Intelligent decision-making method and system for multi-source data fusion

Abstract

The invention relates to the technical field of data processing, and particularly provides an intelligent decision-making method and system for multi-source data fusion, and the method comprises the steps: carrying out the parallel collection of environment data and equipment state data under the same timestamp, obtaining multi-source data, transmitting the multi-source data to a PLC at an equipment end, and carrying out the local preprocessing, performing time sequence feature extraction and anomaly detection on the preprocessed multi-source data to obtain labeled multi-source data and an abnormal event list, and inputting the labeled multi-source data and abnormal event list into a preset weight adjustment model to obtain a dynamically adjusted weight vector; and carrying out weighted fusion on the labeled multi-source data by using the dynamically adjusted weight vector to obtain fused data, and carrying out intelligent decision making based on the fused data. By carrying out dynamic weighted fusion on the environment data and the equipment state data, the overall intelligent regulation and control capability of the system is improved, and the problems that the reliability of multi-source data fusion is insufficient and the weight cannot be adaptively adjusted are solved.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of data processing, and in particular, to a multi-source data fusion intelligent decision method and system. BACKGROUND

[0002] With the continuous improvement of industrial automation, intelligent manufacturing and device networking, a large amount of environmental parameters and device state parameters will be generated during the operation of production equipment. For example, external environmental data such as temperature, humidity, and pressure, and internal device data such as running state identification, speed parameters, and fault codes. With the increase in the number of monitoring objects and the increase in the collection frequency, a multi-source, high-frequency, and heterogeneous data system has gradually formed in the industrial field. In order to ensure production safety, improve device stability, and optimize energy consumption control, enterprises are increasingly relying on comprehensive analysis of these multi-source data to assist real-time control and intelligent decision-making.

[0003] In the prior art, multi-source data fusion intelligent decision-making usually relies on fixed weight linear fusion methods or preset rules for processing. The common practice is to perform independent collection processes on environmental data and device state data respectively, to simply normalize the data without correlation between them, to determine the weight coefficients of each data source according to experience, to statically weight and fuse data from different sources, and to trigger a rule engine or decision model based on the fused data. In addition, existing solutions generally use a single threshold judgment or a simple abnormality screening method based on historical statistics to identify key state changes, thereby achieving basic monitoring and early warning effects.

[0004] Although the above fixed weight fusion and rule-driven decision-making method can achieve joint judgment of multi-source data to a certain extent and complete simple identification and decision triggering of device abnormalities, when a certain data source drifts, the number of abnormalities increases, or the correlation with other data sources weakens, the static fusion method is difficult to adjust the weight in a timely manner, resulting in an increase in the deviation of the fusion result, thereby causing the decision model to make false or missed judgments. SUMMARY

[0005] Embodiments of the present application provide a multi-source data fusion intelligent decision method and system, which can improve the above problems of insufficient reliability of multi-source data fusion and inability to adaptively adjust the weight.

[0006] Other characteristics and advantages of the present application will become apparent from the following detailed description, or will be learned by practice of the present application.

[0007] According to an aspect of an embodiment of the present application, a multi-source data fusion intelligent decision method is provided, including: collecting environment data and device state data in parallel at the same time stamp to obtain multi-source data, transmitting the multi-source data to a PLC at a device end for local preprocessing to obtain preprocessed multi-source data; performing time sequence feature extraction and anomaly detection on the preprocessed multi-source data to obtain labeled multi-source data and an anomaly event list; performing correlation analysis on the environment data and the device state data in the labeled multi-source data to obtain a correlation coefficient matrix between data sources; inputting the anomaly event list and the correlation coefficient matrix into a preset weight adjustment model to obtain a dynamically adjusted weight vector; performing weighted fusion on the labeled multi-source data by using the dynamically adjusted weight vector to obtain fused data, and performing intelligent decision based on the fused data.

[0008] In some embodiments of the present application, based on the foregoing scheme, the step of collecting environment data and device state data in parallel at the same time stamp to obtain multi-source data, and transmitting the multi-source data to a PLC at a device end for local preprocessing to obtain preprocessed multi-source data includes: triggering a data collection instruction at a preset collection period arrival time, controlling an environment data collection module to collect real-time values of a temperature sensor, a humidity sensor and a pressure sensor as environment data according to the collection instruction, and controlling a device state data collection module to collect device running state identifiers, motor speed parameters and fault code information as device state data; attaching a time stamp identifier of a current time to the environment data and the device state data, transmitting the environment data and the device state data with the attached time stamp identifier to a data cache area through a PLC internal bus, arranging and combining the environment data and the device state data in the data cache area in time stamp order to obtain multi-source data; transmitting the multi-source data to an edge computing unit of the PLC, performing data cleaning processing on the multi-source data in the edge computing unit to obtain cleaned multi-source data; performing normalization processing on the cleaned multi-source data to obtain normalized multi-source data, and performing compression storage on the normalized multi-source data with a change amplitude less than a preset change threshold at consecutive time stamps to obtain preprocessed multi-source data.

[0009] In some embodiments of this application, based on the foregoing scheme, the step of extracting time-series features and detecting anomalies from the preprocessed multi-source data to obtain labeled multi-source data and a list of abnormal events includes: arranging the preprocessed multi-source data according to the chronological order of timestamps to construct an environmental data time series and a device status data time series; setting a sliding window with a length of N sampling points, sliding the sliding window across the environmental data time series and the device status data time series with a preset step size, and extracting statistical feature parameters of the data within each window position; and combining the statistical feature parameters of the environmental data time series and the statistical feature parameters of the device status data time series. The parameters are concatenated to obtain a multidimensional feature vector sequence. This multidimensional feature vector sequence is then input into a pre-trained isolated forest anomaly detection model. The average path length of each feature vector in the multidimensional feature vector within the isolated tree is calculated, and an anomaly score is calculated based on the average path length. When the anomaly score exceeds a preset anomaly threshold, the data at the corresponding time point is determined to be anomalous data. An anomaly label is then added to the preprocessed multi-source data for the corresponding time point, resulting in labeled multi-source data. The anomaly label includes an anomaly type code and an anomaly confidence value. The timestamps, anomaly type codes, and anomaly confidence values ​​of all anomalous data are extracted and organized chronologically to construct an anomalous event list.

[0010] In some embodiments of this application, based on the aforementioned scheme, the step of performing correlation analysis on the environmental data and equipment status data in the labeled multi-source data to obtain the correlation coefficient matrix between data sources includes: separating and extracting environmental data and equipment status data from the labeled multi-source data; constructing temperature data, humidity data, and pressure data in the environmental data into independent data vectors; constructing operating status identifiers, motor speed parameters, and fault code information in the equipment status data into independent data vectors; summarizing all independent data vectors to obtain M data source vectors, where M is the total number of data sources; calculating the Pearson correlation coefficient between any two data source vectors in the M data source vectors; when the data types of the two data source vectors are inconsistent, using the Spearman rank correlation coefficient instead of the Pearson correlation coefficient for calculation; traversing all pairwise combinations of data source vectors, calculating the correlation coefficient between each pair of data sources; filling the correlation coefficients into an M-row, M-column matrix according to the order of the data sources; and setting the absolute value of the correlation coefficients in the matrix less than a preset correlation threshold to zero to obtain the correlation coefficient matrix between data sources.

