Artificial intelligence based predictive maintenance system

Abstract

The application relates to the technical field of operation and maintenance monitoring, and discloses a predictive maintenance system based on artificial intelligence. First, equipment load data and buffer placeholder data are collected, vacancy density and vacancy flow are calculated, and a non-equilibrium vacancy load circulation track is obtained. Then, the quantity is used to construct a vacancy convection time delay map and a time rescaling function, and the equipment load data is subjected to time sequence alignment processing. Further, a direction gating matrix is generated based on the non-equilibrium vacancy load circulation track, and the direction gating matrix is embedded into a neural network containing a directional residual branch together with the vacancy convection time delay map. The neural network is jointly trained by time reversal irreversible representation loss and main task loss, so that the neural network can accurately output equipment time risk. In the online inference stage, the above processing procedure is repeatedly performed in real time, real-time risks of all equipment are accumulated within a preset decision window, and finally, equipment with the maximum total risk is output as a fault equipment.

Description

Technical Field

[0001] This invention relates to the field of operation and maintenance monitoring technology, and more specifically, to an artificial intelligence-based predictive maintenance system. Background Technology

[0002] As discrete manufacturing evolves towards flexibility and intelligence, typical production lines generally adopt a multi-station serial structure, with materials connected between stations via conveyor belts or automated storage systems. To alleviate cycle time imbalances and improve capacity utilization, intermediate buffer zones are commonly set up in actual production. However, limited buffer capacity leads to a dynamic phenomenon of alternating material accumulation and equipment standby when downstream cycle time fluctuates. This phenomenon not only alters the equipment's own operating load curve but also introduces short-term energy feedback and production cycle time fluctuations. For example, when downstream equipment stops briefly, accumulated material forms a pressure chain on the conveyor belt, forcing an instantaneous increase in the load of the upstream equipment's motor; when the downstream restarts, the release of the accumulated material causes a sudden drop in load. These dynamic effects typically occur within short periods of milliseconds to seconds and are directional and transiently amplified. Traditional predictive maintenance algorithms extract signal features such as vibration, power, or temperature from a single machine perspective, failing to capture this cross-equipment propagation process, resulting in decreased robustness of the model in complex production environments.

[0003] In real-world production systems, the congestion and starvation caused by limited buffers can create an unsteady transmission chain of operating states between devices through production logistics. If the equipment operation process is linearized, its input-output relationship can be approximated as a dynamic matrix. When the buffer approaches saturation, the offset of this matrix increases dramatically, causing the system's short-term response to exhibit an abnormal energy amplification effect. This means that even if the equipment itself is not faulty, its load signal may still be amplified several times in a short period due to the pressure transmission from upstream and downstream, and then transmitted back upstream. This amplification effect not only changes the temporal structure of the load signal but also makes it difficult for prediction models based on ordinary neural networks to distinguish the source of the anomaly. As a result, the model may mistakenly identify passively responding equipment as actively faulty equipment, leading to incorrect maintenance instructions. Therefore, only by fully considering this cross-device impact caused by buffer dynamics during the modeling process can we avoid fault location errors in predictive maintenance systems in real-world production line scenarios, and achieve accurate prediction of equipment health status and early triggering of maintenance plans. Summary of the Invention

[0004] This invention provides an artificial intelligence-based predictive maintenance system that solves the technical problem of accurately identifying key equipment that causes abnormal propagation under the interactive influence of signals from multiple devices in a production line with limited buffers and dynamic congestion, thereby achieving reliable predictive maintenance decisions.

[0005] This invention provides an artificial intelligence-based predictive maintenance system, comprising: The data acquisition module collects equipment load data and buffer occupancy data, calculates vacancy density and vacancy flow based on the buffer occupancy data, and then calculates the unbalanced vacancy load circulation trace based on the equipment load data and vacancy flow. The data processing module obtains the convection delay map of the unbalanced empty space load circulation trace, constructs a time recalibration function, and uses the time recalibration function to process the equipment load data to obtain an alignment sequence. The feature extraction module generates a directional gating matrix based on the unbalanced vacancy load circulation trace, and applies the directional gating matrix and the vacancy convection delay map to a neural network containing directional residual branches. The training module uses time-reversal irreversible representation loss and main task loss to jointly train the neural network, and outputs the device time risk through the trained neural network. During online inference, the risk calculation module collects real-time device load data and real-time buffer placeholder data, and repeatedly executes the data acquisition module to the risk calculation module to obtain the real-time device time risk. The real-time device time risk of each device is accumulated within a preset decision window, and the device with the highest total risk is output as the faulty device.

[0006] The beneficial effects of this invention include: by introducing unbalanced vacancy load circulation traces in multi-equipment collaborative production line scenarios, real-time identification and correction of load propagation effects caused by finite buffers between equipment are achieved, enabling the neural network to distinguish between external cycle disturbances and equipment degradation signals during predictive maintenance. This invention establishes vacancy density, vacancy flow, vacancy convection delay maps, and a time recalibration function; constructs a direction-gated matrix and embeds a neural network with directional residual branches; and trains it using time-reversal irreversible representation loss, achieving physical consistency alignment and information direction constraints for multi-equipment time-series data. This invention not only improves the accuracy and prediction lead of equipment fault location but also significantly reduces the false alarm rate caused by production cycle fluctuations, enhancing the stability and generalization ability of predictive maintenance models in complex collaborative production environments. Attached Figure Description

[0007] Figure 1 This is a block diagram of the predictive maintenance system based on artificial intelligence of the present invention. Detailed Implementation

[0008] The subject matter described herein will now be discussed with reference to exemplary embodiments. It should be understood that these embodiments are discussed only to enable those skilled in the art to better understand and implement the subject matter described herein, and changes may be made to the function and arrangement of the elements discussed without departing from the scope of this specification. Various processes or components may be omitted, substituted, or added as needed in the examples. Furthermore, features described in some examples may be combined in other examples.

[0009] like Figure 1 As shown, the predictive maintenance system based on artificial intelligence includes: The data acquisition module collects equipment load data and buffer occupancy data, calculates vacancy density and vacancy flow based on the buffer occupancy data, and then calculates the unbalanced vacancy load circulation trace based on the equipment load data and vacancy flow. The data processing module obtains the convection delay map of the unbalanced empty space load circulation trace, constructs a time recalibration function, and uses the time recalibration function to process the equipment load data to obtain an alignment sequence. The feature extraction module generates a directional gating matrix based on the unbalanced vacancy load circulation trace, and applies the directional gating matrix and the vacancy convection delay map to a neural network containing directional residual branches. The training module uses time-reversal irreversible representation loss and main task loss to jointly train the neural network, and outputs the device time risk through the trained neural network. During online inference, the risk calculation module collects real-time device load data and real-time buffer placeholder data, and repeatedly executes the data acquisition module to the risk calculation module to obtain the real-time device time risk. The real-time device time risk of each device is accumulated within a preset decision window, and the device with the highest total risk is output as the faulty device.

[0010] In one embodiment of the present invention, the process of collecting device load data and buffer occupancy data, and calculating vacancy density and vacancy flow based on the buffer occupancy data includes: Collect device load data and buffer placeholder data; For each buffer segment, the empty space density of the buffer segment at the corresponding sampling time is obtained by using the buffer capacity of the buffer segment and the buffer occupancy data of the buffer segment at the corresponding sampling time. For each buffer segment, the space flow of that buffer segment at the current sampling time is obtained by using the space density of that buffer segment at the current sampling time and the space density of that buffer segment at the previous sampling time.

[0011] Equipment load data is signal data that reflects the load intensity of equipment during operation. Specifically, equipment load data includes at least one of the following: equipment servo motor current, motor output torque, equipment operating power, and vibration energy of key components. The sources of equipment load data are the equipment's built-in sensors or external industrial sensors (such as current sensors for current acquisition and torque sensors for torque acquisition).

[0012] Buffer occupancy data is the current status data of work-in-process in the intermediate buffer section of the production line. Specifically, buffer occupancy data is the quantity of work-in-process in the buffer section (e.g., 5 pieces) or the occupancy rate of the buffer section (e.g., 60%). The sources of buffer occupancy data are sensors installed at the entrance / exit of the buffer section, including photoelectric sensors and infrared counters used to count the quantity of work-in-process, and pressure sensors used to calculate the occupancy rate of the buffer section.

[0013] A buffer section is an intermediate storage area in a production line used to connect materials between two adjacent pieces of equipment. Each buffer section has a unique identifier (e.g., buffer section 1 is located between equipment 1 and equipment 2). The purpose of the buffer section is to alleviate the blockage / starvation caused by the uneven cycle time of upstream and downstream equipment. The purpose of the buffer section is to serve as a calculation unit for vacancy density and vacancy flow. Each buffer section needs to calculate its corresponding vacancy density and vacancy flow separately.

