Industrial solid waste treatment equipment remote diagnosis method

Abstract

The present application relates to a kind of industrial solid waste treatment equipment remote diagnosis method, comprising: by the distributed collection equipment of multiple types of sensors different parts of multimodal real-time working condition signal, in combination with geographical position and working condition label, establish position annotation, working condition mapping multi-source original data set;Based on statistical, time domain and frequency domain feature mining, generate high-dimensional feature node that fuses space-time information;Accordingly, construct space-time evolution graph network, and introduce dynamic graph neural network to learn the dynamic association between nodes and working condition evolution law, realize the high-precision identification of complex anomaly and precursor track extraction, finally generate traceable multi-label anomaly early warning.

Description

Technical Field

[0001] This invention relates to the field of "intelligent maintenance technology for industrial solid waste treatment equipment", and more particularly to a remote diagnostic method for industrial solid waste treatment equipment. Background Technology

[0002] Currently, intelligent remote diagnostics and anomaly early warning technologies for industrial solid waste treatment equipment have received widespread attention in the industry. Existing industrial operation and maintenance systems generally rely on the acquisition of signals from various types of sensors, combined with certain signal analysis and intelligent discrimination algorithms, to monitor the equipment's operating status, detect anomalies, and provide preliminary early warnings. Mainstream technical approaches include: time-series feature extraction based on single-point or global sensors, statistical threshold discrimination, rule-based screening, traditional machine learning (such as support vector machines and decision trees), and early neural network models (such as CNN / RNN) for fault detection and classification, as well as edge computing and cloud-integrated diagnostic solutions for remote operation and maintenance. These technologies provide basic fault identification and early warning capabilities for industrial equipment and have been supported in the operational health management systems of most industrial solid waste treatment processes such as incineration, sorting, and crushing equipment, and are gradually being updated towards multi-source data fusion, distributed intelligent analysis, and automated maintenance.

[0003] In terms of application scenarios, as equipment operating conditions become increasingly complex, multiple parts operate collaboratively, and the types of anomalies continue to increase, the industry has begun to explore high-dimensional feature fusion and intelligent analysis methods for multi-source, multi-modal signals. For example, in recent years, there have been practices such as comprehensive diagnosis based on multi-channel vibration and temperature signals, hierarchical discrimination based on multi-dimensional operating condition labels, and remote anomaly data acquisition and simplified pattern recognition for new solid waste treatment equipment. Some advanced systems have introduced operating condition label-driven status recognition, finite-granularity multi-label classification and discrimination, and anomaly trend analysis to improve the automation level of equipment maintenance.

[0004] However, existing technologies still have the following prominent problems and technical shortcomings in practical applications:

[0005] (1) The classification of abnormal patterns has limited subdivision, and the pattern label set is usually difficult to exhaust all types, which makes it easy to confuse early abnormal signals of the same type but different mechanisms. For example, in the case of multiple parts linkage and physical quantities intertwined, single channel or single pattern classification methods cannot effectively decompose and distinguish small and complex early abnormalities, affecting fault tracing and fine-grained prediction.

[0006] (2) There is a lack of intelligent modeling methods that can integrate multi-source spatiotemporal evolution information. Existing methods mostly focus on local statistical features or short time series data, which makes it difficult to characterize the spatiotemporal coupling, evolution path and causal relationship between nodes of multi-channel signals under complex working conditions, and cannot support global anomaly trend monitoring and precursor trajectory tracking of highly complex equipment.

[0007] (3) Insufficient generalization ability for overlapping abnormal patterns and unseen types. Traditional machine learning and single-label classification methods are mostly trained based on historical archived labels. For sudden scenarios such as new processes, equipment upgrades, and extreme changes in operating conditions, they lack robust recognition ability for unknown precursors and anomalies, and are prone to false alarms, missed alarms, or even misjudgments.

[0008] (4) Poor traceability and decision support for anomaly warnings. Most existing systems only output single-point warning signals or coarse-grained anomaly labels, lacking multi-dimensional trajectory visualization of the entire process of anomaly generation, evolution, and propagation, as well as targeted safety measures suggestions. This makes it difficult for operation and maintenance personnel to quickly locate the root cause of anomalies, respond efficiently, and handle them in a closed loop. Summary of the Invention

[0009] This application provides a remote diagnostic method for industrial solid waste treatment equipment, which aims to solve one of the technical problems or technical problems existing in the prior art.

[0010] A remote diagnostic method for industrial solid waste treatment equipment is provided, which specifically includes the following steps:

[0011] S1: Collect multi-source real-time operating condition signals of industrial solid waste treatment equipment at multiple operating locations, including but not limited to vibration, temperature, flow rate and current data, and annotate the specific geographical location and equipment operating condition labels of each sampling channel to establish a multi-source operating condition raw dataset.

[0012] S2: Perform denoising and time-series alignment processing on the original dataset of the multi-source operating conditions. Eliminate the noise effects caused by differences in equipment environment and sensor characteristics through filtering and normalization methods to obtain standardized, multi-channel time-series aligned operating condition data.

[0013] S3: Based on standardized, multi-channel time-aligned operating condition data, calculate the statistics, time-domain features and frequency-domain spectral features of each channel within each sampling window, and generate multi-dimensional spatiotemporal feature node representations by combining the corresponding geographical location and operating condition labels.

[0014] S4: Based on the multi-dimensional spatiotemporal feature node representation, a spatiotemporal evolution graph network is constructed using a preset time sliding window as a unit. In the network, nodes within each window are connected by edges according to temporal sequence and spatial neighborhood relationship to characterize the working condition evolution path and potential causal relationship.

[0015] S5: Input the spatiotemporal evolution graph network into the dynamic graph neural network to learn the dynamic correlation and feature evolution law between nodes, and obtain a spatiotemporal evolution feature description of nodes containing labels of different parts and working conditions, which is compatible with multiple working conditions and heterogeneous signal distributions.

[0016] S6: Based on the spatiotemporal evolution feature description of nodes, the abnormal probability distribution of each node under the abnormal mode type and working condition label is output through a multi-label classification head. At the same time, the evolution trajectory of the nodes is generated and discriminated against using a generative adversarial network to improve the ability to identify complex overlapping and unknown anomalies.

[0017] S7: Based on the obtained anomaly probability distribution and node evolution trajectory, perform precursor trend weighted aggregation on multiple types of early minor anomaly signals, introduce the spatiotemporal neighborhood correlation information before and after, and extract traceable anomaly precursor trajectory data.

[0018] S8: Determine whether the overall anomaly probability of the abnormal precursor trajectory data exceeds the preset warning threshold. If it does, generate a corresponding multi-label early anomaly warning instruction; otherwise, continue to dynamically monitor and adaptively optimize the graph network parameter group.

[0019] S9: The multi-label early anomaly warning command and anomaly precursor trajectory data are visualized and then pushed to the remote operation and maintenance terminal, and targeted safety measures or operation and maintenance work orders are automatically triggered to achieve closed-loop feedback and timely intervention.

[0020] This invention proposes a remote diagnostic method for industrial solid waste treatment equipment, which has the following significant advantages:

[0021] (1) Significantly improves the ability to subdivide and accurately identify abnormal patterns. This invention improves the traditional single-channel, low-dimensional feature input to a multi-physical quantity, spatiotemporal multimodal fusion feature input by distributing multi-source working condition data, aligning spatiotemporal consistency, and mining multi-dimensional features. It characterizes subtle trend changes based on the dynamic spatiotemporal evolution graph structure. The dynamic graph neural network can fully model the time-space-working condition multi-dependency in node evolution, distinguish multiple abnormal categories that are highly overlapping or have similar mechanisms under complex working conditions with fine granularity, significantly improve the discrimination resolution of abnormal categories and the concurrent recognition rate under multi-label states, and effectively overcome the problem that traditional statistical or isolated machine learning methods cannot distinguish complex and heterogeneous abnormal patterns.

