Predictive maintenance method and system for thin-layer cell preparation machine combined with multi-source data

Abstract

The application provides a thin-layer cell preparation machine predictive maintenance method and system combining multi-source data, and belongs to the technical field of intelligent operation and maintenance of equipment, and comprises the following steps: in response to the target preparation machine completing a preset preparation cycle, acquiring a real-time position sequence fed back by a servo motor encoder and a triggering delay sequence of a motion shaft in-place photoelectric sensor at a corresponding moment; calculating a positioning deviation sequence, determining a current gap estimation value of a motion shaft transmission chain; acquiring an internal multi-point temperature sequence, calculating an irreversible wear drift rate; inputting the irreversible wear drift rate into a precision decay trend model to obtain a predicted remaining operation cycle number; combining a steady-state deviation combined amount to correct the predicted remaining operation cycle number to generate a predictive maintenance instruction. The application solves the problem that the positioning precision degradation trend of the motion shaft transmission chain of the thin-layer cell preparation machine caused by the coupling effect of wear and thermal deformation is difficult to accurately predict.

Description

Technical Field

[0001] This invention relates to the field of intelligent equipment operation and maintenance technology, specifically to a predictive maintenance method and system for thin-layer cell slide making machines that combines multi-source data. Background Technology

[0002] Thin-layer cell slide preparation machines are key automated devices used in cytology screening. They contain multiple high-precision motion axes for accurate positioning during sample transfer, smearing, and staining. Each axis is driven by a servo motor, transmitting power to the end-effector via couplings, ball screws, or synchronous belts. The positioning accuracy directly determines the slide quality. In actual operation, the slide preparation machine needs to complete hundreds to thousands of slide preparation cycles daily under high-frequency start-stop and continuous operation conditions. The motion axis transmission chains are subjected to alternating loads over long periods, making wear and tear unavoidable.

[0003] Currently, monitoring the status of motion axes in thin-layer cell slide preparation machines typically relies on the alarm functions built into the servo drives, such as overload alarms and position deviation alarms. However, these only trigger a shutdown when the positioning deviation exceeds the set tolerance, representing a passive, reactive protection approach. Some solutions introduce periodic maintenance strategies, replacing transmission components at fixed intervals or after a fixed number of runs. However, these solutions fail to consider the actual wear differences between different equipment and shafts, easily leading to under-maintenance or over-maintenance. A few solutions detect transmission chain abnormalities by collecting motor current or vibration signals. However, current signals are not sensitive to early, minor wear, and vibration signals are easily interfered with by surrounding equipment, resulting in delayed fault detection or a high false alarm rate.

[0004] The existing methods mentioned above have not effectively addressed the coupling between transmission chain wear and temperature effects. After prolonged operation, heat sources such as motors and bearings in thin-layer cell preparation machines cause thermal expansion in transmission components, introducing reversible temperature-induced positioning drift. This reversible drift and the drift caused by irreversible wear are superimposed in the same measurement channel, making it difficult to distinguish the contribution ratio of the two based solely on positional deviation or time delay data. Without decoupling the temperature effect, it is difficult to accurately extract the true wear rate, resulting in unreliable predictions of the remaining service life of the motion axis. Therefore, there is an urgent need for a maintenance method that can separate thermally induced reversible drift and irreversible wear from multi-source data and quantitatively predict the trend of motion axis accuracy decay, in order to achieve accurate predictive maintenance of the transmission chain of the motion axis in thin-layer cell preparation machines. Summary of the Invention

[0005] This invention provides a predictive maintenance method and system for thin-layer cell slide machines that combines multi-source data, which is used to solve the problem that it is difficult to accurately predict the trend of positioning accuracy degradation caused by wear and thermal deformation coupling in the motion shaft drive chain of thin-layer cell slide machines.

[0006] In view of the above problems, the present invention provides a predictive maintenance method and system for thin-layer cell slide making machines that incorporates multi-source data.

[0007] In a first aspect, the present invention provides a predictive maintenance method for a thin-layer cell slide preparation machine that incorporates multi-source data, comprising: In response to the target film-making machine completing the preset number of film-making cycles, the real-time position sequence fed back by the servo motor encoder and the trigger delay sequence of the motion axis positioning photoelectric sensor at the corresponding moment are obtained; Based on the difference between the real-time position sequence and the command position, the positioning deviation sequence is calculated, and the steady-state deviation of the positioning deviation sequence is extracted using kernel density estimation. Based on the over-limit ratio of the trigger delay sequence, the current clearance estimate of the motion shaft transmission chain is determined. The internal multi-point temperature sequence of the target film-making machine during operation is obtained. Based on the hysteresis relationship between the temperature gradient and the steady-state deviation, the temperature-induced reversible drift component is decoupled. Combined with the current gap estimate, the irreversible wear drift rate is calculated. The irreversible wear drift rate is input into a pre-built accuracy decay trend model, and the current clearance estimate is used as the state update value to extrapolate and obtain the estimated remaining number of running cycles when the motion axis positioning accuracy drifts to the preset maintenance threshold. The estimated remaining operating cycle number is corrected by combining the steady-state deviation, and the corrected remaining operating cycle number is obtained. When the corrected remaining operating cycle number is lower than the maintenance intervention window, a predictive maintenance instruction containing the associated axis number is generated.

[0008] Secondly, the present invention provides a predictive maintenance system for a thin-layer cell slide preparation machine that incorporates multi-source data, comprising: The multi-source signal synchronous acquisition module is used to respond to the target film-making machine completing the preset number of film-making cycles, and to acquire the real-time position sequence fed back by the servo motor encoder and the trigger delay sequence of the motion axis positioning photoelectric sensor at the corresponding moment; The transmission chain clearance quantification and evaluation module is used to calculate the positioning deviation sequence based on the difference between the real-time position sequence and the command position, extract the steady-state deviation of the positioning deviation sequence using kernel density estimation, and determine the current clearance estimate of the motion shaft transmission chain based on the over-limit ratio of the trigger delay sequence. The temperature-wear effect decoupling module is used to obtain the internal multi-point temperature sequence during the operation of the target film-making machine, and based on the hysteresis relationship between the temperature gradient and the steady-state deviation, decouple the temperature-induced reversible drift component, and calculate the irreversible wear drift rate by combining the current gap estimate. The accuracy decay trend extrapolation module is used to input the irreversible wear drift rate into the pre-built accuracy decay trend model, and at the same time use the current gap estimate as the state update amount to extrapolate to obtain the expected number of remaining running cycles when the motion axis positioning accuracy drifts to the preset maintenance threshold. The integrated correction and maintenance decision module is used to correct the expected remaining number of running cycles by combining the steady-state deviation, and obtain the corrected remaining number of running cycles. When the corrected remaining number of running cycles is lower than the maintenance intervention window, a predictive maintenance instruction containing the associated axis number is generated.

[0009] One or more technical solutions provided in this invention have at least the following technical effects or advantages: The technical solution of this invention first automatically injects a step positioning test command when the number of film production cycles reaches the target, and simultaneously acquires two types of dual data: the real-time position of the encoder and the trigger delay of the photoelectric sensor. After commutation transient rejection, a high-quality alignment sequence is output, providing an online data foundation for transmission chain status assessment without disassembly. Automated testing replaces manual inspection, and data acquisition does not interrupt the normal film production process.

[0010] Furthermore, kernel density estimation is performed on the difference sequence between real-time position and commanded position, and the steady-state deviation sum is extracted by weighting the region with the highest probability density. At the same time, the out-of-limit threshold is adaptively determined based on the statistical out-of-limit ratio of the trigger delay and converted into a spatial gap value. The data-driven method replaces the fixed threshold, making the deviation sum more resistant to interference and the gap estimation adaptively change with the wear state, realizing the independent quantification of electrical side following deviation and mechanical side gap.

[0011] Furthermore, the lag time is determined through hysteresis cross-correlation analysis of multi-point temperature sequences and deviation drift sequences. The mapping relationship between temperature gradient and reversible drift is fitted using the least squares method. After removing the temperature-induced reversible thermal drift component, the irreversible deviation component is obtained. Then, the irreversible wear drift rate is calculated by incorporating the cumulative running time and the current gap estimate into the gap correction factor. This eliminates the interference of temperature fluctuations on wear assessment and incorporates the amplification effect of gap deterioration on wear acceleration into the quantification, outputting a pure index that only reflects the rate of mechanical degradation.

[0012] Furthermore, using the irreversible wear drift rate and the current clearance estimate as dual inputs, a pre-built multilayer sensor accuracy decay trend model is extrapolated to output the estimated remaining operating cycles when the motion axis positioning accuracy drifts to a preset maintenance threshold. By learning the nonlinear coupling relationship between wear and clearance through historical multi-cycle data, the status is dynamically updated with real-time clearance values ​​during online prediction, enabling the remaining life prediction to adapt to the accelerating effect of clearance deterioration.

[0013] Finally, a confidence correction coefficient is generated using an S-shaped mapping based on the width of the steady-state deviation confidence interval. This coefficient is used to correct for the confidence loss of the estimated remaining operating cycles. Simultaneously, a maintenance intervention window is constructed based on a preset maintenance threshold, irreversible wear drift rate, and preset production cycle number. When the corrected remaining cycle number falls below this window, a maintenance instruction with axis number is generated. This ensures that the prediction automatically converges to a conservative estimate when the deviation diverges. The maintenance window employs a dual-constraint strategy of test interval and degradation rate, balancing timely early warning with false alarm control.

[0014] In summary, the technical solution of this invention starts with the acquisition of dual data from the transmission chain, and through deviation statistical quantification, temperature-wear effect decoupling, data-driven trend extrapolation, and reliability correction, forms a complete predictive maintenance decision-making chain. It solves problems in traditional methods such as the inability to separate thermal deformation and mechanical wear, poor adaptability of fixed thresholds, failure to consider gap acceleration effects in remaining life prediction, and issues with prediction reliability. This enables online, accurate prediction and timely early warning of the precision degradation of the transmission chain of the motion shaft in a thin-layer cell slide preparation machine. Attached Figure Description

[0015] Figure 1 This is a flowchart illustrating the predictive maintenance method for thin-layer cell slide machines that incorporates multi-source data, provided by the present invention.

[0016] Figure 2 This is a logical schematic diagram of the predictive maintenance method for thin-layer cell slide machines that combines multi-source data, provided by the present invention.

[0017] Figure 3 This is a schematic diagram of the predictive maintenance system for a thin-layer cell slide machine that combines multi-source data, provided by the present invention.

