Cloud computing-based clean room equipment operation and maintenance system

Abstract

The present application relates to cloud computing and clean room equipment operation and maintenance management technical field, specifically to clean room equipment operation and maintenance system based on cloud computing;Contain data access, cloud simulation processing and terminal execution module;The system constructs management benchmark model through collecting business plan and operation and maintenance rule data;Its core is to combine resource account to generate theoretical damaged data through abnormal disturbance simulation, and calculate the dynamic time warping similarity and time series cross-correlation value of actual and theoretical residual vector to generate state decision result;When outputting state decision, the system will issue corresponding operation and maintenance instruction work order, and dynamically correct the disturbance operator weight;The present application realizes the change from single judgment depending on fixed threshold to accurate distinction between normal load deviation and abnormal equipment degradation, which can reduce the false positive rate of unnecessary room inspection and the false negative rate of equipment out of control.

Description

Technical Field

[0001] This invention relates to the field of cloud computing and cleanroom equipment operation and maintenance management technology, specifically a cloud computing-based cleanroom equipment operation and maintenance system. Background Technology

[0002] Cleanroom equipment operation and maintenance is an important means to ensure the continuous stability and compliance of the biopharmaceutical production environment. Existing systems typically collect operating parameters, maintenance records and on-site behavior information around equipment such as incubators, biosafety cabinets and laminar flow hoods, monitor equipment status, issue alarms and manage work orders, and can support the daily operation and maintenance of cleanroom equipment to a certain extent.

[0003] With the dynamic changes in production scheduling, the collaborative operation of equipment groups, and the increasing demand for data access from multiple systems, traditional cleanroom equipment operation and maintenance methods still rely heavily on fixed threshold alarms, single status quantity judgments, or decentralized ledger comparisons. This makes it difficult to distinguish between normal fluctuations caused by changes in business load and abnormal deviations caused by maintenance deficiencies or hidden equipment degradation. This can easily lead to false alarms, missed alarms, and unnecessary in-room inspections, resulting in limited accuracy in determining the operation and maintenance status and delays in processing. Summary of the Invention

[0004] To address the aforementioned technical problems, this invention provides a cloud-based cleanroom equipment operation and maintenance system. Specifically, the technical solution of this invention includes:

[0005] The data access terminal is used to collect business plan data, operation and maintenance rule data, resource ledger data, equipment status time series data and operation and maintenance behavior time series data, and upload them to the cloud server according to a preset unified time window as a unified time benchmark.

[0006] A cloud server is used to construct a management baseline model based on business plan data and operation and maintenance rule data. The management baseline model includes the theoretical resource consumption time series and management baseline parameters of cleanroom equipment under rated performance and meeting preset operation and maintenance rules. Abnormal disturbance parameters and operators are determined based on resource ledger data, and the abnormal disturbance operators are used to simulate management scenario disturbances on the management baseline model to generate theoretical damage data. Based on a preset unified time window, the equipment status time series data and the management baseline model, as well as the theoretical damage data and the management baseline model, are differentially processed to form a difference matrix. Feature extraction is then used to generate actual residual vectors and theoretical residual vectors. A status decision result is generated based on the dynamic time warping similarity and time series cross-correlation value of the actual residual vectors and the theoretical residual vectors.

[0007] The execution terminal is used to receive status judgment results and generate and issue corresponding operation and maintenance work orders, monitoring and maintenance instructions or supplementary data collection tasks, and obtain the work order execution results of the operation and maintenance work orders;

[0008] The cloud server is also used to correct the weight parameters of the abnormal disturbance operator based on the status judgment result and the work order execution result, and to update the work order priority rules.

[0009] Optionally, the data access terminal includes:

[0010] The business interface module is used to receive business plan data;

[0011] The rules interface module is used to receive operation and maintenance rules data;

[0012] The ledger interface module is used to receive resource ledger data;

[0013] The status acquisition module is used to receive device status time-series data and operation and maintenance behavior time-series data, and then send them to the cloud server after timestamp normalization and alignment according to the preset sampling period and unified time window.

[0014] Optionally, the cloud server includes:

[0015] The benchmark building module is used to build management benchmark models based on business plan data and operation and maintenance rule data;

[0016] Among them, the management benchmark model is used to characterize the theoretical operating benchmark when the equipment is in its rated performance state and meets the preset operation and maintenance rules;

[0017] The theoretical operating benchmarks include theoretical resource consumption time series, theoretical material replacement frequency, and theoretical maintenance man-hour consumption.

[0018] Optionally, the cloud server may also include:

[0019] The parameter injection module is used to call the abnormal disturbance operator based on the resource ledger data and inject the abnormal disturbance operator into the management benchmark model to generate theoretical damage data.

[0020] Among them, the abnormal disturbance operators include the maintenance deficiency operator and the implicit aging operator;

[0021] Among them, the maintenance missing operator is used to introduce time delay and amplitude offset relative to the theoretical benchmark into the response curve;

[0022] Among them, the implicit aging operator is used to introduce discrete increments that deviate from the theoretical curve in the mapping relationship between load and operating parameters.

[0023] Optionally, the cloud server may also include:

[0024] The differential extraction module is used to perform differential processing on the device status time series data and the management baseline model based on a unified time window corresponding to the business plan cycle, so as to generate the actual residual vector.

[0025] The differential extraction module is also used to perform differential processing on the theoretically damaged data and the management benchmark model to generate a theoretical residual vector;

[0026] The actual residual vector and the theoretical residual vector are both output vectors obtained by feature extraction from the difference matrix formed by differentiating the multidimensional business state time series at multiple time offsets.

[0027] Optionally, the cloud server may also include:

[0028] The coupled decision module is used to perform coupled calculations on the actual residual vector and the theoretical residual vector based on the dynamic time warping similarity and the temporal cross-correlation value, so as to obtain the normalized temporal similarity and generate the state decision result.

[0029] The status judgment results include alarm confirmation, alarm suppression, and pending review instructions;

[0030] When the normalized temporal similarity is higher than or equal to the preset high threshold, the coupled decision module outputs a confirmation alarm.

[0031] When the normalized temporal similarity is lower than or equal to the preset low threshold, the coupled decision module outputs a suppression alarm.

[0032] When the normalized temporal similarity is between the high threshold and the low threshold, the coupled decision module outputs a review instruction.

[0033] Among them, the high threshold and low threshold are thresholds obtained based on historical work order samples.

[0034] Optionally, the execution terminal includes:

[0035] The work order generation module is used to generate high-priority operation and maintenance work orders based on confirmed alarms, generate monitoring and maintenance instructions based on suppressed alarms, generate supplementary data collection tasks based on instructions pending review, and send high-priority operation and maintenance work orders, monitoring and maintenance instructions or supplementary data collection tasks to the field operation and maintenance terminal and the data collection device corresponding to the status acquisition module.

[0036] The result feedback module is used to send the work order execution results back to the cloud server.

[0037] The cloud server adjusts the weight parameters of the abnormal disturbance operator based on the status judgment result and the work order execution result.

[0038] Optionally, the business plan data is the dynamic production scheduling table of the manufacturing execution system, the resource ledger data is the maintenance material consumption ledger of the enterprise resource planning system, the maintenance rule data is the digital standard operating procedure rule data, the equipment status time series data is the time series of group operation parameters and equipment status parameter time series fluctuations of clean area equipment, and the maintenance behavior time series data is the trajectory check-in log of maintenance personnel.

[0039] Optionally, cleanroom equipment includes biosafety cabinets, incubators, or laminar flow hoods;

[0040] Among them, the management benchmark model generates the theoretical group resource consumption sequence of biosafety cabinets, incubators or clean benches based on the dynamic production scheduling table of the manufacturing execution system;

[0041] The coupling decision module is used to distinguish between load offsets caused by production schedule changes and abnormal offsets caused by maintenance deficiencies or equipment performance degradation, based on the similarity between the theoretical group resource consumption time series and the actual residual vector.

