Hydropower station inspection method and device based on multi-source sensing equipment

Abstract

The invention discloses a hydropower station inspection method and device based on multi-source sensing equipment, relates to the technical field of hydropower station management, and discloses a hydropower station inspection method based on the multi-source sensing equipment, and the method comprises the steps: obtaining multi-source operation data which comprises sensor telemetry data, industrial control operation data and structured visual data; constructing a multi-modal feature vector according to the multi-source operation data; performing equipment state prediction and trend analysis according to a long-term time sequence prediction strategy and the multi-modal feature vector to obtain an equipment operation trend analysis result; generating target early warning information according to the multi-modal feature vector and an isolation forest strategy; and generating a target inspection task according to the equipment operation trend analysis result. The technical effects of accurately monitoring the early failure of the equipment, predicting the running trend of the equipment, generating dynamic early warning information and optimizing the inspection task are achieved.

Description

Technical Field

[0001] This application relates to the field of hydropower station management technology, and in particular to a hydropower station inspection method and device based on multi-source sensing equipment. Background Technology

[0002] Currently, equipment inspection and operation monitoring at hydropower stations are gradually moving towards a digital stage, but significant limitations still exist in practice. Most hydropower stations still rely on fixed measuring points to collect a limited number of operating parameters and judge equipment status through simple threshold alarms. Meanwhile, traditional data acquisition methods have low data frequency and limited content, making it difficult to reflect the multidimensional state changes of equipment under complex operating environments. Because early, minor anomalies are often masked by noise or cross-domain data is difficult to align, latent equipment faults are often only discovered after they accumulate to obvious anomalies, thus limiting the ability of maintenance personnel to perceive and intervene in potential risks in advance.

[0003] Against this backdrop, traditional hydropower station inspection methods typically rely on periodic manual inspections and alarm mechanisms based on fixed thresholds. Maintenance personnel check equipment operation according to the inspection cycle and depend on over-limit alarms from a small number of sensors in the monitoring system to assess fault risk. While this method can detect visible faults, its detection frequency is low, its response is delayed, and data from different sources is isolated, making it unable to support a comprehensive assessment of equipment operating trends and early anomalies.

[0004] The above content is only used to help understand the technical solution of this application and does not represent an admission that the above content is prior art. Summary of the Invention

[0005] The main purpose of this application is to provide a hydropower station inspection method and device based on multi-source sensing equipment, which aims to solve the technical problem that existing hydropower station inspections cannot achieve accurate real-time identification and trend prediction of early, minor, and latent faults in hydropower station equipment.

[0006] To achieve the above objectives, this application proposes a hydropower station inspection method based on multi-source sensing devices, the method comprising: Acquire multi-source operational data, including sensor telemetry data, industrial control operational data, and structured vision data; Construct a multimodal feature vector based on the multi-source operational data; Based on the long-term time series prediction strategy and the multimodal feature vector, equipment status prediction and trend analysis are performed to obtain equipment operation trend analysis results; Target early warning information is generated based on the multimodal feature vectors and the isolated forest strategy; Target inspection tasks are generated based on the analysis results of the equipment operation trends.

[0007] In addition, to achieve the above objectives, this application also proposes a hydropower station inspection device based on multi-source sensing equipment. The hydropower station inspection device based on multi-source sensing equipment includes: a data acquisition module for acquiring multi-source operating data, which includes sensor telemetry data, industrial control operating data and structured vision data. A vector construction module is used to construct multimodal feature vectors based on the multi-source runtime data; The trend analysis module is used to perform equipment status prediction and trend analysis based on the long-term time series prediction strategy and the multimodal feature vector to obtain equipment operation trend analysis results. The early warning generation module is used to generate target early warning information based on the multimodal feature vector and the isolated forest strategy; The task generation module is used to generate target inspection tasks based on the analysis results of the equipment operation trends.

[0008] In addition, to achieve the above objectives, this application also proposes a hydropower station inspection device based on a multi-source sensing device. The device includes a memory, a processor, and a computer program stored in the memory and executable on the processor. The computer program is configured to implement the steps of the hydropower station inspection method based on the multi-source sensing device described above.

[0009] In addition, to achieve the above objectives, this application also proposes a storage medium, which is a computer-readable storage medium, on which a computer program is stored. When the computer program is executed by a processor, it implements the steps of the hydropower station inspection method based on multi-source sensing devices as described above.

[0010] In addition, to achieve the above objectives, this application also provides a computer program product, which includes a computer program that, when executed by a processor, implements the steps of the hydropower station inspection method based on multi-source sensing devices as described above.

[0011] One or more technical solutions proposed in this application have at least the following technical effects: By employing multi-source data acquisition methods, including sensor telemetry data, industrial control operation data, and structured visual data, combined with multimodal feature vector construction, long-term time-series prediction strategies for equipment status prediction and trend analysis, isolated forest strategy for generating target early warning information, and target inspection task generation based on equipment operation trend analysis results, this technology overcomes the data silo problem in traditional hydropower station inspections where operating parameters, environmental quantities, and on-site images are scattered across different systems. Multimodal feature vector construction enables the integration and correlation of multi-source data. Furthermore, the long-term time-series prediction strategy replaces the traditional fixed-cycle preventative maintenance or reactive post-fault handling model, preventing situations where early anomalies go undetected when equipment parameters are far below national or factory standard warning values. Target inspection tasks are generated based on equipment operation trend analysis results. Compared to existing technologies, this approach achieves collaborative integration of hydropower station equipment data, accurate early fault detection, effective prediction of operating trends, and dynamic optimization of inspection tasks, driving a shift in operation and maintenance models from reactive handling and fixed-cycle maintenance to predictive maintenance based on equipment status. Attached Figure Description

[0012] The accompanying drawings, which are incorporated in and form part of this specification, illustrate embodiments consistent with this application and, together with the description, serve to explain the principles of this application.

[0013] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, for those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0014] Figure 1 This is a flowchart illustrating an embodiment of the hydropower station inspection method based on multi-source sensing equipment provided in this application. Figure 2 This is a schematic diagram of the data processing flow provided in Embodiment 1 of the hydropower station inspection method based on multi-source sensing equipment in this application; Figure 3 This is a flowchart illustrating Embodiment 2 of the hydropower station inspection method based on multi-source sensing equipment provided in this application; Figure 4 A simplified flowchart illustrating the hydropower station inspection method based on multi-source sensing devices provided in Embodiment 2 of this application; Figure 5 This is a schematic diagram of the module structure of a hydropower station inspection device based on multi-source sensing equipment, as described in an embodiment of this application. Figure 6 This is a schematic diagram of the equipment structure of the hardware operating environment involved in the hydropower station inspection method based on multi-source sensing devices in the embodiments of this application.

[0015] The purpose, features, and advantages of this application will be further explained in conjunction with the embodiments and with reference to the accompanying drawings. Detailed Implementation

[0016] It should be understood that the specific embodiments described herein are merely illustrative of the technical solutions of this application and are not intended to limit this application.

[0017] To better understand the technical solution of this application, a detailed description will be provided below in conjunction with the accompanying drawings and specific implementation methods.

[0018] The main solution of this application embodiment is as follows: acquiring multi-source operational data, including sensor telemetry data, industrial control operational data, and structured vision data; constructing multimodal feature vectors based on the multi-source operational data; performing equipment status prediction and trend analysis based on a long-term time series prediction strategy and the multimodal feature vectors to obtain equipment operation trend analysis results; generating target early warning information based on the multimodal feature vectors and the isolated forest strategy; and generating target inspection tasks based on the equipment operation trend analysis results.

[0019] In this embodiment, for ease of description, the following description will focus on identifying hydropower station inspection equipment based on multi-source sensing devices.

[0020] Because existing technologies for hydropower station inspections cannot accurately identify and predict early, minor, and latent faults in hydropower station equipment in real time, this application provides a solution. This solution employs multi-source data acquisition methods, including sensor telemetry data, industrial control operation data, and structured visual data. It combines multi-modal feature vector construction, long-term time-series prediction strategies for equipment status prediction and trend analysis, isolated forest strategy for generating target early warning information, and target inspection task generation based on equipment operation trend analysis results. Firstly, multi-source data acquisition breaks down the traditional method of scattering operating parameters, environmental quantities, and on-site images in hydropower station inspections. The problem of data silos between different systems is addressed by constructing multimodal feature vectors to integrate and correlate multi-source data. Secondly, long-term time-series prediction strategies replace traditional fixed-cycle preventive maintenance or reactive post-fault handling, avoiding situations where early anomalies go undetected when equipment parameters are far below national or factory warning values. Target inspection tasks are generated based on equipment operation trend analysis results. Compared with existing technologies, this achieves collaborative integration of hydropower station equipment data, accurate detection of early faults, effective prediction of operation trends, and dynamic optimization of inspection tasks, promoting the transformation of operation and maintenance models from reactive handling and fixed-cycle maintenance to predictive maintenance based on equipment status.

[0021] It should be noted that the executing entity in this embodiment can be a computing service device with data processing, network communication, and program execution functions, such as a tablet computer, personal computer, or mobile phone, or an electronic device capable of performing the above functions, such as a hydropower station inspection device based on multi-source sensing devices. The following description uses a hydropower station inspection device based on multi-source sensing devices as an example to illustrate this embodiment and the subsequent embodiments.

[0022] Based on this, the embodiments of this application provide a hydropower station inspection method based on multi-source sensing devices, referring to... Figure 1 , Figure 1 This is a flowchart illustrating the first embodiment of the hydropower station inspection method based on multi-source sensing equipment in this application.

[0023] In this embodiment, the hydropower station inspection method based on multi-source sensing devices includes steps S10 to S50: Step S10: Acquire multi-source operational data, which includes sensor telemetry data, industrial control operational data, and structured vision data; It should be noted that multi-source operational data is a collection of various data obtained from different acquisition channels of the hydropower station to reflect the operating status of the equipment. It covers data directly monitored by sensors, data recorded by the industrial control system, and data converted from image recognition, and can comprehensively cover multi-dimensional information on equipment operation.

