Scrap processing equipment failure prediction and health management system

Abstract

This invention relates to the field of intelligent operation and maintenance and fault diagnosis technology for industrial equipment, specifically a fault prediction and health management system for scrap steel processing equipment. The system includes: a data acquisition and preprocessing module for acquiring multi-source sensor data within a sliding time window and calculating a data quality index; a feature extraction module for extracting temporal and spatial feature vectors and aligning them based on timestamps; a feature synthesis module for determining fusion weights based on the data quality index and generating fusion features; a fault diagnosis and reasoning module for determining fault types and reasoning fault propagation paths in the fault logic diagram of the scrap steel equipment; and a coordination and update module for determining changes in operating conditions based on the vibration feature distribution under no-load idling conditions and triggering automatic updates of diagnostic model parameters. This invention can suppress low-quality data interference, improve fault diagnosis accuracy, causal interpretability, and adaptability to equipment aging.

Description

Technical Field

[0001] This invention relates to the field of intelligent operation and maintenance and fault diagnosis technology for industrial equipment, specifically a fault prediction and health management system for scrap steel processing equipment. Background Technology

[0002] During the crushing, shearing, and packaging processes, scrap steel processing equipment is subjected to complex conditions of high load, strong impact, heavy dust, and strong electromagnetic interference for extended periods. The equipment's operating status is characterized by multi-source heterogeneity, strong time-varying nature, and complex fault propagation chains. Therefore, accurate fault prediction and health management during equipment operation are crucial for ensuring production continuity, reducing equipment failure rates, and improving equipment operational stability. Existing fault monitoring and diagnosis methods for scrap steel processing equipment have many problems. For example, they often rely on single vibration, current, or temperature signals for independent analysis, making it difficult to fully reflect the coupling relationship between the equipment's physical condition and the feeding conditions. When sensor signals are missing, there is noise interference, or local distortion, the diagnostic results are prone to misjudgment or omission. At the same time, existing methods can usually only output isolated fault types, making it difficult to trace the root cause and propagation path of the fault. Furthermore, after changes in operating conditions caused by equipment tool wear, component aging, etc., the accuracy of static models tends to decrease significantly, failing to continuously meet the application requirements of scrap steel processing sites for high reliability, interpretability, and adaptive health management. Summary of the Invention

[0003] The purpose of this invention is to provide a fault prediction and health management system for scrap steel processing equipment, solving the following technical problems: It avoids false alarms caused by equipment downtime due to distortion or failure of single-point sensors in harsh environments, and can clearly trace the root cause of the fault to directly guide on-site maintenance operations. At the same time, it solves the problem that the accuracy of static diagnostic models will decrease significantly in the later stages of equipment operation due to physical aging.

[0004] The objective of this invention can be achieved through the following technical solutions: The scrap steel processing equipment fault prediction and health management system includes: The data acquisition and preprocessing module connects to multiple sensors mounted on the scrap steel processing equipment via a communication interface to acquire data about the scrap steel processing equipment within a sliding time window. Source sensor data and calculate the data quality index of multi-source sensor data; The feature extraction module is used to receive multi-source sensor data, extract the temporal feature vector and spatial feature vector of the multi-source sensor data respectively, and align the temporal feature vector and spatial feature vector within the same sliding time window based on the timestamp. The feature synthesis module receives the aligned temporal feature vector and spatial feature vector, and determines the fusion weight based on the data quality index to fuse the temporal feature vector and spatial feature vector to generate fused features; The fault diagnosis and reasoning module is used to input the fused features into the preset diagnostic model to determine the fault type, and to use the fault type, temporal feature vector and spatial feature vector as input to reason in the pre-built fault logic relationship diagram of scrap steel equipment to determine the fault propagation path. The coordination and update module is used to determine whether a change in operating conditions has occurred based on the vibration characteristic distribution under no-load idling conditions, and to trigger automatic parameter updates of the diagnostic model if a change in operating conditions is determined to have occurred.

[0005] Furthermore, the data acquisition and preprocessing module includes: The data acquisition unit is used to acquire image data from industrial cameras, vibration data from high-frequency vibration sensors, current data from current transformers, and temperature data from temperature sensors. It also acquires timestamps and aligns them using a synchronous acquisition card with a unified clock source to obtain multi-source sensor data. The missing data rate calculation unit is used to calculate the missing data rate of multi-source sensor data within the sliding time window. The outlier calculation unit is used to calculate the proportion of outliers to total sampling points within a sliding time window using an outlier detection algorithm, thereby obtaining the data outlier rate. The quality index generation unit is used to calculate the data quality index based on the normalized signal-to-noise ratio, data missing rate, data anomaly rate, and preset corresponding weighting coefficients of the multi-source sensor data.

[0006] Furthermore, the feature extraction module includes: The temporal feature extraction unit is used to process multi-source time-series data using a temporal feature extraction model to extract temporal feature vectors. The multi-source time-series data includes vibration data, current data, and temperature data. The temporal feature extraction model uses multi-layer historical data convolution, and the sampling interval of each layer is increased by a factor. The spatial feature extraction unit is used to process image data using an image processing model, extract the volume and stacking density features of scrap steel, and obtain spatial feature vectors. Alignment units are used to combine temporal feature vectors and spatial feature vectors within the same sliding time window based on timestamps.

[0007] Furthermore, the feature synthesis module includes: The early splicing unit is used to extract the effective value of the current data and perform channel-level splicing of the effective values ​​of the temperature data and the current data before inputting them into the time series feature extraction model. The weight calculation unit is used to calculate the fusion weight of each sensor feature based on the data quality index and temperature data; The feature association unit is used to calculate the association features of temporal feature vectors and spatial feature vectors by using a feature weighting processing method and combining fusion weights to obtain fused features; The decision-making unit is used to vote on the output results of different types of sensor features in the classification decision layer of the diagnostic model.

[0008] Furthermore, the fault diagnosis and reasoning module includes: The combined diagnostic unit is used to process the fused features and output the fault type by using a combined diagnostic model consisting of support vector machine, random forest and neural network model as the diagnostic model; The weight dynamic adjustment unit is used to dynamically adjust the final voting weight of each base classifier based on the historical accuracy of each base classifier over a set number of diagnostic cycles.

[0009] Furthermore, the fault diagnosis and reasoning module also includes: The relational graph construction unit is used to map the physical structure of scrap steel processing equipment into graph nodes and the physical transmission relationship and causal relationship into graph lines, thereby pre-constructing the fault logic relational graph of scrap steel equipment; The node association analysis unit is used to perform node association analysis along the graph lines when a fault type triggers a preset alarm tag, starting from the graph node corresponding to the alarm tag and combining the temporal feature vector and spatial feature vector, and output the fault propagation path and maintenance guidance.

[0010] Furthermore, the collaboration and update module includes: The idle running determination unit is used to determine that the equipment is in an unloaded idle running state when the current data is lower than 20% of the rated value of the main motor of the scrap steel processing equipment and there is no material feeding detected by vision. The distance calculation unit is used to extract the vibration characteristic distribution under the current no-load idling state and calculate the Wasserstein distance between the vibration characteristic distribution under the current no-load idling state and the vibration characteristic distribution under the reference no-load idling state. The change determination unit is used to determine that a change in working conditions has occurred if the Wasserstein distance is detected to be greater than or equal to the change trigger threshold after a preset number of consecutive tests. The change trigger threshold is a set percentile of the Wasserstein distance recorded during the break-in period after the equipment is replaced with a new tool, under all no-load idling conditions.

[0011] Furthermore, the collaboration and update module also includes: The field control unit is used to deploy the forward inference process, after parameter compression, in the field processing gateway for emergency shutdown control. The cloud-based adjustment unit is used to upload feature data and feedback labels to the cloud server when a change in operating conditions is detected, start supplementary training to fine-tune the weights of the neural network model in the diagnostic model, and then send the updated data to the field processing gateway.

