Sensor data anomaly detection method based on multi-dimensional fusion of hydrogen-based shaft furnace

Abstract

The invention discloses a multi-dimensional fusion sensor data anomaly detection method based on a hydrogen-based shaft furnace, and belongs to the technical field of metallurgical equipment monitoring. The method comprises the steps of data preprocessing, single-variable anomaly detection, multivariable consistency detection, anomaly type judgment and alarm output. The single variable detection adopts a rolling median absolute deviation method, drift detection, continuous tiny variable threshold and border crossing detection; the multivariate detection calculates reconstruction residuals of each sensor through principal component analysis (PCA) and identifies overall consistency anomalies. And the system judges the fault type according to the comprehensive score of the multiple detection results and outputs a detailed alarm record and a statistical report. According to the method, multi-algorithm fusion analysis is carried out on time sequence data collected by a plurality of sensors in the operation process of the shaft furnace, so that the abnormal states of the sensors are accurately recognized. The method can be widely applied to real-time monitoring of the hydrogen-based shaft furnace smelting process, the accuracy and robustness of anomaly detection are improved, and false alarms and missing alarms are reduced.

Description

Technical Field

[0001] This invention belongs to the field of metallurgical equipment monitoring technology, and relates to a multi-dimensional fusion sensor data anomaly detection method based on hydrogen-based vertical furnace, which is particularly suitable for real-time monitoring and anomaly diagnosis of key parameter sensors in the production process of hydrogen-based vertical furnace. Background Technology

[0002] Hydrogen-based shaft furnaces are core equipment in the modern steel industry's green and low-carbon smelting process. Their production involves real-time monitoring of numerous sensors (such as temperature, pressure, flow, and composition analysis sensors). The reliability and data quality of these sensors directly affect production safety, product quality, and energy consumption levels.

[0003] Currently, anomaly detection methods for industrial sensors mainly include thresholding based on univariate statistics, moving average methods, and principal component analysis (PCA) based on multivariate relationships. However, the hydrogen-based shaft furnace process system is complex, and sensor signals are mutually coupled, making the limitations of a single method obvious. 1. Univariate methods (such as threshold checks and fixed rules) are difficult to identify complex and related faults, resulting in high false alarm and false negative rates.

[0004] 2. Multivariate methods (such as PCA) can detect the disruption of the correlation between variables, but it is difficult to locate the specific faulty sensor and is not sensitive to typical faults such as jamming or drift of a single sensor.

[0005] Therefore, existing technologies lack a solution that can comprehensively judge single-point anomalies and multivariate inconsistencies, and accurately diagnose fault types. Summary of the Invention

[0006] The purpose of this invention is to overcome the shortcomings of existing technologies and provide a multi-dimensional fusion-based sensor data anomaly detection method based on a hydrogen-based vertical shaft furnace. This method combines single-sensor signal feature analysis with multi-sensor consistency detection to quickly and accurately identify various types of sensor anomalies. Furthermore, by fusing univariate time series analysis and multivariate statistical analysis techniques, this method achieves multi-level, high-precision anomaly detection and fault type diagnosis of sensor data.