[0011] In some embodiments of this application, based on the aforementioned scheme, the weight adjustment model includes a reliability assessment module and a weight optimization module. The reliability assessment module is used to calculate the reliability coefficient of each data source, and the weight optimization module is used to reduce and adjust the initial weight of data sources with high anomaly rates. The step of inputting the list of abnormal events and the correlation coefficient matrix into the preset weight adjustment model to obtain the dynamically adjusted weight vector includes: summing the absolute values ​​of each row of the correlation coefficient matrix to obtain the sum of the correlation strengths of each data source with other data sources; normalizing the sum of the correlation strengths to obtain the initial weight vector; counting the frequency of anomalies occurring in each data source within a preset time window from the list of abnormal events; calculating the proportion of the frequency of anomalies to the total number of samples to obtain the anomaly rate of each data source; inputting the anomaly rate of each data source into the reliability assessment module to calculate the reliability coefficient of each data source; and inputting the initial weight vector and the reliability coefficient into the weight optimization module to obtain the dynamically adjusted weight vector.

[0012] In some embodiments of this application, based on the aforementioned scheme, the step of using the dynamically adjusted weight vector to perform weighted fusion of the labeled multi-source data to obtain fused data, and performing intelligent decision-making based on the fused data, includes: extracting the values ​​of each data source corresponding to each timestamp from the labeled multi-source data; multiplying the weight coefficients in the dynamically adjusted weight vector with the values ​​of the corresponding data sources to obtain the weighted values ​​of each data source; summing the weighted values ​​of all data sources at the same timestamp to obtain fused time series data; performing smoothing filtering on the fused time series data to obtain smoothed fused data; inputting the smoothed fused data into a pre-trained intelligent decision-making model, outputting decision results, and adjusting the operating parameters of the device or triggering protection actions according to the decision results.

[0013] In some embodiments of this application, based on the foregoing scheme, the step of adjusting the operating parameters of the equipment or triggering protection actions according to the decision result includes: parsing the decision result to obtain a fault risk level and a recommended operation instruction; when the fault risk level exceeds a preset safety threshold, generating an alarm signal using the intelligent decision model and outputting a corresponding emergency handling instruction; performing correlation analysis between the decision result and the anomaly identification tags in the labeled multi-source data, and executing the recommended operation instruction when the decision result and the anomaly identification tags point to a consistent risk trend; and sending the emergency handling instruction or the recommended operation instruction to the actuator through the output module of the PLC, so that the actuator adjusts the operating parameters of the equipment or triggers protection actions according to the emergency handling instruction or the recommended operation instruction.

[0014] According to another aspect of the embodiments of this application, a multi-source data fusion intelligent decision-making system is provided, comprising: a data acquisition module, used to collect environmental data and equipment status data in parallel at the same timestamp to obtain multi-source data, and transmit the multi-source data to a PLC at the device end for local preprocessing to obtain preprocessed multi-source data; an anomaly detection module, used to extract time-series features and detect anomalies in the preprocessed multi-source data to obtain labeled multi-source data and an anomaly event list; a correlation analysis module, used to perform correlation analysis on the environmental data and equipment status data in the labeled multi-source data to obtain a correlation coefficient matrix between data sources; a data input module, used to input the anomaly event list and the correlation coefficient matrix into a preset weight adjustment model to obtain a dynamically adjusted weight vector; and a weighted fusion module, used to perform weighted fusion of the labeled multi-source data using the dynamically adjusted weight vector to obtain fused data, and to make intelligent decisions based on the fused data.

[0015] Compared with existing technologies, this application has the following beneficial effects: by uniformly collecting, preprocessing, analyzing time-series features, detecting anomalies, performing correlation analysis, and dynamically weighting and fusing environmental data and equipment status data, it can ensure the consistency, reliability, and effectiveness of multi-source data in different stages, so as to make the generated intelligent decision-making results more accurate, adaptable to changes in the equipment operating environment, and improve the overall intelligent control capability of the system, thereby improving the problems of insufficient reliability of multi-source data fusion and the inability to adaptively adjust weights. Attached Figure Description

[0016] Figure 1 This is a flowchart illustrating the intelligent decision-making method for multi-source data fusion provided in an embodiment of the present invention. Figure 2 This is a schematic block diagram of the structure of the intelligent decision-making system for multi-source data fusion provided in the embodiments of the present invention. Detailed Implementation

[0017] Exemplary embodiments will now be described in a more comprehensive manner with reference to the accompanying drawings. However, the exemplary embodiments can be implemented in various forms and should not be construed as limited to these examples; rather, these embodiments are provided so that this application will be more comprehensive and complete, and will fully convey the concept of the exemplary embodiments to those skilled in the art.

[0018] Furthermore, the features, structures, or characteristics described in this application can be combined in any suitable manner in one or more embodiments. Numerous specific details are provided in the following description to provide a full understanding of the embodiments of this application. However, those skilled in the art will recognize that when implementing the technical solutions of this application, not all detailed features in the embodiments may be used, one or more specific details may be omitted, or other methods, elements, devices, steps, etc., may be employed.

[0019] The block diagrams shown in the accompanying drawings are merely functional entities and do not necessarily correspond to physically independent entities. That is, these functional entities can be implemented in software, in one or more hardware modules or integrated circuits, or in different network and / or processor devices and / or microcontroller devices.

[0020] The flowcharts shown in the accompanying drawings are merely illustrative and do not necessarily include all content and operations / steps, nor do they necessarily have to be performed in the described order. For example, some operations / steps can be broken down, while others can be combined or partially combined; therefore, the actual execution order may change depending on the specific circumstances.

[0021] It should be noted that "multiple" in this article refers to two or more. "And / or" describes the relationship between related objects, indicating that three relationships can exist. For example, A and / or B can represent: A alone, A and B simultaneously, or B alone. The character " / " generally indicates that the preceding and following related objects have an "or" relationship.

[0022] The technical solution of the present invention will be further described below with reference to the accompanying drawings and specific embodiments.

[0023] This application provides an example of an intelligent decision-making method that integrates multi-source data. This method can uniformly collect, preprocess, perform correlation analysis, and weighted fusion of environmental data and equipment status data, thereby improving the accuracy and real-time performance of intelligent decision-making. The intelligent decision-making method specifically includes the following steps: Step S100: Environmental data and equipment status data are collected in parallel at the same timestamp to obtain multi-source data. The multi-source data is then transmitted to the PLC at the device end for local preprocessing to obtain preprocessed multi-source data.