[0022] (2) Enhance the trend extraction, traceability, and interpretability of early anomaly precursors. This invention uses window-level node modeling, large-scale parallel feature aggregation, and trend-weighted aggregation mechanisms to extract the trajectory of minute precursor signals under spatiotemporal correlation, automatically trace the occurrence and evolution path of weak anomalies, significantly improve the detection sensitivity of early minute anomaly signals and the spatiotemporal traceability capability of precursor trajectories, provide maintenance personnel with a clear, visualized, and traceable fault precursor evolution map, and provide strong technical support for precise preventive maintenance.

[0023] (3) Significantly enhances the generalization and early warning capabilities for novel anomalies and unknown patterns. By introducing a generative adversarial network to train the generation and discrimination of real and false anomalies in the node evolution trajectory, the model possesses the ability to actively perceive and identify the evolutionary trends of "unseen samples" or boundary samples beyond known anomaly patterns. This effectively overcomes the problems of weak generalization ability and early warning blind spots in existing models when there are insufficient labeled anomaly samples. Experiments show that the overall early warning accuracy is improved by more than 15% compared with the classical method based on static feature learning, and the effective identification rate of novel anomalies is significantly improved.

[0024] (4) Achieving highly efficient intelligent diagnosis and closed-loop remote early warning feedback. This invention improves the real-time processing capability of complex operating data of large-scale equipment through efficient parallel processing of high-dimensional nodes in graph networks, multi-label discrimination, and automated operating condition label mapping. The response time for various types of anomaly detection is significantly shortened, meeting the industrial site's demand for intelligent diagnosis with high real-time performance, low false alarms, and high coverage. Anomaly early warnings can be automatically pushed through the automatic generation of safety measures instructions / work orders and a closed-loop feedback mechanism, achieving seamless and efficient end-to-end operation and maintenance response, greatly reducing the risk of manual judgment and omissions, and improving production safety and the level of automation in equipment management. Attached Figure Description

[0025] Appendix Figure 1 This is a flowchart of a remote diagnostic method for industrial solid waste treatment equipment according to this application. Detailed Implementation

[0026] Embodiments of the present invention are described in detail below, examples of which are shown in the accompanying drawings, wherein the same or similar reference numerals denote the same or similar elements or elements having the same or similar functions throughout. The embodiments described below with reference to the accompanying drawings are exemplary and are only used to explain the present invention, and should not be construed as limiting the present invention.

[0027] The following disclosure provides many different embodiments or examples for implementing various structures of the invention. To simplify the disclosure, specific examples of components and arrangements are described below. These are merely examples and are not intended to limit the invention. Furthermore, reference numerals and / or letters may be repeated in different examples; such repetition is for simplification and clarity and does not in itself indicate a relationship between the various embodiments and / or arrangements discussed. In addition, examples of various specific processes and materials are provided in this invention, but those skilled in the art will recognize the application of other processes and / or the use of other materials.

[0028] like Figure 1 As shown, this application provides a remote diagnostic method for industrial solid waste treatment equipment, which aims to solve one of the technical problems or technical problems existing in the prior art.

[0029] A remote diagnostic method for industrial solid waste treatment equipment is provided, which specifically includes the following steps:

[0030] S1: Collect multi-source real-time operating condition signals from multiple operating locations of industrial solid waste treatment equipment, including vibration, temperature, flow rate and current data, and annotate the specific geographical location and equipment operating condition labels for each sampling channel to establish a multi-source operating condition raw dataset.

[0031] S2: Perform denoising and time-series alignment processing on the original dataset of the multi-source operating conditions. Eliminate the noise effects caused by differences in equipment environment and sensor characteristics through filtering and normalization methods to obtain standardized, multi-channel time-series aligned operating condition data.

[0032] S3: Based on standardized, multi-channel time-aligned operating condition data, calculate the statistics, time-domain features and frequency-domain spectral features of each channel within each sampling window, and generate multi-dimensional spatiotemporal feature node representations by combining the corresponding geographical location and operating condition labels.

[0033] S4: Based on the multi-dimensional spatiotemporal feature node representation, a spatiotemporal evolution graph network is constructed using a preset time sliding window as a unit. In the network, nodes within each window are connected by edges according to temporal sequence and spatial neighborhood relationship to characterize the working condition evolution path and potential causal relationship.

[0034] S5: Input the spatiotemporal evolution graph network into the dynamic graph neural network to learn the dynamic correlation and feature evolution law between nodes, and obtain a spatiotemporal evolution feature description of nodes containing labels of different parts and working conditions, which is compatible with multiple working conditions and heterogeneous signal distributions.

[0035] S6: Based on the spatiotemporal evolution feature description of nodes, the abnormal probability distribution of each node under the abnormal mode type and working condition label is output through a multi-label classification head. At the same time, the evolution trajectory of the nodes is generated and discriminated against using a generative adversarial network to improve the ability to identify complex overlapping and unknown anomalies.

[0036] S7: Based on the obtained anomaly probability distribution and node evolution trajectory, perform precursor trend weighted aggregation on multiple types of early minor anomaly signals, introduce the spatiotemporal neighborhood correlation information before and after, and extract traceable anomaly precursor trajectory data.

[0037] S8: Determine whether the overall anomaly probability of the abnormal precursor trajectory data exceeds the preset warning threshold. If it does, generate a corresponding multi-label early anomaly warning instruction; otherwise, continue to dynamically monitor and adaptively optimize the graph network parameter group.

[0038] S9: The multi-label early anomaly warning command and anomaly precursor trajectory data are visualized and then pushed to the remote operation and maintenance terminal, and targeted safety measures or operation and maintenance work orders are automatically triggered to achieve closed-loop feedback and timely intervention.

[0039] S1: Collect multi-source real-time operating condition signals from multiple operating locations of industrial solid waste treatment equipment, including vibration, temperature, flow rate, and current data, and annotate each sampling channel with specific geographical location and equipment operating condition labels to establish a multi-source operating condition raw dataset, specifically including:

[0040] S1.1: Configure multiple types of sensor nodes, including vibration sensors, temperature sensors, flow sensors and current sensors, on each key operating part of the industrial solid waste treatment equipment to realize distributed physical quantity acquisition of multimodal real-time operating condition signals.

[0041] For each key operating part of the industrial solid waste treatment equipment, based on the equipment structure and monitoring requirements, the functional units and components that need to be acquired in real time are selected as the objects for sensor node placement.

[0042] A distributed sensor node configuration strategy (modular distributed deployment principle) is adopted, selecting multiple types of sensors including but not limited to vibration sensors, temperature sensors, flow sensors and current sensors. Combined with the monitoring characteristics of each part, the corresponding type and number of sensor nodes are configured for each operating part to achieve full coverage sampling for physical quantity categories.

[0043] Furthermore, by using the node parameter configuration method (based on the range, sensitivity, and anti-interference capability of the monitored object), the working range and frequency of change of each type of physical quantity are used as the main setting basis to reasonably determine the sensor specifications, including core elements such as measurement range, sampling frequency, sensitivity level, and signal output type, so as to meet the requirements for real-time acquisition of multimodal signals.

[0044] It adopts an efficient data acquisition terminal and transmission interface integration method, selects corresponding signal processing interfaces (voltage, current, digital bus, etc.) for different types of sensors, and unifies the physical layout of sensor nodes through industrial-grade data acquisition units (such as PLC, embedded acquisition terminals) to achieve efficient local acquisition of distributed signals.