[0007] To achieve the above objectives, the technical solution adopted by this invention is: a multi-dimensional fusion sensor data anomaly detection method based on a hydrogen-based vertical furnace, comprising the following steps: S1: Data Acquisition and Preprocessing of Hydrogen-Based Vertical Shaft Furnace S2: Single-parameter anomaly feature detection For each column of sensor data, anomaly features are calculated using an adaptively sized sliding window to generate four types of anomaly markers: spike anomaly, drift anomaly, jamming anomaly, and out-of-bounds anomaly. S3: Multi-parameter collaborative anomaly detection based on the correlation of parameters in hydrogen-based vertical shaft furnaces: Multi-parameter collaborative anomaly detection aims to discover the disruption of the normal correlation between multiple sensor readings. The effectiveness of the principal component analysis model is highly dependent on the selection of input variables. The following four key parameters are selected to form the input dataset of the PCA model: reduction section temperature, hydrogen concentration, furnace top pressure, and nitrogen flow rate. During normal operation, the four key parameters maintain a dynamic balance. When unexpected and unbalanced changes occur, the PCA model outputs multi-parameter co-parameter anomalies. S4: Anomaly Type Determination and Alarm Generation: S4.1 Multi-evidence fusion and alarm triggering mechanism For each sensor's data at each time point, spike anomalies, drift anomalies, jamming anomalies, out-of-bounds anomalies, and multi-parameter collaborative anomalies are treated as independent Boolean evidence, and summed to obtain a comprehensive "alert score"; and the score is set to be ≥ 2 for the judgment threshold of alarm triggering. S4.2 Fault Type Determination: Once an alarm is triggered, the system will apply the judgment rule engine to diagnose the fault type based on the specific combination pattern of the abnormal flags. S4.3 Output and Response: The system finally outputs a structured list of alarms. Each alarm includes a timestamp, sensor tag number, alarm score, diagnosed fault type, and the status of all underlying anomaly flags. S5: Results Output and Visualization (for Hydrogen-Based Shaft Furnace Operation and Maintenance) Save alarm records to CSV / Excel, including: timestamp, sensor location, anomaly type, and associated parameter values; Generate visualization charts: trend charts showing the time distribution of anomalies in key parameters such as hydrogen concentration and temperature; statistical histograms of anomaly types; and PCA residual distribution charts, which distinguish the correlation between parameters under normal and abnormal operating conditions.

[0008] Furthermore, step S1 includes the following sub-steps: S1.1 Collect sensor time-series data at key locations of the hydrogen-based vertical shaft furnace, including: Temperature sensors are used to collect the temperatures of the pre-reduction section, reduction section, and cooling section of the furnace body. Pressure sensors are used to collect the pressure at the top and bottom of the furnace. Gas sensors are used to collect hydrogen concentration, carbon monoxide concentration, and nitrogen flow rate. Stored as structured data, the data includes a time column; S1.2 Data Preprocessing: The time column was converted to datetime format, sorted by time, and duplicate timestamps were removed. A data frame was then constructed using the time column as an index. Subsequently, target sensor variables related to the hydrogen-based shaft furnace process were identified. A rolling time window is used to ensure that the detection method is adapted to different data sampling frequencies. The specific steps are as follows: Determine the process time scale: Define a physically meaningful reference time length T_window; The dynamic number of window data points, win_n, is calculated and rounded using the following formula: win_n = T_window / Δt_avg; Where Δt_avg is the average sampling interval of the entire dataset.

[0009] Furthermore, step S2 specifically includes the following sub-steps: S2.1 Peak Anomaly Detection: The Hampel algorithm is used to calculate the median and median absolute deviation (MAD) of the data within a scrolling window; when the deviation of a data point from the median exceeds a set threshold, it is marked as a peak anomaly. S2.2 Drift Anomaly Detection: The instantaneous slope is calculated by linear least squares fitting within the scrolling window. The absolute value of all slope values ​​is calculated and a threshold is set. If the absolute value of the slope of the current window exceeds this threshold, the end point of the window is marked as a drift anomaly. S2.3 Deadline Anomaly Detection: If the standard deviation within the rolling window is less than a very small threshold and the percentage of points where the absolute value of the change between consecutive data points within the rolling window is less than another very small threshold exceeds 80%, then it is marked as a deadline anomaly. S2.4 Boundary Exceedance Detection: Set upper and lower limits based on the statistical distribution of the safety operation procedures of the hydrogen-based vertical furnace or historical sensor data; Hard threshold: Directly use process safety limits; Soft threshold: Calculates the quantiles of historical data from the sensor under normal operating conditions; Data that exceeds the soft threshold but does not reach the hard threshold is marked as out-of-bounds anomalies.

[0010] Furthermore, in step 4.1, the score of the alarm trigger determination threshold is set to be ≥ 2.

[0011] Furthermore, in step 4.2, the determination rule is as follows: Rule 1: If the stuck flag is true and the multi-parameter coordination flag is false, then the diagnosis is "sensor hardware failure"; Rule 2: If the spike flag is true and the multi-parameter coordination flag is false, then the diagnosis is "transient interference"; Rule 3: If the drift flag is true and the multi-parameter coordination flag is false, then the diagnosis is "slow parameter degradation"; Rule 4: If the boundary crossing flag is true and the multi-parameter coordination flag is false, then the diagnosis is "exceeding the safe operating threshold"; Rule 5: If the multi-parameter coordination flag is true, then regardless of the single-parameter flag, it should be diagnosed as a "system-level anomaly" first. Rule 6: Any other combination patterns not covered by the above rules shall be diagnosed as "composite suspicious faults".

