification and early warning can be stably realized in the scene of oil and gas production multivariable data quality fluctuation and frequent working condition change, and the alarm accuracy is significantly improved.

Description

Technical Field

[0001] This invention relates to the field of oil and gas production technology, and in particular to an oil and gas production data monitoring system and method. Background Technology

[0002] Current oil and gas production data monitoring typically relies on SCADA / DCS and historical databases to collect multi-variable time-series data such as pressure, temperature, flow rate, water content, valve opening, and equipment electrical parameters. Alarms and trend analysis are then implemented on the monitoring platform using fixed thresholds, rule bases, or simple statistical models. Some systems introduce single time-series prediction or anomaly detection algorithms to predict key indicators or identify deviations in real-time data. Simultaneously, alarm interpretation and handling decisions are made manually by combining text information such as shift reports, operation records, and maintenance records. Finally, monitoring results are output in the form of alarm lists, daily reports, or work orders.

[0003] The shortcomings of the above-mentioned technical solutions are as follows: on the one hand, the quality problems of multivariate data, such as time alignment, unit conversion, missing data, jumps, drift, and bottlenecks, are not uniformly quantified and explicitly incorporated into the monitoring criteria, resulting in false alarms and false negatives coexisting in thresholds and single models under complex working conditions. On the other hand, existing methods often rely only on a single abnormal score or a single predicted value for alarms, lacking a mechanism to integrate "dynamic anomaly evidence based on reconstructed residuals" and "probability deviation evidence based on quantile prediction bands", making it difficult to simultaneously take into account real-time performance and early warning capabilities. Summary of the Invention

[0004] To address the technical problems existing in the background art, this invention proposes an oil and gas production data monitoring system and method.

[0005] This invention proposes an oil and gas production data monitoring system, comprising: The data acquisition module is used to acquire multivariate time-series data of target oil and gas production and preprocess it to generate a multivariate observation vector sequence; The reconstruction and anomaly scoring module is used to construct a sliding window input sequence based on a multivariate observation vector sequence, input the sliding window input sequence into the Mamba-2 model to obtain the reconstruction output vector and construct a dynamic residual vector, and generate Mamba anomaly scores and a ranking list of contributing variables based on the dynamic residual vector; The quantile prediction module is used to obtain key indicator observations and construct historical context sequences based on multivariate observation vector sequences, and output quantile prediction sequences through the Chronos-Bolt model. The alarm fusion judgment module is used to calculate the predicted deviation score based on the quantile prediction sequence and the key indicator observation value, fuse the Mamba anomaly score and the predicted deviation score to generate a fused alarm score, and output the alarm level. The results generation module is used to construct an alarm evidence package based on alarm level, fused alarm score, Mamba anomaly score, predicted deviation score, ranking list of contributing variables, and quantile prediction sequence. The alarm evidence package is then input into the industry big language model to generate structured processing results.

[0006] Furthermore, multivariate time-series data on target oil and gas production are acquired and preprocessed to generate a multivariate observation vector sequence, including: Collect multivariate time-series data, including pressure, temperature, flow rate, moisture content, valve opening, current, and voltage, and perform time alignment on the multivariate time-series data; Based on the aligned multivariate time series data, a multivariate observation vector is constructed at each time step, and the multivariate observation vectors of all time steps are combined to form a multivariate observation vector sequence.

[0007] Furthermore, a sliding window input sequence is constructed based on the multivariate observation vector sequence. This sliding window input sequence is then input into the Mamba-2 model to obtain the reconstructed output vector and construct a dynamic residual vector. Based on the dynamic residual vector, Mamba anomaly scores and a ranking list of contributing variables are generated, including: Construct a sliding window input sequence based on a preset window length; The Mamba-2 model is used to perform a reconstruction calculation based on structured state space duality on the sliding window input sequence to obtain the reconstruction output vector corresponding to the current time step; Calculate the dynamic residual vector based on the multivariate observation vector and the reconstructed output vector corresponding to the current time step; Mamba anomaly scores are generated based on dynamic residual vectors, and a ranking list of contributing variables is generated based on the component magnitudes of the dynamic residual vectors.