[0014] The buffer capacity of a buffer section is a parameter that a single buffer section can hold at the time of design. Specifically, it is the maximum number of work-in-process that can be held (e.g., 20 pieces) or the occupancy rate of the maximum number of work-in-process that can be held (e.g., 100%). The buffer capacity of a buffer section is a fixed parameter. The buffer capacity of a buffer section can be obtained by consulting the production line design drawings, equipment technical manuals, or by obtaining the specifications provided by the buffer section manufacturer.

[0015] The buffer occupancy data of the buffer segment at the corresponding sampling time is the buffer occupancy data collected completely synchronously with the target vacancy density calculation time and belongs to the buffer segment (e.g., to calculate the vacancy density of a buffer segment at 10:00:05 AM, the buffer occupancy data collected at 10:00:05 AM of this buffer segment is required); the data unit of the buffer occupancy data of the buffer segment at the corresponding sampling time must be consistent with the unit of the buffer capacity of the buffer segment (if the buffer capacity of the buffer segment is the maximum number of work-in-process that can be accommodated, the buffer occupancy data of the buffer segment at the corresponding sampling time is the number of work-in-process; if the buffer capacity of the buffer segment is the maximum occupancy rate, the buffer occupancy data of the buffer segment at the corresponding sampling time is the occupancy rate); the purpose of the buffer occupancy data of the buffer segment at the corresponding sampling time is to provide real-time status input data for calculating the vacancy density at the corresponding time.

[0016] The vacancy density of a buffer segment at the corresponding sampling time is the remaining available capacity of a single buffer segment at the target time. The vacancy density of a buffer segment at the corresponding sampling time reflects the idle state of the buffer segment. The data unit of the vacancy density of a buffer segment at the corresponding sampling time is consistent with the unit of the buffer capacity of the buffer segment. Specifically, the vacancy density of a buffer segment at the corresponding sampling time is equal to the buffer capacity of the buffer segment minus the buffer occupancy data of the buffer segment at the corresponding sampling time.

[0017] The vacancy density of a buffer segment at the current sampling time is the vacancy density belonging to that buffer segment at the target time when the vacancy flow needs to be calculated (e.g., when calculating the vacancy flow of a buffer segment at 10:00:10 AM, if the current sampling time is 10:00:10 AM, the vacancy density of the buffer segment at the current sampling time refers to the vacancy density of that buffer segment at 10:00:10 AM).

[0018] The empty space density of the buffer segment at the previous sampling time is the empty space density of the buffer segment at the previous data acquisition time at the current sampling time (e.g., if the current sampling time of a buffer segment is 10:00:10 AM, the sampling interval needs to be set according to 1 / 5 to 1 / 2 of the production line cycle. If the production line cycle is 10 seconds, the sampling interval is set to 2-5 seconds, and the previous sampling time is 10:00:05 AM. The empty space density of the buffer segment at the previous sampling time refers to the empty space density of the buffer segment at 10:00:05 AM).

[0019] The vacancy flow of a buffer segment at the current sampling time is the change in vacancy density of a single buffer segment at the current sampling time relative to the previous sampling time. When the vacancy flow of a buffer segment at the current sampling time is positive, it indicates that the vacancy of the buffer segment has increased (the vacancy level of the buffer has increased), and when it is negative, it indicates that the vacancy of the buffer segment has decreased (the occupancy level of the buffer has increased). The purpose of the vacancy flow of the buffer segment at the current sampling time is to provide key dynamic input data for subsequent calculation of the unbalanced vacancy load circulation trace. Specifically, the vacancy flow of a buffer segment at the current sampling time is equal to the vacancy density of the buffer segment at the current sampling time minus the vacancy density of the buffer segment at the previous sampling time.

[0020] In one embodiment of the present invention, calculating the unbalanced vacancy load circulation trace based on equipment load data and vacancy flow includes: Determine the time window, the number of positive time delay steps, the maximum number of time delay steps, and the hysteresis decay constant; For each device and each buffer segment, the cross-correlation between device load data and space flow at two moments under the positive time delay is calculated by combining the positive time delay step and time window. For each device and each buffer segment, the cross-correlation between device load data and space flow at two moments under the reverse time delay is calculated by combining the reverse time delay steps and time window. Subtracting the cross-correlation between the two time points corresponding to the positive time delay from the cross-correlation between the two time points corresponding to the negative time delay yields the antisymmetric part of the cross-correlation. Determine the downstream distance weight based on the location relationship between the equipment and the buffer section, and only consider the downstream equipment; determine the hysteresis weight function that changes with the number of time delay steps. The numerator is obtained by summing the products of the hysteresis weight function, the downstream distance weight, and the antisymmetric part of the cross-correlation under all positive time delay steps. The device load data energy normalization factor is calculated based on the load data of all devices within the time window. The empty flow energy normalization factor is calculated based on the empty flow of the corresponding buffer segment within the time window. The device load data energy normalization factor and the empty flow energy normalization factor are multiplied together to obtain the denominator. Dividing the numerator by the denominator yields the unbalanced vacancy load circulation trace for each buffer segment within the corresponding time window.

[0021] A time window is a fixed time interval used to capture continuous time-series data. The length of the time window should be set to 1 to 3 times the production line fluctuation cycle (e.g., if the production line cycle fluctuation cycle is 10 seconds, the time window length should be set to 10 to 30 seconds) to ensure that it can completely capture one fluctuation process.

[0022] The forward time delay step is the number of steps that the idle flow moment lags behind the device load data moment when calculating the forward cross-correlation. The value of the forward time delay step must match the sampling interval (e.g., if the sampling interval is 2 seconds, the forward time delay step is set to 1 to 5 steps, corresponding to an actual time delay of 2 to 10 seconds), covering the typical duration of the idle flow's impact on the load data.

[0023] The maximum time delay step is the upper limit of the forward time delay step. The maximum time delay step must be consistent with the range of the forward time delay step (e.g., if the forward time delay step is 1 to 5 steps, the maximum time delay step is set to 5 steps). This is used to limit the range of subsequent summation calculations and avoid interference from invalid long-delay data.

[0024] The hysteresis decay constant is a parameter that controls the decay rate of the hysteresis weight function. The value of the hysteresis decay constant must be greater than the maximum number of time delay steps (e.g., if the maximum number of time delay steps is 5, the hysteresis decay constant should be set to 5 to 10 steps) to ensure that the weight of long delay steps is small enough to highlight the impact of short delays.

[0025] The cross-correlation between equipment load data and idle flow at two moments under positive time delay is an indicator that measures the degree of linear correlation between the value of equipment load data at a certain moment and the value of the corresponding buffer segment's idle flow at that moment minus the positive time delay step value within a time window. The purpose of this indicator is to capture the lagging effect of idle flow on downstream equipment load data. Specifically, for each device and each buffer segment, the load data of the device is first taken at each moment within the time window, and then multiplied by the value of the buffer segment's idle flow at the current moment minus the positive time delay step value. The products of all moments are added together, and then the sum is divided by the number of moments included in the time window to obtain the cross-correlation between the device and the buffer segment at that positive time delay step.

[0026] The reverse time delay step is the number of steps that the empty flow time is ahead of the device load data time when calculating the reverse cross-correlation. The value of the reverse time delay step must be exactly the same as the forward time delay step (e.g., if the forward time delay step is 1 step, the reverse time delay step is also 1 step) to ensure the fairness of the forward and reverse comparison.

[0027] The cross-correlation between equipment load data and idle flow at two moments under the reverse time delay is an indicator that measures the degree of linear correlation between the value of equipment load data at a certain moment and the value of the corresponding buffer segment idle flow at that moment plus the number of reverse time delay steps within a time window. The purpose of this indicator is to extract the time inversion asymmetry features of non-equilibrium systems by comparing it with the positive cross-correlation. Specifically, for each device and each buffer segment, the load data of the device is taken at each moment within the time window, and then multiplied by the value of the idle flow of the buffer segment at the current moment plus the number of reverse time delay steps. The products of all moments are added together, and then the sum is divided by the number of moments included in the time window to obtain the cross-correlation between the device and the buffer segment at the number of reverse time delay steps.

[0028] The antisymmetric part of cross-correlation is the difference between the positive cross-correlation and the negative cross-correlation. When this value is not zero, it reflects the existence of time reversal asymmetry in the system, i.e., non-equilibrium characteristics. The purpose of the antisymmetric part of cross-correlation is to quantify the directional correlation strength between the vacancy flow and the equipment load data. Specifically, the antisymmetric part of cross-correlation is equal to the cross-correlation between the two moments corresponding to the positive time delay and the cross-correlation between the two moments corresponding to the negative time delay, under the same equipment, the same buffer segment, and the same delay step.