[0012] This method includes: data preprocessing, univariate anomaly detection, multivariate consistency detection, anomaly type determination, and alarm output. Univariate detection employs the rolling median absolute deviation method (Hampel method), drift detection (based on the slope distribution of sliding window linear regression), continuous small change thresholds (based on rolling standard deviation and the proportion of small changes), and out-of-bounds detection (based on historical quantile thresholds). Multivariate detection uses principal component analysis (PCA) to calculate the reconstruction residuals of each sensor and identify overall consistency anomalies. The system determines the fault type based on a comprehensive score of multiple detection results and outputs detailed alarm records and statistical reports.

[0013] Compared with existing technologies, the advantages of this invention are as follows: This method achieves accurate identification of sensor anomalies by performing multi-algorithm fusion analysis on time-series data collected by multiple sensors during shaft furnace operation. This method can be widely applied to real-time monitoring of hydrogen-based shaft furnace smelting processes, improving the accuracy and robustness of anomaly detection and reducing false alarms and missed alarms. Specific advantages include: 1. By simultaneously utilizing univariate and multivariate analysis methods, the accuracy and robustness of anomaly detection can be effectively improved; 2. It can provide clear classifications for different anomaly types, making it easier for maintenance personnel to quickly locate the cause of the fault; 3. The method parameters can be flexibly adjusted to adapt to different sensor characteristics and sampling conditions; 4. Supports missing value processing and high-frequency data calculation to meet the real-time monitoring needs of hydrogen-based vertical shaft furnaces. Attached Figure Description

[0014] To make the objectives, technical solutions, and advantages of the present invention clearer, the preferred embodiments of the present invention will be described in detail below with reference to the accompanying drawings, wherein: Figure 1 This is a flowchart of the overall process of the present invention.

[0015] Figure 2 This is a schematic diagram of multi-sensor PCA consistency detection.

[0016] Figure 3 This is a schematic diagram of the comprehensive anomaly determination logic.

[0017] Figure 4 A visualization example of the distribution of fault types. Detailed Implementation

[0018] The present invention will be specifically described below through embodiments. It should be noted that the following embodiments are only used to further illustrate the present invention, but are not limited thereto, unless otherwise stated.

[0019] The specific embodiments of the present invention are described in detail below with reference to the technical solution: S1: Data Acquisition and Preprocessing of Hydrogen-Based Vertical Shaft Furnace S1.1 Collects sensor time-series data at key locations of the hydrogen-based vertical shaft furnace, including: Temperature sensors: Temperature of each section of the furnace body (pre-reduction section, reduction section, cooling section) (unit: °C); Pressure sensors: furnace top pressure, furnace bottom pressure (unit: kPa); Gas sensors: hydrogen concentration (%), carbon monoxide concentration (%), nitrogen flow rate (m³ / h); Stored as structured data (such as Excel format), including a time column (accurate to the second).

[0020] S1.2 Data Preprocessing: The time column is converted to datetime format, sorted by time, and duplicate timestamps are removed. A data frame is then constructed using the time column as an index. Subsequently, target sensor variables related to the hydrogen-based shaft furnace process are selected (their column names usually end with .daca.pv, such as "temp.daca.pv", "press.daca.pv", "h2conc.daca.pv", etc.).

[0021] To ensure the anomaly detection algorithm can adapt to different data sampling frequencies, this invention introduces a time-based rolling window instead of a fixed data point window. The specific steps are as follows: Determine the process timescale: First, based on the process characteristics of the hydrogen-based shaft furnace and expert knowledge, define a physically meaningful baseline time length T_window (e.g., T_window = 1 hour). This timescale should be sufficient to capture dynamic changes in the process (such as gradual temperature changes and flow rate fluctuations) without being too long to mask local anomalies.

[0022] Calculate the dynamic window size win_n: Calculate the average sampling interval Δt_avg for the entire dataset.

[0023] The number of dynamic window data points, win_n, is calculated using the formula win_n = T_window / Δt_avg and then rounded down.