[0018] In the attached diagram, the labels representing each component are as follows: Multi-source signal synchronous acquisition module 11, transmission chain gap quantification evaluation module 12, temperature-wear effect decoupling module 13, accuracy attenuation trend extrapolation module 14, comprehensive correction and maintenance decision module 15. Detailed Implementation

[0019] This invention provides a predictive maintenance method and system for thin-layer cell slide machines that combines multi-source data, solving the problem that it is difficult to accurately predict the trend of positioning accuracy degradation caused by wear and thermal deformation coupling in the motion shaft drive chain of thin-layer cell slide machines.

[0020] It should be noted that the terms "comprising" and "having" are intended to cover non-exclusive inclusion. For example, a process, method, system, product, or server that includes a series of steps or units is not necessarily limited to those steps or units that are explicitly listed, but may include other steps or modules that are not explicitly listed or that are inherent to these processes, methods, products, or devices.

[0021] Example 1, as Figure 1 , Figure 2 As shown, this invention provides a predictive maintenance method for a thin-layer cell preparation machine that incorporates multi-source data, the method comprising: S100: In response to the target film-making machine completing the preset number of film-making cycles, obtain the real-time position sequence fed back by the servo motor encoder and the trigger delay sequence of the motion axis positioning photoelectric sensor at the corresponding moment.

[0022] After the film-making machine completes the preset number of film-making cycles, this step synchronously collects the real-time position sequence of the encoder at the drive end and the trigger delay sequence of the photoelectric sensor at the execution end. This provides a dual data basis for the quantitative evaluation of the transmission chain gap. The real-time position sequence reflects the output accuracy at the motor end, and the trigger delay sequence reflects the terminal positioning deviation. The combination of the two can separate the respective contributions of the motor following error and the mechanical transmission gap.

[0023] Step S100 in the method provided by the present invention includes: Read the film production cycle counter from the target film production machine control software to obtain the current cumulative film production cycle count; When the cumulative number of film production cycles reaches the preset number of film production cycles, a step positioning test command is injected into the motion control device to drive the target motion axis to reciprocate through the reference position calibrated by the photoelectric sensor at a constant speed. During the execution of the step positioning test command, the real-time position sequence fed back by the encoder is recorded synchronously with the internal sampling period of the servo driver, and the command position sequence is recorded accordingly. A positioning deviation sequence is generated based on the difference between the real-time position sequence and the corresponding time-instructed position sequence; The delay values ​​from the issuance of the command to the sensor level flip are collected in each step positioning test, and arranged in the order of collection to form a trigger delay sequence; The positioning deviation sequence and the trigger delay sequence collected within the same test cycle are aligned by timestamp. After removing the transient data segment at the moment of servo motor commutation, the aligned real-time position sequence and trigger delay sequence are output.

[0024] In this step, the film production cycle counter is first read from the target film production machine control software to obtain the current cumulative film production cycle count in real time. When the cumulative film production cycle count reaches the preset film production cycle count, for example, when a test is triggered every 500 film production cycles, the drive train gap detection process is automatically initiated.

[0025] For example, during continuous operation, the control software of a TCT-2000 thin-layer cell slide preparation machine continuously accumulates the slide preparation cycle count in the background. When the count value jumps from 499 to 500, it is determined that the preset slide preparation cycle number has been reached, and the subsequent test process is immediately triggered.

[0026] Furthermore, a preset step positioning test command is injected into the motion control device, driving the target motion axis to reciprocate at a constant speed through the reference position calibrated by the photoelectric sensor. The reference position is the zero point of the photoelectric sensor's installation and calibration. The step size and speed of the step positioning test command are set according to the normal film-making motion range to ensure that the test process does not exceed the mechanical limits.

[0027] For example, a step test command is injected into the X-axis motion control device to drive the X-axis to move forward at a constant speed of 50 mm / s to the trigger position of the photoelectric sensor and then back in the opposite direction. This is repeated 10 times, each time crossing the reference scale point marked by the photoelectric sensor on the guide rail.

[0028] Furthermore, during the execution of the step positioning test command, the real-time position sequence fed back by the encoder is synchronously recorded using the internal sampling period of the servo drive, such as 125μs, and the command position sequence at the corresponding moment is also recorded. The real-time position is the motor shaft end position value measured by the encoder, and the command position is the theoretical position value issued by the motion controller at the corresponding moment.

[0029] For example, during the first forward step of the X-axis, the command position values ​​are continuously recorded at a sampling interval of 125μs as 0.00mm, 0.05mm, and 0.10mm, respectively. The real-time position values ​​fed back by the encoder at the corresponding times are 0.00mm, 0.04mm, and 0.11mm, respectively. The two sets of sequences are stored one-to-one according to the timestamp.

[0030] Furthermore, based on the difference between the real-time position sequence and the commanded position sequence at the corresponding moment, the positioning deviation sequence is calculated point by point. The deviation value at each sampling point reflects the following error between the actual output position of the motor and the commanded position at the current moment.

[0031] For example, using the real-time position sequence and the corresponding command position sequence obtained in the previous example, the first pair of data, 0.00mm and 0.00mm, has a deviation of 0.00mm; the second pair of data, 0.05mm and 0.04mm, has a deviation of +0.01mm, indicating that the command is ahead; the third pair of data, 0.10mm and 0.11mm, has a deviation of -0.01mm, indicating that the command is behind. This process continues to generate a positioning deviation sequence containing 1000 sampling points.

[0032] Furthermore, the delay value from the issuance of the command to the sensor level flipping is collected during each step positioning test. The delay value is calculated as follows: Delay value = Time when the photoelectric sensor is triggered by the motion axis and completes the level flipping - Time when each step command is issued. The calculated delay values ​​are arranged in the order of acquisition to form a trigger delay sequence.

[0033] For example, the first positive step command is issued at 10:00:00.000, and the photoelectric sensor completes the level flip at 10:00:00.152, with a delay of 152ms; the second negative step command is issued at 10:00:02.000, and the photoelectric sensor completes the level flip at 10:00:02.147, with a delay of 147ms. A total of 20 delay values ​​are recorded for 10 cycles, arranged sequentially as [152,147,155,149,151,148,157,146,153,150,154,148,156,147,152,149,155,150,153,148]ms, which is the trigger delay sequence.

[0034] Finally, the positioning deviation sequence and trigger delay sequence collected within the same test cycle are aligned by timestamps, so that each delay value corresponds to a continuously recorded positioning deviation sequence. Transient data segments at the moment of servo motor commutation, i.e., the time period when the motor speed crosses zero and the acceleration is at its maximum, are removed, and the aligned real-time position sequence and trigger delay sequence are output.

[0035] For example, the trigger delay sequence of 20 step jumps is concatenated and aligned with the corresponding positioning deviation sequence by timestamp. Data segments of 50ms before and after the motor commutation moment in each step jump are identified; these segments exhibit abrupt acceleration changes and drastic data fluctuations, and are therefore removed from the sequence. After removal, the remaining valid data segments serve as the final aligned real-time position sequence and trigger delay sequence, which are then sent to the subsequent gap quantization evaluation module for analysis.

[0036] It should be noted that the above values ​​are for illustrative purposes only and do not constitute a limitation on the present invention.

[0037] In summary, this step automatically injects a step positioning test command when the number of film production cycles meets the target, simultaneously acquires the real-time position of the encoder and the trigger delay data of the photoelectric sensor, and outputs a high-quality data sequence after alignment and commutation transient rejection; it enables online quantitative evaluation of the transmission chain clearance state without disassembling the equipment, providing a reliable data foundation for subsequent wear decoupling and accuracy prediction, and reducing downtime and human error caused by manual inspection.

[0038] S200: Based on the difference between the real-time position sequence and the commanded position, calculate the positioning deviation sequence, and use kernel density estimation to extract the steady-state deviation of the positioning deviation sequence. Based on the over-limit ratio of the trigger delay sequence, determine the current clearance estimate of the motion shaft transmission chain.

[0039] This step extracts steady-state characteristic quantities that represent the overall deviation of the transmission chain from the difference between the aligned real-time position sequence and the command position. At the same time, it uses photoelectric sensors to trigger the over-limit ratio of the delay and independently calculate the physical clearance value of the transmission chain, thereby achieving decoupling and quantification of electrical side deviation and mechanical side clearance.

[0040] Step S200 in the method provided by the present invention includes: Based on the difference between the real-time position sequence and the commanded position, a positioning deviation sequence is calculated, and the steady-state deviation sum of the positioning deviation sequence is extracted using kernel density estimation, including: The encoder feedback position is extracted point by point from the aligned real-time position sequence, and the difference is calculated with the command position in the command position sequence at the same timestamp to obtain the positioning deviation sequence containing the forward and reverse travel. The remaining deviation sequence is obtained by removing the deviation values ​​within a fixed range on both sides of the sampling point corresponding to the instant of servo motor commutation in the positioning deviation sequence; According to the constant velocity direction set by the step positioning test command, the remaining deviation value is separated into a positive deviation subsequence and a negative deviation subsequence; The ratio of the average delay of the trigger delay sequence in the reverse travel segment to the average delay in the forward travel segment is used as the reverse gap feature value, and then the reverse gap influence weight is obtained after S-shaped mapping. The positive deviation subsequence and the negative deviation subsequence are weighted and fused according to the negative gap influence weight to generate a comprehensive deviation sequence. Using the comprehensive deviation sequence as input, the deviation probability density distribution is obtained by kernel density estimation, and the deviation value corresponding to the maximum probability density value is extracted as the steady-state deviation sum.

[0041] In this step, the encoder feedback position is first extracted point by point from the aligned real-time position sequence. This is then subtracted from the command position in the command position sequence at the same timestamp to obtain a positioning deviation sequence containing both forward and reverse travel. The forward travel deviation is defined as the command position minus the encoder feedback position, and the reverse travel deviation uses the same definition.

[0042] For example, in a step test, the commanded position of a sampling point in the forward X-axis travel is 10.00mm, and the encoder feedback position is 9.95mm, a difference of +0.05mm, indicating that the actual position of the motor lags behind the commanded position by 0.05mm. In the reverse travel, the commanded position of a sampling point is 10.00mm, and the encoder feedback position is 10.04mm, a difference of -0.04mm, indicating that the actual position of the motor leads the commanded position by 0.04mm. Arranging the differences of all sampling points in chronological order yields a positioning deviation sequence containing 1000 deviation values.

