PCB printer fault prediction method and system based on historical maintenance log

Abstract

The application relates to the technical field of intelligent operation and maintenance management and fault prediction, in particular to a PCB (Printed Circuit Board) printer fault prediction method and system based on historical maintenance logs, which comprises the following steps: obtaining the full life cycle maintenance logs of the PCB printer, performing semantic feature mapping on unstructured texts, generating a benchmark semantic score of a maintenance event, constructing a maintenance depth quantification index model, and calculating a maintenance depth quantification index representing the thoroughness of each maintenance; introducing a machine age discount factor which is nonlinearly attenuated with the cumulative running length of the equipment, establishing a dynamic health degree evolution model of non-complete recovery, calculating the predicted health degree of the PCB printer after each maintenance event ends, and performing fault prediction on the PCB printer based on the predicted health degree. The method eliminates the prediction deviation caused by the traditional maintenance equivalent assumption by evaluating the maintenance depth and the machine age constraint, and realizes accurate prediction of the fault risk of old equipment.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of intelligent operation and maintenance management and fault prediction, in particular to a PCB printer fault prediction method and system based on historical maintenance logs. BACKGROUND

[0002] In the field of PCB (Printed Circuit Board) manufacturing, solder mask printers and vacuum hole plugging machines are key equipment that determine the electrical performance of products. In order to ensure production continuity, enterprises have accumulated a large amount of equipment maintenance logs in the long-term operation and maintenance of equipment. These logs not only contain structured data such as fault occurrence timestamp, alarm code, and downtime length, but also contain a large amount of unstructured text written by maintenance engineers, such as "net plate aperture is slightly blocked, has been cleaned and unblocked" or "squeegee cylinder action is sluggish, adjust air pressure valve" and other descriptions.

[0003] The current equipment management system or fault prediction system generally has a logical defect based on an idealized model when using these data, that is, the maintenance equivalence assumption. The existing algorithm usually defaults that as long as the fault alarm is eliminated, the equipment returns to the initial healthy state, that is, the health degree is reset to 100%.

[0004] However, engineering practice shows that the contribution of different depth of maintenance methods to equipment life is greatly different. For example, simple cleaning and unblocking belongs to shallow maintenance, which can only temporarily alleviate the symptoms, and the equipment health degree may only recover a small proportion; while replacing core components belongs to deep maintenance, and the health degree can be restored to a higher level. In addition, as the running time of the equipment increases, the metal fatigue and precision loss of the basic components are irreversible, and even if the deep maintenance is performed, the recovery effect will be marginal decreasing with the increase of the running time of the equipment.

[0005] Due to the lack of semantic evaluation ability of the maintenance depth in the text log and the neglect of the physical constraint of the age loss, the existing technology often leads to an overestimation of the health degree of the old equipment. The deviation between this data health and the actual operation with illness often leads to sudden downtime of the equipment within the predicted safe period.

[0006] Therefore, there is an urgent need for an intelligent management method that can accurately evaluate the maintenance effect and accurately predict the fault of the PCB printer. SUMMARY

[0007] To solve the problem that the existing technology cannot accurately evaluate the difference in maintenance depth based on the maintenance equivalence assumption when using equipment maintenance logs for fault prediction, ignores the irreversible physical loss of basic components due to the increase of the age in the whole life cycle of the equipment, and leads to an overestimation of the health degree of the old equipment and inaccurate fault prediction, the present application proposes a PCB printer fault prediction method and system based on historical maintenance logs.

[0008] In a first aspect, the present application provides a PCB printer failure prediction method based on historical maintenance logs, comprising:

[0009] Obtaining the full life cycle historical maintenance logs of the PCB printer, extracting keywords using the TF-IDF algorithm to match the pre-constructed maintenance semantic knowledge base, and determining the benchmark semantic score of each maintenance event;

[0010] Extracting the spare parts consumption amount and maintenance time of each maintenance event, combining the benchmark semantic score of the maintenance event to calculate the maintenance depth index of the maintenance event to reflect the recovery degree of the device performance of the maintenance event;

[0011] Obtaining the cumulative running total time and the remaining health degree of the PCB printer at the start time of each maintenance event, and the downtime caused by the maintenance event; based on the cumulative running total time and the design life length of the PCB printer, a age loss factor is constructed to reflect the irreversible loss of the device as the service time increases;

[0012] Combining the age loss factor, the remaining health degree, the downtime and the maintenance depth index, a dynamic health degree evolution model is established to calculate the predicted health degree of the PCB printer after the maintenance event, and the PCB printer is failure predicted based on the predicted health degree.

[0013] This technical solution constructs a closed-loop prediction system that converts unstructured operation and maintenance text into quantifiable decision basis, generates a benchmark semantic score by integrating TF-IDF algorithm and semantic feature mapping technology, effectively associates natural language with industrial control parameters, and on this basis, the method combines multiple modal data to determine the maintenance depth quantitative index to reflect the comprehensive recovery ability of the PCB printer after each maintenance event, and cooperates with the age loss factor representing the irreversible loss of the physical entity. The system introduces a Gaussian decay function as a full life cycle constraint to simulate the metal fatigue characteristics of the basic components of the device accelerating deterioration with increasing cumulative running time, not only eliminates the noise interference of a single data source, but also establishes a dynamic health degree evolution model, uses the age factor as a recovery gain inhibition coefficient, objectively quantifies the actual recovery contribution of the maintenance history to the device performance, solves the problem that the traditional model based on the maintenance equivalent assumption is difficult to distinguish between shallow maintenance and deep repair, and improves the accuracy of failure prediction.