[0045] Furthermore, by using a unique identifier allocation algorithm for sensor nodes (combining device structure tree topology, device number, and node geographical location information), a unique identifier is assigned to each sensor node, enabling traceable management of node-level signal data.

[0046] Through the above processing method, the configuration, parameter setting and interface integration of multiple types of sensor nodes covering multiple key operating parts are transformed into distributed real-time physical signal raw data with diverse physical quantities, clear acquisition locations and unique node identification characteristics, so as to realize the underlying data acquisition for multi-source and multi-modal operating condition monitoring.

[0047] For example, in an industrial solid waste treatment equipment for incinerator slag treatment, vibration sensors (model: IEPE100mV / g, sampling frequency 5kHz, measurement range ±50g), temperature sensors (model: PT100, measurement range -50~+300℃, sampling frequency 1Hz), and flow sensors (model: electromagnetic flowmeter, measurement range 0-50m) are respectively configured for key operating parts such as the main drive motor, slag hopper discharge port, water inlet cooling system, and dust removal fan. 3 Multiple types of nodes, including / h (sampling frequency 10Hz) and current sensors (model: Hall type, measurement range 0-100A, sampling frequency 1kHz).

[0048] The distribution of sensor physical points follows the equipment structure layout diagram and is integrated into the local PLC acquisition unit using a fieldbus protocol (such as Modbus-RTU) to realize data acquisition, real-time uploading and distributed storage.

[0049] By using unique identifier encoding rules in the form of "device ID + part ID + node type code", a single-point fully unique identifier such as "DEV01-DRMTR-VIB01" is generated.

[0050] After executing the above configuration, the multimodal distributed acquisition nodes of each key operating condition part form a linkage, and the output raw multi-source signal data covers the entire life cycle operation of the equipment, supporting the needs of subsequent spatiotemporal synchronization and multi-label abnormal pattern recognition, effectively improving the perceptibility of early minor fault precursors and the systematic monitoring capability.

[0051] S1.2: Based on a time synchronization protocol (such as IEEE 1588PTP), global time synchronization is performed on the data acquisition terminals of each sensor node to obtain the original signal data stream of multi-source acquisition time sequence under the standard time axis, providing a foundation for subsequent spatiotemporal alignment and dynamic modeling.

[0052] For the distributed raw operating condition signals output after the deployment of multiple types of sensor nodes, the setting of various physical quantity acquisition parameters and the integration of signal interfaces have been completed, a unified time reference alignment needs to be implemented before entering the upstream data stream processing chain to ensure the timeliness of synchronous acquisition of data between different nodes and the basis for subsequent spatial-temporal correlation analysis of multi-source asynchronous signals.

[0053] The global timestamp synchronization of data acquisition terminals of each distributed sensor node is achieved by adopting the IEEE 1588 Precision Time Protocol (PTP, parameters: master clock, slave clock, expected nanosecond-level synchronization accuracy).

[0054] Furthermore, by using a clock synchronization management chip or a high-precision clock module (parameters: local oscillator stability, master-slave synchronization polling cycle, network latency compensation value), the latency drift caused by physical wiring and network transmission can be limited, thereby obtaining a highly consistent sampling time reference.

[0055] An automatic synchronization error correction algorithm (parameters: sampling clock deviation tolerance threshold, sliding window verification period, drift fitting compensation) is adopted. Within the multi-node time-division polling window, the local time record of each sensor node is dynamically compared with the system master clock, and the periodic sampling increment is adjusted in real time to converge the synchronization error.

[0056] By using a multi-node timing synchronization calibration method (parameters: node ID, master clock reference, sampling timestamp, root mean square error RMSE), a synchronous acquisition timing raw signal data stream with a standard time axis identifier is generated, which accurately records the instantaneous state changes of various operating condition signals at a unified time.

[0057] Through the above-mentioned timing alignment and synchronization control processing, the local data stream of distributed multi-source acquisition signals is transformed into a timing-consistent multi-source acquisition timing raw signal data stream. This lays a solid data foundation for subsequent steps such as spatiotemporal alignment, node characterization, and dynamic modeling using location and operating condition labels, and achieves global spatiotemporal consistency of cross-node and cross-physical quantity signals in industrial solid waste equipment operating condition monitoring.

[0058] For example, in a set of incinerator slag processing equipment, multiple types of sensors based on the Modbus-RTU protocol are deployed at the main drive motor, slag hopper discharge port, and dust removal fan. An IEEE 1588-PTP hardware time synchronization module is integrated into the local PLC unit, with a master clock source update frequency of 1Hz, supporting nanosecond-level alignment accuracy. Each acquisition channel records signal points with a master clock synchronization stamp (e.g., 2024-03-16 12:34:56.20000) and a locally unique sampling index number in 10ms increments. By setting the synchronization verification period to 5 minutes, comparing the maximum drift value of the master and slave node acquisition times, and applying a sliding drift fitting compensation algorithm, the actual synchronization error RMSE is less than 2µs. Finally, a data stream uniquely identified by "channel ID + master synchronization stamp" is generated, ensuring strict alignment between the main drive vibration signal, current sampling data, and channel data from different locations such as slag hopper temperature / flow rate. Under this scheme, multi-node and multi-modal signals achieve a unified time axis at the nanosecond level, which efficiently eliminates problems such as sample mismatch and response lag caused by time drift in subsequent spatial-temporal aggregation analysis of multi-source signals and intelligent identification of multi-label anomalies.

[0059] S1.3: Based on the equipment structure distribution diagram and digital twin positioning information, the acquired multi-source acquisition timing raw signal data stream is associated with a clear location identifier and operating part code for each signal data, forming a location-marked multi-source acquisition timing raw signal.

[0060] S1.4: Combine the operation status data of the production automation system SCADA or MES system to annotate the original time-series signals acquired from multiple sources at the location with operating condition labels, including operating mode, load characteristics, and operation parameter labels, so as to realize the mapping of operating conditions to the original time-series signals acquired from multiple sources.

[0061] S1.5: Map the operating conditions to the multi-source acquired time-series raw signals, and store them continuously according to the data time window as a multi-source operating condition raw dataset with location identifiers and operating condition labels, to ensure spatiotemporal data consistency and closed-loop labeling information, and provide structured input for the next step of standardization and denoising processing.

[0062] Step S2: Perform denoising and time-series alignment processing on the original multi-source operating condition dataset. This involves eliminating noise caused by differences in equipment environment and sensor characteristics through filtering and normalization methods to obtain standardized, multi-channel time-series aligned operating condition data. Specifically, this includes:

[0063] S2.1: A multi-scale adaptive filtering algorithm is used for each signal channel in the original dataset of the multi-source operating conditions. Based on the signal characteristics, an appropriate filtering window and parameters are selected to achieve preliminary suppression of environmental noise and outliers in signals such as vibration, temperature, flow rate and current, so as to obtain the primary denoised operating condition signal.

[0064] S2.2: Based on the primary denoised operating condition signal, wavelet transform and energy threshold decomposition techniques are used to decompose the signal, further separating and eliminating high-frequency interference and low-frequency drift components in the operating condition signal, retaining the main operating condition characteristic bandwidth, realizing fine denoising processing of multi-modal features, and obtaining high-fidelity denoised operating condition data.

[0065] S2.3: For high-fidelity denoised data, based on the sampling timestamps of each sampling channel, a global time reference alignment algorithm is used to perform time synchronization calibration on all signal sequences, adjusting the sampling lag and drift introduced by different sensors, ensuring strict alignment of multi-channel signals at the same time, and forming a time-synchronized operating condition signal.