[0043] Furthermore, deviation values ​​within a fixed range on either side of the sampling point corresponding to the instant of servo motor commutation are removed from the positioning deviation sequence. The fixed range is centered on the zero-crossing point of the real-time speed signal fed back by the servo driver, and extends forward and backward by a set time window. For example, sampling points within 5ms are removed to obtain the remaining deviation sequence. Following the constant speed direction set by the step positioning test command, the remaining deviation values ​​are separated into a positive deviation subsequence and a negative deviation subsequence. Deviation values ​​collected during the forward stroke are assigned to the positive deviation subsequence, and deviation values ​​collected during the reverse stroke are assigned to the negative deviation subsequence.

[0044] For example, among the 1000 sampling points generated above, the two reversal times were identified at points 198 and 502, respectively. Centered on points 198 and 502, the 40 sampling points before and after each point (i.e., sampling points within 5ms) were removed. The deviation values ​​of points 158 to 238 and 462 to 542 were marked as transient data and removed. The remaining deviation sequence after removal contained 840 valid deviation values. Following the constant velocity direction set by the step positioning test command, points 1 to 157 and 239 to 461 were assigned to the forward subsequence, with deviation values ​​ranging from approximately +0.03mm to +0.08mm; points 463 to 501 and 543 to 840 were assigned to the reverse subsequence, with deviation values ​​ranging from approximately -0.02mm to -0.07mm.

[0045] Furthermore, the ratio of the average delay of the trigger delay sequence in the reverse travel segment to the average delay in the forward travel segment is used as the reverse gap feature value, and then the reverse gap influence weight is obtained after S-shaped mapping. The S-shaped mapping function can be a Logistic function, which, through preset slope factors and flip centers, reflects the influence of the reverse gap feature value and the forward / reverse deviation on the reverse gap influence weight. Generally, the slope factor and flip center can be set to 20 and 1.0, respectively.

[0046] For example, in the trigger delay sequence, the average delay of 10 steps in the forward travel is 152ms, and the average delay of 10 steps in the reverse travel is 157ms. Therefore, the characteristic value of the backlash is approximately 157 / 152 ≈ 1.033. Inputting R = 1.033 into the S-shaped mapping function, with a preset slope factor of 20 and a flip center of 1.0, the output backlash influence weight is 0.66, indicating that the influence of the reverse deviation subsequence on the overall deviation is 66%.

[0047] Furthermore, the positive deviation subsequence and the negative deviation subsequence are weighted and fused according to the influence weight of the negative gap to generate a comprehensive deviation sequence. The weighted fusion method is: comprehensive deviation value = (1 - influence weight of negative gap) × positive deviation value + influence weight of negative gap × absolute value of negative deviation value.

[0048] For example, if the weight of the backlash effect is 0.66, a certain deviation in the positive deviation subsequence is +0.055mm, and the absolute value of a corresponding deviation in the negative deviation subsequence is 0.045mm, then the overall deviation level of the fused comprehensive deviation sequence is (1-0.66)×0.055+0.66×0.045=0.0187+0.0297=0.0484mm. The comprehensive deviation sequence is formed by alternating positive and negative deviation values ​​in chronological order.

[0049] Finally, using the comprehensive deviation sequence as input, kernel density estimation is employed to obtain the deviation probability density distribution. The deviation value corresponding to the maximum probability density value is extracted as the steady-state deviation aggregate. The specific steps include: Sort all deviation values ​​in the comprehensive deviation sequence according to their numerical values, calculate the distance between adjacent deviation values ​​after sorting, and take the median of the distance as the bandwidth of the kernel density estimation. Using the bandwidth as a smoothing parameter, a kernel function is placed for each deviation value in the comprehensive deviation sequence, and all kernel functions are superimposed on the deviation value axis to obtain the deviation probability density distribution curve. Identify the global maximum point from the deviation probability density distribution curve, and use the deviation value corresponding to the global maximum point as the initial steady-state deviation; Centered on the initial steady-state deviation, search to the left and right for positions where the probability density drops to half of the global maximum value, thus forming the steady-state deviation confidence interval; The weighted average of all deviation values ​​within the steady-state deviation confidence interval is extracted as the steady-state deviation sum, and the weighting coefficient is taken as the probability density value corresponding to each deviation value.

[0050] Specifically, firstly, all deviation values ​​in the comprehensive deviation sequence are sorted by numerical value, and the distance between adjacent deviation values ​​after sorting is calculated. The median of the distances is then used as the bandwidth for kernel density estimation. Preferably, the median of the distances between adjacent deviations within the 25th to 75th percentile range of the sorted deviation value sequence can be used to enhance the robustness of bandwidth estimation.

[0051] For example, the comprehensive deviation sequence contains 840 deviation values, which range from 0.010 mm to 0.098 mm after sorting. The interval between adjacent deviations is calculated to obtain 839 interval values. The interval values ​​from the 210th to the 629th are selected, corresponding to the 25% to 75% percentile interval. The median is calculated to be 0.003 mm, and 0.003 mm is used as the kernel density estimation bandwidth.

[0052] Furthermore, using a kernel density estimation bandwidth of 0.003 mm as the smoothing parameter, a Gaussian kernel function is applied to each deviation value in the comprehensive deviation sequence. When calculating the probability density estimate at a certain position, all kernel functions are superimposed on the deviation value axis and normalized to obtain the corresponding probability density estimate. The probability density distribution curve of the deviation is obtained by fitting the probability density at each position. The center position of the kernel function represents each deviation value, and the standard deviation is equal to the kernel density estimation bandwidth.

[0053] For example, a Gaussian kernel function is placed on each of the 840 deviation values ​​along the deviation axis from 0.010 mm to 0.098 mm. Each kernel function expands outwards from its corresponding deviation value by three standard deviations, with a standard deviation of 0.003 mm. After superimposing all the kernel functions, a significant density peak is formed in the range of 0.045 mm to 0.052 mm, while the probability density is lower near 0.02 mm and 0.08 mm, and the curve as a whole exhibits a single-peak shape.

[0054] Furthermore, the global maximum point is identified from the deviation probability density distribution curve, and the deviation value corresponding to the global maximum point is used as the initial steady-state deviation. For example, on the deviation probability density distribution curve generated by superposition, a scan search is performed from 0.01 mm to 0.10 mm, and the global maximum point is found at 0.048 mm, with a corresponding probability density value of 85.3, which is the initial steady-state deviation value. This indicates that the most common comprehensive deviation of the transmission chain in the current state is 0.048 mm.

[0055] Furthermore, using the initial steady-state deviation as the center, the search extends to the left and right to find positions where the probability density drops to half of the global maximum value, forming the steady-state deviation confidence interval. For example, at 0.048 mm, the probability density is 85.3. Searching to the left to 0.043 mm, the probability density drops to 42.1, slightly below the half-peak, so 0.044 mm is taken as the left boundary; searching to the right to 0.054 mm, the probability density drops to 43.0, slightly above the half-peak, so 0.053 mm is taken as the right boundary. The steady-state deviation confidence interval is [0.044 mm, 0.053 mm].

[0056] Finally, a weighted average is used to extract the steady-state deviation summation. This involves extracting the weighted average of all deviation values ​​within the steady-state deviation confidence interval as the steady-state deviation summation, with the weighting coefficients being the probability density values ​​corresponding to each deviation value. For example, the confidence interval [0.044mm, 0.053mm] contains 186 deviation values. Each deviation value is multiplied by its corresponding probability density value, summed, and then divided by the sum of the probability density values ​​to obtain a weighted average of 0.0483mm. This is the steady-state deviation summation under the current test cycle, used for subsequent wear rate calculation and life prediction.

[0057] It should be noted that the above values ​​are for illustrative purposes only and do not constitute a limitation on the present invention.

[0058] Step S200 in the method provided by the present invention further includes: Arrange the delay values ​​in the trigger delay sequence in ascending order, and take the delay value located at the lower quartile after sorting as the baseline response delay; The median absolute deviation of the trigger delay sequence is calculated and superimposed with the reference response delay to form an over-limit discrimination threshold; The number of sampling points whose delay values ​​exceed the over-limit discrimination threshold is counted, and the over-limit ratio is obtained by combining the total length of the trigger delay sequence; Using the aforementioned over-limit ratio as the quantile basis, the delay value corresponding to the high quantile is extracted from the delay values ​​arranged in ascending order as the gap feature delay; The effective gap delay is obtained by comparing the difference between the gap characteristic delay and the reference response delay. Then, the current gap estimate of the motion shaft transmission chain is obtained by combining the constant speed value set in the step positioning test command.

[0059] In this step, all delay values ​​in the trigger delay sequence are first sorted in ascending order, and the delay value at the lower quartile, i.e., the 25th percentile, is taken as the baseline response delay. The basis for selecting the lower quartile is that the smaller value in the delay sequence corresponds to the state of rapid response at the end after the transmission chain gap is compressed, which can represent the ideal response level under gapless or micro-gap conditions.

[0060] For example, using the 20 trigger delay values ​​from the aforementioned 10-cycle X-axis step test, arranged in ascending order as [146, 147, 147, 148, 148, 148, 149, 149, 150, 150, 151, 152, 152, 153, 153, 154, 155, 155, 156, 157] ms, the lower quartile of these 20 values ​​is the 5th value, which is 148 ms, and the baseline response delay is 148 ms.

[0061] Furthermore, the median absolute deviation of the trigger delay sequence is calculated and superimposed with the reference response delay to form an over-limit discrimination threshold. The median absolute deviation reflects the overall dispersion of the delay value. After being superimposed with the reference value, it forms an adaptive threshold line. Delay values ​​exceeding the adaptive threshold line are considered to be affected by additional hysteresis caused by transmission backlash.

[0062] The median absolute deviation is calculated as follows: first, the median of the delayed sequence is taken, and then the median of the absolute values ​​of the differences between each delayed value and the median is calculated. The over-limit discrimination threshold is calculated as follows: over-limit discrimination threshold = reference response delay + k × median absolute deviation, where k is the threshold coefficient, which is 1.5 by default.

[0063] For example, the median of the 20 delay values ​​is 150ms, which is the average of the 10th and 11th values. The absolute deviation of each delay value from the median of 150ms is [4,3,3,2,2,2,1,1,0,0,1,2,2,3,3,4,5,5,6,7]ms. After sorting in ascending order, the median is 2ms, that is, the median absolute deviation is 2ms. The over-limit detection threshold = 148 + 1.5 × 2 = 151ms.

[0064] Furthermore, the number of sampling points whose delay values ​​exceed the over-limit judgment threshold is counted, and the over-limit ratio is calculated by combining this number with the total length of the trigger delay sequence. The over-limit ratio = number of sampling points exceeding the threshold ÷ total sequence length × 100%, reflecting the proportion of delay anomalies caused by the current transmission chain gap.