[0014] Preferably, the maintenance semantic knowledge base is constructed based on the following way:

[0015] The full life cycle history maintenance log is segmented to extract the maintenance action predicate; a three-level recovery force predicate mapping model is established to map the maintenance action predicate to the corresponding weight level: the first-level recovery force predicate represents the replacement of irreversible physical entities, corresponds to spare part replacement behavior, and is assigned a first-level weight; the second-level recovery force predicate represents the reverse calibration of physical parameters, corresponds to accuracy adjustment or fastening behavior, and is assigned a second-level weight; the third-level recovery force predicate represents software logic reset or surface state cleaning, corresponds to restart or dust removal behavior, and is assigned a third-level weight; the maintenance action predicate and its corresponding weight are stored in a key-value database to form the maintenance semantic knowledge base; wherein the first-level weight, the second-level weight and the third-level weight are in a strictly decreasing relationship.

[0016] Preferably, the baseline semantic score of each maintenance event is generated based on the following way: the baseline semantic score of each maintenance event is generated based on the following way:

[0017] The maintenance log of each maintenance event is segmented and stop words are removed to obtain a word sequence; the frequency of each word in the word sequence in the maintenance log and the inverse document frequency of the word in the full life cycle history maintenance log are counted, and a TF-IDF algorithm is performed based on the frequency and the inverse document frequency to extract all core keywords of the word sequence, each core keyword is matched with the pre-constructed maintenance semantic knowledge base to obtain the recovery force predicate level corresponding to the core keyword, the weight corresponding to the recovery force predicate level is taken as the weight of the core keyword, and the maximum value of the weights of all core keywords is taken as the baseline semantic score of the maintenance event.

[0018] Preferably, the maintenance depth index is determined based on the following relationship:

[0019]

[0020] wherein, is the maintenance depth index of the i-th maintenance event, is a logarithmic function, is a natural constant, is the total amount of spare parts consumed by the i-th maintenance event, is a preset basic value constant, is the baseline semantic score of the i-th maintenance event, is the actual time consumption of the i-th maintenance event, is the average maintenance time consumption of all maintenance events with the same fault type as the i-th maintenance event in the full life cycle history maintenance log, ​​​​​This is a preset zero-smoothing constant.

[0021] This technical solution achieves precise quantification of the physical recovery capability of a single maintenance event by constructing a multivariate operator that includes spare parts value, textual semantics, and maintenance time. It introduces a natural logarithm function to handle spare parts costs, conforming to the diminishing marginal effect of economic investment on equipment performance improvement, thus smoothing the nonlinear relationship between high investment and actual results. Simultaneously, combined with relative maintenance time indicators, it effectively reflects the level of detail in the fault diagnosis process. The establishment of this mathematical model objectively evaluates maintenance quality and can keenly capture the differences between component replacement and routine cleaning at the physical level, providing an accurate basis for the system to identify inefficient maintenance behaviors and assess the true extent of equipment health recovery.

[0022] Preferably, the age-related depreciation factor is constructed based on the following method:

[0023]

[0024] in, The depreciation factor is the age of the machine. It is a natural constant. For PCB printing machines in the first The cumulative total operating time at the start time of the maintenance event. This refers to the design lifespan of a PCB printer.

[0025] This technical solution utilizes a full life-cycle constraint model based on the Gaussian decay function to evaluate the nonlinear physical degradation trend that mechanical equipment inevitably exhibits as its cumulative operating time increases. The age-related depreciation factor simulates the accelerated deterioration of metal fatigue and precision loss in the basic components of the equipment as they approach their design life. It corrects the deviation of the idealized recovery assumption from the perspective of physical laws. The introduction of the age-related depreciation factor suppresses the upper limit of the health recovery of old equipment after maintenance, ensuring that the prediction model can truly reflect the objective fact that even after in-depth maintenance, old machines cannot be fully restored.

[0026] Preferably, the dynamic health evolution model for calculating the predicted health of the PCB printer after the end of the maintenance event is based on the following relationship:

[0027]

[0028] in, It is the first Predicted health status after the end of the maintenance event. For the first Remaining health at the start of the next maintenance. The preset natural decay rate, For the first End time of this maintenance incident is the start time of the first maintenance event, is a preset recovery gain coefficient, is the maintenance depth index of the first maintenance event, is an age loss factor.

[0029] The dynamic health evolution model constructed by this technical solution is based on the incomplete recovery theory. By coupling the natural loss term and the recovery gain term modulated by the age, an evolution equation is constructed to describe the dynamic game of the device performance with the passage of time and the maintenance intervention. The model discards the traditional sawtooth-shaped complete recovery assumption, generates a spiral descending curve consistent with the principle of thermodynamic entropy increase, and clearly depicts the gradual trajectory of the device from healthy operation to performance degradation to the brink of failure. This precise situational awareness reveals the true state evolution of the device throughout its life cycle, enabling the maintenance strategy to change from passive response to active condition maintenance based on accurate health threshold, maximizing the effective service life of the device and reducing the total life cycle operation and maintenance cost under the premise of ensuring production safety.