[0008] Furthermore, the Mamba-2 model is used to perform a reconstruction calculation based on structured state-space duality on the sliding window input sequence to obtain the reconstruction output vector corresponding to the current time step, including: For each time step in the sliding window input sequence, the multivariate observation vector is processed by input projection, which maps the multivariate observation vector to a projection vector. Perform one-dimensional depthwise separable convolution on the projection vector to generate a convolution output vector; Based on the convolution output vector, the input driving vector, stride vector, input mapping vector, output mapping vector, and gate vector are generated through linear transformation. Set the initial state vector corresponding to the start time of the sliding window, and recursively update the state vector for each time step in the sliding window in chronological order. Generate a state-output vector based on the output mapping vector, the state vector, and the input driving vector; The state output vector is multiplied element by element to obtain the gated output vector, and then the reconstructed output vector is obtained through linear transformation.

[0009] Furthermore, the state vector is obtained by recursively updating the state at each time step within the sliding window in chronological order, including: For each time step, a state recursion coefficient vector is generated based on the step size vector, and an input injection vector is generated based on the state recursion coefficient vector and the input mapping vector. The sliding window is divided into multiple consecutive blocks. For each block, the cumulative product vector and the cumulative sum vector within the block are calculated based on the state recursion coefficient vector and the input injection vector, respectively. The state vector for each time step is generated based on the intra-block cumulative product vector and the intra-block cumulative sum vector.

[0010] Furthermore, based on the multivariate observation vector sequence, key indicator observations are constructed and historical context sequences are built. The Chronos-Bolt model then outputs quantile prediction sequences, including: The key indicator observation values ​​for each time step are calculated based on the multivariate observation vector, and the key indicator observation sequence is constructed using the key indicator observation values ​​for each time step. A historical context sequence is constructed based on the observation sequence of key indicators. The historical context sequence is input into the Chronos-Bolt model, and the output is a quantile prediction sequence within a preset prediction step size range. The quantile prediction sequence includes the P10 quantile prediction value, P50 quantile prediction value and P90 quantile prediction value corresponding to each prediction step size.

[0011] Furthermore, the predicted deviation score is calculated based on the quantile prediction sequence and the key indicator observations, including: When the observed value of the key indicator is greater than or equal to the predicted value of the 90th quantile, the prediction deviation score is the ratio of the difference between the observed value of the key indicator and the predicted value of the 90th quantile to the difference between the predicted value of the 90th quantile and the predicted value of the 50th quantile. When the observed value of the key indicator is less than or equal to the predicted value of the P10 quantile, the prediction deviation score is the ratio of the difference between the predicted value of the P10 quantile and the observed value of the key indicator to the difference between the predicted value of the P50 quantile and the predicted value of the P10 quantile. When the observed value of a key indicator is greater than the predicted value of the P10 quantile but less than the predicted value of the P90 quantile, the prediction deviation score is 0.

[0012] Furthermore, the Mamba anomaly score and the predicted deviation score are fused to generate a fused alarm score, and the alarm level is output, including: The fused alarm score is obtained by weighted summation of Mamba anomaly score and predicted deviation score; Set the fusion alarm threshold and the prediction deviation threshold. When the fusion alarm score is greater than the fusion alarm threshold and the prediction deviation score is greater than the prediction deviation threshold, output a critical alarm. When the fusion alarm score is less than the fusion alarm threshold and the prediction deviation score is less than the prediction deviation threshold, output no alarm. Otherwise, output a warning alarm.

[0013] Furthermore, the alarm evidence package is input into the industry's large language model to generate structured processing results, including: A pre-defined knowledge base is constructed, which includes the process flow of oil and gas production, process specifications, alarm and interlock logic description documents, alarm handling cards, emergency response plans, equipment manuals, equipment maintenance manuals, equipment ledgers and spare parts lists, well history records, operation measure records, historical alarm records, fault case library, and operation permit documents; The system retrieves process procedures, alarm handling cards, equipment manuals, and well history records that match the alarm evidence package from a preset knowledge base, forming a set of retrieval results. The knowledge base is constructed based on historical oil and gas production data. The alarm evidence package and the retrieval result set are input into the industry big language model for processing to generate structured handling results. The industry big language model is built based on the Tongyi Qianwen 2.5 model and has been fine-tuned. The structured handling results include anomaly description, root cause candidate list, handling step list and risk warning.