[0042] Compared with the prior art, the present invention has the following beneficial effects:

[0043] 1. This system constructs a management baseline model by collecting business plan data and operation and maintenance rule data. Combined with resource ledger data, it uses anomaly disturbance operators to simulate the management baseline model to generate theoretical damage data. The system performs differential processing on the equipment status time series data and theoretical damage data to obtain the actual residual vector and the theoretical residual vector. The system then uses the dynamic time warping similarity and time series cross-correlation value of the two to generate the status judgment result. This mechanism overcomes the shortcomings of traditional systems that rely on fixed thresholds and single status quantity judgments. It can accurately distinguish between normal load deviations caused by production schedule changes and abnormal deviations caused by maintenance deficiencies or equipment performance degradation, effectively reducing the frequency of unnecessary in-room inspections triggered by false alarms and equipment loss of control caused by missed alarms.

[0044] 2. This system uses a status acquisition module to timestamp and align various cross-system access data according to a preset sampling period and a unified time window, laying a unified time benchmark for subsequent analysis. Simultaneously, in the differential extraction module, based on the unified time window, multi-dimensional business status time series are differentially processed at multiple time offsets to form a differential matrix, which is then used for feature extraction to obtain the output vector. This mechanism effectively eliminates time series interference caused by inconsistent sampling periods across multiple systems and reconstructs multi-dimensional indicators into a sequence carrier capable of supporting trend shift comparison, thus maintaining a strong ability to capture abnormal patterns even under complex operating conditions with time series phase shifts.

[0045] 3. This system receives status judgment results from the execution terminal, generates and issues high-priority operation and maintenance work orders, monitoring and maintenance instructions, or supplementary data collection tasks accordingly, and feeds back the work order execution results to the cloud server through the result feedback module. Based on these results, the cloud server corrects the weight parameters of the anomaly disturbance operator and updates the work order priority rules. This mechanism not only directly transforms the backend identification results into differentiated on-site intervention actions, but also uses actual on-site feedback to perform closed-loop learning and penalty updates on theoretical simulation features, making the anomaly diagnosis model gradually closer to the actual operation and maintenance management characteristics of the enterprise, and realizing the adaptive updating of the operation and maintenance model and the automation of anomaly intervention. Attached Figure Description

[0046] The present invention will be further explained below with reference to the accompanying drawings and embodiments:

[0047] Figure 1 This is a structural diagram of the system of the present invention. Detailed Implementation

[0048] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to specific embodiments.

[0049] like Figure 1 As shown, the cloud-based cleanroom equipment operation and maintenance system includes:

[0050] The data access terminal is used to collect business plan data, operation and maintenance rule data, resource ledger data, equipment status time series data and operation and maintenance behavior time series data, and upload them to the cloud server according to a preset unified time window as a unified time benchmark.

[0051] A cloud server is used to construct a management baseline model based on business plan data and operation and maintenance rule data. The management baseline model includes the theoretical resource consumption time series and management baseline parameters of cleanroom equipment under rated performance and meeting preset operation and maintenance rules. Abnormal disturbance parameters and operators are determined based on resource ledger data, and the abnormal disturbance operators are used to simulate management scenario disturbances on the management baseline model to generate theoretical damage data. Based on a preset unified time window, the equipment status time series data and the management baseline model, as well as the theoretical damage data and the management baseline model, are differentially processed to form a difference matrix. Feature extraction is then used to generate actual residual vectors and theoretical residual vectors. A status decision result is generated based on the dynamic time warping similarity and time series cross-correlation value of the actual residual vectors and the theoretical residual vectors.

[0052] The execution terminal is used to receive status judgment results and generate and issue corresponding operation and maintenance work orders, monitoring and maintenance instructions or supplementary data collection tasks, and obtain the work order execution results of the operation and maintenance work orders;

[0053] The cloud server is also used to correct the weight parameters of the abnormal disturbance operator based on the status judgment result and the work order execution result, and to update the work order priority rules.

[0054] This embodiment provides a cloud-based cleanroom equipment operation and maintenance mechanism. Specifically, the mechanism revolves around the same business line: a biopharmaceutical manufacturing company continuously schedules cell culture batches within a week. The incubators, biosafety cabinets, and ultra-clean workbenches in the clean area need to complete continuous operation and compliant maintenance without increasing unnecessary entry into the room. The main function of the system is to distinguish between normal load changes caused by increased production tasks and abnormal deviations caused by maintenance deficiencies or hidden equipment degradation in the context of continuously changing production scheduling, and decide whether to issue operation and maintenance actions accordingly.

[0055] Specifically, the data access terminal first receives five types of data and unifies them to the same business time base. This unified time base can be set in 5-minute windows. Assume that there are three consecutive time windows W1, W2, and W3 between 8:00 and 8:15 a certain morning. The manufacturing execution system shows that W1 plans to run 2 incubators, W2 increases to 3, and W3 maintains 3. The standard operating procedure rules stipulate that the filter components need to be checked every 100 hours of cumulative operation. The enterprise resource planning system ledger shows that one of the devices has not used the planned filter consumables in the past two weeks. The equipment status time sequence records energy consumption, temperature recovery time, and differential pressure fluctuation. The maintenance behavior time sequence records that the on-duty personnel did not enter the partition where the device is located in W2. The data access terminal writes all of the above data into the unified time axis, so that the cloud can perform aligned analysis of plans, rules, ledgers, status, and behavior within the same time window.

[0056] The cloud server constructs a management benchmark model based on business plan data and operation and maintenance rule data. This model does not calculate the average of historical states, but rather assumes that the equipment is at its rated performance, maintenance is fully compliant, and consumables are in normal condition. Under the current production schedule, it shows the theoretical resource consumption sequence. Taking the above three time windows as an example, if the theoretical power consumption per unit window for two incubators under rated conditions is 4 and 4 respectively, then the theoretical total energy consumption of W1 is 8. W2, due to the addition of one incubator, should have a total energy consumption of 12. W3 remains at 12. If the corresponding theoretical temperature recovery times are 2 minutes, 2.3 minutes, and 2.4 minutes respectively, then a theoretical benchmark sequence on the same time axis can be formed. This sequence, combined with the theoretical material replacement frequency and theoretical maintenance man-hours, constitutes the management benchmark parameters.

[0057] Based on this, the cloud determines abnormal disturbance parameters according to resource ledger data and uses abnormal disturbance operators to simulate the management baseline model. Its processing logic is to pre-construct potential abnormal forms and then couple and compare them with actual data. For example, if the enterprise resource planning system ledger indicates that a certain piece of equipment should have had its primary efficiency component replaced last week but no material requisition record is found, the system assigns an initial weight greater than the preset baseline value to the maintenance deficiency. If the cumulative operating hours of the equipment are close to the upper limit of its design life, then implicit aging also receives a corresponding weight. During simulation, the system adds the response delay and energy consumption increase caused by the maintenance deficiency to the theoretical baseline. For example, the theoretical total energy consumption of W1 to W3 is adjusted from 8, 12, 12 to 8.8, 13.5, 13.7, while the temperature recovery time is extended from 2, 2.3, 2.4 minutes to 2.4, 2.9, 3.0 minutes, thereby forming theoretical damage data.

