Multi-source data fault prediction method and system of pneumatic conveying system

Abstract

The invention provides a multi-source data fault prediction method and system for a pneumatic conveying system, and relates to the technical field of industrial automation, and the method comprises the steps: collecting multi-source heterogeneous monitoring data including time sequence pressure, flow, temperature and operation state data during the operation of the pneumatic conveying system, and carrying out the initial time alignment, obtaining a corresponding time alignment mapping relation; the method comprises the following steps: preprocessing multi-source heterogeneous monitoring data to obtain a space-time alignment feature matrix; according to the method, the system state map is dynamically optimized through the real-time pressure data, multi-task learning is carried out through the map neural network and the working condition parameters, and finally accurate type judgment, position positioning and graded early warning of faults are achieved.

Description

Technical Field

[0001] This invention relates to the field of industrial automation technology, and in particular to a multi-source data fault prediction method and system for pneumatic conveying systems. Background Technology

[0002] Pneumatic conveying systems are key equipment for transporting powdery and granular materials in industries such as power and chemicals. Fault prediction is crucial for ensuring production safety and continuous operation. Existing prediction methods mainly fall into two categories: modeling based on physical mechanisms and analysis driven by historical data. The former is difficult to accurately model due to system complexity and nonlinearity; while the latter, although more feasible, has significant shortcomings in processing and fusing multi-source heterogeneous data.

[0003] Specifically, existing data-driven solutions typically process various sensor time-series signals (such as pressure and flow) independently, or perform simple spatiotemporal alignment and feature concatenation before inputting them into a general time-series model (such as LSTM). For example, a power plant's transmission system might use data from multiple pressure monitoring points along the pipeline for threshold alarms, or aggregate the data to predict a single health indicator. The core flaw of this approach is its failure to effectively model and utilize the spatial distribution and dynamic evolution characteristics of pressure signals in the pipeline network. During system operation, changes in materials and local resistance lead to dynamic changes in the pressure gradient distribution, which are key features characterizing the global state and local anomalies. Existing methods treat discrete pressure points as isolated time series, losing their spatial correlation information. Furthermore, the system feature models they construct (such as sensor relationship diagrams) are often static and cannot dynamically adjust the correlation weights of different monitoring points in joint feature learning based on real-time pressure distribution. This results in the model being insensitive to early, localized fault precursors (such as slowly forming local blockages), and the prediction results may struggle to accurately distinguish fault types and locate potential risk areas, thus limiting the accuracy and practicality of early warning systems. Summary of the Invention

[0004] The technical problem to be solved by the present invention is to provide a multi-source data fault prediction method and system for pneumatic conveying systems. By dynamically optimizing the system state spectrum through real-time pressure data, and using graph neural networks and operating parameters for multi-task learning, the method can ultimately achieve accurate fault type judgment, location positioning and graded early warning.

[0005] To solve the above-mentioned technical problems, the technical solution of the present invention is as follows: A first aspect is a multi-source data fault prediction method for a pneumatic conveying system, the method comprising: Collect multi-source heterogeneous monitoring data, including time-series pressure, flow rate, temperature, and operating status data, during the operation of the pneumatic conveying system, and perform initial time alignment to obtain the corresponding time alignment mapping relationship; Preprocessing of multi-source heterogeneous monitoring data yields a spatiotemporally aligned feature matrix; Based on the spatiotemporal alignment feature matrix, a spatiotemporal feature set is constructed; wherein, each element in the spatiotemporal feature set corresponds to a key monitoring point of the system, and its content is aggregated from the time-series monitoring data of the corresponding location; the correlation between elements is defined based on the physical connection of equipment and the dependence of process flow. Based on real-time data from multiple discrete pressure monitoring points on the main pipeline, a distribution feature vector is extracted and used to dynamically correct the spatiotemporal feature set to obtain a dynamic weighted adjacency matrix. The dynamically weighted adjacency matrix is ​​input into a pre-trained graph neural network model to perform spatiotemporal feature joint learning, resulting in a fused feature vector. The fused feature vector is spliced ​​with real-time conveying volume and material characteristic parameters to form an enhanced feature vector, which is then input into the prediction model to simultaneously calculate the probability distribution of fault types within a future set time period, as well as the pipeline section corresponding to the fault type with the highest probability, determined by the internal weight analysis of the prediction model, thus completing the fault location. Based on the probability distribution of fault types and the location of faults, graded early warning information is obtained and issued.

[0006] Secondly, a multi-source data fault prediction system for a pneumatic conveying system includes: The acquisition module is used to collect multi-source heterogeneous monitoring data, including time-series pressure, flow rate, temperature and operating status data, during the operation of the pneumatic conveying system, and to perform initial time alignment to obtain the corresponding time alignment mapping relationship. The preprocessing module is used to preprocess multi-source heterogeneous monitoring data to obtain a spatiotemporally aligned feature matrix; The construction module is used to construct a spatiotemporal feature set based on the spatiotemporal alignment feature matrix; wherein, each element in the spatiotemporal feature set corresponds to a key monitoring point of the system, and its content is aggregated from the time-series monitoring data of the corresponding location; the relationship between elements is defined based on the physical connection of equipment and the dependence of process flow. The dynamic correction module is used to extract the distribution feature vector based on real-time data from multiple discrete pressure monitoring points on the main pipeline, and use it to dynamically correct the spatiotemporal feature set to obtain a dynamic weighted adjacency matrix. The fusion module is used to input the dynamically weighted adjacency matrix into the pre-trained graph neural network model for joint learning of spatiotemporal features to obtain a fused feature vector; The prediction and location module is used to combine the fused feature vector with real-time conveying volume and material characteristic parameters to form an enhanced feature vector, which is then input into the prediction model to simultaneously calculate the probability distribution of fault types within a future set time period, as well as the pipeline section corresponding to the fault type with the highest probability, determined by the internal weight analysis of the prediction model, and complete the fault location. The early warning module is used to obtain and issue graded early warning information based on the probability distribution of fault types and the location of faults.

[0007] Thirdly, a computing device includes: One or more processors; A storage device for storing one or more programs that, when executed by one or more processors, cause the one or more processors to implement the method.

[0008] Fourthly, a computer-readable storage medium storing a program that, when executed by a processor, implements the method.

[0009] The above-described solution of the present invention has at least the following beneficial effects: Multi-dimensional time-series monitoring data of the pneumatic conveying system is collected, integrating heterogeneous data from multiple sources such as pressure, flow, temperature, and operating status. Initial time alignment is completed and a time alignment mapping relationship is generated to achieve time synchronization of different types of data, providing a consistent time reference for subsequent unified data processing. Preprocessing of the multi-source heterogeneous monitoring data eliminates interference factors such as data noise and missing data, mitigates differences in heterogeneity among multiple data sources, and transforms scattered data into a spatiotemporal alignment feature matrix, improving data standardization and completeness. Based on the spatiotemporal alignment feature matrix, a spatiotemporal feature set is constructed, aggregating time-series data according to key monitoring points of the system, determining the monitoring point attributes and feature content of each element, and defining the correlation between elements in conjunction with equipment physical connections and process flow dependencies, making the feature set fit the actual operating architecture of the system and strengthening the spatiotemporal correlation of data. Distribution feature vectors are extracted from real-time data of discrete pressure monitoring points on the main pipeline, and these vectors are used to dynamically correct the spatiotemporal feature set, generating a dynamic additive feature set. The weighted adjacency matrix integrates node relationships with real-time pressure distribution information, giving the adjacency matrix dynamic adaptability and strengthening the representation of the system's real-time operating status. The dynamically weighted adjacency matrix is ​​input into a pre-trained graph neural network for spatiotemporal feature joint learning. Leveraging the model's ability to collaboratively mine spatial relationships and temporal patterns, it integrates multi-dimensional features from dispersed monitoring points, transforming them into a unified fusion feature vector, thus enhancing the comprehensive expression and abstraction capabilities of the features. This fusion feature vector is then combined with real-time transport volume and material characteristic parameters to form an enhanced feature vector, supplementing core operating parameters to expand the feature dimensions. The predictive model synchronously outputs the probability distribution of fault types and corresponding pipeline sections, achieving collaborative analysis of fault types and locations, providing comprehensive and concrete fault information. Finally, the fault type probability distribution and location results are combined to generate and release tiered early warning information, enabling tiered transmission of fault information to adapt to different risk levels and forming a complete closed loop from fault analysis to information dissemination. Attached Figure Description

[0010] Figure 1 This is a flowchart illustrating a multi-source data fault prediction method for a pneumatic conveying system provided by an embodiment of the present invention.

[0011] Figure 2 This is a schematic diagram of a multi-source data fault prediction system for a pneumatic conveying system provided in an embodiment of the present invention. Detailed Implementation

[0012] Exemplary embodiments of the present disclosure will now be described in more detail with reference to the accompanying drawings. While exemplary embodiments of the present disclosure are shown in the drawings, it should be understood that the present disclosure may be implemented in various forms and should not be limited to the embodiments set forth herein. Rather, these embodiments are provided so that this disclosure will be thorough and complete, and will fully convey the scope of the disclosure to those skilled in the art.

[0013] like Figure 1 As shown, an embodiment of the present invention proposes a multi-source data fault prediction method for a pneumatic conveying system, the method comprising the following steps: Step 100: Collect multi-source heterogeneous monitoring data including time-series pressure, flow rate, temperature and operating status data of the pneumatic conveying system during operation, and perform initial time alignment to obtain the corresponding time alignment mapping relationship; Step 200: Preprocess the multi-source heterogeneous monitoring data to obtain the spatiotemporal aligned feature matrix; Step 300: Construct a spatiotemporal feature set based on the spatiotemporal alignment feature matrix; wherein each element in the spatiotemporal feature set corresponds to a key monitoring point of the system, and its content is aggregated from the time-series monitoring data of the corresponding location; the relationship between elements is defined based on the physical connection of equipment and the dependence of process flow. Step 400: Based on real-time data from multiple discrete pressure monitoring points on the main pipeline, extract the distribution feature vector and use it to dynamically correct the spatiotemporal feature set to obtain a dynamic weighted adjacency matrix. Step 500: Input the dynamically weighted adjacency matrix into the pre-trained graph neural network model to perform spatiotemporal feature joint learning and obtain the fused feature vector; Step 600: The fused feature vector is spliced ​​with the real-time conveying volume and material characteristic parameters to form an enhanced feature vector, which is then input into the prediction model to simultaneously calculate the probability distribution of fault types within a future set time period, as well as the pipeline section corresponding to the fault type with the highest probability, determined by the internal weight analysis of the prediction model, thereby completing the fault location. Step 700: Based on the probability distribution of fault types and the location of faults, obtain and issue graded early warning information.

[0014] In this embodiment of the invention, multi-source heterogeneous monitoring data of time-series pressure, flow rate, temperature, and operating status are integrated. An initial time alignment establishes a mapping relationship, eliminating temporal dimension differences between different data types. This provides a unified benchmark for subsequent data collaborative processing and ensures the feasibility of multi-source data correlation analysis. Preprocessing of the multi-source heterogeneous data further optimizes data quality and alignment accuracy, generating a spatiotemporal alignment feature matrix. This removes invalid interference information and strengthens the spatiotemporal correlation of the data, providing standardized and usable basic data support for subsequent feature set construction. A spatiotemporal feature set is constructed based on the spatiotemporal alignment feature matrix, ensuring that each element accurately corresponds to a key monitoring point in the system. Time-series data from corresponding locations are aggregated. Simultaneously, element correlation relationships are defined based on equipment physical connections and process flow dependencies, allowing the feature set to align with the actual operating logic of the system and strengthening the physical orientation and correlation of the features. Discrete pressure monitoring points on the main pipeline are utilized. Real-time data is used to extract distribution feature vectors, dynamically correct spatiotemporal feature sets, and combine real-time operating information of pipeline pressure distribution to calibrate features, compensating for insufficient coverage of local operating condition differences by initial features and making features more adaptable to the real-time operating status of the system. A pre-trained graph neural network model is used to conduct spatiotemporal feature joint learning on a dynamically weighted adjacency matrix, fully exploring the spatiotemporal correlation patterns between features and achieving deep fusion of features from different dimensions. The fused feature vector is then concatenated with real-time transport volume and material characteristic parameters to form an enhanced feature vector, supplementing core operating condition parameters to improve feature dimensions. A prediction model is used to calculate the probability distribution of fault types, and the corresponding pipeline section is located based on the model's internal weight analysis, achieving coordinated fault type judgment and location identification. Based on the fault type probability distribution and location results, graded early warning information is generated, and warnings are issued differentiated according to risk level to avoid homogenization of warning information and improve the targeting and efficiency of operation and maintenance.