[0030] Preferably, the PCB printer is predicted for failure based on the predicted health, comprising:

[0031] A failure threshold is preset. Taking any moment of the PCB printer operation as the current moment, taking the predicted health after the end of the nearest previous maintenance event as the reference, a curve of the predicted health naturally decaying with time is constructed based on the age loss factor. The moment corresponding to the first time when the curve is less than the failure threshold is determined as the predicted failure moment, and a maintenance decision is triggered according to the predicted failure moment, so as to realize the failure prediction of the PCB printer.

[0032] This technical solution realizes the leap from static state evaluation to dynamic risk prediction by constructing a health time-varying decay trajectory based on physical mechanism. The system takes the recovery state after the last maintenance as the energy reference, uses the age loss factor to nonlinearly modulate the natural decay curve, and accurately simulates the differentiated performance degradation rate of the device at different life cycle stages. By calculating the time domain intersection of the decay trajectory and the preset failure threshold, the algorithm can lock the determined predicted failure moment. This mechanism overcomes the defect that the traditional linear prediction model cannot adapt to the accelerated aging characteristics of old devices. It converts the health index into an executable time window decision, so as to avoid the risk of sudden shutdown while accurately arranging the maintenance intervention time according to the remaining service life.

[0033] Preferably, the maintenance decision is triggered according to the predicted failure moment, comprising:

[0034] The predicted completion time of the production order of the PCB printer at the current time is obtained, the predicted failure time is compared with the predicted completion time, if the predicted failure time is less than the predicted completion time, a first-level early warning signal is generated, prompting to carry out deep maintenance in the nearest shift interval, and the deep maintenance includes component replacement operation, if the predicted failure time is not less than the predicted completion time, the attenuation slope of the health degree curve is calculated, if the attenuation slope exceeds the preset slope threshold, a second-level early warning signal is generated, prompting to check the lubrication system or the air pressure system of the equipment, and if the attenuation slope does not exceed the preset slope threshold, no early warning signal is generated.

[0035] Preferably, after obtaining the full life cycle historical maintenance log of the PCB printer, the following preprocessing operations are further performed: performing deduplication verification on the full life cycle historical maintenance log; semantic feature mapping: traversing the full life cycle historical maintenance log, extracting the fault description text and the maintenance measure text corresponding to each maintenance event, encapsulating them as a corpus block, and taking them as source data for determining the benchmark semantic score of each maintenance event; cost feature mapping: extracting the spare part consumption list and the spare part unit price data of each maintenance event, and accumulating and summing the spare part amount to map the total spare part amount consumed by each maintenance event; time sequence feature mapping: extracting the fault occurrence timestamp, the maintenance start timestamp and the maintenance end timestamp associated with each maintenance event to calculate the actual time consumption of each maintenance event, and calculating the total running time of the PCB printer at the start time of each maintenance event based on the time axis.

[0036] The application further provides a PCB printer fault prediction system based on historical maintenance logs, which comprises a memory and a processor, the memory stores a computer program, and the processor executes the computer program to realize the steps of the PCB printer fault prediction method based on historical maintenance logs.

[0037] The application has the following effects:

[0038] The application analyzes the historical maintenance log of the PCB printer, fuses natural language processing technology and multi-modal data analysis, constructs a fault prediction system based on maintenance depth quantification and machine age physical constraints, converts the unstructured log into a calculable benchmark semantic score by using the TF-IDF algorithm and the semantic knowledge base, and constructs a maintenance depth quantification index in combination with the logarithmic smoothing characteristics of the spare part cost and the maintenance work hour ratio, so as to accurately distinguish the essential difference of different maintenance depths on the performance recovery of the equipment, solve the problem that the manual record is subjective and difficult to evaluate the maintenance recovery capability, introduce a machine age depreciation factor that presents Gaussian attenuation with the cumulative running time, establish a non-complete recovery dynamic health degree evolution model conforming to the principle of thermodynamic entropy increase, correct the deviation of the predicted health degree of the equipment, simulate the health degree recovery weakness phenomenon of the old equipment caused by irreversible metal fatigue, and improve the accuracy of fault prediction. Attached Figure Description

[0039] Figure 1 This is a flowchart of the present invention;

[0040] Figure 2 This is a comparison chart of the health evolution prediction of the PCB printing machine throughout its entire life cycle in this invention. Detailed Implementation

[0041] The technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings in the embodiments of the present invention.

[0042] Reference Figure 1 The PCB printer fault prediction method based on historical maintenance logs includes the following steps:

[0043] S1: Obtain the full lifecycle historical maintenance logs of the PCB printer.

[0044] In this embodiment, the system connects to the factory's MES (Manufacturing Execution System) and ERP (Enterprise Resource Planning) systems through a data interface (API). Using the PCB printer's equipment ID as an index, the system extracts the PCB printer's full lifecycle historical maintenance log from the date of manufacture. The full lifecycle historical maintenance log includes information such as equipment ID, fault occurrence time, fault alarm code, description of maintenance measures, and spare parts replacement records.