[0029] The downstream distance weight is calculated based on the physical location number of the device and the buffer section. It has a non-zero value only when the device is downstream of the buffer section (device number is greater than buffer section number). The magnitude of the downstream distance weight is positively correlated with the number of devices between the device and the buffer section, reflecting the cumulative effect of downstream propagation. The purpose of the downstream distance weight is to highlight the contribution of downstream devices to the impact of the buffer section's empty space flow. Specifically, for a buffer section and a device, if the device number is greater than the buffer section number (device is downstream), the downstream distance weight is equal to the device number minus the buffer section number; if the device number is less than or equal to the buffer section number (device is upstream or in the same location), the downstream distance weight is equal to 0.

[0030] The lag weighting function is a function that monotonically decreases as the number of positive time delay steps increases. It is used to reduce the contribution of the antisymmetric part of cross-correlation corresponding to long delay steps, which is consistent with the physical law that the influence of vacancy flow on load data weakens over time. The purpose of the lag weighting function is to adjust the contribution of different delay steps to the numerator. Specifically, the lag weighting function is equal to the negative power of the natural constant, which is the number of positive time delay steps divided by the lag decay constant.

[0031] The numerator is the sum of the products of the lag weight function, the downstream distance weight, and the antisymmetric part of the cross-correlation for all positive time delay steps. The magnitude of this value reflects the total intensity of the non-equilibrium characteristics of directional propagation. The purpose of the numerator is to be used as the numerator in the calculation of the non-equilibrium vacancy load circulation trace. Specifically, for each positive time delay step, the product of the lag weight function, the downstream distance weight, and the antisymmetric part of the cross-correlation is calculated for that step. Then, the products corresponding to all positive time delay steps are added together to obtain the numerator.

[0032] The equipment load data energy normalization factor is the root mean square value of all equipment load data within a time window. It is used to eliminate the influence of differences in the magnitude of load data from different equipment on the results. The purpose of the equipment load data energy normalization factor is to serve as a component of the denominator in the calculation of the unbalanced vacant load circulation trace. Specifically, the square of the load data of each device at each moment within the time window is calculated first. The squares of all the moment values ​​of all devices are added together. The sum is divided by the number of moments included in the time window. The square root of the average value is then taken to obtain the equipment load data energy normalization factor.

[0033] The vacancy flow energy normalization factor is the root mean square value of the vacancy flow in the corresponding buffer section within the time window. It is used to eliminate the influence of differences in vacancy flow rate levels in different buffer sections on the results. The purpose of the vacancy flow energy normalization factor is to serve as a component of the denominator in the calculation of the unbalanced vacancy load circulation trace. Specifically, the square of the vacancy flow value in the corresponding buffer section within the time window is calculated first. The square values ​​of all the time moments are added together, and the sum is divided by the number of time moments included in the time window. Then, the square root of the average value is taken to obtain the vacancy flow energy normalization factor.

[0034] The denominator is the product of the energy normalization factor of the equipment load data and the energy normalization factor of the vacancy flow. It is used to normalize the energy of the numerator to give the non-equilibrium vacancy load circulation track a uniform dimension. The purpose of the denominator is to serve as the denominator for the calculation of the non-equilibrium vacancy load circulation track. Specifically, the denominator is equal to the energy normalization factor of the equipment load data multiplied by the energy normalization factor of the vacancy flow.

[0035] The unbalanced vacancy load circulation trace is the ratio of the numerator to the denominator. The sign of this value reflects the direction of the vacancy flow's influence on the downstream equipment load, and the absolute value reflects the intensity of the unbalanced directional propagation. The purpose of the unbalanced vacancy load circulation trace is to serve as the core basis for subsequent time recalibration and the construction of the direction gating matrix. Specifically, the unbalanced vacancy load circulation trace is equal to the numerator divided by the denominator.

[0036] In one embodiment of the present invention, obtaining a vacancy convection delay map based on the unbalanced vacancy load circulation trace includes: For each device and each buffer segment, extract the positive values ​​from the antisymmetric part of the cross-correlation corresponding to each positive time delay step; The numerator is obtained by summing the products of all positive time delay steps and the corresponding positive antisymmetric values ​​of the cross-correlation. Calculate the sum of the antisymmetric positive values ​​of the cross-correlation under all positive time delay steps to obtain the denominator; Divide the numerator by the denominator to obtain the empty convection delay map element corresponding to each device and each buffer segment. The empty convection delay map is constructed based on all empty convection delay map elements.

[0037] The positive values ​​in the antisymmetric part of the cross-correlation corresponding to each positive time delay step are the values ​​greater than zero selected from the antisymmetric part of the cross-correlation. Since negative values ​​reflect the non-directional interference between the idle flow and the equipment load data, only positive values ​​can effectively characterize the positive directional correlation between the idle flow and the downstream equipment load, thus avoiding invalid data from affecting the time delay calculation.

[0038] The sum of the products (numerator) of all positive time delay steps and the corresponding positive values ​​of the antisymmetric part of the cross-correlation is the result obtained by multiplying all positive time delay steps with the positive values ​​in the antisymmetric part of the cross-correlation for each device and each buffer segment, and then summing all the products. The purpose of this numerator is to calculate the numerator of the vacancy convection delay map element, and its value is positively correlated with the cumulative effect of delay steps × correlation strength. Specifically, the numerator is equal to multiplying all positive time delay steps with the positive values ​​in the antisymmetric part of the cross-correlation for each device and each buffer segment, and then summing all the multiplication results.

[0039] The sum of the positive values ​​of the antisymmetric part of the cross-correlation for all positive time delay steps (denominator) is obtained by directly summing the positive values ​​of the antisymmetric part of the cross-correlation for each device and each buffer segment. The purpose of this denominator is to calculate the denominator term of the empty convection delay map element, and its value reflects the total strength of the effective directional correlation. Specifically, the denominator is equal to the sum of the positive values ​​of the antisymmetric part of the cross-correlation for each device and each buffer segment.

[0040] The empty space convection delay map element corresponding to each device and each buffer segment is the ratio obtained by dividing the numerator by the denominator for that device and that buffer segment. Its unit is consistent with the unit of the forward time delay steps (e.g., steps, which can be converted to actual time by combining the sampling interval; for example, 1 step corresponds to a 2-second sampling interval, so the delay is 2 seconds × element value). It reflects the average time delay of the empty space flow propagating from the buffer segment to the device. The delay is calculated by weighted average of the forward time delay steps (with the positive values ​​of the antisymmetric part of the cross-correlation as the weights). Since the larger the positive value, the stronger the directional correlation under that delay step, the weighted average can accurately capture the dominant propagation delay, thus adapting to the design of non-equilibrium empty space flow propagation law. The purpose of this element is to form the basic unit of the empty space convection delay map. Specifically, the empty space convection delay map element corresponding to each device and each buffer segment is equal to the numerator corresponding to the device and that buffer segment divided by the denominator corresponding to the device and that buffer segment.

[0041] The empty space convection delay map is a two-dimensional matrix formed by arranging the empty space convection delay map elements corresponding to each device and each buffer segment according to the device-buffer segment dimension (the rows of the matrix represent devices, the columns represent buffer segments, and the value of each position in the matrix is ​​the convection delay element corresponding to the device and the buffer segment).

[0042] In one embodiment of the present invention, a time recalibration function is constructed, and the device load data is processed using the time recalibration function to obtain an aligned sequence, including: Set the convection intensity scaling constant; For each buffer segment, the convection intensity weight is calculated based on the unbalanced vacancy load circulation trace and the convection intensity scaling constant corresponding to that buffer segment. For each device, a time window containing the current sampling time is determined. The convection intensity weights and empty space convection delay map elements corresponding to all buffer segments before the device in the time window are extracted. The sum of the products of each convection intensity weight and the corresponding empty space convection delay map element is calculated. The sum of the products is subtracted from the current sampling time to obtain the time recalibration function value of the device at the current sampling time. Extract the decimal part of the time recalibration function value to determine the floor and floor times corresponding to that value. Obtain the load data of the device at the floor and floor times. Subtract the decimal part from 1 to obtain the first weight. Use the decimal part as the second weight. Calculate the first product of the first weight and the load data at the floor time, and the second product of the second weight and the load data at the floor time. Add the first and second products to obtain the alignment sequence data of the device at the current sampling time. The alignment sequence data of all sampling times together constitute the alignment sequence of the device.