[0015] In a preferred embodiment of the present invention, step 100 above involves collecting multi-source heterogeneous monitoring data, including time-series pressure, flow rate, temperature, and operating status data, during the operation of the pneumatic conveying system, and performing initial time alignment to obtain the corresponding time alignment mapping relationship, including: Step 101: Collect time-series pressure data from pressure vessels in the pneumatic conveying system. Specifically, this includes: deploying pressure sensors at preset monitoring locations on the core pressure vessels of the pneumatic conveying system, including silo pumps, buffer tanks, and key pressure-bearing sections of the conveying pipelines. The sensors collect time-series pressure data from each location in real time. During the collection process, a fixed sampling frequency is maintained to ensure the continuity of the data sequence. The specific sampling frequency is set in conjunction with the system's operating rhythm to balance data integrity and redundancy control. This process obtains continuous pressure time-series information that accurately reflects the operating status of each pressure-bearing component, providing basic data support for subsequent analysis of the spatial distribution and dynamic evolution characteristics of pressure signals.

[0016] Step 102: Collect time-series flow data related to material conveying. Specifically, this includes: based on the core requirements of the pneumatic conveying system for material conveying, deploying flow sensors at key nodes in the material conveying path to collect time-series flow data of compressed air and instantaneous flow data of material conveying, which are directly related to material conveying. The compressed air flow data is collected at the fan outlet and the air inlet of the pipeline, while the material flow data is collected at the discharge port of the silo pump and the end of the conveying pipeline. By simultaneously capturing the dynamic changes of the two types of flow data, the material conveying efficiency and the pneumatic power supply status are accurately correlated. This supplements the key time-series information of the material conveying dimension, improves the core coverage of the system's operating conditions by multi-source data, and lays the foundation for subsequent correlation of the dynamic relationship between pressure distribution and material changes.

[0017] Step 103: Collect time-series temperature data of key rotating equipment. Specifically, this includes: for key rotating equipment in the pneumatic conveying system, mainly including components that are prone to overheating and have a high risk of failure, such as conveying fans and drive motors, temperature sensors are installed at the bearings and motor windings of the equipment to collect time-series temperature data of each key component. During the collection process, the rate of temperature change and steady-state temperature value are recorded simultaneously. The time-series characteristics of equipment temperature changes with operating time and load are captured in real time. This provides accurate data for judging abnormal thermal conditions of the equipment and expands the dimensions of multi-source data in representing the operating status of the equipment.

[0018] Step 104: Collect time-series operational status data reflecting equipment operation actions and energy consumption. Specifically, this includes: synchronously collecting time-series operational status data reflecting equipment operation actions and energy consumption through the system controller and energy consumption monitoring module. Among them, operation action data includes equipment action signals such as solenoid valve switching status, unloading valve speed, and fan speed. Energy consumption data includes energy consumption parameters such as motor operating current and voltage. During the collection process, ensure the time synchronization of action signals and energy consumption parameters, take into account the correlation between equipment action logic and energy consumption changes, comprehensively capture the dynamic feedback of system operation, further broaden the operating condition coverage dimension of multi-source data, and enable various types of data to completely map the overall system operation, providing data support for subsequent construction of the correlation between equipment action, energy consumption, pressure, and flow.

[0019] Step 105 involves aggregating and initially aligning the time-series pressure data, flow rate data, temperature data, and operational status data to obtain multi-source heterogeneous monitoring data and corresponding time alignment mapping relationships. Specifically, this includes: based on the acquisition results of various types of time-series data, aggregating the time-series pressure data, flow rate data, temperature data, and operational status data to a unified data acquisition terminal; using the main control clock of the pneumatic conveying system as a reference, performing initial time alignment processing on various types of data; eliminating time dimension deviations of data from different sources by calibrating the timestamps of each sensor data; establishing a one-to-one time alignment mapping relationship between various types of data; achieving preliminary standardization of multi-source heterogeneous data; and providing a unified time reference for subsequent spatiotemporal precise alignment and feature fusion in the data preprocessing stage. This effectively avoids the loss of associated information caused by data time misalignment and ensures the continuity and effectiveness of multi-source data correlation analysis.

[0020] In a preferred embodiment of the present invention, step 200 above, which preprocesses the multi-source heterogeneous monitoring data to obtain a spatiotemporal aligned feature matrix, includes: Step 201 involves performing a quality assessment on the multi-source heterogeneous monitoring data, identifying and labeling missing data segments and noise interference segments to obtain a data quality labeling matrix. Specifically, this includes: based on the multi-source heterogeneous monitoring data and time alignment mapping relationship obtained in step 105, conducting comprehensive quality assessments on time-series pressure data, time-series flow data, time-series temperature data, and time-series operating status data. The assessment process combines the acquisition characteristics of each data source with preset quality thresholds. These preset quality thresholds are pre-set based on the accuracy level of each sensor and the fluctuation range of historical operating data. This accurately identifies and labels missing data segments and noise interference segments in each type of data. Missing data segments are clearly labeled, including the missing duration, corresponding time node, and data source. Noise interference segments are labeled, distinguishing the interference intensity level and the affected data range. This forms a structured data quality labeling matrix, which is synchronously associated with the time reference in the time alignment mapping relationship. This provides a target for targeted preprocessing in subsequent stages, ensuring orderly and accurate preprocessing operations and avoiding the loss of effective information caused by indiscriminate processing. This lays the foundation for subsequent analysis of the spatial distribution and dynamic evolution characteristics of pressure signals.

[0021] Step 202: Based on the data quality labeling matrix, for the segments in the time-series pressure data and time-series flow data marked as noise interference, dynamically adjust the window parameters and smoothing intensity according to the statistical characteristics of the data within each sliding window. This achieves the suppression of high-frequency random noise while selectively retaining and enhancing the low-frequency trend components that characterize equipment condition deterioration and fault evolution, resulting in denoised time-series pressure data and time-series flow data. Specifically, this includes: based on the data quality labeling matrix, identifying the noise interference segments in the time-series pressure data and time-series flow data, and performing targeted noise reduction operations on these segments using a sliding window processing method. The length parameter and smoothing intensity of the sliding window are not fixed but are adjusted according to the mean value of the data within each window. The statistical characteristics such as variance are dynamically adjusted, and the sliding window length can be dynamically adapted within the range of 1 to 10 seconds. The smoothing intensity is adjusted accordingly with the weight coefficient. When the data fluctuation within the window is drastic and exceeds the preset fluctuation threshold, the window length is reduced and the smoothing intensity is increased. When the data shows a stable trend and the fluctuation is within a reasonable range, the window length is increased and the smoothing intensity is reduced. In this way, while effectively suppressing high-frequency random noise and filtering meaningless data fluctuations, the low-frequency trend components that characterize equipment condition deterioration and fault evolution are selectively retained and enhanced. Finally, the time-series pressure data and time-series flow data are obtained after noise reduction, ensuring the purity and integrity of the core features of the pressure and flow data, and laying a solid data foundation for subsequent mining of the spatial distribution and dynamic evolution characteristics of pressure signals.

[0022] Step 203: Based on the data quality labeling matrix, for the time periods marked as having missing data, corresponding completion strategies are applied according to different data source types to obtain completed time-series pressure data, time-series flow data, and time-series temperature data. Specifically, this includes: also based on the data quality labeling matrix obtained in step 201, for the marked data-missing time periods, corresponding completion strategies are applied for classification processing based on the time-series characteristics and cross-dimensional correlation patterns of different data sources. For time-series pressure data and time-series flow data, the change trends and rates of their adjacent valid time periods and the data under the same working conditions are considered. Historical data features within 2 hours are used to complete the trend data, ensuring that the completed data closely matches the actual operating patterns. For time-series temperature data, the data is completed based on the heat conduction characteristics of key rotating equipment, the correlation of operating load, and the equipment's own heat dissipation patterns. During the completion process, the original distribution patterns and trends of the data are followed, and a smooth connection is achieved through gradient transitions between adjacent valid data segments. This ensures that there are no abrupt deviations between the completed data and the valid data segments, thereby obtaining the completed time-series pressure data, time-series flow data, and time-series temperature data. This ensures the integrity and reliability of multi-source data and provides support for subsequent multi-dimensional data fusion analysis.

[0023] Step 204: Based on the switch signals of the time-series operating status data, perform template matching and status encoding processing based on event windows to convert them into a numerical sequence consistent with the continuous time-series data in terms of time and numerical dimensions, obtaining the encoded operating status sequence. Specifically, this includes: for the switch signals in the time-series operating status data, firstly, dividing the event windows according to the operating action cycle of each device in the pneumatic conveying system, with each event window corresponding to a complete device action process. The duration of the event window is typically set to 30 to 60 seconds based on the longest cycle of a single complete action of the device. Then, template matching is performed on the switch signals within the event windows based on a preset standard action template. The standard action template is constructed based on the equipment's factory parameters and action samples during normal operation. It identifies the start and end times, duration, and completeness of various action states. Subsequently, the matched state information is encoded. A gradient numerical encoding method is used to convert discrete switch signals into continuous numerical sequences. The encoding process ensures that the numerical value matches the state level and action amplitude. The more critical the action and the larger the amplitude, the wider the corresponding numerical range. Finally, an encoded operating state sequence is obtained that is consistent with the continuous time series data in both time and numerical dimensions. This completely eliminates the fusion barriers caused by differences in data types and enables collaborative analysis of discrete and continuous signals.

[0024] Step 205 involves integrating the denoised time-series pressure data and time-series flow data, as well as the completed time-series pressure data and time-series flow data, to form the final pressure data and final flow data. Specifically, this includes: using the time alignment mapping relationship established in step 105 as a unified benchmark, integrating the denoised time-series pressure data and time-series flow data obtained in step 202 with the completed time-series pressure data and time-series flow data obtained in step 203 on a time-node-by-time basis. During the integration process, consistency checks are performed on the same source data at the same time node. A deviation threshold comparison method is used to verify data consistency. When the deviation of the same source data exceeds a preset range, the denoised data is used as the benchmark to correct and complete the data. At the same time, redundant data that is repeatedly collected is removed, and deviation data caused by sensor delay is corrected to ensure the continuity of data time sequence and numerical accuracy. Finally, a unified and standardized final pressure data and final flow data are formed, providing high-quality core data support for subsequent multi-dimensional data fusion.

[0025] Step 206 involves fusing the final pressure data, final flow data, completed time-series temperature data, and coded operating status sequence to obtain an integrated feature set. Specifically, this includes: synchronously integrating the final pressure data, final flow data, completed time-series temperature data, and coded operating status sequence based on a unified time axis; synchronously associating the equipment location information and real-time operating condition information corresponding to various data types during the fusion process; the equipment location information corresponds to the pipeline sections and equipment locations where each sensor is deployed; and the operating condition information is associated with real-time delivery volume and material characteristic parameters; supplementing the inherent relationships and physical logic between different data dimensions; and forming an integrated feature set covering pressure, flow, temperature, and equipment operating status. This comprehensively and accurately maps the system operating conditions and the status of each key component, providing complete feature support for subsequent spatiotemporal feature mining.

[0026] Step 207 involves integrating the integrated feature set with the time alignment mapping relationship. A unified resampling operation is used to normalize all data to the same time reference and sampling frequency to obtain a spatiotemporal alignment feature matrix. Specifically, this includes: deeply integrating the integrated feature set with the time alignment mapping relationship established in step 105; firstly, a spatial grid partitioning algorithm is introduced to perform structured grid partitioning of the pneumatic conveying system's pipelines and core equipment areas; the spatial grid partitioning algorithm is a spatial discretization method that relies on physical entity dimensions, structural outlines, and monitoring requirements to decompose the target space into several non-overlapping and non-omission structured units. Its core is to achieve precise binding of data and spatial location through rule-based partitioning, assigning spatial dimension identifiers to multi-source data.

[0027] The division is based on the actual dimensions of the pipeline, the structural outline of the equipment, and the deployment location of the sensors, combined with the specific specifications of the system: the main pipeline is 80 meters long and has an inner diameter of 0.8 meters. Along the pipeline axis, it is equally divided into 80 axial grid units, one unit per meter, numbered from No. 1 at the inlet to No. 80 at the outlet. Radially, the pipeline cross-sectional area is evenly divided into four sector units, each corresponding to a 90-degree central angle, numbered 1 to 4 from the inside to the outside of the pipeline, with each sector having a radial width of 0.2 meters. For special sections such as 90-degree bends at axial distances of 30 to 32 meters and reducers at axial distances of 55 to 57 meters (with inner diameter reduced from 0.8 meters to 0.5 meters), the axial division interval is reduced to 0.5 meters. The bend section is divided into 4 axial units, and the reducer section is divided into 4 axial units. The radial division rules are consistent with those of the straight pipe section. The core pressure vessel is a silo pump with a cavity diameter of 2 meters and a height of 3 meters. The equipment cavity structure is divided into 12 cubic units with a length, width, and height of 1 meter each. Each unit is labeled with three-dimensional coordinate information.

