Artificial Intelligence-Based Fault Prediction Methods and Systems

Abstract

This invention discloses an artificial intelligence-based fault prediction method and system, belonging to the field of fault prediction technology. Through a technical chain of "data quality assessment—multi-dimensional anomaly determination—process-aware correction—labelless time-series modeling," this invention solves the problem of weak model generalization ability caused by the scarcity of fault samples, overcomes false alarms and missed alarms caused by ignoring the dynamic characteristics of the brick-making process, and ensures the accuracy of the prediction input through refined cleaning and reconstruction of sensor data. This method is not only applicable to the health monitoring of the extrusion screw shaft of a brick-making machine, but also provides a transferable intelligent prediction paradigm for other high-noise, strongly process-coupled industrial equipment, significantly improving the feasibility and engineering robustness of predictive maintenance in actual production lines.

Description

Technical Field

[0001] This invention relates to the field of fault prediction technology, and in particular to fault prediction methods and systems based on artificial intelligence. Background Technology

[0002] In the modern building materials industry, brick-making machines are core production equipment, and their operational stability directly affects the quality of brick forming, production efficiency, and the overall operating costs of the enterprise. A brick-making machine typically consists of multiple subsystems, including a feeding system, mixing device, extrusion mechanism, cutting system, and transmission and hydraulic control unit. Among these, the extrusion auger (or main shaft) is a key actuating component, responsible for the core function of pushing the mixed materials under high pressure and forming them through a mold. Due to prolonged exposure to harsh conditions of high load, high wear, high temperature and humidity, and dense dust, the extrusion auger and its associated bearings and seals are highly susceptible to abnormal wear, fatigue cracks, eccentric vibration, and even breakage. Once such a failure occurs, it can not only cause the entire production line to shut down, resulting in a large amount of waste and raw material waste, but may also trigger a chain reaction of equipment damage and even endanger the safety of operators.

[0003] Traditionally, brick manufacturing companies have mostly adopted a strategy of periodic inspections or reactive maintenance, that is, setting fixed maintenance cycles based on experience, or only carrying out emergency repairs after equipment has completely failed. However, this "passive" maintenance model has obvious drawbacks: on the one hand, overly conservative maintenance cycles can cause unnecessary downtime and labor costs; on the other hand, if a failure occurs between two maintenance periods, it can easily lead to unplanned downtime, seriously affecting order delivery. In recent years, some companies have tried to introduce simple condition monitoring methods based on threshold alarms, such as installing vibration sensors to monitor the vibration amplitude of the spindle, triggering an alarm when it exceeds a preset limit. However, such methods can only reflect the state when the failure has developed to a relatively serious stage, lacking sensitivity to early weak abnormal signals, and have difficulty distinguishing between normal operating condition fluctuations and true deterioration trends, resulting in high false alarm and false negative rates.

[0004] To overcome the aforementioned limitations, the industry has already attempted to introduce artificial intelligence (AI) technology into the field of brick-making machine fault prediction, particularly for anomaly identification in extrusion screw shaft systems. A typical approach involves collecting multi-source sensor signals such as spindle vibration, motor current, hydraulic pressure, and temperature, and then using machine learning models (such as support vector machines and random forests) or shallow neural networks to train historical operating data to classify and warn of typical faults such as screw shaft wear, bearing loosening, or imbalance. Some solutions further employ time-series modeling methods (such as LSTM) to capture fault evolution trends, attempting to issue predictive alerts in the early stages of a fault. Compared to traditional threshold alarms, these AI-based methods improve the ability to identify subtle anomalies to a certain extent and can adapt to a certain range of operating condition changes. However, existing AI applications still have significant shortcomings: First, most models rely on a large number of well-labeled fault samples, while actual fault data from brick-making sites is scarce and labels are difficult to obtain, which limits the generalization ability of the models; second, existing methods often treat brick-making machines as general rotating machinery and do not fully integrate the influence of brick-making process characteristics on the spindle state, resulting in poor prediction stability under complex dynamic working conditions; third, raw sensor data is often affected by signal loss, sudden interference, or sensor drift. If it is used directly as model input without effective cleaning and correction, the prediction reliability will be significantly reduced.

[0005] Therefore, there is an urgent need for technical solutions for fault prediction methods and systems based on artificial intelligence. Summary of the Invention

[0006] To address the aforementioned technical problems, this invention provides a fault prediction method based on artificial intelligence, specifically comprising the following steps: S1. Collect multi-source sensor data of the extrusion screw shaft station during the operation of the brick making machine, and simultaneously acquire brick making process parameters aligned with the timestamp of the multi-source sensor data. The multi-source sensor data includes at least: spindle vibration signal, drive motor current signal, hydraulic system pressure signal, and bearing temperature signal; the brick-making process parameters include at least: raw material moisture content, feeding rate, and current production batch identifier; S2. Perform data quality assessment on the multi-source sensor data, filter to obtain a data subset, and judge each sensor data in the data subset. Based on the judgment results, obtain the sensor data with anomalies and the sensor data without anomalies. Perform correction processing on the sensor data with anomalies to generate corrected multi-source sensor data. Integrate the corrected multi-source sensor data with the sensor data without anomalies to obtain the final multi-source sensor data. S21. Divide the multi-source sensor data into segments according to a preset time window, and filter the data subsets based on the segmentation results; S211. Divide the multi-source sensor data into continuous segments according to a fixed-length time window to obtain the sensor data within each time window. S212. Calculate the missing rate for each sensor data within each time window; S213. Calculate the mutation rate for each sensor data within each time window; S214. Set a first threshold for the deletion rate and a second threshold for the mutation rate. S215. Determine whether each time window meets the retention conditions. The retention conditions are: the missing rate of all sensor data within the time window is less than the first threshold, and the mutation rate of all sensor data is less than the second threshold; retain each sensor data within the time window that meets the retention conditions as a data subset. S22. Obtain the statistical distribution characteristics of each sensor data in the data subset, and extract the physical constraint relationships that each sensor data in the data subset should satisfy during the operation of the extrusion screw shaft. The statistical distribution characteristics of each sensor data in the data subset include: the mean, standard deviation, and local outlier factor of each sensor data in the data subset; the physical constraints that each sensor data in the data subset should satisfy during the operation of the extrusion screw shaft include: the positive correlation constraint between the amplitude of the main shaft vibration signal of the extrusion screw shaft and the amplitude of the drive motor current signal of the extrusion screw shaft, the nonlinear mapping constraint between the hydraulic system pressure signal of the extrusion screw shaft and the feeding rate of the extrusion screw shaft, and the thermal balance constraint between the rate of change of the bearing temperature signal of the extrusion screw shaft and the rotational speed of the extrusion screw shaft; S23. Based on the statistical distribution characteristics and physical constraints, determine any sensor data in the data subset, and obtain the sensor data with anomalies and the sensor data without anomalies based on the determination results. S231. For the current sensing data in the data subset, obtain the mean and standard deviation of the statistical distribution characteristics of the sensing data of the same type as the current sensing data in the data subset, calculate the absolute value of the deviation between the current sensing data and the mean, and determine whether the absolute value of the deviation is greater than a preset multiple of the standard deviation; if it is greater, mark the current sensing data as a first type of statistical anomaly candidate. S232. For the current sensing data in the data subset, obtain the local outlier factor in the statistical distribution characteristics of the sensing data of the same type as the current sensing data in the data subset, and determine whether the local outlier factor is greater than a third preset threshold; if it is greater, mark the current sensing data as a second type of statistical anomaly candidate. S233. If the current sensing data is marked as a first-class statistical anomaly candidate or a second-class statistical anomaly candidate, then mark the current sensing data as a statistical anomaly candidate. S234. For the current sensing data in the data subset, perform the following physical constraint verification operation: When the current sensing data is the amplitude of the main shaft vibration signal of the extrusion screw shaft, the amplitude of the drive motor current signal of the extrusion screw shaft at the same timestamp is obtained, and it is determined whether the amplitude of the main shaft vibration signal of the extrusion screw shaft and the amplitude of the drive motor current signal of the extrusion screw shaft satisfy the positive correlation constraint. When the current sensing data is the amplitude of the current signal of the drive motor of the extrusion screw shaft, the amplitude of the main shaft vibration signal of the extrusion screw shaft at the same timestamp is obtained, and it is determined whether the amplitude of the current signal of the drive motor of the extrusion screw shaft and the amplitude of the main shaft vibration signal of the extrusion screw shaft satisfy the positive correlation constraint. When the current sensing data is the hydraulic system pressure signal of the extrusion screw shaft, the feeding rate of the extrusion screw shaft at the same time stamp is obtained, and it is determined whether the hydraulic system pressure signal of the extrusion screw shaft and the feeding rate satisfy the nonlinear mapping constraint. When the current sensing data is the rate of change of the bearing temperature signal of the extrusion screw shaft, the rotational speed of the extrusion screw shaft at the same time stamp is obtained, and it is determined whether the rate of change of the bearing temperature signal of the extrusion screw shaft and the rotational speed of the extrusion screw shaft satisfy the thermal balance constraint. If the current sensing data belongs to any of the above types and does not meet the corresponding physical constraints, then the current sensing data is marked as a physical anomaly candidate. S235. The current sensing data is determined to be abnormal sensing data only if condition one is met, wherein condition one is: if the current sensing data is simultaneously marked as a statistical anomaly candidate and a physical anomaly candidate; if condition one is not met, the current sensing data is determined to be abnormal sensing data. S24. Obtain the sensor type, normal sampling value sequence in the timestamp neighborhood, and current brick-making process parameters of the sensor data with abnormality. Based on the sensor type, normal sampling value sequence, and current brick-making process parameters, perform weighted reconstruction on the current sensor data to generate corrected multi-source sensor data. Then, integrate the corrected multi-source sensor data with the sensor data without abnormality to obtain the final multi-source sensor data. S241. Determine the basic correction mode according to the sensor type. The basic correction mode includes linear interpolation mode, historical similar working condition backfill mode or physical model derivation mode, and obtain the output result based on the determined basic correction mode. S242. Based on the current brick-making process parameters, match the reference dataset under the same or similar process conditions in the historical normal operation database, and calculate the similarity between the normal sampled value sequence in the timestamp neighborhood and the reference dataset. S243. Based on the sensor type, the local trend characteristics of the normal sampled value sequence and the similarity, generate the first weight, the second weight and the third weight respectively, and perform weighted fusion on the output of the basic correction mode to obtain the corrected multi-source sensing data. S244. Replace the original abnormal sensor data with the corrected multi-source sensor data, and merge it with other sensor data without abnormalities in the original timestamp order to form the final multi-source sensor data. S3. The final multi-source sensor data is fused with the brick-making process parameters to construct a joint input vector that includes equipment status features and process context features. S4. Input the joint input vector into the pre-trained semi-supervised temporal anomaly detection model, output the health status score sequence of the extrusion spiral shaft, and determine whether there are early fault signs based on the dynamic change trend of the health status score sequence.