[0014] This invention provides a method for monitoring oil and gas production data, comprising the following steps: S1. Acquire multivariate time-series data of target oil and gas production and preprocess it to generate a multivariate observation vector sequence; S2. Construct a sliding window input sequence based on the multivariate observation vector sequence, input the sliding window input sequence into the Mamba-2 model to obtain the reconstructed output vector and construct a dynamic residual vector, and generate the Mamba anomaly score and contribution variable ranking list based on the dynamic residual vector; S3. Based on the multivariate observation vector sequence, construct the observation values ​​of key indicators and build the historical context sequence, and output the quantile prediction sequence through the Chronos-Bolt model; S4. Calculate the predicted deviation score based on the quantile prediction sequence and the observed values ​​of key indicators, fuse the Mamba anomaly score and the predicted deviation score to generate a fused alarm score, and output the alarm level. S5. Based on alarm level, fused alarm score, Mamba anomaly score, predicted deviation score, contribution variable ranking list and quantile prediction sequence, construct alarm evidence package, input alarm evidence package into industry big language model to generate structured processing results.

[0015] The voiceprint recognition method and system proposed in this invention have the following beneficial technical effects: (1) This invention forms a unified multivariate observation vector sequence by performing time alignment on multivariate time series data. On this basis, the Mamba-2 model is used to reconstruct the sliding window sequence to obtain the reconstructed output vector. The dynamic residual vector, Mamba anomaly score and contribution variable ranking list are calculated by the actual observation vector and the reconstructed output vector. At the same time, the key indicator observation sequence is established and input into the Chronos-Bolt model to output the P10 / P50 / P90 quantile prediction sequence. The prediction deviation score is calculated based on the quantile prediction band and the current key indicator observation value, and the advance amount is determined. Thus, a dual evidence monitoring system of "dynamic residual anomaly evidence + probability prediction deviation evidence" is formed, realizing the unity of real-time identification and early warning of anomalies under complex working conditions.

[0016] (2) After alarm fusion judgment, the present invention constructs an alarm evidence package and combines it with the process procedures, alarm handling cards, equipment manuals and well history records retrieved from the preset knowledge base to form a set of search results. The combined input adopts the industry big language model generated by retrieval enhancement to output structured handling results, so that the alarm is upgraded from a simple numerical limit exceeding the limit to an interpretable diagnosis and standardized handling output based on the evidence chain. This can reduce false alarms and false alarms and improve alarm consistency and auditability, reduce manual interpretation and experience dependence, and realize the closed loop of monitoring-diagnosis-handling. Attached Figure Description

[0017] Figure 1 This is a schematic diagram of an oil and gas production data monitoring system according to the present invention; Figure 2 This is a flowchart of an oil and gas production data monitoring method according to the present invention. Detailed Implementation

[0018] Embodiments of the present invention are described in detail below. Examples of these embodiments are illustrated in the accompanying drawings, wherein the same or similar symbols denote the same or similar elements or elements having the same or similar functions throughout. The embodiments described below with reference to the accompanying drawings are exemplary and are only used to explain the present invention, and should not be construed as limiting the present invention.

[0019] like Figure 1 The oil and gas production data monitoring system shown includes: Data acquisition module 101 is used to acquire multivariate time series data of target oil and gas production and preprocess it to generate multivariate observation vector sequence; The reconstruction and anomaly scoring module 102 is used to construct a sliding window input sequence based on the multivariate observation vector sequence, input the sliding window input sequence into the Mamba-2 model to obtain the reconstruction output vector and construct a dynamic residual vector, and generate Mamba anomaly scores and a ranking list of contributing variables based on the dynamic residual vector; Quantile prediction module 103 is used to obtain key indicator observations and construct historical context sequences based on multivariate observation vector sequences, and output quantile prediction sequences through Chronos-Bolt model. The alarm fusion judgment module 104 is used to calculate the predicted deviation score based on the quantile prediction sequence and the key indicator observation value, fuse the Mamba anomaly score and the predicted deviation score to generate a fused alarm score, and output the alarm level. The result generation module 105 is used to construct an alarm evidence package based on alarm level, fused alarm score, Mamba anomaly score, predicted deviation score, ranking list of contributing variables, and quantile prediction sequence, and input the alarm evidence package into the industry big language model to generate structured processing results.