[0058] The system then enters a dual-track differential processing flow. In the actual data processing branch, the equipment status time series is subtracted from the management baseline model to obtain the actual residual vector. In the theoretical data processing branch, the theoretical damage data is subtracted from the same management baseline model to obtain the theoretical residual vector. For ease of explanation, it is assumed that the system selects two dimensions: total energy consumption and temperature recovery time. The actual observation values ​​of W1 to W3 are 8.7, 13.4, 13.8 and 2.5, 2.8, 3.1 minutes, while the theoretical baselines are 8, 12, 12 and 2.0, 2.3, 2.4 minutes. The actual residuals can be written as [(0.7,0.5),(1.4,0.5),(1.8,0.7)], and the corresponding theoretical damage residuals are [(0.8,0.4),(1.5,0.6),(1.7,0.6)]. The system can further flatten or extract features to transform them into vectors of a unified dimension for subsequent similarity analysis.

[0059] The cloud-based system generates a status judgment result based on dynamic time warping similarity and temporal cross-correlation value. Dynamic time warping is used to tolerate the occurrence of actual anomalies occurring earlier or later than the theoretical simulation within a preset time tolerance range. Temporal cross-correlation is used to observe the synchronicity of the trends of two residual sequences. In the example, if the calculated dynamic time warping similarity is 0.91 and the cross-correlation peak is 0.87, both of which are higher than the confirmation threshold, it is judged as a real anomaly, and a confirmation alarm is output. If the dynamic time warping similarity is only 0.35, it means that although the data is high, the shape does not match the theoretical anomaly, and a suppression alarm is output. If the result is in the middle range, such as 0.62, a verification instruction is output, requiring more data to be collected or non-indoor verification to be performed.

[0060] After receiving the status judgment result from the execution terminal, the system generates and issues corresponding actions. For confirmed alarms, the system generates a high-priority work order, specifying the corresponding area, equipment number, suspected anomaly type, and recommended inspection items. For suppressed alarms, a monitoring hold instruction is issued to maintain the existing sampling density and temporarily suspend indoor entry. For alarms pending review, a supplementary sampling task is sent to the status acquisition side, such as increasing the sampling frequency, adding local vibration acquisition, or verifying attendance logs. After the work order is executed, the field results are sent back to the cloud as the basis for subsequent correction of the anomaly disturbance operator weight and updating the work order priority rules. For example, if it is repeatedly confirmed that the enterprise resource planning system has not issued materials and the energy consumption tail height corresponds to maintenance deficiency, the prior weight of this operator is increased. If a certain type of alarm is frequently reviewed as production scheduling fluctuation, the corresponding priority rule is decreased.

[0061] As a fault-tolerant processing mechanism, when business plan data is missing within a certain time window but equipment status data is continuous, the system can mark this time window as a baseline incomplete window, and instead of directly outputting a confirmation alarm, it will prioritize generating a supplementary data collection task or a pending review instruction. When there is a short-term breakpoint in the equipment status timing, interpolation of adjacent time windows can be used to form a temporary alignment sequence, but if the interpolation ratio exceeds the preset upper limit, it will not participate in the similarity judgment. When the theoretical simulation matches multiple abnormal disturbance modes at the same time and the highest score difference is lower than the preset separation threshold, the system will not directly locate a single fault, but will output a composite abnormality prompt to avoid mis-assigning orders.

[0062] For example, on Wednesday afternoon at this biopharmaceutical manufacturing company, the total energy consumption of the incubator group increased significantly due to an additional batch of cell expansion tasks. If a fixed threshold rule is used, the system would directly identify this upward trend as a fault and require an in-room inspection. However, in this embodiment, the theoretical benchmark related to the business is first reconstructed using the production schedule plan, and then two types of disturbance operators, namely maintenance deficiency and latent aging, are used to simulate the timing deviation pattern presented under abnormal conditions. Only when the actual residual is highly coupled with the maintenance deficiency residual is a high-priority work order issued. If the curve rises as a whole due to the increase in production and the pattern is mismatched, it is suppressed. The purpose of this step is to improve the targeting of anomaly identification in the context of dynamic production and reduce unnecessary in-room inspections caused by false alarms and equipment malfunctions caused by missed alarms.

[0063] In this embodiment, the data access terminal includes:

[0064] The business interface module is used to receive business plan data;

[0065] The rules interface module is used to receive operation and maintenance rules data;

[0066] The ledger interface module is used to receive resource ledger data;

[0067] The status acquisition module is used to receive device status time-series data and operation and maintenance behavior time-series data, and then send them to the cloud server after timestamp normalization and alignment according to the preset sampling period and unified time window.

[0068] This embodiment provides a refined implementation mechanism for the data access end. Specifically, in the aforementioned overall process, if data from different sources is simply aggregated to the cloud without addressing issues such as inconsistent sampling periods, different timestamp formats, and delayed log writing, the subsequently constructed management baseline and residual vector will become misaligned, leading to the misjudgment of timeline misalignment as abnormal morphological differences. Therefore, this embodiment further divides the data access end into a business interface module, a rule interface module, a ledger interface module, and a status collection module, and completes normalization alignment on the access side.

[0069] Specifically, the business interface module can periodically retrieve the dynamic production schedule from the Manufacturing Execution System (MES). Assuming the MES updates every 30 minutes, it returns a production schedule segment for a certain equipment group from 8:00 to 12:00. The rules interface module receives maintenance rules from the Digital Standard Operating Procedures (DSOP) system, and its update frequency may be daily or version-based. The ledger interface module receives data such as consumables in / out and maintenance man-hour registration from the Enterprise Resource Planning (ERP) system, which may include supplementary data. The status acquisition module collects energy consumption, differential pressure, temperature, and maintenance personnel trajectories at intervals of 1 minute or less. Due to the different original time granularities of the four types of data, this embodiment first selects a unified time window, such as 5 minutes. The 30-minute production schedule from the MES system is split and mapped to 6 consecutive 5-minute windows. The DSOP rules are mapped to be constantly valid within the corresponding effective interval. The ERP system ledgers are placed into the nearest business window according to the actual usage time. The equipment status and trajectory logs are aggregated into one or more statistical values ​​per window according to the sampling period.

[0070] To illustrate the alignment method, the following example is provided with specific parameter settings; assuming W1 is 8:00-8:05 and W2 is 8:05-8:10; the Manufacturing Execution System uploads the production schedule for Incubator A to be running and Incubator B to be on standby at 8:00; the Enterprise Resource Planning System enters a record of unused filter material at 8:07; the Status Acquisition Module receives power values ​​of 4.0, 4.2, and 4.1 from Incubator A at 8:01, 8:02, and 8:04 respectively, and at 8:06... At 8:08, data of 4.5 and 4.6 was received; the tracking record at 8:09 showed that the maintenance personnel were still in the outer changing area; after normalization, W1 can form a plan to run 1 unit, the rule is valid, no materials were requisitioned, the average power is 4.1, and no one entered the area; W2 can form a plan to run 1 unit, the rule is valid, there are clues of overdue materials requisitioned, the average power is 4.55, and no one entered the area; in this way, when the cloud analyzes the increased energy consumption of W2, it can simultaneously see its temporal correlation with the unrequited materials ledger and tracking information;

[0071] As a fault-tolerant mechanism, if the business interface module fails to obtain the latest production schedule on time, the previous valid production schedule version is retained and marked as an inheritance plan. This marking will reduce the confidence weight of the confirmed alarm in subsequent decisions. If the enterprise resource planning system has retroactive data entry resulting in a record time later than the actual occurrence time, the system saves both the data entry time and the business occurrence time, and prioritizes alignment based on the business occurrence time. If multiple value conflicts occur in status acquisition, such as two severely discrete values ​​uploaded by the same sensor within the same window, outlier removal is performed first, and then the values ​​within the window are formed using the median or truncated mean. If the number of valid samples within a window is lower than the preset minimum, the window is marked as a low-confidence window and will not participate in high-priority order dispatch triggering.