[0007] This embodiment also discloses an artificial intelligence-based fault prediction system, including the following modules: Data acquisition module: used to collect multi-source sensor data at the station where the extrusion screw shaft is located during the operation of the brick making machine, and simultaneously acquire brick making process parameters aligned with the timestamp of the multi-source sensor data; Data correction module: Connected to the data acquisition module, it is used to assess the data quality of multi-source sensor data, filter out data subsets, and judge each sensor data in the data subset. Based on the judgment results, it obtains sensor data with anomalies and sensor data without anomalies. It performs correction processing on the sensor data with anomalies to generate corrected multi-source sensor data, and integrates the corrected multi-source sensor data with the sensor data without anomalies to obtain the final multi-source sensor data. Data fusion module: Connected to the data correction module and the data acquisition module, it is used to fuse the final multi-source sensor data with the brick-making process parameters to construct a joint input vector containing equipment status features and process context features; Fault identification module: Connected to the data fusion module, it is used to input the joint input vector into the pre-trained semi-supervised time-series anomaly detection model, output the health status score sequence of the extrusion screw shaft, and determine whether there are early fault signs based on the dynamic change trend of the health status score sequence.

[0008] The embodiments of the present invention have the following technical effects: This invention constructs an intelligent fault prediction process for the extrusion screw shaft of a brick-making machine, effectively addressing the shortcomings of existing methods in terms of data quality, process adaptability, and model generalization ability. First, in the data preprocessing stage, a time-window-based missing rate and mutation rate evaluation mechanism is introduced. Combined with statistical distribution characteristics and the unique physical constraints of the extrusion screw shaft (such as the positive correlation between spindle vibration and motor current, the nonlinear mapping between hydraulic pressure and feeding rate, and the thermal balance relationship between bearing temperature rise and rotational speed), the original multi-source sensor data is anomaly identified and weighted for reconstruction correction, thereby obtaining high-fidelity final multi-source sensor data and significantly improving the reliability of the input data. Second, abandoning the reliance on fault labels, a semi-supervised time-series anomaly detection model trained on normal operating data is adopted. The corrected sensor data is integrated with brick-making process parameters such as raw material moisture content and feeding rate into a joint input vector. This enables the model to distinguish between normal state deviations caused by process fluctuations and true anomalies caused by mechanical deterioration, thus achieving stable and accurate identification of early fault symptoms without requiring a large number of fault samples.

[0009] In summary, this invention addresses the problem of weak model generalization ability caused by scarce fault samples through a technical chain of "data quality assessment—multi-dimensional anomaly detection—process-aware correction—labelless time-series modeling." It overcomes false alarms and false negatives caused by ignoring the dynamic characteristics of the brick-making process. Furthermore, through refined cleaning and reconstruction of sensor data, it ensures the accuracy of predictive inputs. This method is not only applicable to health monitoring of the extrusion auger shaft of brick-making machines but also provides a transferable intelligent prediction paradigm for other high-noise, strongly process-coupled industrial equipment, significantly improving the feasibility and engineering robustness of predictive maintenance in actual production lines. Attached Figure Description

[0010] To more clearly illustrate the specific embodiments of the present invention or the technical solutions in the prior art, the drawings used in the description of the specific embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of the present invention. For those skilled in the art, other drawings can be obtained from these drawings without creative effort.

[0011] Figure 1 This is a flowchart of the fault prediction method based on artificial intelligence provided in an embodiment of the present invention; Figure 2 This is a framework diagram of the fault prediction system based on artificial intelligence provided in the embodiments of the present invention. Detailed Implementation

[0012] To make the objectives, technical solutions, and advantages of this invention clearer, the technical solutions of this invention will be clearly and completely described below. Obviously, the described embodiments are only a part of the embodiments of this invention, and not all of them. Based on the embodiments of this invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this invention.

[0013] Example 1: As Figure 1 As shown, the fault prediction method based on artificial intelligence provided by this invention includes the following steps: S1. Collect multi-source sensor data of the extrusion screw shaft station during the operation of the brick making machine, and simultaneously acquire brick making process parameters aligned with the timestamp of the multi-source sensor data. The multi-source sensor data includes at least: spindle vibration signal, drive motor current signal, hydraulic system pressure signal, and bearing temperature signal; the brick-making process parameters include at least: raw material moisture content, feeding rate, and current production batch identifier; It is worth noting that in this embodiment, firstly, corresponding sensors are deployed at the station where the extrusion screw shaft of the brick-making machine is located to collect multi-source sensor data: a three-dimensional vibration sensor is installed near the bearing housing of the extrusion screw shaft to acquire the main shaft vibration signal in real time; a high-precision current transformer is connected to the power supply circuit of the drive motor to acquire the drive motor current signal; a pressure transmitter is installed in the main oil circuit of the hydraulic system to monitor the hydraulic system pressure signal; and a temperature sensor is embedded on the housing surface near the rear bearing of the extrusion screw shaft to acquire the bearing temperature signal. All sensors are synchronously acquired through an industrial-grade data acquisition card at a sampling frequency of not less than 1 kHz, and a time synchronization protocol is used to ensure that each signal has a unified time reference. At the same time, brick-making process parameters that are strictly aligned with the timestamps of the above-mentioned multi-source sensor data are read in real time from the upper control system or process database of the brick-making production line: the moisture content of the raw materials is measured and recorded by an online moisture meter before the mixture enters the extruder; the feeding rate is calculated by the weighing sensor and speed encoder on the feeding belt; and the current production batch identifier is generated by the production management system at the start of each batch and continuously broadcast. All collected multi-source sensor data and brick-making process parameters are paired and stored with a unified timestamp to form time-aligned data records, providing a basic input for subsequent data quality assessment and fusion modeling. The entire data acquisition process is carried out continuously during the continuous operation of the brick-making machine, covering various operating conditions with different raw material ratios, moisture content ranges, and production cycles, ensuring the representativeness and completeness of the data.