[0028] The main pipeline is 80 meters long and has an inner diameter of 0.8 meters. Pressure monitoring points are installed every 10 meters, numbered P1 to P8, corresponding to axial distances of 10 meters, 20 meters, 30 meters, 40 meters, 50 meters, 60 meters, 70 meters, and 80 meters. Each monitoring point is located in radial zone 2 of the pipeline. During the grid division process, the positions of each monitoring point were precisely aligned to ensure that each pressure and flow monitoring point falls precisely into a unique grid cell. For example, pressure monitoring point P3, located at 30 meters axially, is precisely at the beginning of the bend section. Dividing the axial spacing according to the special section of 0.5-meter intervals, it falls into grid cell number 60 (the first cell of the bend section) and radial zone 2. Unit; The temperature monitoring point at the top of the silo pump is located 1 meter above the cavity, corresponding to the cubic unit X1, Y1, Z2 of the silo pump; All grid units cover the entire conveying system without overlap or omission. After the division is completed, the time series data of each integrated feature set are allocated to the grid unit according to the installation position of the corresponding monitoring point, and a one-to-one correspondence between the data and the spatial grid unit is established. The spatial coordinate information of each grid unit and the relationship with the adjacent units are marked simultaneously. For example, the adjacent units of unit 25 in the straight pipe section are units 24 and 26, the adjacent units of unit 2 in the radial direction are units 1 and 3, and the adjacent units of unit 60 in the elbow section are units 59 and 61.

[0029] Subsequently, a unified resampling operation was adopted to normalize the data in all grid cells to the same time base and sampling frequency. The resampling adopted a linear interpolation method to ensure that the data trend was not distorted. The unified sampling frequency was set to a common multiple of the original sampling frequencies of each data source to balance data accuracy and processing efficiency. During the resampling process, the temporal characteristics of the data in each grid cell and the cross-dimensional correlation of data between different grid cells were not destroyed. This enabled the data to achieve accurate alignment under the premise of unified time dimension and clear spatial dimension positioning. Finally, a structured spatiotemporal alignment feature matrix was obtained. This matrix can completely preserve the spatial distribution information and dynamic evolution trend of pressure signals in each grid cell. It provides highly adaptable, standardized basic data with spatial dimension identification for subsequent spatiotemporal feature set construction and joint learning, effectively supporting the mining and utilization of the spatial distribution and dynamic evolution characteristics of pressure signals.

[0030] In this embodiment of the invention, quality assessment is conducted on multi-source heterogeneous monitoring data to identify and label missing data segments and noise interference segments, forming a data quality labeling matrix. This determines the anomaly types and locations of various data types, providing a basis for subsequent targeted preprocessing. Based on the data quality labeling matrix, the sliding window parameters and smoothing intensity are dynamically adjusted for noise interference segments. The processing method is optimized by combining the statistical characteristics of the data within the window. This effectively suppresses high-frequency random noise while selectively retaining and enhancing low-frequency trend components that characterize equipment condition deterioration and fault evolution, balancing data purity and the integrity of core features, and strengthening the data's ability to characterize the system state. Based on the data quality labeling matrix, corresponding completion strategies are adopted according to the data source type for missing data periods, adapting to the temporal characteristics and correlation patterns of different data. This ensures the rationality and continuity of the completed data, filling data gaps while maintaining the original data distribution characteristics, providing complete and reliable basic data for multi-source data fusion. The operation status data is then processed based on event windows. Template matching and status encoding are performed on the switching signals to convert discrete switching quantities into numerical sequences, achieving consistency and adaptation with continuous time-series data in terms of time and numerical dimensions. The denoising and completed time-series pressure and flow data are integrated, and the optimized data results are summarized, invalid interference information is eliminated, and core effective data is refined to form standardized final pressure and flow data, laying a high-quality data foundation for subsequent multi-dimensional data fusion. The final pressure and flow data, completed temperature data, and encoded operating status sequences are fused to construct an integrated feature set, integrating multi-dimensional operating conditions and equipment status information, strengthening the correlation and connection between data, and expanding feature dimensions to comprehensively map the entire system operation. The integrated feature set is uniformly resampled based on time alignment mapping relationships, regularizing all data to the same time base and sampling frequency, generating a spatiotemporal aligned feature matrix, strengthening the spatiotemporal correlation and standardization of the data, and providing highly adaptable basic data support for subsequent spatiotemporal feature set construction and joint learning.

[0031] In a preferred embodiment of the present invention, step 300 above involves constructing a spatiotemporal feature set based on a spatiotemporal alignment feature matrix; wherein each element in the spatiotemporal feature set corresponds to a key monitoring point of the system, and its content is aggregated from time-series monitoring data of the corresponding location; the correlation between elements is defined based on the physical connection of equipment and the dependence on the process flow, including: Step 301: Analyze the spatiotemporal alignment feature matrix. Based on the preset monitoring point identifiers, separate the multivariate time-series data slices corresponding to each monitoring point in the spatiotemporal alignment feature matrix to obtain independent time-series data fragments for each monitoring point. Specifically, this includes: firstly, performing structured analysis on the spatiotemporal alignment feature matrix; the preset monitoring point identifiers need to be pre-set in the early stages of system deployment, taking into account the layout diagram of the conveying system equipment, the monitoring point installation list, and data acquisition requirements. The setting rules are centered on clearly distinguishing the attributes of each monitoring point, using a combination of letters and numbers, with each component corresponding to a clear attribute meaning; among them, equipment types are represented by uppercase letters, with the silo pump corresponding to C. The system uses a combination of numbers to represent the locations of monitoring points. Pipeline axial numbers are incremented from 1 from the inlet to the outlet. Radial partitions are divided into zones 1 to 4 from the inside to the outside of the pipe cross-section. The location identifier format is axial number followed by radial partition. Data categories are represented by lowercase letters: pressure corresponds to y, flow rate to l, temperature to w, and coded operating status to z. Complete identifiers are combined in the order of equipment type, location identifier, and data category to ensure that each type of data collected at each monitoring point has a unique identifier. After pre-setting, the identifier is entered into the system database and synchronously linked to the column dimensions of the spatiotemporal alignment feature matrix for use during matrix parsing.

[0032] The matrix's row dimension corresponds to the time nodes after unified resampling, while the column dimension contains the aforementioned preset unique identifiers for each monitoring point and corresponding multivariate data items. The identifier information respectively represents the deployment location of the monitoring point, such as pipeline axial number and radial partition; the corresponding equipment type, such as silo pump, fan, and elbow; and the data collection category, such as pressure, flow rate, and temperature. Based on this identifier, the data within the matrix is ​​sliced ​​and separated one by one. By filtering the corresponding row data according to the identifier column, the multivariate time series data of each monitoring point is extracted from the matrix separately. During the extraction process, the data integrity is simultaneously verified to ensure that the time series data segment of each monitoring point contains all variable data of the corresponding time node. This achieves accurate isolation and classification of data from different monitoring points, resulting in independent time series data segments for each monitoring point. Each segment completely retains the time series correlation of the corresponding monitoring point's pressure, flow rate, temperature, and coded operating status data, providing targeted basic data for subsequent feature extraction by monitoring point, and laying a solid foundation for mining spatial correlation information between monitoring points.

[0033] Step 302: For each independent time-series data segment at each monitoring point, a sliding time window is used for traversal. Within each time window, multiple predefined statistical features of the data segment are calculated and combined sequentially to form a feature vector. Specifically, this includes: for each independent time-series data segment at each monitoring point, a sliding time window is used for segment-by-segment traversal. The duration of the sliding time window is set to 10 seconds, taking into account the characteristics of the pneumatic conveying system's 10-second operating cycle and the fact that fault precursors typically require 5 to 10 seconds to evolve. The window sliding step size is adapted to a uniform sampling frequency of 5Hz and is set to 5 seconds to avoid feature redundancy and ensure time sequence continuity. The predefined statistical features need to be pre-defined in the early stages of system debugging, taking into account the operating characteristics of the pneumatic conveying system, the characteristic manifestations of fault precursors, and the subsequent feature modeling requirements. The core principle is to fully... The system captures the distribution patterns and temporal trends of data while maintaining the simplicity and specificity of feature calculations. Combining the characteristics of data types such as pressure, flow rate, temperature, and coded operating status, it identifies five core statistical features: mean, variance, peak value, trough value, and trend slope. Each feature's predefined characteristics correspond to clear fault characterization needs: the mean reflects the overall level of data within a window, suitable for steady-state judgments of parameters such as pressure and flow rate; variance reflects the amplitude of data fluctuations, corresponding to the unstable parameter characteristics in the early stages of a fault; peak and trough values ​​capture data extrema, suitable for parameter mutations caused by faults such as local blockages and leaks; and trend slope quantifies the rate of data change, corresponding to the gradual evolution of fault precursors. Once the features are predefined, the calculation rules and combination order are solidified for subsequent window traversal calculations.

[0034] Within each time window, the aforementioned predefined statistical characteristics are calculated for the data segment covered by the window. The mean is obtained by summing all data within the window and dividing by the number of data points. The variance is obtained by summing the squares of the differences between each data point and the mean and dividing by the number of data points. The peak value is the maximum value of the data within the window, and the trough value is the minimum value of the data within the window. The trend slope is calculated by dividing the difference between the first and last data points of the window by the window duration of 10 seconds. This comprehensively captures the distribution characteristics and changing trends of the data within the window. After the calculation is completed, the various statistical characteristics are combined in a fixed predefined order of mean, variance, peak value, trough value, and trend slope to form the feature vector of the corresponding window. This transforms continuous time-series data into a structured feature expression, enhances the data's ability to represent the operational status of monitoring points, and provides core elements for the subsequent construction of node feature matrices.

[0035] Step 303: Stack the feature vectors in chronological order to form a node feature matrix for the monitoring point. The rows of the node feature matrix correspond to time windows, and the columns correspond to the extracted statistical feature dimensions. Specifically, this includes stacking the feature vectors of each monitoring point corresponding to each time window vertically in chronological order from early to late to form a node feature matrix specific to that monitoring point. The number of rows in the matrix is ​​consistent with the total number of time windows. Each row corresponds to the feature vector of one time window, fully recording the multi-dimensional feature information of the monitoring point under that window. Each column corresponds to a predefined statistical feature dimension, clearly presenting the dynamic change trajectory of various features over time. During the matrix construction process, the start timestamp of the corresponding time window is added to the first column of each row to retain the accurate correspondence between the time window and the original time series data, ensuring that the features are closely bound to the time dimension. The matrix dimension is dynamically adjusted according to the actual running time. If the system runs continuously for 1 hour, a node feature matrix of 720 rows and 5 columns is generated, providing a standardized node time series feature carrier for subsequent spatiotemporal joint learning, while also meeting the analysis needs of the dynamic evolution of the system state.

[0036] Step 304: Based on the process flow diagram and equipment layout diagram of the pneumatic conveying system, construct a topology diagram representing the physical connections and process dependencies between system components; convert the topology diagram into an adjacency matrix, where the rows and columns of the matrix correspond to system monitoring points. If there is a direct physical connection or a tightly coupled process dependency between the components represented by two monitoring points, the corresponding position in the adjacency matrix is ​​marked as a connection state. Specifically, based on the process flow diagram and equipment layout diagram of the pneumatic conveying system, systematically analyze the installation position, physical connection path, and upstream and downstream process dependencies of each component in the system. Pipeline sections are numbered sequentially from the silo pump outlet to the preheater inlet. The fan and the corresponding pipeline inlet, and the solenoid valve and the silo pump outlet are all determined to be directly physically connected. The monitoring point at the silo pump outlet and the downstream... The monitoring points of the first pipeline section, the elbow monitoring point, and the monitoring points of the preceding and following straight pipe sections are all determined to have a tightly coupled process dependency relationship. Based on this, a topology graph representing the relationship between components is constructed. Each node in the topology graph corresponds to a key monitoring point in the system, and the edges between nodes correspond to the direct relationship between the components to which the monitoring points belong. Subsequently, the topology graph is transformed into an adjacency matrix. The rows and columns of the matrix are arranged in the order of the monitoring point numbers, corresponding one-to-one with the system monitoring points. If the components represented by two monitoring points have a direct physical connection or a tightly coupled process dependency relationship, they are marked as 1 in the corresponding row and column position of the adjacency matrix. If there is no relationship, it is marked as 0, forming an N×N dimensional adjacency matrix (N is the total number of monitoring points). This transforms the physical and process logic of the system into a data-driven relationship, determines the spatial relationship attributes between monitoring points, and makes up for the shortcomings of the existing static model.