[0012] The beneficial effects of this invention are: 1. The scrap steel processing equipment fault prediction and health management system of the present invention connects to industrial cameras, high-frequency vibration sensors, current transformers and temperature sensors through a data acquisition and preprocessing module, and acquires multi-source sensor data within a sliding time window. It calculates a data quality index by combining the signal-to-noise ratio, data missing rate and data anomaly rate, which can comprehensively reflect the true reliability of various sensor data of scrap steel processing equipment under high load, strong dust, strong electromagnetic interference and severe impact environment. Furthermore, based on the data quality index, it suppresses data with severe interference, effectively reducing misjudgment and missed judgment caused by sensor signal missing, noise pollution or local distortion, and improving the stability and diagnostic reliability of the system under harsh working conditions. 2. The scrap steel processing equipment fault prediction and health management system of the present invention extracts time-series data such as vibration, current and temperature into time-series feature vectors, and extracts feed port image data into spatial feature vectors reflecting scrap steel volume and stacking density. Based on a unified timestamp, the time-series feature vectors and spatial feature vectors within the same sliding time window are aligned and combined. This can fully explore the coupling relationship between the equipment's bottom-level operating status and the upper-level feeding conditions. The system can not only identify abnormalities in the equipment itself, but also identify complex load changes caused by feeding factors such as scrap steel with geometric dimensions exceeding preset thresholds and changes in stacking density. This solves the problem that existing methods rely on independent analysis of a single signal and are difficult to fully reflect the correlation between equipment status and feeding conditions. 3. The scrap steel processing equipment fault prediction and health management system of the present invention determines the fusion weight of temporal feature vector and spatial feature vector based on data quality index, and generates fusion features through feature weighting processing method. In the classification decision layer, it combines the output results of different types of sensor features for voting. It can dynamically adjust the contribution ratio of different modal features in the fusion process according to the real-time physical quality of various sensor data. When a single sensor is blocked by dust, impact distortion or short-term failure, it automatically enhances the role of other high-quality sensor data, avoids the problem of insufficient adaptability of fixed weight fusion method in complex field, and further improves the fault identification fault tolerance, accuracy and anti-interference ability. 4. The scrap steel processing equipment fault prediction and health management system of the present invention uses a combined diagnostic model composed of support vector machine, random forest and neural network model to determine the fault type by integrating feature input. The final voting weight is dynamically adjusted according to the historical accuracy of each base classifier in the past set diagnostic period. It can combine the advantages of different diagnostic algorithms in small sample classification, high-dimensional feature processing and complex nonlinear relationship modeling, so that the system can maintain high fault identification accuracy even when the scrap steel processing equipment is running for a long time and the working conditions are constantly changing. At the same time, the system also inputs the fault type, temporal feature vector and spatial feature vector into the pre-constructed fault logic relationship diagram of scrap steel equipment for reasoning, and outputs the fault propagation path and maintenance guidance. Thus, the diagnostic results are improved from isolated fault labels to propagation analysis results with causal interpretability, which can directly trace the root cause of the fault and guide precise on-site maintenance. 5. The scrap steel processing equipment fault prediction and health management system of the present invention determines that the equipment is in an unloaded idling state when the current is lower than the rated value of the main motor and there is no material feeding. It extracts the vibration feature distribution in the idling state and calculates the distance between it and the vibration feature distribution in the benchmark unloaded idling state. When the change trigger threshold is exceeded multiple times in a row, the system determines the change in working condition and triggers the automatic update of the diagnostic model parameters. It can accurately identify the benchmark state drift caused by tool wear, liner wear, or mechanical component aging, and solve the problem of significant decrease in the accuracy of the static model in the later stage of equipment operation. Combined with the compressed model in the on-site processing gateway to realize emergency shutdown control, and the cloud server to supplement the training and update the weights of the diagnostic model, the system has both rapid on-site response capability and long-term adaptive evolution capability, meeting the application requirements of scrap steel processing site for high reliability, interpretability and adaptive health management. Attached Figure Description

[0013] Other features, objects, and advantages of the invention will become more apparent from the following detailed description of non-limiting embodiments with reference to the accompanying drawings: Figure 1 This is a schematic diagram of the modules of the scrap steel processing equipment fault prediction and health management system provided in the embodiments of this application. Detailed Implementation

[0014] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0015] Please see Figure 1The scrap steel processing equipment fault prediction and health management system includes: a data acquisition and preprocessing module, which connects to multiple sensors arranged on the scrap steel processing equipment through a communication interface to acquire multi-source sensor data of the scrap steel processing equipment within a sliding time window and calculate the data quality index of the multi-source sensor data; The feature extraction module is used to receive multi-source sensor data, extract the temporal feature vector and spatial feature vector of the multi-source sensor data respectively, and align the temporal feature vector and spatial feature vector within the same sliding time window based on the timestamp. The feature synthesis module receives the aligned temporal feature vector and spatial feature vector, and determines the fusion weight based on the data quality index to fuse the temporal feature vector and spatial feature vector to generate fused features; The fault diagnosis and reasoning module is used to input the fused features into the preset diagnostic model to determine the fault type, and to use the fault type, temporal feature vector and spatial feature vector as input to reason in the pre-built fault logic relationship diagram of scrap steel equipment to determine the fault propagation path. The coordination and update module is used to determine whether a change in operating conditions has occurred based on the vibration characteristic distribution under no-load idling conditions, and to trigger automatic parameter updates of the diagnostic model if a change in operating conditions is determined to have occurred.