[0014] S2. Perform data quality assessment on the multi-source sensor data, filter to obtain a data subset, and judge each sensor data in the data subset. Based on the judgment results, obtain the sensor data with anomalies and the sensor data without anomalies. Perform correction processing on the sensor data with anomalies to generate corrected multi-source sensor data. Integrate the corrected multi-source sensor data with the sensor data without anomalies to obtain the final multi-source sensor data. S21. Divide the multi-source sensor data into segments according to a preset time window, and filter the data subsets based on the segmentation results; S211. Divide the multi-source sensor data into continuous segments according to a fixed-length time window to obtain the sensor data within each time window. It is worth noting that in this embodiment, the acquired and time-aligned multi-source sensor data is processed in continuous segments according to a fixed-length time window. Specifically, the time window length is set to 10 seconds, which is determined based on the time-domain response period of typical fault characteristics of the extrusion screw shaft of the brick-making machine, ensuring that the operating state under stable conditions can be completely captured. Taking the data recording start time as a reference, starting from the first sampling point, all sensor data within a continuous 10-second interval are sequentially extracted as a time window, and then advanced forward with a sliding step of 1 second to form the next overlapping time window, until the entire acquisition period is covered. Each time window contains the complete sampling point sequence of the spindle vibration signal, drive motor current signal, hydraulic system pressure signal, and bearing temperature signal within the 10-second interval. For example, if the sampling frequency is 1000 Hz, each time window contains 10,000 sampling points; if a signal is missing some points due to a temporary communication interruption, its empty value position is retained to maintain the time alignment structure. In this way, the original continuous multi-source sensor data is divided into a series of ordered, equal-length and partially overlapping time windows. Each window serves as an independent data unit for subsequent calculation of missing rate and mutation rate.

[0015] S212. Calculate the missing rate for each sensor data within each time window; It is worth noting that in this embodiment, for a certain type of sensor data (such as spindle vibration signal) within a certain time window, the number of valid sampling points actually collected within the window is first counted. Valid sampling points refer to reasonable values ​​that are not null and do not exceed the sensor's range. Then, based on the preset sampling frequency and time window length, the total number of sampling points that the window should theoretically include is determined. For example, when the time window length is 10 seconds and the sampling frequency is 1000 Hz, the theoretical total number of sampling points is 10,000. Next, the number of invalid or missing sampling points is defined as the theoretical total number of sampling points minus the number of valid sampling points. Finally, the number of missing sampling points is divided by the theoretical total number of sampling points, and the resulting ratio is the missing rate of the sensor data within that time window. The above calculation process is performed independently for the spindle vibration signal, drive motor current signal, hydraulic system pressure signal, and bearing temperature signal in each time window, thereby generating a corresponding missing rate value for each type of sensor data in each time window. All missing rates are expressed as decimals, ranging from 0 to 1, and are used for subsequent comparison with the first threshold to determine whether the time window meets the data integrity requirements.

[0016] S213. Calculate the mutation rate for each sensor data within each time window; It is worth noting that in this embodiment, for a certain type of sensor data (such as a drive motor current signal) within a certain time window, the absolute value of the difference between all adjacent sampling points within the window is first iterated sequentially. Then, a jump threshold is set, which is determined based on the maximum allowable rate of change of this type of sensor data under normal operating conditions. For example, for hydraulic system pressure signals, the jump threshold can be set to 0.5 MPa per millisecond. Next, the number of times the absolute value of the difference exceeds the jump threshold within the time window is counted and recorded as the number of abrupt changes. Then, the total number of adjacent sampling point pairs within the time window is determined as the theoretical comparison count, which is equal to the total number of sampling points within the window minus one. Finally, the number of abrupt changes is divided by the theoretical comparison count, and the resulting ratio is the mutation rate of the sensor data within the time window. The above calculation process is performed independently for the spindle vibration signal, drive motor current signal, hydraulic system pressure signal, and bearing temperature signal in each time window, thereby generating a corresponding mutation rate value for each type of sensor data in each time window. All mutation rates are expressed as decimals, ranging from 0 to 1, and are used for subsequent comparison with the second threshold to determine whether the time window meets the data stability requirements.

[0017] S214. Set a first threshold for the deletion rate and a second threshold for the mutation rate. It is worth noting that in this embodiment, the first threshold is determined based on the minimum requirements for data integrity in monitoring the extrusion screw shaft of the brick-making machine: analysis of historical normal operation data reveals that when the missing rate of any sensor data exceeds 5% within a single time window, the data in that window will not accurately reflect the true state of the equipment. Therefore, the first threshold is set to 0.05. The second threshold is calibrated based on the dynamic characteristics of various sensor data under stable operating conditions, and the most stringent condition is taken as the unified threshold. For example, in a week of continuous fault-free operation records, the mutation rate of the spindle vibration signal does not exceed 0.02, the drive motor current signal does not exceed 0.015, the hydraulic system pressure signal does not exceed 0.03, and the bearing temperature signal does not exceed 0.01. Considering the most sensitive bearing temperature signal among all signals and retaining a safety margin, the second threshold is set to 0.012. The above two thresholds are determined by offline analysis of historical normal datasets before system deployment and are fixed in the judgment logic of the prediction process, without being dynamically adjusted with real-time data. Once set, the first threshold is used to determine if the missing rate exceeds the limit, and the second threshold is used to determine if the mutation rate is abnormal. Together, they form the basis for the retention conditions of the time window screening. The entire setting process is based on actual production line operation experience and data statistical patterns, ensuring that the thresholds are neither too lenient, leading to the inclusion of inferior data, nor too stringent, causing excessive rejection of valid data.

[0018] S215. Determine whether each time window meets the retention conditions. The retention conditions are: the missing rate of all sensor data within the time window is less than the first threshold, and the mutation rate of all sensor data is less than the second threshold; retain each sensor data within the time window that meets the retention conditions as a data subset. It is worth noting that in this embodiment, the missing rate and mutation rate of the spindle vibration signal, drive motor current signal, hydraulic system pressure signal, and bearing temperature signal within the time window are first obtained. Then, the missing rate of each type of sensor data is checked sequentially to see if it is less than a first threshold (e.g., 0.05), and the mutation rate is checked to see if it is less than a second threshold (e.g., 0.012). Only when all four types of sensor data simultaneously satisfy both the "missing rate less than the first threshold" and the "mutation rate less than the second threshold" is the time window considered to meet the retention condition. If the missing rate of any sensor data is not less than the first threshold, or the mutation rate of any sensor data is not less than the second threshold, the time window is completely removed, and all sensor data within it is not included in subsequent processing. For time windows that meet the retention condition, the complete sampling sequences of the spindle vibration signal, drive motor current signal, hydraulic system pressure signal, and bearing temperature signal contained within them are preserved as is and aggregated into a set according to the original time order; this set constitutes a data subset. The entire judgment process is executed window by window, ensuring that the data subset consists only of high-quality, low-interference, and stable time-period data, providing a reliable foundation for subsequent extraction of statistical distribution characteristics and verification of physical constraint relationships. All retention or removal operations are logged, but this does not affect the storage of the original collected data to support subsequent backtracking analysis.

[0019] S22. Obtain the statistical distribution characteristics of each sensor data in the data subset, and extract the physical constraint relationships that each sensor data in the data subset should satisfy during the operation of the extrusion screw shaft. The statistical distribution characteristics of each sensor data in the data subset include: the mean, standard deviation, and local outlier factor of each sensor data in the data subset; the physical constraints that each sensor data in the data subset should satisfy during the operation of the extrusion screw shaft include: the positive correlation constraint between the amplitude of the main shaft vibration signal of the extrusion screw shaft and the amplitude of the drive motor current signal of the extrusion screw shaft, the nonlinear mapping constraint between the hydraulic system pressure signal of the extrusion screw shaft and the feeding rate of the extrusion screw shaft, and the thermal balance constraint between the rate of change of the bearing temperature signal of the extrusion screw shaft and the rotational speed of the extrusion screw shaft; It is worth noting that in this embodiment, the statistical distribution characteristics of each type of sensor data in the data subset are first calculated. For the spindle vibration signal, all sampling points within all time windows in the data subset are traversed, and their arithmetic mean is calculated to obtain the mean. Then, the standard deviation is obtained based on the sum of squared deviations between the mean and each sampling point. Simultaneously, a local outlier algorithm is used to calculate the degree of deviation of each sampling point from the local density in its neighborhood, generating the corresponding local outlier value. The above process is also independently applied to the drive motor current signal, hydraulic system pressure signal, and bearing temperature signal, thereby obtaining the mean, standard deviation, and local outlier for each of the four types of sensor data, forming a complete set of statistical distribution characteristics. Subsequently, the physical constraints that each sensor data point in the data subset should satisfy during the operation of the extrusion screw shaft were extracted: Based on the working mechanism of the brick-making machine, it was confirmed that the amplitude of the main shaft vibration signal of the extrusion screw shaft and the amplitude of the drive motor current signal show a positive correlation trend under steady-state conditions, that is, the vibration amplitude should not decrease significantly when the current increases; it was confirmed that there is a nonlinear mapping relationship between the hydraulic system pressure signal of the extrusion screw shaft and the feeding rate, which is manifested as a monotonically increasing hydraulic pressure but a gradually slowing rate of increase when the feeding rate increases; it was confirmed that the rate of change of the bearing temperature signal of the extrusion screw shaft (i.e., the temperature increment per unit time) and the speed of the extrusion screw shaft must satisfy the thermal balance constraint, that is, the rate of change of temperature should tend to zero at a constant speed, and the rate of change of temperature should briefly increase after a sudden increase in speed but then decay. These physical constraints are solidified in the judgment logic in the form of mathematical inequalities or trend rules for subsequent anomaly identification, ensuring that the model judgment conforms to the physical laws of equipment operation.