[0045] Before conducting in-depth fault prediction for PCB printers, it's crucial to consider that raw maintenance logs collected from industrial sites are often multi-source, heterogeneous mixed data. This includes both directly calculable structured data, such as spare parts costs and repair time, and a large amount of unstructured text manually entered by maintenance engineers, such as fault descriptions and repair measures. This text data not only contains noise such as spelling errors and non-standard abbreviations, but also suffers from vastly different descriptions of the same repair action due to varying recording habits among different engineers, such as "replace cylinder" versus "replace with a new cylinder." Directly using this raw data for modeling would result in extremely sparse feature space and semantic ambiguity, severely impacting the model's understanding of the true repair intent.

[0046] Therefore, this step further performs rigorous data cleaning and standardization preprocessing on the historical maintenance logs throughout the entire lifecycle. By removing noisy characters, unifying the timestamp format, and aligning unstructured text logs with structured cost and time data, a clean and reliable data foundation is built for subsequent extraction of accurate maintenance semantic features and calculation of maintenance depth index.

[0047] Specifically, it includes:

[0048] Firstly, data cleaning and deduplication, deduplication verification is performed on the full life cycle historical maintenance log to ensure that each independent maintenance event corresponds to a unique set of data input. For example, use regular expressions to remove duplicate work orders, and uniformly correct the time stamp format that may conflict between different systems.

[0049] Then, extract multi-dimensional feature data, semantic feature mapping: traverse the full life cycle historical maintenance log, extract the fault description text and maintenance measure text corresponding to each maintenance event, for example, the fault description text is "net board aperture is blocked, has been cleaned and dredged" or "scraper cylinder action is delayed, adjust air pressure valve", and the maintenance measure text is "replace cylinder sealing ring", which is encapsulated as a corpus block as the source data for determining the baseline semantic score of each maintenance event; cost feature mapping: extract the spare parts consumption list and spare parts unit price data of each maintenance event through the ERP system, and accumulate the sum of spare parts amount, which is mapped to the total amount of spare parts consumed by each maintenance event. This indicator is a hard economic indicator for measuring the physical depth of maintenance, and directly reflects the scale of material input; time sequence feature mapping: extract the fault occurrence timestamp, maintenance start timestamp and maintenance end timestamp associated with each maintenance event, and use these data to calculate the actual time consumption (unit: hour) of each maintenance event, as well as the fault downtime, and the cumulative total running time of the PCB printer at the start time of each maintenance event based on the time axis.

[0050] S2: Construct a maintenance semantic knowledge base, extract the keywords of each maintenance event using TF-IDF algorithm and match them with the maintenance semantic knowledge base to determine the baseline semantic score of each maintenance event.

[0051] After completing the data cleaning, in order to solve the problem that the computer cannot directly understand the essential differences between different maintenance events in terms of physical recovery, engineering practice shows that different maintenance actions have hierarchical contribution to equipment health, replacing core components is deep maintenance which can greatly restore performance, while parameter adjustment or cleaning is only a short-term maintenance, therefore, a maintenance semantic knowledge base containing physical recovery level is established to represent the difference in physical performance recovery ability of different maintenance actions.

[0052] Firstly, construct a maintenance semantic knowledge base, existing technologies often ignore the essential difference in physical recovery ability of different maintenance actions, this embodiment performs word segmentation processing on the full life cycle historical maintenance log to extract maintenance action predicates, and establishes a three-level recovery force predicate mapping model:

[0053] The first-level recovery force predicate: representing the replacement of irreversible physical entities, corresponding to spare part replacement behavior, such as "replacement", "refurbishment", "overhaul", "replacement", etc., with a first-level weight, the highest weight, set to 1, such actions mean the update of physical components, which completely eliminates the cumulative fatigue of the component at the physical level, and theoretically can bring 100% performance recovery at the component level, giving the maximum weight.

[0054] The second-level recovery force predicate: representing the reverse calibration of physical parameters, corresponding to accuracy adjustment or tightening behavior, such as "calibration", "tightening", "lubrication", "parameter adjustment", etc., with a second-level weight, set to 0.6, such actions can restore operating accuracy and eliminate gaps or friction, but do not eliminate microscopic cracks or wear of the component material itself, so the risk of recurrence is moderate, and therefore a medium weight is assigned.

[0055] The third-level recovery force predicate: representing software logic reset or surface state cleaning, corresponding to reboot or dust removal behavior, such as "cleaning", "reboot", "reset", "unclogging", etc., with a third-level weight, set to 0.2, such actions usually only solve surface blockage or software deadlock, without touching physical entity damage, and the fault is prone to recur, so a lower weight is assigned.

[0056] The above standard maintenance action predicates and their corresponding weight values are stored in a key-value pair database to form the maintenance semantic knowledge base.

[0057] Secondly, the reference semantic score of each maintenance event is generated, specifically by using the TF-IDF (Term Frequency-Inverse Document Frequency) algorithm to extract the keywords of the maintenance event, and determining the reference semantic score of the maintenance event, the specific process is as follows:

[0058] For each maintenance event, the maintenance log of the maintenance event is obtained in the full life cycle historical maintenance log, and the maintenance log of the maintenance event is segmented, and stop words are removed, for example, meaningless stop words such as "of", "is", "engineer" are filtered out, to obtain a word sequence.