[0065] For example, among the 20 delay values, 9 values ​​[152, 152, 153, 153, 154, 155, 155, 156, 157] exceeded the over-limit judgment threshold of 151ms. The over-limit ratio is 9 ÷ 20 × 100% = 45%, indicating that nearly half of the step stroke of the current transmission chain is affected by a non-negligible gap.

[0066] Furthermore, using the over-limit ratio as the quantile basis, the delay value corresponding to the highest quantile is extracted from the delay values ​​arranged in ascending order as the gap feature delay. The position of the highest quantile is 1 minus the over-limit ratio, that is, the delay value at the (1 - over-limit ratio) × 100% quantile is taken. The purpose is to locate the relatively stable upper boundary of the delay values ​​that are most significantly affected by the gap.

[0067] For example, if the over-limit ratio is 45%, then the high quantile is 1% - 45% = 55%, that is, the delay value corresponding to the 55th quantile is taken. Among the 20 values, the 55th quantile is the 11th value. To ensure continuity, the weighted average of the 11th and 12th values ​​is taken, and a linear interpolation of 152ms is used as the gap characteristic delay.

[0068] Finally, based on the difference between the gap characteristic delay and the reference response delay, the effective gap delay is obtained. Then, combined with the constant speed value set in the step positioning test command, the time quantity is converted into a spatial quantity to obtain the current estimated gap value of the motion shaft transmission chain. Wherein, the estimated gap value = (gap characteristic delay - reference response delay) × constant speed value of the step test.

[0069] For example, if the gap characteristic delay is 152ms and the reference response delay is 148ms, then the effective gap delay = 152 - 148 = 4ms. The constant speed set for the step test is 50mm / s, so the current estimated gap value = 4ms × 50mm / s = 0.2mm. This means that under the current wear condition, the drive chain needs to idle for an additional 0.2mm during reversal to start moving the load end.

[0070] It should be noted that the above values ​​are for illustrative purposes only and do not constitute a limitation on the present invention.

[0071] In summary, this step uses a data-driven statistical method to replace fixed threshold judgment, so that the total deviation is not affected by individual abnormal sampling points, and the gap estimation value can adapt to the actual wear state of the transmission chain, providing stable and comparable input characteristics for subsequent temperature decoupling and accuracy life prediction.

[0072] S300: Obtain the internal multi-point temperature sequence during the operation of the target film-making machine, decouple the temperature-induced reversible drift component based on the hysteresis relationship between the temperature gradient and the steady-state deviation, and calculate the irreversible wear drift rate by combining the current gap estimate.

[0073] This step utilizes the thermal hysteresis mapping relationship between the multi-point temperature sequence inside the film-making machine and the steady-state deviation output from the previous step to separate the thermal drift portion of the positioning deviation that changes reversibly with temperature. This part is then fused with the estimated value of the current clearance in the transmission chain to obtain the irreversible drift rate that only reflects mechanical wear. This provides a pure degradation index after eliminating temperature interference for the subsequent extrapolation of the accuracy decay trend.

[0074] Step S300 in the method provided by the present invention includes: Temperature values ​​are collected from multiple locations inside the target film-making machine by temperature sensors, and the average values ​​of multiple points are taken after aligning them according to the sampling time to form the internal multi-point temperature sequence. The internal multi-point temperature sequence is differentially analyzed to obtain a temperature gradient sequence. At the same time, the corresponding steady-state deviation summation is extracted from each step positioning test to form a deviation drift sequence. The temperature gradient sequence and the deviation drift sequence are subjected to hysteresis cross-correlation analysis along the time axis, and the number of hysteresis sampling points at the point where the cross-correlation coefficient is the maximum is taken as the hysteresis time of temperature-induced reversible drift. The deviation drift sequence and temperature gradient sequence are time-shifted and aligned using the lag time to establish the correspondence between the temperature gradient and the reversible drift component. The temperature-induced reversible drift component is then extracted from the current steady-state deviation sum to obtain the irreversible deviation component. By combining the cumulative running time, the irreversible deviation component is converted into an average drift rate per unit time, and the current gap estimate is incorporated as a correction factor to calculate the irreversible wear drift rate.

[0075] In this step, temperature values ​​are first collected from temperature sensors deployed at multiple locations inside the target wafer fabrication machine. After aligning the sampling times, the average values ​​from multiple points are taken to form an internal multi-point temperature sequence. The temperature sensors are typically located near servo motor windings, bearing housings, drive screw nut pairs, and the machine chassis environment; averaging the values ​​from multiple points can comprehensively reflect the overall thermal state of the drivetrain.

[0076] For example, four temperature sensors are deployed on the X-axis drivetrain, located at the motor end cover, front bearing housing, lead screw nut pair, and chassis back plate, respectively. At a certain sampling moment, the readings of the four sensors are 42.1℃, 39.8℃, 38.5℃, and 35.2℃, respectively. The multi-point average is (42.1 + 39.8 + 38.5 + 35.2) / 4 = 38.9℃, which is stored in the internal multi-point temperature sequence. Continuous recording at 1-minute sampling intervals constitutes a temperature sequence containing 480 sampling points, covering an 8-hour operating time.

[0077] Furthermore, the internal multi-point temperature sequence is differentially analyzed, i.e., the temperature values ​​at two different times are subtracted and divided by the time difference to obtain the temperature gradient sequence. Simultaneously, the corresponding steady-state deviation summation is extracted from each step positioning test to form the deviation drift sequence. The temperature gradient reflects the rate of temperature change per unit time; the deviation drift sequence records the change of the steady-state deviation summation obtained from each test over time.

[0078] For example, the average temperature at multiple points in the temperature sequence is 38.9℃ at the 10th minute and 39.2℃ at the 11th minute, with a temperature gradient of (39.2 - 38.9) / 1min = 0.3℃ / min. Within the same time period, the steady-state deviation obtained from the step test triggered at the 10th minute is 0.0483mm, and the result from the test at the 11th minute is 0.0491mm. These two values ​​are stored in the deviation drift sequence in chronological order.

[0079] Furthermore, a hysteresis cross-correlation analysis was performed on the temperature gradient sequence and the deviation drift sequence along the time axis. The number of hysteresis sampling points at the point of maximum cross-correlation coefficient was taken as the hysteresis time of temperature-induced reversible drift. Heat conduction takes time, and temperature changes are not immediately reflected as changes in positioning deviation. The hysteresis time is a quantitative expression of the thermal equilibrium response delay.

[0080] The specific method for hysteresis cross-correlation analysis is as follows: The temperature gradient sequence is kept stationary, while the deviation drift sequence is gradually shifted along the time axis, i.e., shifted by one sampling interval each time. After each shift, the cross-correlation coefficient between the two sequences is calculated. The shift amount corresponding to the maximum cross-correlation coefficient is the number of thermal hysteresis sampling points. The cross-correlation coefficient is calculated as follows: the covariance between the temperature gradient at the statistical time and the deviation drift at the time after the shifted sampling interval, divided by the product of the standard deviation of the temperature gradient and the standard deviation of the deviation drift.

[0081] For example, 480 temperature gradient values ​​over 8 hours are taken, and 48 deviation drift values ​​are obtained by performing a step test every 10 minutes with a sampling interval of 10 minutes. The cross-correlation coefficient of the temperature gradient sequence relative to the deviation drift sequence is calculated at a lag of N sampling points, where N is an integer greater than or equal to 0. It is found that the cross-correlation coefficient reaches its maximum value of 0.82 at a lag of 3 sampling points, i.e., 30 minutes. Therefore, the lag time is 30 minutes.

[0082] Furthermore, the deviation drift sequence and the temperature gradient sequence are time-shifted and aligned using the lag time to establish the correspondence between the temperature gradient and the reversible drift component. The temperature-induced reversible drift component is then extracted from the current steady-state deviation summation to obtain the irreversible deviation component. Specific steps include: Extract the temperature gradient sequence segment from the temperature response cycle preceding the current moment, and simultaneously extract the deviation drift sequence segment after time shift alignment by the lag time. The temperature gradient sequence segment is used as the independent variable, and the time-shifted and aligned deviation drift sequence segment is used as the dependent variable. The least squares method is used to fit the linear mapping coefficient between the temperature gradient and the deviation drift. By combining the instantaneous temperature gradient of the internal multi-point temperature sequence at the current moment with the linear mapping coefficient, the estimated value of the temperature-induced reversible drift component at the current moment is obtained. Based on the steady-state deviation sum extracted by kernel density estimation at the current moment, and combined with the estimated value of the reversible drift component, the irreversible deviation component is calculated.

[0083] Specifically, the deviation drift sequence and temperature gradient sequence are first time-shifted and aligned using a lag time to establish the correspondence between the temperature gradient and the reversible drift component. A temperature gradient sequence segment from the previous temperature response period is extracted, and the deviation drift sequence segment, time-shifted and aligned using the lag time, is extracted simultaneously. The temperature response period is typically 2 to 3 times the lag time to ensure the complete temperature-drift response process is included.

[0084] Since temperature changes are only fully reflected in the positioning deviation after a lag time, a time-shift alignment operation is performed: the entire deviation drift sequence is shifted forward by the lag time so that the temperature gradient value at the original time and the deviation drift value at the original time after the lag time are at the same time position, forming a point-to-point correspondence.

[0085] For example, the current time is the 8th hour of operation, the lag time is 30 minutes, and the temperature response period is twice the lag time, that is, the temperature gradient sequence segment within the 60 minutes before the current time is taken, with a total of 60 sampling points; the deviation drift sequence segment after 30 minutes of time shift alignment is taken simultaneously, that is, the deviation drift value corresponding to each temperature gradient value within the 60-minute window, such as the temperature gradient of the 7th hour corresponding to the deviation drift value of the 7th hour and 30th minute.

[0086] Furthermore, using the temperature gradient sequence segment as the independent variable and the time-shifted, aligned deviation drift sequence segment as the dependent variable, the least squares method is employed to fit the linear mapping coefficient between the temperature gradient and the deviation drift. The linear mapping relationship is: temperature-induced reversible drift component = linear mapping coefficient × temperature gradient. The least squares method is used because, compared to the least squares method, it provides a more stable fit when temperature fluctuations are large or measurement interference occurs.

[0087] For example, among 60 pairs of data points, the temperature gradient values ​​range from -0.5℃ / min to +0.8℃ / min, and the corresponding deviation drift values ​​range from -0.005mm to +0.008mm. Least squares fitting minimizes the sum of the absolute fitting errors for all data points, resulting in a linear mapping coefficient of 0.0102mm / (℃ / min). This means that for every 1℃ / min change in temperature, a reversible positioning drift of approximately 0.0102mm is expected.