[0024] By setting a unified collection cycle and triggering mechanism for environmental data and equipment status data, both can be collected in parallel at the same timestamp, ensuring the consistency of multi-source data in the time dimension and avoiding alignment deviations or data mismatches in subsequent analysis stages.

[0025] A unified data acquisition instruction scheduling logic can be configured in the PLC at the device end to control the environmental data acquisition module to collect real-time information such as temperature, humidity, and pressure, and to control the equipment status data acquisition module to collect equipment operating status, motor speed, and fault information. The collected environmental data and equipment status data are then transmitted to the PLC's buffer after being appended with the timestamp of the current acquisition time, so as to achieve synchronous data aggregation.

[0026] Step S200: Perform time-series feature extraction and anomaly detection on the preprocessed multi-source data to obtain labeled multi-source data and a list of abnormal events.

[0027] The preprocessed multi-source data is organized in a time series format, and its time dimension variation patterns are extracted. An anomaly detection algorithm is used to identify abnormal fluctuations or sudden changes in the multi-source data in different time periods, so as to form labeled multi-source data that can reflect data characteristics and abnormal behavior.

[0028] Based on the sliding observation window, statistical features such as mean, variance, and rate of change of multi-source data are extracted. These statistical features are then input into the anomaly detection model to identify abnormal behavior of multi-source data in different time periods, and finally obtain a list of abnormal events containing anomaly type and anomaly confidence.

[0029] Step S300: Perform correlation analysis on the environmental data and equipment status data in the labeled multi-source data to obtain the correlation coefficient matrix between the data sources.

[0030] By extracting various environmental and equipment status data from labeled multi-source data, and using them as multiple data sources for pairwise correlation calculations, a correlation coefficient matrix reflecting the strength of numerical relationships between different data sources is constructed to provide a basis for subsequent weight adjustments.

[0031] By calculating the correlation coefficients between various data sources, it can be determined whether temperature changes are strongly correlated with increased equipment load, or whether humidity fluctuations have a certain regularity with the occurrence of fault alarms, so that the correlation structure between data sources can be presented in matrix form.

[0032] Step S400: Input the list of abnormal events and the correlation coefficient matrix into the preset weight adjustment model to obtain the dynamically adjusted weight vector.

[0033] By inputting the abnormal behavior information reflected in the abnormal event list and the data source correlation reflected in the correlation coefficient matrix into the weight adjustment model, the model can dynamically correct the initial weights based on the reliability and correlation of the data source, thereby obtaining a weight vector that is more in line with the current operating state.

[0034] When a certain type of environmental data frequently exhibits anomalies within a specific time period, the weight adjustment model can automatically reduce the impact of that data source on the final weighted fusion process, ensuring that the weighted results are more stable and reliable.

[0035] Step S500: Use the dynamically adjusted weight vector to perform weighted fusion of the labeled multi-source data to obtain fused data, and make intelligent decisions based on the fused data.

[0036] By applying dynamically adjusted weight vectors to labeled multi-source data, different types of data sources are weighted and fused according to their reliability and importance, so that the fused data has higher overall credibility. Then, by analyzing and judging the fused data, intelligent decision-making results are generated.

[0037] Intelligent decision-making operations, such as state prediction, risk assessment, and strategy optimization based on fused data, enable the system to dynamically adjust control strategies according to the actual operating environment, thereby achieving refined management.

[0038] In this embodiment, multi-source data is obtained by parallel acquisition of environmental data and equipment status data at the same timestamp. The multi-source data is then transmitted to the PLC at the device end for local preprocessing to obtain preprocessed multi-source data. Based on this, time-series feature extraction and anomaly detection are performed on the preprocessed multi-source data to generate labeled multi-source data and an anomaly event list. Subsequently, environmental data and equipment status data are extracted from the labeled multi-source data, and correlation analysis is performed on the two to construct a correlation coefficient matrix between data sources. The anomaly event list and correlation coefficient matrix are then input into a preset weight adjustment model to generate a dynamically adjusted weight vector. Finally, the dynamically adjusted weight vector is used to perform weighted fusion on the labeled multi-source data to obtain fused data, and intelligent decision-making is completed based on the fused data.

[0039] By collecting environmental and equipment status data at the same timestamp and preprocessing them at the PLC, multi-source data achieves consistent time alignment and basic cleaning. Time-series feature extraction and anomaly detection on the preprocessed multi-source data allow for early identification of data anomalies and generation of anomaly event lists before fusion, improving the reliability of subsequent analysis. Correlation analysis is performed on the labeled multi-source data, and a correlation coefficient matrix between data sources is constructed, quantifying the relationships between them. By inputting the anomaly event list and the correlation coefficient matrix into a weight adjustment model, dynamically adjusted weight vectors are generated based on changes in data source correlations and anomalies, enabling adaptive weight updates. Finally, the dynamically adjusted weight vectors are used to weight and fuse the labeled multi-source data, making the fused data more objective, stable, and aligned with actual operating conditions. This ensures higher accuracy and reliability of intelligent decision-making results even under complex operating conditions, addressing the issues of insufficient reliability in multi-source data fusion and the inability to adaptively adjust weights.

[0040] In another example, step S100 can preferably be implemented as follows: A unified clock source is configured in the PLC main control module as the time reference. Periodic time synchronization calibration is performed with the host computer time server via the NTP protocol. The calibrated system time is then distributed to the environmental data acquisition module and the equipment status data acquisition module. After receiving the system time, the environmental data acquisition module and the equipment status data acquisition module synchronously trigger data acquisition commands at the preset acquisition cycle time. According to the acquisition commands, the environmental data acquisition module is controlled to acquire real-time values ​​from the temperature sensor, humidity sensor, and pressure sensor as environmental data, and the equipment status data acquisition module is controlled to acquire equipment operating status indicators, motor speed parameters, and fault code information as equipment status data.

[0041] The PLC main control module generates a unified clock signal through a hardware timer, which serves as the trigger reference for each acquisition module. The environmental data acquisition module reads analog signals from temperature, humidity, and pressure sensors through the acquisition control register, converts the analog signals into digital signals using an ADC module, and records the current timestamp information. The equipment status data acquisition module reads the equipment operating status identifier and motor speed parameters through a CAN bus or Modbus communication interface, and reads fault code information from the fault record register. All acquisition modules trigger acquisition commands at the same time to ensure that the data acquired in each acquisition cycle has a consistent time reference.

[0042] For example, if the acquisition cycle in the PLC system is configured to be 100 milliseconds, the environmental data acquisition module and the equipment status data acquisition module will be triggered simultaneously every 100 milliseconds. The environmental data acquisition module reads the output of 23.5°C from the temperature sensor, 45%RH from the humidity sensor, and 101.2kPa from the pressure sensor. At the same time, the equipment status acquisition module reads the motor speed as 1500rpm, the running status as "running", and the fault code as 0x00.