[0016] Data acquisition and preprocessing module; This module connects to multiple sensors arranged on the scrap steel processing equipment through a communication interface to acquire multi-source sensor data of the scrap steel processing equipment within a sliding time window; Specifically, the sliding time window is set to a duration of 2 seconds, and multi-source sensor data within the window is extracted. The data quality index is calculated for the multi-source sensor data within each sliding time window. The specific operation for calculating the data quality index is as follows: comprehensively consider the signal-to-noise ratio, data missing rate, and data anomaly rate of the multi-source sensor data within the time window, and sum them together with preset weighting coefficients. Scrap steel processing equipment operates in a harsh physical environment with high load, strong dust, heavy electromagnetic interference and severe impact vibration for a long time. The probability of the sensor losing signal or abnormal noise is higher than the preset threshold. The data quality index reflects the reliability of multi-source sensor data within the current time window; a higher data quality index indicates that the data acquired by the sensor is more accurate and reliable, while a lower data quality index indicates that the sensor signal is affected by interference from the field environment that exceeds the interference tolerance or that there are missing segments of continuous time. At the scrap steel shredder processing site, the sliding time window is set to 2 seconds, which corresponds to a complete cutting cycle of about 30 to 50 rotations of the spindle. When the operator starts the scrap steel shredder, the data acquisition and preprocessing module acquires various sensor data such as vibration, current, temperature and vision within this 2-second period. The system calculates that the current data quality index of the spindle vibration sensor is 0.85, indicating that the vibration data quality is good. When a large amount of metal dust is stirred up on site and blocks the industrial camera at the feed inlet, the system calculates that the data missing rate of the visual data increases sharply, and the corresponding visual data quality index drops to 0.15, indicating that the reliability of the visual data is extremely low at this time, and its proportion in subsequent processing needs to be reduced. Feature extraction module: This module receives multi-source sensor data and extracts the temporal feature vector and spatial feature vector of the multi-source sensor data respectively; Specifically, for time-series data such as vibration, current, and temperature, a temporal feature extraction network is used to extract the transient features from the moment the tool contacts the scrap steel to the periodic features during the continuous cutting process, and outputs the temporal feature vector. For visual image data, an image processing network is used to extract features such as the volume and stacking density of scrap steel at the feed inlet, and output spatial feature vectors. Based on timestamps, the temporal feature vectors and spatial feature vectors within the same sliding time window are aligned and combined; there are significant modal differences between the high-frequency spindle vibration signal, low-frequency bearing temperature signal, and abrupt motor current signal during scrap steel processing and the visual image of scrap steel morphology at the feed inlet. By extracting temporal and spatial feature vectors and aligning them based on timestamps, the precise binding of the equipment's underlying operating state and the upper-level feeding material state on the same physical time scale is achieved, solving the problem of difficult alignment of heterogeneous features. In the scenario of a gantry shear processing large scrap steel, the feature extraction module extracts the vibration time sequence feature vector and current time sequence feature vector of the main shaft within a certain 2-second sliding time window, and at the same time extracts the spatial feature vector of scrap steel volume and stacking density captured by the industrial camera at the feed inlet within the same 2-second period. By comparing timestamps, the system combines the temporal feature vectors and spatial feature vectors that are in the same time window to form a joint feature set that includes the spatial features of large scrap steel feeding and the temporal features of sudden drop in spindle speed, providing an aligned data foundation for subsequent cross-modal complex fault identification. Feature synthesis module: This module receives aligned temporal feature vectors and spatial feature vectors, and determines fusion weights based on the data quality index to fuse the temporal feature vectors and spatial feature vectors to generate fused features; Temperature data is acquired to control the smoothness of weight allocation and used as a temperature scalar parameter. Specifically, an exponential function is used in conjunction with the temperature scalar parameter to convert the data quality index corresponding to each sensor into dynamic fusion weights. Based on the calculated fusion weights, the temporal feature vector and spatial feature vector are weighted and correlated to generate the final fusion feature. The dynamic fusion weight allocation driven by the data quality index changes the traditional fixed-weight splicing fusion method; the magnitude of the dynamic fusion weight is determined by the real-time physical quality of the sensor data. When a sensor is affected by harsh environmental conditions, causing a decline in data quality, the system automatically reduces the weight of that modality feature in the fused features and instead relies on features extracted by other normal sensors. This ensures that the system can maintain high accuracy even when a single sensor fails; when the scrap baler is operating continuously, the spindle vibration sensor generates a large number of transient change signals due to the severe impact on site, and its data quality index drops sharply. Based on the decreased data quality index, the feature synthesis module calculated that the fusion weight of the vibration feature vector decreased from 0.4 to 0.05. At the same time, since the current sensor and the vision camera were not affected by dust and impact, the system automatically increased the fusion weight of the current feature vector and the vision feature vector accordingly. The system weighted and fused the temporal feature vector and the spatial feature vector according to the new fusion weight, generating a fusion feature dominated by current and visual features, which successfully avoided false alarms of equipment shutdown caused by severe distortion of vibration signals. Fault diagnosis and reasoning module; This module is used to input fused features into a preset diagnostic model to determine the fault type, and to use the fault type, temporal feature vector and spatial feature vector as input to reason in a pre-built fault logic relationship graph of scrap steel equipment to determine the fault propagation path; Specifically, the diagnostic model first classifies the fused features and outputs specific fault classification labels; The physical structures of the scrap steel processing equipment, such as the feeding conveyor belt, main motor, reducer, and cutter shaft, are mapped as graph nodes, and the physical transmission relationships and causal relationships are mapped as graph lines to construct a fault logic relationship graph for the scrap steel equipment. Starting from the graph node corresponding to the output fault type, and combining the temporal feature vector and spatial feature vector, a probability walk calculation is performed along the graph lines to output the complete fault propagation path; Existing static diagnostic models can only output isolated fault labels and cannot trace the root cause of the fault. By performing node association analysis and reasoning in the fault logic relationship diagram of scrap steel equipment, the propagation process of the fault in the complex mechanical transmission chain of the equipment can be clearly shown. This makes the diagnostic results causally interpretable, directly pointing out which preceding component's physical abnormality caused the cascading failure, thus directly guiding on-site maintenance work; When the scrap steel shredder processes a batch of scrap steel whose geometric dimensions exceed the preset specifications and whose compressive strength is higher than the preset threshold, the diagnostic model processes the input fusion features and outputs the fault type of main motor overload. In the fault logic diagram of scrap steel equipment, the system takes the main motor node as the starting point, combines the scrap steel accumulation with geometric dimensions exceeding the preset threshold identified by the spatial feature vector and the sudden change in the vibration frequency of the cutter shaft in the temporal feature vector, and performs inference calculations along the graph connection lines. The system ultimately outputs a complete fault propagation path: excessive scrap steel accumulation at the feed inlet causes the cutter shaft to jam due to excessive cutting resistance, resulting in a surge in the reducer's reaction torque, which ultimately causes the main motor to overload and shut down. This propagation path directly generates maintenance instructions that include clearing the jammed scrap steel at the feed inlet and inspecting the cutter shaft blades, allowing on-site operators to execute them precisely. Collaboration and Update Module: This module is used to determine whether a change in operating conditions has occurred based on the vibration characteristic distribution under no-load idling conditions, and to trigger automatic parameter updates of the diagnostic model if a change in operating conditions is determined to have occurred. Specifically, when the equipment is in a no-load idling state, the current vibration feature distribution is extracted, and the distance difference between the current vibration feature distribution and the vibration feature distribution in the initial reference no-load idling state of the equipment is calculated; when the distance difference is detected to be greater than the set change trigger threshold multiple times in a row, it is determined that the operating condition has changed due to equipment aging. At this point, the system packages and extracts recent feature data and starts incremental learning to fine-tune the network weight parameters of the diagnostic model; as scrap steel tools wear out or liners are consumed, the baseline state of the equipment during normal operation will physically drift. By comparing the vibration characteristic distribution under no-load idling conditions, this physical aging phenomenon can be accurately identified; the automatic update of the diagnostic model parameters is triggered, enabling the diagnostic model to adaptively optimize as the scrap steel processing equipment undergoes physical aging, thus solving the problem of the static model's accuracy dropping sharply in the later stages of equipment operation. After the scrap steel shredder replaced the new blades and ran continuously under high load for three months, the blades showed obvious wear. During a period of no-load idling when the motor current was extremely low and there was no material feeding, the collaboration and update module extracted the current vibration characteristic distribution and calculated the difference distance between it and the reference vibration characteristic distribution during the break-in period of the new blades. The system detected that the difference distance was greater than the set change trigger threshold in five consecutive idle runs, and determined that the working condition change was caused by tool wear. At this point, the system automatically packages and uploads the recent feature data to the cloud server, triggering the diagnostic model to perform parameter fine-tuning and updates. After the update is completed, the new model is sent to the field gateway, eliminating false alarms of vibration baseline offset caused by physical wear of the tool.

[0017] In a preferred embodiment of the present invention, the data acquisition and preprocessing module includes: a data acquisition unit, used to acquire image data acquired by an industrial camera, vibration data acquired by a high-frequency vibration sensor, current data acquired by a current transformer, and temperature data acquired by a temperature sensor, and to acquire timestamps and align the timestamps through a synchronous acquisition card with a unified clock source to obtain multi-source sensor data; and a missing rate calculation unit, used to calculate the data missing rate of the multi-source sensor data within a sliding time window. The outlier calculation unit is used to calculate the proportion of outliers to total sampling points within a sliding time window using an outlier detection algorithm, thereby obtaining the data outlier rate. The quality index generation unit is used to calculate the data quality index based on the normalized signal-to-noise ratio, data missing rate, data anomaly rate, and preset corresponding weighting coefficients of the multi-source sensor data.