[0072] For example, during the Thursday night shift at this biopharmaceutical manufacturing company, the Manufacturing Execution System (MES) temporarily added an incubator to the production schedule. The Enterprise Resource Planning (ERP) system had not yet reflected the material requisition in a timely manner, while the on-site power sampling frequency reached once per minute. Through the access and normalization in this embodiment, the system can simultaneously see four types of information within the same 5-minute window: production schedule increase, no material requisition, continuously increasing power, and no maintenance personnel entering the area. This avoids the subsequent incorrect splicing of isolated data fragments. The purpose of this step is to establish a consistent data foundation for residual calculation on the cloud side, thereby achieving comparability of cross-system and cross-frequency data.

[0073] In this embodiment, the cloud server includes:

[0074] The benchmark building module is used to build management benchmark models based on business plan data and operation and maintenance rule data;

[0075] Among them, the management benchmark model is used to characterize the theoretical operating benchmark when the equipment is in its rated performance state and meets the preset operation and maintenance rules;

[0076] The theoretical operating benchmarks include theoretical resource consumption time series, theoretical material replacement frequency, and theoretical maintenance man-hour consumption.

[0077] This embodiment provides a mechanism for constructing a management baseline model. Specifically, in the aforementioned process, if only the historical average value is used as the comparison baseline, non-standard maintenance, slow aging, and non-standard operations contained in the historical data will be misjudged as normal, causing the system baseline itself to shift. Therefore, this embodiment introduces a baseline construction module, which uses business plans and operation and maintenance rules as constraints to reconstruct the theoretical operating baseline that the equipment should achieve when it is at rated performance and fully compliant.

[0078] Specifically, the baseline construction module parses the business plan into the target production load and the operation and maintenance rules into the compliant execution path to undertake the load. For example, the production schedule on a certain morning requires the biosafety cabinet to run for 3 hours, the incubator to be continuously kept warm for 6 hours, and the clean bench to work intermittently 4 times. The corresponding standard operating procedures stipulate that a certain type of filter material should be replaced every 200 hours of cumulative incubator operation, the biosafety cabinet should complete a self-inspection once every working day, and the standard maintenance time for a single maintenance is 20 minutes.

[0079] The baseline construction module outputs three types of content based on this: the first type is the theoretical resource consumption time series, such as the theoretical power, gas consumption or auxiliary resource occupation of each device in each time window; the second type is the theoretical material replacement frequency, such as the replacement expectation when a certain device reaches the 100-hour, 200-hour, and 300-hour milestones; the third type is the theoretical maintenance man-hour consumption, such as how many man-hours should be invested this week under the premise of compliance.

[0080] The following example illustrates a specific operational scenario. Assume that within windows W1, W2, and W3, the planned loads of incubator A are 50%, 80%, and 80%, respectively. Under rated performance, its theoretical power can be set to 4.0, 5.0, and 5.0, respectively. Biosafety cabinet B performs a high-cleanliness operation in W2, theoretically increasing its power by an additional 0.8. Therefore, B's power within the three windows is 3.0, 3.8, and 3.0, respectively. If there are two such devices in the system, the theoretical resource consumption sequence can be represented as total power [7.0, 8.8, 8.0]. Further assuming that incubator A has accumulated 198 hours of operation, and the rule requires filter material replacement every 200 hours, then W3 becomes the trigger point for a theoretical material replacement frequency. If the corresponding standard maintenance time is 15 minutes, then the theoretical maintenance time is also included in the planned interval after W3. This baseline model not only provides the theoretical energy consumption baseline for the current time window but also constructs the baseline for material requisition and maintenance actions that should be triggered during this period.

[0081] As a fault-tolerant processing mechanism, if the business plan only provides production line-level tasks and does not accurately map them to the equipment level, this embodiment can first generate an equipment-level plan based on equipment grouping rules and historical compliant allocation templates, and then add a presumed allocation mark; if the standard operating procedure rules are changed, for example, a new standard is activated at 12 noon, the baseline construction module models the time windows before and after the effective time as the boundary to avoid mixing the old and new rules; if the equipment has a downtime maintenance window, the theoretical resource consumption is reduced to zero or to standby value within the time window, while the theoretical maintenance man-hours increase, and the low energy consumption during downtime is not mistakenly regarded as the low load characteristics of the equipment during normal operation;

[0082] For example, when the company is producing two cell culture batches in parallel on Friday, the manufacturing execution system requires an increase in the frequency of use of the clean bench in the afternoon, while the standard operating procedure simultaneously specifies that one of the incubators has reached its filter material replacement cycle. Based on this, the baseline construction module generates a theoretical baseline that shows an increase in energy consumption in the afternoon, but at the same time, a material replacement and a maintenance period should occur. If the actual monitoring only shows an increase in energy consumption without corresponding ledgers and work hour records, it provides a clear reference for subsequent abnormal disturbance simulation. The purpose of this step is to construct an ideal comparison surface that is not contaminated by historical poor operation and maintenance, so as to accurately locate management negligence and equipment degradation.

[0083] In this embodiment, the cloud server further includes:

[0084] The parameter injection module is used to call the abnormal disturbance operator based on the resource ledger data and inject the abnormal disturbance operator into the management benchmark model to generate theoretical damage data.

[0085] Among them, the abnormal disturbance operators include the maintenance deficiency operator and the implicit aging operator;

[0086] Among them, the maintenance missing operator is used to introduce time delay and amplitude offset relative to the theoretical benchmark into the response curve;

[0087] Among them, the implicit aging operator is used to introduce discrete increments that deviate from the theoretical curve in the mapping relationship between load and operating parameters.

[0088] This embodiment provides a parameterized injection mechanism for anomaly perturbation operators. Specifically, an ideal benchmark alone is insufficient to support accurate judgment because the physical causes of underlying anomalies are not directly and explicitly identified through sensor data. The system needs to first construct the temporal evolution of various anomaly patterns in the current business scenario. Without this process, the system can only compare the actual collected state data with the theoretical limit value, which is still a single fixed threshold alarm method. Therefore, this embodiment adds a parameter injection module on top of the benchmark model to transform ledger information records into simulateable anomaly perturbation features.

[0089] Specifically, the maintenance failure operator mainly simulates the lag and amplitude increase caused by operations that should have been performed but were not, resulting in a delay in system response. For example, if a filter component is not replaced on schedule, the time required for the equipment to reach the target state will be longer under the same workload, and the power required to maintain operation will be higher. The implicit aging operator simulates the deviation in the mapping relationship caused by the performance degradation of components that have not failed. For example, under the same workload of 80%, the required power was 5.0 under the same historical conditions, but the required power has changed to 5.4. This deviation may increase in a step-like or discrete manner. The parameter injection module first determines the strength of these two types of operators based on the resource ledger, and then superimposes them onto the ideal benchmark.

[0090] The following explanation uses specific numerical values; assuming the theoretical power of incubator A at W1, W2, and W3 is [4.0, 5.0, 5.0], and the theoretical temperature recovery time is [2.0, 2.3, 2.4]; the enterprise resource planning system ledger shows that the filter material should have been replaced the day before but was not collected, so the system assigns a value of 0.7 to the maintenance missing operator; when the equipment's cumulative operation reaches a specific proportion of the preset lifespan, the implicit aging operator is assigned a value of 0.3; during parameter injection, the maintenance missing operator can increase the recovery time by [0.3, 0.5, 0.5] and increase the power. [0.4,0.8,0.8]; The implicit aging operator further increases the power discretization [0.1,0.2,0.3]; Thus, the theoretically damaged data can be transformed into power [4.5,6.0,6.1] and recovery time [2.3,2.8,2.9]; If replaced with another device that is maintained on time but has been used for a longer period of time, the weight of the maintenance missing operator can be close to 0, while the weight of the implicit aging operator is higher, resulting in another damage curve; This enables the system to generate candidate anomaly fingerprint maps with multiple feature dimensions, rather than relying solely on a single fixed fault identification template;

[0091] Under the fault tolerance mechanism, if there are conflicts in the ledger data, such as one side showing that consumables have been issued, and the other side showing that the work order has not been closed, the parameter injection module can simultaneously generate two sets of simulation paths: one showing maintenance and the other showing suspected maintenance not being performed. In the subsequent decision stage, the system compares which set is closer to reality. If the ledger data sampling rate is lower than the preset threshold and cannot effectively support the disturbance intensity setting, the system adopts the industry default weight or the equipment family default weight, but adds a low confidence mark to the simulation result and does not directly trigger the highest priority dispatch. If maintenance absence and latent aging coexist strongly, the system can generate a combined operator to avoid fitting only a single-cause anomaly.