[0043] A timestamp is added to the environmental data and equipment status data, including the year, month, day, hour, minute, second, and millisecond information. The environmental data and equipment status data with the added timestamp are then transmitted to the data buffer via the PLC's internal bus. In the data buffer, the environmental data and equipment status data are arranged and combined according to the timestamp order to form a data group with the same timestamp, thus obtaining multi-source data.

[0044] The data buffer stores the collected environmental and equipment status data through a FIFO queue structure and sorts the data according to timestamps. The timestamps are generated by the PLC system clock and are accurate to milliseconds. The multi-source data group construction logic forms a unified structure or record set of environmental and equipment status data by matching data with the same timestamp, so that subsequent processing modules can directly read and analyze them.

[0045] For example, when the timestamp is 20:30:15:123 on November 26, 2025, the temperature of 23.5°C, humidity of 45%RH, pressure of 101.2kPa, equipment status "running", motor speed of 1500rpm and fault code 0x00 are combined into a single data record and written to the data buffer area through the PLC internal bus.

[0046] The multi-source data is transmitted to the edge computing unit of the PLC. In the edge computing unit, the multi-source data is cleaned to remove abnormal data points whose values ​​exceed the preset threshold range, thus obtaining the cleaned multi-source data.

[0047] The edge computing unit performs threshold verification on multi-source data, such as temperature in the range of -20°C to 80°C, humidity in the range of 0%RH to 100%RH, and pressure in the range of 90kPa to 110kPa. Sampling points that do not meet the thresholds are marked as abnormal and removed. The removed data is then re-sorted and indexed in time series to ensure the effectiveness and stability of subsequent time series analysis.

[0048] If the collected temperature is 105°C, the data point is identified as abnormal and removed. The remaining temperature data of 23.5°C, 23.6°C, and 23.4°C are retained and arranged in order of timestamp.

[0049] The cleaned multi-source data is normalized to map environmental data and equipment status data of different dimensions to a unified numerical range of zero to one, resulting in normalized multi-source data. Differential coding is used to compress and store the normalized multi-source data whose change amplitude under continuous timestamps is less than a preset change threshold, resulting in preprocessed multi-source data.

[0050] Normalization is achieved by subtracting the minimum value from the original data and then dividing by the difference between the maximum and minimum values, thus mapping data of different dimensions such as temperature, humidity, pressure, and motor speed to the [0,1] interval. Differential coding is achieved by calculating the increment of continuous timestamp data and recording only the increment identifier rather than the absolute value for points with increment amplitudes below a threshold, thereby compressing the data. The preprocessed multi-source data is stored in the high-speed memory of the edge computing unit for use in temporal feature extraction.

[0051] For example, the temperature of 23.5°C is normalized to 0.47, the humidity of 45%RH is normalized to 0.45, the temperature increments of three consecutive timestamps are 0.01, 0.005, and 0.008, and the increments below the threshold of 0.02 are stored in differential compression form, which finally forms a compressed preprocessed data sequence.

[0052] In another example, step S200 can preferably be implemented as follows: The preprocessed multi-source data are arranged in chronological order according to timestamps to construct environmental data time series and equipment status data time series.

[0053] The PLC or edge computing unit sorts the preprocessed multi-source data according to the timestamp, and constructs independent time series for temperature, humidity, pressure, equipment status indicators, motor speed and fault codes, providing an ordered input sequence for sliding window feature extraction and anomaly detection.

[0054] For example, according to the timestamps 20:30:15:123, 20:30:15:223, and 20:30:15:323 on November 26, 2025, the corresponding environmental data and device status data are arranged into six independent time series.

[0055] Set a sliding window with a length of N sampling points. According to the sliding window, slide it through the environmental data time series and the device status data time series with a preset step size. At each window position, extract the statistical feature parameters of the data within the window. The statistical feature parameters include mean, variance, maximum value, minimum value and rate of change.

[0056] The sliding window can be set to a length of N=10 sampling points and a step size of 2 sampling points. The window slides gradually on each time series. Within each window, the mean, variance, maximum value, minimum value and rate of change are calculated to form window-level features, which are used to capture local time series fluctuations.

[0057] For example, within the window from 20:30:15:123 to 20:30:15:323, the average temperature is 23.52°C, the variance is 0.01, the maximum value is 23.6°C, the minimum value is 23.45°C, and the rate of change is 0.15°C / s.

[0058] The statistical characteristic parameters of environmental data time series and equipment status data time series are concatenated and combined to obtain a multidimensional feature vector sequence.

[0059] Each time window concatenates the environmental data feature vector and the device status data feature vector in a fixed order to form a multidimensional vector, ensuring that the feature information of each data source is completely preserved, forming a continuous multidimensional feature vector sequence for subsequent anomaly detection.

[0060] For example, the concatenated vector is [mean temperature, variance temperature, maximum temperature, mean humidity, variance humidity, maximum motor speed...], with a length of 20 dimensions.

[0061] The multidimensional feature vector sequence is input into a pre-trained isolated forest anomaly detection model. The isolated forest anomaly detection model constructs multiple isolated trees by randomly selecting features and split points, calculates the average path length of each feature vector in the multidimensional feature vector in the isolated trees, and calculates the anomaly score based on the average path length.

[0062] During the training phase, the Isolation Forest model has built 500 isolated trees based on historical normal data. The average path length of each input vector is calculated and mapped to an anomaly score. The higher the score, the more abnormal the data is within that time window.

[0063] If the anomaly score of a certain window is 0.87, which is higher than the anomaly threshold of 0.8, then the data in that window is determined to be abnormal.

[0064] When the anomaly score is greater than the preset anomaly threshold, the data at the corresponding time is determined to be anomalous data, and an anomaly label is added to the preprocessed multi-source data for the corresponding time, resulting in labeled multi-source data. The anomaly label includes an anomaly type code and an anomaly confidence value.

[0065] Anomaly identification labels include type codes (such as temperature anomaly = 0x01, motor speed anomaly = 0x02) and confidence values ​​(between 0 and 1), creating fields in multi-source data records to record anomaly information.

[0066] If the timestamp 20:30:15:123 corresponds to a temperature anomaly, with anomaly type 0x01 and anomaly confidence level of 0.92, the tag will be written to that timestamp record.

[0067] Extract the timestamps, exception type codes, and exception confidence scores of all abnormal data and organize them in chronological order to construct an abnormal event list. Each record in the abnormal event list includes the time when the exception occurred, the duration of the exception, and the identifier of the data source involved.

[0068] The edge computing unit scans the labeled multi-source data, merges consecutive abnormal time periods to form event records, calculates the duration of the anomalies, and identifies the data sources involved (environmental data or device status data) to form a structured list of abnormal events.

[0069] For example, the abnormal event list records: the abnormality started at 20:30:15:123 and ended at 20:30:15:523, lasting for 400 milliseconds, involving a temperature data source, and the abnormality type is 0x01.