[0024] This method ensures that the rolling window always covers a data segment with a fixed process duration, enabling the algorithm to have consistent behavior and comparability across data with different sampling rates.

[0025] S2: Single-parameter anomaly feature detection For each column of sensor data, anomaly features are calculated using the aforementioned adaptive-sized sliding window, generating four types of anomaly markers: S2.1 Peak Anomaly Detection: The Hampel algorithm is used. The median and median absolute deviation (MAD) of the data are calculated within a scrolling window. When the deviation of a data point from the median exceeds... When (the coefficient k is adjustable), it is marked as a spike anomaly. For example, a sudden surge in hydrogen concentration may be due to pipeline leaks or measurement interference, and this method can effectively detect it.

[0026] S2.2 Drift Anomaly Detection: The instantaneous slope is calculated using linear least squares fitting within a rolling window. The absolute value of all slope values ​​is calculated, and the 99th quantile of their historical distribution is used as a threshold. If the absolute value of the slope in the current window exceeds this threshold, the end point of that window is marked as a drift anomaly. For example, a slow rise or fall in the temperature of the reduction section over several hours may indicate aging or dust accumulation in the heating element; this method can effectively identify such slow trends.

[0027] S2.3 Freeze Anomaly Detection: A freeze anomaly is marked when both of the following conditions are met: If the standard deviation within the rolling window is less than a very small threshold (such as 1e-8), it indicates that the data has almost no fluctuation.

[0028] The percentage of points where the absolute value of the change between consecutive data points within the scrolling window is less than another very small threshold (such as 1e-9) exceeds 80%.

[0029] S2.4 Boundary Exceedance Detection: Set upper and lower limits based on the statistical distribution of the safety operation procedures of the hydrogen-based vertical furnace or historical sensor data.

[0030] Hard threshold: Directly use process safety limits (such as hydrogen concentration upper limit of 4.0%, reduction section temperature range of 800-1000℃).

[0031] Soft threshold (statistical threshold): Calculate the quantiles (such as 0.5% and 99.5% quantiles) of historical data of the sensor under normal operating conditions. If the value exceeds this range, it is marked as an out-of-bounds anomaly.

[0032] This method can detect early anomalies that are outside the normal statistical range but have not reached the absolute safety risk value.

[0033] S3: Multi-parameter collaborative anomaly detection (based on the correlation of parameters in a hydrogen-based vertical shaft furnace) Multi-parameter collaborative anomaly detection aims to discover the disruption of the normal correlation between multiple sensor readings, which often indicates deeper systemic problems that are difficult to detect with univariate detection.

[0034] S3.1 Basis for Selecting Key Parameters The validity of principal component analysis (PCA) models is highly dependent on the selection of input variables. Input variables should meet the following two conditions: Strong correlation of process mechanism: The selected variables should have a clear coupling or causal relationship in terms of physicochemical principles, and jointly describe a specific sub-process.

[0035] High statistical correlation of data: Under normal operating conditions, the historical data of these variables should show stable statistical correlation.

[0036] S3.2 Based on the above principles, this invention selects the following four key parameters to form the input dataset for the PCA model, focusing on the core sub-process of the "gas reduction reaction" in a hydrogen-based vertical shaft furnace: Reduction temperature (temp_reduction.daca.pv): The main driving energy source for the reduction reaction.

[0037] Hydrogen concentration (h2_conc.daca.pv): The concentration of the main reducing agent.

[0038] Furnace top pressure (pressure_top.daca.pv): A key physical quantity that affects the flow rate of reactant gases and reaction equilibrium.

[0039] Nitrogen flow rate (n2_flow.daca.pv): As a protective gas and carrier gas, its flow rate changes will affect the partial pressure and concentration of other gases.

[0040] Selection Criteria: These four parameters are highly coupled during the reduction reaction. For example, changes in hydrogen concentration affect the reaction rate and heat release, thus affecting temperature; changes in temperature, in turn, affect reaction equilibrium and gas volume, thereby affecting pressure; changes in nitrogen flow rate dilute the reactant concentration and affect the total pressure. Therefore, they should maintain a dynamic equilibrium during normal operation. Any unexpected coordinated changes that violate this equilibrium (such as hydrogen concentration not decreasing as expected when temperature increases) will be detected by the PCA model.