[0092] For example, on the third consecutive day of production at the company, a certain incubator did not have its planned filter material used, and its cumulative runtime had exceeded the average of similar equipment for the month. Based on this, the parameter injection module simultaneously injected the maintenance deficiency and the hidden aging effect into the ideal baseline, and deduced the theoretical damage data of the power continuously exceeding the baseline and the recovery time being delayed. If the actual sampling subsequently showed the same pattern, the system would be more inclined to judge it as a real operation and maintenance problem rather than a simple business fluctuation. The purpose of this step is to transform empirical anomaly knowledge into calculable and comparable theoretical anomaly spectrum features, thereby realizing the transformation from judging abnormally high indicator values ​​to analyzing the time-series mapping analysis process that causes abnormal fluctuations.

[0093] In this embodiment, the cloud server further includes:

[0094] The differential extraction module is used to perform differential processing on the device status time series data and the management baseline model based on a unified time window corresponding to the business plan cycle, so as to generate the actual residual vector.

[0095] The differential extraction module is also used to perform differential processing on the theoretically damaged data and the management benchmark model to generate a theoretical residual vector;

[0096] The actual residual vector and the theoretical residual vector are both output vectors obtained by feature extraction from the difference matrix formed by differentiating the multidimensional business state time series at multiple time offsets.

[0097] This embodiment provides a differential extraction and residual vector formation mechanism. Specifically, given an ideal benchmark and theoretically damaged data, directly comparing and analyzing the original time-series data is often subject to interference from the absolute magnitude of equipment performance, differences in the dimensions of heterogeneous sensor indicators, and the sampling phase of the time series. This makes it difficult to accurately quantify and characterize the degree of deviation and evolution trend of multi-dimensional state data relative to the theoretical benchmark. Therefore, this embodiment uses a differential extraction module to perform differential calculations between the actual data and the ideal benchmark, and between the theoretically damaged data and the ideal benchmark, respectively, and forms a difference matrix under multiple time offsets, which is then extracted as a vector.

[0098] Specifically, the actual residual reflects the deviation of the real world from the ideal state, which is a mixture of unexplained business fluctuations, abnormal noise, and real damage traces; the theoretical residual reflects the pure trace that a certain assumed anomaly should leave relative to the ideal state; in order to enhance the time series robustness, the system not only calculates the difference within the same window, but also the offset difference of the previous or next time window; for example, within the three windows W1, W2, and W3, the zero time window offset and the adjacent time window offset difference are calculated for the two dimensions of energy consumption and recovery time, respectively, forming a local feature difference matrix;

[0099] Continuing with the aforementioned equipment as an example; the actual observed power is [4.4, 5.9, 6.0], and the theoretical benchmark is [4.0, 5.0, 5.0], so the power difference under zero time window offset is [0.4, 0.9, 1.0]; if we consider the offset of adjacent time windows, that is, compare the current window with the previous benchmark, then the difference between W2 and W1 benchmark is 1.9, and the difference between W3 and W2 benchmark is 1.0; similarly, the recovery time can be obtained as the 0 window difference [0.3, 0.5, 0.5]; after arranging the power 0 offset, power 1 offset, and recovery time 0 offset corresponding to each time window in parallel, a multidimensional time difference matrix can be formed for W1, W2, and W3;

[0100] To adapt these residual characteristics to the subsequent dynamic time warping algorithm, the feature extraction operation here avoids the steps of directly eliminating the time dimension and compressing the data into a single dimension to a global average or extreme value. Instead, it performs cross-sectional weighting or feature dimensionality reduction on the multi-dimensional attribute features along a unified time axis, and merges the multi-dimensional deviation components within each time window into a comprehensive response value. The specific cross-sectional weighting logic is to aggregate the differential components of each dimension within the same window according to information content, stability, and business relevance while keeping the time sequence of W1 to W3 unchanged. Among them, information content is calculated based on the information entropy of each differential component, stability is calculated based on the inverse of the variance of each differential component under historical stationary conditions, and business relevance is calculated based on the Pearson correlation coefficient between each differential component and the corresponding business load variation.

[0101] The final actual residual vector is essentially a time series sequence of comprehensive deviations arranged sequentially according to time steps, such as [0.77, 1.45, 0.43]. The theoretically damaged data is compressed in the same way according to the time section features to obtain a theoretical residual vector that also has time series evolution characteristics, such as [0.80, 1.50, 0.45]. In this way, the subsequent similarity comparison is no longer directly affected by the inconsistency of the original multidimensional dimensions, and the time contour law of deviation evolution as the process develops is completely preserved.

[0102] In the data missing fault tolerance process, if a certain dimension is missing within a specific window, that position in the difference matrix is ​​marked as missing and not immediately replaced with zero. Instead, it is recalculated as a valid dimension during feature extraction. If the variance of a certain dimension in a historically stable state is lower than the preset minimum threshold, it indicates that the parameter may be in a dead zone or have no effective information fluctuation. In this exceptional case, the inverse variance weighting rule is abandoned, and the contribution weight of that dimension to the similarity judgment is directly reduced to zero or the preset minimum value to avoid stable but informationless features drowning out effective features. If the business plan cycle changes, for example, from a 5-minute window to a 10-minute window, the difference extraction module reconstructs the matrix with the new window and does not directly splice data of different window lengths.

[0103] For example, in the incubator monitoring during the company's Saturday morning shift, although the absolute power value increased with the increase in production schedule, after subtracting from the ideal benchmark and extracting serialized features, the system focused on the evolutionary extension feature sequence that appeared successively at multiple time steps, rather than the increase in the absolute value of the parameter at isolated monitoring moments; at the same time, the theoretically damaged simulation sequence also showed a similar temporal correlation evolution structure; through this dual-track differential extraction, the overall magnitude increase caused by normal production schedule was largely eliminated, and the truly time-evolution-oriented abnormal deviation distribution characteristics were stably transmitted; the purpose of this step is to map and reconstruct the original multidimensional parallel complex time-series monitoring values ​​into a sequence carrier that can support trend shift comparison, thereby achieving accurate capture of continuous abnormal evolutionary morphological features;

[0104] In this embodiment, the cloud server further includes:

[0105] The coupled decision module is used to perform coupled calculations on the actual residual vector and the theoretical residual vector based on the dynamic time warping similarity and the temporal cross-correlation value, so as to obtain the normalized temporal similarity and generate the state decision result.

[0106] The status judgment results include alarm confirmation, alarm suppression, and pending review instructions;

[0107] When the normalized temporal similarity is higher than or equal to the preset high threshold, the coupled decision module outputs a confirmation alarm.

[0108] When the normalized temporal similarity is lower than or equal to the preset low threshold, the coupled decision module outputs a suppression alarm.

[0109] When the normalized temporal similarity is between the high threshold and the low threshold, the coupled decision module outputs a review instruction.

[0110] Among them, the high threshold and low threshold are thresholds obtained based on historical work order samples.