[0066] S2.4: Perform channel normalization transformation on the time-synchronous operating condition signal. Referring to the measurement range and historical statistical distribution of each type of sensor, use standard fraction normalization or minimum-maximum normalization algorithm to unify the numerical scale of different channels and obtain the normalized operating condition signal matrix, providing standard input for multi-channel feature equalization analysis.

[0067] S2.5: Bind the normalized operating condition signal matrix with the equipment geographical location and operating condition labels annotated in the original sampling channels to generate multi-channel time-series aligned operating condition data with unique identification, standardized numerical distribution and time-series synchronization characteristics, thus establishing a data foundation for subsequent multi-dimensional spatiotemporal feature node characterization and heterogeneous signal fusion analysis.

[0068] S3: Based on standardized, multi-channel time-aligned operating condition data, calculate the statistics, time-domain features, and frequency-domain spectral features of each channel within each sampling window. Combined with the corresponding geographical location and operating condition labels, generate a multi-dimensional spatiotemporal feature node representation, specifically including:

[0069] S3.1: For each signal channel in the standardized, multi-channel time-aligned operating condition data, divide the interval according to the sampling time window, and obtain the statistical characteristics such as mean, variance, skewness, and kurtosis within the window based on the segmented statistical method, so as to reflect the global state distribution characteristics of different signal channels in a specific interval, and provide basic statistical factors for subsequent time domain and frequency domain feature extraction.

[0070] S3.2: Based on the statistical features generated in S3.1, for each sampling window's single signal channel, a sliding window time-domain signal decomposition algorithm (such as short-time autocorrelation, envelope analysis, instantaneous amplitude and instantaneous phase calculation) is used to mine its time-domain features, further obtaining key time-domain feature parameters such as the signal's extreme points, zero-crossing rate, and impulse factor, to characterize the dynamic change trend and provide time-domain constraint factors for subsequent frequency-domain transformation.

[0071] For each sampling window of the standardized, multi-channel time-aligned operating condition data, the statistical characteristics output in the previous step S3.1 are loaded as the prior parameter set for time-domain analysis.

[0072] A sliding window time-domain signal decomposition algorithm (typical methods include: short-time autocorrelation analysis, envelope analysis, instantaneous amplitude and instantaneous phase calculation; the sliding window length and overlap rate are set according to the signal sampling frequency and the duration of abnormal precursors) is used to extract the instantaneous amplitude trajectory of the data subsequence within each sampling window and obtain the dynamic response of the signal amplitude change.

[0073] Furthermore, the zero-crossing rate (ZCR) is calculated for the signals within the window using the zero-crossing detection method. The ZCR can be determined by the following formula:

[0074]

[0075] Where x[n] is the nth sampling point within the window, N is the window length, and sgn(·) is the sign function. The extreme point density is extracted by statistically analyzing the number of local maxima and minima within each sliding window using an extreme point detection algorithm.

[0076] Furthermore, the pulse factor K is used to measure abnormal events of signal impulse type, and its calculation formula is as follows:

[0077]

[0078] Where max(|x[n]|) is the maximum absolute value of the signal within the window. It is the average of the absolute values ​​of the signal.

[0079] By using envelope analysis methods, such as Hilbert transform, the envelope e[n] of the signal can be obtained, and the dimensions of low-amplitude, low-frequency fluctuations, including the features of minor anomalies, can be further explored.

[0080] The above-mentioned extreme point statistics, zero crossover rate, impulse factor, envelope amplitude, and local root mean square (RMS) and other time-domain characteristic parameters are synthesized into a window-level time-domain feature vector in a standardized form, which serves as a representation of the dynamic changes of each signal channel within the corresponding sampling window.

[0081] Through this multi-level time-domain feature mining and parameter aggregation, a set of feature indicators that can sensitively characterize subtle trends in equipment operating conditions and capture suspected precursor signals is obtained, providing time-domain constraints and feature basis for frequency domain analysis, spectral feature extraction, and advanced spatiotemporal feature embedding in the subsequent S3.3 sub-step.

[0082] For example, in the practical application of the three-axis vibration signal channel of the main drive motor for industrial solid waste crushing, the sampling frequency was set to 2048Hz, the sliding window length was set to 4096 points (corresponding to a duration of 2 seconds), and the window overlap rate was 50%. Short-time autocorrelation analysis revealed an abnormally high peak in the autocorrelation function within some abnormal windows. Within the same batch of sampling windows, the zero-crossing rate, calculated using a formula, suddenly increased from an average of 56 times / window to 80 times / window. Extreme point statistics showed that the number of maximum and minimum points in the local drive abnormal sample windows was 15% higher than the historical baseline under the same operating conditions. The impulse factor, calculated, increased from 1.92 in the normal window to 2.48. The root mean square (RMS) of the envelope amplitude, obtained through Hilbert transform, increased from 0.032 to 0.054. Finally, the above indicators were integrated to form a single-window multidimensional time-domain feature vector, which served as input for subsequent frequency-domain analysis and abnormal pattern label decomposition, achieving high-sensitivity detection of minute precursors. In multiple different production batches of data, the window-level temporal feature distribution can clearly distinguish between early bearing failures and normal operating conditions of the drive motor, achieving high-resolution detection of early abnormal trends.

[0083] S3.3: Using the time-domain feature parameters obtained in S3.2 as input, apply a multi-resolution spectrum analysis algorithm (such as Fast Fourier Transform (FFT) or multi-scale wavelet transform) to the data in each signal channel window to obtain periodic spectrum features such as the main frequency peak, frequency band energy distribution, spectral entropy, and characteristic frequency amplitude, so as to generate frequency domain feature factors for subsequent abnormal mode spectrum identification and high-dimensional node characterization.

[0084] The input consists of standard chemical condition data and its window-level time-domain characteristic parameters after high-fidelity multi-channel time-series alignment processing.

[0085] A multi-resolution spectrum analysis algorithm (parameters: Fast Fourier Transform (FFT), window function type and length are set according to the sampling frequency, multi-scale wavelet transform uses wavelet bases such as Daubechies and Symlet, and is matched with multi-level decomposition depth) is used to perform frequency domain transformation on the data sequence of each signal channel within each data window, thereby realizing signal spectrum mapping within the local time window range.

[0086] Furthermore, the discrete Fourier transform of the signal x[n] within the window is performed using the FFT algorithm, and its calculation formula is as follows:

[0087]

[0088] Where X[k] is the complex spectral component at the k-th frequency point, and N is the window length.

[0089] By analyzing the amplitude distribution of |X[k]| in the frequency domain, the peak frequency f corresponding to the largest amplitude is extracted. peak And based on the frequency distribution interval, the frequency band energy distribution is statistically analyzed:

[0090]

[0091] Where B is the set of frequency point indices for the target frequency band, and E B Accumulate energy for this frequency band.

[0092] The spectral entropy method is used to measure the degree of disorder in the frequency distribution, and its calculation formula is as follows:

[0093]

[0094] in, This represents the normalized spectral power probability.

[0095] Furthermore, a multi-scale wavelet transform algorithm is employed, setting the wavelet basis type and the number of decomposition levels to perform multi-scale decomposition of the signal, extracting the wavelet energy distribution E under each scale bandwidth. w,l Peak value of wavelet coefficients C w,max And diverse frequency domain factors such as characteristic frequency amplitude.

[0096] Through the combined processing of FFT and wavelet analysis, periodic spectral characteristic indicators such as the dominant frequency peak, frequency band energy distribution, spectral entropy, wavelet energy, and characteristic frequency amplitude are generated for each sampling window.