[0041] S4: Anomaly type determination and alarm generation (adapted to hydrogen-based shaft furnace failure modes) To achieve the leap from "anomaly detection" to "fault diagnosis," this invention designs a set of alarm generation and type determination rules based on multi-evidence fusion. This mechanism aims to distinguish between random noise, independent sensor failures, and complex system-level process anomalies, thereby generating high-confidence, actionable alarm information to guide maintenance personnel to take quick and accurate countermeasures.

[0042] S4.1 Multi-evidence fusion and alarm triggering mechanism For each sensor's data at each time point, its five anomaly detection results (single-dimensional jamming, spikes, drift, out-of-bounds, and multi-dimensional multi-parameter coordination) are treated as independent Boolean evidence. These pieces of evidence are summed to obtain a comprehensive "Alert Score".

[0043] The threshold for triggering an alarm is a score ≥ 2. This threshold is set based on the following scientific and engineering considerations: Suppressing false alarms and reducing noise: Instantaneous noise, measurement jitter, or normal extreme fluctuations in industrial field data may cause a single detection algorithm to misjudge (score=1). Requiring at least two independent detection dimensions to trigger simultaneously can effectively filter out the vast majority of occasional false alarms, significantly improve alarm reliability, and avoid "alarm fatigue."

[0044] Enhancing Fault Confidence: When two or more detection methods based on different principles (such as variance-based thresholding of continuous small changes and distribution-based out-of-bounds detection) simultaneously indicate anomalies, they constitute strong cross-validation of evidence, greatly increasing the probability of real faults occurring and reducing the risk of missed detections.

[0045] Clearly defined fault levels: This threshold naturally forms a two-level monitoring system: Score = 1: Potential anomaly or low-confidence event. The system logs this event for long-term trend analysis and equipment health assessment, but does not trigger an immediate alarm.

[0046] A score of 2 or higher indicates a high-confidence anomaly. The system will immediately trigger an alarm, indicating that manual or automated intervention is required.

[0047] Parameter adjustability: A threshold of ≥2 is a default recommended value that has been validated by a large amount of historical data and simulation tests, achieving the optimal balance between detection sensitivity and specificity. In actual deployment, this parameter can be fine-tuned according to the different safety and stability requirements of specific production lines (for example, it can be lowered to ≥1 in scenarios with stringent safety requirements, and raised to ≥3 in scenarios with extremely high noise resistance requirements), reflecting the flexibility and adaptability of the method.

[0048] S4.2 Fault Type Determination Rules Once an alarm is triggered (score ≥ 2), the system will apply the following rule engine to diagnose the fault type based on the specific combination pattern of the abnormal flags, thereby directly pointing to the possible root cause: Rule 1: If the jamming flag is true and the multi-parameter coordination flag is false, the diagnosis is "sensor hardware failure". (For example: a nitrogen flow sensor outputs a constant value due to damage to internal components, causing a jamming alarm).

[0049] Rule 2: If the spike indicator is true and the multi-parameter coordination indicator is false, then it is diagnosed as "transient interference". (For example: a hydrogen concentration laser analyzer may generate transient spike noise due to brief liquid or dust obstruction in the process gas).

[0050] Rule 3: If the drift flag is true and the multi-parameter coordination flag is false, the diagnosis is "slow parameter degradation". (For example: the thermocouple in the reduction section deteriorates due to long-term high-temperature environment, and the measured value drifts slowly).

[0051] Rule 4: If the boundary crossing flag is true and the multi-parameter coordination flag is false, the diagnosis is "exceeding the safe operating threshold". (For example: due to control loop failure, the hydrogen concentration continues to rise and exceeds the process safety interlock value).

[0052] Rule 5: If the multi-parameter coordination indicator is true, then regardless of the single-parameter indicator, it should be diagnosed as a "system-level anomaly". (For example: due to a malfunction in the reducing agent injection system, "abnormal temperature drop", "abnormal hydrogen concentration increase" and "abnormal pressure fluctuation" are observed simultaneously. The normal coupling relationship between multiple parameters is disrupted, indicating a system-level problem of reduction reaction imbalance).