[0024] In addition, sensor telemetry data are real-time operating parameter data collected by various sensors deployed on key equipment in hydropower stations, such as temperature data of the thrust bearing of the turbine generator set and vibration data of the main transformer, which are used to directly reflect the physical operating status of the equipment.

[0025] In addition, industrial control operation data refers to the equipment operation parameter data recorded by the hydropower station's industrial control system, which is used to monitor and control the operation of equipment. This data includes equipment current, voltage, operating load, and other data. These data are the core control parameter records for the normal operation of the equipment.

[0026] In addition, structured visual data is digital data with a clear format and meaning, which is generated by recognizing and processing video or image information of equipment at hydropower stations. Examples include readings on equipment dials and the on / off status of indicator lights, which facilitates subsequent data fusion and analysis.

[0027] It is understandable that sensor telemetry data is collected by deploying sensors on key equipment in hydropower stations, such as deploying sensors near the thrust bearing pads of turbine generator units to collect temperature data; recorded industrial control operation data, such as the operating current and voltage data of the equipment, is extracted from the industrial control system; and unstructured image information is transformed into structured visual data by recognizing and processing equipment videos or images acquired by on-site cameras, such as recognizing dial readings to obtain specific values. Through the above methods, comprehensive multi-source operational data reflecting the operating status of hydropower station equipment is acquired, providing basic data support for the subsequent construction of multimodal feature vectors. In this embodiment, for the turbine generator unit, sensor telemetry data such as the temperature and micro-vibration of its thrust bearing are collected, along with industrial control operation data such as the unit's operating load. Structured visual data is also obtained by recognizing relevant dial readings of the unit through cameras.

[0028] In this embodiment, for the deployment of the multi-dimensional sensor network, a surface-mount micro-electro-mechanical system (MEMS) wireless vibration or temperature sensor is used near the thrust or guide bearing pads of the hydro turbine generator set. The sensor is fixed with industrial strong adhesive or magnetic base and is wirelessly transmitted via Bluetooth Low Energy (BLE) or Long Range Wide Area Network (LoRaWAN). The sensor is powered by a built-in micro battery to avoid wiring.

[0029] Fiber Bragg Grating (FBG) temperature sensors are deployed in the stator coils or core, either directly embedded or attached to the surface of key insulation layers. The high voltage resistance of optical fibers is utilized to centrally acquire signals via a fiber optic demodulator, enhancing electromagnetic interference immunity. For the main transformer, thermistors or infrared temperature sensors are fixed on the oil tank wall and the oil inlet and outlet of the radiator by clamps or magnetic bases. Near the bushing, surface-mount piezoelectric ceramic (Lead Zirconate Titanate, PZT) acoustic or ultrasonic partial discharge sensors are installed on the outer surface of the transformer shell to monitor internal partial discharge, without the need for opening the cover or complicated bushing installation.

[0030] For critical circuits in high-voltage switchgear, small-sized wireless temperature rise sensors are used at bus joints and circuit breaker contacts. These sensors are directly bundled or attached to the heat points and powered by electromagnetic induction or radio frequency energy harvesting to achieve maintenance-free and passive operation. In critical areas of the plant, standard industrial-grade temperature, humidity, or gas integrated sensors are fixed to walls or columns and connected to the network via RS485 bus, wireless ZigBee, or LoRa.

[0031] Regarding the improvement of intelligent video acquisition equipment, no new large mobile devices such as robots or drones will be added. Instead, existing fixed or pan-tilt-zoom (PTZ) cameras will be upgraded with edge intelligence. Edge computing modules will be added to fixed cameras, and low-power micro artificial intelligence processors will be integrated into the camera housing or connected to the camera network interface in the form of an Edge AIBox.

[0032] For PTZ cameras, embedded models and preset point calibrations are performed. The preset cruise function of the PTZ camera is used to set key equipment that needs to be read or observed as fixed inspection points. When the embedded or edge computing module reaches each preset point, it triggers image recognition. The edge computing module only pushes the structured data and the cropped key recognition area image through the Message Queuing Telemetry Transport (MQTT) protocol, without transmitting the complete video stream.

[0033] All deployed nanosensors and optical sensors transmit data to a low-cost wireless gateway within the factory area via built-in wireless modules. Smart cameras push structured data to the edge data acquisition server via industrial Ethernet or Wireless Fidelity (Wi-Fi). The Message Queuing Telemetry Transmission (MQTT) IoT protocol is used uniformly to ensure real-time data transmission and lightweight operation.

[0034] Step S20: Construct multimodal feature vectors based on multi-source operational data; It should be noted that a multimodal feature vector is a comprehensive feature vector formed by converting different types of data from multiple sources into feature vectors and then fusing them in a specific way. It integrates the feature information of sensor telemetry data, industrial control operation data, and structured vision data.

[0035] Understandably, the acquired multi-source operational data is processed to transform sensor telemetry data into telemetry time-series feature vectors. For example, features from temperature and vibration data are extracted and vectorized according to time series. Industrial control operation data is transformed into industrial control operation feature vectors, such as extracting current and voltage change features to form vectors. Structured visual data is transformed into visual feature vectors, such as converting dial readings and indicator light status information into vector elements. Weights for each type of feature vector are calculated based on the historical variance and current outlier degree. Then, the feature vectors are concatenated with their corresponding weights through association operations to ultimately construct a multimodal feature vector.

[0036] In one feasible implementation, step S20 may include steps S21 to S24: Step S21: Perform outlier processing on the multi-source running data to obtain processed multi-source running data; It should be noted that processing multi-source operational data is a clean data set obtained after outlier processing of multi-source operational data. It removes abnormal interference items from the original data and retains data that can truly reflect the operating status of the equipment, providing reliable input for subsequent time synchronization and resampling processing.

[0037] Understandably, sensor telemetry data is selected from multi-source operational data, which is prone to outliers due to sensor errors or external interference. The median absolute deviation method is used to calculate the median and median absolute deviation of the sensor telemetry data. Each data point is then judged to meet the anomaly criteria: whether the absolute value of the difference between the data point and the median, divided by the median absolute deviation, is greater than a preset threshold. If the anomaly criteria are met, the data point is marked as an outlier and replaced with a smoothed value from the previous time step or a predicted value obtained through Kalman filtering. After anomaly processing of the sensor telemetry data, it is integrated with non-anomaly industrial control operational data and structured vision data to obtain processed multi-source operational data.

[0038] In this embodiment, the Median Absolute Deviation (MAD) method is applied to high-frequency T-Data for real-time outlier detection and replacement. MAD is more robust to outliers than standard deviation.

[0039] If data points satisfy (in The preset threshold; This represents the i-th specific data point in the high-frequency telemetry data; The median of the high-frequency telemetry dataset X is used as a benchmark reference value to determine whether a data point is abnormal; MAD represents the absolute deviation of the median of the high-frequency telemetry dataset X. If the median is an anomaly, it is marked as an anomaly and the predicted value is replaced by the smoothed value of the previous time step or the Kalman filter.

[0040] Step S22: Perform time synchronization and resampling processing based on the processed multi-source running data and the preset sliding time window to obtain aligned multi-source running data; It should be noted that the preset sliding time window is a set time interval with a fixed duration that can slide along the time axis in fixed steps. Furthermore, aligned multi-source operational data is a data set obtained after processing multi-source operational data through time synchronization and resampling. It achieves the unification of sensor telemetry data, industrial control operational data, and structured vision data in the time dimension.

[0041] Understandably, an adaptive resampling method is used to address the inherent transmission delays and sampling frequency differences between different data sources. These different data sources can include sensor telemetry time-series data (T-Data), industrial control system operation data (C-Data), and structured visual data (V-Data) output by edge computing modules.

[0042] The timestamp of the industrial control system operation data (C-Data) from the SCADA system Based on this benchmark, other data sources are aligned. For low-frequency structured visual data (V-Data), a zero-order hold strategy is used; for high-frequency sensor telemetry time-series data (T-Data), linear interpolation is used, based on a sliding time window. Perform weighted smoothing. The weighted smoothing expression is as follows:

[0043] Where t represents the time point based on the timestamp of the industrial control operation data (C-Data) of the SCADA system; This represents a sliding time window used for weighted smoothing of data from different data sources; Represents a sliding time window Any time step within; This indicates that the high-frequency telemetry data (T-Data) is at time step The original data value at that time; Indicates time step The corresponding weight, its value varies with The weights decrease as the distance from the reference time point t increases, and the sum of the weights of all time steps is 1. This represents the aligned data value of high-frequency telemetry data (T-Data) at the reference time point t after adaptive time alignment.

[0044] Step S23: Perform deep feature quantization on the image information of the aligned multi-source running data to obtain fused multi-source running data; It should be noted that the image information refers to the original image or video frame information corresponding to the structured visual data in the multi-source operational data, including device dial images, indicator light images, etc.

[0045] Furthermore, the fusion of multi-source operational data is a data set obtained by quantizing the deep features of image information of the aligned multi-source operational data.

[0046] Understandably, image information, including images of key areas such as the device dial and indicator lights, is extracted from aligned multi-source operational data. A lightweight, anchor-free object detection model (such as CenterNet) is used to locate the dial and indicator light regions in the images, outputting bounding boxes for key regions to ensure accurate extraction of target images. For the dial image, the cropped digit region is input into a Convolutional Recurrent Neural Network (CRNN). The convolutional part of this network extracts image feature maps, the bidirectional long short-term memory network part performs temporal modeling, and finally, a decoding module trained with a Connectionist Temporal Classification (CTC) loss function identifies the digit readings on the dial.

[0047] Additionally, for indicator light images, a lightweight classification network based on depthwise separable convolution (such as Xception) is used to extract the color and brightness features of the indicator lights, converting the colors (red, green, yellow) and on / off states into structured status codes. These quantized image features are then integrated with sensor telemetry data and industrial control operation data from aligned multi-source operational data to obtain fused multi-source operational data.