[0020] In one embodiment of the present invention, multivariate time-series data of target oil and gas production are acquired and preprocessed to generate a multivariate observation vector sequence, as follows: Collect multivariate time-series data, including pressure, temperature, flow rate, moisture content, valve opening, current, and voltage. The multivariate time-series data is represented as an original sequence sampled at unequal intervals. Time alignment is performed on the multivariate time-series data. Based on aligned multivariate time series data At each time step Constructing multivariate observation vectors ,in, The number of observed dimensions. For the first Each observation dimension at time step The aligned and converted observations, and , For unit conversion factors, For unit conversion bias, These are the time-aligned observations; All time-step multivariate observation vectors Combine them to form a multivariate observation vector sequence .

[0021] In one embodiment of the present invention, a sliding window input sequence is constructed based on a multivariate observation vector sequence. The sliding window input sequence is then input into the Mamba-2 model to obtain a reconstructed output vector and construct a dynamic residual vector. Based on the dynamic residual vector, a Mamba anomaly score and a ranking list of contributing variables are generated, as follows: Based on the preset window length Constructing a sliding window input sequence , , ; Using the Mamba-2 model, a reconstruction calculation based on structured state-space duality is performed on the sliding window input sequence to obtain the current time step. The corresponding reconstructed output vector ; Based on the current time step Corresponding multivariate observation vector and reconstruct the output vector Calculate the dynamic residual vector , , ; Generating Mamba Anomaly Scores Based on Dynamic Residual Vectors : ; A ranking list of contribution variables is generated based on the component magnitudes of the dynamic residual vector. Contribution variable sorting list For sets according to The index sequence obtained by sorting from largest to smallest. The first index corresponds to The observation dimension.

[0022] In one embodiment of the present invention, the Mamba-2 model is used to perform a reconstruction calculation based on structured state-space duality on the sliding window input sequence to obtain the reconstruction output vector corresponding to the current time step, as follows: Input sequence for sliding window The multivariate observation vector corresponding to each time step within the time step Perform input projection processing to map multivariate observation vectors to projected vectors. : ; in, For the input projection matrix, For input projection bias; Perform one-dimensional depthwise separable convolution on the projection vector to generate a convolution output vector. : ; in, The kernel length is 1. For the first Each convolutional kernel weight vector This indicates element-wise multiplication; The input driving vector is generated by linear transformation based on the convolution output vector. Step size vector Input mapping vector Output mapping vector and gating vectors : ; ; ; ; ; in, , , , , Corresponding linear transformation matrix, , , , , For the corresponding linear transformation bias, , ; Set the initial state vector corresponding to the start time of the sliding window. , The state vector is obtained by recursively updating the state at each time step within the sliding window in chronological order. ; Based on the output mapping vector State vector With input driving vector Generate state output vector : ; in, For the through parameter vector, These are the trainable parameters of the Mamba-2 model; Output vector of state The gating output vector is obtained by performing element-wise multiplication. The reconstructed output vector is obtained through linear transformation. : ; ; in, For the gated vector, To output the mapping matrix, This is the output mapping bias.

[0023] In one embodiment of the present invention, the state vector is obtained by recursively updating the state at each time step within the sliding window in chronological order, as follows: For each time step Generate state recursion coefficient vector based on step size vector Based on the state recursion coefficient vector and the input mapping vector Generate input injection vector ; ; ; in, For the state parameter vector, These are the trainable parameters of the Mamba-2 model; Move the sliding window according to the preset block length. Divided into multiple consecutive blocks, the number of blocks is For each piece Calculate the intra-block cumulative product vector based on the state recursion coefficient vector and the input injection vector, respectively. With the cumulative sum vector within the block : ; ; Based on intra-block cumulative product vector With the cumulative sum vector within the block Generate state vectors for each time step ,when hour, ,when hour, .