[0111] This embodiment provides a coupled decision mechanism. Specifically, relying solely on the static comparison of actual and theoretical residuals has limitations in identification because the abnormal state evolution of cleanroom equipment is usually not completely synchronous. Some system abnormalities may first appear in energy consumption indicators and then deteriorate and propagate in indicators such as equipment recovery time. If only point-by-point alignment and equal value comparison is used, it is easy to cause misidentification or missed judgment when the residual points are similar in shape but have a small relative phase shift in time. Therefore, this embodiment introduces the coupled calculation of dynamic time warping similarity and time-series cross-correlation value, and outputs three types of results based on dual thresholds.

[0112] Specifically, dynamic time warping similarity measures whether two residual sequences satisfy the sequence morphology consistency attribute under the condition of allowing local time axis scaling, while the temporal cross-correlation value reflects the overall synchronization level of their extreme value distribution and evolution trend. Specifically, when calculating the normalized comprehensive temporal similarity index, the system first uses the dynamic time warping algorithm to find the optimal association mapping path with the shortest cumulative distance between the residual sequences, constructing a cumulative alignment distance variable. To reduce the influence of absolute distance numerical dispersion and improve the mathematical interpretability of engineering judgments, the system adopts negative exponential mapping or a linear inversion function based on the maximum permissible benchmark alignment distance. Negative exponential mapping is specifically expressed as... ,in, For the cumulative alignment distance variable, The attenuation coefficient is preset; the linear inversion function is specifically expressed as:

[0113]

[0114] in, The maximum permissible baseline alignment distance is used; this distance result is dimensionless and projected into the [0,1] interval to obtain the dynamic time-warped similarity component. ;

[0115] The system calculates the Pearson cross-correlation coefficient between two sets of vector sequences within the same time window, and denotes this coefficient as... To quantitatively characterize the linear convergent distribution characteristics of the peak and trough amplitudes of the sequence; then, a balance coefficient determined by statistical analysis of offline engineering business data is introduced. The combined temporal similarity is obtained by weighting and converting the two aforementioned components using this coefficient. Its decision model is as follows:

[0116]

[0117] Among them, the balance coefficient This reflects the degree of emphasis the judgment process places on the time-domain deformation delay tolerance and the sensitivity of waveform amplitude state synchronization; for example, assuming the dynamic time warping similarity components obtained through matrix optimization and mapping. The cross-correlation component is 0.88. The value is 0.84, and is determined based on historical latency tolerance configuration. The final comprehensive temporal similarity is obtained. The value is 0.86; if the high threshold obtained from historical work order samples is 0.80 and the low threshold is 0.55, then since 0.86 is higher than the high threshold, the system determines and outputs a confirmation alarm; if the comprehensive value is only 0.42 and lower than the low threshold, then a suppression alarm is output; if the evaluation result falls to 0.68, then a pending review instruction is triggered.

[0118] To verify the alignment logic mechanism, the actual residual sequence vector can be set as [0.4, 0.9, 1.0, 0.8], and the theoretical damaged residual sequence is set as [0.3, 0.8, 1.1, 0.9]. If the abnormal deviation of the equipment suddenly occurs during the W2 period and propagates to the subsequent time series, and the amplitude main envelope in the simulation model based on the parameter construction has a preset phase shift difference, the traditional direct comparison of the same order and the same position often causes a significant decay in the similarity evaluation score. However, in this implementation scheme, the dynamic time warping algorithm adaptively scales and compensates according to the mapping of the time dimension, and can still effectively align the primary peak bands between the sequences, thereby stably maintaining the benchmark score that represents the correlation morphological matching characteristics.

[0119] The above mathematical implementation steps enable the analysis and identification of abnormal features of the operating mechanism to still meet the preset algorithm robustness requirements under the measurement interference of small time-series offset noise. This not only ensures that the alarm monitoring feedback results are free from the limitations of a single static cross-line warning judgment, but also more accurately determines whether the actual signal sequence evolution trend matches the specific fault distribution time-domain context characteristics mapped by the accumulation of core hardware losses of the equipment.

[0120] In the anomaly handling process of coupled computation, when the same actual residual has a high similarity to multiple theoretical residuals, the system first calculates the difference between the highest matching score and the second-highest matching score. If the difference is greater than the identification threshold, since the similarity step difference has a strong confirmation guarantee, the system can identify the classification according to the highest distribution model attribute and trigger the output instruction.

[0121] If the difference fluctuation is within the critical range below the preset difference limit, the system will output a verification instruction and adjust the data categories that need to be enhanced in the adaptive supplementary sampling task sequence. For example, it may add low-cut-frequency vibration feedback records of related workstations or regional inspection identity positioning backtracking parameters to further help analyze the source of multiple cross-damage in the time-domain fuzzy superposition state. If the number of system operation cycle verification work orders covered by the prior calibration database is insufficient, resulting in deviations in the setting points of the upper and lower limit thresholds, the system will attach a robust segment allocation strategy with biased convergence during the reinitialization cycle to increase the adaptability of the confidence critical band range and increase the proportion of the sample set entering the verification sequence to avoid direct release processing before judgment. If the calculated similarity is high, but the corresponding data window has been verified to have a large number of pseudo signal points and jump garbled characters, belonging to the category set where the data quality is lower than the preset judgment threshold, the judgment system will not grant the confirmation and issuance authority, but will transfer to the subsequent manual data correction and review judgment process.

[0122] In this embodiment, the execution terminal includes:

[0123] The work order generation module is used to generate high-priority operation and maintenance work orders based on confirmed alarms, generate monitoring and maintenance instructions based on suppressed alarms, generate supplementary data collection tasks based on instructions pending review, and send high-priority operation and maintenance work orders, monitoring and maintenance instructions or supplementary data collection tasks to the field operation and maintenance terminal and the data collection device corresponding to the status acquisition module.

[0124] The result feedback module is used to send the work order execution results back to the cloud server.

[0125] The cloud server adjusts the weight parameters of the abnormal disturbance operator based on the status judgment result and the work order execution result.

[0126] Specifically, the cloud server adopts a preset step size update mechanism. When the work order execution result is consistent with the status judgment result, the prior weight of the corresponding abnormal disturbance operator is increased by a preset upward step size. When the execution result is inconsistent with the judgment result, the weight parameter of the corresponding operator is attenuated by a preset downward step size.

[0127] This embodiment provides an execution closed-loop and model correction mechanism. Specifically, if the system only stays at the alarm display level after the judgment, without converting the result into differentiated execution actions and without using on-site feedback to correct the model, it will be difficult to optimize and improve the front-end recognition capability, and repeated false alarms and repeated missed alarms will continue to occur. Therefore, this embodiment sets up a work order generation module and a result feedback module in the execution terminal to make the judgment result form a closed loop.

[0128] Specifically, the work order generation module takes different actions for different judgment results; confirmed alarms correspond to high-priority maintenance work orders, which may include information such as equipment identification, anomaly type suggestions, recommended inspection order, whether to recommend entering the room, and recommended materials to bring; suppressed alarms correspond to monitoring and maintenance instructions, requiring the site to maintain the current operation, continue sampling, and not intervene for the time being; pending review corresponds to supplementary sampling tasks, and the system can command the sampling equipment to increase the sampling frequency, add environmental auxiliary parameters, or request maintenance personnel to supplement operation evidence; the result feedback module receives results such as confirmed anomaly, no anomaly found, anomaly type corrected, parts replaced, and human error missed clocking in but maintenance was actually performed after the work order is executed, and feeds them back to the cloud;

[0129] Taking a specific work order execution process as an example; suppose a confirmed alarm is assigned as work order O1. After on-site inspection, it is confirmed that the filter component was not replaced according to the standard operating procedure, and the power returned to normal after replacement. The result is then reported as maintenance deficiency established. Based on this, the cloud increases the weight of the maintenance deficiency operator from 0.70 to 0.78 and increases the priority of work orders with rule combinations such as no material requisition + trajectory not entering the area + energy consumption tailing in the enterprise resource planning system. Another work order O2 is found on-site to be caused by a temporary change in production scheduling that was not synchronized, and there is actually no equipment problem. Therefore, the system lowers the weight of the relevant operators and reduces the tendency for similar condition combinations to directly trigger confirmed alarms. Through this feedback, the model will gradually become closer to the enterprise's own management characteristics.