[0070] In another example, step S300 can preferably be implemented as follows: Environmental data and equipment status data are separated and extracted from the labeled multi-source data. Temperature, humidity and pressure data in the environmental data are constructed into independent data vectors, and operating status identifiers, motor speed parameters and fault code information in the equipment status data are constructed into independent data vectors.

[0071] The edge computing unit scans the labeled multi-source data records and extracts the temperature, humidity, and pressure information from the environmental data fields in timestamp order, constructing them into temperature vectors of length L, T=[T1,T2,…,T…]. L Humidity vector H = [H1, H2, ..., H L Pressure vector P = [P1, P2, ..., P] L Simultaneously, the operating status identifier, motor speed parameters, and fault code information in the equipment status data field are each constructed into a vector of equal length S=[S1,S2,…,S…]. L R = [R1, R2, ..., R] L ] and F=[F1,F2,…,F L All vectors maintain consistent timestamp correspondence to facilitate subsequent correlation analysis and weight calculation.

[0072] For example, within 10 seconds from timestamp 20:30:15.123 to 20:30:25.123, the temperature vector is T=[23.5,23.6,23.4,…], the humidity vector is H=[45.0,44.8,45.2,…], the pressure vector is P=[101.2,101.3,101.1,…], the operating status vector is S=[1,1,1,…] (1 represents “running”, 0 represents “stopped”), the motor speed vector is R=[1500,1505,1498,…], and the fault code vector is F=[0x00,0x00,0x01,…].

[0073] All the independent data vectors are aggregated to obtain M data source vectors, where M is the total number of data sources.

[0074] Environmental data and device status data are integrated into a data source vector set V=[T,H,P,S,R,F], M=6, using a vector list or matrix format. Each vector has a length of L, ensuring that all data sources are fully aligned in the time dimension, facilitating unified calculation and analysis.

[0075] For example, the aggregated data source vector set is V=[T,H,P,S,R,F], where T is the temperature vector, H is the humidity vector, P is the pressure vector, S is the operating status vector, R is the speed vector, and F is the fault code vector, for a total of 6 vectors.

[0076] Calculate the Pearson correlation coefficient between any two data source vectors from the M data source vectors. The Pearson correlation coefficient is calculated by dividing the covariance of the two data source vectors by the product of their respective standard deviations.

[0077] For each pair of continuous vectors V i and V j Calculate the covariance Cov(V) i V j ), and calculate V respectively. i and V j Standard deviation σ i and σ j Use the formula:

[0078] The linear correlation coefficients between pairs of data are obtained; this calculation can be performed by the mathematical library of the PLC edge computing unit or by an embedded Python / Matlab script to achieve automated batch processing.

[0079] If the covariance of the temperature vector T and the humidity vector H is 0.012, the standard deviation of T is 0.05, and the standard deviation of H is 0.06, then the Pearson correlation coefficient ρ(T,H) = 0.012 / (0.05*0.06) = 4.0 (after normalization, it is limited to the interval [-1,1]), indicating a high positive correlation.

[0080] When the data types of the two data source vectors are inconsistent, the Spearman rank correlation coefficient is used instead of the Pearson correlation coefficient for calculation. The Spearman rank correlation coefficient is calculated by sorting the original data and then calculating the correlation between the ranks. All pairwise combinations of data source vectors are traversed, and the correlation coefficient between each pair of data sources is calculated. The correlation coefficients are filled into an M-row M-column matrix according to the order of the data sources. The element in the i-th row and j-th column of the matrix represents the correlation coefficient between the i-th data source and the j-th data source. The diagonal elements of the matrix are 1, and the values ​​of the off-diagonal elements range from -1 to +1.

[0081] For data sources containing categorical or non-linearly distributed data, such as operational status identifiers S and fault codes F, first sort the elements in the vector by size to obtain an ordered vector, then calculate the covariance of the rank differences, and use Spearman's formula to calculate ρ. s Iterate through M(M-1) / 2 pairs of vectors and fill the results into the correlation coefficient matrix CorrMatrix. The diagonal elements of the matrix are automatically set to 1, and the off-diagonal elements are taken as calculated values ​​in the range of [-1,1].

[0082] If the running status vector S=[1,1,0,1] and the fault code vector F=[0x00,0x01,0x00,0x01] are sorted, the resulting rank vector S0 is obtained. r =[3,3,1,3],F r =[1,4,1,4], calculate the Spearman rank correlation coefficient ρ. s =-0.5, fill in the 4th row and 6th column and the 6th row and 4th column of the matrix.

[0083] Elements in the matrix whose absolute values ​​of correlation coefficients are less than a preset correlation threshold are zeroed out, while significantly related data source relationships are retained, resulting in a correlation coefficient matrix between data sources.

[0084] By setting the correlation threshold to 0.3, elements in the CorrMatrix with an absolute value < 0.3 are set to 0, while the rest remain unchanged. This filters out weak correlations, reduces noise interference, and provides a reliable foundation for subsequent feature weighting or data source weight optimization.

[0085] If the correlation coefficient between temperature and fault codes is 0.12, it is set to zero; the correlation coefficient between temperature and humidity of 0.78 is retained, forming the final correlation coefficient matrix: .

[0086] The weight adjustment model includes a reliability assessment module and a weight optimization module. The reliability assessment module is used to calculate the reliability coefficient of each data source, and the weight optimization module is used to reduce and adjust the initial weight of data sources with high anomaly rates.

[0087] The reliability assessment module calculates the reliability coefficient R based on the frequency of anomalies, the proportion of missing data, and the fluctuation range of each data source after annotation of multi-source data. i =1 - (Outlier rate 0.6 + Missing rate 0.3 + Volatility coefficient * 0.1); The weight optimization module will use the initial weight W i Multiplying by the reliability coefficient, if the anomaly rate of a data source exceeds a threshold of 0.2, the weight of that data source is automatically reduced, resulting in an optimized weight vector W. t , used for subsequent fusion calculations.

[0088] If the outlier rate is 0.05, the missing data rate is 0.02, and the volatility is 0.03, calculate the reliability coefficient R. t =1-(0.050.6+0.020.3+0.030.1)=0.966, initial weight 0.2, then the optimized weight W t =0.20.966≈0.193.

[0089] In another example, step S400 can preferably be implemented as follows: The absolute values ​​of each row of the correlation coefficient matrix are summed to obtain the total correlation strength between each data source and other data sources. The total correlation strength is then normalized so that the sum of the weight coefficients of all data sources equals 1, thus obtaining the initial weight vector.

[0090] Summing the absolute values ​​ensures that highly correlated data sources have higher initial weights, while less correlated data sources have lower initial weights, providing a basis for subsequent dynamic adjustments.

[0091] If M=6 data sources, the sum of the absolute values ​​of the correlation coefficients between temperature and other data sources is 2.0, humidity is 1.8, pressure is 1.6, operating status is 1.2, speed is 1.4, and fault code is 0.9, then the initial weight vector = [0.24, 0.216, 0.192, 0.144, 0.168, 0.108] (the sum is normalized to 1).