[0024] In one embodiment of the present invention, key indicator observations are constructed based on multivariate observation vector sequences, and historical context sequences are built. A quantile prediction sequence is then output using the Chronos-Bolt model, as follows: Calculate the key indicator observations at each time step based on multivariate observation vectors. : ; in, , The coefficients corresponding to the key indicators. , For constant terms; And construct a key indicator observation sequence using the key indicator observations at each time step. ; Constructing historical context sequences based on key indicator observation sequences The length of the historical context sequence is : ; Historical context sequence Input a Chronos-Bolt model and output the preset prediction step size range. The quantile prediction sequence includes each prediction step size. Corresponding P10 quantile prediction value P50 quantile predicted value and the predicted value of the P90 quantile The P10 quantile forecast value indicates a 10% probability that the observed value of the key indicator will be lower than this value at the predicted future time. The P50 quantile forecast value indicates a 50% probability that the observed value of the key indicator will be lower than this value at the predicted future time. The P90 quantile forecast value indicates a 90% probability that the observed value of the key indicator will be lower than this value at the predicted future time. The three quantile forecast values ​​provide the expected range of the observed value of the key indicator at different probability levels, thereby helping the system to better assess the possible range of changes and risks. The range between P10 and P90 shows the more extreme possibility of the indicator's change, while P50 is the most likely median value. The Chronos-Bolt model is a time series forecasting model whose architecture is designed to handle the complexity of time series and provide efficient forecasting capabilities. The input layer of the Chronos-Bolt model receives historical data sequences within a time window, and the output layer outputs quantile predictions for multiple future time steps.

[0025] In one embodiment of the present invention, the prediction deviation score is calculated based on the quantile prediction sequence and the observed values ​​of key indicators, as follows: When the observed value of the key indicator is greater than or equal to the predicted value of the P90 quantile, the predicted deviation score is... Take key indicator observations P90 quantile predicted value The difference between the predicted values ​​of the P90 quantile and the P50 quantile The ratio of their differences: ; When the observed value of the key indicator is less than or equal to the predicted value of the P10 quantile, the predicted deviation score is taken as the predicted value of the P10 quantile. The ratio of the difference between the observed value and the predicted value of the key indicator to the difference between the predicted value of the P50 quantile and the predicted value of the P10 quantile: ; When the observed value of a key indicator is greater than the predicted value of the P10 quantile but less than the predicted value of the P90 quantile, the prediction deviation score is 0.

[0026] In one embodiment of the present invention, the Mamba anomaly score and the predicted deviation score are fused to generate a fused alarm score, and the alarm level is output as follows: The fused alarm score is obtained by weighted summation of Mamba anomaly score and predicted deviation score. : ; in, These are the weighting coefficients; Set fusion alarm threshold Deviation from prediction threshold The fusion alarm threshold and prediction deviation threshold are set based on industry experience. When the fusion alarm score is greater than the fusion alarm threshold and the prediction deviation score is greater than the prediction deviation threshold, a critical alarm is output. When the fusion alarm score is less than the fusion alarm threshold and the prediction deviation score is less than the prediction deviation threshold, no alarm is output. Otherwise, a warning alarm is output.

[0027] In one embodiment of the present invention, the alarm evidence package is input into an industry-wide large language model to generate a structured processing result, as follows: A pre-defined knowledge base is constructed, which includes the process flow of oil and gas production, process specifications, alarm and interlock logic description documents, alarm handling cards, emergency response plans, equipment manuals, equipment maintenance manuals, equipment ledgers and spare parts lists, well history records, operation measure records, historical alarm records, fault case library, and operation permit documents; The system retrieves process procedures, alarm handling cards, equipment manuals, and well history records that match the alarm evidence package from a preset knowledge base, forming a set of retrieval results. The knowledge base is constructed based on historical oil and gas production data. The alarm evidence package and the search result set are input into the industry big language model for processing to generate structured handling results. The industry big language model is built based on the Tongyi Qianwen 2.5 model and has been fine-tuned. The structured handling results include anomaly description, root cause candidate list, handling step list and risk warning. The root cause candidate list is the possible root cause of the current anomaly and its corresponding evidence provided by the alarm evidence package and search results. The handling step list is the suggested handling steps generated according to the process specification and root cause candidates.

[0028] like Figure 2The method for monitoring oil and gas production data, as shown, includes the following steps: S1. Acquire multivariate time-series data of target oil and gas production and preprocess it to generate a multivariate observation vector sequence; S2. Construct a sliding window input sequence based on the multivariate observation vector sequence, input the sliding window input sequence into the Mamba-2 model to obtain the reconstructed output vector and construct a dynamic residual vector, and generate the Mamba anomaly score and contribution variable ranking list based on the dynamic residual vector; S3. Based on the multivariate observation vector sequence, construct the observation values ​​of key indicators and build the historical context sequence, and output the quantile prediction sequence through the Chronos-Bolt model; S4. Calculate the predicted deviation score based on the quantile prediction sequence and the observed values ​​of key indicators, fuse the Mamba anomaly score and the predicted deviation score to generate a fused alarm score, and output the alarm level. S5. Based on alarm level, fused alarm score, Mamba anomaly score, predicted deviation score, contribution variable ranking list and quantile prediction sequence, construct alarm evidence package, input alarm evidence package into industry big language model to generate structured processing results.