[0097] By connecting multi-resolution spectral features with time-domain feature output data, an important frequency domain index foundation is generated for subsequent abnormal mode spectral identification and high-dimensional node feature characterization, enabling sensitive spectral characterization of precursor anomalies under complex operating conditions and enhancing the diversity of node features.

[0098] For example, for the triaxial vibration signal of the main drive motor of an industrial solid waste treatment system, after high-fidelity noise reduction and time window segmentation of 2096 points / window - 2 seconds, the FFT algorithm is used to calculate the spectrum of each window, with the window function set to Hanning window and the sampling frequency set to 4096Hz. The main peak frequency f of the spectrum is calculated. peak For example, under a certain type of working condition f peak =46Hz, abnormal precursor stage rises to f peak =79Hz, peak amplitude increased by 41%. Simultaneously, the energy ratio of the low-frequency band (0-40Hz) to the fault characteristic bandwidth (70-120Hz) was statistically analyzed. The early anomaly window was increased to 2.13 times. The spectral entropy index H... s The decrease from the normal window of 1.86 to 1.42 during anomaly monitoring reflects the clustering of energy distribution. Multi-scale wavelet decomposition using the Daubechies-4 wavelet with 5 levels revealed that the bandwidth energy E of the 3rd level wavelet... w,3The value increased from the normal average of 0.016 to 0.029, with the characteristic frequency amplitude exceeding the historical average by 30%. Finally, this processing step generates a set of multi-dimensional periodic spectral feature vectors for each window, including the main frequency peak, spectral entropy, wavelet energy at various scales, and characteristic frequency amplitude. These vectors serve as the channel feature representation output, demonstrating abnormal spectral changes with different patterns in various precursors such as early micro-cracks and uneven wear in drive motor bearings. This provides high-dimensional frequency domain feature inputs for subsequent spatiotemporal indexing of node representation, differentiation of multi-label abnormal patterns, and embedding of graph network nodes.

[0099] S3.4: Combining the periodic spectrum features obtained in S3.3 with the geographical location information of the equipment parts and the operating condition labels of the corresponding sampling windows, a multi-dimensional feature vector containing dimensional information is generated through a feature splicing strategy. This enables the multi-level fusion of spatiotemporal indexes and multimodal features corresponding to each window / channel, providing comprehensive feature support for the spatiotemporal embedding of node representations.

[0100] Multidimensional feature vectors are generated from window-level data after periodic spectrum feature extraction, geographical location information of equipment parts, and operating condition labels through feature splicing and multi-level fusion strategies.

[0101] The feature concatenation method is used to combine the frequency domain periodic spectrum feature vector F spec (Including peak frequency, spectral entropy, bandwidth energy, wavelet coefficients, etc.), time-domain eigenvector F time Statistical eigenvector F st Standardize alignment at the window level by channel.

[0102] Furthermore, by using an embedded feature encoding method, the discrete geographical location information L of the equipment parts is... loc With operating condition label C cond One-hot encoding or label embedding mapping are performed respectively to form positional feature vectors E. loc With the feature vector E of the working condition label cond And concatenate it with the aforementioned time-domain and frequency-domain feature vectors.

[0103] Furthermore, a feature normalization algorithm (parameters: mean standardization, interval scaling) is adopted to unify the scale of each component of the concatenated feature vector, avoid weight imbalance caused by different feature sources, and improve the consistency of feature fusion.

[0104] Furthermore, through a feature hierarchical combination strategy, window-level features are spliced ​​in multiple levels according to "physical quantity level - location spatial level - operating condition label level" to generate a high-dimensional multimodal feature vector V that includes time-domain / frequency-domain / statistical features of each signal channel, spatial geographic location index, and operating condition label encoding. node This enables the complete spatiotemporal and multimodal information encapsulation for each window / channel.

[0105] Through the above processing method, the periodic spectrum features, temporal dynamic change features, statistical features, and spatial / operating condition labels are unified into a structured multidimensional feature vector according to the splicing and fusion rules. This enables the input of multimodal and spatiotemporal index features for each sampling window / channel, providing comprehensive feature support for the spatiotemporal embedding and high-dimensional anomaly pattern differentiation of subsequent node representations.

[0106] For example, in the three-axis vibration channel of the main motor in the crushing section of an industrial solid waste sorting system, the sampling window length is 4096 points (2 seconds). The peak frequency of the extracted periodic spectrum features is 74Hz, the energy ratio of low frequency to mid-high frequency is 1.8, the spectral entropy is 1.23, and the energy of the second wavelet layer is 0.023. After feature concatenation, the time-domain feature vector (extreme point density, zero crossover rate, impulse factor, etc.) is [0.18, 67, 2.72]. The location geographic information number is 5, and the corresponding working condition label is "high load". One-hot encoding is used to map the geographic location [0,0,0,0,1,0,0] and the working condition label [0,1,0]. After multi-level concatenation, the final window-level node multidimensional feature vector V is obtained. node = [0.18,67,2.72,74,1.23,0.023,0,0,0,0,1,0,0,0,1,0], a total of 15 dimensions. This feature vector is scaled to the [0,1] interval using a normalization method and used as input to the spatiotemporal evolution graph network as window nodes. Through this implementation, consistent representation and fusion coding of the feature space can be achieved for multiple different locations, various operating conditions, and multi-channel signals. Verification shows that its node discrimination rate in novel fault precursor identification tasks is more than 20% higher than that of single statistical or frequency domain features.

[0107] S3.5: Based on the multidimensional feature vectors obtained in S3.4, feature normalization and dimensionality reduction mapping methods (such as principal component analysis (PCA) or linear discriminant analysis (LDA)) are applied to unify the feature scales and improve representativeness to obtain standardized multidimensional feature vectors. The standardized multidimensional feature vectors are then transformed into feature node representations that can be used for the initialization of graph network nodes, and the node input dataset for subsequent spatiotemporal evolution graph networks is output.

[0108] S4: Based on the multi-dimensional spatiotemporal feature node representation, a spatiotemporal evolution graph network is constructed using a preset time sliding window as a unit. In the network, nodes within each window are connected by edges according to temporal sequence and spatial neighborhood relationships to characterize the evolution path of working conditions and potential causal relationships, specifically including:

[0109] This step, based on the spatiotemporal evolution features of nodes, outputs the anomaly probability distribution of each node under the anomaly pattern type and operating condition label through a multi-label classification head. Then, a generative adversarial network is used to generate and discriminate the node's evolution trajectory, thereby improving the ability to identify complex, overlapping, and unknown anomalies. As the core of anomaly pattern recognition, this step achieves multi-label subdivision discrimination under varying operating condition complexity, integrating generative adversarial and spatiotemporal evolution features to effectively improve the model's adaptability and generalization warning capabilities for early warning anomaly patterns.

[0110] S4.1: Based on multi-dimensional spatiotemporal feature node representation, obtain the node set within each time sliding window, and use the node's sampling time, geographical location, and other attributes as input conditions to complete the structured window division of the multi-channel working condition feature sequence through sliding window segmentation technology to obtain the time-series distributed node set.

[0111] The input is a set of standardized multidimensional feature vectors output by the S3.5 sub-step. Each vector represents the multi-channel operating condition signal characteristics under a single sampling window, and is accompanied by clear geographical location information and operating condition labels.

[0112] The sliding window segmentation technique (parameters: window length W, sliding step size S) is adopted to segment the multi-channel operating condition feature sequence along the sampling time axis with a fixed step size S, thereby realizing the structured window division in the time dimension.

[0113] Furthermore, through window partitioning mapping, within each time sliding window, feature nodes from different signal channels and different device parts are grouped and assigned to the current window, forming a node set N. t , where t is the index of the starting time of the sliding window.