[0053] Rule 6: Any other combination of patterns not covered by the above rules (such as both spike and drift being true) are diagnosed as "composite suspected faults". It is recommended to initiate higher-level expert system analysis or conduct manual inspection and confirmation immediately.

[0054] S4.3 Output and Response The system ultimately outputs a structured list of alarms, each containing a timestamp, sensor tag number, alarm score, diagnosed fault type, and the status of all underlying anomaly flags. Maintenance personnel can quickly pinpoint the nature of the problem based on the diagnostic results: whether it only requires replacing or calibrating a single instrument, whether it requires eliminating on-site interference, or whether it necessitates immediate adjustments to process operations to address system risks. This invention thus transforms massive amounts of anomaly detection results into clearly structured and targeted maintenance guidance, achieving the ultimate goal of intelligent fault diagnosis.

[0055] S5: Results Output and Visualization (for Hydrogen-Based Shaft Furnace Operation and Maintenance) Save alarm records to CSV / Excel, including: timestamp, sensor location, anomaly type, and associated parameter values; generate visualization charts: anomaly time distribution trend chart (marking anomalies in key parameters such as hydrogen concentration and temperature); anomaly type statistical histogram (e.g., percentage of "exceeding safety threshold"); PCA residual distribution chart (distinguishing parameter correlations under normal / abnormal operating conditions).

[0056] It should be noted that the following detailed description is illustrative and intended to provide further explanation of the invention. Unless otherwise specified, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this invention pertains.

[0057] The real-time monitoring data of a hydrogen-based vertical shaft furnace in a steel plant used in this study includes the following sensor types: 36 temperature sensors. Example 1

[0058] This study uses real-time monitoring data from a hydrogen-based vertical shaft furnace in a steel plant over a period of time for verification. The data comes from a distributed control system (DCS) and includes multiple sensors for temperature, pressure, flow rate, and gas composition, with a sampling interval of approximately 1 minute.

[0059] S1: Data Preparation and Preprocessing Export the raw data file (zong.xlsx) in Excel format from the DCS historical database. The data contains a column of timestamps named "czas" (accurate to the second) and multiple columns of process value variables ending in .daca.pv.

[0060] Preprocessing using Python's pandas library: Read data from the specified worksheet.

[0061] The "czas" column is parsed as a datetime type, set as an index for the data frame, and sorted in ascending order by time.

[0062] Remove data rows with duplicate time indexes.

[0063] All columns ending with .daca.pv were selected as target sensor variables for subsequent analysis, resulting in 17 valid sensor data columns.

[0064] S2: Single-parameter anomaly feature detection For each sensor variable sequence obtained in S1, four types of anomaly detection are performed sequentially. All detections utilize a dynamic rolling window based on process time.

[0065] Dynamic window calculation: The average data sampling interval Δt_avg is calculated to be approximately 60 seconds. The process baseline time length T_window is set to 1 hour. Therefore, the dynamic window size win_n = T_window / Δt_avg = 60 (rounded down).

[0066] Peak detection: A Hampel filter is used with a window size of 60 and a threshold coefficient k=4.0.

[0067] Drift detection: The window size is 60. The instantaneous slope of the data within the window is calculated, and the 99th percentile of the absolute value of all slopes is taken as the threshold.

[0068] Thresholds for continuous small changes: window size is 60, standard deviation threshold eps_std=1e-8, small change threshold flat_eps=1e-9, and continuous small change percentage threshold flat_run_frac=0.8.

[0069] Out-of-bounds detection: For hydrogen concentration (e.g., h2_conc.daca.pv), a hard threshold is used, with upper and lower limits of 4.0% and 0%, respectively. For most other parameters, a statistical threshold is used, employing the 0.5% and 99.5% quantiles of their historical data as the dynamic range.

[0070] S3: Multi-parameter collaborative anomaly detection Key parameter selection: From 17 sensors, based on the correlation of process mechanism and statistical correlation, 4 core parameters were selected to construct the PCA model input dataset: reduction section temperature, hydrogen concentration, furnace top pressure, and nitrogen flow rate.

[0071] PCA Modeling and Detection: For the data of the selected parameters, the median is used to fill in missing values.

[0072] Standardize using StandardScaler.

[0073] Perform PCA analysis, setting the cumulative variance retention ratio to 0.9.