[0029] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. An oil and gas production data monitoring system, characterized in that, include: The data acquisition module is used to acquire multivariate time-series data of target oil and gas production and preprocess it to generate a multivariate observation vector sequence; The reconstruction and anomaly scoring module is used to construct a sliding window input sequence based on a multivariate observation vector sequence, input the sliding window input sequence into the Mamba-2 model to obtain the reconstruction output vector and construct a dynamic residual vector, and generate Mamba anomaly scores and a ranking list of contributing variables based on the dynamic residual vector; The quantile prediction module is used to obtain key indicator observations and construct historical context sequences based on multivariate observation vector sequences, and output quantile prediction sequences through the Chronos-Bolt model. The alarm fusion judgment module is used to calculate the predicted deviation score based on the quantile prediction sequence and the key indicator observation value, fuse the Mamba anomaly score and the predicted deviation score to generate a fused alarm score, and output the alarm level. The results generation module is used to construct an alarm evidence package based on alarm level, fused alarm score, Mamba anomaly score, predicted deviation score, ranking list of contributing variables, and quantile prediction sequence. The alarm evidence package is then input into the industry big language model to generate structured processing results.

2. The oil and gas production data monitoring system according to claim 1, characterized in that, Acquire multivariate time-series data on target oil and gas production and preprocess it to generate a multivariate observation vector sequence, including: Collect multivariate time-series data, including pressure, temperature, flow rate, moisture content, valve opening, current, and voltage, and perform time alignment on the multivariate time-series data; Based on the aligned multivariate time series data, a multivariate observation vector is constructed at each time step, and the multivariate observation vectors of all time steps are combined to form a multivariate observation vector sequence.

3. The oil and gas production data monitoring system according to claim 1, characterized in that, A sliding window input sequence is constructed based on the multivariate observation vector sequence. This sliding window input sequence is then input into the Mamba-2 model to obtain the reconstructed output vector and construct a dynamic residual vector. Based on the dynamic residual vector, Mamba anomaly scores and a ranking list of contributing variables are generated, including: Construct a sliding window input sequence based on a preset window length; The Mamba-2 model is used to perform a reconstruction calculation based on structured state space duality on the sliding window input sequence to obtain the reconstruction output vector corresponding to the current time step; Calculate the dynamic residual vector based on the multivariate observation vector and the reconstructed output vector corresponding to the current time step; Mamba anomaly scores are generated based on dynamic residual vectors, and a ranking list of contributing variables is generated based on the component magnitudes of the dynamic residual vectors.

4. The oil and gas production data monitoring system according to claim 3, characterized in that, The Mamba-2 model is used to perform a structured state-space duality-based reconstruction computation on the sliding window input sequence to obtain the reconstruction output vector corresponding to the current time step, including: For each time step in the sliding window input sequence, the multivariate observation vector is processed by input projection, which maps the multivariate observation vector to a projection vector. Perform one-dimensional depthwise separable convolution on the projection vector to generate a convolution output vector; Based on the convolution output vector, the input driving vector, stride vector, input mapping vector, output mapping vector, and gate vector are generated through linear transformation. Set the initial state vector corresponding to the start time of the sliding window, and recursively update the state vector for each time step in the sliding window in chronological order. Generate a state-output vector based on the output mapping vector, the state vector, and the input driving vector; The state output vector is multiplied element by element to obtain the gated output vector, and then the reconstructed output vector is obtained through linear transformation.

5. The oil and gas production data monitoring system according to claim 4, characterized in that, The state vector is obtained by recursively updating the state at each time step within the sliding window in chronological order, including: For each time step, a state recursion coefficient vector is generated based on the step size vector, and an input injection vector is generated based on the state recursion coefficient vector and the input mapping vector. The sliding window is divided into multiple consecutive blocks. For each block, the cumulative product vector and the cumulative sum vector within the block are calculated based on the state recursion coefficient vector and the input injection vector, respectively. The state vector for each time step is generated based on the intra-block cumulative product vector and the intra-block cumulative sum vector.