[0020] S23. Based on the statistical distribution characteristics and physical constraints, determine any sensor data in the data subset, and obtain the sensor data with anomalies and the sensor data without anomalies based on the determination results. S231. For the current sensing data in the data subset, obtain the mean and standard deviation of the statistical distribution characteristics of the sensing data of the same type as the current sensing data in the data subset, calculate the absolute value of the deviation between the current sensing data and the mean, and determine whether the absolute value of the deviation is greater than a preset multiple of the standard deviation; if it is greater, mark the current sensing data as a first type of statistical anomaly candidate. It is worth noting that in this embodiment, for the current sensing data (e.g., the sampled value of the spindle vibration signal at a certain time point) in the data subset, its type is first identified, that is, it is determined whether it belongs to the spindle vibration signal, drive motor current signal, hydraulic system pressure signal, or bearing temperature signal. Then, all sampling points of the same type as the current sensing data are extracted from the data subset, and the statistical distribution characteristics of this type of sensing data calculated in step S22 are called, specifically obtaining its mean and standard deviation. Next, the difference between the current sensing data and the mean is calculated, and its absolute value is taken to obtain the absolute deviation. The preset multiple is set to 3 based on engineering experience, that is, the three-times-standard-deviation criterion is used as the judgment boundary. The absolute deviation is compared numerically with three times the standard deviation: if the absolute deviation is greater than three times the standard deviation, the current sensing data is considered to have significantly deviated from the normal distribution range and is marked as a first-type statistical anomaly candidate; otherwise, it is not marked. This process is performed one by one for each sampling point in the data subset to ensure that all potential large deviations are identified. For example, when processing a drive motor current signal with a sample value of 85 amps, and the mean of the signal in the data subset is 70 amps and the standard deviation is 4 amps, the absolute value of the deviation is 15 amps, and three times the standard deviation is 12 amps. Since 15 is greater than 12, the sampling point is marked as a candidate for the first type of statistical anomaly.

[0021] S232. For the current sensing data in the data subset, obtain the local outlier factor in the statistical distribution characteristics of the sensing data of the same type as the current sensing data in the data subset, and determine whether the local outlier factor is greater than a third preset threshold; if it is greater, mark the current sensing data as a second type of statistical anomaly candidate. It is worth noting that in this embodiment, for the current sensing data in the data subset (e.g., the sampled value of the bearing temperature signal at a certain moment), the data type to which it belongs is first determined, that is, it is determined whether it belongs to the spindle vibration signal, drive motor current signal, hydraulic system pressure signal, or bearing temperature signal. Then, the local outlier factor value corresponding to the type is retrieved from the statistical distribution characteristics constructed in step S22. This local outlier factor is calculated based on the density distribution relationship between the current sensing data and its neighboring sampling points within the data subset, and is used to characterize its degree of isolation in the local time-series environment. Next, a third preset threshold is set to 1.5. This value is determined by analyzing the distribution of local outliers in historical normal operation data: under fault-free conditions, the local outlier factor of the vast majority of sampling points is less than 1.5, and only a very few exceed it slightly due to minor disturbances. Therefore, 1.5 is used as the discrimination boundary. The local outlier factor of the current sensor data is compared with a third preset threshold: if the local outlier factor is greater than 1.5, it indicates that the sampling point significantly deviates from the normal data cluster in its local time neighborhood, and there is a possibility of an anomaly, so it is marked as a candidate for the second type of statistical anomaly; otherwise, it is not marked. This process is performed independently for each sampling point in the data subset. For example, when the local outlier factor of a pressure signal sampling point in a hydraulic system is 1.8, which exceeds 1.5, it is marked as a candidate for the second type of statistical anomaly.

[0022] S233. If the current sensing data is marked as a first-class statistical anomaly candidate or a second-class statistical anomaly candidate, then mark the current sensing data as a statistical anomaly candidate. It is worth noting that in this embodiment, for each current sensing data in the data subset, after completing the determinations in S231 and S232, it is checked whether it has been marked as a first-type statistical anomaly candidate or a second-type statistical anomaly candidate. Specifically, if the current sensing data is marked in S231 because the absolute value of the deviation exceeds a preset multiple of the standard deviation, or in S232 because its local outlier factor is greater than a third preset threshold, as long as either of these conditions is met, it is considered to have a statistically significant anomaly. In this case, the current sensing data is uniformly classified into the "statistical anomaly candidate" category and a statistical anomaly identifier is attached to it. This identifier does not distinguish the source of the anomaly; it only indicates that the data point deviates from the normal pattern in terms of statistical characteristics, whether due to a large global shift or local isolation. For example, a current signal sampling point of a drive motor may not be marked in S231 (because its deviation is within three standard deviations), but it may be marked in S232 due to a local outlier factor of 1.7 caused by a local mutation, and thus it will still be included in the statistical anomaly candidate. Conversely, if a spindle vibration signal sampling point experiences a slow increase in amplitude due to continuous high load, even if the local outlier factor does not exceed the standard, but the overall deviation from the mean exceeds three standard deviations, it will also be included in the statistical anomaly candidate. This combined judgment mechanism ensures that different types of statistical anomalies can be effectively captured, avoiding missed detections caused by a single criterion. It is worth further explaining that, in this embodiment, S231 to S233 constitute a hierarchical, progressive, complementary and collaborative statistical anomaly judgment mechanism. Its technical effect is that, without relying on fault labels, it quantifies the rationality of sensor data from two dimensions: global distribution and local structure, effectively identifies weak abnormal signals caused by sensor drift, sudden interference or early degradation, and avoids misjudging normal process fluctuations as fault symptoms. Specifically, S231 uses "whether the absolute value of the deviation between the current sensor data and the mean of the same type is greater than a preset multiple of the standard deviation" as the judgment criterion. This judgment is made because during the stable operation of the brick-making machine, various sensor data (such as current and vibration) typically fluctuate slightly around a certain steady-state value, following an approximately normal distribution. The three-times-standard-deviation criterion (i.e., a preset multiple of 3) has been widely verified in engineering practice to cover more than 99.7% of the normal data range. The advantage of this judgment is that it can efficiently filter out large deviations caused by sensor malfunctions or strong interference; the calculation is simple and the physical meaning is clear. The reason why only the mean and standard deviation of data of the same type can be used for judgment is that different sensor signals (such as vibration and temperature) have completely different dimensions, magnitudes, and dynamic characteristics, making cross-type comparisons meaningless; a benchmark must be established based on the historical distribution of similar data to ensure the rationality and comparability of the judgment. S232 introduces a local outlier factor as a second criterion to determine whether the current sensing data is significantly isolated within its temporal neighborhood. This criterion is necessary because some early faults (such as bearing micropitting) may only cause local abrupt changes, with the overall amplitude remaining within the normal range, thus failing to be captured by S231. The local outlier factor, by measuring the density difference between the current point and its neighbors, can sensitively detect such "hidden" anomalies. Its advantage lies in improving the detection capability for non-stationary and transient anomalies, compensating for the blind spots of global statistical methods. The reason it can only be calculated based on a local neighborhood is that the operating state of the extrusion auger shaft has strong temporal correlation; anomalies often manifest as discontinuities from previous and subsequent states, and examining only the global distribution would lose crucial temporal context information. S233 merges the two types of candidates into a unified "statistical anomaly candidate" using a logical "OR" relationship. This design is because the two criteria target different types of anomaly patterns—S231 excels at capturing persistent shifts, while S232 excels at capturing transient mutations; the two complement each other. Using an "AND" relationship would significantly reduce sensitivity, leading to missed detections; while the "OR" relationship maximizes anomaly coverage while ensuring low false positives (due to the need for subsequent physical constraint verification). This hierarchical fusion strategy avoids over-reliance on a single indicator and the introduction of complex models, perfectly meeting the practical needs of brick-making sites for lightweight and highly robust algorithms, and is a key prerequisite for achieving high-precision early fault screening.