[0048] For example, the dial image in the aligned multi-source operation data of the high-voltage switchgear is processed, the dial area is located through the central network, and then the dial reading is obtained as 1.25MPa through the convolutional recurrent neural network; the indicator light image is processed to obtain the "green / bright" status code, and these quantitative features are integrated with the switchgear's temperature, current and other data to form fused multi-source operation data.

[0049] Step S24: Construct a multimodal feature vector based on the fused multi-source running data.

[0050] Understandably, different types of features are extracted from fused multi-source operational data. Sensor telemetry data is extracted according to time series to obtain features such as mean and rate of change, transforming them into telemetry time-series feature vectors. Industrial control operation data is extracted to obtain features such as load fluctuation and parameter stability, transforming them into industrial control operation feature vectors. Quantized image features are transformed into visual feature vectors. Based on historical data for each type of feature vector, the historical sensitivity mean and historical standardized outlier are calculated. Combining the current value of the target feature vector, the historical sensitivity mean, and the historical standardized outlier, anomaly sensitivity weights for each type of target feature vector are calculated. The weight magnitude is positively correlated with the current outlier degree and historical sensitivity of the feature. Each type of target feature vector is associated with its corresponding anomaly sensitivity weight, and the results are concatenated sequentially to obtain a multimodal feature vector.

[0051] In one feasible implementation, step S24 may include: determining a target feature vector based on fused multi-source operational data, the target feature vector including telemetry time-series feature vector, industrial control operation feature vector, and visual feature vector; determining the historical sensitivity mean and historical standardized outlier based on the target feature vector; calculating the anomaly sensitivity weight of the target feature vector based on the target feature vector, the historical sensitivity mean, and the historical standardized outlier; and performing weighted fusion of the target feature vector based on the anomaly sensitivity weight to obtain a multimodal feature vector.

[0052] It should be noted that the target feature vector is a set of vectors extracted from multi-source operational data, representing the core features of different data types.

[0053] Additionally, the telemetry time-series feature vector is a vector formed by extracting features from sensor telemetry data according to a time series; the industrial control operation feature vector is a vector formed by extracting core control parameter features from industrial control operation data; and the visual feature vector is a vector formed by arranging quantified image features such as dial readings and indicator light status codes according to preset rules.

[0054] Furthermore, the historical sensitivity mean is the average sensitivity of various target feature vectors to equipment failure, calculated based on the equipment's historical normal operation data. The higher the sensitivity, the more significant the change in the feature vector before the equipment failure, and the greater its reference value for failure prediction.

[0055] In addition, the historical standardized outlier is a statistical value calculated based on the historical normal operation data of the equipment, which is used to measure the standardization degree of various target feature vectors from the historical mean. It is used as a reference benchmark to measure the deviation of the current value of the feature vector from the normal range.

[0056] It should be understood that the anomaly sensitivity weight is a weight value calculated based on the current value of the target feature vector, the historical sensitivity mean, and the historical standardized outlier. It is used to measure the contribution of various target feature vectors to the abnormal state of the equipment at the current moment. The higher the weight, the greater the influence of that type of feature vector on the judgment of the equipment state.

[0057] Understandably, after obtaining the target feature vectors, the average sensitivity of each type of target feature vector under historical normal operating conditions is calculated to obtain the historical average sensitivity. Simultaneously, the standardized deviation of each type of target feature vector from the historical normal operating conditions is calculated to obtain the historical standardized outlier. Combining the deviation of the current value of the target feature vector from the historical average, the historical average sensitivity, and the historical standardized outlier, the anomaly sensitivity weight of each type of target feature vector is determined through correlation calculation. The greater the current deviation and the higher the historical average sensitivity, the greater the anomaly sensitivity weight. The telemetry time-series feature vector, industrial control operation feature vector, and visual feature vector are respectively correlated with their respective anomaly sensitivity weights. The calculation results are then concatenated in the order of telemetry time-series, industrial control operation, and vision to obtain the multimodal feature vector. The formula for calculating the anomaly sensitivity weight is as follows:

[0058] in, . This represents the weight of the i-th type of feature at time point t (i takes the values ​​T, C, and V, corresponding to telemetry time series features, industrial control operation features, and structured visual features, respectively), and the sum of the weights of all types of features is 1; The output of the anomaly sensitivity function represents the i-th type of feature at time point t, which is used to measure the contribution of this type of feature to the system state. This represents the feature vector of the i-th type of feature at time point t; This represents the three types of features: telemetry timing characteristics (T), industrial control operation characteristics (C), and structured visual characteristics (V). The summation of the results is the denominator of the Softmax normalization, which is used to ensure that the sum of the weights of various features is 1. This represents the anomaly sensitivity function, used to calculate the contribution of various characteristics to the system state. The expression for the anomaly sensitivity function is as follows:

[0059] in, The output of the anomaly sensitivity function represents the i-th type of feature at time t; The sensitivity coefficient represents the characteristic of the i-th type (e.g., MEMS vibration data are more sensitive to faults). Visual data is relatively insensitive. ); This represents the fused feature value of the i-th type of feature at time t; This represents the historical mean of the i-th feature (the statistical mean under normal operating conditions in the past). The historical standard deviation of the i-th feature (statistical standard deviation under normal operating conditions in the past); This represents the standardized outlier of the current value relative to the historical normal distribution (i.e., the absolute value of the Z-score). It is an exponential function used to amplify the sensitivity differences between different features, ensuring that the weights have a more obvious distinguishing effect.

[0060] Final multimodal feature vector By constructing a weighted concatenation of each feature vector, an adaptive emphasis on the real-time importance of each data source is achieved.

[0061] Where t represents the current time point; This represents the processed and aligned telemetry time series feature vector at time point t; This represents the processed and aligned feature vector of industrial control operation at time point t. This represents the structured visual feature vector after processing and alignment at time point t; This represents the weight of the i-th type of feature at time point t (i takes the values ​​T, C, and V, corresponding to telemetry time series features, industrial control operation features, and structured visual features, respectively), and the sum of the weights of all types of features is 1; This represents the multimodal feature vector after adaptive multimodal collaborative feature fusion at time point t; The concatenation operator for eigenvectors.

[0062] Reference Figure 2 , Figure 2 This is a schematic diagram of the data processing flow of the first embodiment of the hydropower station inspection method based on multi-source sensing equipment in this application.

[0063] like Figure 2As shown, the system is mainly divided into a low-cost data acquisition layer, an edge preprocessing layer, and a standardized output layer. The low-cost data acquisition layer includes three parts: wireless sensors, edge cameras, and edge cameras. The wireless sensors are surface-mounted and transmit via Bluetooth Low Energy (BLE) or LoRa (Long Range Radio), offering a temperature accuracy of ±0.1℃ and eliminating the need for wiring. The edge cameras have a recognition accuracy greater than 92%. These devices, after data cleaning, enter the edge preprocessing layer. In the edge preprocessing layer, data cleaning is performed, including outlier detection and Kalman filtering, with a processing latency of less than 100ms. Simultaneously, video structuring is performed, identifying digital meters and classifying indicator lights, outputting structured data combining readings and light colors. The cleaned data is then output to the standardized output layer. In the standardized output layer, the protocol uses lightweight Message Queuing Telemetry Transport (MQTT), with data formatted as JSON object markup language, including device ID, timestamp, and monitored values. The output destination is from the factory's wireless gateway to the data acquisition server.

[0064] Step S30: Based on the long-term time series prediction strategy and multimodal feature vectors, perform equipment status prediction and trend analysis to obtain equipment operation trend analysis results; It should be noted that the long-term time series prediction strategy is a method that uses specific algorithms to predict and analyze the operating status of equipment over a period of time based on the time series characteristics of the data. This strategy can capture the pattern of data changes over time, thereby inferring the development trend of the equipment's operating status. In this embodiment, a long-term time series prediction model based on probabilistic sparse self-attention is adopted.

[0065] Furthermore, the equipment operation trend analysis results are information about the future operating status of the equipment obtained through equipment status prediction and trend analysis, including the predicted value of the remaining service life of the equipment, the prediction confidence interval, and the probability of failure transfer, which can clearly reflect the development direction of the equipment operating status.

[0066] Understandably, a long-term time-series prediction model based on probabilistic sparse self-attention is employed. This model uses multimodal feature vectors containing a predetermined number of historical time steps to form a feature tensor. A predetermined periodic positional encoding is injected into this feature tensor to obtain the target feature tensor. The information dispersion of each query vector in the target feature tensor is calculated. Query vectors with information dispersion meeting predetermined conditions are selected and their corresponding key vectors are used for attention calculation. The attention calculation results are input into subsequent network layers of the model to obtain initial prediction results. A quantile regression loss strategy is used to calculate the target difference value of the initial prediction results. The model parameters are optimized based on the target difference value, and the predicted remaining service life of the equipment and its prediction confidence interval are output. The multimodal feature vectors are input into a predetermined variational autoencoder to obtain the target latent distribution. The Wasserstein distance between this distribution and the standard normal distribution is calculated to determine the equipment health index. The equipment health index and the multimodal feature vectors are then input into a deep survival model to obtain the failure probability and failure time prediction values. The instantaneous failure rate is calculated based on the failure probability, and then the failure transfer probability is obtained. The failure transfer probability is added to the previous prediction results to finally obtain the equipment operation trend analysis results.

[0067] In one feasible implementation, step S30 may include steps S31 to S34: Step S31: Determine the target feature tensor based on the long-term time series prediction strategy, the preset periodic position encoding, and the multimodal feature vector; It should be noted that the preset periodic position encoding is a manually set encoding information used to inject feature tensors to enhance the model's ability to perceive the periodicity and trend of time series.

[0068] Furthermore, the target feature tensor is a tensor-form data formed by integrating multimodal feature vectors along the time dimension and injecting them with preset periodic position encoding.