[0088] Furthermore, by combining the instantaneous temperature gradient of the internal multi-point temperature sequence at the current moment with the linear mapping coefficient, the estimated value of the temperature-induced reversible drift component at the current moment is obtained. Here, the reversible drift component = linear mapping coefficient × temperature gradient at the current moment.

[0089] For example, if the average temperature at multiple points one minute ago was 38.9℃ and the average temperature at the current moment is 39.4℃, then the instantaneous temperature gradient is (39.4 - 38.9) / 1min = 0.5℃ / min. The estimated reversible drift component is 0.0102 × 0.5 = 0.0051mm. This indicates that the positioning deviation increases by approximately 0.0051mm due to thermal expansion caused by the temperature increase.

[0090] Finally, based on the steady-state deviation sum extracted by kernel density estimation at the current moment, and combined with the estimated value of the reversible drift component, the irreversible deviation component is calculated. Wherein, the irreversible deviation component = steady-state deviation sum - estimated value of the reversible drift component.

[0091] For example, the preceding steps yield a steady-state deviation of 0.0483 mm at the current moment, an estimated reversible drift component of 0.0051 mm, and an irreversible deviation component of 0.0483 - 0.0051 = 0.0432 mm, representing the purely permanent positioning deviation caused by mechanical wear of the transmission chain after deducting the temperature effect.

[0092] After obtaining the irreversible deviation component, the irreversible deviation component is converted into the average drift rate per unit time by combining the cumulative running time. Then, the current gap estimate is incorporated as a correction factor to calculate the irreversible wear drift rate.

[0093] The average drift rate is calculated as follows: Average drift rate = Irreversible deviation component / Cumulative running time; the irreversible wear drift rate is calculated as follows: Irreversible wear drift rate = Average drift rate × (1 + Clearance influence coefficient × Current clearance estimate / Reference clearance value), 1 + Clearance influence coefficient × Current clearance estimate / Reference clearance value = Correction factor.

[0094] The reference clearance value refers to the mechanical clearance benchmark value of the motion shaft drive chain in a brand-new, unworn state. It is obtained by the equipment manufacturer through precise calibration before leaving the factory and serves as a reference for measuring subsequent wear. Optionally, the factory calibration method involves measuring the positional change of the drive chain during forward and reverse stroke switching under standard temperature conditions using a step positioning test method or a high-precision method such as a laser interferometer. After multiple measurements and averaging, the value is written into the equipment controller firmware or calibration file and permanently stored as the reference clearance value.

[0095] The gap influence coefficient is a dimensionless parameter used to adjust the sensitivity of the effect of gap deterioration on the wear drift rate. It acts as a scaling factor in the gap correction factor of the irreversible wear drift rate calculation formula. A larger gap influence coefficient results in a greater increase in the correction factor for the same gap deterioration ratio, making the wear drift rate more sensitive to gap changes. Conversely, a smaller gap influence coefficient results in a more gradual effect of gap correction. The default value for the gap influence coefficient is typically 0.5.

[0096] For example, with a cumulative running time of 800 hours, an irreversible deviation component of 0.0432 mm, and an average drift rate of 0.0432 / 800 = 0.000054 mm / h, and a current estimated clearance of 0.2 mm, a factory-calibrated clearance of 0.05 mm, and a clearance influence coefficient of 0.5, then the correction factor = 1 + 0.5 × (0.2 / 0.05) = 3. The irreversible wear drift rate = 0.000054 × 3 = 0.000162 mm / h, indicating that after considering the current clearance deterioration, the moving axis actually experiences approximately 0.000162 mm of irreversible positioning accuracy loss per hour.

[0097] It should be noted that the above values ​​are for illustrative purposes only and do not constitute a limitation on the present invention.

[0098] In summary, this step eliminates the interference of temperature fluctuations on positioning deviation, makes wear assessment unaffected by the hot and cold conditions of the equipment, and incorporates the amplified effect of the deterioration of the transmission chain clearance on accelerated wear into the quantification. The output drift rate can truly reflect the mechanical degradation rate, providing a pure degradation index after excluding environmental and operating condition interference for subsequent accuracy life extrapolation.

[0099] S400: Input the irreversible wear drift rate into the pre-built accuracy decay trend model, and use the current gap estimate as the state update value to extrapolate and obtain the expected number of remaining operating cycles when the motion axis positioning accuracy drifts to the preset maintenance threshold.

[0100] This step uses the irreversible wear drift rate as the degradation rate input to drive the pre-built accuracy decay trend model to perform forward extrapolation. At the same time, the current clearance estimate is injected into the model in real time as a state correction, so that the model output of the expected remaining operating cycle number can synchronously reflect the current real clearance state of the transmission chain, providing dynamically updated remaining life prediction values ​​for maintenance decisions.

[0101] In this step, the pre-construction steps of the accuracy decay trend model include: Collect historical operating data of the same model of thin-layer cell slide preparation machine in multiple complete maintenance cycles. For each step positioning test, record the corresponding irreversible wear drift rate and the current gap estimate, as well as the actual number of slide preparation cycles from the corresponding test time as the prediction starting point until the positioning accuracy drift reaches the preset maintenance threshold. The irreversible wear drift rate corresponding to each step positioning test and the current gap estimate are used as input samples, and the corresponding actual number of film production cycles is used as output labels to construct a sample pair set. A multilayer perceptron network is constructed with irreversible wear drift rate and current gap estimate as two inputs and the expected number of remaining running cycles as the output. The input layer has 2 nodes, the hidden layer uses a nonlinear activation function, the output layer has 1 node, and the output value represents the predicted number of remaining running cycles. The sample set is divided into a fitting set and a validation set, and the mean square error between the output layer's predicted value and the actual number of film production cycles is used as the loss function. The network weights of the multilayer perceptron network are updated using backpropagation of error. The network is iteratively fitted until the loss function converges, and the fitted multilayer perceptron network is used as the accuracy decay trend model.

[0102] Specifically, historical operational data of the same model of thin-layer cell slide preparation machine was first collected over multiple complete maintenance cycles. A complete maintenance cycle refers to the entire process from the start of operation after a certain motion axis of the equipment has been repaired or replaced until the positioning accuracy of that axis degrades to a preset maintenance threshold, triggering maintenance. For each step positioning test within this cycle, three data points were recorded: irreversible wear drift rate, current estimated gap value, and the number of slide preparation cycles from the start of the corresponding test time until the positioning accuracy actually reaches the maintenance threshold.

[0103] The preset maintenance threshold refers to the maximum allowable deviation of the positioning accuracy of the motion axis. When the irreversible wear of the transmission chain accumulates and the positioning deviation reaches the preset maintenance threshold, it is determined that the motion axis can no longer meet the requirements of the film production process and must be repaired or replaced.

[0104] The preset maintenance threshold is determined by the precision requirements of the thin-layer cell slide preparation process. During slide preparation, excessive deviation in the positioning of the motion axis can lead to cell transfer misalignment, uneven smear thickness, or inaccurate staining positioning, directly affecting the slide yield. Therefore, the maintenance threshold is usually set to the maximum allowable positioning deviation value of the slide preparation process, which can be obtained from the equipment technical specifications or process validation report.

[0105] For example, extract complete maintenance cycle records of the X-axis drivetrain from the historical database of five identical TCT-2000 film processing machines. One machine underwent 200 step tests during a specific maintenance cycle. Taking the 50th test record: the irreversible wear drift rate was 0.000135 mm / h, the current estimated clearance was 0.2 mm, and after 45,000 film processing cycles, the positioning accuracy reached the maintenance threshold. The input for this sample would be (0.000135, 0.2), and the output label would be 45,000. Summarize the test records for all machines and all maintenance cycles to construct a dataset containing thousands of samples.

[0106] Furthermore, the irreversible wear drift rate corresponding to each step positioning test and the current gap estimate are used as two feature dimensions of the input sample, and the corresponding actual number of film production cycles is used as the output label. Sample pairs are constructed one by one. Each sample pair consists of a two-dimensional input vector and a one-dimensional scalar label.

[0107] Furthermore, a multilayer perceptron network is constructed with the irreversible wear drift rate and the current gap estimate as dual inputs, and the estimated remaining running cycles as the output. The network structure is as follows: the input layer contains two nodes, which receive the irreversible wear drift rate and the current gap estimate, respectively; there are two to three hidden layers in the middle, each containing several nodes, using nonlinear activation functions such as ReLU or tanh; the output layer contains one node, which outputs the predicted value of the estimated remaining running cycles, without setting an activation function or using linear activation to support continuous value regression.

[0108] For example, a four-layer perceptron network is constructed: the input layer has 2 nodes, the first hidden layer has 32 nodes with ReLU activation, the second hidden layer has 16 nodes with ReLU activation, the third hidden layer has 8 nodes with ReLU activation, and the output layer has 1 node with linear activation. The total number of network parameters is approximately (2×32 + 32×16 + 16×8 + 8×1) plus the bias terms of each layer, totaling approximately 800 trainable parameters.

[0109] Furthermore, the sample set is divided into a fitting set and a validation set. For example, the fitting set and validation set are randomly divided in an 8:2 ratio. The fitting set is used to train the model, and the validation set is used to evaluate the model's generalization ability. The mean squared error between the output layer's predicted values ​​and the actual number of production cycles is used as the loss function. The network weights and biases of the multilayer perceptron network are updated using the error backpropagation algorithm. The Adam or SGD optimizer is used to iteratively fit until the loss function converges, i.e., the validation set loss function no longer decreases or the decrease is lower than the preset loss function convergence threshold. The fitted multilayer perceptron network is used as the accuracy decay trend model, and the network weights and structure are saved for online inference.

[0110] The preset loss function convergence threshold is set within the range of 0.1% to 0.5% of the loss function decrease. The smaller the value, the more sufficient the convergence, but the longer the training time. Since the application scenario of this step is a precision-sensitive scenario, the loss function convergence threshold can be set to 0.1%.

[0111] For example, a batch contains 64 samples. The irreversible wear drift rate and gap value of the input samples are passed through the network for forward propagation, outputting 64 predicted remaining cycle numbers. Three exemplary label values ​​are 45000, 120000, and 18000. The squared differences between the 64 predicted values ​​and the label values ​​are calculated, and the mean squared error of the current batch is obtained as the loss function, serving as the optimization objective for backpropagation.