[0059] The word frequency of each word in the word sequence in the maintenance log and the inverse document frequency of the word in the full life cycle historical maintenance log are counted, and the TF-IDF algorithm is executed based on the word frequency and the inverse document frequency of each word to extract all core keywords in the word sequence, i.e. all core keywords corresponding to the maintenance event. Each core keyword is matched with the pre-constructed maintenance semantic knowledge base. When matching, the words can be converted into vectors first, and then matched based on the principle of maximizing cosine similarity, or fuzzy matching principle is adopted, the recovery force predicate level corresponding to each core keyword in the three-level recovery force predicate mapping model is obtained, and the weight corresponding to the recovery force predicate level is taken as the weight of the core keyword.

[0060] For example, the core keyword extracted from the maintenance log of a certain maintenance event is "wipe". This word does not exist in the maintenance semantic knowledge base, but it has the highest cosine similarity with "clean" in the maintenance semantic knowledge base. In this case, the match is determined to be successful, and the third-level weight value corresponding to "clean" is read as the benchmark semantic score of the core keyword "wipe".

[0061] Considering that a maintenance event usually involves multiple maintenance actions, after obtaining the weights of all the core keywords in the maintenance event, the maximum weight of all the core keywords in the maintenance event is used as the baseline semantic score of the maintenance event.

[0062] For example, in a maintenance event involving cleaning a screen and replacing a scraper, "cleaning the screen" matches "cleaning" in the maintenance semantic knowledge base with a third-level weight of 0.2, and "replacing the scraper" matches "replacing" in the maintenance semantic knowledge base with a first-level weight of 1. Therefore, the baseline semantic score for the core keyword "cleaning the screen" is 0.2, and the baseline semantic score for the core keyword "replacing the scraper" is 1. We select 1 as the baseline semantic score for this maintenance event.

[0063] S3: Extract the spare parts consumption amount and maintenance time for each maintenance event, and calculate the maintenance depth index of the maintenance event based on the baseline semantic score of the maintenance event.

[0064] In order to further objectively reflect the physical depth and economic value of maintenance based on the baseline semantic score, for example, replacing a sealing ring worth only 100 yuan and replacing a precision lead screw worth 10,000 yuan are both semantically replacements, with the same weight of 1 and the same baseline semantic score, but the assessment of the depth of equipment maintenance has different meanings and produces different maintenance effects.

[0065] Therefore, in order to further enhance the differentiation of the same maintenance actions based on the semantic benchmark score and accurately measure the recovery degree of a PCB printer after a maintenance event, this invention constructs a multivariate quantitative model that comprehensively considers the value of spare parts, maintenance time and text semantics to calculate the maintenance depth index of each maintenance event, so as to accurately measure the performance recovery capability of the equipment after the maintenance event.

[0066] In one embodiment, the maintenance depth index It is determined based on the following relation:

[0067]

[0068] in, Indicates the first The maintenance depth index is used to characterize the extent to which maintenance operations restore equipment performance. log function, the total amount of spare parts consumed in the first maintenance event, a preset base value constant, set as 100 yuan, which is the minimum effective maintenance unit cost determined based on the average unit price of common wear-and-tear parts of PCB printers (such as sealing rings, suction nozzles, and solenoid valve coils), the benchmark semantic score of the first maintenance event, the actual time consumed in the first maintenance event, the average maintenance time of all maintenance events with the same fault type as the first maintenance event in the full life cycle historical maintenance log, a preset anti-zero smoothing constant, set as 0.1, which is a very small positive number and is mainly used for mathematical error prevention to prevent division by zero when a new type of fault first occurs, resulting in a natural constant.

[0069] In this relationship, is the economic input item, which uses the natural logarithm function to reflect the diminishing marginal effect of economic input on maintenance effectiveness. When is small, such as when a sealing ring is replaced, the economic input item is , indicating that the basic input has a significant effect. When is large, such as when a lead screw is replaced, the economic input item is , although the amount increases by 100 times, the performance recovery degree evaluation only increases by about 3.5 times, which conforms to the physical reality: the performance improvement brought by high-priced spare parts is not linearly multiplied. The logarithmic function smoothly smooths this order of magnitude difference, avoiding the excessive pull-up of the index by high-priced spare parts. The introduction of the natural constant ensures that when there is no spare part consumption , the value of this item is , which serves as a neutral benchmark for the multiplication factor, ensuring that it will not be zeroed due to no spare part cost.

[0070] In this relationship, is the maintenance operation refinement item, which reflects the refinement of maintenance. If the engineer spends much more time than the average level for troubleshooting and debugging, the ratio is greater than 1, indicating that the maintenance process is meticulous and may solve hidden faults, resulting in a positive gain to the overall index. If the engineer takes replacement measures, , the larger the base under the square root, the more obvious the improvement of the maintenance depth index. The effect of square root is to reduce the excessive sensitivity caused by time fluctuation. For example, even if the time is 4 times more, the contribution to the maintenance depth is only 2 times more, which is in line with the engineering common sense that the maintenance quality is positively correlated with the work time input but the efficiency is limited.

