Metal powder processing equipment fault diagnosis method and system

Abstract

The application relates to the technical field of metal powder processing equipment, and discloses a metal powder processing equipment fault diagnosis method and system. The method comprises the following steps: collecting power factor data and load state data of a motor of a metal powder processing equipment, extracting periodic change characteristics to construct fluctuation characteristic vectors; analyzing peak value distribution density and continuous time length to locate abnormal fluctuation intervals, calculating the coincidence degree of abnormal time sequences and material caking events to mark high-risk points, combining history mutation records to compare to obtain fluctuation mode matching degrees; when the fluctuation mode matching degrees exceed thresholds, deep faults caused by material caking are determined, the caking risk probability is calculated by combining load mutation characteristics and state change trends, and a graded early warning signal is generated as an abnormal intervention basis. The method can realize early and accurate diagnosis and prospective early warning of equipment faults, effectively reduces the risk of unplanned shutdown, and meets the core needs of metal powder processing equipment for operation reliability and production continuity.

Description

Technical Field

[0001] This invention relates to the field of metal powder processing equipment technology, and in particular to a fault diagnosis method and system for metal powder processing equipment. Background Technology

[0002] Currently, in the field of metal powder processing equipment technology, equipment operation fault prediction and health management, as a key link in preventing sudden equipment failures and avoiding production safety and quality risks, has become a key research topic in the field.

[0003] In one existing technology, fault diagnosis of metal powder processing equipment is mainly achieved through post-fault troubleshooting or monitoring of a single operating parameter. This relies primarily on fixed threshold alarms or simple manual inspections, such as using preset upper and lower limits of motor power to determine the equipment's operating status and judging abnormal conditions by fluctuations in basic load data. However, this approach is clearly insufficient in complex operating environments. Fixed thresholds cannot adapt to the dynamic changes in load during powder processing, and single parameter monitoring is easily affected by fluctuations in material characteristics, environmental noise, and mechanical vibration. Furthermore, it is susceptible to the influence of variable processing conditions, random material agglomeration, and gradual equipment wear, which can easily lead to missed early faults, biased judgment of the root causes of abnormalities, and difficulty in early detection of hidden faults and quantitative assessment of equipment health status. This results in frequent unplanned equipment downtime and low production efficiency.

[0004] In summary, existing technologies are insufficient to predict and manage the health of metal powder processing equipment throughout its entire lifecycle, and cannot meet the core requirements of metal powder processing equipment technology for equipment reliability and proactive fault warning when facing complex production conditions. Summary of the Invention

[0005] This invention provides a fault diagnosis method and system for metal powder processing equipment, so as to realize fault prediction and health management throughout the entire life cycle of metal powder processing equipment, and meet the core requirements of metal powder processing equipment technology for equipment operation reliability and fault early warning for complex production conditions.

[0006] In a first aspect, to solve the above-mentioned technical problems, the present invention provides a fault diagnosis method for metal powder processing equipment, comprising: Obtain power factor data and load status data of the motor of the metal powder processing equipment; Periodic variation features are extracted from the power factor data and the load state data, and a fluctuation feature vector is constructed based on the periodic variation features; Perform feature analysis on the fluctuation feature vector to determine the abnormal fluctuation range and extract the abnormal distribution features within the abnormal fluctuation range; Timestamp information is extracted from the abnormal distribution features, and the temporal overlap between the timestamp information and the pre-acquired material clumping event is calculated. If the temporal overlap is greater than the preset overlap judgment threshold, the abnormal fluctuation interval is marked as a high-risk point. The temporal correlation strength is obtained by combining the temporal overlap, the fluctuation feature vector, and the material clumping event. Based on the time correlation strength, the corresponding historical mutation records are filtered, and the fluctuation feature vectors corresponding to the historical mutation records and the high-risk points are dynamically time-normalized to obtain the fluctuation pattern matching degree. If the fluctuation pattern matching degree is greater than the preset abnormal matching threshold, the current abnormality is determined to be a deep fault caused by material agglomeration, and a preliminary diagnosis result is obtained. The load change characteristics are calculated based on the load status data, and the risk probability of material agglomeration is determined in combination with the preliminary diagnosis results. When the risk probability exceeds the preset risk judgment threshold, a graded early warning signal is generated as the basis for abnormal intervention.

[0007] Secondly, the present invention provides a fault diagnosis system for metal powder processing equipment, comprising: The data acquisition module is used to acquire power factor data and load status data of the motor of the metal powder processing equipment; The feature construction module is used to extract periodic variation features from the power factor data and the load state data, and construct a fluctuation feature vector based on the periodic variation features; Anomaly identification module is used to perform feature analysis on the fluctuation feature vector, determine the abnormal fluctuation range, and extract the abnormal distribution features within the abnormal fluctuation range; The time-series matching module is used to extract timestamp information from the abnormal distribution features, calculate the time-series overlap between the timestamp information and the pre-acquired material clumping event, and if the time-series overlap is greater than the preset overlap judgment threshold, mark the abnormal fluctuation interval as a high-risk point, and combine the time-series overlap, the fluctuation feature vector and the material clumping event to obtain the time correlation strength. The pattern comparison module is used to filter corresponding historical mutation records based on the time correlation strength, and to perform dynamic time normalization on the fluctuation feature vectors corresponding to the historical mutation records and the high-risk points to obtain the fluctuation pattern matching degree. The fault determination module is used to determine that the current anomaly is a deep fault caused by material agglomeration if the fluctuation pattern matching degree is greater than the preset anomaly matching threshold, and to obtain a preliminary diagnosis result. The early warning output module is used to calculate the load change characteristics based on the load status data and determine the risk probability of material agglomeration in combination with the preliminary diagnosis results. When the risk probability exceeds the preset risk judgment threshold, a graded early warning signal is generated as the basis for abnormal intervention.

[0008] Compared with the prior art, the present invention has the following beneficial effects: (1) This invention collects power factor data and load status data of the motor of metal powder processing equipment, extracts periodic change features through time series decomposition, integrates fluctuation amplitude and mutation frequency features to construct fluctuation feature vector, breaks through the limitations of traditional single parameter monitoring and fixed threshold alarm, explores multi-dimensional dynamic fluctuation features of equipment operation, eliminates the interference of material characteristic fluctuations and environmental noise, provides high-precision basic data support for fault diagnosis, effectively improves the detection rate of early hidden faults of equipment, and solves the problems of abnormal missed detection and single data features in traditional methods.

[0009] (2) Based on the peak distribution density and duration of fluctuation feature vector analysis, this invention locates abnormal fluctuation intervals, calculates the temporal overlap between abnormal timestamps and material agglomeration events to mark high-risk points, and compares historical mutation records to obtain fluctuation pattern matching degree. This invention breaks through the limitation of traditional methods that cannot associate abnormal fluctuations with specific fault events, accurately captures the abnormal fluctuation characteristics and temporal correlation rules caused by material agglomeration, provides multi-dimensional basis for fault root cause determination, significantly improves the accuracy of fault identification under complex working conditions, and makes up for the defects of existing technologies in root cause judgment bias and high misjudgment rate.

[0010] (3) This invention determines the deep faults caused by material agglomeration based on the fluctuation mode matching degree, calculates the probability of agglomeration risk by combining the frequency of load mutation and the trend of state change, and generates an early warning signal when the threshold is exceeded as the basis for abnormal intervention. It breaks through the limitations of traditional post-fault investigation and lack of forward-looking early warning, and provides accurate fault prediction and intervention basis for the health management of equipment throughout its entire life cycle. It solves the problems of frequent unplanned equipment shutdowns and low production efficiency, takes into account both diagnostic accuracy and early warning foresight, and meets the core requirements of metal powder processing equipment for equipment operation reliability and production continuity. Attached Figure Description

[0011] Figure 1 This is a schematic flowchart of a fault diagnosis method for metal powder processing equipment provided in the first embodiment of the present invention; Figure 2 This is a schematic diagram of a fault diagnosis system for metal powder processing equipment provided in the second embodiment of the present invention. Detailed Implementation

[0012] 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.