[0069] Understandably, multimodal feature vectors containing a preset number of historical time steps are selected and arranged chronologically to form an initial feature tensor. Based on the periodicity of equipment operation, such as the daily load fluctuation cycle of a hydroelectric generator unit, a preset periodic position code is generated. This preset periodic position code is injected into the initial feature tensor. By associating and integrating the encoded information with the feature vectors of the corresponding time steps in the initial feature tensor, it is ensured that the features of each time step carry time and position information, ultimately yielding the target feature tensor. In this embodiment, it can be determined that the data includes historical data. Fusion feature tensor at each time step ,in This represents the multimodal feature vector after adaptive multimodal collaborative feature fusion at time point t. Periodic positional encoding (PE) is injected to enhance the model's ability to perceive the periodicity and trend of time series data. The target feature tensor expression is as follows:

[0070] Wherein, PE represents periodic location encoding, which is used to enhance the model's ability to perceive the periodicity and trend of time series. This represents the target feature tensor expression.

[0071] Step S32: Calculate the information dispersion of the target feature tensor and determine the target dispersion vector based on the information dispersion. It should be noted that information dispersion is an indicator that measures the uniqueness of information in each query vector within the target feature tensor. It is calculated by comparing the distance difference between a single query vector and all other query vectors. The greater the distance difference, the more significant the difference between the device operation information contained in that query vector and other vectors, indicating a higher information dispersion and greater reference value for subsequent device status prediction.

[0072] Furthermore, the target dispersion vector is a set of high-value vectors selected from all query vectors of the target feature tensor. During the selection process, preset conditions for information dispersion are set, and these query vectors that meet the conditions are arranged in their original chronological order, thus centrally retaining the time-series information that is crucial for predicting the device status.

[0073] Understandably, all query vectors are extracted from the target feature tensor. Each query vector corresponds to a comprehensive operational feature of the device at a given time step within the target feature tensor, encompassing key information such as temperature, vibration, and current. The distance between each query vector and all other query vectors is calculated, and the information dispersion of each query vector is calculated based on the distance differences. Query vectors with high information dispersion are selected according to preset filtering criteria and arranged in their original time order to obtain the target dispersion vector.

[0074] Step S33: Calculate the attention based on the target dispersion vector and the key vector to obtain the target attention, and determine the initial prediction result based on the target attention; It should be noted that the key vector is a vector extracted from the target feature tensor that corresponds one-to-one with the query vector. Both the key vector and the query vector originate from the same target feature tensor and together carry operational feature information from different time steps of the device. In attention calculation, the key vector serves as the benchmark for information matching, working in conjunction with the query vector to uncover the correlations between features at different time steps.

[0075] It should be understood that attention calculation is a method to measure the degree of association between query vectors and key vectors. It assigns weights to different key vectors by calculating their similarity. Key vectors with higher similarity have greater weights, thereby highlighting information that is more important for predicting the current device state, while suppressing interference from irrelevant or secondary information and improving the targeting of the calculation.

[0076] Furthermore, target attention is the output of attention calculation, presented in the form of a weight matrix, where each element represents the association weight between the corresponding query vector and key vector.

[0077] Furthermore, the initial prediction result is the preliminary prediction information obtained after inputting the target attention into the subsequent network layers of the long-term time-series prediction model and undergoing operations such as linear transformation and activation function processing. It contains basic prediction data about the future operating state of the device, such as preliminary values ​​for the device's remaining lifespan, but it has not yet undergone error optimization, and its accuracy needs further improvement. The attention calculation formula is as follows:

[0078] in, represents the query matrix after sparsification, containing only the query vectors with the highest information dispersion (i.e., those key time steps with the greatest distance difference from other queries); Q represents the query matrix. Represents the query vector Information dispersion; K represents the key matrix; V represents the value matrix; This indicates the dimensions of the query vector and the key vector. It is the scaling factor; This represents attention, specifically the result of probabilistic sparse self-attention calculation. Step S34: Calculate the target difference value of the initial prediction results according to the quantile regression loss strategy, and perform equipment status prediction and trend analysis based on the target difference value to obtain the equipment operation trend analysis results.

[0079] It should be noted that the quantile regression loss strategy is a method used to calculate the difference between predicted and actual values. It calculates the prediction bias at different quantiles by setting multiple quantiles. This strategy can simultaneously provide the confidence interval of the prediction results, and compared to traditional loss calculation methods, it more comprehensively reflects the uncertainty and error distribution of the prediction results, making it suitable for the risk assessment needs of hydropower station equipment failure prediction.

[0080] Furthermore, the target difference value is the total difference between the initial prediction and the actual value calculated using the quantile regression loss strategy, obtained by summing the prediction biases at different quantiles. It quantifies the overall error level of the initial prediction results, providing a clear error basis for subsequent optimization of long-term time series prediction model parameters and improvement of prediction accuracy.

[0081] Understandably, the preset quantiles and target quantiles are determined based on the quantile regression loss strategy. Preset quantiles typically include low, middle, and high quantiles, while the target quantile is often set as the middle quantile as a benchmark. The deviations between the predicted quantiles and the actual values ​​corresponding to each preset quantile in the initial prediction results are calculated, and these deviations are summarized to obtain the target difference value. Using the target difference value as the optimization objective, the network weights and biases of the long-term time-series prediction model are adjusted. After the model converges, the output includes equipment remaining lifespan and confidence intervals, providing an analysis of the equipment's operating trend.

[0082] In one feasible implementation, step S34, "calculating the target difference value of the initial prediction result according to the quantile regression loss strategy," may include: determining a preset quantile and a target quantile according to the quantile regression loss strategy; determining whether the predicted quantile of the initial prediction result is greater than the target quantile; when the predicted quantile is less than or equal to the target quantile, calculating the target difference value according to the preset quantile, the predicted quantile, and the target quantile; and when the predicted quantile is greater than the target quantile, calculating the target difference value according to the preset quantile and the predicted quantile.

[0083] It should be noted that the preset quantiles are quantile values ​​set in advance in the quantile regression loss strategy to cover different confidence levels of the prediction results. Multiple quantiles are typically set.

[0084] Furthermore, the target quantile is a benchmark quantile used to determine the relationship between the predicted quantile and the true value. Choosing the median as the target quantile can evenly distinguish between the predicted quantile and the true value, ensuring that prediction deviations in different directions can be accurately calculated and avoiding biased error assessment.

[0085] Furthermore, the predicted quantile is the specific predicted value of the preset quantile in the initial prediction result, and each preset quantile has a corresponding predicted quantile.

[0086] In this implementation example, quantile regression loss (Pinball Loss) can be used for calculation. The formula for calculating the target difference value is as follows:

[0087] in, Indicates quantiles; This represents the pinball loss, used to quantify the difference between the model's predicted pinballs and the actual values. Y represents the set of predicted values ​​corresponding to the quantiles predicted by the model; i represents the set of true values; q represents the i-th sample in the dataset; This represents the true value of the i-th sample; This represents the predicted value of the q-th quantile in the i-th sample; This represents the target difference value, which is the absolute error between the true value of the i-th sample and the predicted value of the corresponding quantile.

[0088] In one feasible implementation, step S34 may further include steps S35 to S39: Step S35: Input the multimodal feature vector into a preset variational autoencoder to obtain the target latent distribution; It should be noted that a pre-trained variational autoencoder is a type of deep learning model, belonging to the category of generative models, and consists of two parts: an encoder and a decoder. It can compress high-dimensional input data into a low-dimensional space while learning the probability distribution characteristics of the data, rather than simply performing dimensionality reduction. This allows it to be used to uncover potential, difficult-to-observe patterns in the data.

[0089] Furthermore, the target latent distribution is the probability distribution formed in the low-dimensional latent space after the encoder of the pre-defined variational autoencoder processes the multimodal feature vectors. This distribution can represent the core information of the multimodal feature vectors in probabilistic form, reflecting the potential characteristics of the equipment's operating state and providing a foundation for subsequent calculations of equipment health indicators. The target latent distribution can be understood as follows:

[0090] In the formula, This represents the total loss function of the variational autoencoder (VAE). This represents the reconstruction loss of the variational autoencoder, used to measure the difference between the model-reconstructed data and the original data; The weighting coefficients of the KL divergence loss are used to balance the contributions of the reconstruction loss and the KL divergence loss. This represents the KL divergence loss, used to measure the encoder distribution. With prior distribution Differences; This represents the probability distribution of the encoder in a variational autoencoder, used to map the input feature X to the latent space z; The prior distribution of the latent space z in the variational autoencoder is usually denoted as the standard normal distribution. .

[0091] Step S36: Calculate the Wasserstein distance between the target potential distribution and the standard normal distribution, and determine the equipment health index based on the Wasserstein distance; It's important to note that Wasserstein distance is a metric used to measure the similarity between two probability distributions, also known as Earth movement distance. It quantifies the difference between two probability distributions by calculating the minimum amount of work required to move one distribution to the other. Compared to other distance metrics, it is more sensitive to the shape and overlapping areas of the distributions, and thus more accurately reflects distributional differences.

[0092] Furthermore, the equipment health index is a numerical value used to quantify the current health status of the equipment, derived from the Wasserstein distance. The smaller the Wasserstein distance, the closer the target potential distribution is to the standard normal distribution, the higher the equipment health index, and the closer the equipment's operating status is to normal; conversely, the larger the Wasserstein distance, the worse the equipment health status. The Wasserstein distance calculation formula is as follows:

[0093] In the formula, This is the latent distribution of the encoder output. This metric reflects the degree to which the real-time feature distribution deviates from the normal state distribution. This represents the health indicators of the device at time point t, used to quantify the degree to which the device deviates from its normal state. This represents the multimodal fusion feature vector at time point t; This represents the encoder's response to input features in a variational autoencoder (VAE). The output potential distribution; Represents the latent distribution of encoder output Compared with the standard normal distribution The Wasserstein distance is used to measure the difference in distribution between the two. This represents the preset maximum Wasserstein distance, used for normalizing health indicators. It represents the standard normal distribution and is the prior distribution of the VAE latent space; This represents the center point of the normal state in the potential space z (obtained from the potential distribution statistics of historical normal data).