[0023] S234. For the current sensing data in the data subset, perform the following physical constraint verification operation: When the current sensing data is the amplitude of the main shaft vibration signal of the extrusion screw shaft, the amplitude of the drive motor current signal of the extrusion screw shaft at the same timestamp is obtained, and it is determined whether the amplitude of the main shaft vibration signal of the extrusion screw shaft and the amplitude of the drive motor current signal of the extrusion screw shaft satisfy the positive correlation constraint. When the current sensing data is the amplitude of the current signal of the drive motor of the extrusion screw shaft, the amplitude of the main shaft vibration signal of the extrusion screw shaft at the same timestamp is obtained, and it is determined whether the amplitude of the current signal of the drive motor of the extrusion screw shaft and the amplitude of the main shaft vibration signal of the extrusion screw shaft satisfy the positive correlation constraint. When the current sensing data is the hydraulic system pressure signal of the extrusion screw shaft, the feeding rate of the extrusion screw shaft at the same time stamp is obtained, and it is determined whether the hydraulic system pressure signal of the extrusion screw shaft and the feeding rate satisfy the nonlinear mapping constraint. When the current sensing data is the rate of change of the bearing temperature signal of the extrusion screw shaft, the rotational speed of the extrusion screw shaft at the same time stamp is obtained, and it is determined whether the rate of change of the bearing temperature signal of the extrusion screw shaft and the rotational speed of the extrusion screw shaft satisfy the thermal balance constraint. If the current sensing data belongs to any of the above types and does not meet the corresponding physical constraints, then the current sensing data is marked as a physical anomaly candidate. It is worth noting that in this embodiment, for the current sensing data in the data subset, corresponding physical constraint verification operations are performed according to its specific type. When the current sensing data is the amplitude of the extrusion screw shaft's main shaft vibration signal, the amplitude of the extrusion screw shaft's drive motor current signal with the same timestamp as the data is first extracted from the data subset. Then, it is determined whether the two satisfy the positive correlation constraint: if the drive motor current signal amplitude is in an increasing or stable high-load state, while the main shaft vibration signal amplitude decreases significantly or is at an extremely low level, it is considered a violation of the positive correlation constraint; conversely, if the current increases and the vibration amplitude increases synchronously or maintains a reasonable proportional relationship, it is considered to satisfy the constraint. When the current sensing data is the amplitude of the extrusion screw shaft's drive motor current signal, the amplitude of the extrusion screw shaft's main shaft vibration signal at the same timestamp is also obtained, and the same positive correlation judgment logic is performed to ensure bidirectional consistency.

[0024] When the current sensing data is the hydraulic system pressure signal of the extrusion screw shaft, the feeding rate of the extrusion screw shaft at the same time stamp is obtained, and it is verified whether the two satisfy the nonlinear mapping constraint: According to the working principle of the brick making machine, the increase in feeding rate should lead to an increase in material resistance, thereby causing the hydraulic system pressure to rise monotonically, but the rate of increase gradually slows down with the increase in feeding rate, showing a saturated nonlinear relationship; if the hydraulic system pressure does not rise accordingly or even falls when the feeding rate is significantly increased, it is determined that the constraint is not satisfied.

[0025] When the current sensing data is the rate of change of the bearing temperature signal of the extrusion screw shaft (i.e., the increment of bearing temperature per unit time), the rotational speed of the extrusion screw shaft at the same timestamp is obtained, and it is determined whether the two satisfy the thermal balance constraint: when the rotational speed of the extrusion screw shaft is stable or changes slowly, the rate of change of the bearing temperature signal should approach zero; if the rotational speed suddenly increases, the rate of change of temperature rises briefly but then decays to a steady state, it is considered normal; however, if the rotational speed is constant and the rate of change of temperature remains positive, or if the rate of change of temperature rises sharply when the rotational speed decreases, it is considered a violation of the thermal balance constraint.

[0026] If the current sensor data belongs to any of the above types, and the corresponding physical constraint verification result is not met, then the current sensor data is marked as a physical anomaly candidate. The entire verification process is strictly based on the equipment operating mechanism and process physical laws, ensuring that only data that simultaneously conforms to statistical laws and physical consistency is considered normal, thereby effectively eliminating misjudgments caused by a single signal anomaly or process disturbance; It is worth further explaining that, in this embodiment, S234 introduces a physical constraint verification mechanism based on the operating mechanism of the extrusion screw shaft of the brick making machine to realize the physical consistency verification of the rationality of the sensing data. Its technical effect is that it effectively distinguishes between anomalies caused by real mechanical deterioration and false anomalies caused by process fluctuations, occasional sensor interference or data coupling mismatch, significantly improving the accuracy and robustness of fault diagnosis, especially under complex dynamic working conditions such as sudden changes in raw material moisture content, intermittent start-stop or drastic load adjustment, it can still maintain stable judgment.

[0027] The reason for using the constraint of "positive correlation between the amplitude of the spindle vibration signal and the amplitude of the drive motor current signal" for bidirectional verification is that during the extrusion process, increased material resistance will simultaneously lead to an increase in the motor load current and intensified screw shaft vibration, and the two are physically strongly coupled. If only the current increases without a vibration response, it may indicate a failure of the vibration sensor; if only the vibration increases suddenly while the current remains stable, it may be due to an external impact rather than an internal fault. The advantage of this judgment is that it utilizes the equipment's internal energy transfer logic to form cross-verification, avoiding false alarms from a single signal. The reason for the necessity of bidirectional verification (i.e., checking the other regardless of whether the current data is vibration or current) is that anomalies can occur in either signal channel, and unidirectional checks will miss symmetrical anomalies; only bidirectional mutual verification can ensure the integrity of the physical relationship.

[0028] The "nonlinear mapping constraint between hydraulic system pressure signal and feeding rate" is based on the hydrodynamic characteristics of the extrusion system: the feeding rate determines the volume of material entering the screw cavity per unit time, directly affecting the hydraulic load and hydraulic back pressure, which exhibit a monotonically increasing but decreasing rate of increase nonlinear relationship. The advantage of this method is its ability to identify typical problems such as hydraulic system leakage, feeding blockage, or pressure sensor drift. The reason for using only a nonlinear mapping instead of a nonlinear proportional model is that changes in material compressibility, friction coefficient, and moisture content during actual brick-making cause the pressure response to exhibit saturation characteristics; forcibly applying linear modeling would misjudge normal high-feeding conditions as abnormal.

[0029] The judgment regarding the "thermal balance constraint between bearing temperature change rate and rotational speed" is based on thermodynamic principles: bearing temperature rise is determined by the balance between frictional heat generation and heat dissipation; when the rotational speed is stable, the temperature rise rate should tend to zero; after a sudden change in rotational speed, the temperature rise rate changes briefly but decays as thermal balance is established. The advantage of this judgment is that it allows for early detection of lubrication failure or bearing wear—such faults often manifest as "unchanged rotational speed but continuously rising temperature." The reason for using "temperature change rate" rather than "absolute temperature value" is that ambient temperature, cooling conditions, etc., can cause absolute temperature drift, while the change rate better reflects the dynamic anomalies of internal heat generation. Furthermore, the reason for relating it to rotational speed is that discussing temperature rise without considering rotational speed is physically meaningless—high temperatures at high speeds may be normal, while slight temperature increases at low speeds may be a sign of a fault.

[0030] In summary, the physical constraint verification of S234 is not a simple stacking of rules, but rather closely follows the force-electricity-heat-fluid multi-physics coupling mechanism of the extrusion screw shaft. Its judgment logic is irreplaceable: statistical methods cannot distinguish between "reasonable fluctuations" and "physical contradictions". Only by introducing process mechanism constraints can high-confidence anomaly identification be achieved under unlabeled conditions.