[0112] Furthermore, the initial learning rate was set to 0.001, the batch size to 64, and the maximum number of iterations to 500. During training, the fitted set loss decreased from approximately 1.2 × 10⁻⁶ initially. 9 The loss function gradually decreased, reaching 8.41 × 10⁻⁶ for the validation set in the 155th period. 7 The loss function for the 165th period remains 8.41 × 10⁻⁶. 7 The percentage decrease over the last 10 cycles is (8.41 - 8.41) / 8.41 = 0%, and the decrease has been below the threshold of 0.1% for 10 consecutive cycles, triggering the convergence condition. Training automatically stops at the 165th cycle. The network weights corresponding to the 165th cycle are saved as the final accuracy decay trend model.

[0113] After obtaining the accuracy decay trend model after training, inputting the irreversible wear drift rate and the current clearance estimate into the accuracy decay trend model will yield the estimated remaining number of operating cycles before the motion axis positioning accuracy drifts to the preset maintenance threshold. For example, inputting the irreversible wear drift rate of 0.000162 mm / h under the current working condition and the current clearance estimate of 0.2 mm into the accuracy decay trend model will result in an output of 38,200 estimated remaining operating cycles.

[0114] It should be noted that the above values ​​are for illustrative purposes only and do not constitute a limitation on the present invention.

[0115] In summary, this step constructs an accuracy decay trend model, which learns the nonlinear coupling relationship between wear rate and clearance state through historical multi-cycle data. During online prediction, the static input is dynamically replaced by the current measured clearance value, enabling the remaining life prediction to adapt to the wear acceleration effect caused by the deterioration of the transmission chain clearance. The output results are closer to the actual degradation trajectory than fixed-rate extrapolation or univariate regression, providing a reliable basis for accurate decision-making in the maintenance window.

[0116] S500: Combine the steady-state deviation sum to correct the expected remaining running cycle number to obtain the corrected remaining running cycle number. When the corrected remaining running cycle number is lower than the maintenance intervention window, generate a predictive maintenance instruction containing the associated axis number.

[0117] This step uses the steady-state deviation sum to make a secondary correction to the estimated remaining operating cycle number. It comprehensively considers the impact of the absolute magnitude of the current deviation on the remaining lifespan. When the corrected remaining cycle number is lower than the preset maintenance intervention window, it automatically generates a predictive maintenance instruction with shaft number. This ensures that maintenance decisions take into account both the degradation trend and the current deviation status, avoiding missed alarms caused by the deviation approaching the tolerance boundary but the trend prediction not yet triggering an alarm.

[0118] Step S500 in the method provided by the present invention includes: From the deviation probability density distribution obtained by kernel density estimation at the current moment, extract the upper and lower bound deviation values ​​of the steady-state deviation confidence interval, and calculate the width of the confidence interval; The confidence interval width is mapped in an S-shape to obtain a confidence correction coefficient with a value between zero and one, wherein the S-shape mapping causes the confidence correction coefficient to decrease monotonically as the confidence interval width increases; The estimated remaining number of running cycles is adjusted based on the confidence correction coefficient to obtain the corrected remaining number of running cycles; Based on the ratio of the preset maintenance threshold to the irreversible wear drift rate, and combined with the preset number of film production cycles, a maintenance intervention window is obtained; The remaining number of corrected running cycles is compared with the maintenance intervention window. When the remaining number of corrected running cycles is lower than the maintenance intervention window, a predictive maintenance instruction containing the associated axis number is generated.

[0119] In this step, the upper and lower bound deviation values ​​of the steady-state deviation confidence interval are first extracted from the deviation probability density distribution obtained by kernel density estimation at the current time, and the difference between the two is calculated as the width of the confidence interval.

[0120] For example, after kernel density estimation, the deviation probability density distribution curve of the current step test shows that the global maximum value is located at 0.0483 mm, which is the steady-state net deviation, and the half-peak positions are 0.044 mm and 0.053 mm, respectively. Therefore, the lower limit deviation value is 0.044 mm, the upper limit deviation value is 0.053 mm, and the confidence interval width is 0.053 - 0.044 = 0.009 mm.

[0121] Furthermore, the confidence interval width is processed using a sigmoid mapping function to obtain a confidence correction coefficient with a value between zero and one. The sigmoid mapping function can be a Logistic function. By presetting the steepness of the descent and the confidence interval width threshold, the confidence correction coefficient in the sigmoid mapping function monotonically decreases as the confidence interval width increases. The default steepness of the descent is 0.8, and the default confidence interval width threshold is 0.01 mm, meaning that when the confidence interval width is 0.01 mm, the confidence correction coefficient is 0.5. The smaller the confidence interval width, the closer the correction coefficient is to 1, and the less severe the loss on the remaining iterations; the larger the confidence interval width, the closer the correction coefficient is to 0, and the more severe the loss. In other words, the sigmoid mapping causes the confidence correction coefficient to decrease monotonically as the confidence interval width increases.

[0122] For example, with a confidence interval width of 0.009 mm, substituting a descent steepness of 0.8 and a confidence interval width threshold of 0.01, the calculated confidence correction coefficient is 0.5002, indicating that the dispersion of the current bias distribution is at a moderate level, imposing a loss of approximately 50% on the prediction result.

[0123] Furthermore, the estimated remaining number of runtimes is adjusted based on the confidence correction coefficient to obtain the corrected remaining number of runtimes. The correction method is direct multiplication, i.e., corrected remaining number of runtimes = estimated remaining number of runtimes × confidence correction coefficient. For example, if the estimated remaining number of runtimes output in the previous step is 38200 and the confidence correction coefficient is 0.50, the corrected remaining number of runtimes = 38200 × 0.50 = 19100.

[0124] Furthermore, based on the ratio of a preset maintenance threshold to the irreversible wear drift rate, and in conjunction with a preset number of wafer production cycles, a maintenance intervention window is obtained. The maintenance intervention window represents the buffer margin from the current moment until maintenance is required, expressed in wafer production cycles. The maintenance intervention window is the larger of the preset number of wafer production cycles, the ratio of the preset maintenance threshold to the irreversible wear drift rate, and so on.

[0125] The irreversible wear drift rate needs to be converted into degradation per cycle. For example, if the average cycle time is 3.2 minutes, the degradation per cycle is calculated to be 8.5 × 10⁻⁶ mm / h from 0.000162 mm / h. -6 mm / cycle.

[0126] For example, the preset number of film production cycles is 500, the preset maintenance threshold is 0.1mm, and the irreversible wear drift rate is converted to a degradation amount of 8.5 × 10⁻⁶ per cycle. -6 mm / cycle, then the preset maintenance threshold / irreversible wear drift rate = 0.1 / (8.5×10) -6 )≈11765 cycles. 500<11765, take the larger value of 11765 cycles as the maintenance intervention window.

[0127] Finally, the remaining number of operation cycles is compared with the maintenance intervention window. When the remaining number of operation cycles is lower than the maintenance intervention window, a predictive maintenance instruction containing the associated axis number is generated and pushed to the equipment maintenance terminal.

[0128] For example, if the remaining running cycles are corrected to 19100 and the maintenance intervention window is 11765 cycles, since 19100 > 11765, no maintenance command will be triggered at present, and monitoring will continue. If, in another test, the remaining running cycles are corrected to 4200, which is lower than the maintenance intervention window of 11765, a predictive maintenance command will be generated: "The estimated remaining cycles for the X-axis drivetrain are 4200, which is lower than the maintenance intervention window of 11765. Please arrange maintenance. Associated axis number: X01." After the command is pushed to the maintenance personnel's terminal, the maintenance personnel can arrange a shutdown for maintenance after the film production machine has completed the current batch of film production tasks.

[0129] It should be noted that the above values ​​are for illustrative purposes only and do not constitute a limitation on the present invention.

[0130] In summary, this step enables the remaining lifetime prediction to automatically converge to a more conservative estimate when the deviation distribution diverges, reducing the risk of missed alarms caused by overly optimistic predictions, and ensuring that maintenance warnings are neither less than the detection cycle nor less than the theoretical safety margin derived from the degradation rate, thus achieving a balance between timely warnings and false alarm control.

[0131] In summary, this invention solves the problems of the inability to separate thermal deformation and mechanical wear, poor adaptability of fixed thresholds, and lack of consideration for gap acceleration effect and prediction reliability in the prediction of remaining life in traditional methods. It realizes online accurate prediction and timely early warning of the precision degradation of the motion shaft transmission chain of the thin-layer cell preparation machine.

[0132] Example 2, as Figure 3 As shown, the present invention provides a predictive maintenance system for a thin-layer cell slide preparation machine that incorporates multi-source data, the system comprising: The multi-source signal synchronous acquisition module 11 is used to respond to the target film-making machine completing the preset number of film-making cycles, and to acquire the real-time position sequence fed back by the servo motor encoder and the trigger delay sequence of the motion axis positioning photoelectric sensor at the corresponding moment.

[0133] Among them, in response to the target film-making machine completing a preset number of film-making cycles, the real-time position sequence fed back by the servo motor encoder and the trigger delay sequence of the motion axis positioning photoelectric sensor at the corresponding moment are obtained, including: Read the film production cycle counter from the target film production machine control software to obtain the current cumulative film production cycle count; When the cumulative number of film production cycles reaches the preset number of film production cycles, a step positioning test command is injected into the motion control device to drive the target motion axis to reciprocate through the reference position calibrated by the photoelectric sensor at a constant speed. During the execution of the step positioning test command, the real-time position sequence fed back by the encoder is recorded synchronously with the internal sampling period of the servo driver, and the command position sequence is recorded accordingly. A positioning deviation sequence is generated based on the difference between the real-time position sequence and the corresponding time-instructed position sequence; The delay values ​​from the issuance of the command to the sensor level flip are collected in each step positioning test, and arranged in the order of collection to form a trigger delay sequence; The positioning deviation sequence and the trigger delay sequence collected within the same test cycle are aligned by timestamp. After removing the transient data segment at the moment of servo motor commutation, the aligned real-time position sequence and trigger delay sequence are output.

[0134] The transmission chain clearance quantification and evaluation module 12 is used to calculate the positioning deviation sequence based on the difference between the real-time position sequence and the command position, extract the steady-state deviation of the positioning deviation sequence using kernel density estimation, and determine the current clearance estimate of the motion shaft transmission chain based on the over-limit ratio of the trigger delay sequence.