6. The oil and gas production data monitoring system according to claim 1, characterized in that, Based on multivariate observation vector sequences, key indicator observations are constructed and historical context sequences are built. A quantile prediction sequence is then output using the Chronos-Bolt model, including: The key indicator observation values ​​for each time step are calculated based on the multivariate observation vector, and the key indicator observation sequence is constructed using the key indicator observation values ​​for each time step. A historical context sequence is constructed based on the observation sequence of key indicators. The historical context sequence is input into the Chronos-Bolt model, and the output is a quantile prediction sequence within a preset prediction step size range. The quantile prediction sequence includes the P10 quantile prediction value, P50 quantile prediction value and P90 quantile prediction value corresponding to each prediction step size.

7. The oil and gas production data monitoring system according to claim 1, characterized in that, The predicted deviation score is calculated based on the quantile prediction sequence and the key indicator observations, including: When the observed value of the key indicator is greater than or equal to the predicted value of the 90th quantile, the prediction deviation score is the ratio of the difference between the observed value of the key indicator and the predicted value of the 90th quantile to the difference between the predicted value of the 90th quantile and the predicted value of the 50th quantile. When the observed value of the key indicator is less than or equal to the predicted value of the P10 quantile, the prediction deviation score is the ratio of the difference between the predicted value of the P10 quantile and the observed value of the key indicator to the difference between the predicted value of the P50 quantile and the predicted value of the P10 quantile. When the observed value of a key indicator is greater than the predicted value of the P10 quantile but less than the predicted value of the P90 quantile, the prediction deviation score is 0.

8. The oil and gas production data monitoring system according to claim 1, characterized in that, The Mamba anomaly score and the predicted deviation score are fused to generate a fused alarm score, and the alarm level is output, including: The fused alarm score is obtained by weighted summation of Mamba anomaly score and predicted deviation score; Set the fusion alarm threshold and the prediction deviation threshold. When the fusion alarm score is greater than the fusion alarm threshold and the prediction deviation score is greater than the prediction deviation threshold, output a critical alarm. When the fusion alarm score is less than the fusion alarm threshold and the prediction deviation score is less than the prediction deviation threshold, output no alarm. Otherwise, output a warning alarm.

9. The oil and gas production data monitoring system according to claim 1, characterized in that, The alarm evidence package is input into the industry's large language model to generate structured processing results, including: A pre-defined knowledge base is constructed, which includes the process flow of oil and gas production, process specifications, alarm and interlock logic description documents, alarm handling cards, emergency response plans, equipment manuals, equipment maintenance manuals, equipment ledgers and spare parts lists, well history records, operation measure records, historical alarm records, fault case library, and operation permit documents; The system retrieves process procedures, alarm handling cards, equipment manuals, and well history records that match the alarm evidence package from a preset knowledge base, forming a set of retrieval results. The knowledge base is constructed based on historical oil and gas production data. The alarm evidence package and the retrieval result set are input into the industry big language model for processing to generate structured handling results. The industry big language model is built based on the Tongyi Qianwen 2.5 model and has been fine-tuned. The structured handling results include anomaly description, root cause candidate list, handling step list and risk warning.

10. A method for monitoring oil and gas production data, characterized in that, Implementing an oil and gas production data monitoring system as described in any one of claims 1 to 9 includes the following steps: S1. Acquire multivariate time-series data of target oil and gas production and preprocess it to generate a multivariate observation vector sequence; S2. Construct a sliding window input sequence based on the multivariate observation vector sequence, input the sliding window input sequence into the Mamba-2 model to obtain the reconstructed output vector and construct a dynamic residual vector, and generate the Mamba anomaly score and contribution variable ranking list based on the dynamic residual vector; S3. Based on the multivariate observation vector sequence, construct the observation values ​​of key indicators and build the historical context sequence, and output the quantile prediction sequence through the Chronos-Bolt model; S4. Calculate the predicted deviation score based on the quantile prediction sequence and the observed values ​​of key indicators, fuse the Mamba anomaly score and the predicted deviation score to generate a fused alarm score, and output the alarm level. S5. Based on alarm level, fused alarm score, Mamba anomaly score, predicted deviation score, contribution variable ranking list and quantile prediction sequence, construct alarm evidence package, input alarm evidence package into industry big language model to generate structured processing results.

Images