[0031] S235. The current sensing data is determined to be abnormal sensing data only if condition one is met, wherein condition one is: if the current sensing data is simultaneously marked as a statistical anomaly candidate and a physical anomaly candidate; if condition one is not met, the current sensing data is determined to be abnormal sensing data. It is worth noting that in this embodiment, for each current sensing data in the data subset, after completing the statistical anomaly candidate labeling in S233 and the physical anomaly candidate labeling in S234, the final anomaly determination logic is executed. Specifically, it checks whether the current sensing data is simultaneously labeled as a statistical anomaly candidate and a physical anomaly candidate: if both labels exist, the current sensing data is determined to be an anomaly-prone sensing data; if it is only labeled as one type of candidate, or if neither label exists, the current sensing data is determined to be an anomaly-free sensing data. This determination uses "AND" logic, meaning that both statistical deviation and physical inconsistency conditions must be met simultaneously for it to be considered a true anomaly. For example, if the amplitude of a spindle vibration signal exceeds the local outlier limit due to instantaneous sensor interference (marked as a statistical anomaly candidate), but its amplitude still maintains a positive correlation with the amplitude of the drive motor current signal at the same time point (not marked as a physical anomaly candidate), then it is ultimately determined that there is no anomaly. Conversely, if a hydraulic system pressure signal abnormally increases when the feeding rate is stable (violating nonlinear mapping constraints and marked as a physical anomaly candidate), but its value is still within the historical distribution range (not marked as a statistical anomaly candidate), it is also not determined to be an anomaly. Only when a data point significantly deviates from the normal statistical law and violates the physical mechanism of equipment operation is it confirmed as abnormal sensor data and enters the subsequent correction process. The reason for using "simultaneous satisfaction of statistical anomaly candidates and physical anomaly candidates" as the judgment criterion is that anomaly judgment based on a single dimension is highly susceptible to interference: statistical methods may misjudge normal process fluctuations (such as small current jumps caused by sudden changes in moisture content) as anomalies, while physical constraint verification may suffer from momentary mismatch due to a brief delay in a signal or measurement noise. Only through cross-verification of both can non-fault factors such as occasional interference, process disturbances, or single-point sensor drift be effectively eliminated, ensuring that the identified anomalies have a high degree of confidence. The advantage of this judgment method lies in its significantly improved accuracy and robustness in anomaly detection. In the actual operation of brick-making machines, the working conditions are complex and varied. Relying solely on statistics or physical rules will lead to high false alarm or false negative rates. The "dual-condition joint judgment" mechanism, however, sets up a double verification threshold. It retains sensitivity to early, minor faults (by capturing deviations through statistical methods) while ensuring that the judgment results conform to the inherent laws of equipment operation (by filtering false signals through physical constraints). This achieves a judgment effect close to expert experience even under unlabeled conditions. The reason why only the strict logic of "simultaneous satisfaction" can be used, rather than "OR" relationships or other relaxed strategies, is fundamentally because this invention is geared towards predictive maintenance scenarios, which are extremely sensitive to false alarms—a single false alarm can lead to unnecessary downtime for maintenance, causing production losses. Therefore, high reliability must be the top priority. Engineering practice shows that real mechanical failures usually cause both data distribution shifts and physical relationship disruptions simultaneously, while single anomalies often originate from non-failure factors. Therefore, this "AND" logic is not only a technical choice but also a necessary guarantee for the feasibility of industrial implementation, and cannot be replaced by simplification or relaxation.

[0032] S24. Obtain the sensor type, normal sampling value sequence in the timestamp neighborhood, and current brick-making process parameters of the sensor data with abnormality. Based on the sensor type, normal sampling value sequence, and current brick-making process parameters, perform weighted reconstruction on the current sensor data to generate corrected multi-source sensor data. Then, integrate the corrected multi-source sensor data with the sensor data without abnormality to obtain the final multi-source sensor data. It is worth noting that in this embodiment, for sensor data that has been determined to be abnormal by S235, the sensor type to which it belongs is first identified, that is, it is determined whether the data originates from the spindle vibration signal, drive motor current signal, hydraulic system pressure signal, or bearing temperature signal. Then, with the timestamp of the abnormal data as the center, 50 consecutive sampling points are selected forward and backward from the data subset (i.e., a total of 101 sampling points, the specific number of which can be adjusted according to the sampling frequency and signal dynamic characteristics) to form a timestamp neighborhood. Sensor data of the same type as the abnormal data that has been determined to be non-abnormal are extracted from it to form a normal sampling value sequence within the timestamp neighborhood. If there are other data points marked as abnormal in the neighborhood, they are removed, and only the sampling values ​​confirmed by S235 to be non-abnormal are retained to ensure that the sequence is composed entirely of reliable data. At the same time, the current brick-making process parameters that are strictly aligned with the timestamp of the abnormal data are obtained from the brick-making process parameter record, including the raw material moisture content, feeding rate, and current production batch identifier. The three types of information mentioned above—sensor type, normal sampled value sequence within the timestamp neighborhood, and current brick-making process parameters—are all organized according to a unified time reference and serve as input for subsequent weighted reconstruction processing. The entire acquisition process ensures that all data are precisely synchronized in time, and that the normal sampled value sequence accurately reflects the stable operating status of the equipment under similar operating conditions before and after the anomaly, providing a reliable context for high-fidelity correction.

[0033] S241. Determine the basic correction mode according to the sensor type. The basic correction mode includes linear interpolation mode, historical similar working condition backfill mode or physical model derivation mode, and obtain the output result based on the determined basic correction mode. It is worth noting that in this embodiment, for current sensing data that has been determined to be abnormal, the corresponding basic correction mode is first determined based on the type of sensor to which it belongs. Specifically, if the current sensing data originates from the spindle vibration signal or bearing temperature signal, a linear interpolation mode is adopted: based on the normal sampling value sequence within the timestamp neighborhood, a linear fit is performed between the two nearest normal sampling points before and after the anomaly point, and the interpolation result at that timestamp is calculated as the basic correction output; if the current sensing data originates from the drive motor current signal, a historical similar working condition backfilling mode is adopted: using the current brick-making process parameters (including raw material moisture content, feeding rate, and current production batch identifier), a working condition segment with the same production batch and a raw material moisture content deviation of no more than ±2% and a feeding rate deviation of no more than ±5% is retrieved from the historical normal operation database, and the current signal value with the same relative position as the current timestamp is extracted as the basic correction output; if the current sensing data originates from the hydraulic system pressure signal, a physical model derivation mode is adopted: based on the nonlinear mapping relationship between the feeding rate of the extrusion screw shaft and the hydraulic system pressure (this relationship has been established as a piecewise function or lookup table form through calibration experiments during the equipment debugging phase), the current feeding rate is substituted into this mapping relationship, and the corresponding theoretical pressure value is calculated as the basic correction output. Each basic correction mode relies solely on the physical characteristics and engineering experience corresponding to the sensor type, ensuring that the correction logic matches the essence of the signal. After completing the above operations, the output result of that basic correction mode is obtained, which is used for subsequent weighted fusion.

[0034] S242. Based on the current brick-making process parameters, match the reference dataset under the same or similar process conditions in the historical normal operation database, and calculate the similarity between the normal sampled value sequence in the timestamp neighborhood and the reference dataset. It is worth noting that in this embodiment, the current brick-making process parameters are first obtained, including the raw material moisture content, feeding rate, and the current production batch identifier. Then, a matching search is performed in the historical normal operation database, which stores all multi-source sensor data determined to be without anomalies within the past three consecutive months, along with their corresponding brick-making process parameters. During matching, records with the same production batch identifier are prioritized; if data from the same batch is insufficient, other batches are considered. Based on this, further screening is performed on operating condition segments where the absolute value of the difference between the raw material moisture content and the current value does not exceed 2 percentage points, and the absolute value of the difference between the feeding rate and the current value does not exceed 0.5 tons / hour. Sensor data of the same type as the current sensor data from these operating condition segments that meet these conditions are then aggregated by time to form a reference dataset. Next, for the time window where the current anomaly occurs, neighborhood segments with the same relative time position (e.g., 50 sampling points before and after) are extracted from the reference dataset to form a reference neighborhood sequence. Simultaneously, a local sequence of the same length is extracted from the normal sampling value sequence within the timestamp neighborhood obtained in step S24. Finally, the similarity between this local sequence and the reference neighborhood sequence is calculated: the Pearson correlation coefficient is used as the metric, calculated as the ratio of the product of the covariance and standard deviation of the two sequences. The resulting value is the similarity, ranging from -1 to 1. The closer the value is to 1, the more similar the dynamic behavior of the operating condition. This similarity is used for subsequent weight generation, reflecting the degree of consistency between the current local trend and historical normal patterns.

[0035] S243. Based on the sensor type, the local trend characteristics of the normal sampled value sequence and the similarity, generate the first weight, the second weight and the third weight respectively, and perform weighted fusion on the output of the basic correction mode to obtain the corrected multi-source sensing data. It is worth noting that in this embodiment, for the current sensing data that has been determined to be abnormal, after obtaining the output result of the basic correction mode, three weights are further generated to achieve weighted fusion. First, the first weight is determined according to the sensor type: if the sensor type is a spindle vibration signal or a bearing temperature signal, because it is sensitive to local continuity, the linear interpolation result is given a high degree of reliability, and the first weight is set to 0.6; if it is a drive motor current signal, because it is highly dependent on the process, the reliability of linear interpolation is low, and the first weight is set to 0.3; if it is a hydraulic system pressure signal, because it mainly depends on the physical model, the first weight is set to 0.2. Second, the local trend characteristics of the normal sampled value sequence in the timestamp neighborhood are analyzed—by calculating the first difference mean of the sequence, its overall upward, downward or stable trend is determined, and compared with the trend of the output result of the basic correction mode at the neighborhood endpoint; if the trend is consistent, the second weight is set to 0.7; if the trend is opposite, it is set to 0.2; if it is basically stable, it is set to 0.5. This weight reflects the degree of support of local temporal behavior for the correction result. Next, the similarity calculated in S242 is used as the direct basis for the third weight: the similarity value is linearly mapped to the range of 0.1 to 0.9. For example, when the similarity is 0.85, the third weight is 0.85, which is used to characterize the confidence contribution of historical operating condition matching degree to the correction result. Subsequently, the above three weights are normalized so that their sum is 1, resulting in the normalized first weight, second weight, and third weight. Finally, the output result of the basic correction mode is multiplied by each of the three weights and summed to obtain the final weighted fusion value. This value is used as the correction result of the current sensing data, constituting part of the corrected multi-source sensing data. The entire weighting process dynamically integrates three types of information: sensor prior characteristics, local temporal consistency, and process context similarity, ensuring that the correction result conforms to physical laws and fits the current operating state, which is significantly better than a single correction strategy.