[0018] Data acquisition unit; This unit is used to acquire image data from industrial cameras, vibration data from high-frequency vibration sensors, current data from current transformers, and temperature data from temperature sensors, and achieves timestamp alignment through a synchronous acquisition card with a unified clock source to obtain multi-source sensor data. Specifically, an industrial camera is deployed above the feed inlet of the scrap steel shredder or gantry shear to capture image data of the scrap steel material shape and feed amount; a three-axis high-frequency vibration sensor is deployed in the main shaft bearing housing to collect vibration data at a sampling rate of 10kHz; and a current transformer is connected to the stator of the drive motor to obtain three-phase current data. A patch-type temperature sensor is deployed in the gearbox housing to acquire temperature data; all data acquired by the hardware is connected to a synchronous acquisition card with a unified clock source to achieve microsecond-level timestamp alignment. This operation unifies the sensor data from different parts and with different physical properties of the scrap steel processing equipment to the same time reference. Since the scrap steel processing operation is a continuous high-speed physical process of feeding, shearing, crushing and discharging, the timestamp alignment of the unified clock source ensures that the overall physical state of the equipment at the same instant can be accurately restored during subsequent analysis, avoiding data timing misalignment caused by the independent timing of each sensor. During the on-site operation of the scrap steel gantry shear, the operator starts the shear and the industrial camera above the feed inlet captures image data of scrap steel with geometric dimensions larger than the preset specifications entering the shearing area. Within the same microsecond time dimension, the high-frequency vibration sensor of the spindle collected vibration impact data at the moment the blade came into contact with the scrap steel, and the current transformer of the motor stator collected data on the surge in current. The data acquisition unit uses a synchronous acquisition card to stamp these sets of data with completely consistent microsecond-level timestamps and package them for output, ensuring that the subsequent processing module can accurately correspond the large scrap steel feeding with vibration impact and current surge in physical time. Missing rate calculation unit; This unit is used to calculate the missing rate of multi-source sensor data within a sliding time window; Specifically, within a sliding time window set to 2 seconds, the number of sampling points actually collected by a certain data source is counted, and it is compared with the total theoretical number of sampling points of the sensor within the window to calculate the proportion of missing sampling points to the total theoretical number of sampling points, which is the missing rate. The data missing rate reflects the continuity of the sensor signal within the current time period; the higher the data missing rate, the more serious the data loss caused by the sensor due to loose wiring or strong electromagnetic interference at the scene, and the lower the reference value of the data; when the scrap steel shredder is operating under high load, strong electromagnetic interference is generated at the scene; Within the current 2-second sliding time window, the theoretical number of sampling points for the spindle vibration sensor should be 20,000. Based on a 10kHz sampling rate, the missing rate calculation unit found that only 18,000 valid sampling points were actually received. The system calculated that the data missing rate of vibration data within this window was 10%, indicating that the vibration signal was lost due to interference during this period. Anomaly rate calculation unit; This unit is used to calculate the proportion of outliers to total sampling points within a sliding time window using an outlier detection algorithm, thus obtaining the data anomaly rate; Specifically, the system calls a pre-trained isolated forest algorithm to analyze the distribution of data points within the sliding time window, identify outliers that deviate from the normal distribution pattern, and calculate the ratio of the number of these outliers to the total number of sampling points as the data anomaly rate; there is a large amount of metal dust and severe impact at the scrap steel processing site, and the sensor is very likely to generate instantaneous burr signals or abnormal noise. Calculating the data anomaly rate can quantify the degree of contamination of the overall signal by such abnormal noise; the higher the data anomaly rate, the more high-frequency abrupt interference and invalid signals are contained in the signal segment, and the worse the data quality. When the scrap steel shear cuts a piece of alloy steel with a hardness exceeding the preset range, the violent mechanical rebound causes a momentary poor contact at the wiring of the surface-mount temperature sensor in the reducer housing, resulting in multiple sudden spikes in the temperature signal with amplitudes exceeding the normal operating range. The outlier calculation unit identifies these outliers that clearly do not conform to the physical law of slow temperature change through the isolated forest algorithm, and calculates that the data outlier rate of the temperature data in the current time window is 5%, indicating that there are many noise spikes in the temperature data. The data quality index generation unit calculates the data quality index based on the normalized signal-to-noise ratio, missing data rate, outlier rate, and corresponding weighting coefficients of the multi-source sensor data. The system uses the following formula: in, This refers to the data source sequence number of the multi-source sensor data; For the first Data quality index of each data source; This is the range normalization function; during the calculation, the signal-to-noise ratio of the multi-source sensor data within this window is used. The data is processed using a range normalization function, combined with the missing data rate within the window. and data anomaly rate Multiply by the weighting coefficient corresponding to the signal-to-noise ratio respectively. Weighting coefficients corresponding to missing rates Weighting coefficients corresponding to the anomaly rate Perform a summation calculation; initially set the aforementioned weighting coefficients. , , The values ​​are 0.4, 0.3, and 0.3 respectively. During the system trial operation phase, the weighting coefficients are calibrated according to the characteristics of different physical sensors; for example, for slowly changing temperature sensors, the missing rate weighting coefficient is adjusted. Increase to 0.5; For vibration sensors that are susceptible to impact and burr formation, their anomaly rate weighting coefficient is adjusted. Increase it to 0.5; through this formula, the three dimensions of signal-to-noise ratio, missing rate, and outlier rate are combined into a single data quality index; This index comprehensively and objectively quantifies the reliability of data collected by sensors under harsh working conditions, providing a physical basis for weight allocation during subsequent feature fusion. At the scrap baler operation site, the normalized signal-to-noise ratio of the vibration sensor within the current time window is 0.8, the missing data rate is 0.1, and the anomaly rate is 0.2. Since the vibration sensor is susceptible to impact, the system calls the calibrated weighting coefficients, namely, signal-to-noise ratio weight 0.2, missing data rate weight 0.3, and anomaly rate weight 0.5. The quality index generation unit substitutes the above values ​​into the formula to calculate the data quality index of the current vibration data: This value indicates that the current vibration data is of good quality and can serve as a reliable input for subsequent fault diagnosis.

[0019] The feature extraction module includes a time-series feature extraction unit, which is used to process multi-source time-series data using a time-series feature extraction model to extract time-series feature vectors. The multi-source time-series data includes vibration data, current data, and temperature data. The time-series feature extraction model uses multi-layer historical data convolution, and the sampling interval of each layer is increased by a factor. The spatial feature extraction unit is used to process image data using an image processing model, extract the volume and stacking density features of scrap steel, and obtain spatial feature vectors. Alignment units are used to combine temporal feature vectors and spatial feature vectors within the same sliding time window based on timestamps.

[0020] The specific operational logic and technical details of the feature extraction module in this scrap steel processing equipment fault prediction and health management system are as follows: Temporal feature extraction unit; This unit is used to process multi-source time-series data using a temporal feature extraction model to extract temporal feature vectors, where the multi-source time-series data includes vibration data, current data and temperature data; Specifically, a temporal convolutional network is used as the temporal feature extraction model. In view of the physical characteristics of large instantaneous impact during scrap steel shearing, the model adopts a dilated causal convolutional structure, that is, multi-layer historical data convolution is performed, and the dilation rate of each layer, that is, the sampling interval, increases exponentially, such as by 1, 2, 4, and 8 times. The vibration, current, and temperature data within the sliding time window are input into the model. The model captures instantaneous abrupt changes through low-level convolutions with small dilation rates and captures long-term periodic features through high-level convolutions with large dilation rates, ultimately outputting a unified temporal feature vector. By using multi-layer historical data convolution with sampling intervals increased exponentially, the model has a larger receptive field without increasing computational burden; This can accurately adapt to the time span of the scrap steel processing process, extracting both the transient current surge at the moment the tool contacts the scrap steel, which is a sudden change in time from microseconds to milliseconds, and the periodic vibration in the continuous cutting process, which is a stable feature over a longer time span. During the process of cutting a thick-walled steel pipe with a scrap steel gantry shear, the current data showed a sharp peak within tens of milliseconds after the cutter just contacted the steel pipe, and the vibration data showed regular periodic fluctuations during the more than 1 second of cutting. The temporal feature extraction unit inputs the temporal data within these two seconds into the temporal convolutional network; the convolutional layers at the bottom of the network accurately capture the initial transient current spike features, while the convolutional layers at the higher levels fully extract the subsequent periodic vibration features, and finally output a temporal feature vector that integrates multi-scale dynamic information. Spatial feature extraction unit; This unit is used to process image data using an image processing model, extract the volume and stacking density features of scrap steel, and obtain spatial feature vectors; Specifically, a lightweight convolutional neural network is used as the image processing model. The scrap steel image at the feed inlet captured by the industrial camera is input into the network. The network identifies the outline edge, stacking level and area ratio of the scrap steel in the image through multi-layer two-dimensional convolution operations, calculates the volume characteristics and stacking density characteristics of the scrap steel, and finally encodes and outputs them as a spatial feature vector. The shape, size, and stacking of scrap steel directly determine the load on subsequent shearing and crushing. Extracting the volume and stacking density characteristics of scrap steel transforms intuitive visual images into quantitative feature vectors that can be used for mathematical calculations, providing a preliminary physical basis for the system to predict equipment load and potential jamming failures in advance. An industrial camera captured the scene as a loader fed an entire scrapped car body into the feed inlet of a scrap metal shredder. The spatial feature extraction unit inputs the image into a lightweight convolutional neural network. The network identifies that the scrap steel in the image has a large volume and low density hollow structure, and converts it into a spatial feature vector output containing numerical information on volume and stacking density. This provides visual data support for the subsequent system to predict the cutting resistance of the main motor. Alignment unit; this unit is used to combine temporal feature vectors and spatial feature vectors within the same sliding time window based on timestamps; Specifically, read the timestamp information attached to the temporal feature vector and the spatial feature vector, and compare the time bases of the two. When it is confirmed that both belong to the same 2-second sliding time window, the temporal feature vector and spatial feature vector are concatenated and combined to form a complete joint feature set; the scrap steel processing process is a continuous physical action, and there is a strong causal relationship between the feeding state and the equipment operating state. By aligning and combining features of different modes based on timestamps, the time difference between data acquisition and independent extraction is eliminated, ensuring the absolute correspondence between the feeding of large pieces of material and the instantaneous rise in current on the time axis, laying the foundation for subsequent cross-modal deep fusion. During the continuous operation of the scrap steel baler, the alignment unit receives the spatial feature vector with extremely high scrap steel stacking density output by the spatial feature extraction unit, and at the same time receives the temporal feature vector with abrupt change in the main shaft vibration frequency output by the temporal feature extraction unit. By checking the timestamps of both, the alignment unit confirmed that both features were generated within the same sliding time window from 10:05:12 to 10:05 on the same day. Therefore, the alignment unit combined the two feature vectors and submitted them to the downstream module, ensuring that the system could jointly analyze high-density feeding and vibration anomalies as the same physical event.