[0074] Calculate the squared prediction error (SPE) for each time point and use its 99.5th percentile as the outlier threshold. If the SPE value exceeds this threshold, mark the multi-parameter collaborative anomaly as true.

[0075] S4: Anomaly Type Determination and Alarm Generation Evidence fusion: For each sensor at each time point, five Boolean flags are summed to obtain the alarm score, which includes criteria for jamming, spikes, drift, out-of-bounds, and multi-parameter coordination.

[0076] Alarm triggering: Set the alarm score threshold score_trigger = 2. When the score is ≥2, an alarm record is generated.

[0077] Fault diagnosis: The predefined rule engine is used to perform type diagnosis on the triggered alarms. For example: If the hydrogen concentration sensor's out-of-range flag is true (value > 4.0%) at a certain moment, and the multi-parameter coordination flag is false, then it is diagnosed as "exceeding the safe operating threshold".

[0078] If the multi-parameter coordination flag is true at a certain moment, it is diagnosed as a "system-level anomaly" even if not all individual sensor flags are true.

[0079] S5: Results Output and Visualization Data Output: Save detailed alarm records (including timestamps, sensor tags, fault types, alarm scores, status of each anomaly flag, PCA reconstruction errors, etc.) as an alerts_detail.csv file. Save a statistical table summarizing by time and fault type as alerts_summary_by_time.csv.

[0080] Visualization: A trend chart of the number of generated alarms over time (e.g.) Figure 1 As shown in the figure, it visually displays the periods when abnormalities occur most frequently.

[0081] Generate a statistical chart of the distribution of fault types (e.g.) Figure 4 As shown in the figure, the frequency and percentage of various faults are clearly displayed.

[0082] Generate a histogram of PCA reconstruction error distribution and mark the abnormal threshold in the graph to distinguish between normal and abnormal operating conditions. Implementation effect

[0083] Analyzing the data from the above embodiments using this method successfully detected various faults, including "excessive hydrogen concentration," "nitrogen flow meter jamming," "temperature drift in the reduction section," and multiple "system-level anomalies." Compared with manual inspection records, this method effectively reduces false alarms caused by a single algorithm and provides maintenance personnel with clear handling directions through precise type diagnosis, verifying the practicality and effectiveness of this method in hydrogen-based vertical shaft furnace monitoring scenarios.

[0084] Those skilled in the art will understand that embodiments of the present invention can be provided as methods, systems, or computer program products. Therefore, the present invention can take the form of hardware embodiments, software embodiments, or embodiments combining software and hardware aspects. Furthermore, the present invention can take the form of a computer program product embodied on one or more computer-usable storage media (including, but not limited to, disk storage and optical storage) containing computer-usable program code.

[0085] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the present invention, and all such modifications or substitutions should be covered within the scope of the claims of the present invention.

Claims

1. A method for detecting anomalies in sensor data based on multi-dimensional fusion of hydrogen-based vertical shaft furnaces, characterized in that, Includes the following steps: S1: Data Acquisition and Preprocessing of Hydrogen-Based Vertical Shaft Furnace S2: Single-parameter anomaly feature detection For each column of sensor data, anomaly features are calculated using an adaptively sized sliding window to generate four types of anomaly markers: spike anomaly, drift anomaly, jamming anomaly, and out-of-bounds anomaly. S3: Multi-parameter collaborative anomaly detection based on the correlation of parameters in hydrogen-based vertical shaft furnaces: Multi-parameter collaborative anomaly detection aims to discover the disruption of the normal correlation between multiple sensor readings. The effectiveness of the principal component analysis model is highly dependent on the selection of input variables. The following four key parameters are selected to form the input dataset of the PCA model: reduction section temperature, hydrogen concentration, furnace top pressure, and nitrogen flow rate. During normal operation, the four key parameters maintain a dynamic balance. When unexpected and unbalanced changes occur, the PCA model outputs multi-parameter co-parameter anomalies. S4: Anomaly Type Determination and Alarm Generation: S4.1 Multi-evidence fusion and alarm triggering mechanism For each sensor's data at each time point, spike anomalies, drift anomalies, jamming anomalies, out-of-bounds anomalies, and multi-parameter collaborative anomalies are treated as independent Boolean evidence, summed to obtain a comprehensive "alert score"; and the score is set to be ≥ 2 for the judgment threshold of alarm triggering. S4.2 Fault Type Determination: Once an alarm is triggered, the system will apply the judgment rule engine to diagnose the fault type based on the specific combination pattern of the abnormal flags. S4.3 Output and Response: The system finally outputs a structured list of alarms. Each alarm includes a timestamp, sensor tag number, alarm score, diagnosed fault type, and the status of all underlying anomaly flags. S5: Results Output and Visualization (for Hydrogen-Based Shaft Furnace Operation and Maintenance) Save alarm records to CSV / Excel, including: timestamp, sensor location, anomaly type, and associated parameter values; Generate visualization charts: trend charts showing the time distribution of anomalies in key parameters such as hydrogen concentration and temperature; and histograms showing the statistical distribution of anomaly types. PCA residual distribution plot, which distinguishes the parameter correlation under normal / abnormal operating conditions.