[0071] The relationship formula couples the two dimensions of material input and technical behavior through multiplication. Only when the input spare parts cost is high, the maintenance work time is long, and the semantic matching is high, the maintenance depth index will be larger, thereby objectively evaluating the real contribution of a maintenance to the performance recovery of the equipment.

[0072] For better understanding, a specific calculation example is given below:

[0073] Suppose a maintenance event is "replace the scraper cylinder", which consumes spare parts amount , the base value constant is , this maintenance event contains the "replace" action, so the benchmark semantic score is , the actual maintenance time is hours, the average time of this type of fault history is hours, .

[0074] Economic input items: The use of logarithmic function reflects the diminishing marginal effect, that is, the increase of input amount will not linearly and infinitely increase the maintenance depth index.

[0075] Maintenance operation refinement items: Through the square root processing, the influence of operation time and semantic weight on the overall index is smoothed.

[0076] Final maintenance depth index calculation: Take 3.73 as the maintenance depth index of this maintenance event, the larger the value, the more thorough the maintenance.

[0077] In this way, by constructing the maintenance depth index model, the maintenance effect that is difficult to measure can be digitized. Not only the economic factor of spare parts cost is considered, but also the time input of maintenance personnel and the semantic meaning of specific maintenance operation are combined, making the evaluation result more scientific and objective.

[0078] S4: Introduce age depreciation factor to establish a dynamic health degree evolution model of non-complete recovery to calculate the predicted health degree after each maintenance event.

[0079] Existing technologies often assume complete repair, ignoring irreversible damage caused by equipment aging. This step introduces the concept of thermodynamic entropy increase and establishes a dynamic model constrained by the age of the equipment. This model no longer simply resets the health level to 100%, but dynamically calculates the recovery rate based on the depth of maintenance and the current age of the equipment, ensuring that the prediction results conform to the physical degradation law of the entire life cycle of the equipment.

[0080] Specifically, the process includes the following:

[0081] First, construct the age-related depreciation factor: ,in, The depreciation factor is the age of the machine. For PCB printing machines in the first The cumulative total operating time at the start time of the maintenance event. For the design life of PCB printers, Let be the natural constant, set for The value is 1 hour, which is an empirical value determined based on the fatigue life limit of the spindle and lead screw of a general printing press in the PCB industry.

[0082] The age-related depreciation factor simulates the irreversible degradation characteristics of basic equipment components (such as frame rigidity and circuit board aging) over time, especially in the new machine stage. , At this point, the equipment is in good condition, and the recovery in health brought about by maintenance is almost unimpeded, resulting in excellent maintenance outcomes, consistent with the equipment's strong recoverability in its early stages; in the middle stage of its service life, , As equipment performance begins to decline, the same repair methods can only restore 77% of the performance of a new machine; in the old stage, , Rapidly decayed to .

[0083] It is evident that when the cumulative operating time of the equipment approaches its design life, the age-related depreciation factor decreases significantly. This means that even if engineers perform in-depth component replacement and maintenance with a high maintenance depth index, the overall health can only be restored to about 36% of its original level due to irreversible factors such as the overall rigidity of the machine and circuit aging. The age-related depreciation factor accurately simulates the physical reality of the decline in the recovery ability of aging equipment, preventing the system from being blindly optimistic about the condition of old equipment.

[0084] Secondly, a dynamic health evolution model is established to calculate the predicted health of the PCB printer after each maintenance event:

[0085]

[0086] in, It is the first Predicted health status after the end of the maintenance event. For the first Remaining health at the start of the next maintenance. Set the preset natural decay rate. The rate is 0.01 / hour, which represents the natural depreciation rate of the equipment under non-production conditions, such as dried-out lubricating oil and dust accumulation. For the first End time of this maintenance incident For the first Start time of the maintenance incident Set the preset recovery gain coefficient. The value is 15, which is an empirical adjustment coefficient used to map the dimensionless depth index to a health rating system ranging from 0 to 100. For example, when performing a deep repair on a new machine... and Health score increased by approximately 60 points. For the first Repair depth index of the next repair incident. This is a depreciation factor based on machine age, applied when the equipment is in its new machine stage. A value close to 1 indicates excellent repair results, especially when the equipment is nearing its design life. Rapidly decayed to The value is close to 0.36 when the equipment exceeds its service life. Approaching 0, this simulates the physical law that old equipment is difficult to repair. For machines that are on the verge of being scrapped, even if they undergo in-depth repairs, the repair depth index is very large. However, due to the irreversible fatigue of the basic components, the recovery rate of their health will be strongly suppressed by the age-related depreciation factor, thus achieving accurate health prediction.

[0087] In this relation, This is a natural loss / depletion term. The duration of maintenance downtime reflects the slight depreciation and degradation of equipment due to idleness during maintenance. This aligns with the principle of entropy increase in the physical world, ensuring the model avoids the erroneous logic of "no loss during downtime." It refers to the amount of health restored and the depth of repair. Depreciation factor based on age The combined effect is beneficial for older equipment. When the value is very small, this term will be significantly compressed, reflecting the characteristics of incomplete recovery.

[0088] The following demonstration of the model's calculation process compares the health recovery of new and old equipment during the same level of maintenance.