[0036] S244. Replace the original abnormal sensor data with the corrected multi-source sensor data, and merge it with other sensor data without abnormalities in the original timestamp order to form the final multi-source sensor data. S3. The final multi-source sensor data is fused with the brick-making process parameters to construct a joint input vector that includes equipment status features and process context features. It is worth noting that in this embodiment, the final multi-source sensor data obtained after S2 processing is fused with the synchronously acquired brick-making process parameters through time alignment to construct a joint input vector. Specifically, firstly, key state features of each sensor signal are extracted from the final multi-source sensor data: the amplitude (such as RMS value or peak value) of the spindle vibration signal, the steady-state amplitude of the drive motor current signal, the instantaneous pressure value of the hydraulic system pressure signal, and the rate of change of the bearing temperature signal per unit time; these feature values ​​all correspond to the same sampling time to ensure time consistency. At the same time, brick-making process parameters strictly aligned with this time are acquired, including raw material moisture content, feeding rate, and current production batch identifier. Subsequently, the aforementioned features are arranged in a fixed order to form a feature vector: for example, [spindle vibration signal amplitude, drive motor current signal amplitude, hydraulic system pressure signal value, bearing temperature signal change rate, raw material moisture content, feeding rate]. The current production batch identifier is converted into a numerical vector using a preset one-hot encoding method (e.g., batch A is encoded as [1,0,0], batch B as [0,1,0], etc.), and concatenated to the end of the feature vector. The resulting joint input vector contains both equipment status features reflecting the health of the extrusion screw shaft and process context features affecting equipment operation, with fixed dimensions and a unified structure. This vector is organized in a sliding window manner according to the time series, forming continuous input samples for inference in the subsequent semi-supervised time-series anomaly detection model.

[0037] It is important to note that in this embodiment, the final multi-source sensor data used to fuse with brick-making process parameters to construct a joint input vector specifically takes the form of the amplitude of the spindle vibration signal, the amplitude of the drive motor current signal, the instantaneous value of the hydraulic system pressure signal, and the rate of change of the bearing temperature signal. These values ​​are all directly derived from the original multi-source sensor data collected in S1: the spindle vibration signal and drive motor current signal typically reflect the equipment load and vibration intensity through their amplitudes in engineering applications; the hydraulic system pressure signal is itself a scalar process quantity, and its value represents the current pressure state; the bearing temperature signal is calculated by the time difference between adjacent sampling points to obtain the rate of change of temperature per unit time, which is used to characterize the trend of thermal anomalies. The above values ​​are consistent with the original acquisition in the time dimension and undergo quality assessment, anomaly judgment, and correction processing in S2 to ensure their accuracy and reliability. The selection of this data format is based on engineering experience of the operating mechanism and fault sensitivity characteristics of the extrusion screw shaft, which can effectively characterize the health status of the equipment and provide a physically meaningful and stable input basis for the subsequent semi-supervised time-series anomaly detection model.

[0038] S4. Input the joint input vector into the pre-trained semi-supervised temporal anomaly detection model, output the health status score sequence of the extrusion spiral shaft, and determine whether there are early fault signs based on the dynamic change trend of the health status score sequence. It is worth noting that in this embodiment, a semi-supervised temporal anomaly detection model is first constructed. This model has been pre-trained before deployment, and its training process is entirely based on historical normal operation data, without any fault labels. Specifically, high-quality, anomaly-free multi-source sensor data verified by the S2 process and their corresponding brick-making process parameters are selected from the brick-making machine's operation records over the past six months to form a training dataset. This dataset covers different raw material ratios, moisture content ranges (8%–18%), feeding rates (15–30 tons / hour), and multiple production batches to ensure diverse operating conditions. The same fusion operation as in S3 is performed on each record to generate a joint input vector containing equipment state features and process context features, which is then organized chronologically into a sliding window sequence of 100 time steps as model input samples. Due to the semi-supervised paradigm, all training samples are labeled "normal," meaning the model only learns the temporal evolution patterns under normal operating conditions.

[0039] The model architecture employed is a temporal reconstruction network based on an autoencoder: the encoder consists of two stacked Long Short-Term Memory (LSTM) layers, each containing 64 hidden units, used to compress the input sequence into a low-dimensional latent representation; the decoder also consists of two LSTM layers, responsible for reconstructing the original input sequence from the latent representation. The model is optimized by minimizing the reconstruction error, with the loss function being the weighted mean square error of each feature dimension. The weights for equipment state features (such as vibration amplitude and current amplitude) are set to 1.0, and the weights for process parameters (such as moisture content and feeding rate) are set to 0.5, to emphasize the importance of the equipment's own state. During training, the Adam optimizer is used with an initial learning rate of 0.001, a batch size of 32, and 200 training epochs. Training is terminated early when the reconstruction error on the validation set fails to decrease for 10 consecutive epochs to prevent overfitting. After training, the model is frozen and no longer updated, used only for inference.

[0040] In the actual prediction phase, the joint input vector sequence generated by S3 is fed into the pre-trained model through a sliding window of the same length (100 time steps, step size 1). The model output is the reconstruction result of the input vector at each time step. Subsequently, the point-by-point reconstruction error between the original input and the reconstructed output is calculated. This error vector is then weighted and summed according to the feature dimension (weights are consistent with those during training) to obtain a single scalar value, which is the health status score for that time step—the lower the score, the closer the current state is to the normal pattern learned by the model; the higher the score, the more significant the deviation from normal behavior. The health status scores of consecutive time steps are arranged in chronological order to form a health status score sequence.

[0041] Finally, dynamic trend analysis was performed on the scoring sequence: a sliding window (window length of 50 time steps) was used to calculate the local mean and slope. If the mean of three consecutive windows exceeded a preset threshold (determined by the 99th percentile of the statistical training set), and the slope remained positive (i.e., the score showed an upward trend), then the extrusion auger shaft was judged to have early signs of failure. If the score occasionally increased but quickly fell back below the threshold, or the slope was negative / stable, it was considered normal fluctuation. The entire judgment process did not rely on fault samples, but only identified potential degradation through normal data modeling and trend evolution, achieving a highly sensitive early warning under label-less conditions.

[0042] Example 2: Figure 2 As shown, the present invention also discloses an artificial intelligence-based fault prediction system, comprising the following modules: Data acquisition module: used to collect multi-source sensor data at the station where the extrusion screw shaft is located during the operation of the brick making machine, and simultaneously acquire brick making process parameters aligned with the timestamp of the multi-source sensor data; Data correction module: Connected to the data acquisition module, it is used to assess the data quality of multi-source sensor data, filter out data subsets, and judge each sensor data in the data subset. Based on the judgment results, it obtains sensor data with anomalies and sensor data without anomalies. It performs correction processing on the sensor data with anomalies to generate corrected multi-source sensor data, and integrates the corrected multi-source sensor data with the sensor data without anomalies to obtain the final multi-source sensor data. Data fusion module: Connected to the data correction module and the data acquisition module, it is used to fuse the final multi-source sensor data with the brick-making process parameters to construct a joint input vector containing equipment status features and process context features; Fault identification module: Connected to the data fusion module, it is used to input the joint input vector into the pre-trained semi-supervised time-series anomaly detection model, output the health status score sequence of the extrusion screw shaft, and determine whether there are early fault signs based on the dynamic change trend of the health status score sequence.

[0043] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and are not intended to limit them. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some or all of the technical features therein. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the technical solutions of the embodiments of the present invention.

Claims

1. A method for failure prediction based on artificial intelligence, characterized in that, Includes the following steps: S1. Collect multi-source sensor data of the extrusion screw shaft station during the operation of the brick making machine, and simultaneously acquire brick making process parameters aligned with the timestamp of the multi-source sensor data. S2. Perform data quality assessment on the multi-source sensor data, filter to obtain a data subset, and judge each sensor data in the data subset. Based on the judgment results, obtain the sensor data with anomalies and the sensor data without anomalies. Perform correction processing on the sensor data with anomalies to generate corrected multi-source sensor data. Integrate the corrected multi-source sensor data with the sensor data without anomalies to obtain the final multi-source sensor data. S3. The final multi-source sensor data is fused with the brick-making process parameters to construct a joint input vector that includes equipment status features and process context features. S4. Input the joint input vector into the pre-trained semi-supervised temporal anomaly detection model, output the health status score sequence of the extrusion spiral shaft, and determine whether there are early fault signs based on the dynamic change trend of the health status score sequence. 2.The artificial intelligence-based failure prediction method of claim 1, wherein, The multi-source sensor data includes at least: spindle vibration signal, drive motor current signal, hydraulic system pressure signal, and bearing temperature signal; the brick-making process parameters include at least: raw material moisture content, feeding rate, and current production batch identifier.