[0130] As a fault-tolerant mechanism, if the field terminal goes offline after a high-priority work order is issued, the executing terminal will first forward the request to the backup terminal or the person in charge on duty and record the number of retries; if the acquisition equipment does not support remote parameter adjustment after the supplementary acquisition task is issued, it will automatically switch to manual review prompts to avoid the task being in an unresponsive or unexecuted state; if the work order execution result is not returned for a long time, the cloud will not perform positive learning on the sample and will only retain it as a sample to be confirmed to prevent incorrect model updates; if the deviation between the returned result and the original decision exceeds the preset tolerance threshold and occurs continuously, the system can trigger threshold recalibration or prompt the management personnel to review the rule configuration.

[0131] For example, in a night shift alarm at the company, the system issued a high-priority work order to the on-site maintenance terminal, requesting the inspection of the incubator filter component status and verification of the most recent standard operating procedure execution record. After on-site confirmation that a replacement was missed and the replacement was completed, the equipment power dropped back to near the baseline within the next two time windows. After the results were returned, the cloud increased the prior weight of this abnormal mode, enabling faster confirmation in similar situations in the future. The purpose of this step is to transform the judgment result into an executable action and use the execution consequences to continuously correct the simulation model, thereby achieving an adaptive improvement in maintenance identification capabilities.

[0132] In this embodiment, the business plan data is the dynamic production scheduling table of the manufacturing execution system, the resource ledger data is the maintenance material consumption ledger of the enterprise resource planning system, the maintenance rule data is the digital standard operating procedure rule data, the equipment status time series data is the time series of group operation parameters and equipment status parameter time series fluctuations of clean area equipment, and the maintenance behavior time series data is the trajectory check-in log of maintenance personnel.

[0133] This embodiment provides a mechanism for implementing data from specific business sources. Specifically, in the aforementioned solution, if the data is only abstractly described as business plan resource ledger rules data, although the logic is complete, it may still face problems such as unclear field mapping between different systems and unclear collection scope during engineering deployment. Therefore, this embodiment further places the main data sources on the manufacturing execution system, enterprise resource planning system, digital standard operating procedure system, and clean area collection and trajectory recording system.

[0134] Specifically, the Manufacturing Execution System's dynamic production scheduling table describes the allocation of production load in different time periods, such as which batch of cell culture tasks enters which cleanroom, which equipment is required, and how long it is expected to last; the Enterprise Resource Planning System's maintenance material consumption ledger describes the planned requisition and actual consumption of filter materials, seals, lubricants, etc.; digital standard operating procedure rule data describes maintenance cycles, standard working hours, allowable deviations, and execution sequence; equipment status time-series data includes not only single equipment energy consumption, temperature, humidity, pressure difference, vibration, etc., but also group-level operating parameter time-series; and maintenance behavior time-series data can come from personnel trajectory check-in logs, access control records, or location base station records.

[0135] For example, in a complete data chain, the Manufacturing Execution System shows that the second batch of cell culture will start at 14:00; the Enterprise Resource Planning System shows that the filter material for the corresponding incubator should have been issued the day before but was not issued; the Digital Standard Operating Procedures stipulate that the equipment must be replaced within 24 hours after it reaches 200 hours of operation; the equipment status timeline shows that the power increased and the recovery time was extended after 14:10; the tracking check-in shows that the maintenance personnel on duty did not enter the target clean area from 14:00 to 16:00; by combining this information, the system can not only extract the abnormal features of a single power parameter, but also build a multi-dimensional abnormal scenario model that relates business, management, materials and behaviors.

[0136] As a fault-tolerant mechanism, if the data record level of the material ledger in the enterprise resource planning system is higher than that of the equipment level, and can only record the requisition at the shift level rather than the equipment level, then the system can perform secondary splitting through the work order association number or material batch number; if the timing of maintenance behavior comes from access control rather than continuous positioning, and its time resolution is low, then the system will not treat non-entry into the area as absolute evidence, but only as a weight correction factor; if there are exemptions for manual approval in the digital standard operating procedures, then the corresponding exemption records need to be entered into the rule data simultaneously to avoid the system misjudging legitimate delays as violations;

[0137] For example, during deployment in this enterprise, the Manufacturing Execution System (MES), Enterprise Resource Planning (ERP), and Standard Operating Procedure (SOP) are provided by different vendors, and their field naming is inconsistent. Through the mapping in this embodiment, the equipment group number in production scheduling is converted into a unified equipment identifier, the material requisition form in the ERP is associated with the planned maintenance task in the SOP, and the trajectory check-in log is aligned to the equipment status curve according to the same time window. In this way, all subsequent theoretical modeling and anomaly judgment are based on specific and implementable data sources. The purpose of this step is to clarify the business entities and source boundaries of various input data, thereby achieving stable access of the solution in the enterprise's real information environment.

[0138] In this embodiment, the cleanroom equipment includes a biosafety cabinet, an incubator, or a laminar flow hood;

[0139] Among them, the management benchmark model generates the theoretical group resource consumption sequence of biosafety cabinets, incubators or clean benches based on the dynamic production scheduling table of the manufacturing execution system;

[0140] The coupling decision module is used to distinguish between load offsets caused by production schedule changes and abnormal offsets caused by maintenance deficiencies or equipment performance degradation, based on the similarity between the theoretical group resource consumption time series and the actual residual vector.

[0141] This embodiment provides an implementation mechanism for typical equipment groups in clean areas. Specifically, in the aforementioned general solution, without further combining the group collaboration characteristics of equipment such as biosafety cabinets, incubators, and laminar flow hoods, the system may only be able to see fluctuations at the level of a single device, and it may be difficult to distinguish between the synchronous increase in the load of the entire clean area due to production schedule changes and the abnormal deviation of individual devices from the group behavior. Therefore, this embodiment further extends the management benchmarks and decisions to the group level.

[0142] Specifically, biosafety cabinets, incubators, and laminar flow hoods often operate in tandem around the same batch of production tasks. Once the manufacturing execution system adjusts the production schedule, the theoretical resource consumption of the three equipment groups usually changes in tandem. For example, incubators may continuously increase their load first, laminar flow hoods may increase their load briefly during the operating window, and biosafety cabinets may increase their load intermittently during sampling or transfer phases. The management baseline model can generate a theoretical group resource consumption time series for each type of equipment based on the production schedule, and then synthesize it into a total theoretical resource consumption time series at the clean area level. In this way, when the total energy consumption increases within a certain period of time, the system first determines whether this increase is consistent with the group linkage driven by the production schedule. If it is consistent, it is more likely to be a normal load deviation. If only one incubator deviates significantly from the same group of equipment while other equipment meets the production schedule expectations, it is more likely to be an equipment malfunction or maintenance problem.

[0143] To illustrate with a micro-level example: Suppose the theoretical group power time series for three types of equipment in the afternoon are as follows: incubator group [10,12,12], clean bench group [3,5,3], and biosafety cabinet group [2,3,2]. Then the total theoretical time series for the clean area is [15,20,17]; the actual observed total power is [15.2,20.4,17.1]. ​​The macro-level group performance matches, but when reduced to the level of individual equipment, the actual power of a certain incubator changes from the theoretical 5.0, 5.0, 5.0 to 5.6, 6.2, 6.3, while the power of other incubators in the same group remains basically consistent with the theory. At this point, the system's troubleshooting mechanism is not limited to the overall group matching status, but will further compare the actual residual of that equipment with the theoretical damaged residual after the business plan is corrected. If the two are highly similar, it is judged as an abnormal offset; if all equipment increases synchronously with the group theory, it is more likely that the load offset is caused by changes in production scheduling.