[0135] Specifically, based on the difference between the real-time position sequence and the commanded position, a positioning deviation sequence is calculated, and the steady-state deviation sum of the positioning deviation sequence is extracted using kernel density estimation, including: The encoder feedback position is extracted point by point from the aligned real-time position sequence, and the difference is calculated with the command position in the command position sequence at the same timestamp to obtain the positioning deviation sequence containing the forward and reverse travel. The remaining deviation sequence is obtained by removing the deviation values ​​within a fixed range on both sides of the sampling point corresponding to the instant of servo motor commutation in the positioning deviation sequence; According to the constant velocity direction set by the step positioning test command, the remaining deviation value is separated into a positive deviation subsequence and a negative deviation subsequence; The ratio of the average delay of the trigger delay sequence in the reverse travel segment to the average delay in the forward travel segment is used as the reverse gap feature value, and then the reverse gap influence weight is obtained after S-shaped mapping. The positive deviation subsequence and the negative deviation subsequence are weighted and fused according to the negative gap influence weight to generate a comprehensive deviation sequence. Using the comprehensive deviation sequence as input, the deviation probability density distribution is obtained by kernel density estimation, and the deviation value corresponding to the maximum probability density value is extracted as the steady-state deviation sum.

[0136] Specifically, using the comprehensive deviation sequence as input, kernel density estimation is employed to obtain the deviation probability density distribution, and the deviation value corresponding to the maximum probability density value is extracted as the steady-state deviation aggregate, including: Sort all deviation values ​​in the comprehensive deviation sequence according to their numerical values, calculate the distance between adjacent deviation values ​​after sorting, and take the median of the distance as the bandwidth of the kernel density estimation. Using the bandwidth as a smoothing parameter, a kernel function is placed for each deviation value in the comprehensive deviation sequence, and all kernel functions are superimposed on the deviation value axis to obtain the deviation probability density distribution curve. Identify the global maximum point from the deviation probability density distribution curve, and use the deviation value corresponding to the global maximum point as the initial steady-state deviation; Centered on the initial steady-state deviation, search to the left and right for positions where the probability density drops to half of the global maximum value, thus forming the steady-state deviation confidence interval; The weighted average of all deviation values ​​within the steady-state deviation confidence interval is extracted as the steady-state deviation sum, and the weighting coefficient is taken as the probability density value corresponding to each deviation value.

[0137] Based on the over-limit ratio of the trigger delay sequence, determine the current backlash estimate of the motion shaft drive chain, including: Arrange the delay values ​​in the trigger delay sequence in ascending order, and take the delay value located at the lower quartile after sorting as the baseline response delay; The median absolute deviation of the trigger delay sequence is calculated and superimposed with the reference response delay to form an over-limit discrimination threshold; The number of sampling points whose delay values ​​exceed the over-limit discrimination threshold is counted, and the over-limit ratio is obtained by combining the total length of the trigger delay sequence; Using the aforementioned over-limit ratio as the quantile basis, the delay value corresponding to the high quantile is extracted from the delay values ​​arranged in ascending order as the gap feature delay; The effective gap delay is obtained by comparing the difference between the gap characteristic delay and the reference response delay. Then, the current gap estimate of the motion shaft transmission chain is obtained by combining the constant speed value set in the step positioning test command.

[0138] The temperature-wear effect decoupling module 13 is used to obtain the internal multi-point temperature sequence during the operation of the target film-making machine, decouple the temperature-induced reversible drift component based on the hysteresis relationship between the temperature gradient and the steady-state deviation, and calculate the irreversible wear drift rate by combining the current gap estimate.

[0139] Specifically, the process involves acquiring the internal multi-point temperature sequence during the operation of the target film-making machine, decoupling the temperature-induced reversible drift component based on the hysteresis relationship between the temperature gradient and the steady-state deviation, and calculating the irreversible wear drift rate by combining the current gap estimate. Temperature values ​​are collected from multiple locations inside the target film-making machine by temperature sensors, and the average values ​​of multiple points are taken after aligning them according to the sampling time to form the internal multi-point temperature sequence. The internal multi-point temperature sequence is differentially analyzed to obtain a temperature gradient sequence. At the same time, the corresponding steady-state deviation summation is extracted from each step positioning test to form a deviation drift sequence. The temperature gradient sequence and the deviation drift sequence are subjected to hysteresis cross-correlation analysis along the time axis, and the number of hysteresis sampling points at the point where the cross-correlation coefficient is the maximum is taken as the hysteresis time of temperature-induced reversible drift. The deviation drift sequence and temperature gradient sequence are time-shifted and aligned using the lag time to establish the correspondence between the temperature gradient and the reversible drift component. The temperature-induced reversible drift component is then extracted from the current steady-state deviation sum to obtain the irreversible deviation component. By combining the cumulative running time, the irreversible deviation component is converted into an average drift rate per unit time, and the current gap estimate is incorporated as a correction factor to calculate the irreversible wear drift rate.

[0140] Specifically, by removing the temperature-induced reversible drift component from the current steady-state deviation, the irreversible deviation component is obtained, including: Extract the temperature gradient sequence segment from the temperature response cycle preceding the current moment, and simultaneously extract the deviation drift sequence segment after time shift alignment by the lag time. The temperature gradient sequence segment is used as the independent variable, and the time-shifted and aligned deviation drift sequence segment is used as the dependent variable. The least squares method is used to fit the linear mapping coefficient between the temperature gradient and the deviation drift. By combining the instantaneous temperature gradient of the internal multi-point temperature sequence at the current moment with the linear mapping coefficient, the estimated value of the temperature-induced reversible drift component at the current moment is obtained. Based on the steady-state deviation sum extracted by kernel density estimation at the current moment, and combined with the estimated value of the reversible drift component, the irreversible deviation component is calculated.

[0141] The precision decay trend extrapolation module 14 is used to input the irreversible wear drift rate into the pre-built precision decay trend model, and at the same time use the current gap estimate as the state update amount to extrapolate to obtain the expected number of remaining running cycles when the positioning accuracy of the motion axis drifts to the preset maintenance threshold.

[0142] The pre-construction steps of the accuracy decay trend model include: Collect historical operating data of the same model of thin-layer cell slide preparation machine in multiple complete maintenance cycles. For each step positioning test, record the corresponding irreversible wear drift rate and the current gap estimate, as well as the actual number of slide preparation cycles from the corresponding test time as the prediction starting point until the positioning accuracy drift reaches the preset maintenance threshold. The irreversible wear drift rate corresponding to each step positioning test and the current gap estimate are used as input samples, and the corresponding actual number of film production cycles is used as output labels to construct a sample pair set. A multilayer perceptron network is constructed with irreversible wear drift rate and current gap estimate as two inputs and the expected number of remaining running cycles as the output. The input layer has 2 nodes, the hidden layer uses a nonlinear activation function, the output layer has 1 node, and the output value represents the predicted number of remaining running cycles. The sample set is divided into a fitting set and a validation set, and the mean square error between the output layer's predicted value and the actual number of film production cycles is used as the loss function. The network weights of the multilayer perceptron network are updated using backpropagation of error. The network is iteratively fitted until the loss function converges, and the fitted multilayer perceptron network is used as the accuracy decay trend model.

[0143] The integrated correction and maintenance decision module 15 is used to correct the expected remaining number of running cycles by combining the steady-state deviation total to obtain the corrected remaining number of running cycles. When the corrected remaining number of running cycles is lower than the maintenance intervention window, a predictive maintenance instruction containing the associated axis number is generated.

[0144] Specifically, the estimated remaining operating cycle number is corrected by combining the steady-state deviation sum to obtain the corrected remaining operating cycle number. When the corrected remaining operating cycle number is lower than the maintenance intervention window, a predictive maintenance instruction containing the associated axis number is generated, including: From the deviation probability density distribution obtained by kernel density estimation at the current moment, extract the upper and lower bound deviation values ​​of the steady-state deviation confidence interval, and calculate the width of the confidence interval; The confidence interval width is mapped in an S-shape to obtain a confidence correction coefficient with a value between zero and one, wherein the S-shape mapping causes the confidence correction coefficient to decrease monotonically as the confidence interval width increases; The estimated remaining number of running cycles is adjusted based on the confidence correction coefficient to obtain the corrected remaining number of running cycles; Based on the ratio of the preset maintenance threshold to the irreversible wear drift rate, and combined with the preset number of film production cycles, a maintenance intervention window is obtained; The remaining number of corrected running cycles is compared with the maintenance intervention window. When the remaining number of corrected running cycles is lower than the maintenance intervention window, a predictive maintenance instruction containing the associated axis number is generated.

[0145] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

[0146] This specification and accompanying drawings are merely illustrative examples of the invention and are intended to cover any and all modifications, variations, combinations, or equivalents within the scope of the invention. Clearly, those skilled in the art can make various alterations and modifications to the invention without departing from its scope. Therefore, if such modifications and modifications fall within the scope of the invention and its equivalents, the invention is intended to include these modifications and modifications.

Claims

1. A predictive maintenance method for thin-layer cell slide preparation machines that combines multi-source data, characterized in that, include: In response to the target film-making machine completing the preset number of film-making cycles, the real-time position sequence fed back by the servo motor encoder and the trigger delay sequence of the motion axis positioning photoelectric sensor at the corresponding moment are obtained; Based on the difference between the real-time position sequence and the command position, the positioning deviation sequence is calculated, and the steady-state deviation of the positioning deviation sequence is extracted using kernel density estimation. Based on the over-limit ratio of the trigger delay sequence, the current clearance estimate of the motion shaft transmission chain is determined. The internal multi-point temperature sequence of the target film-making machine during operation is obtained. Based on the hysteresis relationship between the temperature gradient and the steady-state deviation, the temperature-induced reversible drift component is decoupled. Combined with the current gap estimate, the irreversible wear drift rate is calculated. The irreversible wear drift rate is input into a pre-built accuracy decay trend model, and the current clearance estimate is used as the state update value to extrapolate and obtain the estimated remaining number of running cycles when the motion axis positioning accuracy drifts to the preset maintenance threshold. The estimated remaining operating cycle number is corrected by combining the steady-state deviation, and the corrected remaining operating cycle number is obtained. When the corrected remaining operating cycle number is lower than the maintenance intervention window, a predictive maintenance instruction containing the associated axis number is generated.

2. The predictive maintenance method for a thin-layer cell slide preparation machine combining multi-source data according to claim 1, characterized in that, In response to the target film-making machine completing a preset number of film-making cycles, the system acquires the real-time position sequence fed back by the servo motor encoder and the trigger delay sequence of the motion axis positioning photoelectric sensor at the corresponding moment, including: Read the film production cycle counter from the target film production machine control software to obtain the current cumulative film production cycle count; When the cumulative number of film production cycles reaches the preset number of film production cycles, a step positioning test command is injected into the motion control device to drive the target motion axis to reciprocate through the reference position calibrated by the photoelectric sensor at a constant speed. During the execution of the step positioning test command, the real-time position sequence fed back by the encoder is recorded synchronously with the internal sampling period of the servo driver, and the command position sequence is recorded accordingly. A positioning deviation sequence is generated based on the difference between the real-time position sequence and the corresponding time-instructed position sequence; The delay values ​​from the issuance of the command to the sensor level flip are collected in each step positioning test, and arranged in the order of collection to form a trigger delay sequence; The positioning deviation sequence and the trigger delay sequence collected within the same test cycle are aligned by timestamp. After removing the transient data segment at the moment of servo motor commutation, the aligned real-time position sequence and trigger delay sequence are output.