[0037] Step 305 involves using the node feature matrix as the feature content of nodes in the spatiotemporal feature set, and the adjacency matrix as the definition of the association relationship between nodes in the spatiotemporal feature set, together forming the spatiotemporal feature set. Specifically, this includes: organizing the node feature matrices of each monitoring point into a set of node feature matrices according to the monitoring point number, serving as the core feature content of the corresponding nodes in the spatiotemporal feature set, fully carrying the temporal feature information of each monitoring point, with each node feature matrix including the location and equipment association information of its respective monitoring point; using the adjacency matrix as the definition basis for the association relationship between nodes in the spatiotemporal feature set, clearly defining the spatial association logic between each monitoring point and other monitoring points, and adding a lookup table of monitoring point numbers and topology graph nodes to facilitate subsequent data retrieval and association analysis; the two are integrated to jointly construct the spatiotemporal feature set, which not only retains the temporal dynamic features of each monitoring point, but also accurately depicts the spatial association attributes between monitoring points, achieving deep fusion of spatiotemporal dimension data, providing highly adaptable feature inputs for subsequent spatiotemporal feature joint learning of graph neural network models, and effectively supporting the mining and utilization of the spatial distribution and dynamic evolution characteristics of pressure signals.

[0038] In this embodiment of the invention, the spatiotemporal alignment feature matrix is ​​analyzed, and the multivariate time-series data of each monitoring point are separated according to the preset monitoring point identifier slices to obtain independent time-series data segments. This achieves accurate splitting and isolation of data from different monitoring points, ensuring the independence and specificity of data from each monitoring point, and laying the foundation for subsequent node feature extraction. By traversing the independent time-series data segments of each monitoring point through a sliding time window, multiple predefined statistical features of the data within the window are calculated and combined sequentially to form feature vectors. This captures the characteristic patterns of time-series data in the local time dimension, transforming continuous time-series data into structured feature expressions and enhancing the data's ability to represent the state of the monitoring points. The feature vectors are stacked in chronological order to form a node feature matrix, so that the matrix rows correspond to time windows and the columns correspond to statistical features. The feature dimension standardizes the arrangement of feature data, clearly presenting the dynamic change trajectory of monitoring point features over time, providing a standardized node feature carrier for subsequent spatiotemporal correlation analysis. A topology graph is constructed based on the system process flow diagram and equipment layout diagram, transforming it into an adjacency matrix to represent the physical connections and process dependencies between monitoring points. This converts the physical logic and process correlation of the actual system operation into a data-driven correlation form, determining the intrinsic connections between nodes and aligning with the essence of system operation. Using the node feature matrix as the node feature content and the adjacency matrix to define the correlation between nodes, a spatiotemporal feature set is constructed, integrating the temporal and spatial correlation features of nodes to achieve deep fusion of spatiotemporal dimension data, providing comprehensive and system-logic-aligned feature support for subsequent joint learning.

[0039] In a preferred embodiment of the present invention, step 400, based on real-time data from multiple discrete pressure monitoring points on the main pipeline, extracts a distribution feature vector and uses it to dynamically correct the spatiotemporal feature set to obtain a dynamically weighted adjacency matrix, including: Step 401: Identify nodes corresponding to multiple preset discrete pressure monitoring points on the main pipeline from the spatiotemporal feature set, and extract real-time pressure time-series data segments of the corresponding nodes in the current time window. Specifically, this includes: based on the spatiotemporal feature set, combined with the equipment layout diagram of the main pipeline and the list of preset pressure monitoring points; the preset discrete pressure monitoring points need to be pre-set during the system installation and commissioning phase, taking into account the structural characteristics of the main pipeline, high-fault areas, and pressure transmission patterns. The core principle is to fully cover the key sections of the main pipeline and accurately capture changes in the spatial distribution of pressure; specifically, when setting them, priority is given to setting monitoring points in sections where pressure is prone to sudden changes and faults are frequent, such as the feed end, discharge end, before and after bends, and both sides of reducers. At the same time, the continuity of pressure transmission is taken into account, and other monitoring points are added at equal intervals. Finally, eight discrete pressure monitoring points are preset on the main pipeline, numbered P1 to P8 in sequence from the feed end to the discharge end; each monitoring point is equipped with a pressure sensor, which is synchronously configured for it. Each node is assigned a unique identifier, which is prefixed with a corresponding monitoring point number. After presetting, a list of pre-set pressure monitoring points is compiled, entered into the system database, and associated with the main pipeline equipment layout diagram for subsequent node matching. Nodes corresponding to these pre-set discrete pressure monitoring points on the main pipeline are identified one by one, with accurate matching achieved by comparing the node identifier prefix with the monitoring point number. Subsequently, real-time pressure time-series data segments of these corresponding nodes are extracted within the current time window, which is the 10-second duration set in step 302. Calculated at a uniform sampling frequency of 5Hz, each data segment contains 50 continuous pressure sampling points. During the extraction process, data integrity is simultaneously verified to ensure no missing sampling points and that the timestamp is perfectly aligned with the current window. At the same time, outliers exceeding the normal pressure range are removed to ensure that the obtained pressure data accurately reflects the current pressure operating status of the main pipeline, providing a high-quality core data source for subsequent analysis of pressure spatial distribution characteristics and dynamic adjustment of node association weights.

[0040] Step 402: Based on the real-time pressure time-series data segment, calculate the real-time pressure gradient value between every two adjacent pressure monitoring points, and construct a pressure gradient vector based on the real-time pressure gradient values ​​of all adjacent point pairs. Specifically, this includes: based on the real-time pressure time-series data segment, first calculate the average pressure value of each monitoring point within the current 10-second window. The calculation method is to sum the pressure values ​​of 50 sampling points and divide by the number of sampling points to obtain the average real-time pressure of each monitoring point; then, according to the physical arrangement order of P1 to P8 on the main pipeline, calculate the real-time pressure gradient value between every two adjacent monitoring points sequentially. The calculation method is to subtract the average pressure value of the previous monitoring point from the average pressure value of the subsequent monitoring point. The average pressure value is calculated, and the result retains the positive or negative sign to represent the pressure rise and fall trend. For example, if the average pressure value of P2 is 0.7 MPa and P1 is 0.6 MPa, then the gradient value between them is 0.1 MPa. If the average pressure value of P3 is 0.65 MPa, then the gradient value between P3 and P2 is -0.05 MPa. After the pressure gradient values ​​of all 7 pairs of adjacent monitoring points are calculated, these gradient values ​​are arranged in the order of P1 and P2, P2 and P3, and so on up to P7 and P8 to construct a 7-dimensional pressure gradient vector. This vector transforms the discrete average pressure data into a quantifiable spatial distribution feature of pressure, accurately presenting the pressure variation pattern along the main pipeline.

[0041] Step 403: Normalize the pressure gradient vector to obtain a normalized gradient feature vector; input the normalized gradient feature vector into a preset exponential mapping function. This exponential mapping function maps the gradient elements in the vector to corresponding dynamic adjustment coefficients through exponential operations based on the magnitude and sign of each gradient element. The calculation rule of the exponential mapping function is as follows: the adjustment coefficient value calculated for the connecting edges corresponding to adjacent monitoring points with larger absolute gradient values ​​is correspondingly increased. Specifically, this includes: first, performing linear normalization on the pressure gradient vector; finding the maximum and minimum values ​​in the vector; then subtracting the minimum value from each gradient value and dividing by the maximum and minimum values. The difference between the smallest values ​​is used to map all gradient values ​​to the range of 0 to 1, eliminating the dimensional influence caused by different pressure levels. Based on the actual working conditions of the pneumatic conveying system, the gradient value range of this pressure gradient vector is set to -0.05MPa to 0.1MPa, with a maximum value of 0.1MPa and a minimum value of -0.05MPa. The original gradient values ​​corresponding to the 7 pairs of adjacent monitoring points are as follows: P1 and P2 = 0.02MPa, P2 and P3 = 0.08MPa, P3 and P4 = -0.03MPa, P4 and P5 = 0.1MPa, P5 and P6 = 0.04MPa, P6 and P7 = -0.05MPa, and P7 and P8 = 0.06MPa.

[0042] The formula used for linear normalization in this study is: The meanings of each parameter are as follows: The normalized gradient value has a mapping range of 0 to 1; The original gradient value for each pair of adjacent monitoring points This is the minimum value in the pressure gradient vector; Let P be the maximum value in the pressure gradient vector. Calculate using this formula (results rounded to two decimal places, approximate values ​​are marked with ≈, exact values ​​are marked with =). The normalized values ​​for each gradient are as follows: P1 and P2: (0.02+0.05) / (0.1+0.05)≈0.47; P2 and P3: (0.08+0.05) / (0.1+0.05)≈0.87; P3 and P4: (-0.03+0.05) / (0.1+0.05)≈ 0.13; P4 and P5 are (0.1+0.05) / (0.1+0.05)=1.0; P5 and P6 are (0.04+0.05) / (0.1+0.05)=0.6; P6 and P7 are (-0.05+0.05) / (0.1+0.05)=0; P7 and P8 are (0.06+0.05) / (0.1+0.05)≈0.73, thus obtaining a standardized normalized gradient feature vector containing 7 elements.

[0043] After obtaining the standardized normalized gradient feature vector, it is input into a preset exponential mapping function. The core function of the preset exponential mapping function is to transform the normalized gradient feature vector into a dynamic adjustment coefficient that adapts to the adjustment of node association weights. By strengthening the differences between different gradients, it highlights the importance of associations in areas with significant pressure changes. At the same time, it uses the positive and negative signs of the gradient to distinguish the impact of pressure rise and fall trends on node associations, making the weight adjustment more in line with the real-time pressure distribution characteristics of the system. The value generated by this function is the dynamic adjustment coefficient, which is used to scale the connection weights of corresponding adjacent monitoring point pairs in the adjacency matrix element by element, so as to realize the dynamic update of the association weights between nodes with the real-time pressure gradient, breaking the limitations of the static association model.

[0044] The calculation rules for this function have been pre-defined based on the pressure change characteristics of the pneumatic conveying system. The core logic is based on the natural exponent, and the exponent mapping function expression is as follows: Simultaneously, the fine-tuning term Δ is determined by the sign of the gradient, thus completing the calculation of the final dynamic adjustment coefficient; the meanings of each parameter are as follows: This represents a dynamic adjustment coefficient used to scale the connection weights of adjacent monitoring point pairs. It is a natural constant, used as the basis for exponential operations, and is used to enhance the difference in normalized gradient values; This is a preset coefficient amplification factor, with a value of 2, used to further enhance the influence of the normalized gradient value on the dynamic adjustment coefficient. is the value of a single element in the normalized gradient feature vector, that is, the result of normalizing the original gradient value for each pair of adjacent monitoring points; Δ is the gradient sign fine-tuning term, which is 0.1 when the original gradient value is positive and -0.1 when it is negative, and is used to distinguish the influence of pressure rise and fall trends on weight adjustment.

[0045] Substituting the seven normalized values ​​and corresponding fine-tuning terms for precise calculation (natural constant e≈2.71828, base coefficients rounded to four decimal places, final results rounded to two decimal places), the base coefficients are obtained. After fine-tuning, the final seven dynamic adjustment coefficients are: P1 and P2 (positive gradient, =0.47), ω≈e (2×0.47) +0.1≈e 0.94 +0.1≈2.5611+0.1≈2.66; P2 and P3 (positive gradient, =0.87), ω≈e (2×0.87) +0.1≈e 1.74 +0.1≈5.7031+0.1≈5.80; P3 and P4 (negative gradient, =0.13), ω≈e (2×0.13) -0.1≈e 0.26 -0.1≈1.2969 -0.1≈1.20; P4 and P5 (positive gradient, =1.0), ω≈e (2×1.0) +0.1≈e 2.0 +0.1≈7.3891 +0.1≈7.49; P5 and P6 (positive gradient, =0.6), ω≈e (2×0.6) +0.1≈e 1.2 +0.1≈3.3201+0.1≈3.42; P6 and P7 (negative gradient, =0), ω≈e (2×0) -0.1 ≈ 1.0 - 0.1 = 0.90; P7 and P8 (positive gradient, =0.73), ω≈e (2×0.73) +0.1≈e 1.46 +0.1≈4.3175+0.1≈4.42; This strengthens the weight of areas with significant pressure changes in node correlation, while also distinguishing the impact of pressure rise and fall trends.

[0046] Step 404: Using the dynamic adjustment coefficients, the connection weights of corresponding edges in the adjacency matrix are scaled element-wise to obtain a corrected adjacency matrix. Specifically, this includes: based on the seven dynamic adjustment coefficients generated in step 403, and combined with the N×N adjacency matrix obtained in step 304 (where N=8, corresponding to eight discrete pressure monitoring points P1 to P8 on the main pipeline, with the matrix dimension matching the number of monitoring points), the connection weights of corresponding edges in the matrix are scaled element-wise. The weight values ​​of the original connection states in the adjacency matrix are uniformly set to 1.0. Weight adjustment is only applied to adjacent monitoring points, while non-adjacent monitoring points retain their original zero values, ensuring the matrix structure matches the actual connection relationship of the monitoring points. Specifically, the scaling method is as follows: the seven dynamic adjustment coefficients generated in step 403 correspond to P1 and P2, P2 and P3, P3 and P4 respectively. 4. The seven pairs of adjacent monitoring points, P4 and P5, P5 and P6, P6 and P7, and P7 and P8, correspond to 14 symmetric elements in the N×N adjacency matrix: (1,2), (2,1), (2,3), (3,2), (3,4), (4,3), (4,5), (5,4), (5,6), (6,5), (6,7), (7,6), (7,8), and (8,7). (The adjacency matrix is ​​an undirected matrix with symmetric element values.) The corresponding dynamic adjustment coefficient is multiplied by the original weight value of 1.0 to obtain the scaled weight value. For example, the weight of the element corresponding to P1 and P2 is adjusted to 1.0×2.66=2.66, and the weight of the element corresponding to P2 and P3 is adjusted to 1.0×5.80=5.80. The scaling of the weights of all adjacent edges is completed in sequence to obtain the corrected adjacency matrix, realizing the dynamic update of node association weights with real-time pressure gradient.