[0013] Reference Figure 1 The first embodiment of the present invention provides a fault diagnosis method for metal powder processing equipment, comprising the following steps: S101, acquire the power factor data and load status data of the motor of the metal powder processing equipment; S102 extracts periodic variation features from the power factor data and the load state data, and constructs a fluctuation feature vector based on the periodic variation features; S103, Perform feature analysis on the fluctuation feature vector to determine the abnormal fluctuation range and extract the abnormal distribution features within the abnormal fluctuation range; S104, extract timestamp information from the abnormal distribution features, calculate the temporal overlap between the timestamp information and the pre-acquired material clumping event, if the temporal overlap is greater than the preset overlap judgment threshold, mark the abnormal fluctuation interval as a high-risk point, and combine the temporal overlap, the fluctuation feature vector and the material clumping event to obtain the time correlation strength. S105, based on the time correlation strength, filter the corresponding historical mutation records, and perform dynamic time normalization on the fluctuation feature vectors corresponding to the historical mutation records and the high-risk points to obtain the fluctuation pattern matching degree. S106, if the fluctuation pattern matching degree is greater than the preset abnormal matching threshold, then the current abnormality is determined to be a deep fault caused by material agglomeration, and a preliminary diagnosis result is obtained. S107, calculate the load mutation characteristics based on the load status data, and determine the risk probability of material agglomeration based on the preliminary diagnosis results. When the risk probability exceeds the preset risk judgment threshold, generate a graded early warning signal as the basis for abnormal intervention.

[0014] In step S101, the power factor data and load status data of the motor of the metal powder processing equipment are obtained.

[0015] It should be noted that the power factor data is obtained by collecting real-time power factor values ​​during motor operation. Data acquisition relies on a high-precision three-phase power sensor, which is directly connected to the three-phase power supply circuit of the drive motor in the metal powder processing equipment. The sensor can simultaneously collect the motor's active power, reactive power, and voltage-current phase difference parameters, and obtain the real-time power factor value through standard electrical conversion formulas. The basic sampling frequency is set to once per second. In high-load fluctuation scenarios such as ball milling and ultrafine processing, this can be increased to 10 times per second, while in conventional mixing processing scenarios, it can be decreased to once every 5 seconds.

[0016] All collected power factor values ​​are mapped to the range of 0 to 1 using the minimum-maximum normalization method. The maximum and minimum values ​​used for normalization are determined based on the theoretical physical boundaries calibrated at the motor's factory (i.e., the minimum value is 0 and the maximum value is 1) or the measured absolute extreme values ​​in the equipment's historical operating data over the past year. For example, during the continuous operation of a metal powder ball mill, the real-time power factor of the drive motor is collected by a three-phase power sensor at a frequency of once per second, resulting in a continuous numerical sequence of 0.85, 0.87, 0.83, and 0.88. After normalization, standardized power factor data is generated.

[0017] Real-time load rate values ​​during equipment processing are collected to obtain load status data. Data acquisition relies on the built-in load monitoring unit of the motor drive controller and a dynamic torque sensor mounted on the drive shaft. The load rate value is calculated by converting the actual output torque of the motor to the rated torque of the equipment, reflecting the real-time load changes and dynamic situation of powder processing resistance during equipment processing. The data acquisition frequency is kept completely consistent with the sampling frequency of the power factor data to ensure that the timing nodes of the two types of data are perfectly aligned. The collected load rate values ​​are simultaneously subjected to minimum-maximum normalization processing. The minimum value in the normalization process is set as the measured baseline value when the equipment is running under no-load (usually 0%), and the maximum value is set as the limit overload protection threshold allowed by the motor drive system (such as 120% or 150% of the rated load). After processing, the values ​​are in the same 0 to 1 range as the power factor data, eliminating the need for repeated normalization operations.

[0018] In step S102, the step of extracting periodic variation features from the power factor data and the load state data, and constructing a fluctuation feature vector based on the periodic variation features, includes: Analyze the time series of the power factor data and extract the fluctuation amplitude characteristics; Analyze the time series of the load state data and extract the mutation frequency features; Based on the time-series periodic patterns of the power factor data and the load state data, extract periodic variation characteristics; By integrating the fluctuation amplitude feature, the mutation frequency feature, and the periodic change feature, a fluctuation feature vector is constructed.

[0019] It should be noted that the time series analysis employs wavelet transform algorithm for multi-scale decomposition, with a decomposition level of 4. Fluctuation amplitude characteristics are obtained by calculating the standard deviation and peak deviation of the power factor values ​​within a sliding window. The base window length within the sliding window is set to 30 minutes, and the sliding step size is set to 5 minutes. For high-sensitivity monitoring scenarios involving ultrafine powder processing, the window length can be shortened to 10 minutes, while for conventional coarse powder processing scenarios, it can be extended to 60 minutes. All fluctuation amplitude characteristic values ​​utilize the previously completed minimum-maximum normalization results, maintaining a value range of 0 to 1, and the feature data is aligned with the time series.

[0020] Subsequently, the time series of load status data was analyzed to extract abrupt change frequency characteristics. Load abrupt change identification was achieved using a first-order backward difference algorithm. By calculating the absolute difference in load rate values ​​at adjacent sampling times, load abrupt change points were identified. The threshold for determining abrupt change was set based on statistical data of load fluctuations during the equipment's normal operation over the past year. The basic threshold was set at 8% of the rated load rate, which could be increased to 12% for ball milling scenarios with high load fluctuations and decreased to 5% for mixing scenarios with low load fluctuations. The abrupt change frequency characteristics were obtained by counting the number of effective abrupt change points within a sliding window. The sliding window parameters were kept completely consistent with the window parameters for power factor analysis to ensure temporal alignment of the two types of characteristics. The characteristic values ​​underwent minimax normalization processing synchronously, maintaining a consistent numerical range with the fluctuation amplitude characteristics.

[0021] Next, based on the time-series periodic patterns of power factor and load state data, periodic variation features are extracted. The periodic pattern analysis employs a Fast Fourier Transform (FFT) algorithm to perform frequency domain conversion. Frequency domain decomposition is performed on the time-series power factor and load state data respectively. Since power factor fluctuations in motors typically reflect the micro-motion rhythm of equipment (such as agitation or feeding), while load state changes typically reflect the macro-motion rhythm of production batch switching, the dominant power factor frequency component corresponding to the micro-motion and the dominant load state cycle length corresponding to the macro-motion are extracted from the sequence. Simultaneously, autocorrelation analysis is used to verify the significance of the periodic patterns, ensuring that the extracted periodic features closely match the actual production operation patterns of the equipment. The sampling length for frequency domain analysis is set according to the equipment's production batch switching cycle. The basic analysis length is set to 72 hours, which can be extended to 168 hours for multi-batch continuous production scenarios and shortened to 24 hours for single-batch intermittent production scenarios.

[0022] The extracted periodic features include two core indicators: power factor periodic frequency and load state period length. To eliminate calculation discrepancies caused by the different physical dimensions of Hz and h, a minimum-maximum normalization method is used for independent mapping. Specifically, the baseline maximum value for the power factor periodic frequency is set to the upper limit of 0.5Hz determined by the sampling theorem, and the baseline minimum value is set to 0Hz; the baseline maximum value for the load state period length is set to the maximum design duration of 24h per batch, and the baseline minimum value is set to 0h. All indicator values ​​are independently mapped to the dimensionless interval of 0 to 1 based on their respective baselines. The normalized power factor periodic frequency and load state period length will be used as two independent feature dimensions with fixed positions in the vector. These dimensions will be sequentially and structurally concatenated with the aforementioned fluctuation amplitude and abrupt change frequency to ensure that the dimension of the final generated fluctuation feature vector remains constant, thus meeting the fixed-length requirement for subsequent multi-dimensional vector similarity calculations.

[0023] Furthermore, the wavelet transform employs the Daubechies db4 wavelet basis function for a 4-level multi-scale decomposition. The db4 wavelet has a 4th-order vanishing moment and 7 filter coefficients, with a moderate support length (filter length is 8), exhibiting good localization characteristics in both the time and frequency domains. The power factor signal of the motor in the metal powder processing equipment is a typical non-stationary time-varying signal. The vanishing moment order of the db4 wavelet is sufficient to effectively capture the difference between the instantaneous fluctuation component and the trend component in the signal. At the same time, its tight support characteristic ensures the computational efficiency of the decomposition process and meets the time delay requirements of online real-time monitoring. After decomposition, 4 levels of detail coefficients (d1, d2, d3, d4) and 1 level of approximation coefficients (a4) are obtained. Among them, the high-frequency noise components corresponding to the d1-d2 levels are removed, the mid-to-low frequency fluctuation components caused by material agglomeration in the d3-d4 levels are retained, and the load trend component in the a4 level is retained. The retained components are reconstructed to obtain the denoised fluctuation signal.

[0024] Further calculations of the fluctuation amplitude characteristics are performed, starting with the standard deviation of the power factor values ​​within a sliding window. Then calculate the peak deviation. The peak deviation Defined as the maximum power factor within the window With window mean The absolute value of the difference; the standard deviation Deviation from peak value The volatility characteristic value A is obtained by weighted combination, and the calculation formula is as follows: ,in and This is the weighting coefficient, with a default value. =0.6、 =0.4.