[0092] The abnormality rate of each data source is obtained by counting the frequency of abnormal events in the list of abnormal events within a preset time window, calculating the proportion of abnormal frequency to the total number of samples.

[0093] The PLC edge computing unit or host computer groups the list of abnormal events by data source and counts the number of times n abnormal events occur for each data source within the most recent N sampling time steps. e Divide the frequency of anomalies by the total number of sampling points N. t Calculate the anomaly rate R e =n e / N t This yields an anomaly rate vector R = [R1, R2, ..., R] within the range [0, 1]. M ].

[0094] For example, in the most recent 1000 sampling points, the temperature was abnormal 15 times, the humidity was abnormal 10 times, the pressure was abnormal 8 times, the operating status was abnormal 5 times, the speed was abnormal 12 times, and the fault code was abnormal 20 times. Then the abnormality rate vector R = [0.015, 0.01, 0.008, 0.005, 0.012, 0.02].

[0095] The anomaly rate of each data source is input into the reliability assessment module to calculate the reliability coefficient of each data source. The reliability coefficient is inversely proportional to the anomaly rate; the higher the anomaly rate, the lower the reliability coefficient. The initial weight vector and reliability coefficient are input into the weight optimization module. The weight optimization module uses a penalty function to reduce the initial weight of data sources with high anomaly rates. The penalty force of the penalty function increases exponentially with the anomaly rate. A time decay factor is introduced into the weight optimization module to assign lower weights to historical anomalies that are far removed from the current time. The time decay factor uses an exponential decay function, so that the impact of recent anomalies on weight adjustment is greater than that of historical anomalies. The weight optimization module normalizes the weight coefficients of each data source after adjustment to ensure that the sum of the adjusted weight coefficients is still equal to 1, resulting in a dynamically adjusted weight vector. When there are related anomalies in multiple data sources in the anomaly event list, the weight optimization module further reduces the weight coefficients of the data sources involved in the related anomalies and distributes the reduced weights proportionally to other data sources with weaker correlations based on the correlation coefficient matrix.

[0096] In this step, the reliability assessment module uses formula R t =exp(-k*R e *D t Calculate the reliability coefficient, where k is the adjustment constant and D t The time decay factor is the near real-time anomalous event D. t ≈1, Historical Anomaly D t <1; The weight optimization module will combine the initial weights with the reliability coefficient R. i Multiplication results in a penalty adjustment weight W a =W i *R iAfter normalization, a dynamic weight vector W is formed. y =W a / ΣW a If there are associated abnormal events, then the W data source involved will be affected. y The proportion is reduced, and the reduced portion is allocated to data sources with low correlation according to the correlation coefficient matrix at a ratio of 1 - correlation to achieve overall balance.

[0097] For example, with an initial temperature weight of 0.24, an anomaly rate of 0.015, k=5, and time decay D... t =1, then the reliability coefficient R t =exp(-5*0.015*1)=0.927, dynamic weight W y =0.24*0.927≈0.222; If temperature and humidity both show abnormal correlation, the dynamic weight will be further reduced by 10%, and the reduced portion will be allocated to these data sources according to the correlation ratio between temperature and pressure, and rotation speed.

[0098] In another example, step S500 can preferably be implemented as follows: Extract the values ​​of each data source corresponding to each timestamp from the labeled multi-source data, and multiply the weight coefficients in the dynamically adjusted weight vector with the values ​​of the corresponding data sources to obtain the weighted values ​​of each data source.

[0099] For each timestamp t, the values ​​V from each data source are... i (t) and the corresponding dynamic weight W yi Multiply to obtain the weighted value V. w (t)=V i (t)*W yi This operation can be implemented in the PLC edge computing unit or the host computer batch processing module to form a weighted matrix, with each row corresponding to a timestamp and each column corresponding to a data source.

[0100] For example, at timestamp t=20:30:15.123, with a temperature value of 23.5°C and a dynamic weight of 0.222, the weighted value is 23.5*0.222≈5.217; the weighted value for humidity is calculated similarly.

[0101] The weighted values ​​of all data sources at the same timestamp are summed to obtain the merged value of each timestamp. The weighted summation process is repeated for all timestamps to form the merged time series data.

[0102] The summation of all weighted values ​​at timestamp t is F(t) = ΣV w (t), iterate through all L timestamps to generate a fusion sequence F=[F(1),F(2),…,F(L)], which maintains the time order and can be directly input into subsequent smoothing filters or intelligent decision models.

[0103] For example, when t=20:30:15.123, the weighted average is 5.217 for temperature, 0.432 for humidity, 0.312 for pressure, 0.05 for operating status, 0.21 for speed, and 0.36 for fault code. Therefore, the fused value F=5.217+0.432+0.312+0.05+0.21+0.36≈6.581.

[0104] The fused time-series data undergoes smoothing filtering, employing a moving average filter to eliminate high-frequency noise interference. The window length of the moving average filter is dynamically determined based on the data sampling frequency and the desired smoothness. The resulting smoothed fused data is then input into a pre-trained intelligent decision-making model. This model utilizes a Long Short-Term Memory (LSTM) neural network structure, comprising an input layer, multiple LSTM hidden layers, and an output layer. The input layer receives fused data from multiple consecutive time steps as the input sequence. The LSTM hidden layers extract the temporal dependency features of the fused data through a gating mechanism. The output layer outputs the decision result based on the temporal dependency features extracted by the LSTM hidden layers, adjusting the equipment's operating parameters or triggering protective actions accordingly.

[0105] Set the moving average window length w, based on the sampling frequency f. s And expected smoothing time T s w=ceil(f s *T s The smoothed sequence is input into the LSTM network, and the input sequence length is L. s eq, with H hidden layer units, the LSTM network calculates time-dependent features through input gate, forget gate, and output gate. The output layer adopts a fully connected layer to obtain the prediction decision result D(t), which can represent the adjustment amount of equipment operating parameters or the trigger signal of protection action.

[0106] For example, with a sampling frequency of 100Hz, a desired smoothing time of 0.5 seconds, and a window length of w=50, the smoothed fusion value of 20:30:15.123 is approximately 6.502. The smoothed fusion values ​​of the past 50 time steps are used as the input sequence to input the LSTM, with hidden layer units H=64. The output decision D(20:30:15.123)=1 indicates that a mild adjustment operation is triggered.

[0107] The preferred steps for adjusting equipment operating parameters or triggering protection actions based on the decision results are as follows: The decision results are analyzed to obtain the fault risk level and recommended operation instructions.

[0108] The intelligent decision-making model outputs D(t) and converts it into a fault risk level ∈{low, medium, high} through a mapping table, and generates recommended operation instructions ∈{adjust speed, adjust temperature, shut down for protection, continue running}.