[0043] The convection intensity scaling constant is a hyperparameter that adjusts the range of convection intensity weights. The value of the convection intensity scaling constant should be set according to the typical amplitude of the unbalanced vacant load circulation trace (e.g., the typical value of the unbalanced vacant load circulation trace is 0.1-1, and the convection intensity scaling constant is set to 0.5-2) to avoid deviations in time shift calculation caused by excessively large or small weights.

[0044] The convection intensity weight is a weight that quantifies the impact of the directional propagation intensity of the vacant flow in the buffer zone on the equipment time-shift. This weight is non-negative (due to the use of the softplus function for mapping). The larger the unbalanced vacant load circulation trace, the larger the convection intensity weight, reflecting the stronger the contribution of the buffer zone to the equipment time-shift. The purpose of the convection intensity weight is to provide a weight for the contribution of the buffer zone in the subsequent calculation of the sum of products. Specifically, the convection intensity weight is equal to the unbalanced vacant load circulation trace corresponding to the buffer zone multiplied by the convection intensity scaling constant, and then input into the softplus function (the softplus function is the logarithm of the product raised to the power of the natural constant plus one).

[0045] The convection intensity weights corresponding to all buffer sections before this device refer to the convection intensity weights of buffer sections with device numbers greater than the buffer section number (i.e., buffer sections downstream of the physical location of the device) within this time window. Only the weights of such buffer sections are extracted because the empty flow in the upstream buffer section will not have a time shift effect on the downstream device, thus avoiding interference from invalid weights.

[0046] The empty convection delay map elements corresponding to all buffer sections before this device refer to the empty convection delay map elements of the buffer section (i.e., the buffer section downstream of the physical location of the device) under this time window, which correspond one-to-one with the convection intensity weights corresponding to all buffer sections before this device.

[0047] The sum of the products of each convection intensity weight and its corresponding empty convection delay map element is the result obtained by multiplying the convection intensity weights of all buffer segments preceding the device with their corresponding empty convection delay map elements one by one, and then summing all the products. The unit of this result is time (e.g., seconds), reflecting the total time shift contribution of all upstream buffer segments to the device. The purpose of this sum of products is to calculate the core intermediate quantity of the time recalibration function value. Specifically, the sum of products is equal to multiplying the convection intensity weight of each buffer segment preceding the device by the empty convection delay map element corresponding to that buffer segment, and then summing all the multiplication results.

[0048] The time recalibration function value of the device at the current sampling time is the corrected endogenous degradation time, and its numerical unit is consistent with the sampling time unit (such as seconds). By subtracting the total time shift contribution, the exogenous time shift caused by the propagation of the space flow is offset, so that the load data at the corrected time is more consistent with the device's own degradation state. Specifically, the time recalibration function value of the device at the current sampling time is equal to the current sampling time minus the sum of the products of each convection intensity weight and the corresponding space convection delay map element.

[0049] The decimal part of the time recalibration function value is the value obtained by subtracting its integer part from the function value (e.g., if the function value is 10.3 seconds, the decimal part is 0.3), and the range of this value is [0,1). The purpose of the decimal part of the time recalibration function value is to calculate the interpolation weights (first weight, second weight) to avoid load data errors caused by truncation of integer time points.

[0050] The floor time corresponding to the time recalibration function value is the integer part of the function value (e.g., if the function value is 10.3 seconds, the floor time is 10 seconds). This floor time is the historical time when the load data has been collected.

[0051] The rounding up time corresponding to the time recalibration function value is the time obtained by adding one to the rounding down time (e.g., the rounding down time is 10 seconds and the rounding up time is 11 seconds). This time is also the historical time when the load data has been collected.

[0052] The first weight is the interpolation weight that matches the load data of the device at the time of rounding down. The value range is (0,1]. The smaller the decimal part, the larger the first weight, reflecting the stronger contribution of the load data at the time of rounding down to the aligned sequence data. The purpose of the first weight is to adjust the proportion of the load data at the time of rounding down in the interpolation. Specifically, the first weight is equal to 1 minus the decimal part of the time recalibration function value.

[0053] The second weight is the interpolation weight that matches the load data of the device at the rounding up time. The value range is [0,1). The larger the decimal part, the larger the second weight, reflecting the stronger contribution of the load data at the rounding up time to the aligned sequence data. The purpose of the second weight is to adjust the proportion of the load data at the rounding up time in the interpolation. Specifically, the second weight is equal to the decimal part of the time recalibration function value.

[0054] The aligned sequence data of the device at the current sampling time is obtained by linear interpolation and is load data after eliminating exogenous time shift. The data unit is consistent with the load data unit, and its value can more accurately reflect the intrinsic degradation state of the device. Specifically, the aligned sequence data of the device at the current sampling time is equal to the first product plus the second product.

[0055] The device alignment sequence is a time-series data formed by arranging the alignment sequence data of the device at each sampling time in chronological order. This sequence has been stripped of exogenous time shifts caused by the propagation of the void flow and only retains the load changes related to the device's own degradation.

[0056] In one embodiment of the present invention, generating a direction-gated matrix based on an unbalanced vacancy load circulation track includes: Set the direction-gated gain constant; Calculate the product of the directional gating gain constant, the unbalanced vacant load circulation trace, and the downstream distance weight. Input this product into a logic function to obtain the directional gating matrix elements for the corresponding device and buffer section. The logic function is a monotonic function that maps real numbers to zero and one. The direction gating matrix is ​​formed by the combination of all the direction gating matrix elements corresponding to all devices and buffer sections.

[0057] The directional gating gain constant is a hyperparameter that adjusts the range of the product of the directional gating gain constant, the unbalanced vacant load circulation trace, and the downstream distance weight. The value of the directional gating gain constant should be combined with the typical amplitude settings (such as 0.1-1) of the unbalanced vacant load circulation trace (typical value 0.1-1) and the downstream distance weight (typical value 1-5) to avoid the product being too large, which would cause the logic function output to saturate (close to 1) or too small, which would cause the output to be invalid (close to 0).

[0058] The product of the directional gating gain constant, the unbalanced empty load circulation trace, and the downstream distance weight is obtained for each device and each buffer segment by multiplying the directional gating gain constant, the unbalanced empty load circulation trace corresponding to that buffer segment, and the downstream distance weight corresponding to that device and that buffer segment. Its magnitude reflects the combined effect of directional propagation intensity × position contribution × gain adjustment. The purpose of this product is to serve as input to the logic function, providing the original values ​​for the elements of the directional gating matrix. Specifically, this product equals the directional gating gain constant multiplied by the unbalanced empty load circulation trace corresponding to that buffer segment, and then multiplied by the downstream distance weight corresponding to that device and that buffer segment.

[0059] The logic function is a monotonically increasing function that converts the product (real number range) of the direction-gated gain constant, the unbalanced vacant load circulation trace, and the downstream distance weight into values ​​in the interval [0,1]. Here, the Sigmoid function is preferred (because this function can smoothly map real numbers to [0,1], and outputs 0.5 when the product is 0, approaches 1 when the product increases, and approaches 0 when the product decreases, which is suitable for the gradual change requirement of weak-coupling to strong-coupling in gating). The purpose of the logic function is to normalize the product to [0,1]. Specifically, if the Sigmoid function is selected, the result of the logic function is equal to 1 divided by (1 plus the negative product power of the natural constant).

[0060] The direction gating matrix element corresponding to the device and buffer section is the output value obtained by inputting the product of the direction gating gain constant, the unbalanced empty load circulation trace, and the downstream distance weight into a logic function. The value range is [0,1]. The larger the value, the stronger the cross-device information coupling strength of the empty flow of the buffer section to the device. The smaller the value, the weaker the coupling (no coupling when it is 0). Specifically, the direction gating matrix element corresponding to the device and buffer section is equal to the result obtained by inputting the product of the direction gating gain constant, the unbalanced empty load circulation trace, and the downstream distance weight of the device and the buffer section into a logic function.

[0061] The direction gating matrix is ​​a two-dimensional matrix formed by arranging the direction gating matrix elements of the corresponding devices and buffer sections according to the device-buffer section dimension (the rows of the matrix represent devices, the columns represent buffer sections, and the value of each position in the matrix is ​​the gating element of the corresponding device and buffer section). Only the elements corresponding to downstream devices and upstream buffer sections have non-zero values ​​(because the downstream distance weight is only non-zero for downstream devices), and the elements corresponding to upstream devices and downstream buffer sections are 0, ensuring that information is coupled only along the direction of space flow propagation (downstream).