[0047] Step 405: Keeping the node feature matrix unchanged, the modified adjacency matrix is ​​used as the optimized inter-node association relationship and combined with the node feature matrix to obtain a dynamic weighted adjacency matrix. Specifically, considering that the node feature matrix constructed in step 303 already fully carries the temporal feature information of each monitoring point, with each node feature matrix consisting of 720 rows and 5 columns, each row corresponding to a feature vector of a 10-second time window, and each column corresponding to a statistical feature, this node feature matrix is ​​kept unchanged to preserve the integrity and accuracy of the original temporal features; the 8×8 modified adjacency matrix obtained in step 404 is used as the optimized inter-node association relationship, replacing... The original adjacency matrix is ​​collaboratively combined with the node feature matrices of each monitoring point. During the combination, the node feature matrices are bound one-to-one with the row and column positions in the modified adjacency matrix according to the monitoring point number order, forming a dynamic weighted adjacency matrix. This not only preserves the temporal dynamic characteristics of each monitoring point over time, but also accurately portrays the evolution of the relationship between nodes with the change of pressure distribution through dynamic weight updates. The association weights of corresponding nodes are strengthened for the local pressure gradient anomaly area in the main pipeline, providing more feature inputs that fit the real-time operating state of the system for subsequent spatiotemporal feature joint learning of graph neural network models, effectively supporting the mining of the spatial distribution and dynamic evolution characteristics of pressure signals.

[0048] In this embodiment of the invention, nodes corresponding to preset discrete pressure monitoring points of the main pipeline are accurately identified from the spatiotemporal feature set. Real-time pressure time-series data segments under the current time window are extracted, focusing on the core pressure monitoring area of ​​the system and screening key data sources to provide targeted data support for subsequent pressure distribution feature mining. Based on the real-time pressure time-series data segments, the real-time pressure gradient values ​​between adjacent monitoring points are calculated. Gradient quantization is used to capture the spatial distribution differences of pressure along the pipeline. Then, a pressure gradient vector is constructed from the gradient values ​​of all adjacent points, transforming the discrete pressure data into a structured vector representing the spatial distribution law, thus strengthening the concrete expression of pressure spatial correlation information. The pressure gradient vector is normalized to eliminate the influence of dimensions, resulting in a standardized normalized gradient feature vector. The gradient is then mapped using a preset exponential mapping function. The size and sign of the degree elements are converted into dynamic adjustment coefficients. The coefficient weights are assigned according to the absolute value of the gradient, highlighting the importance of correlations corresponding to significant pressure changes and achieving precise adaptation of the coefficients to the real-time pressure distribution. The connection weights of corresponding edges in the adjacency matrix are scaled element by element using the dynamic adjustment coefficients to generate a modified adjacency matrix. This allows the correlation weights between nodes to be dynamically updated with the real-time pressure gradient, enabling the matrix to reflect the changes in correlations caused by the system pressure distribution in real time and giving the correlation definition temporal adaptability. The node feature matrix is ​​kept unchanged to preserve the integrity of the original temporal features. The modified adjacency matrix is ​​used as the optimized correlation between nodes and is recombined with the node feature matrix to form a dynamic weighted adjacency matrix. This achieves the fusion of temporal features and dynamic spatial correlation features, improving the fit of the feature set to the real-time operating state of the system.

[0049] In a preferred embodiment of the present invention, step 500, which involves inputting the dynamically weighted adjacency matrix into a pre-trained graph neural network model for joint learning of spatiotemporal features to obtain a fused feature vector, includes: Step 501: Based on the dynamically weighted adjacency matrix, initialize the node feature matrix to the node features of each monitoring point in the graph neural network, and load the modified adjacency matrix into the message propagation topology of the model. Specifically, this includes: the graph neural network model is pre-built and pre-trained based on historical multi-source monitoring data of similar pneumatic conveying systems. During the construction phase, a graph convolutional network framework is used, with an initial layer count of 3. The feature dimension is gradually adjusted to 256 dimensions according to data processing requirements. The number of nodes is adapted to the total number of monitoring points in the system. The message propagation mechanism is preset to weighted propagation based on the adjacency matrix, while reserving... An interface adapted to the modified adjacency matrix is ​​provided to ensure that dynamic correlation weights can be loaded. During the training phase, historical data on the normal operation of the system and the evolution of various faults are collected. The spatiotemporal feature set for training is obtained according to the preprocessing steps 100 to 405 of this invention. The training set and validation set are divided into a 7:3 ratio. A stochastic gradient descent optimizer is used to optimize the model parameters with the cross-entropy loss function. The model is trained iteratively for 50 rounds. After each round of training, the model's feature extraction capability is verified using the validation set. The number of network layers, activation function parameters, and convergence method are adjusted until the model can stably capture the spatiotemporal correlation patterns of the system. The pre-training is then completed and the model parameters are saved.

[0050] Based on the dynamic weighted adjacency matrix, the input initialization operation of the pre-trained graph neural network model is carried out. The node feature matrix in the set is initialized one by one according to the monitoring point number order as the initial features of each monitoring point node in the model. The node feature dimension is consistent with the number of matrix columns to ensure that each node fully carries the temporal statistical features of the corresponding monitoring point. At the same time, the modified adjacency matrix is ​​loaded as the message propagation topology structure inside the model. The weight values ​​of each position in the matrix are directly used as the basic weights for message propagation between corresponding nodes. The core parameters and propagation rules of the pre-trained graph neural network model are loaded to make the model propagation logic accurately match the node association relationship driven by the real-time pressure distribution of the system. Based on the spatiotemporal feature patterns of similar systems learned in the pre-training, a basic framework that fits the dynamic operation state of the system is built for subsequent spatiotemporal feature joint learning.

[0051] Step 502: Using the node features and message propagation topology as input, in the first layer of the graph neural network, for each node corresponding to a monitoring point, identify its directly connected first-order neighbor node set according to the message propagation topology; aggregate the features of all neighbor nodes in the first-order neighbor node set to obtain a neighbor feature vector; concatenate the neighbor feature vector with the feature vector of the current node itself to obtain the aggregated feature output of the node in the first layer. Specifically, this includes: using the initialized node features, message propagation topology, and basic parameters loaded by the pre-trained graph neural network model as input, in the first layer of the graph neural network, for each node corresponding to a monitoring point, based on the modified adjacency matrix and combined with the neighbor identification rules fixed in the pre-trained graph neural network model, identifying its directly connected first-order neighbor node set, that is, nodes with non-zero row weights in the adjacency matrix are all included in the current node set. The first-order neighbor range of the previous node is selected, while excluding the node itself to avoid feature redundancy. This identification rule has been verified through pre-training to effectively filter core related nodes. Subsequently, the features of all neighbor nodes in the first-order neighbor node set are aggregated. The average aggregation method determined during pre-training is used to calculate the neighbor feature vector. The features of all neighbor nodes are summed according to their corresponding dimensions, and then divided by the number of neighbor nodes to obtain the mean of each dimension, forming a neighbor feature vector with the same feature dimensions as the current node. This aggregation method can balance the local feature weights and improve the stability of feature fusion. The neighbor feature vector is then concatenated with the current node's own feature vector in the order of feature dimensions. After concatenation, the feature dimensions are doubled, resulting in the aggregated feature output of the node in the first layer. This achieves deep fusion of the node's own information and local neighbor association information, which is consistent with the local feature extraction logic learned by the pre-trained graph neural network model.

[0052] Step 503: The aggregated features are input into a nonlinear transformation layer for processing, updating the feature representation of each node to obtain the node embedding vector of the graph neural network layer. Specifically, this includes: inputting the aggregated features of each node into a preset nonlinear transformation layer for processing. This nonlinear transformation layer is constructed using the ReLU activation function with optimal fit verified during pre-training, and the ReLU activation function is concatenated after the fully connected layer to form a complete feature processing chain of linear transformation and nonlinear activation; the core expression of the ReLU activation function is: The parameters and function mechanisms are explained in detail below: The linear transformation feature vector representing the output of the fully connected layer has the same dimension as the output dimension of the fully connected layer (i.e., 256 dimensions). Each element corresponds to the feature value of the aggregated feature after linear weighting, covering multi-dimensional information such as pressure, flow rate, and spatial location of the pneumatic conveying system. This represents the output feature vector after processing by the ReLU activation function, compared to the input. Having the same dimension is the core intermediate variable for updating the node feature representation; the core logic of the function is to process the input feature vector. Threshold judgment is performed on each element, only non-negative feature values ​​are retained and their original size is maintained, and all negative feature values ​​are directly set to 0, thereby realizing nonlinear feature screening and reconstruction, which is adapted to the sparsity and non-negativity characteristics of fault features in pneumatic conveying systems.

[0053] The ReLU activation function itself has no additional trainable parameters; it only achieves feature transformation through fixed threshold logic. The parameters determined through training iterations and optimization mentioned here refer to the weight matrix (denoted as W) and bias vector (denoted as b) of the fully connected layer before the ReLU activation function in the nonlinear transformation layer. The optimization process is as follows: In the parameter initialization stage, the Xavier initialization method is used to randomly initialize the weight matrix W of the fully connected layer. The dimensions are [1024, 256] (corresponding to 1024-dimensional aggregated features as input and 256-dimensional linear features as output), ensuring that the initial weights follow a normal distribution with a mean of 0 and a variance of 2 / (1024+256), so that the variances of the input and output features remain consistent; the bias vector The dimension is [256,1], and it is uniformly initialized to 0.01 to avoid gradient explosion caused by excessively large initial parameters and slow feature updates caused by excessively small initial parameters, thus ensuring stability in the early stage of training.

[0054] During the training iterations, a dataset constructed from historical fault data (including typical faults such as blockages and leaks) and normal operation data is used as the basis. A 1024-dimensional aggregated feature is input into the fully connected layer, and a linear transformation formula is applied. 256-dimensional linear features were obtained. ,in As the input aggregated feature vector, This represents the dot product operation between a matrix and a vector, and then... As input to the ReLU function ,through The processed output is a 256-dimensional nonlinear feature. Combining the fault type prediction results of subsequent layers of the graph neural network, the cross-entropy loss value is calculated. The loss gradient is then propagated back through the output layer, ReLU layer, and finally the fully connected layer to the weight matrix W and bias vector b. The Adam optimizer is used to iteratively update W and b, with an initial learning rate of 0.001. The learning rate decreases by a factor of 0.9 every 100 iterations. During iteration, the validation set loss is monitored in real time. When the validation set loss shows no decrease for 20 consecutive iterations and the fluctuation is less than 0.0001, parameter updates are stopped, and the optimal value is finally determined. and The optimization process, combined with the nonlinear characteristics of the ReLU function, can effectively avoid the gradient vanishing problem common in deep networks. At the same time, it allows the linear transformation and nonlinear activation to accurately adapt to the feature distribution of the pneumatic conveying system, ensuring the stability and effectiveness of the feature transformation.

[0055] The ReLU activation function is used to perform dimensionality compression and feature reconstruction on the aggregated features. Specifically, the aggregated features at each node are composed of multi-source monitoring data, spatial features, and correlation features, with an original dimension of 1024, containing a large amount of redundant features (such as repetitive spatial coordinate information) and noise information (such as small fluctuation signals from sensors). The optimized fully connected layer is used to perform initial dimensionality compression, mapping the 1024-dimensional aggregated features to a 256-dimensional feature space through a linear transformation. This process is achieved using a weight matrix. Strengthen the weights of fault-related features (such as pressure gradient anomalies), weaken redundant features, and adjust the bias vector. Fine-tuning the feature baseline to offset distribution offset interference enables preliminary screening and dimensionality reduction of high-dimensional features.

[0056] Then, a nonlinear transformation is performed using the ReLU activation function, to... For input Execute element by element Operation: For Elements greater than 0 (corresponding to valid fault characteristics and normal operation characteristics) retain their original values; for Elements less than 0 (corresponding to invalid information such as sensor noise and data acquisition errors) are directly set to 0 to complete feature reconstruction. This process not only eliminates interference from invalid negative features by setting them to zero, but also transforms high-dimensional complex features into compact 256-dimensional features through dimensionality compression. Furthermore, the sparse activation property of the ReLU function can reduce redundant correlations between features and improve the model's sensitivity to fault feature recognition. This 256-dimensional feature has been verified through multiple rounds of training and can not only fully retain the key feature information required for fault recognition, but also reduce the subsequent computational load of the model, balancing feature representation ability and model computational efficiency, without the need for additional adjustments.