[0025] The mutation frequency feature is extracted, and a first-order backward difference operation is performed on the time series of load state data to calculate the absolute difference ΔL(t) between adjacent sampling time points. When ΔL(t) exceeds the preset mutation judgment threshold (the base value is 8% of the rated load rate), time t is marked as a load mutation point. The number of all marked mutation points is counted within a sliding window, and the mutation frequency F_mut is obtained by dividing by the window duration. The mutation judgment threshold is adaptively selected according to the current working condition type. Three typical working condition templates are preset: ball mill ultrafine processing condition (threshold 12%), conventional mixing processing condition (threshold 8%), and light load fine processing condition (threshold 5%). The corresponding working condition template is automatically matched according to the average level of the current load rate.

[0026] Periodic variation characteristics were extracted, and Fast Fourier Transform (FFT) was performed on the time series of power factor data and load state data, respectively, with an FFT analysis length of 72 hours. In the FFT amplitude spectrum of the power factor data, the frequency component with the largest amplitude was taken as the dominant frequency of the power factor period. The dominant frequency reflects the rhythmic period of microscopic actions such as feeding and stirring of the equipment. It is typically in the range of 0.01Hz to 0.1Hz. In the FFT amplitude spectrum of the load state data, the reciprocal of the period corresponding to the frequency component with the largest amplitude is taken as the load state period length T_load, reflecting the macroscopic rhythm of production batch switching, typically in the range of 4 hours to 24 hours. Simultaneously, the significance of the periodic pattern is verified through autocorrelation analysis. If the significant period obtained from the autocorrelation analysis deviates from the period corresponding to the FFT dominant frequency by less than 10%, the periodic feature is confirmed to be valid; otherwise, the weighted average of the two is taken.

[0027] Finally, the fluctuation amplitude features, mutation frequency features, and periodic variation features are integrated to construct a fluctuation feature vector. Feature integration is achieved using a temporally aligned concatenation method. Based on a sliding window at the same time point, the fluctuation amplitude features, mutation frequency features, and periodic variation features within the corresponding window are concatenated in a fixed order to form a one-dimensional feature vector with uniform dimensions. Each dimension of the vector corresponds to a clear physical feature, eliminating dimensional confusion or feature redundancy. The feature concatenation process strictly adheres to the principle of temporal synchronization, ensuring that all features within the vector originate from the same time window. The temporal sequence of the feature vector completely corresponds to the sampling time sequence of the original data, allowing it to be directly used for subsequent abnormal fluctuation interval identification and pattern matching analysis.

[0028] In step S103, the step of performing feature analysis on the fluctuation feature vector to determine the abnormal fluctuation range and extracting the abnormal distribution features within the abnormal fluctuation range includes: Based on the fluctuation feature vector, the distribution density of abnormal peaks and the duration of fluctuations are statistically analyzed to obtain fluctuation statistics. Based on the fluctuation statistics, the time interval in which the fluctuation value exceeds the preset fluctuation threshold is located, and the abnormal fluctuation interval is determined. Extract the data distribution characteristics of the abnormal fluctuation range to obtain the abnormal distribution characteristics.

[0029] It should be noted that the abnormal peak distribution density statistics are implemented using a kernel density estimation algorithm. Taking the fluctuation amplitude characteristics in the fluctuation feature vector as the main input, the algorithm fits the probability density distribution of abnormal peaks in the time series, accurately quantifying the degree of peak clustering over time. The determination of abnormal peaks is based on the statistical settings of fluctuation characteristics from the normal operation of metal powder processing equipment over the past year.

[0030] The process involves extracting historical baseline fluctuation data from equipment operating continuously for at least 72 hours under rated load conditions. During data preprocessing, non-steady-state data from equipment start-up / shutdown phases, periods of drastic batch switching, and known anomaly alarm periods are strictly removed, retaining only clean sequences from stable processing operations. The overall standard deviation of all values ​​in the clean sequences is calculated using statistical formulas to obtain the equipment's steady-state operating fluctuation standard deviation. Based on this standard deviation, the basic criterion is a fluctuation amplitude exceeding three times the equipment's steady-state operating fluctuation standard deviation; this can be tightened to two times for high-sensitivity fault early warning scenarios and relaxed to four times for routine operation monitoring scenarios.

[0031] The duration of fluctuations is calculated using a sliding window statistical method, with the window parameters being completely consistent with those extracted from previous features. The time difference between the start and end timestamps of each consecutive abnormal peak sequence is statistically analyzed to obtain the duration of a single abnormal fluctuation. All statistically obtained distribution density and duration values ​​are mapped to the interval between 0 and 1 using the min-max normalization method, and integrated to form complete fluctuation statistics. The bandwidth of the kernel density estimation is set based on the standard deviation of the equipment's steady-state operation fluctuation data over the past year. The basic bandwidth is set to 0.5 times the steady-state standard deviation, which can be tightened to 0.3 times for high-sensitivity early warning scenarios and relaxed to 0.7 times for routine monitoring scenarios.

[0032] Furthermore, the kernel density estimation algorithm employs a Gaussian kernel function, and its bandwidth h is selected using a modified Silverman rule, with the specific calculation formula as follows: ,in denoted as , where is the standard deviation of the steady-state operating fluctuation of the equipment; IQR is the interquartile range (the difference between the 75th and 25th percentiles) of the fluctuation amplitude characteristics in the equipment's steady-state operating data over the past year. The purpose of introducing IQR is to enhance the robustness of bandwidth estimation to outliers; n is the total number of samples in the steady-state operating data. This bandwidth selection method effectively preserves the local clustering characteristics of outlier peaks while ensuring the smoothness of density estimation.

[0033] Get The specific acquisition method is as follows: From the equipment's operating data over the past year, data for periods when the equipment has continuously operated under rated load conditions for no less than 72 hours are selected; based on the start-up and shutdown logs recorded by the equipment's PLC system, data from the equipment startup and shutdown phases, data from the batch switching phase (based on a load rate change rate exceeding 20% / min), and data from known abnormal alarm periods are removed; for the clean steady-state sequence retained after preprocessing, the standard deviation σ of all fluctuation amplitude characteristic values ​​is calculated, and this standard deviation is the equipment's steady-state operating fluctuation standard deviation. .

[0034] The specific method for determining the volatility threshold is as follows: basic volatility threshold. Where k is a multiplier, with a base value of k=3. When the equipment is operating normally in a steady state, the probability that the power factor fluctuation exceeds 3 times the standard deviation is less than 0.3% (assuming an approximate normal distribution). The adjustment rules for different operating conditions are: k=2.5 for ultrafine powder processing scenarios and k=3.5 for coarse powder conventional processing scenarios.

[0035] Excluding sporadic fluctuations lasting less than 10 minutes, the shortest confirmed duration of abnormal power factor fluctuations caused by material agglomeration was 12.3 minutes (statistical sample size n=47 agglomeration events); while fluctuations caused by sporadic disturbances during normal production typically last no more than 5 minutes; therefore, 10 minutes was set as the time threshold for distinguishing between sporadic fluctuations and persistent faults. This threshold is dynamically updated as equipment operating data accumulates.

[0036] The interval positioning is achieved by a sliding window traversal method, which traverses the entire time series in 5-minute increments, identifies sequence segments whose fluctuation statistics continuously exceed the fluctuation threshold, records the start and end timestamps of the segments, removes occasional fluctuation segments with a duration of less than 10 minutes, and determines the remaining continuous time series intervals as abnormal fluctuation intervals. The interval markers contain complete time series boundaries and fluctuation characteristic information.

[0037] Finally, the distribution characteristic parameters of the abnormal fluctuation intervals are extracted to obtain the abnormal distribution characteristics. For a given abnormal fluctuation interval, multi-dimensional distribution characteristic parameters are extracted within the interval. These parameters include five main indicators: abnormal peak distribution density, average duration of fluctuations, frequency of occurrence per unit time within the abnormal interval, average fluctuation amplitude within the interval, and peak frequency of load mutations. The feature extraction process is strictly limited to the time-series boundaries of the abnormal fluctuation interval to ensure that all feature parameters originate from the abnormal fluctuation period and are free from interference from normal operation data.

[0038] It should be noted that all extracted distribution feature parameters follow the previous min-max normalization method, uniformly mapping them to the interval between 0 and 1, and splicing them in a fixed order to form a one-dimensional abnormal distribution feature. The feature dimension is clear and the physical meaning is clear, which can be directly used for subsequent time series overlap analysis and risk point marking.