3. The fault prediction method based on artificial intelligence according to claim 2, characterized in that, The process involves evaluating the quality of multi-source sensor data, filtering to obtain a data subset, and then judging each sensor data within the subset. Based on the judgment results, sensor data with anomalies and sensor data without anomalies are identified. Correction processing is performed on the sensor data with anomalies to generate corrected multi-source sensor data. Finally, the corrected multi-source sensor data is integrated with the sensor data without anomalies to obtain the final multi-source sensor data, including: S21. Divide the multi-source sensor data into segments according to a preset time window, and filter the data subsets based on the segmentation results; S22. Obtain the statistical distribution characteristics of each sensor data in the data subset, and extract the physical constraint relationships that each sensor data in the data subset should satisfy during the operation of the extrusion screw shaft. S23. Based on the statistical distribution characteristics and physical constraints, determine any sensor data in the data subset, and obtain the sensor data with anomalies and the sensor data without anomalies based on the determination results. S24. Obtain the sensor type, normal sampling value sequence in the timestamp neighborhood, and current brick-making process parameters of the sensor data with abnormality. Based on the sensor type, normal sampling value sequence, and current brick-making process parameters, perform weighted reconstruction on the current sensor data to generate corrected multi-source sensor data. Then, integrate the corrected multi-source sensor data with the sensor data without abnormality to obtain the final multi-source sensor data.

4. The fault prediction method based on artificial intelligence according to claim 3, characterized in that, The process of segmenting multi-source sensor data according to a preset time window and filtering the data subsets based on the segmentation results includes: S211. Divide the multi-source sensor data into continuous segments according to a fixed-length time window to obtain the sensor data within each time window. S212. Calculate the missing rate for each sensor data within each time window; S213. Calculate the mutation rate for each sensor data within each time window; S214. Set a first threshold for the deletion rate and a second threshold for the mutation rate. S215. Determine whether each time window meets the retention conditions. The retention conditions are: the missing rate of all sensor data within the time window is less than the first threshold, and the mutation rate of all sensor data is less than the second threshold; retain each sensor data within the time window that meets the retention conditions as a data subset.

5. The fault prediction method based on artificial intelligence according to claim 4, characterized in that, The statistical distribution characteristics of each sensor data in the data subset include: the mean, standard deviation, and local outlier factor of each sensor data in the data subset; the physical constraints that each sensor data in the data subset should satisfy during the operation of the extrusion screw shaft include: the positive correlation constraint between the amplitude of the main shaft vibration signal of the extrusion screw shaft and the amplitude of the drive motor current signal of the extrusion screw shaft, the nonlinear mapping constraint between the hydraulic system pressure signal of the extrusion screw shaft and the feeding rate of the extrusion screw shaft, and the thermal balance constraint between the rate of change of the bearing temperature signal of the extrusion screw shaft and the rotational speed of the extrusion screw shaft.

6. The fault prediction method based on artificial intelligence according to claim 5, characterized in that, The process of judging any sensor data in the data subset based on statistical distribution characteristics and physical constraints, and obtaining sensor data with anomalies and sensor data without anomalies based on the judgment results, includes: S231. For the current sensing data in the data subset, obtain the mean and standard deviation of the statistical distribution characteristics of the sensing data of the same type as the current sensing data in the data subset, calculate the absolute value of the deviation between the current sensing data and the mean, and determine whether the absolute value of the deviation is greater than a preset multiple of the standard deviation; if it is greater, mark the current sensing data as a first type of statistical anomaly candidate. S232. For the current sensing data in the data subset, obtain the local outlier factor in the statistical distribution characteristics of the sensing data of the same type as the current sensing data in the data subset, and determine whether the local outlier factor is greater than a third preset threshold; if it is greater, mark the current sensing data as a second type of statistical anomaly candidate. S233. If the current sensing data is marked as a first-class statistical anomaly candidate or a second-class statistical anomaly candidate, then mark the current sensing data as a statistical anomaly candidate. S234. For the current sensing data in the data subset, perform the following physical constraint verification operation: When the current sensing data is the amplitude of the main shaft vibration signal of the extrusion screw shaft, the amplitude of the drive motor current signal of the extrusion screw shaft at the same timestamp is obtained, and it is determined whether the amplitude of the main shaft vibration signal of the extrusion screw shaft and the amplitude of the drive motor current signal of the extrusion screw shaft satisfy the positive correlation constraint. When the current sensing data is the amplitude of the current signal of the drive motor of the extrusion screw shaft, the amplitude of the main shaft vibration signal of the extrusion screw shaft at the same timestamp is obtained, and it is determined whether the amplitude of the current signal of the drive motor of the extrusion screw shaft and the amplitude of the main shaft vibration signal of the extrusion screw shaft satisfy the positive correlation constraint. When the current sensing data is the hydraulic system pressure signal of the extrusion screw shaft, the feeding rate of the extrusion screw shaft at the same time stamp is obtained, and it is determined whether the hydraulic system pressure signal of the extrusion screw shaft and the feeding rate satisfy the nonlinear mapping constraint. When the current sensing data is the rate of change of the bearing temperature signal of the extrusion screw shaft, the rotational speed of the extrusion screw shaft at the same time stamp is obtained, and it is determined whether the rate of change of the bearing temperature signal of the extrusion screw shaft and the rotational speed of the extrusion screw shaft satisfy the thermal balance constraint. If the current sensing data belongs to any of the above types and does not meet the corresponding physical constraints, then the current sensing data is marked as a physical anomaly candidate. S235. The current sensing data is determined to be abnormal only if condition one is met, wherein condition one is: if the current sensing data is simultaneously marked as a statistical anomaly candidate and a physical anomaly candidate; if condition one is not met, the current sensing data is determined to be abnormal.

7. The fault prediction method based on artificial intelligence according to claim 6, characterized in that, The process involves acquiring the sensor type of the abnormal sensor data, the normal sample value sequence within the timestamp neighborhood, and the current brick-making process parameters. Based on these parameters, the current sensor data is weighted and reconstructed to generate corrected multi-source sensor data. This corrected multi-source sensor data is then integrated with the abnormal sensor data to obtain the final multi-source sensor data, including: S241. Determine the basic correction mode according to the sensor type. The basic correction mode includes linear interpolation mode, historical similar working condition backfill mode or physical model derivation mode, and obtain the output result based on the determined basic correction mode. S242. Based on the current brick-making process parameters, match the reference dataset under the same or similar process conditions in the historical normal operation database, and calculate the similarity between the normal sampled value sequence in the timestamp neighborhood and the reference dataset. S243. Based on the sensor type, the local trend characteristics of the normal sampled value sequence and the similarity, generate the first weight, the second weight and the third weight respectively, and perform weighted fusion on the output of the basic correction mode to obtain the corrected multi-source sensing data. S244. Replace the original abnormal sensor data with the corrected multi-source sensor data, and merge it with other sensor data without abnormalities in the original timestamp order to form the final multi-source sensor data.

8. An artificial intelligence-based fault prediction system, used to execute the artificial intelligence-based fault prediction method according to any one of claims 1-7, characterized in that, Includes the following modules: Data acquisition module: used to collect multi-source sensor data at the station where the extrusion screw shaft is located during the operation of the brick making machine, and simultaneously acquire brick making process parameters aligned with the timestamp of the multi-source sensor data; Data correction module: Connected to the data acquisition module, it is used to assess the data quality of multi-source sensor data, filter out data subsets, and judge each sensor data in the data subset. Based on the judgment results, it obtains sensor data with anomalies and sensor data without anomalies. It performs correction processing on the sensor data with anomalies to generate corrected multi-source sensor data, and integrates the corrected multi-source sensor data with the sensor data without anomalies to obtain the final multi-source sensor data. Data fusion module: Connected to the data correction module and the data acquisition module, it is used to fuse the final multi-source sensor data with the brick-making process parameters to construct a joint input vector containing equipment status features and process context features; Fault identification module: Connected to the data fusion module, it is used to input the joint input vector into the pre-trained semi-supervised time-series anomaly detection model, output the health status score sequence of the extrusion screw shaft, and determine whether there are early fault signs based on the dynamic change trend of the health status score sequence.

Images