[0021] In a preferred embodiment of the present invention, the feature synthesis module includes: an early splicing unit, used to extract the effective value of the current data and to perform channel-level splicing of the effective values ​​of the temperature data and the current data before inputting them into the time-series feature extraction model; The weight calculation unit is used to calculate the fusion weight of each sensor feature based on the data quality index and temperature data; The feature association unit is used to calculate the association features of temporal feature vectors and spatial feature vectors by using a feature weighting processing method and combining fusion weights to obtain fused features; the decision judgment unit is used to vote on the output results of different types of sensor features in the classification decision layer of the diagnostic model.

[0022] Early splicing unit; This unit is used to extract the effective value of the current data and perform channel-level splicing of the effective values ​​of the temperature data and the current data before inputting them into the time series feature extraction model. Specifically, after obtaining the current data at the motor stator, the effective value of the current data within the current sliding time window is calculated; The effective values ​​of the temperature data of the reducer housing and the calculated current data are timestamped and then directly combined in the dimension of the data channel before inputting these data into the time series feature extraction model. The effective value of the current reflects the overall work load of the main motor over a period of time, while the temperature data reflects the continuous heating state of mechanical components such as the reducer. Both types of data belong to low-dimensional physical quantities with relatively slow changes. Early channel-level splicing before inputting into a complex feature extraction network can preserve the most original physical causal relationship between motor load and mechanical heat generation to the greatest extent, and avoid losing this intuitive low-frequency feature relationship when extracting in subsequent deep networks. In the scenario of continuous processing of high-hardness scrap steel by the scrap steel shredder, the load of the main motor remains high. The early splicing unit calculates that the effective value of the three-phase current of the main motor within the current sliding time window is 110% of the rated value, and at the same time, the temperature data of the reducer housing is 65℃. The system directly concatenates and combines this high current RMS value with the relatively high temperature data at the channel dimension, and inputs it into the downstream time-series feature extraction model as a set of data containing the initial correlation between high load and high heat generation, thus laying the data foundation for accurately determining whether the equipment is in an overload and overheating state. Weighting calculation unit; this unit is used to calculate the fusion weights of each sensor feature based on the data quality index and temperature data; Specifically, the system extracts the data quality indices of each data source calculated in real time, and obtains temperature data (i.e., temperature scalar parameters) used to control the smoothness of weight allocation. It then calculates the dynamic fusion weights of each sensor feature using an exponential function; the system uses the formula: Calculate the first Fusion weights of sensor features ;in, For the natural constant An exponential function with base 0; The total number of sensors participating in feature fusion; The sensor serial number in the summation process ( ); For the first Data quality index of each sensor; The dimensionless temperature scalar parameter is obtained by mapping from temperature data, thus ensuring the independent variable of the exponential function. Dimensionless; Specifically, this fusion weight By the Data quality index of individual sensors Divide by the temperature scalar parameter corresponding to the temperature data Then take the exponent and divide by all the participants in the fusion. The sum of the corresponding exponential terms of each sensor is obtained; Set this temperature scalar parameter The initial value is 1.0; during the model training phase in the cloud, this temperature scalar parameter is used as a learnable parameter and is iteratively optimized through the backpropagation algorithm with the goal of minimizing the cross-entropy loss function of the final fault classification. By introducing a data quality index and a temperature scalar parameter for calculation, the traditional splicing method where the weights remain unchanged during the fusion of multi-source heterogeneous data is changed; the data quality index directly determines the proportion of the corresponding feature in the final fused feature, and the better the data quality, the greater the calculated fusion weight; The worse the data quality, the smaller the weight; the temperature scalar parameter plays a buffer role in adjusting the weight allocation, preventing the features of a sensor from being completely discarded due to a slight decrease in the data quality. This dynamic weighting mechanism ensures that even when a single sensor is severely interfered with or fails under harsh operating conditions, the system can still maintain normal operation by relying on data from other high-quality sensors. During the on-site operation of the scrap steel gantry shear, the surface of the spindle vibration sensor was covered with a large amount of metal dust, which caused the signal it collected to be severely distorted, and the vibration data quality index calculated in real time plummeted to 0.2. Meanwhile, the current sensor and industrial camera were unaffected, with data quality indices of 0.9 and 0.85, respectively; the weighting calculation unit substituted the above data quality indices and the current temperature scalar parameter into the calculation formula; The calculation results automatically and significantly reduced the fusion weight of vibration features from 0.35 in the normal state to 0.05, while increasing the fusion weights of current features and visual features to 0.50 and 0.45 respectively; this allows the system to still operate stably and output reliable fusion features even when the vibration sensor fails briefly. Feature association unit; this unit is used to calculate the association features of temporal feature vectors and spatial feature vectors by using a feature weighting processing method and combining fusion weights to obtain fused features; Specifically, after obtaining the fusion weights of the features of each sensor, the attention mechanism and other feature weighting processing methods are used to cross-correlate the temporal feature vector output by the temporal feature extraction model and the spatial feature vector output by the image processing model. The calculated fusion weights are multiplied by the corresponding temporal feature vectors and spatial feature vectors respectively. The features of different modalities are weighted, summed and correlated to generate a unified fusion feature. Failures of scrap steel processing equipment are often complex manifestations across physical modalities. Simply splicing features directly cannot reflect the linkage effect inside the equipment. By combining the correlation features of the time-series feature vector and the spatial feature vector calculated by the fusion weight, a deep binding between the feeding space morphology and the underlying electrical and mechanical time-series state of the equipment is achieved; This solves the problem of identifying complex cross-modal faults, such as large material jamming, which cannot be accurately identified by a single sensor, and ensures that the fused features can comprehensively and accurately characterize the overall physical health status of the scrap steel processing equipment within the current time window. When a steel plate with compressive strength and thickness both exceeding the preset threshold falls into the feed inlet of the scrap steel baler, the industrial camera extracts the spatial feature vector representing the large piece of scrap steel being fed in, while the main motor current transformer extracts the temporal feature vector representing the instantaneous surge in current. After receiving the two feature vectors, the feature association unit performs a weighted cross-association calculation on the two feature vectors in combination with the high fusion weights assigned by the weight calculation unit. The system deeply binds the large pieces of scrap steel in space with the surge of current in time, generating a fusion feature that clearly points to the physical condition of heavy-load cutting obstruction, and passes it to the downstream diagnostic model for fault classification. Decision-making unit; this unit is used to vote on the output results of different types of sensor features in the classification decision layer of the diagnostic model. Specifically, when the forward inference of the diagnostic model reaches the final classification decision layer, the independent classification probability results obtained based on the features of different types of sensors are extracted. These different types of classification probability results are statistically analyzed according to a preset probability level voting logic, and the fault classification label with the highest number of votes or the highest overall probability is used as the final diagnostic output result. Based on deep feature fusion, an additional round of voting is added at the classification decision layer, which constitutes the late fusion mechanism of the system. This mechanism is equivalent to receiving and integrating independent results from different physical perception perspectives before finally issuing a conclusion on equipment failure. It can effectively filter out misjudgments caused by accidental mutations of a single feature, further improving the fault tolerance and reliability of the final diagnostic decision, and greatly reducing the false alarm rate of the system in the harsh physical environment of scrap steel processing. During the continuous operation of the scrap steel shredder, a piece of scrap steel bounced irregularly in the crushing chamber. Based on the independent classification results of the main shaft vibration characteristics, it was determined that the main shaft was abnormal and a high alarm probability was given. However, both the independent classification results based on motor current characteristics and the independent classification results based on reducer temperature characteristics indicate that the current state is normal operation. After the decision-making unit collects these three results at the classification decision level, it performs a vote. Since two votes indicate normality and one vote indicates abnormality, the system comprehensively determines that the current vibration change is a normal physical impact during processing rather than a fault in the equipment itself, and finally outputs a diagnosis that the equipment is normal, successfully avoiding an unnecessary shutdown for inspection.