[0039] In step S104, the extraction of timestamp information from the abnormal distribution features, calculation of the temporal overlap between the timestamp information and pre-acquired material agglomeration events, and marking the abnormal fluctuation interval as a high-risk point if the temporal overlap is greater than a preset overlap judgment threshold, and combining the temporal overlap, the fluctuation feature vector, and the material agglomeration events to obtain the time correlation strength, includes: Extract the timestamp information of the collection period corresponding to the abnormal distribution features to generate a timestamp sequence; Compare the timestamp sequence with the occurrence time periods of pre-acquired material agglomeration events to calculate the time sequence overlap. If the time series overlap is greater than the preset overlap judgment threshold, the corresponding abnormal fluctuation interval is marked as a high-risk point. The fluctuation intensity is extracted from the fluctuation feature vector, and the frequency of agglomeration events is statistically analyzed from the pre-acquired material agglomeration events. The time correlation strength is calculated by combining the temporal overlap, the fluctuation intensity, and the frequency of the block event.

[0040] It should be noted that the timestamp information corresponding to the sampling period for the abnormal distribution characteristics is extracted to generate a timestamp sequence. The timestamp extraction is limited by the temporal boundaries of the abnormal fluctuation interval, accurately extracting the millisecond-level timestamp corresponding to each sampling point within the interval. The sampling period is completely consistent with the sampling frequency of previous power factor and load state data, ensuring complete temporal alignment between the timestamps and the original sampled data. The extracted timestamps are arranged in ascending order according to the sampling sequence, forming a continuous and uninterrupted timestamp sequence. Each timestamp in the sequence is associated with the abnormal distribution characteristics and fluctuation feature vectors of the corresponding sampling point, eliminating data misalignment or temporal disorder.

[0041] Subsequently, the time-series overlap is calculated by comparing the timestamp sequence with the occurrence periods of pre-acquired material caking events. The pre-acquired material caking events are derived from material caking fault records confirmed by maintenance during equipment operation over the past year. Power factor and load status data are continuously collected throughout the equipment's operation. When the equipment shuts down due to a material caking fault and is confirmed by on-site maintenance, complete operating data for 30 minutes before and after the fault occurrence is automatically extracted. The occurrence timestamp, fault type, severity level, and fluctuation intensity parameters within the fault period are extracted, packaged into standardized material caking event records, and stored in the local event database. The fluctuation intensity parameter is the ratio of the maximum value of the fluctuation amplitude characteristic within the fault period to the average fluctuation amplitude during steady-state operation of the equipment, mapped to a 0-1 interval using a minimum-maximum normalization method before storage. For newly installed or initial equipment that has been in operation for less than one year and lacks local fault records, a pre-built industry standard clumping fault library or a benchmark fault template of the same model of equipment is used as the initial benchmark data for time series comparison. The template is stored in the format of standardized material clumping event records and includes the occurrence time of typical faults, fault type, severity level and normalized fluctuation intensity parameters built into the template.

[0042] As equipment operating time increases, the system automatically collects locally confirmed fault records, gradually replacing and improving the local historical fault database. Each record contains the start and end timestamps of the fault, which can be directly used for time-series comparison. The time-series overlap calculation adopts a method based on time interval overlap determination, that is, comparing the first time interval formed by the start and end timestamps of the abnormal fluctuation interval with the second time interval formed by the start and end timestamps of the material agglomeration event record, and calculating the proportion of the overlap duration of the two intervals to the total duration of the abnormal fluctuation interval. To improve the robustness of the comparison, a matching tolerance window is introduced to extend the boundary of the abnormal fluctuation interval. The start timestamp of the first time interval is extended forward and the end timestamp is extended backward by half the matching tolerance window duration, and then the overlap duration is calculated with the second time interval. The size of the matching tolerance window is set according to the equipment fault response characteristics. The basic tolerance is set to 10 minutes, which can be reduced to 5 minutes in high-sensitivity early warning scenarios and expanded to 15 minutes in regular monitoring scenarios. The numerical range of temporal overlap is between 0 and 1. The higher the value, the greater the degree of temporal overlap between the two.

[0043] Furthermore, the automated acquisition mechanism for material agglomeration events specifically involves continuously monitoring the operating status of the equipment and triggering the fault event acquisition process when any of the following conditions are met: unplanned equipment shutdown; load rate continuously exceeding 120% of the rated value for more than 30 seconds; power factor experiencing a step drop (a drop exceeding 20% ​​and lasting for more than 10 seconds); or abnormal vibration detected by the equipment vibration sensor (acceleration effective value exceeding 3 times the normal average).

[0044] Upon triggering, the system automatically captures complete operational data for 30 minutes before and after the trigger time, including power factor time-series data, load status time-series data, equipment vibration data, and ambient temperature and humidity data, and automatically calculates the fluctuation characteristic vector sequence for that period. A pre-record of the fault event is automatically generated and pushed to the mobile terminal or host computer interface of the equipment maintenance personnel. After on-site verification and confirmation of material clumping, the maintenance personnel confirm the fault type and severity level (minor / moderate / severe) through the human-machine interface. Upon receiving confirmation, the pre-record is converted into a formal material clumping event record. If the fault is determined to be non-clumping, it is marked as a negative sample and stored in the non-clumping anomaly database.

[0045] The standardized material clumping event record data structure includes the following fields: event_id (unique event number), start_timestamp and end_timestamp (millisecond-level ISO 8601 format timestamps), event_type (fault type code), severity_level (severity level), device_id (device number), operating_condition (operating condition type), power_factor_sequence (power factor time-series JSON array), load_sequence (load status time-series JSON array), feature_vector_sequence (fluctuation feature vector time-series sequence), peak_fluctuation_amplitude (maximum fluctuation amplitude), mutation_frequency (load mutation frequency), confirmed_by (confirmed personnel ID), and environmental_data (environmental temperature and humidity data).

[0046] For newly installed equipment, typical material agglomeration fault template data for the same model of equipment is downloaded from the equipment manufacturer's fault case database via a cloud data service interface. Each template contains no fewer than 10 sets of fault cases confirmed by the manufacturer. After the equipment is put into operation, when the cumulative number of confirmed fault records for the same local operating conditions reaches more than 20 sets, the system automatically reduces the retrieval weight of the template data (linearly decaying from 1.0 to 0), eventually relying entirely on local data for pattern matching.

[0047] In this implementation case, the overlap judgment threshold is set based on historical operating data of material agglomeration failures in metal powder processing equipment over the past year. For the initial operating phase lacking historical data, this threshold defaults to the factory-preset empirical parameters. Once the cumulative number of valid local fault samples exceeds a preset number (e.g., 5 times), it automatically switches to a dynamic threshold set based on local historical data. The lowest temporal overlap value corresponding to all confirmed material agglomeration failures is statistically analyzed, and the basic overlap judgment threshold is set to 0.7. For high-precision fault warning scenarios in ultrafine powder processing, the threshold can be increased to 0.75, while for conventional coarse powder processing scenarios, it can be decreased to 0.65. This threshold has been validated through long-term operation under multiple conditions, accurately distinguishing between abnormal fluctuations related to material agglomeration and irrelevant random interference fluctuations, effectively reducing the fault misjudgment rate.

[0048] If the temporal overlap is greater than the preset overlap judgment threshold, the corresponding abnormal fluctuation interval is marked as a high-risk point. The marking content includes the interval temporal boundary, the temporal overlap value and the corresponding abnormal distribution characteristics, which can be directly used for subsequent correlation strength calculation.

[0049] Subsequently, the fluctuation intensity is extracted from the fluctuation feature vector, and the frequency of clumping events is statistically analyzed from pre-acquired material clumping events. The fluctuation intensity is the ratio of the maximum value of the fluctuation amplitude characteristic within the abnormal fluctuation range to the average fluctuation amplitude during steady-state operation. The average fluctuation amplitude during steady-state operation is obtained through a combination of on-site calibration and dynamic statistics. During the initial trial operation phase, after confirming that the equipment is operating smoothly under rated load without clumping or faults, the system continuously collects at least 72 hours of continuous operating data, calculating the arithmetic mean of the fluctuation amplitude during this period as the initial calibration benchmark. In subsequent formal production, a sliding time window mechanism is used to periodically (e.g., monthly) extract data from confirmed periods without anomalies, and the steady-state average is adaptively updated and calibrated. Based on this ratio calculation, the severity of abnormal fluctuations can be directly quantified, and the value is mapped to the 0-1 range using a minimum-maximum normalization method, maintaining a consistent numerical range with the previous feature data.

[0050] The frequency of clumping events is obtained by statistically analyzing the number of material clumping events occurring per unit of equipment runtime over the past three months. When the equipment is initially commissioned and there is no local historical data for the past three months, the system uses the baseline fault frequency (e.g., 0.05 times per 100 hours) provided by the equipment manufacturer under standard operating conditions as the initial default parameter. After the equipment has been running for three months, the system fully transitions to dynamic statistics based on the actual number of occurrences locally. The basic statistical period is hourly, and the statistical results undergo min-max normalization to eliminate calculation biases caused by differences in numerical magnitude.