[0062] In one embodiment of the present invention, applying the directional gating matrix and the empty convection delay map to a neural network containing directional residual branches includes: Neural networks consist of a backbone branch and directional residual branches; Define the linear transformation matrix of the residual branches; For each device, the hidden state of the device at the previous moment is called and the aligned sequence data of the device is called. The intermediate state after the backbone is updated is obtained through the single-step transformation function of the neural network. Extract the empty convection delay map elements corresponding to all buffer segments before the device as the residual branch time delay corresponding to each buffer segment, extract the fractional part of each residual branch time delay, and determine the rounding down and rounding up times corresponding to each residual branch time delay. Obtain the hidden state of the upstream device corresponding to each buffer segment at the time of rounding down and the time of rounding up. Subtract the corresponding decimal part from 1 to obtain the first weight, and use the corresponding decimal part as the second weight. Calculate the third product of the first weight and the hidden state at the time of rounding down, and the fourth product of the second weight and the hidden state at the time of rounding up. Add the third product and the fourth product to obtain the residual branch aligned hidden state of the corresponding buffer segment. The upstream device of each buffer segment is the device that is adjacent to the buffer segment and located on the upstream side of the buffer segment. Extract the directional gating matrix elements corresponding to each buffer segment, calculate the fifth product of each directional gating matrix element, the linear transformation matrix of the residual branch, and the corresponding residual branch aligned hidden state, and the sum of all fifth products constitutes the residual term of the directional residual branch output; The hidden state of the device at the current moment is obtained by adding the intermediate state after the trunk update to the residual term.

[0063] The backbone of the neural network is the core branch responsible for learning the intrinsic degradation dynamics of the device. It only takes into account the hidden state of the device at the previous moment and the aligned sequence data of the device, without introducing cross-device information, thus avoiding interference from exogenous cascades (such as space flow propagation) on the learning of the device's own degradation.

[0064] The directional residual branch of the neural network is a branch that only transmits information about the equipment corresponding to the upstream buffer section adjacent to the previous equipment. It is constrained by the direction gating matrix and the empty space convection delay map, and does not transmit downstream or non-adjacent upstream information, adapting to the law of directional propagation of empty space flow along the process.

[0065] The residual branch linear transformation matrix is ​​a square matrix that maps the hidden state of the upstream device to the hidden state of the current device. The dimension is the hidden layer dimension of the neural network × the hidden layer dimension of the neural network (e.g., if the hidden layer dimension is 64, the matrix will be 64 rows and 64 columns). The initial value of this matrix is ​​randomly initialized by Xavier (to avoid gradient explosion caused by the initial value being too large or gradient vanishing caused by the initial value being too small), and is subsequently updated as the neural network is jointly trained.

[0066] The hidden state of the device at the previous moment is a vector calculated by the neural network at the previous sampling moment, reflecting the degenerate state of the device up to the previous moment, and its dimension is consistent with the dimension of the hidden layer of the neural network.

[0067] The neural network single-step transformation function is a time-series model function used to update the hidden states of the main branch. The single-step update function of a gated recurrent unit (GRU) is preferred (because GRU can efficiently capture time dependencies and adapt to the degradation time-series characteristics of device load). This function's purpose is to fuse historical hidden states with real-time aligned sequence data and output the intermediate states of the main branch. Specifically, the first step is to calculate the reset gate: the reset gate equals the logic function (1 divided by 1 plus the negative 'reset gate weight matrix of the natural constant multiplied by the hidden state of the device at the previous time step plus the reset gate weight matrix multiplied by the aligned sequence data of the device plus the reset gate bias raised to the power of '). The second step... The first step is to calculate the candidate hidden state: the candidate hidden state is equal to the hyperbolic tangent function (candidate state weight matrix multiplied by the hidden state at the previous time step of the device multiplied by the reset gate plus the candidate state weight matrix multiplied by the aligned sequence data of the device plus the candidate state bias); the second step is to calculate the update gate: the update gate is equal to the logic function (1 divided by 1 plus the negative of the natural constant, update gate weight matrix multiplied by the hidden state at the previous time step of the device plus the update gate weight matrix multiplied by the aligned sequence data of the device plus the update gate bias); the third step is to calculate the update gate: the update gate is equal to the logic function (1 divided by 1 plus the negative of the natural constant, update gate weight matrix multiplied by the hidden state at the previous time step of the device plus update gate bias); the fourth step is to output the intermediate state: the intermediate state after the trunk update is equal to the hidden state at the previous time step of the device multiplied by (1 minus the update gate) plus the candidate hidden state multiplied by the update gate.

[0068] The intermediate state after the backbone update is the output vector after the single-step transformation of the backbone branch of the neural network. Its dimension is the same as that of the hidden layer and it only contains the intrinsic degradation information of the device. Specifically, the intermediate state after the backbone update is equal to the result obtained by processing the hidden state of the device at the previous time step and the aligned sequence data of the device through the neural network single-step transformation function (GRU single-step update function).

[0069] The residual branch time delay corresponding to all buffer segments before the device is extracted from the empty convection delay map and corresponds to all buffer segments before the device (device number is greater than buffer segment number). The numerical unit is the number of sampling steps (which can be converted to actual time in combination with the sampling interval).

[0070] The fractional part of the time delay for each residual branch is the value obtained by subtracting its integer part from the time delay (e.g., if the time delay is 2.4 steps, the fractional part is 0.4), and the value range is [0,1).

[0071] The rounding down time corresponding to the time delay of each residual branch is the integer part of the time delay (e.g., if the time delay is 2.4 steps, the rounding down time is the current time minus 2 steps), and the rounding up time is the rounding down time minus 1 step (e.g., the rounding down time is t-2, and the rounding up time is t-3); both times are historical times when the hidden state has been calculated by the upstream device.

[0072] Each buffer segment corresponds to an upstream device that is directly adjacent to the physical location of the buffer segment. Specifically, the upstream device of buffer segment k (located between device k and device k+1) is device k (e.g., the upstream device of buffer segment 1 is device 1, and the upstream device of buffer segment 2 is device 2), to avoid ambiguity in the range of upstream devices.

[0073] The hidden state of the upstream device at the time of rounding down is the hidden state calculated by the upstream device at the time of rounding down, and the hidden state of the upstream device at the time of rounding up is the hidden state calculated by the upstream device at the time of rounding up.

[0074] The first weight for the hidden state of the residual branch is the interpolation weight that matches the hidden state of the upstream device at the time of rounding down. The value range is (0,1]. The smaller the fractional part, the larger the first weight, reflecting the stronger contribution of the hidden state at the time of rounding down to the interpolation result. Specifically, the first weight is equal to 1 minus the fractional part of the time delay of each residual branch.

[0075] The second weight for the hidden state of the residual branch is an interpolation weight that matches the hidden state of the upstream device at the rounding up time. The value range is [0,1). The larger the fractional part, the larger the second weight, reflecting the stronger contribution of the hidden state at the rounding up time to the interpolation result. Specifically, the second weight is equal to the fractional part of the time delay of each residual branch.

[0076] The third product is for the residual branch, and is the result of element-wise multiplication of the first weight and the hidden state of the upstream device at the round-down time. The vector dimension is the same as the hidden layer dimension. Specifically, the third product is equal to the first weight multiplied by the hidden state of the upstream device at the round-down time (element-wise multiplication).

[0077] The fourth product is for the residual branch, and is the result of element-wise multiplication of the second weight and the hidden state of the upstream device at the rounding-up time. The vector dimension is the same as the hidden layer dimension. Specifically, the fourth product is equal to the second weight multiplied by the hidden state of the upstream device at the rounding-up time (element-wise multiplication).

[0078] The residual branch aligned hidden state is the upstream device hidden state that is aligned with the current device timing through linear interpolation, eliminating the impact of the time shift of the empty flow propagation on cross-device information; the purpose of this hidden state is to serve as the input data for the directional residual branch; specifically, the residual branch aligned hidden state is equal to the third product plus the fourth product.

[0079] The fifth product is the result of multiplying the elements of the direction-gated matrix, the linear transformation matrix of the residual branch, and the hidden state aligned with the residual branch in sequence for each buffer segment (matrix multiplication of the hidden state first, then element-wise multiplication of the gated elements). The vector dimension is consistent with the hidden layer dimension. Specifically, the fifth product is equal to the element of the direction-gated matrix multiplied by (the linear transformation matrix of the residual branch multiplied by the hidden state aligned with the residual branch) (matrix multiplication first, then element-wise multiplication).

[0080] The residual term output by the directional residual branch is a vector obtained by adding the fifth products corresponding to all buffer segments before the device element by element, with the same dimension as the hidden layer dimension; specifically, the residual term output by the directional residual branch is equal to the sum of the fifth products corresponding to all buffer segments before the device (element by element).

[0081] The current hidden state of the device is the final state that integrates endogenous degradation information (intermediate state of the backbone) and exogenous directional concatenation information (residual terms), and its dimension is consistent with the dimension of the hidden layer. Specifically, the current hidden state of the device is equal to the intermediate state after the backbone is updated plus the residual terms output by the directional residual branches (adding element by element).