[0057] During the processing, feature values ​​are simultaneously pruned to limit feature values ​​that exceed the reasonable range determined by pre-training ([0, 10], which is derived from the statistical analysis of historical feature data of the pneumatic conveying system) to within the range, avoiding interference from extreme values ​​on feature representation. This pruning range can retain effective features and eliminate abnormal interference, thereby updating the feature representation of each node and obtaining the node embedding vector of the graph neural network layer. This makes the node features more adaptable to the feature characterization requirements under the complex operating conditions of the system, and continues the feature optimization capability of the pre-trained graph neural network model.

[0058] Step 504: The node embedding vector is used as input and iteratively processed through a pre-defined number of graph neural network layers. Each layer performs message aggregation and nonlinear transformation operations, ultimately outputting high-order node features containing deep spatiotemporal dependencies of the system. Specifically, this includes: using the node embedding vector as input to the next layer of the graph neural network, and iteratively processing through a pre-trained and optimized 3-layer graph neural network. This number of layers is set in conjunction with the structural complexity and feature mining depth of the pneumatic transport system, and training verification has shown that it can fully explore deep correlations without overfitting; each layer repeatedly performs message aggregation and nonlinear transformation operations. The message aggregation stage uses the average convergence method of the first layer, based on the current layer's topology... The clustering structure aggregates neighbor node features and concatenates them with its own features. This process is consistent with the iterative logic during pre-training, ensuring the continuity of feature extraction. In the nonlinear transformation stage, the ReLU activation function is also used for feature mapping and dimension adjustment, with the parameters retaining the optimal values ​​from pre-training. During the iteration process, the node embedding vectors output by the previous layer serve as the initial node features for the next layer. Relying on the hierarchical feature propagation rules learned by the pre-trained graph neural network model, the deep spatiotemporal dependencies between nodes are progressively mined, gradually strengthening the abstractness and hierarchy of features. Finally, high-order node features containing deep global and local correlations of the system are output, fully leveraging the feature mining advantages of the pre-trained graph neural network model.

[0059] Step 505: Perform global pooling on the high-order node features to aggregate the features of all nodes into a fixed-dimensional graph-level feature vector, resulting in a fused feature vector. Specifically, this involves: performing global pooling on the high-order node features, using a combination of max pooling and average pooling as validated during pre-training. This combination preserves both extreme value information and overall distribution characteristics, making it more suitable for capturing system fault features compared to a single pooling method; calculating the maximum and average values ​​for each feature dimension of the high-order features of all nodes, then concatenating the maximum and average values ​​of all dimensions sequentially to form a 512-dimensional fixed-dimensional graph-level feature vector pre-trained. This dimension, validated through training, can fully encompass global spatiotemporal features and facilitates subsequent model processing; finally, the fused feature vector is obtained. This vector, relying on the feature learning capability of the pre-trained graph neural network model, integrates the temporal dynamic features, spatial correlation features, and deep dependencies of each monitoring point in the system, providing standardized comprehensive feature support for subsequent fault type analysis and location positioning, ensuring that the feature quality matches the generalization capability of the pre-trained graph neural network model.

[0060] In this embodiment of the invention, model initialization is completed based on a dynamically weighted adjacency matrix. The node feature matrix is ​​assigned the corresponding node features of each monitoring point in the graph neural network. The corrected adjacency matrix is ​​loaded into the message propagation topology structure inside the model, achieving accurate adaptation between feature data and model propagation logic. This provides a basic framework for subsequent spatiotemporal feature joint learning that fits the real-time correlation state of the system. The first-order neighbor set of each node is identified based on the message propagation topology structure. The neighbor node features are aggregated to form a neighbor feature vector, which is then concatenated with the current node's own feature vector to obtain aggregated features. This fully integrates the node's own information with local neighbor correlation information, strengthening the feature's ability to represent local spatiotemporal relationships. The aggregated features are then input into a nonlinear transformation layer for processing, and the node feature representation is updated through nonlinear mapping to obtain... Node embedding vectors overcome the limitations of linear feature processing, enhancing the adaptability of features to complex system operating states and allowing node features to more accurately reflect their comprehensive state with local relationships. Using node embedding vectors as input, message aggregation and nonlinear transformations are iteratively performed through a pre-defined number of graph neural network layers, progressively uncovering deep spatiotemporal dependencies between nodes and gradually strengthening the hierarchy and abstraction of features. This results in higher-order node features that contain more comprehensive information about system operation. Global pooling is then performed on these higher-order node features, aggregating all node features into a fixed-dimensional graph-level feature vector, i.e., a fused feature vector. This eliminates the impact of differences in the number of nodes on subsequent processing, resulting in a standardized comprehensive feature carrier that provides unified and comprehensive spatiotemporal fusion feature support for subsequent fault analysis.

[0061] In a preferred embodiment of the present invention, step 600 involves concatenating the fused feature vector with real-time conveying volume and material characteristic parameters to form an enhanced feature vector, which is then input into the prediction model to simultaneously calculate the probability distribution of fault types within a future set time period, and the pipeline segment corresponding to the fault type with the highest probability, determined based on the weight analysis within the prediction model, thereby completing the fault location, including: Step 601: Receive the fused feature vector and concatenate it with the real-time collected conveying capacity setpoint and material characteristic parameters to obtain an enhanced feature vector. Specifically, this includes: First, receiving the 512-dimensional fused feature vector output in step 505. This vector integrates the deep spatiotemporal correlation features of each monitoring point in the system, while simultaneously collecting the real-time conveying capacity setpoint and material characteristic parameters of the pneumatic conveying system. The real-time conveying capacity setpoint is read from the system's main control unit and corresponds to the rated material conveying value under the current operating conditions. The material characteristic parameters include: material particle size distribution, bulk density, and moisture content, all of which are collected in real-time by a preset detection module. The data is converted into numerical parameters to ensure consistency with the data format of the fused feature vector. Then, the fused feature vector is concatenated with the aforementioned real-time parameters according to their feature dimensions. The concatenation order is: fused feature vector first, conveying capacity setpoint in the middle, and material characteristic parameters last. After concatenation, the feature dimensions are expanded to 515, resulting in an enhanced feature vector. This supplements the core operating parameters of the system, enabling the feature vector to simultaneously encompass spatiotemporal fusion features and real-time operating condition information. Based on the influence of material characteristics and conveying capacity on pressure distribution and fault evolution, it provides more comprehensive feature support for subsequent accurate fault analysis based on predictive models.

[0062] Step 602: The enhanced feature vector is input into the multi-layer fully connected network of the prediction model for nonlinear feature transformation and dimensionality reduction to obtain high-dimensional abstract features. Specifically, the prediction model is pre-built and trained based on historical fault data and normal operation data of similar pneumatic conveying systems. During the construction phase, a multi-layer fully connected network with a dual-branch structure is adopted. The core fully connected network has three layers: the first layer has 1024 neurons to adapt to high-dimensional input features, the second layer has 512 neurons for feature depth mining, and the third layer has 256 neurons to achieve dimensionality compression and abstraction. Feature transfer between layers is achieved through a trainable weight matrix, and the initial weight values ​​are set using a random initialization method. The dual branches are a fault classification branch and a location regression branch. The fault classification branch has a built-in Softmax classifier, and the location regression branch is equipped with a weight analysis module. Both are connected to the output of the fully connected network, forming an overall architecture of feature extraction and dual-branch parallel inference, while reserving space for feature dimension adaptation. The interface is designed to ensure compatibility with 515-dimensional enhanced feature vector input. During the training phase, historical data from the past three years for this type of system are collected, covering the complete evolution process data of four types of faults: normal operation and partial blockage, pipeline leakage, fan malfunction, and solenoid valve jamming. Following the preprocessing procedures from steps 100 to 405 and the spatiotemporal feature joint learning procedure from step 500 of this invention, enhanced feature vectors and corresponding labels for training are generated. The labels include the fault type and the actual location of occurrence. The data is divided into training, validation, and test sets in a 7:2:1 ratio. A stochastic gradient descent optimizer is used with a learning rate of 0.001. The weights of the fully connected network and the bi-branch parameters are optimized using the cross-entropy loss function. The training is iterated for 80 rounds. After each round of training, the model's inference accuracy is verified using the validation set, and the number of neurons, activation function parameters, and branch weight allocation ratio are adjusted. The test set is used for final performance verification to ensure that the model can stably output the probability distribution of fault types and the node contribution value. After training is completed, the core parameters of the model are saved.

[0063] The enhanced feature vector is input into the multi-layer fully connected network of the pre-trained prediction model. The network loads the trained weight matrix and activation function parameters, first performing a nonlinear feature transformation on the enhanced feature vector, and then using the ReLU activation function to process the output features of each layer, breaking the linear correlation limitation of features and exploring the deep coupling relationship between working condition parameters and spatiotemporal features. At the same time, dimensionality reduction is achieved through the network's built-in dimensionality adjustment mechanism. Relying on the training-fixed weight decay coefficient, redundant features are eliminated, and core information strongly correlated with fault type and fault location is retained. The 515-dimensional feature vector is gradually mapped to 256 dimensions, finally obtaining high-dimensional abstract features, which not only meets the needs of subsequent dual-branch processing, but also enhances the targeted representation ability of features for fault features.

[0064] Step 603: The high-dimensional abstract features are simultaneously input into the fault classification branch and the location regression branch of the prediction model. The fault classification branch calculates the probability of occurrence of various faults within a set future time period using a Softmax classifier, forming a fault type probability distribution. The location regression branch calculates the contribution value of each node in the spatiotemporal feature set based on the high-dimensional abstract features. Specifically, this includes: simultaneously inputting the 256-dimensional high-dimensional abstract features into the fault classification branch and the location regression branch of the prediction model, relying on the parallel inference mechanism fixed by the model training to achieve synchronous processing of fault type judgment and location analysis, improving overall computational efficiency and result synergy. The Softmax classifier built into the fault classification branch has been trained and adapted to four common fault types of pneumatic conveying systems. The classifier loads the optimized probability calculation parameters and calculates the probability of occurrence within a given time period. The probabilistic processing of high-dimensional abstract features calculates the probability of occurrence of various faults within a future set time period. The set time period is preset to 5 minutes, combining the fault evolution rate and the fault development cycle in the training data. Then, the probabilities of various types are processed according to the normalization rules determined in the training, so that the sum of the probabilities of all fault types is 1, forming a standardized fault type probability distribution. The location regression branch, based on high-dimensional abstract features, loads the trained weight analysis parameters and, combined with the mapping relationship between nodes and pipeline sections in the spatiotemporal feature set mentioned above, calculates the contribution value of each node to the occurrence of the fault. The calculation process relies on the weight analysis mechanism built into the branch, combining the node feature weights with the operating condition sensitivity of the corresponding pipeline section. This sensitivity coefficient is determined by calibration with training data. The higher the contribution value, the closer the correlation between the node and the corresponding section and the fault, ensuring that the calculation results are consistent with the actual fault evolution law.

[0065] Step 604: Identify the fault type with the highest probability from the fault type probability distribution; sort all nodes according to the contribution value calculated by the location regression branch, and determine the pipeline segment mapped by the node with the highest ranking as the location of the fault type with the highest probability, thereby completing the fault location. Specifically, this includes: first, traversing and comparing the fault type probability distribution to identify the fault type with the highest probability, and recording the specific probability value of the fault type as the core basis for subsequent graded early warning. This traversal and comparison logic has been verified by model training to accurately lock the core fault risk; then, based on the contribution value of each node calculated by the location regression branch, sort the nodes corresponding to all monitoring points in descending order according to the sorting rules fixed by model training, and take the top three nodes with the highest ranking as candidate nodes. This setting has been tested and can effectively avoid the positioning error caused by the calculation deviation of a single node; combining the system equipment layout diagram and the mapping relationship of monitoring points, query the specific pipeline segment corresponding to the candidate node, and at the same time call the pressure gradient data calculated in step 402 to carry out multi-algorithm joint processing to optimize the judgment of pressure gradient abnormal areas, further improving the positioning accuracy.

[0066] First, the pressure gradient vector data is processed using the Kriging spatial interpolation algorithm. This algorithm is an interpolation method based on spatial correlation analysis. Its core principle is to infer the values ​​of unknown regions by fitting the spatial distribution patterns of known discrete points, generating a continuous and realistic distribution surface. The main pipeline is 80 meters long, and discrete pressure monitoring points P1 to P8 are set every 10 meters from the inlet to the outlet, corresponding to axial positions from 10 meters to 80 meters. Based on the gradient values ​​of these eight monitoring points, a spherical variogram model is constructed, with a range of 12 meters and a sill value of 0.0004 MPa. 2 The nugget has a strength of 0.00005 MPa. 2 By calculating the spatial distance between adjacent monitoring points and fitting the spatial correlation with the gradient value difference, the model is then used to perform gradient value interpolation calculations on the sections of the main pipeline without monitoring points at a step size of 0.5 meters. This generates a continuous pressure gradient distribution surface covering the entire axial and radial directions of the main pipeline, making up for the deficiency that discrete monitoring points cannot reflect the gradient changes across the entire area, and clearly presenting the gradient fluctuations in each subtle section.