[0023] In a preferred embodiment of the present invention, the fault diagnosis and reasoning module includes: a combined diagnosis unit, used to use a combined diagnosis model composed of support vector machine, random forest and neural network model as a diagnosis model, process the fused features and output the fault type; The weight dynamic adjustment unit is used to dynamically adjust the final voting weight of each base classifier based on the historical accuracy of each base classifier over a set number of diagnostic cycles.

[0024] Combined diagnostic unit; This unit is used to use a combined diagnostic model consisting of support vector machine, random forest and neural network model as the diagnostic model, process the fused features, and output the fault type; Specifically, the fused features output by the feature synthesis module are received and simultaneously input into three base classifiers: support vector machine, random forest, and deep neural network model. Each base classifier independently calculates and classifies the input fusion features based on its learned classification boundaries and feature mapping rules, and outputs its own fault classification results. The combined diagnostic model summarizes the three independent classification results as the basis for determining the final fault type; support vector machines are good at handling small samples and nonlinear boundary problems, random forests have good anti-overfitting ability for high-dimensional features, and deep neural network models perform well in extracting complex deep nonlinear correlations. The combined diagnostic model composed of these three models can integrate the advantages of different algorithms, avoid the problem of insufficient generalization ability of a single diagnostic model when facing the complex and ever-changing harsh working conditions of scrap steel processing equipment, and improve the overall accuracy and robustness of fault type identification. When a batch of complex scrap steel containing thin iron sheets and thick steel blocks is processed by a scrap steel shredder, the feature synthesis module outputs fused features that include information on violently fluctuating currents and irregular vibrations. The combined diagnostic unit inputs the fused features into three base classifiers respectively; the classification result output by the support vector machine determines that the main axis is slightly stuck, the classification result output by the random forest determines that it is a normal impact, and the deep neural network model determines that the cutting resistance of the tool axis exceeds the limit based on its sensitivity to deep temporal correlation. The combined diagnostic unit collected the preliminary classification results of the three base classifiers, providing multi-dimensional diagnostic references for the subsequent comprehensive voting. The weight dynamic adjustment unit is used to dynamically adjust the final voting weight of each base classifier based on its historical accuracy over a set number of diagnostic cycles. Specifically, the sliding window size is set to 100 diagnostic cycles. During continuous system operation, the classification results of support vector machine, random forest and deep neural network models in the past 100 diagnostic cycles are continuously recorded and compared with the actual confirmed equipment status labels to calculate the historical accuracy of each base classifier in these 100 diagnostic cycles. Based on the calculated historical accuracy values, the final voting weight is dynamically assigned to each base classifier. The base classifier with the higher historical accuracy value receives the greater final voting weight. The classification results currently output by the three base classifiers are multiplied by their corresponding final voting weights and summed. The fault classification label with the highest overall probability is output as the final fault type. The final voting weight assigned to the base classifier reflects the diagnostic reliability of the base classifier under the current equipment operating conditions. The operating conditions of scrap steel processing equipment will change constantly with the change of feed material and the wear of mechanical parts, and the static weight model cannot always remain optimal; Dynamically adjusting the voting weights allows the combined diagnostic model to automatically tilt its confidence level according to the characteristics of recent operating conditions. If the deep neural network model performs better under recent complex operating conditions, its final voting weight will automatically increase, thereby ensuring the system's continuous high accuracy during long-term equipment operation. Under the condition that the scrap steel gantry shear machine processes high-hardness alloy scrap steel for several consecutive days, the historical accuracy of the deep neural network model reached 95% in the past 100 diagnostic cycles, while the historical accuracy of the support vector machine and random forest dropped to 80% and 82%, respectively. Based on this historical accuracy performance, the weight dynamic adjustment unit dynamically increases the final voting weight of the deep neural network model to 0.6, and decreases the weights of the support vector machine and random forest to 0.2 and 0.2, respectively. When the system receives new fused features, although both the support vector machine and random forest determine that the equipment status is normal, the deep neural network model determines that the gears of the reducer are worn. Since the deep neural network model has the dominant weight, the system finally outputs the fault type as gear wear of the reducer, successfully warning of potential mechanical faults.

[0025] The fault diagnosis and reasoning module also includes: a relational graph construction unit, which maps the physical structure of the scrap steel processing equipment into graph nodes and maps the physical transmission relationship and causal relationship into graph lines, thereby pre-constructing a fault logic relational graph of the scrap steel equipment; The node association analysis unit is used to perform node association analysis along the graph lines when a fault type triggers a preset alarm tag, starting from the graph node corresponding to the alarm tag and combining the temporal feature vector and spatial feature vector, and output the fault propagation path and maintenance guidance.

[0026] Relationship Graph Construction Unit: This unit is used to map the physical structure of scrap steel processing equipment into graph nodes and the physical transmission and causal relationships into graph lines, thereby pre-constructing a fault logic relationship graph of scrap steel equipment. Specifically, during the system initialization and deployment phase, the mechanical transmission chain of the scrap steel processing equipment is sorted out, and specific physical components such as the feeding conveyor belt, main motor, reducer high-speed shaft, flywheel, and cutter shaft are defined and mapped as various graph nodes in the graph. Based on the physical transmission relationships between components, such as the driving relationship between the main motor and the high-speed shaft of the reducer, the rigid connection relationship between the flywheel and the cutter shaft, and the causal relationship of fault propagation, such as abnormal vibration caused by obstruction of the cutter shaft, directed graph lines are established between the corresponding graph nodes to construct a fault logic relationship diagram of scrap steel equipment that includes the complete physical structure of the equipment. Pre-constructing a fault logic diagram for scrap steel equipment essentially transforms the maintenance experience of on-site mechanical experts and equipment structural drawings into a computer-computable structured knowledge ontology; the nodes in the diagram represent physical entities that may fail, and the lines in the diagram represent the physical channels through which the fault propagates within the equipment. This changes the traditional model's ability to output isolated fault labels, providing a clear physical structure map for tracing the root cause of faults; when deploying the health management system for scrap steel shredders, technicians input equipment structural information by constructing units using relationship diagrams; The system maps the main motor, reducer, and cutter shaft as three independent graph nodes. It adds graph lines representing the driving relationship between the main motor and reducer nodes, graph lines representing torque transmission between the reducer and cutter shaft nodes, and graph lines representing the causal relationship between the state nodes representing cutter shaft jamming and main motor overload. Finally, the system generates a fault logic diagram of scrap steel equipment that accurately reflects the physical transmission structure of the scrap steel shredder. Node association analysis unit; This unit is used to perform node association analysis along the graph lines when a fault type triggers a preset alarm tag, starting from the graph node corresponding to the alarm tag, and combining the temporal feature vector and spatial feature vector, to output the fault propagation path and maintenance guidance. Specifically, when the fault type output by the combined diagnostic unit belongs to a preset alarm tag, such as main motor overload, the system calls the graph neural network algorithm to locate the main motor graph node corresponding to the alarm tag in the scrap steel equipment fault logic relationship graph as the starting point for inference. The spatial feature vector and temporal feature vector passed in by the feature extraction module are used as node attribute inputs. The probability walk calculation is performed in reverse along the graph connection to evaluate the probability of physical anomalies occurring in adjacent preceding nodes. Finally, the nodes with probabilities exceeding the limit are connected in series to output a complete fault propagation path, and corresponding text repair instructions are automatically generated based on the root cause node at the end of the path. Combining spatial feature vectors and temporal feature vectors to perform node association analysis on a logical relationship graph is a process of causally binding abstract diagnostic classification results with specific physical phenomena. It can clearly explain a surface-level fault phenomenon, such as motor overload, and which series of underlying physical anomalies, such as the increased resistance caused by the feeding of large pieces of scrap steel, trigger a chain reaction. This reasoning result, which is causally explainable, can directly guide on-site maintenance work, avoid maintenance personnel from performing untargeted disassembly and troubleshooting, and significantly improve maintenance efficiency. During the scrap steel shearing operation, the combined diagnostic unit outputs the main motor overload fault type that triggers the alarm tag; The node association analysis unit responds immediately, taking the main motor graph node in the scrap steel equipment fault logic relationship graph as the starting point, and combining the currently received visual spatial features to identify two feature vectors: scrap steel with geometric dimensions exceeding the preset threshold and sudden changes in the vibration frequency of the cutter shaft. It then performs probability walk calculations along the graph lines. The calculation results show that the abnormal state is transmitted from the feed end to the motor end; the system finally outputs a complete and explainable fault propagation path: the excessive accumulation of scrap steel at the feed inlet causes the cutting resistance of the cutter shaft to exceed the limit and become stuck, which causes a surge in the reaction torque of the reducer, and finally causes the main motor to overload and stop. Based on this fault propagation path, the system directly generated maintenance instructions that included clearing the stuck scrap steel at the feed inlet and inspecting the cutter shaft and blades. On-site operators quickly removed the stuck scrap steel according to the instructions, enabling the equipment to quickly return to normal operation.