[0114] Furthermore, based on the geographical location information and working condition labels carried by the feature vector of each node, a node filtering and mapping algorithm is used to process the node set N. t Perform attribute reorganization to ensure that nodes in each set are compatible and consistent in terms of temporal, spatial, and operational attributes.

[0115] Furthermore, record the set N of each window-level node. t The structured index includes the sampling time interval [t, t+W-1] and the geolocation code L. loc Operating condition label C cond The node IDs are used to form a standardized list of node tuples, laying the structural foundation for subsequent spatial neighborhood inference and graph edge establishment.

[0116] By using sliding window segmentation and node set archiving, the original multi-channel continuous spatiotemporal feature sequence is organized into a discrete temporal distribution of node sets, realizing an explicit node representation of operating condition data under a three-dimensional index of time, space, and operating condition.

[0117] For example, during the signal acquisition process of the main drive system of industrial solid waste treatment equipment, the standardized feature vector sequences of sensors such as triaxial vibration, temperature, and current are input into the sliding window segmentation module. The window length W = 4096 points (2 seconds) and the step size S = 1024 points (0.5 seconds) are set to segment the 24-hour acquisition sequence. Within each time window, the corresponding multi-channel node features from multiple components such as the main drive, feeding mechanism, and sorting module are automatically aggregated, outputting a node set N. t A total of 172,800 groups were generated. For each group of nodes, the geographic location number (e.g., number 3: main drive shaft, number 7: sorting belt) and operating condition label (e.g., "high load", "changing operating condition") were used to complete attribute grouping. Ultimately, a standard node tuple set table containing {sampling time interval, geographic location code, operating condition label, node feature vector} was formed, providing structured input for subsequent spatial relationship inference, temporal evolution edge construction, and complex spatiotemporal graph network generation. This process effectively supports multi-machine parallel node processing, flexible switching between multiple operating conditions, and provides efficient navigation for the accurate extraction of high-dimensional operating condition anomaly trends and multi-label patterns.

[0118] S4.2: For a time-series distributed set of nodes, a spatial neighborhood relationship inference algorithm is used to establish spatial neighborhood connections between nodes based on geographical location information, upstream and downstream topological relationships of equipment parts, and operating condition labels, thereby obtaining a preliminary spatial structure map.

[0119] S4.3: Based on the preliminary spatial structure graph, execute the time series dynamic edge construction algorithm to establish the node time evolution edge of the same equipment part and the same node type under adjacent time sliding windows through the time sequence logic, so as to output a spatial-temporal hybrid structure graph containing the time evolution relationship.

[0120] S4.4: For the space-time hybrid structure graph, apply the causal modeling method of operating condition influence, take the known operating condition events between nodes (such as load change, equipment switching, etc.) as triggering factors, supplement and construct causal edges for specific operating condition events, and enhance the causal expression capability of the operating condition evolution of the graph network.

[0121] S4.5: Taking the structured graph containing spatial neighborhood edges, temporal evolution edges, and working condition causal edges as input, the spatiotemporal correlation weight optimization algorithm is used to assign different connection weights to the edges of different types in the graph to achieve dynamic weighting, so as to optimize the final spatiotemporal evolution graph network and output a high-dimensional spatiotemporal evolution graph data structure that meets the needs of subsequent dynamic graph neural network processing.

[0122] S5: Input the spatiotemporal evolution graph network into the dynamic graph neural network to learn the dynamic correlation and feature evolution laws between nodes, and obtain a spatiotemporal evolution feature description of nodes containing labels of different parts and working conditions, compatible with multiple working conditions and heterogeneous signal distributions, specifically including:

[0123] S5.1: For the spatiotemporal evolution graph network established based on the multidimensional spatiotemporal feature node representation, a dynamic graph neural network is used for input processing to structurally map the spatiotemporal feature node representation into node embedding representation, thereby obtaining the original node embedding features that can retain the window temporal sequence and spatial neighborhood topology, providing an input basis for subsequent correlation modeling and evolution law analysis.

[0124] S5.2: Based on the original embedding features of nodes, the temporal embedding update mechanism in the dynamic graph neural network is used to perform dynamic temporal information fusion on the node sequence, capture the dynamic evolution characteristics of nodes in the time dimension, and realize the dynamic perception of the spatiotemporal evolution characteristics of nodes to adapt to multi-condition switching and signal fluctuations.

[0125] S5.3: Based on the dynamic spatiotemporal evolution characteristics, the spatial message passing mechanism of the dynamic graph neural network is used to aggregate the information of each node in the spatial neighborhood (including nodes in different parts of the same equipment and nodes in similar working conditions of different equipment), optimize the representation ability of nodes in the spatial distribution of multiple parts and the spatial heterogeneity of different equipment, thereby improving the expression accuracy of the spatial correlation characteristics of multiple parts.

[0126] S5.4: Combining the dynamic spatiotemporal evolution characteristics of nodes, the label condition regularization module of the dynamic graph neural network introduces the corresponding working condition label and location label information of each node, realizes the compatibility of node embedded features with multiple working conditions and heterogeneous signal distribution, enables node feature representation to have label discrimination ability and improves the discriminability of subsequent multi-label anomaly discrimination.

[0127] The input consists of node dynamic spatiotemporal evolution feature data processed by spatial message aggregation. Each node feature integrates micro-variable signals of operating conditions, spatial correlation features, and spatial coordinates and operating condition codes of physical components such as the main drive shaft, feeding mechanism, and sorting module. A label conditional regularization method based on a dynamic graph neural network (parameter settings: label feature dimension 16, regularization loss adjustment factor 0.05) is used to enhance the discriminative power of signals from multiple operating conditions and multiple components in the embedded space.

[0128] For example, in the diagnostic scenario of a multi-parallel system of industrial solid waste treatment equipment, three parts are set up, including the main drive shaft. 256-dimensional dynamic spatiotemporal evolution features are collected for each part. Using three operating condition labels ("normal," "high load," and "equipment switching") and three part labels, one-hot encoding is used to form 16-dimensional label features. After node embedding, 272-dimensional combined features are obtained. The training set constructs 5000 pairs of nodes with consistent labels and 5000 pairs of nodes with different labels. The loss function parameter is set with an adjustment factor of 0.05, and training is performed for 60 epochs. After model convergence, within the embedding space, the average distance between nodes with consistent labels is 0.14, and that between nodes with different labels is 2.01. The validation set shows a 17% improvement in the accuracy of multi-label anomaly discrimination and a 10% decrease in the false positive rate for cross-operating conditions. The final output is the conditional embedding features of node labels, providing basic data for subsequent multi-label anomaly probability estimation and anomaly trajectory aggregation, enabling effective identification of early anomaly patterns under complex and heterogeneous operating conditions.

[0129] A tag information embedding algorithm is employed to input the node's operating condition tag and operating location tag into a one-hot encoding module, outputting a tag feature vector. Furthermore, this tag feature is concatenated with dynamic spatiotemporal evolution features at the feature level to form tag-conditional dynamic spatiotemporal node features, thereby improving the distinguishability of identifying complex precursor trend features.

[0130] S5.5: The spatiotemporal evolution features of nodes after spatial message aggregation and label regularization are subjected to high-order feature abstraction through the global induction layer of the dynamic graph neural network. The final output is a spatiotemporal evolution feature description of nodes containing fused information of multiple operating parts and multiple working conditions labels, which serves as the core input for subsequent multi-label anomaly probability distribution calculation and anomaly precursor trajectory extraction.