[0082] In one embodiment of the present invention, a neural network is jointly trained using time-reversal irreversible representation loss and main task loss. The trained neural network outputs device time risk, including: Set the joint loss weight constant, risk mapping vector, risk bias constant, and sampling time interval; For each time window, extract the start and end times of the time window, calculate the average value of the hidden state of all devices at each sampling time within the time window, and obtain the window representation of the time window; Extract the average value of the hidden state of all devices within the time window at the corresponding sampling time in reverse chronological order to obtain the time inversion window representation of the time window; Extract the convection intensity weight corresponding to each buffer segment, calculate the ratio of each convection intensity weight to 1 plus the corresponding convection intensity weight, and use the average of all ratios as the time inversion weight for that time window; The difference between the calculated window representation and the time-reversal window representation is multiplied by the square of the difference and the time-reversal weight as the time-reversal irreversible representation loss for that time window. The sum of the time-reversal irreversible representation losses for all time windows constitutes the total time-reversal irreversible representation loss. For each device's hidden state at each sampling time, the device time risk at that sampling time is calculated using the risk mapping vector and the risk bias constant; The probability of the device occurring at a given sampling time is calculated based on the device's time risk and the sampling time interval. The main task loss of all devices at each sampling time within each time window is calculated by combining the supervision label and the occurrence probability. The sum of the main task losses of all time windows constitutes the total main task loss. Among them, the supervision label is zero or one. The joint loss is obtained by multiplying the total main task loss and the total time inversion irreversible representation loss by the joint loss weight constant. The neural network is trained based on joint loss. After training, the relevant data of the devices are input into the neural network, and the neural network outputs the device time risk of each device at the corresponding sampling time.

[0083] The joint loss weight constant is a hyperparameter that adjusts the proportion of the total main task loss and the total time inversion irreversible representation loss in the joint loss. The value of the joint loss weight constant ranges from 0.1 to 0.5 (a smaller value is taken if the main task loss fluctuates greatly, and a larger value is taken if the time inversion irreversible representation loss fluctuates greatly), to avoid a single loss dominating the training.

[0084] The risk mapping vector is a vector that maps the hidden state of a device to the temporal risk of the device. The dimension of the risk mapping vector is the same as the dimension of the hidden layer of the neural network (e.g., if the hidden layer dimension is 64, the vector is 64-dimensional). The initial value of this vector is randomly initialized by Xavier and subsequently updated as the neural network is jointly trained.

[0085] The risk bias constant is a scalar constant used to adjust the equipment time risk baseline value. The initial value is set to 0.01 (to avoid the initial risk value being too small).

[0086] The sampling time interval is the time difference between two consecutive acquisitions of device load data and buffer placeholder data. The sampling time interval is consistent with the data acquisition time interval (e.g., 2-5 seconds).

[0087] The start and end times of a time window are two points in time that define the data range of a single time window (e.g., the start time is t1, the end time is t2, and the window length is t2-t1).

[0088] The window representation of a time window is a vector obtained by globally averaging the hidden states (vectors) of all devices at all sampling times within the time window, with the same dimension as the hidden layer. Specifically, the window representation is equal to the sum of the hidden states of all devices at each sampling time within the time window, and then divided by (the number of devices included in the time window multiplied by the number of sampling times included in the time window).

[0089] The time-reversal window representation of a time window is to arrange the sampling moments within the time window in reverse order (e.g., if the original moments are t1 < t2 < t3, after reversing, it becomes t3 < t2 < t1), and then calculate the global average vector of the hidden states of all devices at the reversed moments, with the dimension being the same as that of the hidden layer; specifically, the time-reversal window representation is equal to the sum of the hidden states of all devices at each reversed sampling moment within the time window, and then divided by (the number of devices included in the time window multiplied by the number of sampling moments included in the time window).

[0090] The time-reversal weight of a time window is a scalar value, ranging from [0, 0.5) (because the convection intensity weight is non-negative). The greater the convection intensity weight, the greater this ratio, and the greater the time-reversal weight, indicating that the stronger the non-equilibrium, the heavier the penalty for time-reversal asymmetry; specifically, the time-reversal weight is equal to first calculating the convection intensity weight for each buffer segment divided by (1 plus the convection intensity weight), then summing up this ratio for all buffer segments, and finally dividing by the total number of buffer segments.

[0091] The difference between the window representation and the time-reversal window representation is the element-wise difference of two vectors, with the dimension being the same as that of the hidden layer; the greater this difference, the stronger the time-reversal asymmetry; specifically, the difference between the window representation and the time-reversal window representation is equal to the window representation minus the time-reversal window representation element-wise.

[0092] The square of the difference between the window representation and the time-reversal window representation is the element-wise square of the difference vector between the window representation and the time-reversal window representation, with the dimension being the same as that of the hidden layer; specifically, this square of the difference is equal to the difference between the window representation and the time-reversal window representation multiplied by itself element-wise.

[0093] The time-reversal irreversible representation loss of a time window is a scalar value. The stronger the non-equilibrium (the greater the time-reversal weight and the greater the square of the difference), the greater the loss value, which is used to constrain the neural network to learn the non-equilibrium representation with a time arrow; specifically, the time-reversal irreversible representation loss of this time window is equal to the sum of all elements of the square of this difference multiplied by the time-reversal weight.

[0094] The total time-reversal irreversible representation loss is a scalar value obtained by summing up the time-reversal irreversible representation losses of all training time windows; specifically, the total time-reversal irreversible representation loss is equal to the sum of the time-reversal irreversible representation losses of all time windows.

[0095] The device time risk at this sampling moment is a scalar value, reflecting the instantaneous risk intensity of the device failing at this moment. The greater the value, the higher the probability of failure; specifically, the device time risk is equal to the dot product (the sum of the products of vector elements) of the risk mapping vector and the hidden state of the device at this sampling moment plus the risk bias constant.

[0096] The probability of the device occurring at this sampling time is a scalar value, ranging from [0,1], used to convert the risk value into an index that conforms to the definition of probability; specifically, the probability of occurrence is equal to 1 divided by (1 plus the negative device time risk of the natural constant multiplied by the sampling time interval raised to the power of 1).

[0097] The monitoring label is a fault status label of the device at the sampling time, recorded manually or by the system. A value of 1 indicates that the device is faulty at that time, and a value of 0 indicates that the device is normal at that time.

[0098] The loss of the main task within each time window is a scalar value, using binary cross-entropy loss (since the supervision label is 0 / 1), to quantify the difference between the occurrence probability and the supervision label. Specifically, the loss of the main task within each time window is equal to the logarithm of the supervision label multiplied by the negative probability power of the natural constant at each sampling time of all devices within that window, plus the logarithm of (1 minus supervision label) multiplied by the negative (1 minus probability) power of the natural constant, and then the average of all the negative results.

[0099] The total principal task loss is a scalar value that is the sum of the principal task losses for all training time windows; the total principal task loss is used to constrain the neural network's ability to predict fault risk; specifically, the total principal task loss is equal to the sum of the principal task losses for all time windows.

[0100] The joint loss is a scalar value that combines the accuracy of fault risk prediction (main task loss) with the rationality of unbalanced representation (time inversion loss). The smaller the value, the better the performance of the neural network. Specifically, the joint loss is equal to the total main task loss plus (total time inversion irreversible representation loss multiplied by the joint loss weight constant).

[0101] Training a neural network based on joint loss employs gradient descent (such as the Adam optimizer). It calculates the gradient of the joint loss with respect to all neural network parameters (such as GRU weights and risk mapping vectors) through backpropagation, updates the parameters in the gradient direction, and iterates until the joint loss value changes by less than 0.001 for 10 consecutive iterations (convergence threshold). Specifically, the steps are: first, initialize the neural network parameters (weights, vectors, constants); second, input training data in batches (aligned sequences, initial hidden state values) and calculate the joint loss; third, use the Adam optimizer to calculate the loss gradient and update the parameters (learning rate set to 0.001); fourth, repeat steps two and three until the joint loss changes by less than 0.001 for 10 consecutive iterations, at which point training stops.

[0102] The device time risk output after training is the scalar risk value obtained by inputting the real-time aligned sequence of the device into the trained neural network, calculating the hidden state through the main branch and the directional residual branch, and then using the risk mapping vector and the bias constant. Specifically, the device time risk output after training is equal to the dot product of the trained risk mapping vector and the hidden state of the device at that sampling time, plus the trained risk bias constant.