[0027] In a preferred embodiment of the present invention, the coordination and updating module includes: an idle determination unit, used to determine that the equipment is in an unloaded idle state when the current data is lower than 20% of the rated value of the main motor of the scrap steel processing equipment and there is no material feeding in visual recognition. The distance calculation unit is used to extract the vibration characteristic distribution under the current no-load idling state and calculate the Wasserstein distance between the vibration characteristic distribution under the current no-load idling state and the vibration characteristic distribution under the reference no-load idling state. The change determination unit is used to determine that a change in working conditions has occurred if the Wasserstein distance is detected to be greater than or equal to the change trigger threshold after a preset number of consecutive tests. The change trigger threshold is a set percentile of the Wasserstein distance recorded during the break-in period after the equipment is replaced with a new tool, under all no-load idling conditions.

[0028] Idle running determination unit; This unit is used to determine that the equipment is in an unloaded idle running state when the current data is lower than 20% of the rated value of the main motor of the scrap steel processing equipment and there is no material feeding detected by vision. Specifically, the system acquires the main motor current data collected by the current transformer and the visual recognition results of the industrial camera in real time. When it is determined that the current data in the current time window is less than 20% of the rated value of the main motor of the scrap steel processing equipment, and the visual recognition results confirm that no scrap steel material enters the feed inlet, the equipment is determined to be in a no-load idling state. The no-load idling state reflects the pure mechanical operating state of the equipment without external physical load interference such as cutting or crushing. By using both current data and visual recognition for joint judgment, the extremely complex impact of stress during scrap steel processing can be accurately identified, providing a data environment free from external load noise interference for the subsequent extraction of pure mechanical and physical aging characteristics. During continuous operation of the scrap steel shredder, when the previous batch of scrap steel has been processed and the next batch of scrap steel has not yet been fed into the feed inlet by the steel grabber, the industrial camera above the feed inlet detects that the image is empty and no scrap steel is seen stacked. At the same time, the system detected that the three-phase current of the main motor dropped rapidly to 15% of the rated value due to the lack of cutting resistance. The idle determination unit accurately determined that the scrap steel shredder was in an unloaded idle state when both conditions were met, and sent a status confirmation signal to the subsequent units. Distance calculation unit; This unit is used to extract the vibration characteristic distribution under the current no-load idling state and calculate the Wasserstein distance between the vibration characteristic distribution under the current no-load idling state and the vibration characteristic distribution under the reference no-load idling state. Specifically, after receiving the determination signal of no-load idling state, high-frequency vibration data within the current idling time period is extracted, and the vibration characteristic distribution of the current no-load idling state is statistically formed. The vibration characteristic distribution under the baseline no-load idling state in the model established in the early stage of the equipment is retrieved, and the Wasserstein distance between the two is calculated; the Wasserstein distance reflects the degree of overall physical difference between the current vibration distribution pattern and the initial healthy state distribution pattern. The larger the distance value, the more serious the deviation of the mechanical operating condition of the equipment spindle and the cutter from the healthy benchmark, corresponding to the deeper physical aging of cutter wear, liner consumption or loosening of mechanical parts; after the scrap steel shearing machine replaced the new cutter and ran continuously under high load for three months, it entered a 2-minute no-load idling state. The distance calculation unit extracts the current idling vibration characteristic distribution during this period and calculates the Wasserstein distance between it and the baseline idling vibration characteristic distribution when the tool was replaced three months ago. The system calculated that the current Wasserstein distance is 2.5, which is significantly higher than the value at the beginning of the equipment's operation, indicating that with the physical wear of the tool, the normal cutting vibration baseline of the spindle has been physically shifted. Change determination unit; This unit is used to determine that a change in operating conditions has occurred when the Wasserstein distance is detected to be greater than or equal to the change trigger threshold a preset number of times. Among them, the change trigger threshold is the set percentile of all Wasserstein distances under no-load idling conditions recorded during the break-in period after the equipment is replaced with a new tool; Specifically, the system continuously monitors the Wasserstein distance calculated during each no-load idling state and compares it with a pre-set change trigger threshold; The method for setting the change trigger threshold is as follows: During the first 72 hours of break-in period after the equipment is replaced with a new tool, the system automatically records all Wasserstein distances under no-load idling conditions, sorts these distance values ​​by size, and takes the 95th percentile as the change trigger threshold. When the preset number of consecutive times is set to 5, and the Wasserstein distance is detected to be greater than the change trigger threshold for 5 consecutive times, it is determined that the equipment has experienced a change in operating conditions caused by physical aging such as tool wear. The 95th percentile of the distance recorded during the break-in period was taken as the threshold for triggering changes. This not only accommodates reasonable mechanical vibration fluctuations during the normal physical break-in period of the new tool, but also sets a precise physical limit for excessive wear in the long term. The requirement that the limit be exceeded multiple times consecutively before a judgment is triggered effectively eliminates single data mutations caused by accidental mechanical vibrations or on-site dust interference, ensuring the absolute reliability of the judgment of changes in working conditions. Within the first three days after the scrap metal shredder was fitted with new blades, the system recorded the Wasserstein distance during hundreds of idling runs and calculated its 95th percentile to be 1.8, automatically setting it as the change trigger threshold. After the equipment had been running for six months, the system found that the calculated Wasserstein distances were 2.1, 2.2, 2.1, 2.3 and 2.2 in five consecutive idling tests, all of which were greater than 1.8. Based on this, the change determination unit decisively determines that the equipment has undergone a change in operating conditions due to severe tool wear, and outputs the determination result to trigger subsequent model adaptive updates.

[0029] The collaboration and update module also includes: a field control unit, which is used to deploy the forward inference process after parameter compression in the field processing gateway for emergency shutdown control; The cloud-based adjustment unit is used to upload feature data and feedback labels to the cloud server when a change in operating conditions is detected, start supplementary training to fine-tune the weights of the neural network model in the diagnostic model, and then send the updated data to the field processing gateway.