[0094] Step S37: Input the equipment health indicators and multimodal feature vectors into the deep survival model to obtain the failure probability and failure time prediction values. The deep survival model is used to represent the relationship between the equipment health indicators, failure probability and failure time prediction values. It should be noted that deep survival models are models that integrate deep learning and survival analysis theory, capable of handling both high-dimensional features and time-related survival data. They achieve quantitative prediction of equipment failure risk by learning the correlation between equipment health indicators, multimodal feature vectors, and failure occurrence time and probability, rather than simply performing classification or regression.

[0095] Furthermore, the failure probability is the probability value output by the deep survival model of the device failing within a preset time window in the future. The failure time prediction value is the predicted time length from the current moment to the expected failure moment output by the deep survival model.

[0096] Understandably, equipment health indicators and multimodal feature vectors are integrated to ensure their formats conform to the input requirements of the deep survival model. The integrated input data is then fed into a pre-trained deep survival model, which has been trained using historical health data, multimodal feature data, and fault record data from hydropower station equipment. This model accurately captures the correlation between equipment status and faults. After the model runs, it outputs two core results: the probability of equipment failure within multiple preset future time windows, and the predicted failure time from the current moment to the expected failure time.

[0097] Step S38: Calculate the instantaneous failure rate based on the failure probability and the predicted failure time. It should be noted that the instantaneous failure rate refers to the probability density of a device failing within a very short period after a certain point in its normal operation. It dynamically reflects the changes in the failure risk of the device at different times, rather than a fixed failure probability.

[0098] Understandably, based on the failure probabilities and corresponding failure time predictions for multiple future time windows, and according to a pre-defined instantaneous failure rate calculation method, the instantaneous failure rate is calculated using the difference in failure probabilities and time intervals between adjacent time windows, and corrected by the failure time predictions. The instantaneous failure rate is then obtained at different times. The formula for calculating the instantaneous failure rate is as follows:

[0099] in, Indicates in features Below, the equipment at the time point The instantaneous failure rate, i.e., the time it takes for the equipment to survive. The conditional probability of a failure occurring at that point in time; Indicates the time point at which the device survives. And the characteristics are Under certain conditions, the conditional probability of equipment failure is equivalent to the instantaneous failure rate. ; Indicates in features Below, the equipment at the time point The probability of failure; Indicates in features Below, the equipment at the time point The sum of probabilities of failure occurring at or after a certain point in time, i.e., the probability of the equipment failing at that point in time. and the probability that no failure (survival) will occur afterwards. The complement (because) ); The k-th discrete time point is used to divide the time interval for fault probability; K represents the total number of discrete time points, covering all time nodes where fault probability needs to be evaluated. This represents the feature vector input to the DeepHit model, which typically includes health indicators. and multimodal fusion features .

[0100] Step S39: Calculate the fault transfer probability based on the instantaneous failure rate, and update the equipment operation trend analysis results based on the fault transfer probability.

[0101] It should be noted that failover probability refers to the probability that equipment will transition from its current operating state to a faulty state. It is calculated based on the instantaneous failure rate and is typically for a predetermined future time period. It quantifies the likelihood that equipment will deteriorate and fail within a specific timeframe, and is a key indicator for assessing the short-term failure risk of equipment.

[0102] Understandably, based on the instantaneous failure rate and the length of a preset future time period, the fault transfer probability is calculated to obtain the specific probability value of the equipment transitioning from a normal state to a fault state within that time period. The formula for calculating the fault transfer probability is as follows:

[0103] In the formula, : Indicates the device in a future time window The probability of a state transition (i.e., a failure) occurring within the system. : Indicates the device in a future time window The probability of internal survival (without failure). t: represents the current time point. : Indicates the end time of the future time window, that is, starting from the current time t and continuing for... The point in time after the duration. : indicates in features Below, the equipment at the time point Instantaneous failure rate. : Indicates the time from the current time t to the end of the time window. Inside, the equipment at each point in time The product of the probabilities of no failures, i.e., the survival probability. .

[0104] Furthermore, the failure transfer probability is added to the output equipment operation trend analysis results. At the same time, the original remaining service life prediction value and confidence interval are adjusted according to the magnitude of the failure transfer probability. For example, if the failure transfer probability is high, the remaining service life prediction value is appropriately shortened and the lower limit of the confidence interval is expanded, ultimately forming an updated and more accurate equipment operation trend analysis result.

[0105] Step S40: Generate target early warning information based on multimodal feature vectors and the isolated forest strategy; It should be noted that the isolation forest strategy is an algorithmic strategy used to detect data anomalies. By randomly selecting features and randomly partitioning values, abnormal data is quickly isolated at a small tree depth, thereby identifying anomalies in the data. In this embodiment, a spatiotemporal context adaptive isolation forest model strategy is adopted.

[0106] Furthermore, the target early warning information is based on multimodal feature vector and isolated forest strategy analysis to indicate that the device may be abnormal. It includes the final anomaly score and the early warning threshold. When the final anomaly score exceeds the early warning threshold, it indicates that the device may have an abnormal situation that needs attention.

[0107] Understandably, a spatiotemporally adaptive isolated forest model is constructed. Multimodal feature vectors are input into this model, and the initial anomaly score is calculated through random feature selection and random partitioning. The similarity between the multimodal feature vectors of equipment in the same unit or region and the current multimodal feature vector is calculated, and spatial weights are determined based on this similarity. Multimodal feature vectors under similar historical operating conditions are matched, and their differences from the current multimodal feature vector are calculated, with temporal weights determined based on these differences. The difference between the initial anomaly score and the spatial weights is then summed with the temporal weights to obtain the final anomaly score. Subsequently, a warning threshold is determined based on the historical distribution of final anomaly scores and a preset false alarm rate. When the final anomaly score exceeds the warning threshold, a target warning message containing both the final anomaly score and the warning threshold is generated.

[0108] In one feasible implementation, step S40 may include steps S41 to S45: Step S41: Calculate the original anomaly score based on the multimodal feature vector and the isolated forest strategy; It should be understood that the raw anomaly score is a value directly output by the isolated forest model after analyzing the multimodal feature vectors, used to initially determine the degree of anomaly in the equipment's operating data. The higher the value, the greater the deviation of the multimodal feature vectors from the normal data distribution, and the higher the probability of equipment anomalies. However, this score does not consider the influence of spatial and temporal dimensions, and the accuracy of the judgment needs further correction.

[0109] Understandably, a pre-trained isolation forest model is invoked. This model has been trained and converged using multimodal feature vector data from hydropower station equipment operating under normal conditions, and can accurately distinguish between normal and abnormal data patterns. The multimodal feature vectors are input into the isolation forest model, which randomly partitions the feature vectors using multiple isolation trees, calculating the path length of the feature vectors in each tree. Based on the average path length of all isolation trees, a raw anomaly score is obtained, completing the initial anomaly judgment of the equipment's current operating state.

[0110] Step S42: Calculate the similarity based on the multimodal feature vectors, and determine the spatial weights based on the similarity. It should be noted that the spatial weight is a weight value determined based on the degree of similarity and used to correct the original anomaly score. The higher the degree of similarity, the more consistent the current device's operating status is with the surrounding normal devices, and the larger the spatial weight is, which is used to reduce the risk of misjudgment caused by the data fluctuation of a single device in the original anomaly score; conversely, the smaller the spatial weight is, the more sensitive the original anomaly score is to potential anomalies.

[0111] It should be understood that similarity is an indicator that measures the similarity between the multimodal feature vectors of other normally operating equipment in the same unit or region and the multimodal feature vector of the current equipment. It can be obtained by calculating the cosine similarity between the two types of feature vectors.

[0112] Understandably, the cosine similarity between the current device's multimodal feature vector and the feature vector of each normal device is calculated, and the average of all similarities is taken as the final similarity score. The similarity score is then converted into spatial weights according to a predefined mapping rule.

[0113] Step S43: Calculate the degree of difference based on the multimodal feature vectors, and determine the time weight based on the degree of difference; It should be understood that the degree of difference is an indicator that measures the difference between the current multimodal feature vector of the equipment and the multimodal feature vector of the equipment under similar historical operating conditions. By calculating the Euclidean distance between the two types of feature vectors, the larger the distance value, the more significant the deviation between the current operating state of the equipment and the state under similar historical operating conditions; the smaller the distance value, the closer the current state is to the state under similar historical operating conditions, and the more it can reflect the time dimension change of the equipment's operating state.

[0114] It should be noted that the time weight is a weight value determined based on the degree of difference and used to correct the original anomaly score. The higher the degree of difference, the greater the deviation between the current equipment operating state and similar historical operating conditions, and the greater the time weight, which is used to improve the sensitivity of the original anomaly score to potential time-dimensional anomalies; conversely, the smaller the time weight, the lower the misjudgment of anomalies caused by fluctuations in normal operating conditions.

[0115] Understandably, the Euclidean distance between the current device feature vector and the historical feature vector is calculated to obtain the degree of difference, and the degree of difference is converted into time weights according to preset rules.

[0116] Step S44: Calculate the target anomaly score based on the original anomaly score, spatial weight, and temporal weight; It should be noted that the target anomaly score is the final anomaly quantification value obtained by combining the original anomaly scores with spatial and temporal weights. The formula for calculating the target anomaly score is as follows:

[0117] in, This represents the final anomaly score at time point t after adjustment by spatiotemporal context weights; This represents the original anomaly score calculated based on the basic isolated forest model at time point t; The spatial weight is determined based on the similarity of the current operating status of equipment in the same unit or the same area (obtained through the cosine similarity of multimodal fusion feature vectors) and is used to correct the original anomaly score. The time weight is determined based on the differences in features between the current time and similar historical operating conditions (through SCADA data matching), and is used to correct the original anomaly score.