[0109] D(t)=2 corresponds to a medium risk level, and the recommended operation command is to adjust the speed to decrease by 10%.

[0110] When the fault risk level exceeds the preset safety threshold, the intelligent decision-making model generates an alarm signal and outputs corresponding emergency handling instructions.

[0111] If the fault risk level is greater than or equal to the set safety threshold, an alarm signal will be generated, and an emergency command will be generated at the same time, such as immediately stopping the machine or triggering the protection relay.

[0112] For example, if the risk level is 1, the emergency response instruction is to shut down immediately.

[0113] The decision results are correlated with the anomaly labels in the labeled multi-source data. When the decision results and anomaly labels point to the same risk trend, the confidence of the decision is increased and the recommended operation instructions are executed first.

[0114] The anomaly label under timestamp t is compared with the decision output. If they match (e.g., both show that the temperature is too high), the decision confidence is increased and the corresponding operation instruction is executed first.

[0115] An anomaly label indicates a temperature anomaly, prompting a temperature adjustment command to be output. The decision confidence level increases from 0.85 to 1.02, and the adjustment is executed first.

[0116] Emergency handling instructions or recommended operation instructions are sent to the actuator through the PLC's output module, enabling the actuator to adjust the equipment's operating parameters or trigger protection actions according to the emergency handling instructions or recommended operation instructions. At the same time, feedback data after execution is collected and sent back to the intelligent decision model for closed-loop verification and online updating of model parameters.

[0117] The PLC output module converts instructions into control signals, such as PWM signals and relay control signals, and sends them to the corresponding actuators via the fieldbus. After execution, the sensors collect feedback data such as temperature, humidity, and motor status and send it back to the edge computing unit. The intelligent decision model updates the LSTM parameters based on the feedback to achieve online closed-loop optimization.

[0118] The speed adjustment command is sent to the motor drive module, and the actual motor speed is adjusted from 1500rpm to 1350rpm. After the feedback data is updated, the model fine-tunes the weights to improve the accuracy of the next decision.

[0119] In this application example, a multi-source data fusion intelligent decision-making system 10 is also provided, including a data acquisition module 11, an anomaly detection module 12, a correlation analysis module 13, a data input module 14, and a weighted fusion module 15.

[0120] The data acquisition module 11 is used to collect environmental data and equipment status data in parallel at the same timestamp to obtain multi-source data. The multi-source data is then transmitted to the PLC at the device end for local preprocessing to obtain preprocessed multi-source data.

[0121] The anomaly detection module 12 is used to extract time-series features and detect anomalies in the preprocessed multi-source data to obtain labeled multi-source data and a list of anomaly events.

[0122] The correlation analysis module 13 is used to perform correlation analysis on environmental data and equipment status data in the labeled multi-source data to obtain the correlation coefficient matrix between data sources.

[0123] The data input module 14 is used to input the list of abnormal events and the correlation coefficient matrix into the preset weight adjustment model to obtain the dynamically adjusted weight vector.

[0124] The weighted fusion module 15 is used to perform weighted fusion on the labeled multi-source data using dynamically adjusted weight vectors to obtain fused data, and to make intelligent decisions based on the fused data.

[0125] In this embodiment, the data acquisition module 11 collects environmental data and equipment status data in parallel at the same timestamp, and performs local preprocessing on the PLC device to generate preprocessed data that can be used for subsequent analysis; the anomaly detection module 12 performs time-series feature extraction and anomaly detection on the preprocessed multi-source data to obtain labeled multi-source data and an anomaly event list, which is used to identify abnormal data points and anomaly event characteristics; the correlation analysis module 13 performs pairwise correlation analysis on the environmental data and equipment status data in the labeled multi-source data, calculates the correlation coefficient matrix between data sources, and quantifies the degree of mutual influence between each data source; the data input module 1... 4. Input the list of abnormal events and the correlation coefficient matrix into the preset weight adjustment model. Through reliability assessment and weight optimization, the initial weights of each data source are dynamically adjusted to generate a dynamically adjusted weight vector, thereby realizing weight adjustment for data sources with frequent anomalies or strong correlations. The weighted fusion module 15 multiplies the dynamically adjusted weight vector with the corresponding labeled multi-source data and sums them to obtain the fused data sequence. After smoothing the fused data, it is input into the intelligent decision model for analysis and outputs optimized equipment operating parameters or protection commands. This enables real-time monitoring of equipment operating status, anomaly warning, and intelligent decision adjustment, thereby improving the system's decision accuracy, stability, and response efficiency.

[0126] It should be noted that although several modules or units of the system for executing actions have been mentioned in the detailed description above, this division is not mandatory. In fact, according to the embodiments of this application, the features and functions of two or more modules or units described above can be embodied in one module or unit. Conversely, the features and functions of one module or unit described above can be further divided and embodied by multiple modules or units.

[0127] Through the above description of the embodiments, those skilled in the art will readily understand that the exemplary embodiments described herein can be implemented by software or by combining software with necessary hardware. Therefore, the technical solutions according to the embodiments of this application can be embodied in the form of a software product, which can be stored in a non-volatile storage medium (such as a CD-ROM, USB flash drive, external hard drive, etc.) or on a network, including several instructions to cause an electronic device (such as a personal computer, server, touch terminal, or network device, etc.) to execute the methods according to the embodiments of this application.

[0128] Other embodiments of this application will readily occur to those skilled in the art upon consideration of the specification and practice of the embodiments disclosed herein. This application is intended to cover any variations, uses, or adaptations of this application that follow the general principles of this application and include common knowledge or customary techniques in the art not disclosed herein.

[0129] It should be understood that this application is not limited to the precise structure described above and shown in the accompanying drawings, and various modifications and changes can be made without departing from its scope. The scope of this application is limited only by the appended claims.

Claims

1. An intelligent decision-making method based on multi-source data fusion, characterized in that, include: Environmental data and equipment status data are collected in parallel at the same timestamp to obtain multi-source data. The multi-source data is then transmitted to the PLC at the equipment end for local preprocessing to obtain preprocessed multi-source data. Temporal feature extraction and anomaly detection are performed on the preprocessed multi-source data to obtain labeled multi-source data and a list of abnormal events; Correlation analysis was performed on the environmental data and equipment status data in the labeled multi-source data to obtain the correlation coefficient matrix between the data sources. Input the list of abnormal events and the correlation coefficient matrix into a preset weight adjustment model to obtain a dynamically adjusted weight vector; The labeled multi-source data is weighted and fused using the dynamically adjusted weight vector to obtain fused data, and intelligent decision-making is performed based on the fused data.