[0144] As a fault-tolerant mechanism, if a certain type of equipment does not participate in production tasks within a specific time window, such as a clean bench being in standby mode, its theoretical group resource consumption is treated as the standby baseline and is not forcibly included in high-load comparisons; if a piece of equipment in the group is not included in the statistics due to planned shutdown, the theoretical time sequence of the group will synchronously remove that equipment to avoid false anomalies caused by changes in quantity; if there is a contradiction between the total group quantity and the quantity of a single machine, such as a small deviation in the total quantity but multiple single machines having opposite deviation directions, the system can prioritize transferring it to pending review and prompting a check for metering mapping errors or equipment number confusion.

[0145] For example, when a production batch changes at the company, the manufacturing execution system arranges for a short-term coordinated increase in the operation of a clean bench and a biosafety cabinet, causing a sudden increase in the total load of the clean area. The system identifies this linkage as a normal load shift based on the theoretical group timing after the production scheduling correction, and therefore does not issue a work order.

[0146] Another incubator, although under the same production scheduling background, consistently exceeded the theoretical curve of the same group and was highly coupled with the theoretical residual of maintenance deficiency, and was eventually identified as an anomaly. The purpose of this step is to identify the source of the deviation at both the equipment group and individual equipment levels, so as to effectively distinguish between production scheduling fluctuations and actual operation and maintenance anomalies.

[0147] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention.

Claims

1. A cloud-based cleanroom equipment operation and maintenance system, characterized in that, include: The data access terminal is used to collect business plan data, operation and maintenance rule data, resource ledger data, equipment status time series data and operation and maintenance behavior time series data, and upload them to the cloud server according to a preset unified time window as a unified time benchmark. The cloud server is used to build a management baseline model based on business plan data and operation and maintenance rule data. The management baseline model includes the theoretical resource consumption time series and management baseline parameters of cleanroom equipment when it is in rated performance state and meets the preset operation and maintenance rules. Based on the resource ledger data, abnormal disturbance parameters and abnormal disturbance operators are determined, and the abnormal disturbance operators are used to simulate the management scenario disturbance on the management baseline model to generate theoretical damage data; Based on a preset unified time window, the equipment status time series data and the management benchmark model, and the theoretical damage data and the management benchmark model are respectively differentially processed to form a difference matrix, and the actual residual vector and the theoretical residual vector are generated through feature extraction. The state decision result is generated based on the dynamic time warping similarity and temporal cross-correlation value between the actual residual vector and the theoretical residual vector; The execution terminal is used to receive status judgment results and generate and issue corresponding operation and maintenance work orders, monitoring and maintenance instructions or supplementary data collection tasks, and obtain the work order execution results of the operation and maintenance work orders; The cloud server is also used to correct the weight parameters of the abnormal disturbance operator based on the status judgment result and the work order execution result, and to update the work order priority rules.

2. The cloud-based cleanroom equipment operation and maintenance system according to claim 1, characterized in that, Data access terminals include: The business interface module is used to receive business plan data; The rules interface module is used to receive operation and maintenance rules data; The ledger interface module is used to receive resource ledger data; The status acquisition module is used to receive device status time-series data and operation and maintenance behavior time-series data, and then send them to the cloud server after timestamp normalization and alignment according to the preset sampling period and unified time window.

3. The cloud-based cleanroom equipment operation and maintenance system according to claim 2, characterized in that, Cloud servers include: The benchmark building module is used to build management benchmark models based on business plan data and operation and maintenance rule data; Among them, the management benchmark model is used to characterize the theoretical operating benchmark when the equipment is in its rated performance state and meets the preset operation and maintenance rules; The theoretical operating benchmarks include theoretical resource consumption time series, theoretical material replacement frequency, and theoretical maintenance man-hour consumption.

4. The cloud-based cleanroom equipment operation and maintenance system according to claim 3, characterized in that, Cloud servers also include: The parameter injection module is used to call the abnormal disturbance operator based on the resource ledger data and inject the abnormal disturbance operator into the management benchmark model to generate theoretical damage data. Among them, the abnormal disturbance operators include the maintenance deficiency operator and the implicit aging operator; Among them, the maintenance missing operator is used to introduce time delay and amplitude offset relative to the theoretical benchmark into the response curve; Among them, the implicit aging operator is used to introduce discrete increments that deviate from the theoretical curve in the mapping relationship between load and operating parameters.

5. The cloud-based cleanroom equipment operation and maintenance system according to claim 4, characterized in that, Cloud servers also include: The differential extraction module is used to perform differential processing on the device status time series data and the management baseline model based on a unified time window corresponding to the business plan cycle, so as to generate the actual residual vector. The differential extraction module is also used to perform differential processing on the theoretically damaged data and the management benchmark model to generate a theoretical residual vector; The actual residual vector and the theoretical residual vector are both output vectors obtained by feature extraction from the difference matrix formed by differentiating the multidimensional business state time series at multiple time offsets.

6. The cloud-based cleanroom equipment operation and maintenance system according to claim 5, characterized in that, Cloud servers also include: The coupled decision module is used to perform coupled calculations on the actual residual vector and the theoretical residual vector based on the dynamic time warping similarity and the temporal cross-correlation value, so as to obtain the normalized temporal similarity and generate the state decision result. The status judgment results include alarm confirmation, alarm suppression, and pending review instructions; When the normalized temporal similarity is higher than or equal to the preset high threshold, the coupled decision module outputs a confirmation alarm. When the normalized temporal similarity is lower than or equal to the preset low threshold, the coupled decision module outputs a suppression alarm. When the normalized temporal similarity is between the high threshold and the low threshold, the coupled decision module outputs a review instruction. Among them, the high threshold and low threshold are thresholds obtained based on historical work order samples.

7. The cloud-based cleanroom equipment operation and maintenance system according to claim 6, characterized in that, The execution terminal includes: The work order generation module is used to generate high-priority operation and maintenance work orders based on confirmed alarms, generate monitoring and maintenance instructions based on suppressed alarms, generate supplementary data collection tasks based on instructions pending review, and send high-priority operation and maintenance work orders, monitoring and maintenance instructions or supplementary data collection tasks to the field operation and maintenance terminal and the data collection device corresponding to the status acquisition module. The result feedback module is used to send the work order execution results back to the cloud server. The cloud server adjusts the weight parameters of the abnormal disturbance operator based on the status judgment result and the work order execution result.

8. The cloud-based cleanroom equipment operation and maintenance system according to claim 7, characterized in that, Business plan data is the dynamic production scheduling table of the Manufacturing Execution System; resource ledger data is the maintenance material consumption ledger of the Enterprise Resource Planning System; maintenance rule data is the digital standard operating procedure rule data; equipment status time sequence data is the time sequence of group operation parameters and equipment status parameter time sequence fluctuations of clean area equipment; and maintenance behavior time sequence data is the trajectory check-in log of maintenance personnel.

9. The cloud-based cleanroom equipment operation and maintenance system according to claim 8, characterized in that, Cleanroom equipment includes biosafety cabinets, incubators, or laminar flow hoods; Among them, the management benchmark model generates the theoretical group resource consumption sequence of biosafety cabinets, incubators or clean benches based on the dynamic production scheduling table of the manufacturing execution system; The coupling decision module is used to distinguish between load offsets caused by production schedule changes and abnormal offsets caused by maintenance deficiencies or equipment performance degradation, based on the similarity between the theoretical group resource consumption time series and the actual residual vector.

Images