[0118] Step S45: Determine the warning threshold based on the historical anomaly distribution and the preset false alarm rate, and generate target warning information based on the target anomaly score and the warning threshold.

[0119] It should be noted that the historical anomaly distribution is a distribution pattern formed by statistically analyzing all target anomaly scores of the equipment over a past period. It presents the range and probability density of target anomaly scores under normal and abnormal operating conditions.

[0120] Understandably, the statistical equipment calculates historical anomaly distributions by plotting frequency distribution histograms or probability density curves for all target anomaly scores over a past period. Based on a preset false alarm rate, the corresponding quantile value in this historical anomaly distribution is used as a warning threshold. The current target anomaly score is compared to the warning threshold. If the target anomaly score exceeds the warning threshold, a target warning message containing both the target anomaly score and the warning threshold is generated; otherwise, no warning is generated.

[0121] Step S50: Generate target inspection tasks based on the equipment operation trend analysis results.

[0122] It should be noted that the target inspection task is generated based on the results of equipment operation trend analysis and is used to guide inspection personnel in carrying out equipment inspections. It includes inspection priorities and optimized inspection routes, enabling inspection personnel to carry out inspection work in a targeted manner.

[0123] Understandably, the process involves obtaining the target anomaly score, combining it with the predicted remaining service life and fault transfer probability from the equipment operation trend analysis results, assigning weights to these three indicators, and performing correlation calculations to obtain the equipment risk index. The inspection priority for different equipment is determined based on the equipment risk index, with higher risk indices resulting in higher inspection priority. An improved ant colony strategy is employed, treating the equipment points to be inspected as nodes in a graph theory framework. The pheromone concentration and heuristic factor of the ant colony algorithm are initialized, where the heuristic factor is related to the distance between two nodes and the corresponding inspection priority. Each ant selects the next node to be inspected based on the current node's pheromone concentration and heuristic factor, records the inspection path, and calculates the total inspection time, total inspection priority coverage, and personnel load distribution difference for that path. Each inspection path is evaluated using a path evaluation function; the optimal path receives pheromone enhancement, while other paths undergo pheromone evaporation. This process is repeated iteratively a preset number of times until a convergent optimal inspection path, i.e., an optimized inspection route, is obtained. The target inspection task is then generated by combining the inspection priority and the optimized inspection route.

[0124] This embodiment provides a hydropower station inspection method based on multi-source sensing devices. By constructing multimodal feature vectors, applying long-term time-series prediction strategies, using isolated forest strategies to generate target early warning information, and generating target inspection tasks based on equipment operation trend analysis results, it solves the technical problems of incomplete equipment operation status monitoring, inaccurate fault prediction, and low inspection efficiency in existing technologies. It achieves the beneficial effects of improving the comprehensiveness of equipment operation status monitoring, the accuracy of fault prediction, and the efficiency of inspection task execution. In turn, it realizes accurate real-time monitoring and trend prediction of hydropower station equipment, optimizes the dynamic adjustment of inspection tasks, and improves the intelligence level of equipment maintenance and risk prevention capabilities.

[0125] Based on the first embodiment of this application, in the second embodiment of this application, the content that is the same as or similar to that in the first embodiment described above can be referred to the above description, and will not be repeated hereafter. Based on this, please refer to... Figure 3 The steps S50 of the hydropower station inspection method based on multi-source sensing devices include steps S51 to S53: Step S51: Obtain the target anomaly score, and based on the equipment operation trend analysis results, obtain the predicted value of the remaining service life of the equipment and the failure transfer probability. It should be noted that the predicted remaining service life of the equipment is one of the core data in the equipment operation trend analysis results. It refers to the expected length of time the equipment can continue to operate normally from the current moment. This predicted value is derived from equipment health indicators, multimodal feature vectors, and historical fault data. It can intuitively reflect the remaining available time of the equipment and is an important reference for judging the urgency of inspections.

[0126] Additionally, failover probability is a key probability indicator in equipment operation trend analysis results, referring to the probability that equipment will transition from its current normal operating state to a fault state within a preset time period. A higher probability value indicates a greater risk of equipment failure in the short term, and therefore requires priority inclusion in routine inspections.

[0127] Understandably, the calculated target anomaly score obtained from the initial anomaly detection phase has already been corrected for spatial and temporal dimensions, accurately reflecting the current anomaly status of the equipment. The equipment operation trend analysis results output from the initial equipment status prediction and trend analysis phase are retrieved, from which two core data points are extracted: first, the predicted remaining service life of the equipment, clarifying the expected time the equipment can continue operating normally; and second, the fault transfer probability, determining the likelihood of the equipment failing within a preset future timeframe.

[0128] Step S52: Calculate the equipment risk index based on the predicted remaining service life of the equipment, the failure transfer probability, and the target anomaly score, and determine the inspection priority based on the equipment risk index. It can be noted that the equipment risk index is a quantitative risk indicator calculated by comprehensively considering the predicted remaining useful life of the equipment, the probability of failure transfer, and the target anomaly score.

[0129] In this embodiment, the risk indicators output by the prediction model are transformed into quantified inspection priorities, realizing the shift from fixed cycles to prediction-driven approaches. The equipment risk index is calculated as follows:

[0130] in, This represents the equipment risk index of the equipment at time point t, used to measure the risk level of the equipment. This represents the weighting coefficient of the DeepHit state transition probability in the overall risk index; This represents the device failure state transition probability output by the DeepHit model; This represents the weighting coefficient of the final anomaly score in the overall risk index; This represents the final anomaly score at time point t; This represents the weighting coefficient of the reciprocal of the remaining useful life in the comprehensive risk index; This represents the remaining lifespan of the device as predicted by the Informer model. It is the reciprocal of the remaining useful life (the higher the risk, the larger the value); This represents the weighting coefficient of the uncertainty of the forecast result in the comprehensive risk index; This indicates the uncertainty of the prediction results output by the Informer model (measured by the width of the confidence interval output by PinballLoss).

[0131] It should be understood that task priority is directly proportional to the comprehensive risk index. When the comprehensive risk index exceeds the preset attention threshold, relevant target warning information is automatically generated.

[0132] Furthermore, the inspection priority levels are divided according to the numerical range of the equipment risk index. For example, a risk index of 0.8~1.0 is the first priority (emergency inspection), 0.5~0.8 is the second priority (priority inspection), 0.2~0.5 is the third priority (routine inspection), and 0~0.2 is the fourth priority (delayed inspection), thus completing the determination of the inspection priority.

[0133] Step S53: Determine the optimized inspection route based on the improved ant colony strategy, and generate the target inspection task by combining the inspection priority and the optimized inspection route.

[0134] It should be noted that the improved ant colony strategy is a path planning strategy optimized based on the traditional ant colony algorithm. By simulating the pheromone transmission mechanism when ants search for food, and adjusting the pheromone update rules and heuristic factors according to the characteristics of the inspection scenario, it can quickly plan the path with the shortest total inspection time, the most comprehensive inspection priority coverage, and the most balanced personnel load among multiple inspection nodes.

[0135] Understandably, this involves defining the location coordinates of all equipment to be inspected within the hydropower station, the travel paths and distances between each piece of equipment, and the estimated inspection time for each piece of equipment. This data is then input into a computational model to improve the ant colony strategy. The model initializes the ant colony size, initial pheromone concentration, and heuristic factor. Through iterative path searching and pheromone concentration updates by the ants, after a preset number of iterations, it outputs the converged optimal path, i.e., the optimized inspection route. The optimized inspection route clearly marks the inspection priority of each piece of equipment to be inspected, supplements the key inspection items for each piece of equipment, and forms a target inspection task that includes the inspection sequence, priority, key inspection points, and estimated time.

[0136] In one feasible implementation, step S53, "determining the optimized inspection route based on the improved ant colony strategy," may include steps S531 to S535: Step S531: Obtain the initial node to be inspected, and determine the pheromone concentration and heuristic factor according to the improved ant colony strategy. The heuristic factor is related to the distance between the initial node to be inspected and the inspection priority corresponding to the initial node to be inspected. It should be noted that the initial inspection nodes refer to the set of locations of all equipment within the hydropower station that needs to be inspected, with each node corresponding to one piece of equipment to be inspected. These nodes are the basic units for planning the inspection route, and their location information, corresponding equipment numbers, and inspection priorities must all be clearly defined in advance to ensure coverage of all risky equipment that needs to be inspected.

[0137] Furthermore, pheromone concentration is a key parameter used to guide ants (the main agents of path searching) in improving ant colony strategies. Initially, the pheromone concentration of all paths between the nodes to be inspected is set to the same baseline value; as the search iterates, the better-performing paths accumulate higher pheromone concentrations, attracting more ants to choose them.

[0138] Additionally, the heuristic factor is a parameter used in ant colony optimization strategies to assist ants in selecting paths. Its value is directly related to the distance between the initial nodes to be inspected and the inspection priority of the nodes. The closer the nodes are, the larger the heuristic factor; the higher the priority of the nodes, the larger the heuristic factor, which guides ants to prioritize short-distance, high-priority paths.

[0139] Understandably, the location information, equipment number, and corresponding inspection priority of all equipment to be inspected are extracted from the hydropower station equipment management system. Each piece of equipment is considered an initial node to be inspected, forming an initial set of nodes to be inspected. Based on an improved ant colony strategy, an initial pheromone concentration is set for the paths between all initial nodes to be inspected. Heuristic factors are calculated according to a preset formula: first, the physical distance between any two initial nodes to be inspected is calculated; the smaller the distance, the larger the basic heuristic factor. Then, the priority level of the higher-priority node is considered; the higher the priority, the larger the additional heuristic factor value. Finally, the heuristic factors for the paths between each node are obtained.

[0140] Step S532: Select intermediate nodes to be inspected based on the initial nodes to be inspected, pheromone concentration, and heuristic factor. It should be noted that intermediate inspection nodes refer to the inspection nodes selected by the ant during its movement from the initial inspection node to the target initial inspection node.