[0051] Finally, the temporal correlation strength is calculated by combining the temporal overlap degree, fluctuation intensity, and clumping event frequency. The temporal overlap degree is the core criterion described in the independent claim, while fluctuation intensity and clumping event frequency are auxiliary weighting parameters further refined in the dependent claims. The temporal correlation strength is obtained by weighted summation of the three parameters, with the weight of temporal overlap degree set at 0.6, fluctuation intensity at 0.25, and clumping event frequency at 0.15. This weighting combination is based on historical statistical data from metal powder processing equipment fault diagnosis. The temporal overlap degree directly reflects the temporal correlation between abnormal fluctuations and material clumping events, providing the strongest indication for fault root cause judgment and thus receiving a high weight. Fluctuation intensity reflects the severity of abnormal fluctuations and is given a medium weight. The clumping event frequency reflects the probability of historical equipment faults and has a relatively weak indication for current faults, thus receiving a low weight. The calculated temporal correlation strength value ranges from 0 to 1; a higher value indicates a stronger temporal correlation between abnormal fluctuations and material clumping events.

[0052] In step S105, the step of filtering corresponding historical mutation records based on the time correlation strength, and performing dynamic time warping on the fluctuation feature vectors corresponding to the historical mutation records and the high-risk points to obtain the fluctuation pattern matching degree includes: Based on the time correlation strength, retrieve the pre-acquired historical mutation records, which include historical load mutation response data and state change trend data; Calculate the temporal similarity between the fluctuation feature vector corresponding to the high-risk point and the historical mutation record to obtain the correlation index; The fluctuation pattern matching degree is determined by weighted fusion calculation based on the time correlation strength and the correlation index.

[0053] It should be noted that, based on the strength of time correlation, historical mutation records that best match the waveform morphology of high-risk points are retrieved from a pre-stored historical mutation database. This retrieval process prioritizes historical records based on the strength of time correlation, rather than directly incorporating them into waveform similarity calculations. Historical mutation records include historical load mutation response data and state change trend data. These records are derived from operational data corresponding to all material agglomeration faults confirmed during equipment operation over the past year. For equipment in its initial state with no historical mutation records, pre-stored factory-calibrated fault simulation data on the local controller, or standardized mutation characteristic curves from similar mature equipment distributed from the cloud, are used as the initial matching master record.

[0054] Each record contains four main data categories within 30 minutes before and after the fault: power factor fluctuation time series, load change response speed, load state change trend slope, and equipment operating condition parameters. All data undergoes min-max normalization to maintain a consistent numerical range and dimensional structure with the current fluctuation feature vector. The retrieval rules employ a time correlation strength-based hierarchical filtering mechanism. When the time correlation strength is greater than 0.5, historical change records under the same operating conditions within the past 3 months are retrieved. When the time correlation strength is greater than 0.7, the retrieval scope is expanded to historical change records under all operating conditions within the past 6 months. When the time correlation strength is greater than 0.9, all historical change records within the past year are retrieved. For the initial equipment transition phase, the system uses a weighted fusion matching mode of simulated master data and gradually accumulated local measured data. As the number of local measured fault samples increases, the system gradually increases the retrieval weight of local measured records, eventually switching entirely to matching based on the equipment's own historical data. This ensures that the retrieved records have the highest matching degree with the current high-risk point's operating conditions, without any interference from irrelevant data.

[0055] Subsequently, the temporal similarity between the fluctuation feature vector corresponding to the high-risk point and the historical mutation record is calculated to obtain the correlation index. The temporal similarity calculation is implemented using a dynamic time warping algorithm, which can effectively eliminate length differences and phase shifts in time series sequences, accurately matching the similarity of fluctuation patterns between two time series sequences. Before calculation, the time series sequence of the fluctuation feature vector corresponding to the high-risk point is aligned in length with the feature time series sequence of each historical mutation record. The alignment window size is consistent with the duration of the abnormal fluctuation interval of the high-risk point to ensure sequence dimension matching.

[0056] The calculation process uses fluctuation amplitude characteristics, mutation frequency characteristics, and periodic change characteristics as the main matching dimensions, with weights set to 0.5, 0.3, and 0.2 for each dimension, respectively. This weight combination is based on statistical data of historical failure mode recognition accuracy. Newly deployed equipment uses this preset default empirical weight, which is then dynamically fine-tuned and optimized based on actual recognition accuracy feedback. The fluctuation amplitude characteristic has the highest distinguishability for material agglomeration failure modes and is assigned the highest weight. The distinguishability of mutation frequency characteristics and periodic change characteristics decreases sequentially, with corresponding weights decreasing accordingly. The temporal similarity value of each historical record ranges from 0 to 1. The maximum similarity value among all historical records is taken as the final correlation index; a higher value indicates a higher degree of similarity between the current fluctuation mode and the historical agglomeration failure mode.

[0057] Furthermore, the specific implementation of the Dynamic Time Warping (DTW) algorithm is as follows: Let the time series sequence of the fluctuation feature vector corresponding to the high-risk point be... The characteristic time sequence of historical mutation records is , where each q_i and c_j is a four-dimensional feature vector.

[0058] The weighted Euclidean distance is used as the local distance metric between point pairs, and the calculation formula is as follows: ,in =0.5 (fluctuation range) =0.3 (mutation frequency) =0.1 (power factor cycle frequency) =0.1 (load state cycle length).

[0059] The Sakoe-Chiba constraint is used to limit the search range of the regularized path. The constraint condition is |i - j| ≤ R, where R is the bandwidth parameter, set to max(m, n) × 0.15 (i.e., not exceeding 15% of the sequence length). This also satisfies the three basic constraints of DTW: boundary conditions, monotonicity, and continuity.

[0060] When calculating the local distance between each time step pair, the weighted Euclidean distance described above is directly used to fuse the multidimensional features into a single scalar distance value. Then, based on this scalar distance matrix, standard one-dimensional DTW dynamic programming is performed. The recursive formula is as follows: The boundary conditions are D(0,0)=0, D(i,0)=D(0,j)=∞. The final DTW distance is obtained by dividing by min(m,n) to get the normalized DTW distance. .

[0061] The final temporal similarity is defined as: ,in The preset maximum distance threshold (taken as an empirical value of 2.0). The value range is [0, 1].

[0062] The matching degree of fluctuation patterns is determined based on the correlation index. The matching degree of fluctuation patterns is calculated by weighted fusion of the correlation index and the temporal correlation strength, where the weight of the correlation index is set to 0.6 and the weight of the temporal correlation strength is set to 0.4. This weight combination is given based on the historical accuracy data of fault diagnosis. The temporal similarity of fluctuation patterns is more indicative of the root cause of the fault and is given a higher weight, while the temporal correlation strength is used as an auxiliary verification index and is given a lower weight.

[0063] It should be noted that all values ​​were normalized to the minimum and maximum values ​​within the range of 0 to 1 before the fusion calculation. The final fluctuation pattern matching degree value also falls within the range of 0 to 1. A higher value indicates a higher degree of matching between the current abnormal fluctuation and the historical material agglomeration fault pattern, and a stronger reliability of the fault determination. After the matching degree calculation is completed, the corresponding historical fault types and fault severity are matched simultaneously, providing a complete pattern reference for subsequent fault determination.

[0064] In step S106, if the fluctuation pattern matching degree is greater than a preset anomaly matching threshold, then the current anomaly is determined to be a deep fault caused by material agglomeration, and a preliminary diagnostic result is obtained, including: The real-time signal update frequency for obtaining the load status data; The fluctuation pattern matching degree is compared with a preset abnormal matching threshold; If the fluctuation pattern matching degree is greater than the preset abnormal matching threshold, then by combining the preset load peak triggering condition and the real-time signal update frequency, it is determined that the current abnormality belongs to a deep fault caused by material agglomeration, and a fault determination result is obtained. By integrating the fault determination results with the abnormal distribution characteristics, a preliminary diagnostic result is obtained.

[0065] It should be noted that the real-time signal update frequency for acquiring the load status data is as follows: The pre-acquired load peak trigger condition is set based on the equipment's rated load parameters and the load characteristics of historical blockage faults. The basic trigger condition is that the load rate peak within the abnormal fluctuation range exceeds 95% of the equipment's rated load rate. For high-load processing scenarios, this can be increased to 100% of the rated load rate, and for low-load processing scenarios, it can be decreased to 90% of the rated load rate. The synchronously acquired real-time signal update frequency is directly derived from the sampling frequency setting of the load status data acquisition stage and is completely consistent with the actual sampling frequency of the load monitoring sensor. The frequency value can be updated in real time according to the equipment's operating conditions. Higher frequency sampling in high-load fluctuation scenarios corresponds to a higher update frequency, while lower frequency sampling in normal steady-state operation scenarios corresponds to a lower update frequency.