[0030] Field control unit; This unit is used to deploy the forward inference process, after parameter compression, in the field processing gateway for emergency shutdown control; Specifically, the network parameters of the diagnostic model, which includes feature extraction, feature synthesis, and fault diagnosis, are compressed, converting the original 32-bit floating-point parameters into 8-bit integer parameters; the compressed model is then deployed in a field processing gateway located near the scrap shearing machine or crusher. When the on-site processing gateway receives multi-source sensor data and infers serious fault types such as tool jamming, it directly sends an instruction to the equipment control cabinet to perform emergency shutdown control. Parameter compression of the model greatly reduces the computational resource consumption and memory usage of the diagnostic model, enabling complex diagnostic models to run smoothly in field processing gateways with limited computing power. By deploying the forward inference process on the field processing gateway, an ultra-low latency response of 50ms was achieved. This is crucial for preventing jamming in physical scenarios in scrap steel processing where power needs to be cut off quickly, thus avoiding network communication delays caused by uploading large-scale, high-frequency data to the cloud and waiting for the results to return. When a scrap steel shear accidentally cuts into a solid alloy steel column whose shearing resistance exceeds the rated load of the equipment and whose cross-sectional area is greater than a preset threshold, the compression model deployed in the on-site processing gateway identifies the jamming characteristics the instant it receives the sudden current and vibration data, and directly outputs an emergency stop command within 50ms. The field control unit immediately cut off the main motor power. At this time, the spindle had just begun to show signs of stalling, successfully preventing the main motor windings from burning out and the cutter shaft from breaking due to prolonged overload. Cloud-based adjustment unit; This unit is used to upload feature data and feedback labels to the cloud server when a change in working conditions is detected, start supplementary training to fine-tune the weights of the first layer of the diagnostic model, and send the updated data to the field processing gateway after completion. Specifically, when the change determination unit determines that a change in working condition has occurred due to tool wear, the field processing gateway packages and uploads the feature data extracted within the recent sliding time window, along with the feedback tags confirmed by the field operator, to the cloud server. The cloud server retains the complete, uncompressed diagnostic model. After receiving the above data, it starts supplementary training, keeping the weights of the feature extraction network at the front end of the model unchanged, and only fine-tuning the network weights of the first layer of the diagnostic model. After the fine-tuning and update are completed, the new model parameters are compressed again and sent to the on-site processing gateway to replace the original old model. By supplementing the training of the weights of only one layer of the diagnostic model, the model can quickly learn and adapt to the new vibration characteristic distribution after tool wear and aging, without destroying the basic mechanical and physical laws that the model has learned from a large amount of data in the early stage. This edge and cloud collaborative model adaptive update mechanism enables the diagnostic model to evolve synchronously with the physical aging of scrap steel processing equipment, solving the problem that the false alarm rate of static models increases significantly due to the drift of the baseline state after long-term operation of the equipment. After the system determines that the scrap steel shredder has experienced a change in operating conditions, the cloud adjustment unit receives the recent wear and tear characteristic data and the feedback tag of normal equipment aging uploaded from the field. Based on this, the cloud server initiated supplementary training and fine-tuned the weights of the first layer of the diagnostic model. After the updated model parameters were sent to the on-site processing gateway, the system re-identified the new vibration baseline after wear and classified it as a normal cutting state, eliminating the frequent false alarms of spindle abnormalities caused by the previous vibration baseline offset, and ensuring the stable operation of the scrap steel processing equipment in the later stages of its life cycle.

[0031] The foregoing has provided a detailed description of one embodiment of the present invention, but this description is merely a preferred embodiment and should not be construed as limiting the scope of the invention. All equivalent variations and modifications made within the scope of the claims of this invention should still fall within the patent coverage of this invention.

Claims

1. A fault prediction and health management system for scrap steel processing equipment, characterized in that, The system includes: The data acquisition and preprocessing module is connected to multiple sensors arranged on the scrap steel processing equipment through a communication interface to acquire multi-source sensor data of the scrap steel processing equipment within a sliding time window and to calculate the data quality index of the multi-source sensor data. The feature extraction module is used to receive the multi-source sensing data, extract the temporal feature vector and spatial feature vector of the multi-source sensing data respectively, and align the temporal feature vector and spatial feature vector within the same sliding time window based on the timestamp; The feature synthesis module is used to receive the aligned temporal feature vector and the spatial feature vector, and determine the fusion weight based on the data quality index, so as to fuse the temporal feature vector and the spatial feature vector to generate fused features; The fault diagnosis and reasoning module is used to input the fused features into a preset diagnostic model to determine the fault type, and to use the fault type, the temporal feature vector and the spatial feature vector as inputs to perform reasoning in a pre-constructed fault logic relationship diagram of scrap steel equipment to determine the fault propagation path; The coordination and update module is used to determine whether a change in operating conditions has occurred based on the vibration characteristic distribution under no-load idling conditions, and to trigger automatic parameter updates of the diagnostic model if a change in operating conditions is determined to have occurred.

2. The scrap steel processing equipment fault prediction and health management system according to claim 1, characterized in that, The data acquisition and preprocessing module includes: The data acquisition unit is used to acquire image data acquired by an industrial camera, vibration data acquired by a high-frequency vibration sensor, current data acquired by a current transformer, and temperature data acquired by a temperature sensor. It also acquires timestamps and aligns the timestamps through a synchronous acquisition card with a unified clock source to obtain the multi-source sensor data. The missing data rate calculation unit is used to calculate the missing data rate of the multi-source sensor data within the sliding time window. An anomaly rate calculation unit is used to calculate the proportion of outliers to total sampling points within the sliding time window using an outlier detection algorithm, thereby obtaining the data anomaly rate. The quality index generation unit is used to calculate the data quality index based on the normalized signal-to-noise ratio of the multi-source sensor data, the data missing rate, the data anomaly rate, and the preset corresponding weight coefficients.

3. The scrap steel processing equipment fault prediction and health management system according to claim 2, characterized in that, The feature extraction module includes: The time-series feature extraction unit is used to process multi-source time-series data using a time-series feature extraction model to extract the time-series feature vector. The multi-source time-series data includes the vibration data, the current data, and the temperature data. The time-series feature extraction model uses multi-layer historical data convolution, and the sampling interval of each layer is increased by a factor. The spatial feature extraction unit is used to process the image data using an image processing model to extract the volume and stacking density features of the scrap steel, thereby obtaining the spatial feature vector. An alignment unit is used to combine the temporal feature vector and the spatial feature vector within the same sliding time window based on the timestamp.

4. The scrap steel processing equipment fault prediction and health management system according to claim 3, characterized in that, The feature synthesis module includes: An early splicing unit is used to extract the effective value of the current data and perform channel-level splicing of the effective value of the temperature data and the current data before inputting them into the time-series feature extraction model. The weight calculation unit is used to calculate the fusion weight of each sensor feature based on the data quality index and the temperature data. The feature association unit is used to calculate the association features of the temporal feature vector and the spatial feature vector by using a feature weighting processing method and combining the fusion weights to obtain the fused features; The decision-making unit is used to vote on the output results of different types of sensor features in the classification decision layer of the diagnostic model.

5. The scrap steel processing equipment fault prediction and health management system according to claim 1, characterized in that, The fault diagnosis and reasoning module includes: The combined diagnostic unit is used to process the fused features by employing a combined diagnostic model composed of support vector machine, random forest and neural network model as the diagnostic model, and output the fault type; The weight dynamic adjustment unit is used to dynamically adjust the final voting weight of each base classifier based on the historical accuracy of each base classifier over a set number of diagnostic periods.

6. The scrap steel processing equipment fault prediction and health management system according to claim 5, characterized in that, The fault diagnosis and reasoning module also includes: The relationship graph construction unit is used to map the physical structure of the scrap steel processing equipment into graph nodes and map the physical transmission relationship and causal relationship into graph lines, thereby pre-constructing the fault logic relationship graph of the scrap steel equipment. The node association analysis unit is used to perform node association analysis along the graph lines when the fault type triggers a preset alarm tag, starting from the graph node corresponding to the alarm tag, and combining the temporal feature vector and the spatial feature vector, to output the fault propagation path and maintenance guidance.

7. The scrap steel processing equipment fault prediction and health management system according to claim 2, characterized in that, The collaboration and update module includes: The idling determination unit is used to determine that the equipment is in the no-load idling state when the current data is lower than 20% of the rated value of the main motor of the scrap steel processing equipment and there is no material feeding detected by vision. The distance calculation unit is used to extract the vibration feature distribution under the current no-load idling state and calculate the Wasserstein distance between the vibration feature distribution under the current no-load idling state and the vibration feature distribution under the reference no-load idling state. The change determination unit is used to determine that the change in working condition has occurred if the Wasserstein distance is detected to be greater than or equal to the change trigger threshold a number of consecutive preset number of times; wherein, the change trigger threshold is a set percentile of all Wasserstein distances under no-load idling conditions recorded during the break-in period after the equipment replaces the new tool.

8. The scrap steel processing equipment fault prediction and health management system according to claim 7, characterized in that, The collaboration and update module also includes: The field control unit is used to deploy the forward inference process, after parameter compression, in the field processing gateway for emergency shutdown control. The cloud-based adjustment unit is used to upload feature data and feedback labels to the cloud server when the change in the working condition is determined to have occurred, start supplementary training to fine-tune the weights of the neural network model in the diagnostic model, and send the updated data to the field processing gateway after completion.

Images