[0141] Understandably, an initial node to be inspected for an ant can be set. This initial node can be randomly selected or set as the starting point for the inspector. Based on the pheromone concentration and heuristic factor of the paths between the current initial nodes to be inspected, the probability of the ant moving from the current node to other unvisited initial nodes to be inspected is calculated using a preset probability formula: paths with higher pheromone concentrations and larger heuristic factors have a higher probability of being selected. Based on the calculated probabilities, the unvisited node with the highest probability is selected as the intermediate node to be inspected, and this node is marked as visited to avoid duplicate selection.

[0142] Step S533: Determine the initial inspection path based on the intermediate nodes to be inspected, and calculate the total inspection time, inspection priority coverage, and personnel load distribution difference value of the initial inspection path. It should be noted that the total inspection time refers to the total time required for inspection personnel to perform inspection tasks according to the initial inspection path, including the time spent moving between nodes, which can be calculated based on the distance between nodes and the speed of personnel movement, and the inspection operation time at each node, which can be preset according to the equipment type.

[0143] Additionally, the inspection priority coverage sum is the sum of the quantified inspection priority values ​​corresponding to all nodes to be inspected in the initial inspection path. Inspection priorities are divided into different levels according to the equipment risk index and quantified into numerical values. The sum of the quantified values ​​of all nodes in the path is the inspection priority coverage sum. The higher the value, the higher the proportion of high-priority equipment in the path.

[0144] The personnel load distribution difference value is an indicator that measures the degree of task load balance among multiple inspection personnel after the initial inspection path is split. The calculation first divides the initial inspection path into a corresponding number of sub-paths based on the geographical distribution of the nodes to be inspected and the number of personnel. Then, the total inspection time for each sub-path is calculated. The difference between the total time of the longest sub-path and the total time of the shortest sub-path is used as the personnel load distribution difference value; the smaller the difference value, the more balanced the personnel load distribution.

[0145] Understandably, the starting node to be inspected is determined, and the starting node and all intermediate nodes to be inspected are sequentially connected according to the order in which the ants visit the intermediate nodes to be inspected, forming an initial inspection path that covers all nodes to be inspected, ensuring that each node to be inspected is visited only once.

[0146] Furthermore, the distance between adjacent nodes is first calculated using the node coordinates, and the movement time is obtained by combining the movement speed. Then, the preset inspection time of each node is added to obtain the total inspection time.

[0147] Retrieve the inspection priority quantification value corresponding to each node to be inspected within the initial inspection path from the equipment management system, and add all the quantification values ​​to obtain the inspection priority coverage sum.

[0148] Based on the number of available inspection personnel at the hydropower station, the initial inspection route is divided into several sub-routes. The total inspection time for each sub-routes is calculated, and the maximum and minimum values ​​are found. The difference between the two values ​​is the personnel load allocation difference value.

[0149] Step S534: Calculate the pheromone change value based on the total inspection time, inspection priority coverage, and personnel workload distribution difference value; It should be noted that the pheromone change value is a parameter used in the improved ant colony strategy to update the pheromone concentration of a path. Its value is determined by the total inspection time of the initial inspection path, the inspection priority coverage, and the difference in personnel workload allocation. The better the path performs, that is, the shorter the total time, the higher the coverage, and the smaller the difference, the larger the pheromone change value, and the greater the increase in the pheromone concentration of the corresponding path.

[0150] Understandably, weights are assigned to the total inspection time, inspection priority coverage, and personnel workload distribution difference. These three indicators are standardized: a shorter total inspection time corresponds to a higher standardized score, higher inspection priority coverage corresponds to a higher standardized score, and a smaller personnel workload distribution difference corresponds to a higher standardized score. The standardized scores of the three indicators are multiplied by their respective weights, summed, and then multiplied by a preset coefficient to obtain the pheromone change value corresponding to the initial inspection path.

[0151] Step S535: Update the initial inspection path based on the pheromone change value to obtain the optimized inspection route.

[0152] Understandably, based on the pheromone change value, the pheromone concentration of each node in the initial inspection path is updated: the original pheromone concentration is added to the pheromone change value, and pheromone is evaporated according to a preset ratio to obtain the updated pheromone concentration. Steps S532-S534 are repeated, allowing multiple ants to search and form different initial inspection paths, and the pheromone concentration is updated sequentially until the preset number of iterations is completed. Among all the initial inspection paths generated in the iterations, the path with the shortest total inspection time, the highest inspection priority coverage, and the smallest difference in personnel workload allocation is selected as the optimized inspection route.

[0153] Based on an improved ant colony strategy, combining pheromone concentration and heuristic factors for precise path selection, and incorporating multi-dimensional evaluation indicators such as total inspection time, inspection priority, and personnel workload, an optimized inspection route that is both efficient and tailored to actual needs can be planned. This route ensures that inspection personnel prioritize covering high-priority equipment, reduce ineffective movement time, and balance personnel workload, significantly improving inspection efficiency and resource utilization compared to traditional manually planned routes.

[0154] This embodiment provides a hydropower station inspection method based on multi-source sensing devices. By acquiring target anomaly scores, calculating the predicted value of the remaining service life of the equipment and the probability of fault transfer, determining the equipment risk index and inspection priority, and using an improved ant colony strategy to determine the optimized inspection route, it solves the technical problems of unreasonable allocation of inspection tasks, unscientific planning of inspection routes, and inability to effectively cope with changes in equipment operating status in traditional inspection methods. It achieves the beneficial effects of improving inspection efficiency, optimizing resource allocation, and reducing equipment failure risk, thereby realizing more precise equipment maintenance management.

[0155] For example, to help understand the implementation process of the hydropower station inspection method based on multi-source sensing devices obtained by combining this embodiment with the above embodiment one, please refer to... Figure 4 , Figure 4 A simplified flowchart of a hydropower station inspection method based on multi-source sensing devices is provided, specifically: The system first inputs risk prediction data into the system through a predictive input step. The next step is the early warning decision step, which determines whether a high risk exists. This is a decision point where the subsequent process is determined based on the level of risk. If the judgment is yes, indicating a high risk, the system proceeds to the task optimization step, which involves optimizing the inspection task to address the high-risk situation. If the judgment is no, indicating no high risk, the system proceeds to the standard inspection step, executing the standard inspection process. Both task optimization and standard inspection involve execution and feedback steps, including executing the corresponding inspection task and collecting feedback information. This feedback information is used for data iteration to improve the risk prediction input, forming a closed-loop management system. The entire flowchart, through clearly defined steps and decision points, guides how to adjust inspection strategies according to risk levels, ensuring timely and effective response to potential risks.

[0156] It should be noted that the above examples are only for understanding this application and do not constitute a limitation on the hydropower station inspection method based on multi-source sensing equipment in this application. Any further modifications based on this technical concept are within the protection scope of this application.

[0157] This application also provides a hydropower station inspection device based on multi-source sensing equipment; please refer to... Figure 5 The hydropower station inspection device based on multi-source sensing equipment includes: Data acquisition module 10 is used to acquire multi-source operation data, including sensor telemetry data, industrial control operation data and structured vision data; Vector construction module 20 is used to construct multimodal feature vectors based on the multi-source running data; Trend analysis module 30 is used to perform equipment status prediction and trend analysis based on long-term time series prediction strategy and the multimodal feature vector to obtain equipment operation trend analysis results; Early warning generation module 40 is used to generate target early warning information based on the multimodal feature vector and the isolated forest strategy; The task generation module 50 is used to generate target inspection tasks based on the analysis results of the equipment operation trend.

[0158] The hydropower station inspection device based on multi-source sensing equipment provided in this application, employing the hydropower station inspection method based on multi-source sensing equipment described in the above embodiments, can solve the technical problem that existing hydropower station inspection methods cannot achieve accurate real-time identification and trend prediction of early, minor, and latent faults in hydropower station equipment. Compared with the prior art, the beneficial effects of the hydropower station inspection device based on multi-source sensing equipment provided in this application are the same as those of the hydropower station inspection method based on multi-source sensing equipment provided in the above embodiments, and other technical features in the hydropower station inspection device based on multi-source sensing equipment are the same as those disclosed in the methods of the above embodiments, and will not be repeated here.

[0159] This application provides a hydropower station inspection device based on a multi-source sensing device. The hydropower station inspection device based on a multi-source sensing device includes: at least one processor; and a memory communicatively connected to the at least one processor; wherein the memory stores instructions that can be executed by the at least one processor, and the instructions are executed by the at least one processor to enable the at least one processor to execute the hydropower station inspection method based on a multi-source sensing device in the above embodiment 1.

[0160] The following is for reference. Figure 6 It shows a structural schematic diagram of a hydropower station inspection device based on a multi-source sensing device suitable for implementing the embodiments of this application. Figure 6 The hydropower station inspection equipment based on multi-source sensing devices shown is merely an example and should not impose any limitations on the functionality and scope of use of the embodiments of this application.

[0161] like Figure 6As shown, the hydropower station inspection equipment based on multi-source sensing devices may include a processing unit 1001 (e.g., a central processing unit, a graphics processing unit, etc.), which can perform various appropriate actions and processes according to the program stored in ROM (Read Only Memory) 1002 or the program loaded from storage device 1003 into RAM (Random Access Memory) 1004. RAM 1004 also stores various programs and data required for the operation of the hydropower station inspection equipment based on multi-source sensing devices. The processing unit 1001, ROM 1002, and RAM 1004 are interconnected via bus 1005. Input / output (I / O) interface 1006 is also connected to the bus. Typically, the following systems can be connected to I / O interface 1006: input devices 1007 including, for example, touchscreens, touchpads, keyboards, mice, image sensors, microphones, accelerometers, gyroscopes, etc.; output devices 1008 including, for example, liquid crystal displays (LCDs), speakers, vibrators, etc.; storage devices 1003 including, for example, magnetic tapes, hard disks, etc.; and communication devices 1009. Communication device 1009 allows the hydropower station inspection equipment based on multi-source sensing devices to exchange data wirelessly or via wired communication with other devices. Although the figure shows hydropower station inspection equipment based on multi-source sensing devices with various systems, it should be understood that it is not required to implement or possess all the systems shown. More or fewer systems can be implemented or possessed alternatively.