[0067] Subsequently, the convex hull algorithm was used to determine the geometric boundary of the pressure gradient anomaly region. This algorithm is a geometric algorithm used to define the minimum enclosing region of a point set. It can form a convex polygon by connecting the outermost points of the point set to accurately define the core area. The preset normal pressure gradient threshold is ±0.02MPa. Monitoring points with gradient values ​​exceeding this threshold are selected as P3, P4, and P5, corresponding to axial positions of 20 meters, 30 meters, and 40 meters, respectively. These three points are used as the core point set. The spatial coordinates of the three points in the pipeline coordinate system are extracted and connected sequentially to form the minimum convex polygon. The vertices of the polygon correspond to the radial midpoints of P3, P4, and P5, respectively. The area enclosed by this convex polygon is defined as the initial boundary of the pressure gradient anomaly region, determining the core area of ​​the anomaly region covering the axial range of 20 meters to 40 meters and the radial range of zones 1 to 3.

[0068] Next, a region growing algorithm is used to refine and expand the initial abnormal region. This algorithm is an image segmentation algorithm that starts from a specified seed point and gradually merges adjacent regions that meet the requirements according to preset similarity conditions, which can achieve accurate expansion and optimization of the target region. The seed point is set at the position corresponding to the candidate node with the highest contribution, namely P4 at 30 meters axially and zone 2 radially. The gradient difference threshold is set to 0.01 MPa. The growth termination condition is that the gradient value of the adjacent region falls into the normal threshold range or reaches the pipe endpoint. Based on the continuous gradient distribution surface generated by Kriging interpolation, the regions with gradient values ​​exceeding ±0.02 MPa in the adjacent 0.5-meter segments are gradually merged from the seed point. At the same time, combined with the pipe topology, the growth range stops immediately when it touches the radial boundary zone 1 or 4 of the pipe or the endpoint of the axial 10 meters or 80 meters. Invalid regions that exceed the pipe entity range are removed. Finally, a complete and accurate fault-prone region is obtained, covering 18 meters to 42 meters axially and zones 1 to 3 radially.

[0069] Cross-validation is performed between the candidate node's corresponding segment and the pressure gradient anomaly region optimized by multiple algorithms. The candidate node's corresponding segment is axially 25 meters to 35 meters and radially 2 to 3 meters. If it falls completely within the above-mentioned suspected fault area, the segment is determined as the location of the highest probability fault type. If there is a deviation, the segment corresponding to the node with the highest contribution value is used. At the same time, the specific location and length of the deviation are marked by combining the convex hull boundary and the region growth range, thereby completing the fault location and achieving a precise correspondence between the fault type and the occurrence segment, solving the problem that existing methods are difficult to locate potential risk segments.

[0070] In this embodiment of the invention, a fused feature vector is received and concatenated with the real-time collected conveying volume setpoint and material characteristic parameters to supplement the core operating parameters of the system, enriching the operating condition representation dimension of the feature vector. This allows the enhanced feature vector to simultaneously cover spatiotemporal fusion features and real-time operating condition information, strengthening the feature's comprehensive mapping capability to the system's operating state. The enhanced feature vector is then input into a multi-layer fully connected network of the prediction model. Through nonlinear feature transformation, deep correlations between features are mined, and dimensionality reduction and redundant information removal are simultaneously completed to obtain high-dimensional abstract features. This retains the core effective features while adapting to subsequent dual-branch processing requirements, improving the feature's relevance and adaptability. The high-dimensional... Abstract features are synchronously input into the fault classification branch and the location regression branch to achieve parallel processing of fault type judgment and location analysis, thereby improving overall computational efficiency. The fault classification branch quantifies the probability of occurrence of various faults through a Softmax classifier, forming a standardized fault type probability distribution. The location regression branch calculates the contribution value of each node, providing a quantitative basis for fault location. The highest probability fault type is identified from the fault type probability distribution to determine the core fault risk. Key nodes are locked based on the ranking of node contribution values ​​and mapped to the corresponding pipeline sections, achieving a precise correspondence between fault type and occurrence location, providing clear and specific core information support for subsequent graded early warning.

[0071] In a preferred embodiment of the present invention, step 700, which involves obtaining and issuing graded early warning information based on the fault type probability distribution and fault location, includes: Step 701: Receive the fault type with the highest probability and its corresponding fault location. Based on the fault type probability distribution, determine the current warning level according to preset risk assessment rules. Specifically, this includes: First, receiving the fault type with the highest probability and its corresponding fault location, while simultaneously acquiring the complete fault type probability distribution. Then, determine the current warning level based on preset risk assessment rules. The preset risk assessment rules must be formulated before the system is officially launched. The formulation process is based on historical fault handling records throughout the pneumatic conveying system's lifecycle, equipment operation and maintenance technical manuals, and production downtime loss calculation data. It revolves around three core dimensions: fault impact range, handling difficulty, and downtime loss. First, systematically analyze the actual production impact of common system fault types such as local blockage, pipeline leakage, fan malfunction, and solenoid valve jamming to determine the impact boundaries of different faults on continuous production operation and equipment safety. Then, based on the impact level, divide the system into a three-level risk hierarchy framework: Level 1 (low risk), Level 2 (medium risk), and Level 3 (high risk). Simultaneously, combine on-site operation and maintenance experience with historical fault data statistics to set the warning level. The rules for adjusting the risk level based on the probability range of the fault and the criticality of the fault location are solidified after multiple rounds of review and verification by operation and maintenance experts and backtracking of historical fault cases. These rules are then entered into the system's risk assessment module for real-time early warning level determination. Pre-defined risk assessment rules are based on the impact range, handling difficulty, and downtime losses of pneumatic conveying system faults, dividing the early warning level into three levels: Level 1 (low risk), Level 2 (medium risk), and Level 3 (high risk). During the assessment, the basic risk level is first determined based on the highest probability fault type. For example, partial blockage and pipeline leakage correspond to basic Level 2 risk, while fan malfunction and solenoid valve jamming correspond to basic Level 1 risk. The level is then adjusted based on the highest fault probability value: a probability higher than 80% raises the basic level by one level; a probability between 50% and 80% maintains the basic level; and a probability lower than 50% lowers the basic level by one level. Simultaneously, the criticality of the fault location is considered. If the location is a core section such as a pipeline bend or a silo pump outlet, the risk level is increased by half a level after the above adjustments. Finally, the current early warning level is determined comprehensively, achieving quantitative classification and precise definition of risk, providing a basis for subsequent differentiated early warning issuance.

[0072] Step 702: Based on the warning level, the most probable fault type, and the fault location, match the corresponding text description template from the pre-set warning template library to obtain structured warning content containing fault type, possible location, risk level, and suggested handling measures. Specifically, this includes: matching the corresponding text description template from the pre-set warning template library based on the determined warning level, the most probable fault type, and the corresponding fault location; the pre-set warning template library needs to be built before the system goes online, and its construction is based on common fault types of pneumatic conveying systems, hierarchical warning management specifications, and on-site operation and maintenance handling standard procedures. First, a classification framework is built according to the two dimensions of warning level and fault type. Level 1 warning corresponds to low-risk fault types, Level 2 warning corresponds to medium-risk fault types, and Level 3 warning corresponds to high-risk fault types. Each dimension intersection node is configured with an independent template; then, a general template structure is designed, divided into two parts: a fixed text frame and variable fill items. The fixed text frame includes the warning information title, information validity description, and responsible entity prompt, etc. The variable fill items are preset as follows: The system identifies fault types, possible locations, risk levels, and recommended handling measures. It also considers the handling priorities and practical points for different faults, matching specific handling measures for each type. After multiple rounds of review by maintenance experts and verification with historical fault handling cases, all templates are entered into the system database to form a complete early warning template library. The early warning template library is pre-built according to early warning levels and fault types. Each template contains a fixed text frame and variable fill items. The variable fill items correspond to core information such as fault type, possible location, risk level, and recommended handling measures. During matching, a subset of templates is first filtered according to the early warning level. Then, within the subset, a unique template is precisely matched according to the fault type. Subsequently, the specific fault type, the located pipeline section, and the determined early warning level are filled into the template. Simultaneously, appropriate recommended handling measures are added based on the fault type and location. For example, partial blockage corresponds to recommendations to reduce the flow rate and activate the backflushing device; pipeline leakage corresponds to recommendations to immediately shut down the system and investigate the leak point. Finally, a complete and standardized structured early warning content is generated, ensuring that the early warning information is both comprehensive and practical.

[0073] Step 703 involves converting the structured warning content into a data format conforming to the target publishing interface protocol to obtain the graded warning information to be published. Specifically, this includes converting the structured warning content into a data format conforming to the target publishing interface protocol. The target publishing interface protocol is preset to the industrial Ethernet communication protocol according to the system communication architecture, and the data format adopts the JSON format specification. During the conversion process, the structured warning content is first split into fields, mapping the fault type, possible location, risk level, and suggested handling measures to the standard fields specified by the protocol. Then, the field content is formatted to ensure that the text length and character encoding meet the interface requirements. At the same time, a timestamp field is added to record the warning generation time, and a check code field is added for the receiving end to verify the data integrity to avoid loss or distortion during information transmission. Finally, a graded warning information to be published that is compliant in format and can be directly transmitted is obtained.

[0074] Step 704: Publish the tiered early warning information to be published to the designated target receiving end through a preset publishing interface. Specifically, this includes: publishing the tiered early warning information to be published to the designated target receiving end through a preset publishing interface. The preset publishing interface needs to be built during the system deployment phase, based on the communication architecture specifications of the pneumatic conveying system, the real-time requirements for early warning information transmission, and the multi-terminal compatibility requirements. First, the industrial-grade TCP / IP communication protocol is selected as the underlying communication standard for the interface, adapting to the communication protocols of various receiving ends such as maintenance terminals, central control system consoles, and enterprise production management platforms. Then, a two-way communication function module is constructed, integrating core functions such as real-time data push, reception feedback, and communication status monitoring. Simultaneously, a transmission timeout judgment threshold, retry mechanism parameters, and fault log recording rules are preset. After multiple rounds of communication integration testing to verify the interface's stability and compatibility... The interface is permanently deployed to the system communication layer for use in the early warning information release process. The preset release interface is a pre-deployed bidirectional communication interface that supports real-time data push and reception feedback. The designated target receiving end is set differently according to the early warning level. Level 1 early warnings are pushed to the mobile terminals of maintenance personnel, Level 2 early warnings are pushed to both the maintenance terminal and the central control system console, and Level 3 early warnings are pushed to the maintenance terminal, the central control console, and the enterprise production management platform. During the release process, the interface communication status is monitored synchronously. If a transmission timeout or failure occurs, a retry mechanism is automatically triggered. The retry interval is set to 1 second, 3 seconds, and 5 seconds respectively. If three consecutive retry failures occur, a fault log is recorded and a local audible and visual alarm is triggered. After the information is successfully pushed, a confirmation receipt is received from the target end, completing the early warning release loop and ensuring that relevant personnel can obtain risk information in a timely manner and quickly initiate the corresponding handling process.

[0075] In this embodiment of the invention, the fault type with the highest reception probability and its corresponding location are combined with the complete fault type probability distribution and a pre-set risk assessment rule to determine the warning level. This transforms quantified fault information into a clear risk hierarchy, achieving standardized risk definition and providing a basis for subsequent differentiated warning issuance and handling. Based on the warning level, fault type, and location, a pre-set template library of text description templates is matched to quickly generate structured warning content containing core fault information and suggested handling measures. This ensures that the warning information is standardized and complete, balancing the integrity and practicality of information transmission, and providing clear guidance for efficient handling by the receiving end. The structured warning content is converted into a data format that conforms to the target release interface protocol, completing the standardized adaptation of the warning information format, eliminating release obstacles caused by format incompatibility, ensuring that the warning information can be smoothly connected to the release channel, and improving the stability and adaptability of data transmission. The graded warning information is pushed to the designated target receiving end through a pre-set release interface, achieving precise targeted transmission of warning information, ensuring that relevant personnel can obtain the corresponding risk information in a timely manner, quickly initiate the handling process, and form a complete business link from fault analysis to warning implementation.