2. The intelligent decision-making method based on multi-source data fusion according to claim 1, characterized in that, The steps of parallel acquisition of environmental data and equipment status data at the same timestamp to obtain multi-source data, and transmitting the multi-source data to the PLC at the equipment end for local preprocessing to obtain preprocessed multi-source data include: At the preset collection cycle time, a data collection command is triggered synchronously. According to the collection command, the environmental data collection module is controlled to collect the real-time values ​​of the temperature sensor, humidity sensor and pressure sensor as environmental data, and the equipment status data collection module is controlled to collect the equipment operating status identifier, motor speed parameters and fault code information as equipment status data. The environmental data and the equipment status data are appended with a timestamp of the current time. The environmental data and equipment status data with the appended timestamp are then transmitted to the data buffer via the PLC internal bus. In the data buffer, the environmental data and equipment status data are arranged and combined according to the timestamp order to obtain multi-source data. The multi-source data is transmitted to the edge computing unit of the PLC, where the multi-source data is cleaned to obtain cleaned multi-source data. The cleaned multi-source data is normalized to obtain normalized multi-source data. The normalized multi-source data whose change amplitude under consecutive timestamps is less than a preset change threshold is compressed and stored to obtain preprocessed multi-source data.

3. The intelligent decision-making method based on multi-source data fusion according to claim 1, characterized in that, The step of performing time-series feature extraction and anomaly detection on the preprocessed multi-source data to obtain labeled multi-source data and a list of anomalous events includes: The preprocessed multi-source data are arranged in chronological order according to timestamps to construct environmental data time series and equipment status data time series; Set a sliding window with a length of N sampling points, and slide it through the environmental data time series and the device status data time series with a preset step size according to the sliding window, and extract the statistical feature parameters of the data in the window at each window position; The statistical feature parameters of the environmental data time series and the statistical feature parameters of the equipment status data time series are concatenated and combined to obtain a multidimensional feature vector sequence. The multidimensional feature vector sequence is input into a pre-trained isolated forest anomaly detection model. The average path length of each feature vector in the multidimensional feature vector in the isolated tree is calculated, and the anomaly score is calculated based on the average path length. When the anomaly score is greater than the preset anomaly threshold, the data at the corresponding time is determined to be abnormal data, and an anomaly identification label is added to the preprocessed multi-source data for the corresponding time to obtain labeled multi-source data. The anomaly identification label includes an anomaly type code and an anomaly confidence value. Extract the timestamps, exception type codes, and exception confidence scores of all exception data and organize them in chronological order to construct a list of exception events.

4. The intelligent decision-making method based on multi-source data fusion according to claim 1, characterized in that, The step of performing correlation analysis on the environmental data and equipment status data in the labeled multi-source data to obtain the correlation coefficient matrix between the data sources includes: Environmental data and equipment status data are separated and extracted from the labeled multi-source data. Temperature data, humidity data and pressure data in the environmental data are respectively constructed into independent data vectors. Operating status identifiers, motor speed parameters and fault code information in the equipment status data are respectively constructed into independent data vectors. All the independent data vectors are aggregated to obtain M data source vectors, where M is the total number of data sources; Calculate the Pearson correlation coefficient between any two data source vectors from the M data source vectors; When the data types of two data source vectors are inconsistent, Spearman's rank correlation coefficient is used instead of Pearson's correlation coefficient for calculation. All pairwise combinations of data source vectors are traversed, and the correlation coefficient between each pair of data sources is calculated. The correlation coefficient is then filled into an M-row, M-column matrix according to the order of the data sources. Elements in the matrix whose absolute values ​​of correlation coefficients are less than a preset correlation threshold are zeroed out to obtain a correlation coefficient matrix between data sources.

5. The intelligent decision-making method based on multi-source data fusion according to claim 1, characterized in that, The weight adjustment model includes a reliability assessment module and a weight optimization module. The reliability assessment module is used to calculate the reliability coefficient of each data source, and the weight optimization module is used to reduce and adjust the initial weight of data sources with high anomaly rates. The step of inputting the list of abnormal events and the correlation coefficient matrix into a preset weight adjustment model to obtain a dynamically adjusted weight vector includes: The absolute values ​​of each row of the correlation coefficient matrix are summed to obtain the total correlation strength between each data source and other data sources. The total correlation strength is then normalized to obtain the initial weight vector. The abnormality rate of each data source is obtained by counting the frequency of abnormalities occurring in each data source within a preset time window from the list of abnormal events and calculating the proportion of the abnormality frequency to the total number of samples. The anomaly rate of each data source is input into the reliability assessment module to calculate the reliability coefficient of each data source. The initial weight vector and the reliability coefficient are then input into the weight optimization module to obtain the dynamically adjusted weight vector.

6. The intelligent decision-making method based on multi-source data fusion according to claim 1, characterized in that, The step of using the dynamically adjusted weight vector to perform weighted fusion of the labeled multi-source data to obtain fused data, and then making intelligent decisions based on the fused data, includes: Extract the values ​​of each data source corresponding to each timestamp from the labeled multi-source data, and multiply the weight coefficients in the dynamically adjusted weight vector with the values ​​of the corresponding data sources to obtain the weighted values ​​of each data source. The weighted values ​​of all data sources at the same timestamp are summed to obtain the merged time series data; The fused time series data is smoothed and filtered to obtain smoothed fused data. The smoothed fused data is then input into a pre-trained intelligent decision-making model, which outputs decision results. Based on the decision results, the operating parameters of the equipment are adjusted or protection actions are triggered.

7. The intelligent decision-making method for multi-source data fusion according to claim 6, characterized in that, The step of adjusting the operating parameters of the equipment or triggering protection actions based on the decision result includes: The decision results are analyzed to obtain the fault risk level and recommended operation instructions; When the fault risk level exceeds the preset safety threshold, the intelligent decision-making model is used to generate an alarm signal and output the corresponding emergency handling instructions. The decision result is correlated with the anomaly identification labels in the labeled multi-source data. When the decision result and the anomaly identification labels point to the same risk trend, the recommended operation instruction is executed. The emergency handling instruction or the recommended operation instruction is sent to the actuator through the output module of the PLC, so that the actuator can adjust the operating parameters of the equipment or trigger protection actions according to the emergency handling instruction or the recommended operation instruction.

8. An intelligent decision-making system that integrates multi-source data, characterized in that, include: The data acquisition module is used to collect environmental data and equipment status data in parallel at the same timestamp to obtain multi-source data. The multi-source data is then transmitted to the PLC at the device end for local preprocessing to obtain preprocessed multi-source data. Anomaly detection module is used to extract time-series features and detect anomalies in the preprocessed multi-source data to obtain labeled multi-source data and a list of anomaly events; The correlation analysis module is used to perform correlation analysis on environmental data and equipment status data in the labeled multi-source data to obtain the correlation coefficient matrix between data sources; The data input module is used to input the list of abnormal events and the correlation coefficient matrix into a preset weight adjustment model to obtain a dynamically adjusted weight vector. The weighted fusion module is used to perform weighted fusion on the labeled multi-source data using the dynamically adjusted weight vector to obtain fused data, and to make intelligent decisions based on the fused data.

Images