[0103] In one embodiment of the present invention, during online inference, the real-time device time risk includes: For each buffer segment, the real-time empty space density of the buffer segment at the corresponding real-time sampling time is obtained by using the buffer capacity of the buffer segment and the real-time buffer occupancy data of the buffer segment at the corresponding real-time sampling time. The real-time empty space flow of the buffer segment at the current real-time sampling time is obtained by using the real-time empty space density of the buffer segment at the current real-time sampling time and the real-time empty space density of the buffer segment at the previous real-time sampling time. Determine the current time window, forward time delay steps, maximum time delay steps, and hysteresis decay constant. For each device and each buffer segment, calculate the real-time cross-correlation between real-time device load data and real-time space flow under forward time delay by combining the forward time delay steps and the current time window. Calculate the real-time cross-correlation between real-time device load data and real-time space flow under reverse time delay by combining the reverse time delay steps and the current time window. Subtract the real-time cross-correlation corresponding to the reverse time delay from the real-time cross-correlation corresponding to the forward time delay to obtain the antisymmetric part of the real-time cross-correlation. Based on the antisymmetric part of real-time cross-correlation, the lag weight function, the downstream distance weight, the energy normalization factor of real-time equipment load data, and the energy normalization factor of real-time empty space flow, the real-time unbalanced empty space load circulation trace of each buffer segment under the current time window is calculated. Based on the antisymmetric part of real-time cross correlation, the positive values ​​corresponding to each positive time delay step are extracted, and the real-time space convection delay map elements corresponding to each device and each buffer segment are calculated. All real-time space convection delay map elements together constitute the real-time space convection delay map. Determine the convection intensity scaling constant, and for each buffer segment, calculate the real-time convection intensity weight based on the real-time unbalanced vacancy load circulation trace and the convection intensity scaling constant corresponding to that buffer segment. For each device, a current time window containing the current real-time sampling time is determined. The real-time convection intensity weights and real-time empty space convection delay map elements corresponding to all buffer segments before the device under the current time window are extracted. The real-time time recalibration function value of the device at the current real-time sampling time is calculated. The real-time alignment sequence of the device is obtained based on the real-time time recalibration function value. The real-time alignment sequence of each device is input into the trained neural network, and the neural network outputs the real-time device time risk of each device at the corresponding real-time sampling time.

[0104] Real-time device time risk is the instantaneous risk value output by the trained neural network during online inference, reflecting the probability of device failure at the current moment. The input is replaced by the real-time alignment sequence. Specifically, the real-time device time risk is equal to the dot product of the trained risk mapping vector and the hidden state of the device at the current real-time sampling moment (calculated by the neural network from the real-time alignment sequence), plus the trained risk bias constant.

[0105] In one embodiment of the present invention, the real-time device time risk of each device is accumulated within a preset decision window, and the device with the highest total risk is output as the faulty device, including: Set the decision window length; The decision window is determined based on the decision window length and the current time. For each device, extract the real-time device time risk corresponding to all real-time sampling times within the decision window, calculate the product of each real-time device time risk and the sampling time interval, and the sum of all products constitutes the total risk of the device within the preset decision window; Compare the total risk of all devices and output the device with the highest total risk as the faulty device.

[0106] The decision window length is a fixed time span used to accumulate real-time time risks of equipment. The decision window length needs to be set in combination with the production line fault warning cycle (e.g., 5-15 minutes) and the fluctuation characteristics of real-time equipment time risks (take a longer value for large risk fluctuations and a shorter value for small fluctuations) to ensure that the accumulation process of fault risks can be fully captured and to avoid misjudgment caused by the randomness of risks at a single moment.

[0107] The decision window is a time interval determined by the current time and the decision window length. The start time of the interval is the current time minus the decision window length, and the end time is the current time (e.g., if the current time is 10:15 and the decision window length is 5 minutes, the decision window is 10:10-10:15). This window must contain an integer number of real-time sampling times (if this is not met, the start time is slightly adjusted to include an integer number of sampling times) to ensure the integrity of the risk data. The purpose of the decision window is to define the data boundary for extracting real-time device time risks. Specifically, the start time of the decision window is equal to the current time minus the decision window length. If the number of real-time sampling times included from the start time to the current time is not an integer, the start time is adjusted to the current time minus (an integer number of sampling times × the sampling time interval). The final decision window is the interval from the adjusted start time to the current time.

[0108] The product of the time risk of each real-time device and the sampling time interval is the time-weighted risk contribution corresponding to the time risk of the real-time device at a single moment (the time risk of the real-time device is an instantaneous value, which is multiplied by the sampling time interval to obtain the cumulative risk contribution within the sampling period, thus transforming the instantaneous risk into the periodic cumulative risk); specifically, the product of the time risk of each real-time device and the sampling time interval is equal to the real-time device time risk of the device at a certain real-time sampling moment within the decision window multiplied by the sampling time interval.

[0109] The total risk of a device within a preset decision window is a scalar value that is the sum of the products of the time risks of all real-time devices within the decision window and the sampling time interval. It reflects the cumulative intensity of the fault risk of the device within the decision window. The larger the value, the higher the probability of failure. Specifically, the total risk of the device within the preset decision window is equal to the sum of the products of the real-time device time risk of the device and the sampling time interval corresponding to all real-time sampling moments within the decision window.

[0110] The device with the highest total risk is the one with the highest total risk value after comparing the total risk values ​​of all devices within the preset decision window. If multiple devices have the same total risk and both are at the maximum value (e.g., two devices have a total risk of 15), then the average real-time device time risk value of the last three sampling times within the decision window is further compared, and the device with the larger average value is identified as the faulty device.

[0111] The embodiments of this example have been described above. However, this example is not limited to the specific implementation methods described above. The specific implementation methods described above are merely illustrative and not restrictive. Those skilled in the art can make many other forms based on the guidance of this example, and all of them are within the protection scope of this example.

Claims

1. A predictive maintenance system based on artificial intelligence, characterized in that, include: The data acquisition module collects equipment load data and buffer occupancy data, calculates vacancy density and vacancy flow based on the buffer occupancy data, and then calculates the unbalanced vacancy load circulation trace based on the equipment load data and vacancy flow. The data processing module obtains the convection delay map of the unbalanced empty space load circulation trace, constructs a time recalibration function, and uses the time recalibration function to process the equipment load data to obtain an alignment sequence. The feature extraction module generates a directional gating matrix based on the unbalanced vacancy load circulation trace, and applies the directional gating matrix and the vacancy convection delay map to a neural network containing directional residual branches. The training module uses time-reversal irreversible representation loss and main task loss to jointly train the neural network, and outputs the device time risk through the trained neural network. During online inference, the risk calculation module collects real-time device load data and real-time buffer placeholder data, and repeatedly executes the data collection module to the risk calculation module to obtain the real-time device time risk. Accumulate the real-time device time risk of each device within the preset decision window, and output the device with the highest total risk as the faulty device.

2. The predictive maintenance system based on artificial intelligence according to claim 1, characterized in that, Collect equipment load data and buffer occupancy data, and calculate vacancy density and vacancy flow based on the buffer occupancy data, including: Collect device load data and buffer placeholder data; For each buffer segment, the empty space density of the buffer segment at the corresponding sampling time is obtained by using the buffer capacity of the buffer segment and the buffer occupancy data of the buffer segment at the corresponding sampling time. For each buffer segment, the space flow of that buffer segment at the current sampling time is obtained by using the space density of that buffer segment at the current sampling time and the space density of that buffer segment at the previous sampling time.

3. The predictive maintenance system based on artificial intelligence according to claim 2, characterized in that, The unbalanced vacancy load circulation trace is calculated based on equipment load data and vacancy flow, including: Determine the time window, the number of positive time delay steps, the maximum number of time delay steps, and the hysteresis decay constant; For each device and each buffer segment, the cross-correlation between device load data and space flow at two moments under the positive time delay is calculated by combining the positive time delay step and time window. For each device and each buffer segment, the cross-correlation between device load data and space flow at two moments under the reverse time delay is calculated by combining the reverse time delay steps and time window. Subtracting the cross-correlation between the two time points corresponding to the positive time delay from the cross-correlation between the two time points corresponding to the negative time delay yields the antisymmetric part of the cross-correlation. Determine the downstream distance weight based on the location relationship between the equipment and the buffer section, and only consider the downstream equipment; determine the hysteresis weight function that changes with the number of time delay steps. The numerator is obtained by summing the products of the hysteresis weight function, the downstream distance weight, and the antisymmetric part of the cross-correlation under all positive time delay steps. The device load data energy normalization factor is calculated based on the load data of all devices within the time window. The empty flow energy normalization factor is calculated based on the empty flow of the corresponding buffer segment within the time window. The device load data energy normalization factor and the empty flow energy normalization factor are multiplied together to obtain the denominator. Dividing the numerator by the denominator yields the unbalanced vacancy load circulation trace for each buffer segment within the corresponding time window.