2. The detection method according to claim 1, characterized in that, Step S1 includes the following sub-steps: S1.1 Collect sensor time-series data at key locations of the hydrogen-based vertical shaft furnace, including: Temperature sensors are used to collect the temperatures of the pre-reduction section, reduction section, and cooling section of the furnace body. Pressure sensors are used to collect the pressure at the top and bottom of the furnace. Gas sensors are used to collect hydrogen concentration, carbon monoxide concentration, and nitrogen flow rate. Stored as structured data, the data includes a time column; S1.2 Data Preprocessing: The time column was converted to datetime format, sorted by time, and duplicate timestamps were removed. A data frame was then constructed using the time column as an index. Subsequently, target sensor variables related to the hydrogen-based shaft furnace process were identified. A rolling time window is used to ensure that the detection method is adapted to different data sampling frequencies. The specific steps are as follows: Determine the process time scale: Define a physically meaningful reference time length T_window; The dynamic number of window data points, win_n, is calculated and rounded using the following formula: win_n = T_window / Δt_avg; Where Δt_avg is the average sampling interval of the entire dataset.

3. The detection method according to claim 1, characterized in that, Step S2 specifically includes the following sub-steps: S2.1 Peak Anomaly Detection: The Hampel algorithm is used to calculate the median and median absolute deviation (MAD) of the data within a scrolling window; when the deviation of a data point from the median exceeds a set threshold, it is marked as a peak anomaly. S2.2 Drift Anomaly Detection: The instantaneous slope is calculated by linear least squares fitting within the scrolling window. The absolute value of all slope values ​​is calculated and a threshold is set. If the absolute value of the slope of the current window exceeds this threshold, the end point of the window is marked as a drift anomaly. S2.3 Deadline Anomaly Detection: If the standard deviation within the rolling window is less than a very small threshold and the percentage of points where the absolute value of the change between consecutive data points within the rolling window is less than another very small threshold exceeds 80%, then it is marked as a deadline anomaly. S2.4 Boundary Exceedance Detection: Set upper and lower limits based on the statistical distribution of the safety operation procedures of the hydrogen-based vertical furnace or historical sensor data; Hard threshold: Directly use process safety limits; Soft threshold: Calculates the quantiles of historical data from the sensor under normal operating conditions; Data that exceeds the soft threshold but does not reach the hard threshold is marked as out-of-bounds anomalies.

4. The detection method according to claim 1, characterized in that, In step 4.1, the score of the alarm trigger determination threshold is set to be ≥ 2.

5. The detection method according to claim 1, characterized in that, In step 4.2, the determination rule is as follows: Rule 1: If the stuck flag is true and the multi-parameter coordination flag is false, then the diagnosis is "sensor hardware failure"; Rule 2: If the spike flag is true and the multi-parameter coordination flag is false, then the diagnosis is "transient interference"; Rule 3: If the drift flag is true and the multi-parameter coordination flag is false, then the diagnosis is "slow parameter degradation"; Rule 4: If the boundary crossing flag is true and the multi-parameter coordination flag is false, then the diagnosis is "exceeding the safe operating threshold"; Rule 5: If the multi-parameter coordination flag is true, then regardless of the single-parameter flag, it should be diagnosed as a "system-level anomaly" first. Rule 6: Any other combination patterns not covered by the above rules shall be diagnosed as "composite suspicious faults".

Images