[0089] Assume the first The remaining health degree of the equipment before the start of the minor repair event , the natural loss during shutdown is negligible, i.e. partly ignored, the recovery gain coefficient , the repair depth index , the design life hours.

[0090] A new PCB printer, cumulative running hours, ; for , the age loss factor is , the health recovery amount: , the predicted health: , it shows that the performance of the new equipment after repair is excellent.

[0091] An old PCB printer, cumulative running hours, the age loss factor is , the health recovery amount: , the predicted health: , it shows that even if the same depth repair is performed on the old equipment, due to the fatigue of the basic components, the health cannot be restored to a high level.

[0092] In this way, by introducing the age loss factor, the physical law that old equipment is difficult to repair can be accurately simulated, the safety hazard caused by overestimating the health status of old equipment after repair is avoided, and the dynamic and accurate correction of the health status of the equipment throughout its life cycle is realized.

[0093] S5: Based on the predicted health after each repair event, fault prediction is carried out, and hierarchical maintenance decision is triggered.

[0094] Based on the established dynamic health evolution model of incomplete recovery, the system can analyze the future health status of the equipment like state deduction. This step converts theoretical calculation into specific operation and maintenance decision, solving the contradiction between over-repair and insufficient repair in traditional maintenance.

[0095] Specifically, it includes:

[0096] Firstly, the decay curve is constructed, the failure threshold is set to 20, according to the statistical distribution of historical failure data, when the predicted health is lower than 20, the probability of sudden shutdown of the equipment is more than 85%, which belongs to the high-risk interval, taking any moment of the PCB printer as the current moment, taking the predicted health after the end of the last repair event closest to the current moment as the reference, according to the age loss factor, the curve of the predicted health naturally decaying with time is constructed.

[0097] Then, the predicted failure time is determined: the time corresponding to the first time when the curve is less than 20 is determined as the predicted failure time.

[0098] Finally, a maintenance decision is triggered: the predicted completion time of the production order of the PCB printer at the current time is obtained, the predicted failure time is compared with the predicted completion time, if the predicted failure time is less than the predicted completion time, it means that the equipment is extremely likely to have a shutdown failure in the middle of production, a first-level warning signal is generated, deep maintenance is needed, and the deep maintenance is prompted to be performed in the nearest shift interval, the deep maintenance includes component replacement operation, and this decision logic can maximize the avoidance of unplanned shutdown loss. If the predicted failure time is greater than the predicted completion time, the attenuation slope of the health degree curve is calculated, the slope threshold is set to-0.5 points / hour, if the absolute value of the attenuation slope exceeds the slope threshold, it is indicated that although no failure has occurred, the equipment state is rapidly deteriorating, a second-level warning signal is generated, and inspection and maintenance are needed, and the equipment lubrication system or air pressure system is prompted to be checked to prevent the state from further deteriorating, and if the slope threshold does not exceed the slope threshold, it is indicated that the equipment state is in a normal operating state, and no warning or maintenance is needed.

[0099] Referring to Figure 2 , the device full life cycle health degree evolution prediction comparison is shown, the predicted health degree curve of the prior art model presents a mechanical sawtooth shape feature, the health degree is forced to rebound to 100 points after each maintenance, and the score still maintains at a high position of more than 80 points at the end of the graph (close to 1000 hours), which causes a false impression that the equipment is very healthy for the management personnel, and there is a great risk of missed detection. The predicted health degree curve of the present application presents a linear attenuation trend in line with the physical law, and accelerates attenuation at the end. At the shallow maintenance points of 200 hours and 450 hours, the curve only has a small amplitude of rebound, which truly reflects the limitation of the shallow means. At the middle maintenance of 700 hours, the health degree rebound amplitude is limited due to the suppression of the age loss factor in the formula, and cannot return to a high position. Finally, the health degree breaks through the failure critical line at about 810 hours, the preventive overhaul order is triggered at this moment, and the failure risk is locked in advance.

[0100] The second aspect of the embodiment of the present application also provides a PCB printer failure prediction system based on historical maintenance logs, the system comprises a memory and a processor, the memory stores a computer program, and the processor executes the computer program to realize the steps of the PCB printer failure prediction method based on historical maintenance logs. The processor can be a central processing unit (CPU), a digital signal processor (DSP) or a field programmable gate array (FPGA), and the memory can be a volatile memory or a non-volatile memory.

[0101] The above merely provides the preferred embodiment of the present application, and is not used to limit the present application. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present application should be included in the protection scope of the present application.