4. The predictive maintenance system based on artificial intelligence according to claim 3, characterized in that, Based on the unbalanced vacancy load circulation trace, a vacancy convection delay map is obtained, including: For each device and each buffer segment, extract the positive values ​​from the antisymmetric part of the cross-correlation corresponding to each positive time delay step; The numerator is obtained by summing the products of all positive time delay steps and the corresponding positive antisymmetric values ​​of the cross-correlation. Calculate the sum of the antisymmetric positive values ​​of the cross-correlation under all positive time delay steps to obtain the denominator; Divide the numerator by the denominator to obtain the empty convection delay map element corresponding to each device and each buffer segment. The empty convection delay map is constructed based on all empty convection delay map elements.

5. The predictive maintenance system based on artificial intelligence according to claim 4, characterized in that, Construct a time recalibration function, and use the time recalibration function to process the device load data to obtain an aligned sequence, including: Set the convection intensity scaling constant; For each buffer segment, the convection intensity weight is calculated based on the unbalanced vacancy load circulation trace and the convection intensity scaling constant corresponding to that buffer segment. For each device, a time window containing the current sampling time is determined. The convection intensity weights and empty space convection delay map elements corresponding to all buffer segments before the device in the time window are extracted. The sum of the products of each convection intensity weight and the corresponding empty space convection delay map element is calculated. The sum of the products is subtracted from the current sampling time to obtain the time recalibration function value of the device at the current sampling time. Extract the decimal part of the time recalibration function value to determine the floor and floor times corresponding to that value. Obtain the load data of the device at the floor and floor times. Subtract the decimal part from 1 to obtain the first weight. Use the decimal part as the second weight. Calculate the first product of the first weight and the load data at the floor time, and the second product of the second weight and the load data at the floor time. Add the first and second products to obtain the alignment sequence data of the device at the current sampling time. The alignment sequence data of all sampling times together constitute the alignment sequence of the device.

6. The predictive maintenance system based on artificial intelligence according to claim 5, characterized in that, The direction-gated matrix is ​​generated based on the unbalanced vacancy load circulation trajectory, including: Set the direction-gated gain constant; Calculate the product of the directional gating gain constant, the unbalanced vacant load circulation trace, and the downstream distance weight. Input this product into a logic function to obtain the directional gating matrix elements for the corresponding device and buffer section. The logic function is a monotonic function that maps real numbers to zero and one. The direction gating matrix is ​​formed by the combination of all the direction gating matrix elements corresponding to all devices and buffer sections.

7. The predictive maintenance system based on artificial intelligence according to claim 6, characterized in that, Applying directional gating matrices and vacant convection delay maps to neural networks containing directional residual branches includes: Neural networks consist of a backbone branch and directional residual branches; Define the linear transformation matrix of the residual branches; For each device, the hidden state of the device at the previous moment is called and the aligned sequence data of the device is called. The intermediate state after the backbone is updated is obtained through the single-step transformation function of the neural network. Extract the empty convection delay map elements corresponding to all buffer segments before the device as the residual branch time delay corresponding to each buffer segment, extract the fractional part of each residual branch time delay, and determine the rounding down and rounding up times corresponding to each residual branch time delay. Obtain the hidden state of the upstream device corresponding to each buffer segment at the time of rounding down and the time of rounding up. Subtract the corresponding decimal part from 1 to obtain the first weight, and use the corresponding decimal part as the second weight. Calculate the third product of the first weight and the hidden state at the time of rounding down, and the fourth product of the second weight and the hidden state at the time of rounding up. Add the third product and the fourth product to obtain the residual branch aligned hidden state of the corresponding buffer segment. The upstream device of each buffer segment is the device that is adjacent to the buffer segment and located on the upstream side of the buffer segment. Extract the directional gating matrix elements corresponding to each buffer segment, calculate the fifth product of each directional gating matrix element, the linear transformation matrix of the residual branch, and the corresponding residual branch aligned hidden state, and the sum of all fifth products constitutes the residual term of the directional residual branch output; The hidden state of the device at the current moment is obtained by adding the intermediate state after the trunk update to the residual term.

8. The predictive maintenance system based on artificial intelligence according to claim 7, characterized in that, A neural network is jointly trained using time-reversal irreversible representation loss and main task loss. The trained neural network outputs device time risk, including: Set the joint loss weight constant, risk mapping vector, risk bias constant, and sampling time interval; For each time window, extract the start and end times of the time window, calculate the average value of the hidden state of all devices at each sampling time within the time window, and obtain the window representation of the time window; Extract the average value of the hidden state of all devices within the time window at the corresponding sampling time in reverse chronological order to obtain the time inversion window representation of the time window; Extract the convection intensity weight corresponding to each buffer segment, calculate the ratio of each convection intensity weight to 1 plus the corresponding convection intensity weight, and use the average of all ratios as the time inversion weight for that time window; The difference between the calculated window representation and the time-reversal window representation is multiplied by the square of the difference and the time-reversal weight as the time-reversal irreversible representation loss for that time window. The sum of the time-reversal irreversible representation losses for all time windows constitutes the total time-reversal irreversible representation loss. For each device's hidden state at each sampling time, the device time risk at that sampling time is calculated using the risk mapping vector and the risk bias constant; The probability of the device occurring at a given sampling time is calculated based on the device's time risk and the sampling time interval. The main task loss of all devices at each sampling time within each time window is calculated by combining the supervision label and the occurrence probability. The sum of the main task losses of all time windows constitutes the total main task loss. Among them, the supervision label is zero or one. The joint loss is obtained by multiplying the total main task loss and the total time inversion irreversible representation loss by the joint loss weight constant. The neural network is trained based on joint loss. After training, the relevant data of the devices are input into the neural network, and the neural network outputs the device time risk of each device at the corresponding sampling time.

9. The predictive maintenance system based on artificial intelligence according to claim 8, characterized in that, During online inference, real-time device time risks include: For each buffer segment, the real-time empty space density of the buffer segment at the corresponding real-time sampling time is obtained by using the buffer capacity of the buffer segment and the real-time buffer occupancy data of the buffer segment at the corresponding real-time sampling time. The real-time empty space flow of the buffer segment at the current real-time sampling time is obtained by using the real-time empty space density of the buffer segment at the current real-time sampling time and the real-time empty space density of the buffer segment at the previous real-time sampling time. Determine the current time window, forward time delay steps, maximum time delay steps, and hysteresis decay constant. For each device and each buffer segment, calculate the real-time cross-correlation between real-time device load data and real-time space flow under forward time delay by combining the forward time delay steps and the current time window. Calculate the real-time cross-correlation between real-time device load data and real-time space flow under reverse time delay by combining the reverse time delay steps and the current time window. Subtract the real-time cross-correlation corresponding to the reverse time delay from the real-time cross-correlation corresponding to the forward time delay to obtain the antisymmetric part of the real-time cross-correlation. Based on the antisymmetric part of real-time cross-correlation, the lag weight function, the downstream distance weight, the energy normalization factor of real-time equipment load data, and the energy normalization factor of real-time empty space flow, the real-time unbalanced empty space load circulation trace of each buffer segment under the current time window is calculated. Based on the antisymmetric part of real-time cross correlation, the positive values ​​corresponding to each positive time delay step are extracted, and the real-time space convection delay map elements corresponding to each device and each buffer segment are calculated. All real-time space convection delay map elements together constitute the real-time space convection delay map. Determine the convection intensity scaling constant, and for each buffer segment, calculate the real-time convection intensity weight based on the real-time unbalanced vacancy load circulation trace and the convection intensity scaling constant corresponding to that buffer segment. For each device, a current time window containing the current real-time sampling time is determined. The real-time convection intensity weights and real-time empty space convection delay map elements corresponding to all buffer segments before the device under the current time window are extracted. The real-time time recalibration function value of the device at the current real-time sampling time is calculated. The real-time alignment sequence of the device is obtained based on the real-time time recalibration function value. The real-time alignment sequence of each device is input into the trained neural network, and the neural network outputs the real-time device time risk of each device at the corresponding real-time sampling time.

10. The predictive maintenance system based on artificial intelligence according to claim 9, characterized in that, Accumulate the real-time device time risk of each device within a preset decision window, and output the device with the highest total risk as the faulty device, including: Set the decision window length; The decision window is determined based on the decision window length and the current time. For each device, extract the real-time device time risk corresponding to all real-time sampling times within the decision window, calculate the product of each real-time device time risk and the sampling time interval, and the sum of all products constitutes the total risk of the device within the preset decision window; Compare the total risk of all devices and output the device with the highest total risk as the faulty device.

Images