[0131] S6: Based on the spatiotemporal evolution feature description of nodes, the abnormal probability distribution of each node under the abnormal mode type and working condition label is output through a multi-label classification head. At the same time, a generative adversarial network is used to generate and discriminate the evolution trajectory of the nodes to improve the ability to identify complex overlapping and unknown anomalies. Specifically, it includes:

[0132] S6.1: Based on the spatiotemporal evolution feature description of nodes, load working condition labels and abnormal mode labels, perform multi-label vector initialization for each node to provide multi-dimensional supervision signals and clarify the task objectives of classification and discrimination.

[0133] S6.2: Using a multi-label classification head to describe the spatiotemporal evolution characteristics of the signal input nodes, a multi-task loss function is used to calculate the abnormal probability distribution of nodes under abnormal mode type and working condition label, so as to obtain fine-grained multi-label abnormal probability features.

[0134] S6.3: Based on the spatiotemporal evolution characteristics and anomaly probability characteristics of nodes, the historical trajectory and future predicted trajectory of nodes are input into the generator module of the generative adversarial network to generate potential abnormal evolution trajectories and uncover hidden high-order dynamic evolution laws.

[0135] S6.4: Input the potential abnormal evolution trajectory generated by the generative adversarial network and the actual node evolution trajectory into the discriminator module. Utilize the discriminant loss function to optimize the discriminator's ability to distinguish between real and generated trajectories, thereby enhancing the robustness of abnormal trajectory recognition.

[0136] S6.5: Integrate the multi-label anomaly probability features of nodes with the output results of the generative discriminative adversarial network. Through a joint loss optimization strategy, adjust the parameters of the multi-label classification head and the generative adversarial branch to achieve a simultaneous improvement in discriminative and generalization capabilities. The final output can cover the multi-label probability distribution of node anomalies that covers complex overlaps and unknown anomalies.

[0137] S7: Based on the obtained anomaly probability distribution and node evolution trajectory, perform precursor trend weighted aggregation on multiple types of early minor anomaly signals, introduce spatiotemporal neighborhood correlation information before and after, and extract traceable anomaly precursor trajectory data, specifically including:

[0138] S7.1: Synchronously acquire the node evolution trajectory and corresponding anomaly probability distribution to form a node anomaly state set containing temporal and probability information, providing basic data for subsequent weighted aggregation of precursor trends.

[0139] S7.2: Based on the node abnormal state set, a weighted sliding window aggregation technique is used to perform probability weighting processing on the early small abnormal signals of multiple nodes in the same spatiotemporal evolution graph network to obtain a preliminary node precursor trend aggregation vector.

[0140] S7.3: Utilize the spatiotemporal neighborhood dependency relationship between nodes to perform spatiotemporal dependency feature enhancement on the node precursor trend aggregation vector, and integrate the attributes, anomaly probability and trajectory evolution information of neighboring nodes to improve the spatiotemporal consistency of anomalous signals.

[0141] S7.4: Based on the enhanced precursor trend aggregation vector, a multi-class anomaly probability gating mechanism is adopted to perform multi-label weight normalization on different types of early anomaly signals, generating anomaly precursor aggregation state descriptions with clear type differentiation and low interference.

[0142] S7.5: Using the aggregated state description of abnormal precursors as trajectory nodes, and through spatiotemporal dynamic trajectory sorting and tracing algorithms, abnormal precursor trajectory data with traceability and continuity are extracted, providing data support for subsequent early warning decisions and result visualization.

[0143] Step S8: Determine whether the overall anomaly probability of the abnormal precursor trajectory data exceeds a preset warning threshold. If it does, generate a corresponding multi-label early anomaly warning instruction; otherwise, continue dynamic monitoring and adaptive optimization of the graph network parameter group, specifically including:

[0144] S8.1: Obtain the multi-label anomaly probability distribution result of the anomaly precursor trajectory data from the previous step, and use it as the input condition for anomaly probability determination; based on the anomaly probability distribution data, use a preset probability aggregation algorithm (such as weighted normalization, confidence fusion, etc.) to calculate the comprehensive anomaly probability index of the anomaly precursor trajectory data for subsequent early warning determination.

[0145] S8.2: Based on the calculated comprehensive anomaly probability index, the internally set adaptive warning threshold parameter is applied to compare and analyze the comprehensive anomaly probability and the warning threshold, determine whether there are early anomaly signals that exceed the set safety limit, and output the determination result as the early anomaly warning trigger condition.

[0146] S8.3: If the judgment result indicates that the comprehensive anomaly probability index exceeds the adaptive early warning threshold, then based on the multi-label anomaly probability distribution and anomaly precursor trajectory data, a multi-label early anomaly warning instruction matching the equipment location, operating condition label and anomaly type is generated, output to the next level decision module, and the anomaly source tracing and handling process is initiated.

[0147] S8.4: If the judgment result does not reach the adaptive warning threshold, the abnormal probability distribution and trajectory features of this round will be used as feedback input to update the parameter set of the dynamic graph neural network model in real time. Through the adaptive online optimization mechanism, the model's discrimination performance for new abnormalities or boundary samples will be improved.

[0148] S8.5: Perform dynamic statistical analysis on the historical judgment results of the comprehensive anomaly probability index, and use time series autocorrelation technology and trend detection algorithm to periodically optimize the adaptive early warning threshold parameters, so as to achieve a dynamic balance between early warning sensitivity and false alarm rate and ensure the stability of anomaly identification under different operating conditions.

[0149] S9: The multi-tag early anomaly warning command and anomaly precursor trajectory data are visualized and then pushed to the remote operation and maintenance terminal, automatically triggering targeted safety measures or generating operation and maintenance work orders to achieve closed-loop feedback and timely intervention, specifically including:

[0150] S9.1: For multi-label early warning instructions and anomalous precursor trajectory data, interactive data parsing and map generation are performed based on multi-dimensional working condition visualization algorithms to obtain dynamic visualized working condition maps that include anomalous type distribution, anomalous probability weights, and historical precursor trajectories.

[0151] S9.2: Based on the dynamic visualization of the working condition map, the working condition digital twin template is used to spatially map the anomaly type, geographical location, time window and equipment location information, and generate a visual interface of equipment distribution with anomaly markers, which makes it easier for maintenance personnel to locate the specific anomaly occurrence point and trend.

[0152] S9.3: The generated anomaly distribution visualization interface and precursor trajectory data are encrypted and packaged using remote operation and maintenance communication protocols (such as MQTT, OPC UA, etc.) and pushed to the remote operation and maintenance terminal to achieve highly reliable long-distance distribution of anomaly information.

[0153] S9.4: On the remote operation and maintenance terminal, the received abnormal warning instructions are automatically parsed based on the distributed operating condition abnormality management system. Relying on the multiple tags of abnormality types and location data, the system automatically matches the set safety measure templates or operation and maintenance response strategies to generate targeted automated safety measure instructions or operation and maintenance work orders.

[0154] S9.5: For automatically generated safety measures instructions / maintenance work orders, a synchronous closed-loop feedback mechanism is implemented to transmit the processing progress, response execution results, and work order status back to the intelligent diagnostic system in real time through the working condition information feedback module, thereby realizing a closed-loop technology for the entire process of intelligent fault early warning, response, and backtracking.

[0155] For those skilled in the art, various other corresponding changes and modifications can be made based on the technical solutions and concepts described above, and all such changes and modifications should fall within the protection scope of the claims of this invention.

[0156] Unless otherwise defined, the technical or scientific terms used herein shall have the ordinary meaning as understood by one of ordinary skill in the art to which this application pertains. The terms “first,” “second,” “third,” and similar terms used in this patent application specification and claims do not indicate any order, quantity, or importance, but are merely used to distinguish different components. Similarly, the terms “an” or “a” and similar terms do not indicate a quantity limitation, but rather indicate the presence of at least one. The terms “comprising” or “including” and similar terms mean that the elements or objects preceding “comprising” or “including” encompass the elements or objects listed following “comprising” or “including” and their equivalents, and do not exclude other elements or objects. The “multiple” mentioned in the embodiments of this application refers to two or more. A and / or B indicate three possibilities: A; B; and A and B.