3. The predictive maintenance method for a thin-layer cell slide preparation machine combining multi-source data according to claim 1, characterized in that, Based on the difference between the real-time position sequence and the commanded position, a positioning deviation sequence is calculated, and the steady-state deviation sum of the positioning deviation sequence is extracted using kernel density estimation, including: The encoder feedback position is extracted point by point from the aligned real-time position sequence, and the difference is calculated with the command position in the command position sequence at the same timestamp to obtain the positioning deviation sequence containing the forward and reverse travel. The remaining deviation sequence is obtained by removing the deviation values ​​within a fixed range on both sides of the sampling point corresponding to the instant of servo motor commutation in the positioning deviation sequence; According to the constant velocity direction set by the step positioning test command, the remaining deviation value is separated into a positive deviation subsequence and a negative deviation subsequence; The ratio of the average delay of the trigger delay sequence in the reverse travel segment to the average delay in the forward travel segment is used as the reverse gap feature value, and then the reverse gap influence weight is obtained after S-shaped mapping. The positive deviation subsequence and the negative deviation subsequence are weighted and fused according to the negative gap influence weight to generate a comprehensive deviation sequence. Using the comprehensive deviation sequence as input, the deviation probability density distribution is obtained by kernel density estimation, and the deviation value corresponding to the maximum probability density value is extracted as the steady-state deviation sum.

4. The predictive maintenance method for a thin-layer cell slide preparation machine combining multi-source data according to claim 3, characterized in that, Using the comprehensive deviation sequence as input, kernel density estimation is used to obtain the deviation probability density distribution. The deviation value corresponding to the maximum probability density value is extracted as the steady-state deviation sum, including: Sort all deviation values ​​in the comprehensive deviation sequence according to their numerical values, calculate the distance between adjacent deviation values ​​after sorting, and take the median of the distance as the bandwidth of the kernel density estimation. Using the bandwidth as a smoothing parameter, a kernel function is placed for each deviation value in the comprehensive deviation sequence, and all kernel functions are superimposed on the deviation value axis to obtain the deviation probability density distribution curve. Identify the global maximum point from the deviation probability density distribution curve, and use the deviation value corresponding to the global maximum point as the initial steady-state deviation; Centered on the initial steady-state deviation, search to the left and right for positions where the probability density drops to half of the global maximum value, thus forming the steady-state deviation confidence interval; The weighted average of all deviation values ​​within the steady-state deviation confidence interval is extracted as the steady-state deviation sum, and the weighting coefficient is taken as the probability density value corresponding to each deviation value.

5. The predictive maintenance method for a thin-layer cell slide preparation machine combining multi-source data according to claim 1, characterized in that, Based on the over-limit ratio of the trigger delay sequence, determine the current backlash estimate of the motion shaft drive chain, including: Arrange the delay values ​​in the trigger delay sequence in ascending order, and take the delay value located at the lower quartile after sorting as the baseline response delay; The median absolute deviation of the trigger delay sequence is calculated and superimposed with the reference response delay to form an over-limit discrimination threshold; The number of sampling points whose delay values ​​exceed the over-limit discrimination threshold is counted, and the over-limit ratio is obtained by combining the total length of the trigger delay sequence; Using the aforementioned over-limit ratio as the quantile basis, the delay value corresponding to the high quantile is extracted from the delay values ​​arranged in ascending order as the gap feature delay; The effective gap delay is obtained by comparing the difference between the gap characteristic delay and the reference response delay. Then, the current gap estimate of the motion shaft transmission chain is obtained by combining the constant speed value set in the step positioning test command.

6. The predictive maintenance method for a thin-layer cell slide preparation machine combining multi-source data according to claim 1, characterized in that, The internal multi-point temperature sequence of the target film-making machine during operation is obtained. Based on the hysteresis relationship between the temperature gradient and the steady-state deviation, the temperature-induced reversible drift component is decoupled. Combined with the current gap estimate, the irreversible wear drift rate is calculated, including: Temperature values ​​are collected from multiple locations inside the target film-making machine by temperature sensors, and the average values ​​of multiple points are taken after aligning them according to the sampling time to form the internal multi-point temperature sequence. The internal multi-point temperature sequence is differentially analyzed to obtain a temperature gradient sequence. At the same time, the corresponding steady-state deviation summation is extracted from each step positioning test to form a deviation drift sequence. The temperature gradient sequence and the deviation drift sequence are subjected to hysteresis cross-correlation analysis along the time axis, and the number of hysteresis sampling points at the point where the cross-correlation coefficient is the maximum is taken as the hysteresis time of temperature-induced reversible drift. The deviation drift sequence and temperature gradient sequence are time-shifted and aligned using the lag time to establish the correspondence between the temperature gradient and the reversible drift component. The temperature-induced reversible drift component is then extracted from the current steady-state deviation sum to obtain the irreversible deviation component. By combining the cumulative running time, the irreversible deviation component is converted into an average drift rate per unit time, and the current gap estimate is incorporated as a correction factor to calculate the irreversible wear drift rate.

7. The predictive maintenance method for a thin-layer cell slide preparation machine combining multi-source data according to claim 6, characterized in that, By stripping the temperature-induced reversible drift component from the current steady-state deviation, we obtain the irreversible deviation components, including: Extract the temperature gradient sequence segment from the temperature response cycle preceding the current moment, and simultaneously extract the deviation drift sequence segment after time shift alignment by the lag time. The temperature gradient sequence segment is used as the independent variable, and the time-shifted and aligned deviation drift sequence segment is used as the dependent variable. The least squares method is used to fit the linear mapping coefficient between the temperature gradient and the deviation drift. By combining the instantaneous temperature gradient of the internal multi-point temperature sequence at the current moment with the linear mapping coefficient, the estimated value of the temperature-induced reversible drift component at the current moment is obtained. Based on the steady-state deviation sum extracted by kernel density estimation at the current moment, and combined with the estimated value of the reversible drift component, the irreversible deviation component is calculated.

8. The predictive maintenance method for a thin-layer cell slide preparation machine combining multi-source data according to claim 1, characterized in that, The pre-construction steps of the accuracy decay trend model include: Collect historical operating data of the same model of thin-layer cell slide preparation machine in multiple complete maintenance cycles. For each step positioning test, record the corresponding irreversible wear drift rate and the current gap estimate, as well as the actual number of slide preparation cycles from the corresponding test time as the prediction starting point until the positioning accuracy drift reaches the preset maintenance threshold. The irreversible wear drift rate corresponding to each step positioning test and the current gap estimate are used as input samples, and the corresponding actual number of film production cycles is used as output labels to construct a sample pair set. A multilayer perceptron network is constructed with irreversible wear drift rate and current gap estimate as two inputs and the expected number of remaining running cycles as the output. The input layer has 2 nodes, the hidden layer uses a nonlinear activation function, the output layer has 1 node, and the output value represents the predicted number of remaining running cycles. The sample set is divided into a fitting set and a validation set, and the mean square error between the output layer's predicted value and the actual number of film production cycles is used as the loss function. The network weights of the multilayer perceptron network are updated using backpropagation of error. The network is iteratively fitted until the loss function converges, and the fitted multilayer perceptron network is used as the accuracy decay trend model.

9. The predictive maintenance method for a thin-layer cell slide preparation machine combining multi-source data according to claim 1, characterized in that, The estimated remaining operating cycles are corrected by combining the steady-state deviation, resulting in a corrected remaining operating cycles. When the corrected remaining operating cycles are lower than the maintenance intervention window, a predictive maintenance instruction containing the associated axis number is generated, including: From the deviation probability density distribution obtained by kernel density estimation at the current moment, extract the upper and lower bound deviation values ​​of the steady-state deviation confidence interval, and calculate the width of the confidence interval; The confidence interval width is mapped in an S-shape to obtain a confidence correction coefficient with a value between zero and one, wherein the S-shape mapping causes the confidence correction coefficient to decrease monotonically as the confidence interval width increases; The estimated remaining number of running cycles is adjusted based on the confidence correction coefficient to obtain the corrected remaining number of running cycles; Based on the ratio of the preset maintenance threshold to the irreversible wear drift rate, and combined with the preset number of film production cycles, a maintenance intervention window is obtained; The remaining number of corrected running cycles is compared with the maintenance intervention window. When the remaining number of corrected running cycles is lower than the maintenance intervention window, a predictive maintenance instruction containing the associated axis number is generated.

10. A predictive maintenance system for thin-layer cell slide preparation machines that integrates multi-source data, characterized in that: For implementing the predictive maintenance method for a thin-layer cell slide preparation machine combining multi-source data as described in any one of claims 1-9, the system comprises: The multi-source signal synchronous acquisition module is used to respond to the target film-making machine completing the preset number of film-making cycles, and to acquire the real-time position sequence fed back by the servo motor encoder and the trigger delay sequence of the motion axis positioning photoelectric sensor at the corresponding moment; The transmission chain clearance quantification and evaluation module is used to calculate the positioning deviation sequence based on the difference between the real-time position sequence and the command position, extract the steady-state deviation of the positioning deviation sequence using kernel density estimation, and determine the current clearance estimate of the motion shaft transmission chain based on the over-limit ratio of the trigger delay sequence. The temperature-wear effect decoupling module is used to obtain the internal multi-point temperature sequence during the operation of the target film-making machine, and based on the hysteresis relationship between the temperature gradient and the steady-state deviation, decouple the temperature-induced reversible drift component, and calculate the irreversible wear drift rate by combining the current gap estimate. The accuracy decay trend extrapolation module is used to input the irreversible wear drift rate into the pre-built accuracy decay trend model, and at the same time use the current gap estimate as the state update amount to extrapolate to obtain the expected number of remaining running cycles when the motion axis positioning accuracy drifts to the preset maintenance threshold. The integrated correction and maintenance decision module is used to correct the expected remaining number of running cycles by combining the steady-state deviation, and obtain the corrected remaining number of running cycles. When the corrected remaining number of running cycles is lower than the maintenance intervention window, a predictive maintenance instruction containing the associated axis number is generated.

Images