[0066] The acquisition process synchronously verifies the continuity and integrity of signal updates, eliminating invalid time periods corresponding to sampling interruptions and data packet loss, ensuring that the acquired update frequency value perfectly matches the actual sampling characteristics of the current load status data, without any frequency value distortion. For example, if the real-time signal update frequency of the current data acquired from the load status data acquisition module is once per second, synchronous verification confirms that there is no data packet loss during the sampling process, and the update frequency value is accurate and valid.

[0067] Next, the fluctuation pattern matching degree is compared with a preset anomaly matching threshold. The anomaly matching threshold is set based on historical operational data from the past year regarding material agglomeration fault diagnosis in metal powder processing equipment. The lowest fluctuation pattern matching degree corresponding to all confirmed material agglomeration faults is statistically analyzed, and the basic anomaly matching threshold is set to 0.7. For high-precision fault warning scenarios in ultrafine powder processing, the threshold can be increased to 0.75, while for conventional coarse powder processing scenarios, it can be decreased to 0.65. This threshold has been validated through long-term operation under multiple conditions and can reliably distinguish fault modes caused by material agglomeration from normal operating condition fluctuations and other non-agglomeration-related equipment anomaly modes, effectively reducing the false alarm rate and missed detection rate. The comparison process uses a point-by-point numerical comparison method to clarify the relationship between the fluctuation pattern matching degree and the threshold, providing a crucial basis for subsequent fault determination. For example, if the currently calculated fluctuation pattern matching degree is 0.815, comparing it with the basic anomaly matching threshold of 0.7 confirms that the matching degree value is greater than the preset threshold.

[0068] Subsequently, if the fluctuation pattern matching degree is greater than the preset anomaly matching threshold, then, combined with the preset load peak trigger condition and the real-time signal update frequency, the current anomaly is determined to be a deep fault caused by material agglomeration, and a fault determination result is obtained. The fault determination employs a multi-condition joint verification mechanism. The key determination conditions are that the fluctuation pattern matching degree is greater than the anomaly matching threshold, and the load peak trigger condition is met. The confidence level of the determination result is then corrected by combining the real-time signal update frequency; the higher the signal update frequency, the higher the confidence level of the determination result. When all key conditions are met, the current anomaly is determined to be a deep fault caused by material agglomeration. The fault determination result includes three key information categories: fault type, fault occurrence time, and fault confidence level. If any key condition is not met, the current anomaly is determined to be a non-agglomeration fluctuation, and fault determination is not triggered. For example, the fluctuation pattern matching degree of 0.815 is greater than the abnormal matching threshold of 0.7. At the same time, the peak load rate in the abnormal range reaches 96.2%, which meets the triggering condition of 95% rated load rate. Combined with the signal update frequency of 1 per second to complete the confidence correction, it is finally determined that the current abnormality belongs to the deep fault caused by material agglomeration, and the fault confidence is 92%.

[0069] Finally, the fault determination results are integrated with the abnormal distribution characteristics to obtain preliminary diagnostic results. The integration process adopts a time-aligned structured splicing method, using the time boundary of the abnormal fluctuation interval as a unified benchmark. The fault type, occurrence time, and confidence level in the fault determination results are completely spliced ​​with parameters such as peak distribution density, fluctuation duration, abnormal occurrence frequency, and average fluctuation amplitude in the abnormal distribution characteristics to form a preliminary diagnostic result that is dimensionally complete and logically coherent.

[0070] The integration process simultaneously supplements two types of auxiliary verification parameters: time correlation strength and fluctuation pattern matching degree, corresponding to high-risk points. This ensures that the preliminary diagnostic results not only include a clear determination of the root cause of the fault but also cover quantitative indicators of the fault's severity, time-series characteristics, and correlation verification information, which can be directly used for subsequent risk probability calculations and early warning signal generation. All integrated parameters use the previously completed minimum-maximum normalization results, maintaining a numerical range of 0 to 1, without any data dimension mismatch issues. For example, the judgment result of deep material agglomeration fault is integrated with the abnormal distribution characteristics of peak density of 0.12 and duration of 15 minutes in the abnormal interval, and parameters of time correlation strength of 0.785 and fluctuation pattern matching degree of 0.815 are simultaneously added to generate a complete preliminary diagnostic result.

[0071] In step S107, the calculation of load mutation characteristics based on the load status data, combined with the preliminary diagnostic results to determine the risk probability of material agglomeration, and the generation of a graded early warning signal as a basis for abnormal intervention when the risk probability exceeds a preset risk judgment threshold, includes: Based on the load status data, calculate the frequency of load status abrupt changes and the slope of the status change trend; Based on the preliminary diagnostic results, the frequency of load state abrupt changes, and the slope of the state change trend, the probability of material agglomeration is calculated. If the risk probability exceeds the preset risk judgment threshold, a warning signal for material agglomeration is generated; The warning signal is integrated with the risk probability to determine the basis for abnormal intervention.

[0072] It should be noted that the load state change frequency is obtained by identifying load change points using a first-order backward difference algorithm, and then calculating the number of effective changes per unit time using a sliding window statistical method. The sliding window parameters are completely consistent with the window parameters extracted in the previous feature extraction. The basic window length is set to 30 minutes and the sliding step size is set to 5 minutes. In the high-sensitivity monitoring scenario of ultrafine powder processing, the window length can be shortened to 10 minutes, and in the conventional coarse powder processing scenario, it can be extended to 60 minutes.

[0073] It is worth noting that the threshold for determining load abrupt change points is set based on statistical data of load fluctuations during normal equipment operation over the past year. The basic threshold is set at 8% of the rated load rate. For ball milling scenarios with high load fluctuations, this threshold can be increased to 12%, while for mixing scenarios with low load fluctuations, it can be decreased to 5%. The slope of the state change trend is calculated using a univariate linear regression algorithm. Using the load state time-series data for 30 minutes before and after the abnormal fluctuation interval as input, a linear trend line of load change over time is fitted. The slope value can directly quantify the deterioration or mitigation trend of the load state; a positive value represents a continuous increase in load, while a negative value represents a gradual decrease in load. All calculated abrupt change frequencies and trend slope values ​​are mapped to the range of 0 to 1 using the min-max normalization method, maintaining a consistent numerical range with the previously processed data and eliminating calculation bias caused by differences in data magnitude.

[0074] The risk probability of material agglomeration is calculated through cross-verification using two dimensions: time overlap correlation verification and waveform similarity correlation verification. The resulting fluctuation pattern matching degree serves as the core input. The risk probability calculation employs a multi-dimensional weighted fusion algorithm. The fusion weights are set based on statistical data of the accuracy of material agglomeration fault warnings over the past year. Specifically, the fault confidence score corresponding to the preliminary diagnosis result is weighted at 0.5, the load state change frequency at 0.3, and the state change trend slope at 0.2. This weighting is suitable for the fault warning requirements of metal powder processing equipment. The fault confidence score of the preliminary diagnosis result directly reflects the reliability of root cause determination and has the most direct impact on the risk probability, thus receiving the highest weight. The load change frequency reflects the frequency of abnormal fluctuations and is given a medium weight. The trend slope reflects the development trend of the fault and has a relatively weak indicative effect on the current risk, thus receiving a lower weight.

[0075] It should be noted that all input parameters have undergone min-max normalization within the range of 0 to 1 before calculation. The final material agglomeration risk probability value ranges from 0 to 1, with higher values ​​indicating a higher probability and severity of material agglomeration failure. A support vector machine (SVM) model can be used for secondary verification of the risk probability. The model training set is constructed from over 5000 pieces of equipment operation data and fault records from the past year, with abnormal fault data accounting for 15%. The input features of the SVM model are a three-dimensional feature vector composed of the fault confidence of the preliminary diagnosis results, the frequency of load state abrupt changes, and the slope of the state change trend. The output target is a binary label representing the true state of the equipment (i.e., agglomeration failure or normal operation). The Platt scaling method is used to convert the output boundary distance into a calibrated actual agglomeration probability. The training process uses a radial basis function kernel with a penalty coefficient (C) set to 1.0 and a kernel parameter (Gamma) set to 0.1. The model optimization algorithm employs Sequential Minimum Optimization (SMO), with an error tolerance of 0.001 and a maximum iteration limit of 1000. Model training is considered complete when the convergence change of the objective function falls below the aforementioned error tolerance. Verified risk probabilities can further improve the accuracy of early warnings.