[0162] The hydropower station inspection equipment based on multi-source sensing devices provided in this application, employing the hydropower station inspection method based on multi-source sensing devices described in the above embodiments, can solve the technical problem that existing hydropower station inspection methods cannot achieve accurate real-time identification and trend prediction of early, minor, and latent faults in hydropower station equipment. Compared with the prior art, the beneficial effects of the hydropower station inspection equipment based on multi-source sensing devices provided in this application are the same as those of the hydropower station inspection method based on multi-source sensing devices provided in the above embodiments, and other technical features of this hydropower station inspection equipment based on multi-source sensing devices are the same as those disclosed in the previous embodiment method, and will not be repeated here.

[0163] This application provides a computer-readable storage medium having computer-readable program instructions (i.e., a computer program) stored thereon, which are used to execute the hydropower station inspection method based on multi-source sensing devices in the above embodiments.

[0164] The aforementioned computer-readable storage medium carries one or more programs. When these programs are executed by the hydropower station inspection equipment based on multi-source sensing devices, the hydropower station inspection equipment based on multi-source sensing devices enables the following actions: acquiring multi-source operational data, including sensor telemetry data, industrial control operational data, and structured visual data; constructing multi-modal feature vectors based on the multi-source operational data; performing equipment status prediction and trend analysis based on long-term time-series prediction strategies and multi-modal feature vectors to obtain equipment operation trend analysis results; generating target early warning information based on multi-modal feature vectors and isolated forest strategies; and generating target inspection tasks based on the equipment operation trend analysis results.

[0165] The flowcharts and block diagrams in the accompanying drawings illustrate the architecture, functionality, and operation of possible implementations of systems, methods, and computer program products according to various embodiments of this application. In this regard, each block in a flowchart or block diagram may represent a module, segment, or portion of code containing one or more executable instructions for implementing a specified logical function. In some alternative implementations, the functions indicated in the blocks may occur in a different order than those indicated in the drawings.

[0166] The readable storage medium provided in this application is a computer-readable storage medium that stores computer-readable program instructions (i.e., a computer program) for executing the above-described hydropower station inspection method based on multi-source sensing devices. This solves the technical problem that existing hydropower station inspection methods cannot achieve accurate real-time identification and trend prediction of early, minor, and latent faults in hydropower station equipment. Compared with the prior art, the beneficial effects of the computer-readable storage medium provided in this application are the same as those of the hydropower station inspection method based on multi-source sensing devices provided in the above embodiments, and will not be repeated here.

[0167] This application also provides a computer program product, including a computer program that, when executed by a processor, implements the steps of the hydropower station inspection method based on multi-source sensing devices as described above.

[0168] The computer program product provided in this application can solve the technical problem that existing hydropower station inspection methods cannot achieve accurate real-time identification and trend prediction of early, minor, and latent faults in hydropower station equipment. Compared with the prior art, the beneficial effects of the computer program product provided in this application are the same as those of the hydropower station inspection method based on multi-source sensing devices provided in the above embodiments, and will not be repeated here.

[0169] The above description is only a part of the embodiments of this application and does not limit the patent scope of this application. All equivalent structural transformations made under the technical concept of this application and using the contents of the specification and drawings of this application, or direct / indirect applications in other related technical fields, are included in the patent protection scope of this application.

Claims

1. A method for inspecting a hydropower station based on multi-source sensing devices, characterized in that, The method includes: Acquire multi-source operational data, including sensor telemetry data, industrial control operational data, and structured vision data; Construct a multimodal feature vector based on the multi-source operational data; Based on the long-term time series prediction strategy and the multimodal feature vector, equipment status prediction and trend analysis are performed to obtain equipment operation trend analysis results; Target early warning information is generated based on the multimodal feature vectors and the isolated forest strategy; Target inspection tasks are generated based on the analysis results of the equipment operation trends.

2. The method as described in claim 1, characterized in that, The step of performing equipment status prediction and trend analysis based on the long-term time series prediction strategy and the multimodal feature vector to obtain the equipment operation trend analysis results includes: The target feature tensor is determined based on the long-term time series prediction strategy, the preset periodic position encoding, and the multimodal feature vector; Calculate the information dispersion of the target feature tensor, and determine the target dispersion vector based on the information dispersion; Attention is calculated based on the target dispersion vector and the key vector to obtain the target attention, and the initial prediction result is determined based on the target attention; The target difference value of the initial prediction result is calculated based on the quantile regression loss strategy, and the equipment status prediction and trend analysis are performed based on the target difference value to obtain the equipment operation trend analysis result.

3. The method as described in claim 2, characterized in that, The steps for calculating the target difference value of the initial prediction results based on the quantile regression loss strategy include: Determine the preset quantile and target quantile based on the quantile regression loss strategy; Determine whether the predicted quantile of the initial prediction result is greater than the target quantile; When the predicted quantile is less than or equal to the target quantile, the target difference value is calculated based on the preset quantile, the predicted quantile, and the target quantile; When the predicted quantile is greater than the target quantile, the target difference value is calculated based on the preset quantile and the predicted quantile.

4. The method as described in claim 2, characterized in that, The step of calculating the target difference value of the initial prediction result based on the quantile regression loss strategy, and performing equipment status prediction and trend analysis based on the target difference value to obtain the equipment operation trend analysis result further includes: The multimodal feature vectors are input into a preset variational autoencoder to obtain the target latent distribution; Calculate the Wasserstein distance between the target potential distribution and the standard normal distribution, and determine the equipment health index based on the Wasserstein distance; The equipment health indicators and the multimodal feature vectors are input into a deep survival model to obtain the failure probability and failure time prediction value, wherein the deep survival model is used to represent the relationship between the equipment health indicators, the failure probability and the failure time prediction value. The instantaneous failure rate is calculated based on the failure probability and the predicted failure time. The fault transfer probability is calculated based on the instantaneous failure rate, and the equipment operation trend analysis results are updated based on the fault transfer probability.

5. The method as described in claim 1, characterized in that, The step of generating target early warning information based on the multimodal feature vector and the isolated forest strategy includes: Based on the multimodal feature vectors and the isolated forest strategy, the original anomaly score is calculated; The similarity is calculated based on the multimodal feature vectors, and the spatial weights are determined based on the similarity. The degree of difference is calculated based on the multimodal feature vectors, and the time weight is determined based on the degree of difference. The target anomaly score is calculated based on the original anomaly score, the spatial weight, and the temporal weight. The warning threshold is determined based on the historical anomaly distribution and the preset false alarm rate, and the target warning information is generated based on the target anomaly score and the warning threshold.

6. The method as described in claim 1, characterized in that, The step of generating the target inspection task based on the equipment operation trend analysis results includes: Obtain the target anomaly score, and based on the equipment operation trend analysis results, obtain the predicted value of the remaining service life of the equipment and the probability of failure transfer; The equipment risk index is calculated based on the predicted remaining service life of the equipment, the failure transfer probability, and the target anomaly score, and the inspection priority is determined based on the equipment risk index. The optimized inspection route is determined based on the improved ant colony strategy, and the target inspection task is generated by combining the inspection priority and the optimized inspection route.

7. The method as described in claim 6, characterized in that, The steps for determining the optimized inspection route based on the improved ant colony strategy include: Obtain initial nodes to be inspected, and determine pheromone concentration and heuristic factor according to the improved ant colony strategy, wherein the heuristic factor is associated with the distance between the initial nodes to be inspected and the inspection priority corresponding to the initial nodes to be inspected. The intermediate nodes to be inspected are selected based on the initial nodes to be inspected, the pheromone concentration, and the heuristic factor. The initial inspection path is determined based on the intermediate nodes to be inspected, and the total inspection time, inspection priority coverage, and personnel load distribution difference value of the initial inspection path are calculated. The pheromone change value is calculated based on the total inspection time, the inspection priority coverage, and the personnel workload allocation difference value. The initial inspection path is updated based on the pheromone change value to obtain an optimized inspection route.

8. The method as described in claim 1, characterized in that, The step of constructing a multimodal feature vector based on the multi-source operational data includes: Outlier processing is performed on the multi-source operational data to obtain processed multi-source operational data; Based on the processed multi-source running data and the preset sliding time window, time synchronization and resampling are performed to obtain aligned multi-source running data; The image information of the aligned multi-source running data is subjected to deep feature quantization to obtain fused multi-source running data; A multimodal feature vector is constructed based on the fused multi-source operational data.

9. The method as described in claim 8, characterized in that, The step of constructing a multimodal feature vector based on the fused multi-source operational data includes: The target feature vector is determined based on the fused multi-source operation data. The target feature vector includes a telemetry time series feature vector, an industrial control operation feature vector, and a visual feature vector. The historical sensitivity mean and historical standardized outlier are determined based on the target feature vector; The anomaly sensitivity weight of the target feature vector is calculated based on the target feature vector, the historical sensitivity mean, and the historical standardized outlier. The target feature vector is weighted and fused according to the anomaly sensitivity weight to obtain a multimodal feature vector.

10. A hydropower station inspection device based on multi-source sensing equipment, characterized in that, The device includes: The data acquisition module is used to acquire multi-source operational data, including sensor telemetry data, industrial control operational data, and structured vision data. A vector construction module is used to construct multimodal feature vectors based on the multi-source runtime data; The trend analysis module is used to perform equipment status prediction and trend analysis based on the long-term time series prediction strategy and the multimodal feature vector to obtain equipment operation trend analysis results. The early warning generation module is used to generate target early warning information based on the multimodal feature vector and the isolated forest strategy; The task generation module is used to generate target inspection tasks based on the analysis results of the equipment operation trends.

Images