Claims

1. A PCB printing press fault prediction method based on historical maintenance logs, characterized in that, include: Obtain the full lifecycle historical maintenance logs of the PCB printer, use the TF-IDF algorithm to extract keywords to match the pre-built maintenance semantic knowledge base, and determine the baseline semantic score for each maintenance event; Extract the spare parts consumption amount and maintenance time for each maintenance event, and calculate the maintenance depth index for that maintenance event based on the baseline semantic score of that maintenance event. This index reflects the degree of equipment performance recovery achieved by the maintenance event, and satisfies the following: ； For the first Repair depth index of the next repair incident. It is a logarithmic function. It is a natural constant. For the first Total cost of spare parts consumed in this maintenance incident. As a preset basic value constant, , The first The baseline semantic score and actual time consumed for each maintenance event For the entire lifecycle historical maintenance log and the first The average repair time for all repair events with the same fault type in this repair event. This is a preset zero-prevention smoothing constant; Obtain the total cumulative running time and remaining health of the PCB printer at the start of each maintenance event, as well as the downtime caused by that maintenance event; An age-related depreciation factor is constructed based on the total cumulative operating time and the design life of the PCB printer to reflect the irreversible wear and tear that occurs as the equipment ages, satisfying the following: , The depreciation factor is the age of the machine. For PCB printing machines in the first The cumulative total operating time at the start time of the maintenance event. The design lifespan of the PCB printer; By combining the machine age depreciation factor, remaining health, downtime, and maintenance depth index, a dynamic health evolution model is established to calculate the predicted health of the PCB printer after the end of the maintenance event. Based on the predicted health, fault prediction is performed on the PCB printer to meet the following conditions: ； It is the first Predicted health status after the end of the maintenance event. For the first Remaining health at the start of the next maintenance. The preset natural decay rate, , The first The end time and start time of this maintenance event. This is the preset recovery gain coefficient.

2. The PCB printing machine fault prediction method according to claim 1, characterized in that, The maintenance semantic knowledge base is constructed based on the following method: The entire lifecycle historical maintenance log is segmented to extract maintenance action predicates; a three-level resilience predicate mapping model is established to map the maintenance action predicates to the corresponding weight levels: Level 1 resilience predicate: represents the replacement of irreversible physical entities, corresponding to spare parts replacement behavior, and is assigned Level 1 weight; Level 2 resilience predicate: represents the reverse calibration of physical parameters, corresponding to precision adjustment or tightening behavior, and is assigned Level 2 weight. The third-level resilience predicate represents the software logic reset or surface state cleanup, corresponding to restart or dust removal behavior, and is assigned a third-level weight. The maintenance action predicate and its corresponding weight are stored in a key-value pair database to form the maintenance semantic knowledge base. The first-level weight, second-level weight and third-level weight have a strictly decreasing relationship.

3. The PCB printing machine fault prediction method according to claim 2, characterized in that, The baseline semantic score for each maintenance event is generated based on the following method: The maintenance logs for each maintenance event are segmented and stop words are removed to obtain a word sequence. The word frequency of each word in the word sequence in the maintenance log and the inverse document frequency of the word in the entire lifecycle of the historical maintenance logs are calculated. Based on the word frequency and the inverse document frequency, the TF-IDF algorithm is executed to extract all core keywords of the word sequence. Each core keyword is matched with a pre-built maintenance semantic knowledge base to obtain the resilience predicate level corresponding to the core keyword. The weight corresponding to the resilience predicate level is used as the weight of the core keyword. The maximum value of the weights of all core keywords is used as the baseline semantic score of the maintenance event.

4. The PCB printing machine fault prediction method according to claim 1, characterized in that, Based on the predicted health status, fault prediction for the PCB printing machine is performed, including: A preset fault threshold is set. The current time is any moment when the PCB printer is running. The predicted health status after the last maintenance event closest to the current time is used as the benchmark. Based on the age depreciation factor, a curve is constructed to show the natural decay of the predicted health status over time. The moment when the curve first falls below the fault threshold is determined as the predicted fault moment. Maintenance decisions are triggered based on the predicted fault moment to achieve fault prediction for the PCB printer.

5. The PCB printing press fault prediction method based on historical maintenance logs according to claim 4, characterized in that, Maintenance decisions are triggered based on the predicted failure time, including: The system obtains the estimated completion time of the production orders for the PCB printer at the current moment and compares the predicted failure time with the estimated completion time. If the predicted failure time is less than the estimated completion time, a Level 1 warning signal is generated, prompting a deep repair during the nearest shift change, which includes component replacement. If the predicted failure time is not less than the estimated completion time, the system calculates the decay slope of the health curve. If the decay slope exceeds a preset slope threshold, a Level 2 warning signal is generated, prompting a check of the equipment's lubrication or air pressure system. If the decay slope does not exceed the preset slope threshold, no warning signal is generated.

6. The PCB printing machine fault prediction method according to claim 1, characterized in that, After obtaining the full lifecycle historical maintenance logs of the PCB printer, the following preprocessing operations are performed: deduplication verification is performed on the full lifecycle historical maintenance logs; Semantic feature mapping: Traverse the entire lifecycle of historical maintenance logs, extract the fault description text and maintenance measure text corresponding to each maintenance event, encapsulate them into corpus blocks, and use them as source data to determine the benchmark semantic score of each maintenance event; Cost feature mapping: Extract the spare parts consumption list and spare parts unit price data for each maintenance event, sum the spare parts amount, and map it to the total spare parts consumption amount for each maintenance event; Time-series feature mapping: Extract the fault occurrence timestamp, maintenance start timestamp, and maintenance end timestamp associated with each maintenance event to calculate the actual time consumed for each maintenance event, and estimate the total cumulative running time of the PCB printer at the start time of each maintenance event based on the time axis.

7. A PCB printing press fault prediction system based on historical maintenance logs, characterized in that, The system includes a memory and a processor, the memory storing a computer program, and the processor executing the computer program to implement the steps of the PCB printing machine fault prediction method as described in any one of claims 1-6.

Images