[0076] If the risk probability exceeds the preset risk assessment threshold, a warning signal for material agglomeration is generated. The risk assessment threshold is set based on historical data of warnings and interventions for material agglomeration failures in metal powder processing equipment over the past year. The lowest risk probability value for all successful interventions that prevented unplanned equipment downtime is calculated, and the basic risk assessment threshold is set to 0.7. For high-value machine head component processing scenarios with high safety requirements, the threshold can be increased to 0.6, while for conventional coarse powder processing scenarios, it can be decreased to 0.75. This threshold has been validated through long-term operation under multiple conditions, effectively reducing the false alarm rate of invalid warnings while ensuring the foresight of the warnings. The warning signal adopts a hierarchical generation mechanism: a level 1 warning signal is generated when the risk probability is between 0.7 and 0.8, corresponding to a suggestive warning; a level 2 warning signal is generated when the risk probability is between 0.8 and 0.9, corresponding to a planned maintenance warning; and a level 3 warning signal is generated when the risk probability is above 0.9, corresponding to an emergency shutdown warning. Different levels of warning signals correspond to different trigger actions. When there is no risk probability exceeding the threshold, no warning signal is generated; only the equipment health status record is updated.

[0077] Finally, the warning signals and risk probability data are integrated to form the basis for anomaly intervention. The integration process employs structured data encapsulation, using the warning signal's generation timestamp as a unique identifier. Five key information categories—warning signal level, material agglomeration risk probability, fault occurrence time, anomaly distribution characteristics, and equipment operating parameters—are fully encapsulated to form standardized anomaly intervention criteria. Simultaneously, the encapsulation process supplements two key information categories from the preliminary diagnostic results: fault root cause determination and fault confidence level. This ensures that the anomaly intervention criteria not only include clear warning levels and risk quantification indicators but also cover complete diagnostic information such as fault root cause, location, and development trend. This allows for direct guidance of on-site maintenance personnel to perform corresponding intervention operations and can also be directly input into the equipment control system to execute automated parameter adjustments and shutdown protection actions. All integrated data maintains temporal alignment and dimensional consistency, with no missing information or data misalignment issues. The complete anomaly intervention criteria are also synchronously archived into the equipment health management database for subsequent iterative optimization of the fault diagnosis model.

[0078] It should be noted that the methods for determining the preset thresholds involved in this invention uniformly follow the following calibration process: The initial thresholds are determined as follows for newly installed equipment or equipment that has been in operation for less than one year: Overlap judgment threshold: 0.7, based on the temporal overlap distribution of 20 known cluster faults and 20 normal fluctuations, taking the threshold that maximizes the Youden index; Anomaly matching threshold: 0.7, based on the matching distribution of 30 known cluster faults and 30 non-cluster anomalies, taking the threshold corresponding to the maximum F1-score; Risk judgment threshold: 0.75, determined based on the statistical equilibrium point of successful interventions and false alarms in historical early warning records; Fluctuation threshold… Based on the 3σ principle under the assumption of normal distribution; the mutation judgment threshold is 8% of the rated load rate, determined based on the 99.5 percentile of the first-order difference of the load rate when the equipment is operating normally. The above initial thresholds were determined by ROC curve analysis.

[0079] Subsequently, the threshold is adaptively adjusted based on local data. When the cumulative number of valid fault samples accumulated locally by the device exceeds the preset number... After this, automatic threshold adaptive adjustment is initiated, triggering a threshold update every N_min confirmed fault records; the classification threshold is redefined using the method of maximizing the Youden exponent, and the old and new thresholds are weighted and smoothed. ,in Each time the threshold is updated, it is verified on backtesting data. If the diagnostic accuracy or recall drops by more than 5 percentage points, it is reverted to the old threshold.

[0080] The thresholds are adaptively adjusted for different operating conditions. Three standard operating condition types are preset: ball mill ultrafine processing condition (overlap detection threshold +0.05, abnormal matching threshold +0.05, risk detection threshold -0.05); conventional mixing processing condition (all threshold offsets are 0, serving as the baseline); and light-load precision processing condition (overlap detection threshold -0.05, abnormal matching threshold -0.05, risk detection threshold +0.05). The operating condition type is automatically determined by a pre-trained decision tree classifier based on the current average load rate and power factor fluctuation frequency.

[0081] To verify the effectiveness of the method of the present invention, a field test was conducted on three ball mills in a metal powder processing enterprise for a period of 6 months. The equipment parameters were QM-300 planetary ball mills, with a drive motor power of 30kW, a rated speed of 1450rpm, and the processed material was metal powder (average particle size 50-200μm), operating for 16 hours a day.

[0082] 4320 hours of normal operation data (approximately 180 days, continuously collected from 3 devices); 47 confirmed material agglomeration fault events (confirmed by on-site maintenance personnel); 23 non-agglomeration abnormal events (8 mechanical wear events, 6 electrical fault events, and 9 overload events); 12 accidental shutdown events.

[0083] Of the 47 confirmed material agglomeration faults, the method of this invention successfully diagnosed 44, with a diagnostic accuracy of 93.6%, and missed 3 (missed detection rate of 6.4%). Of the 23 non-agglomeration-related anomalies, 21 were correctly excluded (specificity 91.3%), and 2 were falsely identified as agglomeration faults. Compared with the fixed threshold alarm method (accuracy 61.7%, recall 57.4%, false alarm rate 18.5%, average warning lead time 12 minutes, F1-score 0.532), the method of this invention is comprehensively superior to traditional methods. Compared with the single power factor monitoring method (accuracy 72.3%, recall 68.1%, false alarm rate 12.8%, average warning lead time 21 minutes, F1-score 0.679), it also shows a significant improvement.

[0084] When the overlap determination threshold varies from 0.5 to 0.9, the accuracy ranges from 85.1% to 82.9%, and the recall ranges from 97.9% to 76.6%, with the F1-score reaching its optimum near 0.7. When the anomaly matching threshold varies from 0.5 to 0.9, the accuracy ranges from 80.9% to 83.1%, and the recall ranges from 95.7% to 74.5%. When the fluctuation threshold coefficient k varies from 2.0 to 4.0, the false alarm rate increases significantly when k < 2.5. These data indicate that within a reasonable fluctuation range of ±0.1 for each threshold, the diagnostic accuracy changes by no more than 5 percentage points, demonstrating strong robustness.

[0085] Compared to fixed threshold alarm methods, the early warning lead time of this invention is increased from an average of 12 minutes to 38 minutes, providing maintenance personnel with an additional 26-minute response window. During a 6-month testing period, the number of unplanned shutdowns of equipment using this invention was 2, a 94.8% reduction compared to the same model of equipment using traditional fixed threshold alarms (an average of 3.2 shutdowns per month). Compared to single power factor monitoring methods, through multi-dimensional feature fusion and time-series correlation verification of material agglomeration events, the false alarm rate was reduced from 12.8% to 4.3%.

[0086] The comparison results with existing technologies are as follows: In summary, this invention discloses a fault diagnosis method for metal powder processing equipment. The method includes collecting power factor data and load status data of the motor of the metal powder processing equipment; extracting periodic variation features to construct a fluctuation feature vector; analyzing peak distribution density and duration to locate abnormal fluctuation intervals; calculating the overlap between abnormal timing sequences and material agglomeration events to mark high-risk points; and comparing with historical mutation records to obtain the fluctuation pattern matching degree. When a threshold is exceeded, a deep fault caused by material agglomeration is determined; the probability of agglomeration risk is calculated by combining load mutation characteristics and state change trends; and a graded early warning signal is generated as the basis for abnormal intervention. This invention solves the problems of traditional single-parameter monitoring being susceptible to operating condition interference, missing early hidden faults, and biased fault root cause judgment, achieving early and accurate diagnosis and proactive early warning of equipment faults.

[0087] Reference Figure 2 The second embodiment of the present invention provides a fault diagnosis system for metal powder processing equipment, comprising: The data acquisition module is used to acquire power factor data and load status data of the motor of the metal powder processing equipment; The feature construction module is used to extract periodic variation features from the power factor data and the load state data, and construct a fluctuation feature vector based on the periodic variation features; Anomaly identification module is used to perform feature analysis on the fluctuation feature vector, determine the abnormal fluctuation range, and extract the abnormal distribution features within the abnormal fluctuation range; The time-series matching module is used to extract timestamp information from the abnormal distribution features, calculate the time-series overlap between the timestamp information and the pre-acquired material clumping event, and if the time-series overlap is greater than the preset overlap judgment threshold, mark the abnormal fluctuation interval as a high-risk point, and combine the time-series overlap, the fluctuation feature vector and the material clumping event to obtain the time correlation strength. The pattern comparison module is used to filter corresponding historical mutation records based on the time correlation strength, and to perform dynamic time normalization on the fluctuation feature vectors corresponding to the historical mutation records and the high-risk points to obtain the fluctuation pattern matching degree. The fault determination module is used to determine that the current anomaly is a deep fault caused by material agglomeration if the fluctuation pattern matching degree is greater than the preset anomaly matching threshold, and to obtain a preliminary diagnosis result. The early warning output module is used to calculate the load change characteristics based on the load status data and determine the risk probability of material agglomeration in combination with the preliminary diagnosis results. When the risk probability exceeds the preset risk judgment threshold, a graded early warning signal is generated as the basis for abnormal intervention.