[0157] The above description is merely an exemplary embodiment of this application, but the scope of protection of this application is not limited thereto. Any person skilled in the art can easily conceive of various equivalent modifications or substitutions within the technical scope disclosed in this application, and such modifications or substitutions should all be covered within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

Claims

1. A remote diagnostic method for industrial solid waste treatment equipment, specifically including the following steps: S1: Collect multi-source real-time operating condition signals of industrial solid waste treatment equipment at multiple operating locations, and annotate the specific geographical location and equipment operating condition labels of each sampling channel to establish a multi-source operating condition raw dataset. S2: Perform denoising and time-series alignment processing on the original multi-source operating condition dataset to obtain standardized, multi-channel time-series aligned operating condition data; S3: Based on standardized, multi-channel time-series aligned working condition data, calculate the statistics, time-domain features and frequency-domain spectral features of each channel within each sampling window, and generate multi-dimensional spatiotemporal feature node representations by combining the corresponding geographical location and working condition labels. S4: Based on the multidimensional spatiotemporal feature node representation, a spatiotemporal evolution graph network is constructed using a preset time sliding window as a unit. In the network, nodes within each window are connected by edges according to temporal sequence and spatial neighborhood relationship. S5: Input the spatiotemporal evolution graph network into the dynamic graph neural network to learn the dynamic correlation and feature evolution law between nodes, and obtain a spatiotemporal evolution feature description of nodes containing labels of different parts and working conditions. S6: Based on the spatiotemporal evolution feature description of nodes, the abnormal probability distribution of each node under the abnormal mode type and working condition label is output through a multi-label classification head. At the same time, the evolution trajectory of the nodes is generated and discriminated against adversarially using a generative adversarial network. S7: Obtain the node abnormal state set based on the obtained abnormal probability distribution and node evolution trajectory. Based on the node abnormal state set, perform precursor trend weighted aggregation on multiple types of early small abnormal signals to obtain the node precursor trend aggregation vector. Introduce the spatiotemporal neighborhood correlation information before and after to perform spatiotemporal dependency feature enhancement on the node precursor trend aggregation vector. Generate an abnormal precursor aggregation state description based on the enhanced precursor trend aggregation vector. Extract traceable abnormal precursor trajectory data based on the abnormal precursor aggregation state description. S8: Determine whether the overall anomaly probability of the abnormal precursor trajectory data exceeds the preset warning threshold. If it does, generate a corresponding multi-label early anomaly warning instruction.

2. The remote diagnostic method for industrial solid waste treatment equipment according to claim 1, characterized in that: Following step S8, the following is also included: S9: The multi-label early anomaly warning command and anomaly precursor trajectory data are visualized and then pushed to the remote operation and maintenance terminal, and targeted safety measures or operation and maintenance work orders are automatically triggered to achieve closed-loop feedback and timely intervention.

3. The remote diagnostic method for industrial solid waste treatment equipment according to claim 1, characterized in that: The multi-source real-time operating condition signals in step S1 include: vibration, temperature, flow rate, and current data.

4. The remote diagnostic method for industrial solid waste treatment equipment according to claim 1, characterized in that: In the configuration of multiple types of sensor nodes in the key operating parts of industrial solid waste treatment equipment, the multiple types of sensors include vibration sensors, temperature sensors, flow sensors and current sensors.

5. The remote diagnostic method for industrial solid waste treatment equipment according to claim 1, characterized in that: The equipment condition labels in step S1 include: operating mode, load characteristics, and operating parameter labels.

6. The remote diagnostic method for industrial solid waste treatment equipment according to claim 1, characterized in that: Step S3 specifically includes: For each signal channel in the standardized, multi-channel time-aligned operating condition data, the interval is divided according to the sampling time window, and the mean, variance, skewness, and kurtosis statistical characteristics within the window are obtained based on the segmented statistical method. Based on the generated statistical features, time-domain features are mined for a single signal channel in each sampling window to further obtain key time-domain feature parameters of the signal, such as extreme points, zero crossover rate, and impulse factor. Using the obtained time-domain characteristic parameters as input, a multi-resolution spectrum analysis algorithm is applied to the data within each signal channel window to obtain the main frequency peak, frequency band energy distribution, spectral entropy, and characteristic frequency amplitude periodic spectrum characteristics. By combining the obtained periodic spectrum features with the geographical location information of the equipment parts and the working condition labels of the corresponding sampling window, a multidimensional feature vector containing dimensional information is generated through a feature splicing strategy. Based on the obtained multidimensional feature vectors, the scales of each feature are unified and the representativeness is improved to obtain standardized multidimensional feature vectors. The standardized multidimensional feature vectors are transformed into feature node representations that can be used for the initialization of graph network nodes, and the node input dataset for subsequent spatiotemporal evolution graph networks is output.

7. The remote diagnostic method for industrial solid waste treatment equipment according to claim 1, characterized in that: Step S4 specifically includes: Based on multi-dimensional spatiotemporal feature node representation, the node set within each time sliding window is obtained. Using the sampling time and geographical location attributes of the nodes as input conditions, the multi-channel working condition feature sequence is divided into structured windows through sliding window segmentation technology to obtain a time-series distributed node set. For a time-series distributed set of nodes, spatial neighbor-to-neighbor connections are established between nodes based on geographical location information, upstream and downstream topological relationships of equipment parts, and operating condition labels to obtain a preliminary spatial structure diagram. Based on the preliminary spatial structure diagram, nodes of the same equipment part and the same node type under adjacent time sliding windows are used to establish node temporal evolution edges through temporal sequence logic, so as to output a spatial-temporal hybrid structure diagram containing temporal evolution relationships. For a space-time hybrid structure graph, known working condition events between nodes are used as triggering factors to supplement the construction of causal edges for known working condition events between nodes, thereby enhancing the causal expression capability of the working condition evolution of the graph network. The structured graph containing spatial neighborhood edges, temporal evolution edges, and operational causal edges is used as input. Different connection weights are assigned to the various types of edges in the graph to achieve dynamic weighting and output a high-dimensional spatiotemporal evolution graph data structure that meets the requirements of subsequent dynamic graph neural network processing.

8. The remote diagnostic method for industrial solid waste treatment equipment according to claim 1, characterized in that: Step S5 specifically includes: For the spatiotemporal evolution graph network established based on the multidimensional spatiotemporal feature node representation, the spatiotemporal feature node representation is structurally mapped into node embedding representation, thereby obtaining the original node embedding features that can preserve the window temporal sequence and spatial neighborhood topology. Based on the original embedding features of nodes, dynamic temporal information fusion is performed on node sequences to capture the dynamic evolution characteristics of nodes in the time dimension; Based on the dynamic spatiotemporal evolution characteristics, the information of each node within its spatial neighborhood is aggregated; By combining the dynamic spatiotemporal evolution characteristics of nodes, the label condition regularization module of the dynamic graph neural network introduces the corresponding working condition label and location label information of each node, so as to realize the compatibility of node embedding features with multiple working conditions and heterogeneous signal distribution, enabling node feature representation to have label discrimination ability and improving the discriminativeness of subsequent multi-label anomaly discrimination. The spatiotemporal evolution features of nodes, after spatial message aggregation and label regularization, are subjected to high-order feature abstraction through the global induction layer of a dynamic graph neural network, outputting a final spatiotemporal evolution feature description that includes fused information of multiple operating parts and multiple working conditions labels.

Images