[0076] like Figure 2 As shown, embodiments of the present invention also provide a multi-source data fault prediction system for a pneumatic conveying system, comprising: The acquisition module is used to collect multi-source heterogeneous monitoring data, including time-series pressure, flow rate, temperature and operating status data, during the operation of the pneumatic conveying system, and to perform initial time alignment to obtain the corresponding time alignment mapping relationship. The preprocessing module is used to preprocess multi-source heterogeneous monitoring data to obtain a spatiotemporally aligned feature matrix; The construction module is used to construct a spatiotemporal feature set based on the spatiotemporal alignment feature matrix; wherein, each element in the spatiotemporal feature set corresponds to a key monitoring point of the system, and its content is aggregated from the time-series monitoring data of the corresponding location; the relationship between elements is defined based on the physical connection of equipment and the dependence of process flow. The dynamic correction module is used to extract the distribution feature vector based on real-time data from multiple discrete pressure monitoring points on the main pipeline, and use it to dynamically correct the spatiotemporal feature set to obtain a dynamic weighted adjacency matrix. The fusion module is used to input the dynamically weighted adjacency matrix into the pre-trained graph neural network model for joint learning of spatiotemporal features to obtain a fused feature vector; The prediction and location module is used to combine the fused feature vector with real-time conveying volume and material characteristic parameters to form an enhanced feature vector, which is then input into the prediction model to simultaneously calculate the probability distribution of fault types within a future set time period, as well as the pipeline section corresponding to the fault type with the highest probability, determined by the internal weight analysis of the prediction model, and complete the fault location. The early warning module is used to obtain and issue graded early warning information based on the probability distribution of fault types and the location of faults.

[0077] It should be noted that this system is a system corresponding to the above method. All implementation methods in the above method embodiments are applicable to this embodiment and can achieve the same technical effect.

[0078] Embodiments of the present invention also provide a computing device, including: a processor and a memory storing a computer program, wherein the computer program, when executed by the processor, performs the method described above. All implementations in the above method embodiments are applicable to this embodiment and can achieve the same technical effects.

[0079] The above description represents the preferred embodiments of the present invention. It should be noted that those skilled in the art can make various improvements and modifications without departing from the principles of the present invention, and these improvements and modifications should also be considered within the scope of protection of the present invention.

Claims

1. A multi-source data fault prediction method for a pneumatic conveying system, characterized in that, The method includes: Step 100: Collect multi-source heterogeneous monitoring data including time-series pressure, flow rate, temperature and operating status data of the pneumatic conveying system during operation, and perform initial time alignment to obtain the corresponding time alignment mapping relationship; Step 200: Preprocess the multi-source heterogeneous monitoring data to obtain a spatiotemporal aligned feature matrix; Step 300: Construct a spatiotemporal feature set based on the spatiotemporal alignment feature matrix; wherein each element in the spatiotemporal feature set corresponds to a key monitoring point of the system, and its content is aggregated from the time-series monitoring data of the corresponding location; the relationship between elements is defined based on the physical connection of equipment and the dependence of process flow. Step 400: Based on real-time data from multiple discrete pressure monitoring points on the main pipeline, extract the distribution feature vector and use it to dynamically correct the spatiotemporal feature set to obtain a dynamic weighted adjacency matrix. Step 500: Input the dynamically weighted adjacency matrix into the pre-trained graph neural network model to perform spatiotemporal feature joint learning and obtain the fused feature vector; Step 600: The fused feature vector is spliced ​​with the real-time conveying volume and material characteristic parameters to form an enhanced feature vector, which is then input into the prediction model to simultaneously calculate the probability distribution of fault types within a future set time period, as well as the pipeline section corresponding to the fault type with the highest probability, determined by the internal weight analysis of the prediction model, thereby completing the fault location. Step 700: Based on the probability distribution of fault types and the location of faults, obtain and issue graded early warning information.

2. The multi-source data fault prediction method for a pneumatic conveying system according to claim 1, characterized in that, Step 100 includes: Collect time-series pressure data from pressure vessels in a pneumatic conveying system; Collect time-series flow data related to material conveying; Collect time-series temperature data of key rotating equipment; Collect time-series operating status data reflecting equipment operation and energy consumption; The time-series pressure data, time-series flow data, time-series temperature data, and time-series operating status data are aggregated and aligned with the initial time to obtain multi-source heterogeneous monitoring data and the corresponding time alignment mapping relationship.

3. The multi-source data fault prediction method for a pneumatic conveying system according to claim 2, characterized in that, Step 200 includes: The quality of the multi-source heterogeneous monitoring data is assessed, and missing data segments and noise interference segments are identified and labeled to obtain a data quality labeling matrix. Based on the data quality labeling matrix, for the segments in the time-series pressure data and time-series flow data that are labeled as noise interference, the window parameters and smoothing intensity are dynamically adjusted according to the statistical characteristics of the data in each sliding window. This achieves the suppression of high-frequency random noise while selectively retaining and enhancing the low-frequency trend components that characterize equipment condition deterioration and fault evolution, thus obtaining the noise-reduced time-series pressure data and time-series flow data. Based on the data quality labeling matrix, for the time periods marked as missing data, corresponding completion strategies are adopted according to different data source types to obtain the completed time-series pressure data, time-series flow data, and time-series temperature data. Based on the switch signals of the time-series running status data, template matching and status encoding processing based on event windows are performed to convert them into a numerical sequence that is consistent with the continuous time-series data in terms of time and numerical dimensions, thus obtaining the encoded running status sequence.

4. The multi-source data fault prediction method for a pneumatic conveying system according to claim 3, characterized in that, Step 200 further includes: The denoised time-series pressure data and time-series flow data, as well as the completed time-series pressure data and time-series flow data, are integrated to form the final pressure data and the final flow data. By integrating the final pressure data, the final flow data, the completed time-series temperature data, and the encoded operating status sequence, an integrated feature set is obtained. The integrated feature set and time alignment mapping are combined, and all data are normalized to the same time base and sampling frequency through a unified resampling operation to obtain the spatiotemporal aligned feature matrix.

5. The multi-source data fault prediction method for a pneumatic conveying system according to claim 4, characterized in that, Step 300 includes: The spatiotemporal alignment feature matrix is ​​analyzed, and based on the preset monitoring point identifiers therein, the multivariate time series data slices corresponding to each monitoring point in the spatiotemporal alignment feature matrix are separated to obtain independent time series data fragments for each monitoring point; For each independent time-series data segment of a monitoring point, a sliding time window is used for traversal; within each time window, multiple predefined statistical features of the data segment in that window are calculated and combined in sequence to form a feature vector; The feature vectors are stacked in chronological order to form a node feature matrix for the monitoring point; the rows of the node feature matrix correspond to the time window, and the columns correspond to the extracted statistical feature dimensions. Based on the process flow diagram and equipment layout diagram of the pneumatic conveying system, a topology diagram is constructed to characterize the physical connections and process dependencies between system components. The topology diagram is then transformed into an adjacency matrix, where the rows and columns of the matrix correspond to system monitoring points. If there is a direct physical connection or a tightly coupled process dependency between the components represented by two monitoring points, the corresponding position in the adjacency matrix is ​​marked as a connection state. The node feature matrix is ​​used as the feature content of the nodes in the spatiotemporal feature set, and the adjacency matrix is ​​used as the definition of the association relationship between the nodes in the spatiotemporal feature set, together forming the spatiotemporal feature set.

6. The multi-source data fault prediction method for a pneumatic conveying system according to claim 5, characterized in that, Step 400 includes: From the spatiotemporal feature set, nodes corresponding to multiple preset discrete pressure monitoring points on the main pipeline are identified, and real-time pressure time series data segments of the corresponding nodes in the current time window are extracted. Based on the real-time pressure time series data segment, calculate the real-time pressure gradient value between every two adjacent pressure monitoring points, and construct a pressure gradient vector based on the real-time pressure gradient values ​​of all adjacent point pairs. The pressure gradient vector is normalized to obtain a normalized gradient feature vector. The normalized gradient feature vector is then input into a preset exponential mapping function. This exponential mapping function maps the gradient elements in the vector to corresponding dynamic adjustment coefficients through exponential operations based on the magnitude and sign of each element. The calculation rule of the exponential mapping function is as follows: the adjustment coefficient value of the connecting edge corresponding to the adjacent monitoring point with a larger absolute gradient value is increased accordingly. Using the aforementioned dynamic adjustment coefficient, the connection weights of corresponding edges in the adjacency matrix are scaled element-wise to obtain a modified adjacency matrix; Keeping the node feature matrix unchanged, the modified adjacency matrix is ​​used as the optimized inter-node association relationship and combined with the node feature matrix to obtain a dynamic weighted adjacency matrix.

7. The multi-source data fault prediction method for a pneumatic conveying system according to claim 6, characterized in that, Step 500 includes: Based on the dynamic weighted adjacency matrix, the node feature matrix is ​​initialized with the node features of each monitoring point in the graph neural network, and the modified adjacency matrix is ​​loaded as the message propagation topology structure inside the model; Using the node features and message propagation topology as input, in the first layer of the graph neural network, for each node corresponding to a monitoring point, the set of its directly connected first-order neighbor nodes is identified according to the message propagation topology; the features of all neighbor nodes in the first-order neighbor node set are aggregated to obtain a neighbor feature vector; the neighbor feature vector is concatenated with the feature vector of the current node itself to obtain the aggregated feature as the output of the node in the first layer. The aggregated features are input into a nonlinear transformation layer for processing, and the feature representation of each node is updated to obtain the node embedding vector of the graph neural network layer. The node embedding vector is taken as input and iteratively processed through a graph neural network of a predetermined number of layers; each layer performs message aggregation and nonlinear transformation operations, and finally outputs high-order node features containing deep spatiotemporal dependencies of the system. A global pooling operation is performed on the higher-order node features to aggregate the features of all nodes into a fixed-dimensional graph-level feature vector, resulting in a fused feature vector.

8. The multi-source data fault prediction method for a pneumatic conveying system according to claim 7, characterized in that, Step 600 includes: The fused feature vector is received and concatenated with the real-time collected conveying capacity setpoint and material characteristic parameters by feature dimension to obtain an enhanced feature vector; The enhanced feature vector is input into the multilayer fully connected network of the prediction model for nonlinear feature transformation and dimensionality reduction to obtain high-dimensional abstract features; The high-dimensional abstract features are simultaneously input into the fault classification branch and the location regression branch of the prediction model; the fault classification branch calculates the probability of occurrence of various faults within a set future time period using a Softmax classifier, forming a fault type probability distribution; the location regression branch calculates the contribution value of each node in the spatiotemporal feature set based on the high-dimensional abstract features. The fault type with the highest probability is identified from the fault type probability distribution; all nodes are sorted according to the contribution value calculated by the location regression branch, and the pipeline segment mapped by the node with the highest ranking is determined as the location of the fault type with the highest probability, thereby completing the fault location.

9. The multi-source data fault prediction method for a pneumatic conveying system according to claim 8, characterized in that, Step 700 includes: The system receives the fault type with the highest probability and its corresponding location, and, based on the probability distribution of the fault type, determines the current warning level according to preset risk assessment rules. Based on the warning level, the most probable fault type, and the location of the fault, the corresponding text description template is matched from the preset warning template library to obtain structured warning content that includes fault type, possible location, risk level, and suggested handling measures. The structured warning content is converted into a data format that conforms to the target publishing interface protocol to obtain the hierarchical warning information to be published; The tiered early warning information to be published is published to the designated target receiving end through a preset publishing interface.

10. A multi-source data fault prediction system for a pneumatic conveying system, the system implementing the method as described in any one of claims 1 to 9, characterized in that, include: The acquisition module is used to collect multi-source heterogeneous monitoring data, including time-series pressure, flow rate, temperature and operating status data, during the operation of the pneumatic conveying system, and to perform initial time alignment to obtain the corresponding time alignment mapping relationship. The preprocessing module is used to preprocess multi-source heterogeneous monitoring data to obtain a spatiotemporally aligned feature matrix; The construction module is used to construct a spatiotemporal feature set based on the spatiotemporal alignment feature matrix; wherein, each element in the spatiotemporal feature set corresponds to a key monitoring point of the system, and its content is aggregated from the time-series monitoring data of the corresponding location; the relationship between elements is defined based on the physical connection of equipment and the dependence of process flow. The dynamic correction module is used to extract the distribution feature vector based on real-time data from multiple discrete pressure monitoring points on the main pipeline, and use it to dynamically correct the spatiotemporal feature set to obtain a dynamic weighted adjacency matrix. The fusion module is used to input the dynamically weighted adjacency matrix into the pre-trained graph neural network model for joint learning of spatiotemporal features to obtain a fused feature vector; The prediction and location module is used to combine the fused feature vector with real-time conveying volume and material characteristic parameters to form an enhanced feature vector, which is then input into the prediction model to simultaneously calculate the probability distribution of fault types within a future set time period, as well as the pipeline section corresponding to the fault type with the highest probability, determined by the internal weight analysis of the prediction model, and complete the fault location. The early warning module is used to obtain and issue graded early warning information based on the probability distribution of fault types and the location of faults.

Images