[0088] It should be noted that the metal powder processing equipment fault diagnosis system provided in this embodiment of the invention is used to execute all the process steps of the metal powder processing equipment fault diagnosis method in the above embodiment. The working principle and beneficial effect of the two are one-to-one, so they will not be described again.

[0089] It should be noted that the system embodiments described above are merely illustrative. The units described as separate components may or may not be physically separate, and the components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of this embodiment according to actual needs. Furthermore, in the accompanying drawings of the system embodiments provided by this invention, the connection relationships between modules indicate that they have communication connections, which can be specifically implemented as one or more communication buses or signal lines. Those skilled in the art can understand and implement this without any creative effort.

[0090] The specific embodiments described above further illustrate the purpose, technical solution, and beneficial effects of the present invention. It should be understood that the above descriptions are merely specific embodiments of the present invention and are not intended to limit the scope of protection of the present invention. In particular, it should be noted that any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention for those skilled in the art.

Claims

1. A fault diagnosis method for metal powder processing equipment, characterized in that, include: Obtain power factor data and load status data of the motor of the metal powder processing equipment; Extracting periodic variation features from the power factor data and the load state data, and constructing a fluctuation feature vector based on the periodic variation features, includes: analyzing the time series of the power factor data to extract fluctuation amplitude features; analyzing the time series of the load state data to extract abrupt change frequency features; extracting periodic variation features based on the time series periodic patterns of the power factor data and the load state data; and integrating the fluctuation amplitude features, the abrupt change frequency features, and the periodic variation features to construct a fluctuation feature vector. Perform feature analysis on the fluctuation feature vector to determine the abnormal fluctuation range and extract the abnormal distribution features within the abnormal fluctuation range; Timestamp information is extracted from the abnormal distribution features, and the temporal overlap between the timestamp information and the pre-acquired material clumping event is calculated. If the temporal overlap is greater than the preset overlap judgment threshold, the abnormal fluctuation interval is marked as a high-risk point. The temporal correlation strength is obtained by combining the temporal overlap, the fluctuation feature vector, and the material clumping event. Based on the time correlation strength, the corresponding historical mutation records are filtered, and the fluctuation feature vectors corresponding to the historical mutation records and the high-risk points are dynamically time-normalized to obtain the fluctuation pattern matching degree. If the fluctuation pattern matching degree is greater than the preset abnormal matching threshold, the current abnormality is determined to be a deep fault caused by material agglomeration, and a preliminary diagnosis result is obtained. The load change characteristics are calculated based on the load status data, and the risk probability of material agglomeration is determined in combination with the preliminary diagnosis results. When the risk probability exceeds the preset risk judgment threshold, a graded early warning signal is generated as the basis for abnormal intervention.

2. The fault diagnosis method for metal powder processing equipment according to claim 1, characterized in that, The step of performing feature analysis on the fluctuation feature vector to determine the abnormal fluctuation range and extracting the abnormal distribution features within the abnormal fluctuation range includes: Based on the fluctuation feature vector, the distribution density of abnormal peaks and the duration of fluctuations are statistically analyzed to obtain fluctuation statistics. Based on the fluctuation statistics, the time interval in which the fluctuation value exceeds the preset fluctuation threshold is located, and the abnormal fluctuation interval is determined. Extract the data distribution characteristics of the abnormal fluctuation range to obtain the abnormal distribution characteristics.

3. The fault diagnosis method for metal powder processing equipment according to claim 1, characterized in that, The step of extracting timestamp information from the abnormal distribution features, calculating the temporal overlap between the timestamp information and pre-acquired material clumping events, and marking the abnormal fluctuation interval as a high-risk point if the temporal overlap is greater than a preset overlap judgment threshold, and combining the temporal overlap, the fluctuation feature vector, and the material clumping events to obtain the time correlation strength, includes: Extract the timestamp information of the collection period corresponding to the abnormal distribution features to generate a timestamp sequence; Compare the timestamp sequence with the occurrence time periods of pre-acquired material agglomeration events to calculate the time sequence overlap. If the time series overlap is greater than the preset overlap judgment threshold, the corresponding abnormal fluctuation interval is marked as a high-risk point. The fluctuation intensity is extracted from the fluctuation feature vector, and the frequency of agglomeration events is statistically analyzed from the pre-acquired material agglomeration events. The time correlation strength is calculated by combining the temporal overlap, the fluctuation intensity, and the frequency of the block event.

4. The fault diagnosis method for metal powder processing equipment according to claim 1, characterized in that, The step of filtering corresponding historical mutation records based on the time correlation strength, and performing dynamic time warping on the fluctuation feature vectors corresponding to the historical mutation records and the high-risk points to obtain the fluctuation pattern matching degree includes: Based on the time correlation strength, retrieve the pre-acquired historical mutation records, which include historical load mutation response data and state change trend data; Calculate the temporal similarity between the fluctuation feature vector corresponding to the high-risk point and the historical mutation record to obtain the correlation index; The fluctuation pattern matching degree is determined by weighted fusion calculation based on the time correlation strength and the correlation index.

5. The fault diagnosis method for metal powder processing equipment according to claim 1, characterized in that, If the fluctuation pattern matching degree is greater than a preset anomaly matching threshold, then the current anomaly is determined to be a deep fault caused by material agglomeration, and a preliminary diagnostic result is obtained, including: The real-time signal update frequency for obtaining the load status data; The fluctuation pattern matching degree is compared with a preset abnormal matching threshold; If the fluctuation pattern matching degree is greater than the preset abnormal matching threshold, then by combining the preset load peak triggering condition and the real-time signal update frequency, it is determined that the current abnormality belongs to a deep fault caused by material agglomeration, and a fault determination result is obtained. By integrating the fault determination results with the abnormal distribution characteristics, a preliminary diagnostic result is obtained.

6. The fault diagnosis method for metal powder processing equipment according to claim 1, characterized in that, The process involves calculating load mutation characteristics based on the load status data and determining the risk probability of material agglomeration in conjunction with the preliminary diagnostic results. When the risk probability exceeds a preset risk judgment threshold, a graded early warning signal is generated as a basis for abnormal intervention, including: Based on the load status data, calculate the frequency of load status abrupt changes and the slope of the status change trend; Based on the preliminary diagnostic results, the frequency of load state abrupt changes, and the slope of the state change trend, the probability of material agglomeration is calculated. If the risk probability exceeds the preset risk judgment threshold, a warning signal for material agglomeration is generated; The warning signal is integrated with the risk probability to determine the basis for abnormal intervention.

7. A fault diagnosis system for metal powder processing equipment, characterized in that, include: The data acquisition module is used to acquire power factor data and load status data of the motor of the metal powder processing equipment; The feature construction module is used to extract periodic variation features from the power factor data and the load state data, and construct a fluctuation feature vector based on the periodic variation features. This includes: analyzing the time series of the power factor data to extract fluctuation amplitude features; analyzing the time series of the load state data to extract abrupt change frequency features; extracting periodic variation features based on the time series periodic patterns of the power factor data and the load state data; and integrating the fluctuation amplitude features, the abrupt change frequency features, and the periodic variation features to construct a fluctuation feature vector. Anomaly identification module is used to perform feature analysis on the fluctuation feature vector, determine the abnormal fluctuation range, and extract the abnormal distribution features within the abnormal fluctuation range; The time-series matching module is used to extract timestamp information from the abnormal distribution features, calculate the time-series overlap between the timestamp information and the pre-acquired material clumping event, and if the time-series overlap is greater than the preset overlap judgment threshold, mark the abnormal fluctuation interval as a high-risk point, and combine the time-series overlap, the fluctuation feature vector and the material clumping event to obtain the time correlation strength. The pattern comparison module is used to filter corresponding historical mutation records based on the time correlation strength, and to perform dynamic time normalization on the fluctuation feature vectors corresponding to the historical mutation records and the high-risk points to obtain the fluctuation pattern matching degree. The fault determination module is used to determine that the current anomaly is a deep fault caused by material agglomeration if the fluctuation pattern matching degree is greater than the preset anomaly matching threshold, and to obtain a preliminary diagnosis result. The early warning output module is used to calculate the load change characteristics based on the load status data and determine the risk probability of material agglomeration in combination with the preliminary diagnosis results. When the risk probability exceeds the preset risk judgment threshold, a graded early warning signal is generated as the basis for abnormal intervention.

Images